diff --git a/CODE_OF_CONDUCT.md b/CODE_OF_CONDUCT.md new file mode 100644 index 0000000000000000000000000000000000000000..a0cbeaab7650bf08267fbdbc9bb54e845c88f392 --- /dev/null +++ b/CODE_OF_CONDUCT.md @@ -0,0 +1,77 @@ +# Code of Conduct + +## Our Pledge + +In the interest of fostering an open and welcoming environment, we as +contributors and maintainers pledge to make participation in our project and +our community a harassment-free experience for everyone, regardless of age, body +size, disability, ethnicity, sex characteristics, gender identity and expression, +level of experience, education, socio-economic status, nationality, personal +appearance, race, religion, or sexual identity and orientation. + +## Our Standards + +Examples of behavior that contributes to creating a positive environment +include: + +* Using welcoming and inclusive language +* Being respectful of differing viewpoints and experiences +* Gracefully accepting constructive criticism +* Focusing on what is best for the community +* Showing empathy towards other community members + +Examples of unacceptable behavior by participants include: + +* The use of sexualized language or imagery and unwelcome sexual attention or + advances +* Trolling, insulting/derogatory comments, and personal or political attacks +* Public or private harassment +* Publishing others' private information, such as a physical or electronic + address, without explicit permission +* Other conduct which could reasonably be considered inappropriate in a + professional setting + +## Our Responsibilities + +Project maintainers are responsible for clarifying the standards of acceptable +behavior and are expected to take appropriate and fair corrective action in +response to any instances of unacceptable behavior. + +Project maintainers have the right and responsibility to remove, edit, or +reject comments, commits, code, wiki edits, issues, and other contributions +that are not aligned to this Code of Conduct, or to ban temporarily or +permanently any contributor for other behaviors that they deem inappropriate, +threatening, offensive, or harmful. + +## Scope + +This Code of Conduct applies within all project spaces, and it also applies when +an individual is representing the project or its community in public spaces. +Examples of representing a project or community include using an official +project e-mail address, posting via an official social media account, or acting +as an appointed representative at an online or offline event. Representation of +a project may be further defined and clarified by project maintainers. + +## Enforcement + +Instances of abusive, harassing, or otherwise unacceptable behavior may be +reported by contacting the project team at . All +complaints will be reviewed and investigated and will result in a response that +is deemed necessary and appropriate to the circumstances. The project team is +obligated to maintain confidentiality with regard to the reporter of an incident. +Further details of specific enforcement policies may be posted separately. + +Project maintainers who do not follow or enforce the Code of Conduct in good +faith may face temporary or permanent repercussions as determined by other +members of the project's leadership. + +## Attribution + +This Code of Conduct is adapted from the [Contributor Covenant][homepage], version 1.4, +available at https://www.contributor-covenant.org/version/1/4/code-of-conduct.html + +[homepage]: https://www.contributor-covenant.org + +For answers to common questions about this code of conduct, see +https://www.contributor-covenant.org/faq + diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md new file mode 100644 index 0000000000000000000000000000000000000000..4d7ca6a98ebdabd7a6770ea616ee355ffb4a41e1 --- /dev/null +++ b/CONTRIBUTING.md @@ -0,0 +1,28 @@ +# Contributing to Facebook AI Research Sequence-to-Sequence Toolkit (fairseq) +We want to make contributing to this project as easy and transparent as +possible. + +## Pull Requests +We actively welcome your pull requests. + +1. Fork the repo and create your branch from `master`. +2. If you've added code that should be tested, add tests. +3. If you've changed APIs, update the documentation. +4. Ensure the test suite passes. +5. Make sure your code lints. +6. If you haven't already, complete the Contributor License Agreement ("CLA"). + +## Contributor License Agreement ("CLA") +In order to accept your pull request, we need you to submit a CLA. You only need +to do this once to work on any of Facebook's open source projects. + +Complete your CLA here: + +## Issues +We use GitHub issues to track public bugs. Please ensure your description is +clear and has sufficient instructions to be able to reproduce the issue. + +## License +By contributing to Facebook AI Research Sequence-to-Sequence Toolkit (fairseq), +you agree that your contributions will be licensed under the LICENSE file in +the root directory of this source tree. diff --git a/LICENSE b/LICENSE new file mode 100644 index 0000000000000000000000000000000000000000..b96dcb0480a0b0be0727976e5202a1e7b23edc3f --- /dev/null +++ b/LICENSE @@ -0,0 +1,21 @@ +MIT License + +Copyright (c) Facebook, Inc. and its affiliates. + +Permission is hereby granted, free of charge, to any person obtaining a copy +of this software and associated documentation files (the "Software"), to deal +in the Software without restriction, including without limitation the rights +to use, copy, modify, merge, publish, distribute, sublicense, and/or sell +copies of the Software, and to permit persons to whom the Software is +furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all +copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE +SOFTWARE. diff --git a/app.py b/app.py new file mode 100644 index 0000000000000000000000000000000000000000..60c0281605082721b92f77dece4cd661052ac810 --- /dev/null +++ b/app.py @@ -0,0 +1,14 @@ +import gradio as gr + + + +description = "HuBERT: Self-Supervised Speech Representation Learning. To use it, simply add your audio or click one of the examples to load them. Read more at the links below." +article = "

HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units | Github Repo

" + +gr.Interface.load("huggingface/facebook/hubert-large-ls960-ft", + description=description, + article=article, + examples=[ + ["./audio1.mp3"], + ["./audio2.mp3"] +]).launch() diff --git a/audio1.mp3 b/audio1.mp3 new file mode 100644 index 0000000000000000000000000000000000000000..c9ae29453730edfdb2d1edf131b4417277c8235a Binary files /dev/null and b/audio1.mp3 differ diff --git a/audio2.mp3 b/audio2.mp3 new file mode 100644 index 0000000000000000000000000000000000000000..202e356991204ed5d228b7b88d501cc582efe40e Binary files /dev/null and b/audio2.mp3 differ diff --git a/docs/Makefile b/docs/Makefile new file mode 100644 index 0000000000000000000000000000000000000000..c2f5b1a89cfc9e02d1bb09027d9e1e520ba53d53 --- /dev/null +++ b/docs/Makefile @@ -0,0 +1,20 @@ +# Minimal makefile for Sphinx documentation +# + +# You can set these variables from the command line. +SPHINXOPTS = +SPHINXBUILD = python -msphinx +SPHINXPROJ = fairseq +SOURCEDIR = . +BUILDDIR = _build + +# Put it first so that "make" without argument is like "make help". +help: + @$(SPHINXBUILD) -M help "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O) + +.PHONY: help Makefile + +# Catch-all target: route all unknown targets to Sphinx using the new +# "make mode" option. $(O) is meant as a shortcut for $(SPHINXOPTS). +%: Makefile + @$(SPHINXBUILD) -M $@ "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O) \ No newline at end of file diff --git a/docs/_static/theme_overrides.css b/docs/_static/theme_overrides.css new file mode 100644 index 0000000000000000000000000000000000000000..2a0764193625e1a6fd66ff8af2ccdd0ad6369188 --- /dev/null +++ b/docs/_static/theme_overrides.css @@ -0,0 +1,9 @@ +.wy-table-responsive table td kbd { + white-space: nowrap; +} +.wy-table-responsive table td { + white-space: normal !important; +} +.wy-table-responsive { + overflow: visible !important; +} diff --git a/docs/command_line_tools.rst b/docs/command_line_tools.rst new file mode 100644 index 0000000000000000000000000000000000000000..c16300ff5cd42d9a6c0070c2d9bec3a802eacfad --- /dev/null +++ b/docs/command_line_tools.rst @@ -0,0 +1,85 @@ +.. _Command-line Tools: + +Command-line Tools +================== + +Fairseq provides several command-line tools for training and evaluating models: + +- :ref:`fairseq-preprocess`: Data pre-processing: build vocabularies and binarize training data +- :ref:`fairseq-train`: Train a new model on one or multiple GPUs +- :ref:`fairseq-generate`: Translate pre-processed data with a trained model +- :ref:`fairseq-interactive`: Translate raw text with a trained model +- :ref:`fairseq-score`: BLEU scoring of generated translations against reference translations +- :ref:`fairseq-eval-lm`: Language model evaluation + + +.. _fairseq-preprocess: + +fairseq-preprocess +~~~~~~~~~~~~~~~~~~ +.. automodule:: fairseq_cli.preprocess + + .. argparse:: + :module: fairseq.options + :func: get_preprocessing_parser + :prog: fairseq-preprocess + + +.. _fairseq-train: + +fairseq-train +~~~~~~~~~~~~~ +.. automodule:: fairseq_cli.train + + .. argparse:: + :module: fairseq.options + :func: get_training_parser + :prog: fairseq-train + + +.. _fairseq-generate: + +fairseq-generate +~~~~~~~~~~~~~~~~ +.. automodule:: fairseq_cli.generate + + .. argparse:: + :module: fairseq.options + :func: get_generation_parser + :prog: fairseq-generate + + +.. _fairseq-interactive: + +fairseq-interactive +~~~~~~~~~~~~~~~~~~~ +.. automodule:: fairseq_cli.interactive + + .. argparse:: + :module: fairseq.options + :func: get_interactive_generation_parser + :prog: fairseq-interactive + + +.. _fairseq-score: + +fairseq-score +~~~~~~~~~~~~~ +.. automodule:: fairseq_cli.score + + .. argparse:: + :module: fairseq_cli.score + :func: get_parser + :prog: fairseq-score + + +.. _fairseq-eval-lm: + +fairseq-eval-lm +~~~~~~~~~~~~~~~ +.. automodule:: fairseq_cli.eval_lm + + .. argparse:: + :module: fairseq.options + :func: get_eval_lm_parser + :prog: fairseq-eval-lm diff --git a/docs/conf.py b/docs/conf.py new file mode 100644 index 0000000000000000000000000000000000000000..440784bfae96c14e9050542b1b1921a75a3b4b27 --- /dev/null +++ b/docs/conf.py @@ -0,0 +1,134 @@ +#!/usr/bin/env python3 +# -*- coding: utf-8 -*- +# +# fairseq documentation build configuration file, created by +# sphinx-quickstart on Fri Aug 17 21:45:30 2018. +# +# This file is execfile()d with the current directory set to its +# containing dir. +# +# Note that not all possible configuration values are present in this +# autogenerated file. +# +# All configuration values have a default; values that are commented out +# serve to show the default. + +# If extensions (or modules to document with autodoc) are in another directory, +# add these directories to sys.path here. If the directory is relative to the +# documentation root, use os.path.abspath to make it absolute, like shown here. + +import os +import sys +from fairseq import __version__ + + +# source code directory, relative to this file, for sphinx-autobuild +sys.path.insert(0, os.path.abspath("..")) + +source_suffix = [".rst"] + +# -- General configuration ------------------------------------------------ + +# If your documentation needs a minimal Sphinx version, state it here. +# +# needs_sphinx = '1.0' + +# Add any Sphinx extension module names here, as strings. They can be +# extensions coming with Sphinx (named 'sphinx.ext.*') or your custom +# ones. +extensions = [ + "sphinx.ext.autodoc", + "sphinx.ext.intersphinx", + "sphinx.ext.viewcode", + "sphinx.ext.napoleon", + "sphinxarg.ext", +] + +# Add any paths that contain templates here, relative to this directory. +templates_path = ["_templates"] + +# The master toctree document. +master_doc = "index" + +# General information about the project. +project = "fairseq" +copyright = "Facebook AI Research (FAIR)" +author = "Facebook AI Research (FAIR)" + +github_doc_root = "https://github.com/pytorch/fairseq/tree/master/docs/" + +# The version info for the project you're documenting, acts as replacement for +# |version| and |release|, also used in various other places throughout the +# built documents. +# +# The short X.Y version. +version = __version__ +# The full version, including alpha/beta/rc tags. +release = __version__ + +# The language for content autogenerated by Sphinx. Refer to documentation +# for a list of supported languages. +# +# This is also used if you do content translation via gettext catalogs. +# Usually you set "language" from the command line for these cases. +language = None + +# List of patterns, relative to source directory, that match files and +# directories to ignore when looking for source files. +# This patterns also effect to html_static_path and html_extra_path +exclude_patterns = ["_build", "Thumbs.db", ".DS_Store"] + +# The name of the Pygments (syntax highlighting) style to use. +pygments_style = "sphinx" +highlight_language = "python" + +# If true, `todo` and `todoList` produce output, else they produce nothing. +todo_include_todos = False + + +# -- Options for HTML output ---------------------------------------------- + +# The theme to use for HTML and HTML Help pages. See the documentation for +# a list of builtin themes. +# +html_theme = "sphinx_rtd_theme" + +# Theme options are theme-specific and customize the look and feel of a theme +# further. For a list of options available for each theme, see the +# documentation. +# +# html_theme_options = {} + +# Add any paths that contain custom static files (such as style sheets) here, +# relative to this directory. They are copied after the builtin static files, +# so a file named "default.css" will overwrite the builtin "default.css". +html_static_path = ["_static"] + +html_context = { + "css_files": [ + "_static/theme_overrides.css", # override wide tables in RTD theme + ], +} + +# Custom sidebar templates, must be a dictionary that maps document names +# to template names. +# +# This is required for the alabaster theme +# refs: http://alabaster.readthedocs.io/en/latest/installation.html#sidebars +# html_sidebars = { +# '**': [ +# 'about.html', +# 'navigation.html', +# 'relations.html', # needs 'show_related': True theme option to display +# 'searchbox.html', +# 'donate.html', +# ] +# } + + +# Example configuration for intersphinx: refer to the Python standard library. +intersphinx_mapping = { + "numpy": ("http://docs.scipy.org/doc/numpy/", None), + "python": ("https://docs.python.org/", None), + "torch": ("https://pytorch.org/docs/master/", None), +} diff --git a/docs/criterions.rst b/docs/criterions.rst new file mode 100644 index 0000000000000000000000000000000000000000..d6b8ca6b671a32d0da4aca7b18626e0df58a7258 --- /dev/null +++ b/docs/criterions.rst @@ -0,0 +1,31 @@ +.. role:: hidden + :class: hidden-section + +.. _Criterions: + +Criterions +========== + +Criterions compute the loss function given the model and batch, roughly:: + + loss = criterion(model, batch) + +.. automodule:: fairseq.criterions + :members: + +.. autoclass:: fairseq.criterions.FairseqCriterion + :members: + :undoc-members: + +.. autoclass:: fairseq.criterions.adaptive_loss.AdaptiveLoss + :members: + :undoc-members: +.. autoclass:: fairseq.criterions.composite_loss.CompositeLoss + :members: + :undoc-members: +.. autoclass:: fairseq.criterions.cross_entropy.CrossEntropyCriterion + :members: + :undoc-members: +.. autoclass:: fairseq.criterions.label_smoothed_cross_entropy.LabelSmoothedCrossEntropyCriterion + :members: + :undoc-members: diff --git a/docs/data.rst b/docs/data.rst new file mode 100644 index 0000000000000000000000000000000000000000..6a390cb336ab3c5fb28edec7448abc35a8e22bbb --- /dev/null +++ b/docs/data.rst @@ -0,0 +1,58 @@ +.. role:: hidden + :class: hidden-section + +.. module:: fairseq.data + +Data Loading and Utilities +========================== + +.. _datasets: + +Datasets +-------- + +**Datasets** define the data format and provide helpers for creating +mini-batches. + +.. autoclass:: fairseq.data.FairseqDataset + :members: +.. autoclass:: fairseq.data.LanguagePairDataset + :members: +.. autoclass:: fairseq.data.MonolingualDataset + :members: + +**Helper Datasets** + +These datasets wrap other :class:`fairseq.data.FairseqDataset` instances and +provide additional functionality: + +.. autoclass:: fairseq.data.BacktranslationDataset + :members: +.. autoclass:: fairseq.data.ConcatDataset + :members: +.. autoclass:: fairseq.data.ResamplingDataset + :members: +.. autoclass:: fairseq.data.RoundRobinZipDatasets + :members: +.. autoclass:: fairseq.data.TransformEosDataset + :members: + + +Dictionary +---------- + +.. autoclass:: fairseq.data.Dictionary + :members: + + +Iterators +--------- + +.. autoclass:: fairseq.data.CountingIterator + :members: +.. autoclass:: fairseq.data.EpochBatchIterator + :members: +.. autoclass:: fairseq.data.GroupedIterator + :members: +.. autoclass:: fairseq.data.ShardedIterator + :members: diff --git a/docs/docutils.conf b/docs/docutils.conf new file mode 100644 index 0000000000000000000000000000000000000000..526acffd32d16217160aee917db2b120354f20f0 --- /dev/null +++ b/docs/docutils.conf @@ -0,0 +1,2 @@ +[writers] +option-limit=0 diff --git a/docs/fairseq.gif b/docs/fairseq.gif new file mode 100644 index 0000000000000000000000000000000000000000..5782fdbc7e0014564725c3ad0fc6be5c6bcd9983 Binary files /dev/null and b/docs/fairseq.gif differ diff --git a/docs/fairseq_logo.png b/docs/fairseq_logo.png new file mode 100644 index 0000000000000000000000000000000000000000..75472cbb5ff78acc8716ad9121ed421f17f96c9a Binary files /dev/null and b/docs/fairseq_logo.png differ diff --git a/docs/getting_started.rst b/docs/getting_started.rst new file mode 100644 index 0000000000000000000000000000000000000000..745ad7763cee67a8dec25bdd7ba7b79cbe0b7754 --- /dev/null +++ b/docs/getting_started.rst @@ -0,0 +1,216 @@ +Evaluating Pre-trained Models +============================= + +First, download a pre-trained model along with its vocabularies: + +.. code-block:: console + + > curl https://dl.fbaipublicfiles.com/fairseq/models/wmt14.v2.en-fr.fconv-py.tar.bz2 | tar xvjf - + +This model uses a `Byte Pair Encoding (BPE) +vocabulary `__, so we'll have to apply +the encoding to the source text before it can be translated. This can be +done with the +`apply\_bpe.py `__ +script using the ``wmt14.en-fr.fconv-cuda/bpecodes`` file. ``@@`` is +used as a continuation marker and the original text can be easily +recovered with e.g. ``sed s/@@ //g`` or by passing the ``--remove-bpe`` +flag to :ref:`fairseq-generate`. Prior to BPE, input text needs to be tokenized +using ``tokenizer.perl`` from +`mosesdecoder `__. + +Let's use :ref:`fairseq-interactive` to generate translations interactively. +Here, we use a beam size of 5 and preprocess the input with the Moses +tokenizer and the given Byte-Pair Encoding vocabulary. It will automatically +remove the BPE continuation markers and detokenize the output. + +.. code-block:: console + + > MODEL_DIR=wmt14.en-fr.fconv-py + > fairseq-interactive \ + --path $MODEL_DIR/model.pt $MODEL_DIR \ + --beam 5 --source-lang en --target-lang fr \ + --tokenizer moses \ + --bpe subword_nmt --bpe-codes $MODEL_DIR/bpecodes + | loading model(s) from wmt14.en-fr.fconv-py/model.pt + | [en] dictionary: 44206 types + | [fr] dictionary: 44463 types + | Type the input sentence and press return: + Why is it rare to discover new marine mammal species? + S-0 Why is it rare to discover new marine mam@@ mal species ? + H-0 -0.0643349438905716 Pourquoi est-il rare de découvrir de nouvelles espèces de mammifères marins? + P-0 -0.0763 -0.1849 -0.0956 -0.0946 -0.0735 -0.1150 -0.1301 -0.0042 -0.0321 -0.0171 -0.0052 -0.0062 -0.0015 + +This generation script produces three types of outputs: a line prefixed +with *O* is a copy of the original source sentence; *H* is the +hypothesis along with an average log-likelihood; and *P* is the +positional score per token position, including the +end-of-sentence marker which is omitted from the text. + +Other types of output lines you might see are *D*, the detokenized hypothesis, +*T*, the reference target, *A*, alignment info, *E* the history of generation steps. + +See the `README `__ for a +full list of pre-trained models available. + +Training a New Model +==================== + +The following tutorial is for machine translation. For an example of how +to use Fairseq for other tasks, such as :ref:`language modeling`, please see the +``examples/`` directory. + +Data Pre-processing +------------------- + +Fairseq contains example pre-processing scripts for several translation +datasets: IWSLT 2014 (German-English), WMT 2014 (English-French) and WMT +2014 (English-German). To pre-process and binarize the IWSLT dataset: + +.. code-block:: console + + > cd examples/translation/ + > bash prepare-iwslt14.sh + > cd ../.. + > TEXT=examples/translation/iwslt14.tokenized.de-en + > fairseq-preprocess --source-lang de --target-lang en \ + --trainpref $TEXT/train --validpref $TEXT/valid --testpref $TEXT/test \ + --destdir data-bin/iwslt14.tokenized.de-en + +This will write binarized data that can be used for model training to +``data-bin/iwslt14.tokenized.de-en``. + +Training +-------- + +Use :ref:`fairseq-train` to train a new model. Here a few example settings that work +well for the IWSLT 2014 dataset: + +.. code-block:: console + + > mkdir -p checkpoints/fconv + > CUDA_VISIBLE_DEVICES=0 fairseq-train data-bin/iwslt14.tokenized.de-en \ + --optimizer nag --lr 0.25 --clip-norm 0.1 --dropout 0.2 --max-tokens 4000 \ + --arch fconv_iwslt_de_en --save-dir checkpoints/fconv + +By default, :ref:`fairseq-train` will use all available GPUs on your machine. Use the +``CUDA_VISIBLE_DEVICES`` environment variable to select specific GPUs and/or to +change the number of GPU devices that will be used. + +Also note that the batch size is specified in terms of the maximum +number of tokens per batch (``--max-tokens``). You may need to use a +smaller value depending on the available GPU memory on your system. + +Generation +---------- + +Once your model is trained, you can generate translations using +:ref:`fairseq-generate` **(for binarized data)** or +:ref:`fairseq-interactive` **(for raw text)**: + +.. code-block:: console + + > fairseq-generate data-bin/iwslt14.tokenized.de-en \ + --path checkpoints/fconv/checkpoint_best.pt \ + --batch-size 128 --beam 5 + | [de] dictionary: 35475 types + | [en] dictionary: 24739 types + | data-bin/iwslt14.tokenized.de-en test 6750 examples + | model fconv + | loaded checkpoint trainings/fconv/checkpoint_best.pt + S-721 danke . + T-721 thank you . + ... + +To generate translations with only a CPU, use the ``--cpu`` flag. BPE +continuation markers can be removed with the ``--remove-bpe`` flag. + +Advanced Training Options +========================= + +Large mini-batch training with delayed updates +---------------------------------------------- + +The ``--update-freq`` option can be used to accumulate gradients from +multiple mini-batches and delay updating, creating a larger effective +batch size. Delayed updates can also improve training speed by reducing +inter-GPU communication costs and by saving idle time caused by variance +in workload across GPUs. See `Ott et al. +(2018) `__ for more details. + +To train on a single GPU with an effective batch size that is equivalent +to training on 8 GPUs: + +.. code-block:: console + + > CUDA_VISIBLE_DEVICES=0 fairseq-train --update-freq 8 (...) + +Training with half precision floating point (FP16) +-------------------------------------------------- + +.. note:: + + FP16 training requires a Volta GPU and CUDA 9.1 or greater + +Recent GPUs enable efficient half precision floating point computation, +e.g., using `Nvidia Tensor Cores +`__. +Fairseq supports FP16 training with the ``--fp16`` flag: + +.. code-block:: console + + > fairseq-train --fp16 (...) + +Distributed training +-------------------- + +Distributed training in fairseq is implemented on top of ``torch.distributed``. +The easiest way to launch jobs is with the `torch.distributed.launch +`__ tool. + +For example, to train a large English-German Transformer model on 2 nodes each +with 8 GPUs (in total 16 GPUs), run the following command on each node, +replacing ``node_rank=0`` with ``node_rank=1`` on the second node and making +sure to update ``--master_addr`` to the IP address of the first node: + +.. code-block:: console + + > python -m torch.distributed.launch --nproc_per_node=8 \ + --nnodes=2 --node_rank=0 --master_addr="192.168.1.1" \ + --master_port=12345 \ + $(which fairseq-train) data-bin/wmt16_en_de_bpe32k \ + --arch transformer_vaswani_wmt_en_de_big --share-all-embeddings \ + --optimizer adam --adam-betas '(0.9, 0.98)' --clip-norm 0.0 \ + --lr-scheduler inverse_sqrt --warmup-init-lr 1e-07 --warmup-updates 4000 \ + --lr 0.0005 \ + --dropout 0.3 --weight-decay 0.0 --criterion label_smoothed_cross_entropy --label-smoothing 0.1 \ + --max-tokens 3584 \ + --max-epoch 70 \ + --fp16 + +On SLURM clusters, fairseq will automatically detect the number of nodes and +GPUs, but a port number must be provided: + +.. code-block:: console + + > salloc --gpus=16 --nodes 2 (...) + > srun fairseq-train --distributed-port 12345 (...). + +Sharding very large datasets +---------------------------- + +It can be challenging to train over very large datasets, particularly if your +machine does not have much system RAM. Most tasks in fairseq support training +over "sharded" datasets, in which the original dataset has been preprocessed +into non-overlapping chunks (or "shards"). + +For example, instead of preprocessing all your data into a single "data-bin" +directory, you can split the data and create "data-bin1", "data-bin2", etc. +Then you can adapt your training command like so: + +.. code-block:: console + + > fairseq-train data-bin1:data-bin2:data-bin3 (...) + +Training will now iterate over each shard, one by one, with each shard +corresponding to an "epoch", thus reducing system memory usage. diff --git a/docs/hydra_integration.md b/docs/hydra_integration.md new file mode 100644 index 0000000000000000000000000000000000000000..6a15298382a6a16dfc4c5a4a812ea1cd0477ed52 --- /dev/null +++ b/docs/hydra_integration.md @@ -0,0 +1,284 @@ +## Hydra + +[Hydra](https://github.com/facebookresearch/hydra) is an open-source Python +framework that simplifies the development of research and other complex +applications. The key feature is the ability to dynamically create a +hierarchical configuration by composition and override it through config files +and the command line. The name Hydra comes from its ability to run multiple +similar jobs - much like a Hydra with multiple heads. + +## Motivation + +Until recently, all components in fairseq were configured through a shared +`args` namespace that was created at application startup. Components declared +their own `add_args` method to update the argparse parser, hoping that the names +would not clash with arguments from other components. While this model works for +smaller applications, as fairseq grew and became integrated into other +applications, this became problematic. In order to determine how to configure +each component, one needed to a) examine what args were added by this component, +and b) read the code to figure out what shared arguments it is using that were +added in other places. Reproducing models involved sharing commands that often +contained dozens of command line switches. + +The model described above is still supported by fairseq for backward +compatibility, but will be deprecated some time in the future. + +New components in fairseq should now create a dataclass that encapsulates all +parameters required to configure this component. The dataclass is registered +along with the component, and fairseq takes care of constructing and providing +this configuration object to the component's constructor. Note that sharing +parameters can optionally still work, but one has to explicitly point to the +"source of truth" (see inheritance example below). These changes make components +in fairseq more independent and re-usable by other applications: all that is +needed to create a component is to initialize its dataclass and overwrite some +of the defaults. + +While configuring fairseq through command line (using either the legacy argparse +based or the new Hydra based entry points) is still fully supported, you can now +take advantage of configuring fairseq completely or piece-by-piece through +hierarchical YAML configuration files. These files can also be shipped as +examples that others can use to run an identically configured job. + +Additionally, Hydra has a rich and growing [library of +plugins](https://github.com/facebookresearch/hydra/tree/master/plugins) that +provide functionality such as hyperparameter sweeping (including using bayesian +optimization through the [Ax](https://github.com/facebook/Ax) library), job +launching across various platforms, and more. + +## Creating or migrating components + +In general, each new (or updated) component should provide a companion +[dataclass](https://www.python.org/dev/peps/pep-0557/). These dataclass are +typically located in the same file as the component and are passed as arguments +to the `register_*()` functions. Top-level configs that should be present in +every fairseq application are placed in the +[global](fairseq/dataclass/configs.py) config file and added to the +`FairseqConfig` object. + +Each dataclass is a plain-old-data object, similar to a `NamedTuple`. These +classes are decorated with a `@dataclass` decorator, and typically inherit from +`FairseqDataclass` (which adds some functionality for backward compatibility). +Each field must have a type, and generally has metadata (such as a help string) +and a default value. Only primitive types or other config objects are allowed as +data types for each field. + +#### Example: + +```python +from dataclasses import dataclass, field +from fairseq.dataclass import FairseqDataclass + +@dataclass +class InteractiveConfig(FairseqDataclass): + buffer_size: int = field( + default=0, + metadata={ + "help": "read this many sentences into a buffer before processing them" + }, + ) + input: str = field( + default="-", + metadata={"help": "file to read from; use - for stdin"}, + ) +``` + +### Inherting values + +Some components require sharing a value. For example, a learning rate scheduler +and an optimizer may both need to know the initial learning rate value. One can +declare a field that, by default, will inherit its value from another config +node in the same hierarchy: + +```python +@dataclass +FairseqAdamConfig(FairseqDataclass): + ... + lr: List[float] = II("optimization.lr") + ... +``` + +`II("optimization.lr")` is syntactic sugar for `"${optimization.lr}"`, which is +the value one can use in a YAML config file or through command line to achieve +the same effect. Note that this assumes that there is an "optimization" config +object in the root config and it has a field called "lr". + +### Tasks and Models + +Creating Tasks and Models works same as before, except that legacy +implementations now inherit from `LegacyFairseq*` base classes, while new +components inherit from `FairseqTask` and `FairseqModel` and provide a dataclass +to the `register_*()` functions. + +#### Task example: + +```python +@dataclass +class LanguageModelingConfig(FairseqDataclass): + data: Optional[str] = field( + default=None, metadata={"help": "path to data directory"} + ) + ... + +@register_task("language_modeling", dataclass=LanguageModelingConfig) +class LanguageModelingTask(FairseqTask): + ... + @classmethod + def setup_task(cls, cfg: LanguageModelingConfig): + ... +``` + +#### Model example: + +```python +@dataclass +class TransformerLanguageModelConfig(FairseqDataclass): + activation_fn: ChoiceEnum(utils.get_available_activation_fns()) = field( + default="relu", metadata={"help": "activation function to use"} + ) + dropout: float = field(default=0.1, metadata={"help": "dropout probability"}) + ... + +@register_model("transformer_lm", dataclass=TransformerLanguageModelConfig) +class TransformerLanguageModel(FairseqLanguageModel): + ... + @classmethod + def build_model(cls, cfg: TransformerLanguageModelConfig, task: FairseqTask): + ... +``` + +### Other components + +Other components work as before, but they now take their configuration dataclass +as the only constructor argument: + +```python +@dataclass +class MosesTokenizerConfig(FairseqDataclass): + source_lang: str = field(default="en", metadata={"help": "source language"}) + ... + +@register_tokenizer("moses", dataclass=MosesTokenizerConfig) +class MosesTokenizer(object): + def __init__(self, cfg: MosesTokenizerConfig): + ... +``` + +Note that if you are adding a new registry for a new set of components, you need +to add it to the `FairseqConfig` object in `fairseq/dataclass/configs.py`: + +```python +@dataclass +class FairseqConfig(object): + ... + my_new_registry: Any = None +``` + +## Training with `fairseq-hydra-train` + +To fully take advantage of configuration flexibility offered by Hydra, you may +want to train new models using the `fairseq-hydra-train` entry point. Legacy CLI +tools such as `fairseq-train` will remain supported for the foreseeable future +but will be deprecated eventually. + +On startup, Hydra will create a configuration object that contains a hierarchy +of all the necessary dataclasses populated with their default values in the +code. The default values are overwritten by values found in YAML files in +`fairseq/config` directory (which currently sets minimal defaults) and then +further overwritten by values provided through command line arguments. + +Some of the most common use cases are shown below: + +### 1. Override default values through command line: + +```shell script +$ fairseq-hydra-train \ + distributed_training.distributed_world_size=1 \ + dataset.batch_size=2 \ + task.data=data-bin \ + model=transformer_lm/transformer_lm_gpt \ + task=language_modeling \ + optimization.max_update=5000 +``` + +Note that along with explicitly providing values for parameters such as +`dataset.batch_size`, this also tells Hydra to overlay configuration found in +`fairseq/config/model/transformer_lm/transformer_lm_gpt.yaml` over the default +values in the dataclass. If you want to train a model without specifying a +particular architecture you can simply specify `model=transformer_lm`. This only +works for migrated tasks and models. + +### 2. Replace bundled configs with an external config: + +```shell script +$ fairseq-hydra-train \ + --config-dir /path/to/external/configs \ + --config-name wiki103 +``` + +where `/path/to/external/configs/wiki103.yaml` contains: + +```yaml +# @package _group_ + +model: + _name: transformer_lm +distributed_training: + distributed_world_size: 1 +dataset: + batch_size: 2 +task: + _name: language_modeling + data: /path/to/data + add_bos_token: false + max_target_positions: 1024 +optimization: + max_update: 50000 + lr: [ 0.25 ] +criterion: cross_entropy +optimizer: adam +lr_scheduler: + _name: cosine +``` + +Note that here bundled configs from `fairseq/config` directory are not used, +however the defaults from each dataclass will still be used (unless overwritten +by your external config). + +Additionally you can choose to break up your configs by creating a directory +structure in the same location as your main config file, with the names of the +top-level fields (such as "model", "dataset", etc), and placing config files +with meaningful names that would populate that specific section of your +top-level config file (for example, you might have +`model/small_transformer_lm.yaml`, `model/big_transformer_lm.yaml`, etc). You +can then specify the correct configuration via command line, defaults in the +main config, or even launch all of them as a sweep (see Hydra documentation on +how to do this). + +### 3. Add an external config directory to Hydra search path: + +This allows combining default configuration (including using any bundled config +files), while specifying your own config files for some parts of the +configuration. + +```shell script +$ fairseq-hydra-train \ + distributed_training.distributed_world_size=1 \ + dataset.batch_size=2 \ + task.data=/path/to/data/ \ + model=transformer_lm/2_layers \ + task=language_modeling \ + optimization.max_update=5000 \ + --config-dir /path/to/external/configs +``` + +where `/path/to/external/configs` has the following structure: +``` +. ++-- model +| +-- transformer_lm +| | +-- 2_layers.yaml +``` + +and `2_layers.yaml` contains a copy of `transformer_lm_gpt.yaml` but with +`decoder_layers` set to 2. You can add other configs to configure other +components as well. diff --git a/docs/index.rst b/docs/index.rst new file mode 100644 index 0000000000000000000000000000000000000000..591db86cdf49e6f0a7a6686df2150f11418e90d0 --- /dev/null +++ b/docs/index.rst @@ -0,0 +1,49 @@ +.. fairseq documentation master file, created by + sphinx-quickstart on Fri Aug 17 21:45:30 2018. + You can adapt this file completely to your liking, but it should at least + contain the root `toctree` directive. + +:github_url: https://github.com/pytorch/fairseq + + +fairseq documentation +===================== + +Fairseq is a sequence modeling toolkit written in `PyTorch +`_ that allows researchers and developers to +train custom models for translation, summarization, language modeling and other +text generation tasks. + +.. toctree:: + :maxdepth: 1 + :caption: Getting Started + + getting_started + command_line_tools + +.. toctree:: + :maxdepth: 1 + :caption: Extending Fairseq + + overview + tutorial_simple_lstm + tutorial_classifying_names + +.. toctree:: + :maxdepth: 2 + :caption: Library Reference + + tasks + models + criterions + optim + lr_scheduler + data + modules + + +Indices and tables +================== + +* :ref:`genindex` +* :ref:`search` diff --git a/docs/lr_scheduler.rst b/docs/lr_scheduler.rst new file mode 100644 index 0000000000000000000000000000000000000000..bbc09dc22e6a7ac05137954e0b9c80ca030f62f4 --- /dev/null +++ b/docs/lr_scheduler.rst @@ -0,0 +1,34 @@ +.. role:: hidden + :class: hidden-section + +.. _Learning Rate Schedulers: + +Learning Rate Schedulers +======================== + +Learning Rate Schedulers update the learning rate over the course of training. +Learning rates can be updated after each update via :func:`step_update` or at +epoch boundaries via :func:`step`. + +.. automodule:: fairseq.optim.lr_scheduler + :members: + +.. autoclass:: fairseq.optim.lr_scheduler.FairseqLRScheduler + :members: + :undoc-members: + +.. autoclass:: fairseq.optim.lr_scheduler.cosine_lr_scheduler.CosineSchedule + :members: + :undoc-members: +.. autoclass:: fairseq.optim.lr_scheduler.fixed_schedule.FixedSchedule + :members: + :undoc-members: +.. autoclass:: fairseq.optim.lr_scheduler.inverse_square_root_schedule.InverseSquareRootSchedule + :members: + :undoc-members: +.. autoclass:: fairseq.optim.lr_scheduler.reduce_lr_on_plateau.ReduceLROnPlateau + :members: + :undoc-members: +.. autoclass:: fairseq.optim.lr_scheduler.triangular_lr_scheduler.TriangularSchedule + :members: + :undoc-members: diff --git a/docs/make.bat b/docs/make.bat new file mode 100644 index 0000000000000000000000000000000000000000..35c5085de318190514ee3b48d10060aa57a4fa50 --- /dev/null +++ b/docs/make.bat @@ -0,0 +1,36 @@ +@ECHO OFF + +pushd %~dp0 + +REM Command file for Sphinx documentation + +if "%SPHINXBUILD%" == "" ( + set SPHINXBUILD=python -msphinx +) +set SOURCEDIR=. +set BUILDDIR=_build +set SPHINXPROJ=fairseq + +if "%1" == "" goto help + +%SPHINXBUILD% >NUL 2>NUL +if errorlevel 9009 ( + echo. + echo.The Sphinx module was not found. Make sure you have Sphinx installed, + echo.then set the SPHINXBUILD environment variable to point to the full + echo.path of the 'sphinx-build' executable. Alternatively you may add the + echo.Sphinx directory to PATH. + echo. + echo.If you don't have Sphinx installed, grab it from + echo.http://sphinx-doc.org/ + exit /b 1 +) + +%SPHINXBUILD% -M %1 %SOURCEDIR% %BUILDDIR% %SPHINXOPTS% +goto end + +:help +%SPHINXBUILD% -M help %SOURCEDIR% %BUILDDIR% %SPHINXOPTS% + +:end +popd diff --git a/docs/models.rst b/docs/models.rst new file mode 100644 index 0000000000000000000000000000000000000000..054622d587c3b7f01f17f442919140755acd8f9e --- /dev/null +++ b/docs/models.rst @@ -0,0 +1,104 @@ +.. role:: hidden + :class: hidden-section + +.. module:: fairseq.models + +.. _Models: + +Models +====== + +A Model defines the neural network's ``forward()`` method and encapsulates all +of the learnable parameters in the network. Each model also provides a set of +named *architectures* that define the precise network configuration (e.g., +embedding dimension, number of layers, etc.). + +Both the model type and architecture are selected via the ``--arch`` +command-line argument. Once selected, a model may expose additional command-line +arguments for further configuration. + +.. note:: + + All fairseq Models extend :class:`BaseFairseqModel`, which in turn extends + :class:`torch.nn.Module`. Thus any fairseq Model can be used as a + stand-alone Module in other PyTorch code. + + +Convolutional Neural Networks (CNN) +----------------------------------- + +.. module:: fairseq.models.fconv +.. autoclass:: fairseq.models.fconv.FConvModel + :members: +.. autoclass:: fairseq.models.fconv.FConvEncoder + :members: + :undoc-members: +.. autoclass:: fairseq.models.fconv.FConvDecoder + :members: + + +Long Short-Term Memory (LSTM) networks +-------------------------------------- + +.. module:: fairseq.models.lstm +.. autoclass:: fairseq.models.lstm.LSTMModel + :members: +.. autoclass:: fairseq.models.lstm.LSTMEncoder + :members: +.. autoclass:: fairseq.models.lstm.LSTMDecoder + :members: + + +Transformer (self-attention) networks +------------------------------------- + +.. module:: fairseq.models.transformer +.. autoclass:: fairseq.models.transformer.TransformerModel + :members: +.. autoclass:: fairseq.models.transformer.TransformerEncoder + :members: +.. autoclass:: fairseq.models.transformer.TransformerEncoderLayer + :members: +.. autoclass:: fairseq.models.transformer.TransformerDecoder + :members: +.. autoclass:: fairseq.models.transformer.TransformerDecoderLayer + :members: + + +Adding new models +----------------- + +.. currentmodule:: fairseq.models +.. autofunction:: fairseq.models.register_model +.. autofunction:: fairseq.models.register_model_architecture +.. autoclass:: fairseq.models.BaseFairseqModel + :members: + :undoc-members: +.. autoclass:: fairseq.models.FairseqEncoderDecoderModel + :members: + :undoc-members: +.. autoclass:: fairseq.models.FairseqEncoderModel + :members: + :undoc-members: +.. autoclass:: fairseq.models.FairseqLanguageModel + :members: + :undoc-members: +.. autoclass:: fairseq.models.FairseqMultiModel + :members: + :undoc-members: +.. autoclass:: fairseq.models.FairseqEncoder + :members: +.. autoclass:: fairseq.models.CompositeEncoder + :members: +.. autoclass:: fairseq.models.FairseqDecoder + :members: + + +.. _Incremental decoding: + +Incremental decoding +-------------------- + +.. autoclass:: fairseq.models.FairseqIncrementalDecoder + :members: + :undoc-members: diff --git a/docs/modules.rst b/docs/modules.rst new file mode 100644 index 0000000000000000000000000000000000000000..9631c93d4682286e1cea1ddd961d3f6ab06f2589 --- /dev/null +++ b/docs/modules.rst @@ -0,0 +1,9 @@ +Modules +======= + +Fairseq provides several stand-alone :class:`torch.nn.Module` classes that may +be helpful when implementing a new :class:`~fairseq.models.BaseFairseqModel`. + +.. automodule:: fairseq.modules + :members: + :undoc-members: diff --git a/docs/optim.rst b/docs/optim.rst new file mode 100644 index 0000000000000000000000000000000000000000..c3326456bd9291a1d05bd3316bef5c9fb25c6c49 --- /dev/null +++ b/docs/optim.rst @@ -0,0 +1,38 @@ +.. role:: hidden + :class: hidden-section + +.. _optimizers: + +Optimizers +========== + +Optimizers update the Model parameters based on the gradients. + +.. automodule:: fairseq.optim + :members: + +.. autoclass:: fairseq.optim.FairseqOptimizer + :members: + :undoc-members: + +.. autoclass:: fairseq.optim.adadelta.Adadelta + :members: + :undoc-members: +.. autoclass:: fairseq.optim.adagrad.Adagrad + :members: + :undoc-members: +.. autoclass:: fairseq.optim.adafactor.FairseqAdafactor + :members: + :undoc-members: +.. autoclass:: fairseq.optim.adam.FairseqAdam + :members: + :undoc-members: +.. autoclass:: fairseq.optim.fp16_optimizer.FP16Optimizer + :members: + :undoc-members: +.. autoclass:: fairseq.optim.nag.FairseqNAG + :members: + :undoc-members: +.. autoclass:: fairseq.optim.sgd.SGD + :members: + :undoc-members: diff --git a/docs/overview.rst b/docs/overview.rst new file mode 100644 index 0000000000000000000000000000000000000000..026b3b5c7b21d071d8b8a3405898977c760d05b8 --- /dev/null +++ b/docs/overview.rst @@ -0,0 +1,74 @@ +Overview +======== + +Fairseq can be extended through user-supplied `plug-ins +`_. We support five kinds of +plug-ins: + +- :ref:`Models` define the neural network architecture and encapsulate all of the + learnable parameters. +- :ref:`Criterions` compute the loss function given the model outputs and targets. +- :ref:`Tasks` store dictionaries and provide helpers for loading/iterating over + Datasets, initializing the Model/Criterion and calculating the loss. +- :ref:`Optimizers` update the Model parameters based on the gradients. +- :ref:`Learning Rate Schedulers` update the learning rate over the course of + training. + +**Training Flow** + +Given a ``model``, ``criterion``, ``task``, ``optimizer`` and ``lr_scheduler``, +fairseq implements the following high-level training flow:: + + for epoch in range(num_epochs): + itr = task.get_batch_iterator(task.dataset('train')) + for num_updates, batch in enumerate(itr): + task.train_step(batch, model, criterion, optimizer) + average_and_clip_gradients() + optimizer.step() + lr_scheduler.step_update(num_updates) + lr_scheduler.step(epoch) + +where the default implementation for ``task.train_step`` is roughly:: + + def train_step(self, batch, model, criterion, optimizer, **unused): + loss = criterion(model, batch) + optimizer.backward(loss) + return loss + +**Registering new plug-ins** + +New plug-ins are *registered* through a set of ``@register`` function +decorators, for example:: + + @register_model('my_lstm') + class MyLSTM(FairseqEncoderDecoderModel): + (...) + +Once registered, new plug-ins can be used with the existing :ref:`Command-line +Tools`. See the Tutorial sections for more detailed walkthroughs of how to add +new plug-ins. + +**Loading plug-ins from another directory** + +New plug-ins can be defined in a custom module stored in the user system. In +order to import the module, and make the plugin available to *fairseq*, the +command line supports the ``--user-dir`` flag that can be used to specify a +custom location for additional modules to load into *fairseq*. + +For example, assuming this directory tree:: + + /home/user/my-module/ + └── __init__.py + +with ``__init__.py``:: + + from fairseq.models import register_model_architecture + from fairseq.models.transformer import transformer_vaswani_wmt_en_de_big + + @register_model_architecture('transformer', 'my_transformer') + def transformer_mmt_big(args): + transformer_vaswani_wmt_en_de_big(args) + +it is possible to invoke the :ref:`fairseq-train` script with the new architecture with:: + + fairseq-train ... --user-dir /home/user/my-module -a my_transformer --task translation diff --git a/docs/requirements.txt b/docs/requirements.txt new file mode 100644 index 0000000000000000000000000000000000000000..c734a1f04f1c108d84d3a07643ac93adf6485f13 --- /dev/null +++ b/docs/requirements.txt @@ -0,0 +1,2 @@ +sphinx<2.0 +sphinx-argparse diff --git a/docs/tasks.rst b/docs/tasks.rst new file mode 100644 index 0000000000000000000000000000000000000000..5f65c3c866865e50332d8e6ca012a4a81e7bea74 --- /dev/null +++ b/docs/tasks.rst @@ -0,0 +1,61 @@ +.. role:: hidden + :class: hidden-section + +.. module:: fairseq.tasks + +.. _Tasks: + +Tasks +===== + +Tasks store dictionaries and provide helpers for loading/iterating over +Datasets, initializing the Model/Criterion and calculating the loss. + +Tasks can be selected via the ``--task`` command-line argument. Once selected, a +task may expose additional command-line arguments for further configuration. + +Example usage:: + + # setup the task (e.g., load dictionaries) + task = fairseq.tasks.setup_task(args) + + # build model and criterion + model = task.build_model(args) + criterion = task.build_criterion(args) + + # load datasets + task.load_dataset('train') + task.load_dataset('valid') + + # iterate over mini-batches of data + batch_itr = task.get_batch_iterator( + task.dataset('train'), max_tokens=4096, + ) + for batch in batch_itr: + # compute the loss + loss, sample_size, logging_output = task.get_loss( + model, criterion, batch, + ) + loss.backward() + + +Translation +----------- + +.. autoclass:: fairseq.tasks.translation.TranslationTask + +.. _language modeling: + +Language Modeling +----------------- + +.. autoclass:: fairseq.tasks.language_modeling.LanguageModelingTask + + +Adding new tasks +---------------- + +.. autofunction:: fairseq.tasks.register_task +.. autoclass:: fairseq.tasks.FairseqTask + :members: + :undoc-members: diff --git a/docs/tutorial_classifying_names.rst b/docs/tutorial_classifying_names.rst new file mode 100644 index 0000000000000000000000000000000000000000..b02fec0489a86e7b1ccec481342fa4fbd93a80ae --- /dev/null +++ b/docs/tutorial_classifying_names.rst @@ -0,0 +1,415 @@ +Tutorial: Classifying Names with a Character-Level RNN +====================================================== + +In this tutorial we will extend fairseq to support *classification* tasks. In +particular we will re-implement the PyTorch tutorial for `Classifying Names with +a Character-Level RNN `_ +in fairseq. It is recommended to quickly skim that tutorial before beginning +this one. + +This tutorial covers: + +1. **Preprocessing the data** to create dictionaries. +2. **Registering a new Model** that encodes an input sentence with a simple RNN + and predicts the output label. +3. **Registering a new Task** that loads our dictionaries and dataset. +4. **Training the Model** using the existing command-line tools. +5. **Writing an evaluation script** that imports fairseq and allows us to + interactively evaluate our model on new inputs. + + +1. Preprocessing the data +------------------------- + +The original tutorial provides raw data, but we'll work with a modified version +of the data that is already tokenized into characters and split into separate +train, valid and test sets. + +Download and extract the data from here: +`tutorial_names.tar.gz `_ + +Once extracted, let's preprocess the data using the :ref:`fairseq-preprocess` +command-line tool to create the dictionaries. While this tool is primarily +intended for sequence-to-sequence problems, we're able to reuse it here by +treating the label as a "target" sequence of length 1. We'll also output the +preprocessed files in "raw" format using the ``--dataset-impl`` option to +enhance readability: + +.. code-block:: console + + > fairseq-preprocess \ + --trainpref names/train --validpref names/valid --testpref names/test \ + --source-lang input --target-lang label \ + --destdir names-bin --dataset-impl raw + +After running the above command you should see a new directory, +:file:`names-bin/`, containing the dictionaries for *inputs* and *labels*. + + +2. Registering a new Model +-------------------------- + +Next we'll register a new model in fairseq that will encode an input sentence +with a simple RNN and predict the output label. Compared to the original PyTorch +tutorial, our version will also work with batches of data and GPU Tensors. + +First let's copy the simple RNN module implemented in the `PyTorch tutorial +`_. +Create a new file named :file:`fairseq/models/rnn_classifier.py` with the +following contents:: + + import torch + import torch.nn as nn + + class RNN(nn.Module): + + def __init__(self, input_size, hidden_size, output_size): + super(RNN, self).__init__() + + self.hidden_size = hidden_size + + self.i2h = nn.Linear(input_size + hidden_size, hidden_size) + self.i2o = nn.Linear(input_size + hidden_size, output_size) + self.softmax = nn.LogSoftmax(dim=1) + + def forward(self, input, hidden): + combined = torch.cat((input, hidden), 1) + hidden = self.i2h(combined) + output = self.i2o(combined) + output = self.softmax(output) + return output, hidden + + def initHidden(self): + return torch.zeros(1, self.hidden_size) + +We must also *register* this model with fairseq using the +:func:`~fairseq.models.register_model` function decorator. Once the model is +registered we'll be able to use it with the existing :ref:`Command-line Tools`. + +All registered models must implement the :class:`~fairseq.models.BaseFairseqModel` +interface, so we'll create a small wrapper class in the same file and register +it in fairseq with the name ``'rnn_classifier'``:: + + from fairseq.models import BaseFairseqModel, register_model + + # Note: the register_model "decorator" should immediately precede the + # definition of the Model class. + + @register_model('rnn_classifier') + class FairseqRNNClassifier(BaseFairseqModel): + + @staticmethod + def add_args(parser): + # Models can override this method to add new command-line arguments. + # Here we'll add a new command-line argument to configure the + # dimensionality of the hidden state. + parser.add_argument( + '--hidden-dim', type=int, metavar='N', + help='dimensionality of the hidden state', + ) + + @classmethod + def build_model(cls, args, task): + # Fairseq initializes models by calling the ``build_model()`` + # function. This provides more flexibility, since the returned model + # instance can be of a different type than the one that was called. + # In this case we'll just return a FairseqRNNClassifier instance. + + # Initialize our RNN module + rnn = RNN( + # We'll define the Task in the next section, but for now just + # notice that the task holds the dictionaries for the "source" + # (i.e., the input sentence) and "target" (i.e., the label). + input_size=len(task.source_dictionary), + hidden_size=args.hidden_dim, + output_size=len(task.target_dictionary), + ) + + # Return the wrapped version of the module + return FairseqRNNClassifier( + rnn=rnn, + input_vocab=task.source_dictionary, + ) + + def __init__(self, rnn, input_vocab): + super(FairseqRNNClassifier, self).__init__() + + self.rnn = rnn + self.input_vocab = input_vocab + + # The RNN module in the tutorial expects one-hot inputs, so we can + # precompute the identity matrix to help convert from indices to + # one-hot vectors. We register it as a buffer so that it is moved to + # the GPU when ``cuda()`` is called. + self.register_buffer('one_hot_inputs', torch.eye(len(input_vocab))) + + def forward(self, src_tokens, src_lengths): + # The inputs to the ``forward()`` function are determined by the + # Task, and in particular the ``'net_input'`` key in each + # mini-batch. We'll define the Task in the next section, but for + # now just know that *src_tokens* has shape `(batch, src_len)` and + # *src_lengths* has shape `(batch)`. + bsz, max_src_len = src_tokens.size() + + # Initialize the RNN hidden state. Compared to the original PyTorch + # tutorial we'll also handle batched inputs and work on the GPU. + hidden = self.rnn.initHidden() + hidden = hidden.repeat(bsz, 1) # expand for batched inputs + hidden = hidden.to(src_tokens.device) # move to GPU + + for i in range(max_src_len): + # WARNING: The inputs have padding, so we should mask those + # elements here so that padding doesn't affect the results. + # This is left as an exercise for the reader. The padding symbol + # is given by ``self.input_vocab.pad()`` and the unpadded length + # of each input is given by *src_lengths*. + + # One-hot encode a batch of input characters. + input = self.one_hot_inputs[src_tokens[:, i].long()] + + # Feed the input to our RNN. + output, hidden = self.rnn(input, hidden) + + # Return the final output state for making a prediction + return output + +Finally let's define a *named architecture* with the configuration for our +model. This is done with the :func:`~fairseq.models.register_model_architecture` +function decorator. Thereafter this named architecture can be used with the +``--arch`` command-line argument, e.g., ``--arch pytorch_tutorial_rnn``:: + + from fairseq.models import register_model_architecture + + # The first argument to ``register_model_architecture()`` should be the name + # of the model we registered above (i.e., 'rnn_classifier'). The function we + # register here should take a single argument *args* and modify it in-place + # to match the desired architecture. + + @register_model_architecture('rnn_classifier', 'pytorch_tutorial_rnn') + def pytorch_tutorial_rnn(args): + # We use ``getattr()`` to prioritize arguments that are explicitly given + # on the command-line, so that the defaults defined below are only used + # when no other value has been specified. + args.hidden_dim = getattr(args, 'hidden_dim', 128) + + +3. Registering a new Task +------------------------- + +Now we'll register a new :class:`~fairseq.tasks.FairseqTask` that will load our +dictionaries and dataset. Tasks can also control how the data is batched into +mini-batches, but in this tutorial we'll reuse the batching provided by +:class:`fairseq.data.LanguagePairDataset`. + +Create a new file named :file:`fairseq/tasks/simple_classification.py` with the +following contents:: + + import os + import torch + + from fairseq.data import Dictionary, LanguagePairDataset + from fairseq.tasks import FairseqTask, register_task + + + @register_task('simple_classification') + class SimpleClassificationTask(LegacyFairseqTask): + + @staticmethod + def add_args(parser): + # Add some command-line arguments for specifying where the data is + # located and the maximum supported input length. + parser.add_argument('data', metavar='FILE', + help='file prefix for data') + parser.add_argument('--max-positions', default=1024, type=int, + help='max input length') + + @classmethod + def setup_task(cls, args, **kwargs): + # Here we can perform any setup required for the task. This may include + # loading Dictionaries, initializing shared Embedding layers, etc. + # In this case we'll just load the Dictionaries. + input_vocab = Dictionary.load(os.path.join(args.data, 'dict.input.txt')) + label_vocab = Dictionary.load(os.path.join(args.data, 'dict.label.txt')) + print('| [input] dictionary: {} types'.format(len(input_vocab))) + print('| [label] dictionary: {} types'.format(len(label_vocab))) + + return SimpleClassificationTask(args, input_vocab, label_vocab) + + def __init__(self, args, input_vocab, label_vocab): + super().__init__(args) + self.input_vocab = input_vocab + self.label_vocab = label_vocab + + def load_dataset(self, split, **kwargs): + """Load a given dataset split (e.g., train, valid, test).""" + + prefix = os.path.join(self.args.data, '{}.input-label'.format(split)) + + # Read input sentences. + sentences, lengths = [], [] + with open(prefix + '.input', encoding='utf-8') as file: + for line in file: + sentence = line.strip() + + # Tokenize the sentence, splitting on spaces + tokens = self.input_vocab.encode_line( + sentence, add_if_not_exist=False, + ) + + sentences.append(tokens) + lengths.append(tokens.numel()) + + # Read labels. + labels = [] + with open(prefix + '.label', encoding='utf-8') as file: + for line in file: + label = line.strip() + labels.append( + # Convert label to a numeric ID. + torch.LongTensor([self.label_vocab.add_symbol(label)]) + ) + + assert len(sentences) == len(labels) + print('| {} {} {} examples'.format(self.args.data, split, len(sentences))) + + # We reuse LanguagePairDataset since classification can be modeled as a + # sequence-to-sequence task where the target sequence has length 1. + self.datasets[split] = LanguagePairDataset( + src=sentences, + src_sizes=lengths, + src_dict=self.input_vocab, + tgt=labels, + tgt_sizes=torch.ones(len(labels)), # targets have length 1 + tgt_dict=self.label_vocab, + left_pad_source=False, + # Since our target is a single class label, there's no need for + # teacher forcing. If we set this to ``True`` then our Model's + # ``forward()`` method would receive an additional argument called + # *prev_output_tokens* that would contain a shifted version of the + # target sequence. + input_feeding=False, + ) + + def max_positions(self): + """Return the max input length allowed by the task.""" + # The source should be less than *args.max_positions* and the "target" + # has max length 1. + return (self.args.max_positions, 1) + + @property + def source_dictionary(self): + """Return the source :class:`~fairseq.data.Dictionary`.""" + return self.input_vocab + + @property + def target_dictionary(self): + """Return the target :class:`~fairseq.data.Dictionary`.""" + return self.label_vocab + + # We could override this method if we wanted more control over how batches + # are constructed, but it's not necessary for this tutorial since we can + # reuse the batching provided by LanguagePairDataset. + # + # def get_batch_iterator( + # self, dataset, max_tokens=None, max_sentences=None, max_positions=None, + # ignore_invalid_inputs=False, required_batch_size_multiple=1, + # seed=1, num_shards=1, shard_id=0, num_workers=0, epoch=1, + # data_buffer_size=0, disable_iterator_cache=False, + # ): + # (...) + + +4. Training the Model +--------------------- + +Now we're ready to train the model. We can use the existing :ref:`fairseq-train` +command-line tool for this, making sure to specify our new Task (``--task +simple_classification``) and Model architecture (``--arch +pytorch_tutorial_rnn``): + +.. note:: + + You can also configure the dimensionality of the hidden state by passing the + ``--hidden-dim`` argument to :ref:`fairseq-train`. + +.. code-block:: console + + > fairseq-train names-bin \ + --task simple_classification \ + --arch pytorch_tutorial_rnn \ + --optimizer adam --lr 0.001 --lr-shrink 0.5 \ + --max-tokens 1000 + (...) + | epoch 027 | loss 1.200 | ppl 2.30 | wps 15728 | ups 119.4 | wpb 116 | bsz 116 | num_updates 3726 | lr 1.5625e-05 | gnorm 1.290 | clip 0% | oom 0 | wall 32 | train_wall 21 + | epoch 027 | valid on 'valid' subset | valid_loss 1.41304 | valid_ppl 2.66 | num_updates 3726 | best 1.41208 + | done training in 31.6 seconds + +The model files should appear in the :file:`checkpoints/` directory. + + +5. Writing an evaluation script +------------------------------- + +Finally we can write a short script to evaluate our model on new inputs. Create +a new file named :file:`eval_classifier.py` with the following contents:: + + from fairseq import checkpoint_utils, data, options, tasks + + # Parse command-line arguments for generation + parser = options.get_generation_parser(default_task='simple_classification') + args = options.parse_args_and_arch(parser) + + # Setup task + task = tasks.setup_task(args) + + # Load model + print('| loading model from {}'.format(args.path)) + models, _model_args = checkpoint_utils.load_model_ensemble([args.path], task=task) + model = models[0] + + while True: + sentence = input('\nInput: ') + + # Tokenize into characters + chars = ' '.join(list(sentence.strip())) + tokens = task.source_dictionary.encode_line( + chars, add_if_not_exist=False, + ) + + # Build mini-batch to feed to the model + batch = data.language_pair_dataset.collate( + samples=[{'id': -1, 'source': tokens}], # bsz = 1 + pad_idx=task.source_dictionary.pad(), + eos_idx=task.source_dictionary.eos(), + left_pad_source=False, + input_feeding=False, + ) + + # Feed batch to the model and get predictions + preds = model(**batch['net_input']) + + # Print top 3 predictions and their log-probabilities + top_scores, top_labels = preds[0].topk(k=3) + for score, label_idx in zip(top_scores, top_labels): + label_name = task.target_dictionary.string([label_idx]) + print('({:.2f})\t{}'.format(score, label_name)) + +Now we can evaluate our model interactively. Note that we have included the +original data path (:file:`names-bin/`) so that the dictionaries can be loaded: + +.. code-block:: console + + > python eval_classifier.py names-bin --path checkpoints/checkpoint_best.pt + | [input] dictionary: 64 types + | [label] dictionary: 24 types + | loading model from checkpoints/checkpoint_best.pt + + Input: Satoshi + (-0.61) Japanese + (-1.20) Arabic + (-2.86) Italian + + Input: Sinbad + (-0.30) Arabic + (-1.76) English + (-4.08) Russian diff --git a/docs/tutorial_simple_lstm.rst b/docs/tutorial_simple_lstm.rst new file mode 100644 index 0000000000000000000000000000000000000000..f52988507c5da5125668e143bd2bfe4df117b41c --- /dev/null +++ b/docs/tutorial_simple_lstm.rst @@ -0,0 +1,518 @@ +Tutorial: Simple LSTM +===================== + +In this tutorial we will extend fairseq by adding a new +:class:`~fairseq.models.FairseqEncoderDecoderModel` that encodes a source +sentence with an LSTM and then passes the final hidden state to a second LSTM +that decodes the target sentence (without attention). + +This tutorial covers: + +1. **Writing an Encoder and Decoder** to encode/decode the source/target + sentence, respectively. +2. **Registering a new Model** so that it can be used with the existing + :ref:`Command-line tools`. +3. **Training the Model** using the existing command-line tools. +4. **Making generation faster** by modifying the Decoder to use + :ref:`Incremental decoding`. + + +1. Building an Encoder and Decoder +---------------------------------- + +In this section we'll define a simple LSTM Encoder and Decoder. All Encoders +should implement the :class:`~fairseq.models.FairseqEncoder` interface and +Decoders should implement the :class:`~fairseq.models.FairseqDecoder` interface. +These interfaces themselves extend :class:`torch.nn.Module`, so FairseqEncoders +and FairseqDecoders can be written and used in the same ways as ordinary PyTorch +Modules. + + +Encoder +~~~~~~~ + +Our Encoder will embed the tokens in the source sentence, feed them to a +:class:`torch.nn.LSTM` and return the final hidden state. To create our encoder +save the following in a new file named :file:`fairseq/models/simple_lstm.py`:: + + import torch.nn as nn + from fairseq import utils + from fairseq.models import FairseqEncoder + + class SimpleLSTMEncoder(FairseqEncoder): + + def __init__( + self, args, dictionary, embed_dim=128, hidden_dim=128, dropout=0.1, + ): + super().__init__(dictionary) + self.args = args + + # Our encoder will embed the inputs before feeding them to the LSTM. + self.embed_tokens = nn.Embedding( + num_embeddings=len(dictionary), + embedding_dim=embed_dim, + padding_idx=dictionary.pad(), + ) + self.dropout = nn.Dropout(p=dropout) + + # We'll use a single-layer, unidirectional LSTM for simplicity. + self.lstm = nn.LSTM( + input_size=embed_dim, + hidden_size=hidden_dim, + num_layers=1, + bidirectional=False, + batch_first=True, + ) + + def forward(self, src_tokens, src_lengths): + # The inputs to the ``forward()`` function are determined by the + # Task, and in particular the ``'net_input'`` key in each + # mini-batch. We discuss Tasks in the next tutorial, but for now just + # know that *src_tokens* has shape `(batch, src_len)` and *src_lengths* + # has shape `(batch)`. + + # Note that the source is typically padded on the left. This can be + # configured by adding the `--left-pad-source "False"` command-line + # argument, but here we'll make the Encoder handle either kind of + # padding by converting everything to be right-padded. + if self.args.left_pad_source: + # Convert left-padding to right-padding. + src_tokens = utils.convert_padding_direction( + src_tokens, + padding_idx=self.dictionary.pad(), + left_to_right=True + ) + + # Embed the source. + x = self.embed_tokens(src_tokens) + + # Apply dropout. + x = self.dropout(x) + + # Pack the sequence into a PackedSequence object to feed to the LSTM. + x = nn.utils.rnn.pack_padded_sequence(x, src_lengths, batch_first=True) + + # Get the output from the LSTM. + _outputs, (final_hidden, _final_cell) = self.lstm(x) + + # Return the Encoder's output. This can be any object and will be + # passed directly to the Decoder. + return { + # this will have shape `(bsz, hidden_dim)` + 'final_hidden': final_hidden.squeeze(0), + } + + # Encoders are required to implement this method so that we can rearrange + # the order of the batch elements during inference (e.g., beam search). + def reorder_encoder_out(self, encoder_out, new_order): + """ + Reorder encoder output according to `new_order`. + + Args: + encoder_out: output from the ``forward()`` method + new_order (LongTensor): desired order + + Returns: + `encoder_out` rearranged according to `new_order` + """ + final_hidden = encoder_out['final_hidden'] + return { + 'final_hidden': final_hidden.index_select(0, new_order), + } + + +Decoder +~~~~~~~ + +Our Decoder will predict the next word, conditioned on the Encoder's final +hidden state and an embedded representation of the previous target word -- which +is sometimes called *teacher forcing*. More specifically, we'll use a +:class:`torch.nn.LSTM` to produce a sequence of hidden states that we'll project +to the size of the output vocabulary to predict each target word. + +:: + + import torch + from fairseq.models import FairseqDecoder + + class SimpleLSTMDecoder(FairseqDecoder): + + def __init__( + self, dictionary, encoder_hidden_dim=128, embed_dim=128, hidden_dim=128, + dropout=0.1, + ): + super().__init__(dictionary) + + # Our decoder will embed the inputs before feeding them to the LSTM. + self.embed_tokens = nn.Embedding( + num_embeddings=len(dictionary), + embedding_dim=embed_dim, + padding_idx=dictionary.pad(), + ) + self.dropout = nn.Dropout(p=dropout) + + # We'll use a single-layer, unidirectional LSTM for simplicity. + self.lstm = nn.LSTM( + # For the first layer we'll concatenate the Encoder's final hidden + # state with the embedded target tokens. + input_size=encoder_hidden_dim + embed_dim, + hidden_size=hidden_dim, + num_layers=1, + bidirectional=False, + ) + + # Define the output projection. + self.output_projection = nn.Linear(hidden_dim, len(dictionary)) + + # During training Decoders are expected to take the entire target sequence + # (shifted right by one position) and produce logits over the vocabulary. + # The *prev_output_tokens* tensor begins with the end-of-sentence symbol, + # ``dictionary.eos()``, followed by the target sequence. + def forward(self, prev_output_tokens, encoder_out): + """ + Args: + prev_output_tokens (LongTensor): previous decoder outputs of shape + `(batch, tgt_len)`, for teacher forcing + encoder_out (Tensor, optional): output from the encoder, used for + encoder-side attention + + Returns: + tuple: + - the last decoder layer's output of shape + `(batch, tgt_len, vocab)` + - the last decoder layer's attention weights of shape + `(batch, tgt_len, src_len)` + """ + bsz, tgt_len = prev_output_tokens.size() + + # Extract the final hidden state from the Encoder. + final_encoder_hidden = encoder_out['final_hidden'] + + # Embed the target sequence, which has been shifted right by one + # position and now starts with the end-of-sentence symbol. + x = self.embed_tokens(prev_output_tokens) + + # Apply dropout. + x = self.dropout(x) + + # Concatenate the Encoder's final hidden state to *every* embedded + # target token. + x = torch.cat( + [x, final_encoder_hidden.unsqueeze(1).expand(bsz, tgt_len, -1)], + dim=2, + ) + + # Using PackedSequence objects in the Decoder is harder than in the + # Encoder, since the targets are not sorted in descending length order, + # which is a requirement of ``pack_padded_sequence()``. Instead we'll + # feed nn.LSTM directly. + initial_state = ( + final_encoder_hidden.unsqueeze(0), # hidden + torch.zeros_like(final_encoder_hidden).unsqueeze(0), # cell + ) + output, _ = self.lstm( + x.transpose(0, 1), # convert to shape `(tgt_len, bsz, dim)` + initial_state, + ) + x = output.transpose(0, 1) # convert to shape `(bsz, tgt_len, hidden)` + + # Project the outputs to the size of the vocabulary. + x = self.output_projection(x) + + # Return the logits and ``None`` for the attention weights + return x, None + + +2. Registering the Model +------------------------ + +Now that we've defined our Encoder and Decoder we must *register* our model with +fairseq using the :func:`~fairseq.models.register_model` function decorator. +Once the model is registered we'll be able to use it with the existing +:ref:`Command-line Tools`. + +All registered models must implement the +:class:`~fairseq.models.BaseFairseqModel` interface. For sequence-to-sequence +models (i.e., any model with a single Encoder and Decoder), we can instead +implement the :class:`~fairseq.models.FairseqEncoderDecoderModel` interface. + +Create a small wrapper class in the same file and register it in fairseq with +the name ``'simple_lstm'``:: + + from fairseq.models import FairseqEncoderDecoderModel, register_model + + # Note: the register_model "decorator" should immediately precede the + # definition of the Model class. + + @register_model('simple_lstm') + class SimpleLSTMModel(FairseqEncoderDecoderModel): + + @staticmethod + def add_args(parser): + # Models can override this method to add new command-line arguments. + # Here we'll add some new command-line arguments to configure dropout + # and the dimensionality of the embeddings and hidden states. + parser.add_argument( + '--encoder-embed-dim', type=int, metavar='N', + help='dimensionality of the encoder embeddings', + ) + parser.add_argument( + '--encoder-hidden-dim', type=int, metavar='N', + help='dimensionality of the encoder hidden state', + ) + parser.add_argument( + '--encoder-dropout', type=float, default=0.1, + help='encoder dropout probability', + ) + parser.add_argument( + '--decoder-embed-dim', type=int, metavar='N', + help='dimensionality of the decoder embeddings', + ) + parser.add_argument( + '--decoder-hidden-dim', type=int, metavar='N', + help='dimensionality of the decoder hidden state', + ) + parser.add_argument( + '--decoder-dropout', type=float, default=0.1, + help='decoder dropout probability', + ) + + @classmethod + def build_model(cls, args, task): + # Fairseq initializes models by calling the ``build_model()`` + # function. This provides more flexibility, since the returned model + # instance can be of a different type than the one that was called. + # In this case we'll just return a SimpleLSTMModel instance. + + # Initialize our Encoder and Decoder. + encoder = SimpleLSTMEncoder( + args=args, + dictionary=task.source_dictionary, + embed_dim=args.encoder_embed_dim, + hidden_dim=args.encoder_hidden_dim, + dropout=args.encoder_dropout, + ) + decoder = SimpleLSTMDecoder( + dictionary=task.target_dictionary, + encoder_hidden_dim=args.encoder_hidden_dim, + embed_dim=args.decoder_embed_dim, + hidden_dim=args.decoder_hidden_dim, + dropout=args.decoder_dropout, + ) + model = SimpleLSTMModel(encoder, decoder) + + # Print the model architecture. + print(model) + + return model + + # We could override the ``forward()`` if we wanted more control over how + # the encoder and decoder interact, but it's not necessary for this + # tutorial since we can inherit the default implementation provided by + # the FairseqEncoderDecoderModel base class, which looks like: + # + # def forward(self, src_tokens, src_lengths, prev_output_tokens): + # encoder_out = self.encoder(src_tokens, src_lengths) + # decoder_out = self.decoder(prev_output_tokens, encoder_out) + # return decoder_out + +Finally let's define a *named architecture* with the configuration for our +model. This is done with the :func:`~fairseq.models.register_model_architecture` +function decorator. Thereafter this named architecture can be used with the +``--arch`` command-line argument, e.g., ``--arch tutorial_simple_lstm``:: + + from fairseq.models import register_model_architecture + + # The first argument to ``register_model_architecture()`` should be the name + # of the model we registered above (i.e., 'simple_lstm'). The function we + # register here should take a single argument *args* and modify it in-place + # to match the desired architecture. + + @register_model_architecture('simple_lstm', 'tutorial_simple_lstm') + def tutorial_simple_lstm(args): + # We use ``getattr()`` to prioritize arguments that are explicitly given + # on the command-line, so that the defaults defined below are only used + # when no other value has been specified. + args.encoder_embed_dim = getattr(args, 'encoder_embed_dim', 256) + args.encoder_hidden_dim = getattr(args, 'encoder_hidden_dim', 256) + args.decoder_embed_dim = getattr(args, 'decoder_embed_dim', 256) + args.decoder_hidden_dim = getattr(args, 'decoder_hidden_dim', 256) + + +3. Training the Model +--------------------- + +Now we're ready to train the model. We can use the existing :ref:`fairseq-train` +command-line tool for this, making sure to specify our new Model architecture +(``--arch tutorial_simple_lstm``). + +.. note:: + + Make sure you've already preprocessed the data from the IWSLT example in the + :file:`examples/translation/` directory. + +.. code-block:: console + + > fairseq-train data-bin/iwslt14.tokenized.de-en \ + --arch tutorial_simple_lstm \ + --encoder-dropout 0.2 --decoder-dropout 0.2 \ + --optimizer adam --lr 0.005 --lr-shrink 0.5 \ + --max-tokens 12000 + (...) + | epoch 052 | loss 4.027 | ppl 16.30 | wps 420805 | ups 39.7 | wpb 9841 | bsz 400 | num_updates 20852 | lr 1.95313e-05 | gnorm 0.218 | clip 0% | oom 0 | wall 529 | train_wall 396 + | epoch 052 | valid on 'valid' subset | valid_loss 4.74989 | valid_ppl 26.91 | num_updates 20852 | best 4.74954 + +The model files should appear in the :file:`checkpoints/` directory. While this +model architecture is not very good, we can use the :ref:`fairseq-generate` script to +generate translations and compute our BLEU score over the test set: + +.. code-block:: console + + > fairseq-generate data-bin/iwslt14.tokenized.de-en \ + --path checkpoints/checkpoint_best.pt \ + --beam 5 \ + --remove-bpe + (...) + | Translated 6750 sentences (153132 tokens) in 17.3s (389.12 sentences/s, 8827.68 tokens/s) + | Generate test with beam=5: BLEU4 = 8.18, 38.8/12.1/4.7/2.0 (BP=1.000, ratio=1.066, syslen=139865, reflen=131146) + + +4. Making generation faster +--------------------------- + +While autoregressive generation from sequence-to-sequence models is inherently +slow, our implementation above is especially slow because it recomputes the +entire sequence of Decoder hidden states for every output token (i.e., it is +``O(n^2)``). We can make this significantly faster by instead caching the +previous hidden states. + +In fairseq this is called :ref:`Incremental decoding`. Incremental decoding is a +special mode at inference time where the Model only receives a single timestep +of input corresponding to the immediately previous output token (for teacher +forcing) and must produce the next output incrementally. Thus the model must +cache any long-term state that is needed about the sequence, e.g., hidden +states, convolutional states, etc. + +To implement incremental decoding we will modify our model to implement the +:class:`~fairseq.models.FairseqIncrementalDecoder` interface. Compared to the +standard :class:`~fairseq.models.FairseqDecoder` interface, the incremental +decoder interface allows ``forward()`` methods to take an extra keyword argument +(*incremental_state*) that can be used to cache state across time-steps. + +Let's replace our ``SimpleLSTMDecoder`` with an incremental one:: + + import torch + from fairseq.models import FairseqIncrementalDecoder + + class SimpleLSTMDecoder(FairseqIncrementalDecoder): + + def __init__( + self, dictionary, encoder_hidden_dim=128, embed_dim=128, hidden_dim=128, + dropout=0.1, + ): + # This remains the same as before. + super().__init__(dictionary) + self.embed_tokens = nn.Embedding( + num_embeddings=len(dictionary), + embedding_dim=embed_dim, + padding_idx=dictionary.pad(), + ) + self.dropout = nn.Dropout(p=dropout) + self.lstm = nn.LSTM( + input_size=encoder_hidden_dim + embed_dim, + hidden_size=hidden_dim, + num_layers=1, + bidirectional=False, + ) + self.output_projection = nn.Linear(hidden_dim, len(dictionary)) + + # We now take an additional kwarg (*incremental_state*) for caching the + # previous hidden and cell states. + def forward(self, prev_output_tokens, encoder_out, incremental_state=None): + if incremental_state is not None: + # If the *incremental_state* argument is not ``None`` then we are + # in incremental inference mode. While *prev_output_tokens* will + # still contain the entire decoded prefix, we will only use the + # last step and assume that the rest of the state is cached. + prev_output_tokens = prev_output_tokens[:, -1:] + + # This remains the same as before. + bsz, tgt_len = prev_output_tokens.size() + final_encoder_hidden = encoder_out['final_hidden'] + x = self.embed_tokens(prev_output_tokens) + x = self.dropout(x) + x = torch.cat( + [x, final_encoder_hidden.unsqueeze(1).expand(bsz, tgt_len, -1)], + dim=2, + ) + + # We will now check the cache and load the cached previous hidden and + # cell states, if they exist, otherwise we will initialize them to + # zeros (as before). We will use the ``utils.get_incremental_state()`` + # and ``utils.set_incremental_state()`` helpers. + initial_state = utils.get_incremental_state( + self, incremental_state, 'prev_state', + ) + if initial_state is None: + # first time initialization, same as the original version + initial_state = ( + final_encoder_hidden.unsqueeze(0), # hidden + torch.zeros_like(final_encoder_hidden).unsqueeze(0), # cell + ) + + # Run one step of our LSTM. + output, latest_state = self.lstm(x.transpose(0, 1), initial_state) + + # Update the cache with the latest hidden and cell states. + utils.set_incremental_state( + self, incremental_state, 'prev_state', latest_state, + ) + + # This remains the same as before + x = output.transpose(0, 1) + x = self.output_projection(x) + return x, None + + # The ``FairseqIncrementalDecoder`` interface also requires implementing a + # ``reorder_incremental_state()`` method, which is used during beam search + # to select and reorder the incremental state. + def reorder_incremental_state(self, incremental_state, new_order): + # Load the cached state. + prev_state = utils.get_incremental_state( + self, incremental_state, 'prev_state', + ) + + # Reorder batches according to *new_order*. + reordered_state = ( + prev_state[0].index_select(1, new_order), # hidden + prev_state[1].index_select(1, new_order), # cell + ) + + # Update the cached state. + utils.set_incremental_state( + self, incremental_state, 'prev_state', reordered_state, + ) + +Finally, we can rerun generation and observe the speedup: + +.. code-block:: console + + # Before + + > fairseq-generate data-bin/iwslt14.tokenized.de-en \ + --path checkpoints/checkpoint_best.pt \ + --beam 5 \ + --remove-bpe + (...) + | Translated 6750 sentences (153132 tokens) in 17.3s (389.12 sentences/s, 8827.68 tokens/s) + | Generate test with beam=5: BLEU4 = 8.18, 38.8/12.1/4.7/2.0 (BP=1.000, ratio=1.066, syslen=139865, reflen=131146) + + # After + + > fairseq-generate data-bin/iwslt14.tokenized.de-en \ + --path checkpoints/checkpoint_best.pt \ + --beam 5 \ + --remove-bpe + (...) + | Translated 6750 sentences (153132 tokens) in 5.5s (1225.54 sentences/s, 27802.94 tokens/s) + | Generate test with beam=5: BLEU4 = 8.18, 38.8/12.1/4.7/2.0 (BP=1.000, ratio=1.066, syslen=139865, reflen=131146) diff --git a/examples/.gitignore b/examples/.gitignore new file mode 100644 index 0000000000000000000000000000000000000000..1ef816f2cd7b4a9aa7adf8bd5635a644834738f1 --- /dev/null +++ b/examples/.gitignore @@ -0,0 +1,2 @@ +!*/*.sh +!*/*.md diff --git a/examples/__init__.py b/examples/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..44bb24ae614941f23fea29c56d60167650c39bcb --- /dev/null +++ b/examples/__init__.py @@ -0,0 +1,9 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +try: + from fairseq.version import __version__ # noqa +except ImportError: + pass diff --git a/examples/adaptive_span/README.md b/examples/adaptive_span/README.md new file mode 100644 index 0000000000000000000000000000000000000000..913a87338633f8a790d70fe4133b8bd8b95a4c50 --- /dev/null +++ b/examples/adaptive_span/README.md @@ -0,0 +1,90 @@ +# Adaptive Span + +Adaptive Span is a novel self-attention mechanism that can learn its optimal +attention span. This allows us to extend significantly the maximum context size +used in Transformer, while maintaining control over their memory footprint +and computational time. It uses the Truncated BPTT technique for training, +as in [transformerXL](https://github.com/pytorch/fairseq/blob/master/examples/truncated_bptt/README.md). + +Adaptive Span was introduced by paper: +[Adaptive Attention Span in Transformers](https://arxiv.org/abs/1905.07799), +which achieved state-of-the-art language modeling results at the time of publication. + +We manage to reproduce their result in fairseq and keep most of the +[original implementation](https://github.com/facebookresearch/adaptive-span) untouched. +You can refer to the their sweep file as well if any combination of hyperparameter is not clear. + +##### 0. Setup + +First you need to process the Enwik8 dataset, we use the pre-tokenized dataset +from [adaptive span paper](https://github.com/facebookresearch/adaptive-span/blob/master/get_data.sh). +You can download the dataset, and then run: +```bash +fairseq-preprocess --only-source --trainpref ~/data/enwik8/train.txt \ + --validpref ~/data/enwik8/valid.txt --testpref ~/data/enwik8/test.txt \ + --destdir ~/data/enwik8/data-bin/ --joined-dictionary --workers 20 +``` + +##### 1. Train a Adaptive Span model on Enwik8 + +We will train a 12-layer Adaptive Span model following the [hyperparameters +used in the original +paper](https://github.com/facebookresearch/adaptive-span/blob/master/experiments/enwik8.sh). + +The following command assumes 4 GPUs, so that the total batch size is 64 +sequences (4 x 16). Training should take 2-3 days on 4 V100 GPUs: +```bash +CUDA_VISIBLE_DEVICES=0,1,2,3 fairseq-train \ + --user-dir examples/adaptive_span \ + --data ~/data/enwik8/data-bin/ \ + --fp16 --fp16-no-flatten-grads --max-update 600000 \ + --task truncated_bptt_lm --tokens-per-sample 512 --arch adaptive_span \ + --n-layer 12 --d-model 512 --n-head 8 --d-inner 2048 --dropout 0.3 \ + --attn-span 8192 --optimizer adagrad_with_grad_clip --adagrad-clip 0.03 \ + --validate-interval-updates 1000 \ + --lr-scheduler fixed --warmup-updates 32000 --batch-size-valid 32 \ + --lr 0.07 --criterion adaptive_span_loss --batch-size 16 --update-freq 1 \ + --seed 2 --log-format json --log-interval 25 --aux-loss-scaler 5e-07 +``` +This should land around 1.05 on validation, 1.03 on test. You can lower the +--aux-loss-scaler for better performance (longer span). It gives ~0.03 bpc +improvement to the transformerXL baseline here. +If training on a single GPU, set `--update-freq=4` to accumulate 4x gradients +and simulate training on 4 GPUs. +You can also reproduce the transformerXL result on enwik8 using this code base. +It should land around 1.06 on test,matching the [original paper](https://github.com/kimiyoung/transformer-xl/blob/master/pytorch/run_enwik8_base.sh). +You can try by +```bash +CUDA_VISIBLE_DEVICES=0,1,2,3 fairseq-train \ + --user-dir examples/truncated_bptt \ + ~/data/enwik8/data-bin/ \ + --task truncated_bptt_lm --fp16 --max-update 400000 \ + --tokens-per-sample 512 --arch transformer_xl --n-layer 12 \ + --d-model 512 --n-head 8 --d-head 64 --d-inner 2048 --dropout 0.1 \ + --dropatt 0.0 --mem-len 512 --optimizer adam --clip-norm 0.25 \ + --lr-scheduler cosine --warmup-updates 0 \ + --lr 0.0 --lr 0.00025 --batch-size 15 \ + --update-freq 1 --seed 2 --log-format json --log-interval 25 \ + --fp16 +``` + +##### 2. Evaluate +For Adaptive Span: +```bash +fairseq-eval-lm ~/data/enwik8/data-bin/ --path model/checkpoint_best.pt \ + --user-dir examples/adaptive_span \ + --task truncated_bptt_lm --batch-size 8 --tokens-per-sample 512 --gen-subset test +``` +For Transformer-XL evaluation: +```bash +fairseq-eval-lm ~/data/enwik8/data-bin/ --path model/checkpoint_best.pt \ + --user-dir examples/truncated_bptt/ --task truncated_bptt_lm --batch-size 8 \ + --tokens-per-sample 80 \ + --model-overrides '{"mem_len":2100,"clamp_len":820,"same_length":True}' \ + --gen-subset valid +``` + +*Note:* During training the model saw 512 tokens of context +(``--tokens-per-sample=512``), with batch size 8. These settings match the evaluation +settings from [the original +paper](https://github.com/facebookresearch/adaptive-span/blob/master/experiments/enwik8.sh). diff --git a/examples/adaptive_span/__init__.py b/examples/adaptive_span/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e0a142a769360e1140bf814c532eaf841f1d52d8 --- /dev/null +++ b/examples/adaptive_span/__init__.py @@ -0,0 +1,19 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import importlib +import os + +# automatically import any Python files in the current directory +cur_dir = os.path.dirname(__file__) +for file in os.listdir(cur_dir): + path = os.path.join(cur_dir, file) + if ( + not file.startswith("_") + and not file.startswith(".") + and (file.endswith(".py") or os.path.isdir(path)) + ): + mod_name = file[: file.find(".py")] if file.endswith(".py") else file + module = importlib.import_module(__name__ + "." + mod_name) diff --git a/examples/adaptive_span/adagrad_with_grad_clip.py b/examples/adaptive_span/adagrad_with_grad_clip.py new file mode 100644 index 0000000000000000000000000000000000000000..585ce184ab2d6bbde0d2f7fcafd6536fa8f6d8b6 --- /dev/null +++ b/examples/adaptive_span/adagrad_with_grad_clip.py @@ -0,0 +1,128 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from torch.optim import Adagrad + +from fairseq.optim import LegacyFairseqOptimizer, register_optimizer + + +@register_optimizer("adagrad_with_grad_clip") +class FairseqAdagradWithGradClip(LegacyFairseqOptimizer): + def __init__(self, args, params): + super().__init__(args) + self._optimizer = AdagradWithGradClip(params, **self.optimizer_config) + + @staticmethod + def add_args(parser): + """Add optimizer-specific arguments to the parser.""" + # fmt: off + parser.add_argument('--weight-decay', '--wd', default=0.0, type=float, metavar='WD', + help='weight decay') + parser.add_argument('--adagrad-clip', default=0.0, type=float, metavar='D', + help='internal grad clip') + # fmt: on + + @property + def optimizer_config(self): + """ + Return a kwarg dictionary that will be used to override optimizer + args stored in checkpoints. This allows us to load a checkpoint and + resume training using a different set of optimizer args, e.g., with a + different learning rate. + """ + return { + "lr": self.args.lr[0], + "weight_decay": self.args.weight_decay, + "grad_clip": self.args.adagrad_clip, + } + + @property + def supports_flat_params(self): + return False + + +def _clip_grad(clr, grad, group_grad_clip): + if group_grad_clip > 0: + norm = grad.norm(2).item() + if norm > group_grad_clip: + clr *= group_grad_clip / (norm + 1e-10) + return clr + + +class AdagradWithGradClip(Adagrad): + """Adagrad algorithm with custom gradient clipping""" + + def __init__( + self, + params, + lr=1e-2, + lr_decay=0, + weight_decay=0, + initial_accumulator_value=0, + grad_clip=0, + ): + Adagrad.__init__( + self, + params, + lr=lr, + lr_decay=lr_decay, + weight_decay=weight_decay, + initial_accumulator_value=initial_accumulator_value, + ) + self.defaults["grad_clip"] = grad_clip + self.param_groups[0].setdefault("grad_clip", grad_clip) + + def step(self, closure=None): + loss = None + if closure is not None: + loss = closure() + + for group in self.param_groups: + for p in group["params"]: + if p.grad is None: + continue + + grad = p.grad.data + state = self.state[p] + + state["step"] += 1 + + if group["weight_decay"] != 0: + if p.grad.data.is_sparse: + raise RuntimeError( + "weight_decay option is " + "not compatible with sparse " + "gradients" + ) + grad = grad.add(group["weight_decay"], p.data) + + clr = group["lr"] / (1 + (state["step"] - 1) * group["lr_decay"]) + + # clip + clr = _clip_grad(clr=clr, grad=grad, group_grad_clip=group["grad_clip"]) + + if grad.is_sparse: + # the update is non-linear so indices must be unique + grad = grad.coalesce() + grad_indices = grad._indices() + grad_values = grad._values() + size = grad.size() + + def make_sparse(values): + constructor = grad.new + if grad_indices.dim() == 0 or values.dim() == 0: + return constructor().resize_as_(grad) + return constructor(grad_indices, values, size) + + state["sum"].add_(make_sparse(grad_values.pow(2))) + std = state["sum"]._sparse_mask(grad) + std_values = std._values().sqrt_().add_(1e-10) + p.data.add_(-clr, make_sparse(grad_values / std_values)) + else: + state["sum"].addcmul_(1, grad, grad) + std = state["sum"].sqrt().add_(1e-10) + p.data.addcdiv_(-clr, grad, std) + + return loss diff --git a/examples/adaptive_span/adaptive_span_attention.py b/examples/adaptive_span/adaptive_span_attention.py new file mode 100644 index 0000000000000000000000000000000000000000..07f757bb8e1a8a67b1124175ee338c8735aa8d65 --- /dev/null +++ b/examples/adaptive_span/adaptive_span_attention.py @@ -0,0 +1,160 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +import math + +import torch +import torch.nn as nn +import torch.nn.functional as F + + +class AdaptiveMask(nn.Module): + """Soft masking function for adaptive size. + It masks out the last K values of an input. The masking value + goes from 1 to 0 gradually, so K can be learned with + back-propagation. + Args: + max_size: maximum size (i.e. input dimension) + ramp_size: size of the ramp going from 0 to 1 + init_val: initial size proportion not to be masked out + shape: learn multiple sizes independent of each other + """ + + def __init__(self, max_size, ramp_size, init_val=0, shape=(1,)): + nn.Module.__init__(self) + self._max_size = max_size + self._ramp_size = ramp_size + self.current_val = nn.Parameter(torch.zeros(*shape) + init_val) + mask_template = torch.linspace(1 - max_size, 0, steps=max_size) + self.register_buffer("mask_template", mask_template) + + def forward(self, x): + mask = self.mask_template.float() + self.current_val.float() * self._max_size + mask = mask / self._ramp_size + 1 + mask = mask.clamp(0, 1) + if x.size(-1) < self._max_size: + # the input could have been trimmed beforehand to save computation + mask = mask.narrow(-1, self._max_size - x.size(-1), x.size(-1)) + x = (x * mask).type_as(x) + return x + + def get_current_max_size(self, include_ramp=True): + current_size = math.ceil(self.current_val.max().item() * self._max_size) + if include_ramp: + current_size += self._ramp_size + current_size = max(0, min(self._max_size, current_size)) + return current_size + + def get_current_avg_size(self, include_ramp=True): + current_size = math.ceil( + self.current_val.float().mean().item() * self._max_size + ) + if include_ramp: + current_size += self._ramp_size + current_size = max(0, min(self._max_size, current_size)) + return current_size + + def clamp_param(self): + """this need to be called after each update""" + self.current_val.data.clamp_(0, 1) + + +class AdaptiveSpan(nn.Module): + """Adaptive attention span for Transformerself. + This module learns an attention span length from data for each + self-attention head. + Args: + attn_span: maximum attention span + adapt_span_loss: loss coefficient for the span length + adapt_span_ramp: length of the masking ramp + adapt_span_init: initial size ratio + adapt_span_cache: adapt cache size to reduce memory usage + """ + + def __init__( + self, + attn_span, + adapt_span_ramp, + adapt_span_init, + n_head, + adapt_span_layer, + **kargs + ): + nn.Module.__init__(self) + self._max_span = attn_span + self._n_head = n_head + self._adapt_span_layer = adapt_span_layer + if self._adapt_span_layer: + self._mask = AdaptiveMask( + max_size=self._max_span, + ramp_size=adapt_span_ramp, + init_val=adapt_span_init, + ) + else: + self._mask = AdaptiveMask( + max_size=self._max_span, + ramp_size=adapt_span_ramp, + init_val=adapt_span_init, + shape=(n_head, 1, 1), + ) + + def forward(self, attn, normalize=True): + """mask attention with the right span""" + # batch and head dimensions are merged together, so separate them first + self.clamp_param() + if self._adapt_span_layer: + attn = self._mask(attn) + else: + B = attn.size(0) # batch size + M = attn.size(1) # block size + attn = attn.reshape(B // self._n_head, self._n_head, M, -1) + attn = self._mask(attn) + attn = attn.view(B, M, -1) + return attn + + def get_trim_len(self): + """how much of memory can be trimmed to reduce computation""" + L = self._max_span + trim_len = min(L - 1, L - self._mask.get_current_max_size()) + # too fine granularity might be bad for the memory management + trim_len = math.floor(trim_len / 64) * 64 + return trim_len + + def trim_memory(self, query, key, value, key_pe): + """trim out unnecessary memory beforehand to reduce computation""" + trim_len = self.get_trim_len() + cache_size = key.size(1) - query.size(1) + trim_len_cache = trim_len - (self._max_span - cache_size) + if trim_len_cache > 0: + key = key[:, trim_len_cache:, :] + value = value[:, trim_len_cache:, :] + elif trim_len_cache < 0: + # cache is too short! this happens when validation resumes + # after a lot of updates. + key = F.pad(key, [0, 0, -trim_len_cache, 0]) + value = F.pad(value, [0, 0, -trim_len_cache, 0]) + if trim_len > 0: + if key_pe is not None: + key_pe = key_pe[:, :, trim_len:] + return key, value, key_pe + + def get_cache_size(self): + """determine how long the cache should be""" + trim_len = self.get_trim_len() + # give a buffer of 64 steps since a span might increase + # in future updates + return min(self._max_span, self._max_span - trim_len + 64) + + def get_loss(self): + """a loss term for regularizing the span length""" + return self._max_span * self._mask.current_val.float().mean() + + def get_current_max_span(self): + return self._mask.get_current_max_size() + + def get_current_avg_span(self): + return self._mask.get_current_avg_size() + + def clamp_param(self): + self._mask.clamp_param() diff --git a/examples/adaptive_span/adaptive_span_loss.py b/examples/adaptive_span/adaptive_span_loss.py new file mode 100644 index 0000000000000000000000000000000000000000..056245807e5f8d313a8ad5be68aea4e285f4f580 --- /dev/null +++ b/examples/adaptive_span/adaptive_span_loss.py @@ -0,0 +1,106 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +from dataclasses import dataclass + +import torch.nn.functional as F +from fairseq import metrics, utils +from fairseq.criterions import register_criterion +from fairseq.criterions.cross_entropy import CrossEntropyCriterion +from fairseq.dataclass import FairseqDataclass +from omegaconf import II + + +@dataclass +class AdaptiveSpanCriterionConfig(FairseqDataclass): + sentence_avg: bool = II("optimization.sentence_avg") + + +@register_criterion("adaptive_span_loss", dataclass=AdaptiveSpanCriterionConfig) +class AdaptiveSpanCriterion(CrossEntropyCriterion): + def __init__(self, task, sentence_avg): + super().__init__(task, sentence_avg) + + def forward(self, model, sample, reduce=True): + """Compute the loss for the given sample. + + Returns a tuple with three elements: + 1) the loss here is summed, different from the adaptive span code + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + net_output = model(**sample["net_input"]) + loss, aux_loss, avg_span, max_span = self.compute_loss( + model, net_output, sample, reduce=reduce + ) + sample_size = ( + sample["target"].size(0) if self.sentence_avg else sample["ntokens"] + ) + loss /= sample_size + total_loss = loss + aux_loss + sample_size = 1 + + logging_output = { + "loss": loss.data, + "ntokens": sample["ntokens"], + "nsentences": sample["target"].size(0), + "sample_size": sample_size, + "total_loss": total_loss.data, + "avg_span": avg_span * sample_size, + "max_span": max_span * sample_size, + } + return total_loss, sample_size, logging_output + + def compute_loss(self, model, net_output, sample, reduce=True): + loss, _ = super().compute_loss(model, net_output, sample, reduce) + aux_loss = model.get_aux_loss() + avg_span = model.get_current_avg_span() + max_span = model.get_current_max_span() + return loss, aux_loss, avg_span, max_span + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + """Aggregate logging outputs from data parallel training.""" + loss_sum = sum(log.get("loss", 0) for log in logging_outputs) + ntokens = sum(log.get("ntokens", 0) for log in logging_outputs) + sample_size = sum(log.get("sample_size", 0) for log in logging_outputs) + total_loss_sum = sum(log.get("total_loss", 0) for log in logging_outputs) + avg_span_sum = sum(log.get("avg_span", 0) for log in logging_outputs) + max_span_sum = sum(log.get("max_span", 0) for log in logging_outputs) + + # we divide by log(2) to convert the loss from base e to base 2 + metrics.log_scalar( + "loss", loss_sum / sample_size / math.log(2), sample_size, round=3 + ) + metrics.log_scalar("avg_span", avg_span_sum / sample_size, sample_size, round=3) + metrics.log_scalar("max_span", max_span_sum / sample_size, sample_size, round=3) + # total loss contains the L1 norm on adaptive-span + metrics.log_scalar( + "total_loss", + total_loss_sum / sample_size / math.log(2), + sample_size, + round=3, + ) + if sample_size != ntokens: + metrics.log_scalar( + "nll_loss", loss_sum / ntokens / math.log(2), ntokens, round=3 + ) + metrics.log_derived( + "ppl", lambda meters: utils.get_perplexity(meters["nll_loss"].avg) + ) + else: + metrics.log_derived( + "ppl", lambda meters: utils.get_perplexity(meters["loss"].avg) + ) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return True diff --git a/examples/adaptive_span/adaptive_span_model.py b/examples/adaptive_span/adaptive_span_model.py new file mode 100644 index 0000000000000000000000000000000000000000..d96c95b85dbcf29e9384cc6d8d9630d2489991b2 --- /dev/null +++ b/examples/adaptive_span/adaptive_span_model.py @@ -0,0 +1,263 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# All rights reserved. +# +# This source code is licensed under the license found in the +# LICENSE file in the root directory of this source tree. + +import math + +import torch +import torch.nn as nn +import torch.nn.functional as F + +from fairseq.modules.layer_norm import LayerNorm + +from .adaptive_span_attention import AdaptiveSpan + +# Size notations: +# B = batch_size, H = d_model, M = block_size, L = attn_span + + +def _skew(X, pad_value): + """shift every row 1 step to right""" + # X = B x M x L + B, M, L = X.size() + X = F.pad(X, (0, M + 1), value=pad_value) # B x M x (L+M+1) + X = X.view(B, -1) # B x ML+MM+M + X = X[:, :-M] # B x ML+MM + X = X.view(B, M, M + L) # B x M x L+M + return X + + +def _unskew(X): + """reverse _skew operation""" + # X = B x M x L+M + B, M, L = X.size() + L -= M + X = X.view(B, -1) # B x ML+MM + X = F.pad(X, (0, M)) # B x ML+MM+M + X = X.view(B, M, M + L + 1) # B x M x L+M+1 + X = X[:, :, :L] # B x M x L + return X + + +class SeqAttention(nn.Module): + """Sequential self-attention layer. + Each token will attend to its previous fixed number of steps. + Note that attention doesn't include the current step itself. + """ + + def __init__(self, d_model, n_head, attn_span, dropout, adapt_span_layer, **kargs): + nn.Module.__init__(self) + self.dropout = nn.Dropout(dropout) + self.d_model = d_model # size of a single head + self.attn_span = attn_span + self.adaptive_span = AdaptiveSpan( + attn_span=attn_span, + n_head=n_head, + adapt_span_layer=adapt_span_layer, + **kargs + ) + + def forward(self, query, key, value, key_pe): + # query size = B x M x H + # key, value sizes = B x (M+L) x H + + key, value, key_pe = self.adaptive_span.trim_memory(query, key, value, key_pe) + + # compute attention from context + # B x M (dest) x (M+L) (src) + attn_cont = torch.matmul(query, key.transpose(-1, -2)) + attn_cont = _unskew(attn_cont) # B x M x L + + # compute the effect of position embedding + attn_pos = torch.matmul(query, key_pe) # B x M x L_pos + attn = attn_cont + attn_pos + + attn = attn / math.sqrt(self.d_model) # B x M X L_pos + + attn = F.softmax(attn.float(), dim=-1).type_as(attn) + + # trim attention lengths according to the learned span + attn = self.adaptive_span(attn) + + attn = self.dropout(attn) # B x M X L_pos + + attn_cont = _skew(attn, 0) # B x M X (L+M) + out = torch.matmul(attn_cont, value) # B x M x H + return out + + def get_cache_size(self): + return self.adaptive_span.get_cache_size() + + +class MultiHeadSeqAttention(nn.Module): + def __init__(self, d_model, n_head, **kargs): + nn.Module.__init__(self) + assert d_model % n_head == 0 + self.n_head = n_head + self.head_dim = d_model // n_head + self.attn = SeqAttention(d_model=self.head_dim, n_head=n_head, **kargs) + self.proj_query = nn.Linear(d_model, d_model, bias=False) + nn.init.xavier_normal_(self.proj_query.weight) + self.proj_out = nn.Linear(d_model, d_model, bias=False) + nn.init.xavier_normal_(self.proj_out.weight) + self.proj_val = nn.Linear(d_model, d_model, bias=False) + nn.init.xavier_normal_(self.proj_val.weight) + self.proj_key = nn.Linear(d_model, d_model, bias=False) + nn.init.xavier_normal_(self.proj_key.weight) + + def head_reshape(self, x): + K = self.n_head + D = self.head_dim + x = x.view(x.size()[:-1] + (K, D)) # B x (M+L) x K x D + x = x.transpose(1, 2).contiguous() # B x K x (M+L) x D + x = x.view(-1, x.size(-2), x.size(-1)) # B_K x (M+L) x D + return x + + def forward(self, query, key, value, key_pe): + B = query.size(0) + K = self.n_head + D = self.head_dim + M = query.size(1) + + query = self.proj_query(query) + query = self.head_reshape(query) + value = self.proj_val(value) + value = self.head_reshape(value) + key = self.proj_key(key) + key = self.head_reshape(key) + + out = self.attn(query, key, value, key_pe) # B_K x M x D + out = out.view(B, K, M, D) # B x K x M x D + out = out.transpose(1, 2).contiguous() # B x M x K x D + out = out.view(B, M, -1) # B x M x K_D + out = self.proj_out(out) + return out + + +class FeedForwardLayer(nn.Module): + def __init__(self, d_model, d_inner, dropout, **kargs): + nn.Module.__init__(self) + self.fc1 = nn.Linear(d_model, d_inner) + self.fc2 = nn.Linear(d_inner, d_model) + nn.init.xavier_uniform_(self.fc1.weight) + nn.init.xavier_uniform_(self.fc2.weight) + self.dropout = nn.Dropout(dropout) + + def forward(self, h): + h1 = F.relu(self.fc1(h)) + h1 = self.dropout(h1) + h2 = self.fc2(h1) + return h2 + + +class TransformerSeqLayer(nn.Module): + def __init__(self, d_model, **kargs): + nn.Module.__init__(self) + self.attn = MultiHeadSeqAttention(d_model=d_model, **kargs) + self.norm1 = LayerNorm(d_model) + self.ff = FeedForwardLayer(d_model=d_model, **kargs) + self.norm2 = LayerNorm(d_model) + + def forward(self, h, h_cache, key_pe): + # h = B x M x H + # h_cache = B x L x H + h_all = torch.cat([h_cache, h], dim=1) # B x (M+L) x H + attn_out = self.attn(h, h_all, h_all, key_pe) + h = self.norm1(h + attn_out) # B x M x H + if self.ff is not None: + ff_out = self.ff(h) + out = self.norm2(h + ff_out) # B x M x H + else: + out = h + return out + + def get_cache_size(self): + return self.attn.attn.get_cache_size() + + +class TransformerSeq(nn.Module): + def __init__( + self, + vocab_size, + d_model, + n_head, + n_layer, + attn_span, + emb_dropout, + aux_loss_scaler, + adapt_span_layer, + **kargs + ): + nn.Module.__init__(self) + # token embeddings + self.in_emb = nn.Embedding(vocab_size, d_model) + nn.init.normal_(self.in_emb.weight, mean=0, std=d_model ** -0.5) + self.out_emb = nn.Linear(d_model, vocab_size) + self.aux_loss_scaler = aux_loss_scaler + if emb_dropout > 0: + self.emb_dropout = nn.Dropout(emb_dropout) + else: + self.emb_dropout = None + # position embeddings + self.key_pe = nn.Parameter(torch.randn(1, d_model // n_head, attn_span)) + + self.layers = nn.ModuleList() + self.layers.extend( + TransformerSeqLayer( + d_model=d_model, + n_head=n_head, + attn_span=attn_span, + adapt_span_layer=adapt_span_layer, + **kargs + ) + for _ in range(n_layer) + ) + + def forward(self, x, h_cache, target=None): + # x size = B x M + block_size = x.size(1) + h = self.in_emb(x) # B x M x H + if self.emb_dropout is not None: + h = self.emb_dropout(h) + + h_cache_next = [] + for l, layer in enumerate(self.layers): + cache_size = layer.attn.attn.get_cache_size() + if cache_size > block_size: + h_cache_next_l = torch.cat( + [h_cache[l][:, -cache_size + block_size :, :], h], dim=1 + ).detach() + else: + h_cache_next_l = h[:, -cache_size:, :].detach() + h_cache_next.append(h_cache_next_l) + h = layer(h, h_cache[l], self.key_pe) # B x M x H + + if self.emb_dropout is not None: + h = self.emb_dropout(h) + + out = F.log_softmax(self.out_emb(h).float(), dim=-1).type_as(h) + dummy_loss = None + + return out, h_cache_next, dummy_loss + + def get_aux_loss(self): + loss = 0.0 + for layer in self.layers: + loss += layer.attn.attn.adaptive_span.get_loss() + return self.aux_loss_scaler * loss + + def get_current_max_span(self): + max_span = 0.0 + for layer in self.layers: + max_span = max( + max_span, layer.attn.attn.adaptive_span.get_current_max_span() + ) + return max_span + + def get_current_avg_span(self): + avg_span = 0.0 + for layer in self.layers: + avg_span += layer.attn.attn.adaptive_span.get_current_avg_span() + return avg_span / len(self.layers) diff --git a/examples/adaptive_span/adaptive_span_model_wrapper.py b/examples/adaptive_span/adaptive_span_model_wrapper.py new file mode 100644 index 0000000000000000000000000000000000000000..5b147fe11f9d730438d036321a2d4a5d776efaa2 --- /dev/null +++ b/examples/adaptive_span/adaptive_span_model_wrapper.py @@ -0,0 +1,145 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +from dataclasses import dataclass +from typing import Dict, List, Optional + +import torch +from fairseq.dataclass import FairseqDataclass +from fairseq.models import ( + FairseqIncrementalDecoder, + FairseqLanguageModel, + register_model, +) +from .adaptive_span_model import TransformerSeq as AdaptiveSpanTransformerModel + + +logger = logging.getLogger(__name__) + + +@dataclass +class AdaptiveSpanSmallConfig(FairseqDataclass): + # defaults come from https://github.com/facebookresearch/adaptive-span/blob/master/experiments/enwik8_small.sh + vocab_size: int = 50 + d_model: int = 256 + n_head: int = 4 + d_inner: int = 1024 + n_layer: int = 8 + attn_span: int = 1024 + dropout: float = 0.0 + emb_dropout: float = 0.0 + adapt_span_ramp: int = 32 + adapt_span_init: float = 0.0 + aux_loss_scaler: float = 0.000002 + adapt_span_layer: bool = False + + +@register_model("adaptive_span", dataclass=AdaptiveSpanSmallConfig) +class AdaptiveSpanTransformer(FairseqLanguageModel): + @classmethod + def build_model(cls, cfg: AdaptiveSpanSmallConfig, task): + return cls(AdaptiveSpanDecoder(cfg, task)) + + def get_aux_loss(self): + return self.decoder.get_aux_loss() + + def get_current_max_span(self): + return self.decoder.get_current_max_span() + + def get_current_avg_span(self): + return self.decoder.get_current_avg_span() + + +class AdaptiveSpanDecoder(FairseqIncrementalDecoder): + def __init__(self, cfg, task): + + super().__init__(task.target_dictionary) + + self.config = cfg + config = AdaptiveSpanSmallConfig( + vocab_size=len(task.target_dictionary), + d_model=cfg.d_model, + n_head=cfg.n_head, + d_inner=cfg.d_inner, + n_layer=cfg.n_layer, + attn_span=cfg.attn_span, + dropout=cfg.dropout, + emb_dropout=cfg.emb_dropout, + adapt_span_ramp=cfg.adapt_span_ramp, + adapt_span_init=cfg.adapt_span_init, + aux_loss_scaler=cfg.aux_loss_scaler, + adapt_span_layer=cfg.adapt_span_layer, + ) + logger.info(config) + self.model = AdaptiveSpanTransformerModel(**config.__dict__) + + self._mems = None + + def forward( + self, + src_tokens, + incremental_state: Optional[Dict[str, List[torch.Tensor]]] = None, + encoder_out=None, + ): + bsz = src_tokens.size(0) + if incremental_state is not None: # used during inference + mems = self.get_incremental_state("mems") + src_tokens = src_tokens[:, -1:] # only keep the most recent token + else: + mems = self._mems + + if mems is None: + # first time init + mems = self.init_hid_cache(bsz) + output = self.model(x=src_tokens, h_cache=mems,) + if incremental_state is not None: + self.set_incremental_state(incremental_state, "mems", output[1]) + else: + self._mems = output[1] + return (output[0],) + + def max_positions(self): + return self.config.attn_span + + def init_hid_cache(self, batch_sz): + hid = [] + for layer in self.model.layers: + param = next(self.model.parameters()) + h = torch.zeros( + batch_sz, + layer.get_cache_size(), + self.config.d_model, + dtype=param.dtype, + device=param.device, + ) + hid.append(h) + return hid + + def get_aux_loss(self): + return self.model.get_aux_loss() + + def get_current_max_span(self): + return self.model.get_current_max_span() + + def get_current_avg_span(self): + return self.model.get_current_avg_span() + + def reorder_incremental_state( + self, + incremental_state: Dict[str, Dict[str, Optional[torch.Tensor]]], + new_order: torch.Tensor, + ): + """Reorder incremental state. + + This will be called when the order of the input has changed from the + previous time step. A typical use case is beam search, where the input + order changes between time steps based on the selection of beams. + """ + raise NotImplementedError("This is required for generation/beam search") + # mems = self.get_incremental_state(incremental_state, "mems") + # if mems is not None: + # new_mems = [mems_i.index_select(1, new_order) for mems_i in mems] + # self.set_incremental_state(incremental_state, "mems", new_mems) diff --git a/examples/adaptive_span/truncated_bptt_lm_task.py b/examples/adaptive_span/truncated_bptt_lm_task.py new file mode 120000 index 0000000000000000000000000000000000000000..a92da3a298e21528b7007df3f8198bb3af94a485 --- /dev/null +++ b/examples/adaptive_span/truncated_bptt_lm_task.py @@ -0,0 +1 @@ +../truncated_bptt/truncated_bptt_lm_task.py \ No newline at end of file diff --git a/examples/backtranslation/README.md b/examples/backtranslation/README.md new file mode 100644 index 0000000000000000000000000000000000000000..73675f1125d80f58aa824db67d8970504d4d6b2a --- /dev/null +++ b/examples/backtranslation/README.md @@ -0,0 +1,297 @@ +# Understanding Back-Translation at Scale (Edunov et al., 2018) + +This page includes pre-trained models from the paper [Understanding Back-Translation at Scale (Edunov et al., 2018)](https://arxiv.org/abs/1808.09381). + +## Pre-trained models + +Model | Description | Dataset | Download +---|---|---|--- +`transformer.wmt18.en-de` | Transformer
([Edunov et al., 2018](https://arxiv.org/abs/1808.09381))
WMT'18 winner | [WMT'18 English-German](http://www.statmt.org/wmt18/translation-task.html) | [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/wmt18.en-de.ensemble.tar.gz)
See NOTE in the archive + +## Example usage (torch.hub) + +We require a few additional Python dependencies for preprocessing: +```bash +pip install subword_nmt sacremoses +``` + +Then to generate translations from the full model ensemble: +```python +import torch + +# List available models +torch.hub.list('pytorch/fairseq') # [..., 'transformer.wmt18.en-de', ... ] + +# Load the WMT'18 En-De ensemble +en2de_ensemble = torch.hub.load( + 'pytorch/fairseq', 'transformer.wmt18.en-de', + checkpoint_file='wmt18.model1.pt:wmt18.model2.pt:wmt18.model3.pt:wmt18.model4.pt:wmt18.model5.pt', + tokenizer='moses', bpe='subword_nmt') + +# The ensemble contains 5 models +len(en2de_ensemble.models) +# 5 + +# Translate +en2de_ensemble.translate('Hello world!') +# 'Hallo Welt!' +``` + +## Training your own model (WMT'18 English-German) + +The following instructions can be adapted to reproduce the models from the paper. + + +#### Step 1. Prepare parallel data and optionally train a baseline (English-German) model + +First download and preprocess the data: +```bash +# Download and prepare the data +cd examples/backtranslation/ +bash prepare-wmt18en2de.sh +cd ../.. + +# Binarize the data +TEXT=examples/backtranslation/wmt18_en_de +fairseq-preprocess \ + --joined-dictionary \ + --source-lang en --target-lang de \ + --trainpref $TEXT/train --validpref $TEXT/valid --testpref $TEXT/test \ + --destdir data-bin/wmt18_en_de --thresholdtgt 0 --thresholdsrc 0 \ + --workers 20 + +# Copy the BPE code into the data-bin directory for future use +cp examples/backtranslation/wmt18_en_de/code data-bin/wmt18_en_de/code +``` + +(Optionally) Train a baseline model (English-German) using just the parallel data: +```bash +CHECKPOINT_DIR=checkpoints_en_de_parallel +fairseq-train --fp16 \ + data-bin/wmt18_en_de \ + --source-lang en --target-lang de \ + --arch transformer_wmt_en_de_big --share-all-embeddings \ + --dropout 0.3 --weight-decay 0.0 \ + --criterion label_smoothed_cross_entropy --label-smoothing 0.1 \ + --optimizer adam --adam-betas '(0.9, 0.98)' --clip-norm 0.0 \ + --lr 0.001 --lr-scheduler inverse_sqrt --warmup-updates 4000 \ + --max-tokens 3584 --update-freq 16 \ + --max-update 30000 \ + --save-dir $CHECKPOINT_DIR +# Note: the above command assumes 8 GPUs. Adjust `--update-freq` if you have a +# different number of GPUs. +``` + +Average the last 10 checkpoints: +```bash +python scripts/average_checkpoints.py \ + --inputs $CHECKPOINT_DIR \ + --num-epoch-checkpoints 10 \ + --output $CHECKPOINT_DIR/checkpoint.avg10.pt +``` + +Evaluate BLEU: +```bash +# tokenized BLEU on newstest2017: +bash examples/backtranslation/tokenized_bleu.sh \ + wmt17 \ + en-de \ + data-bin/wmt18_en_de \ + data-bin/wmt18_en_de/code \ + $CHECKPOINT_DIR/checkpoint.avg10.pt +# BLEU4 = 29.57, 60.9/35.4/22.9/15.5 (BP=1.000, ratio=1.014, syslen=63049, reflen=62152) +# compare to 29.46 in Table 1, which is also for tokenized BLEU + +# generally it's better to report (detokenized) sacrebleu though: +bash examples/backtranslation/sacrebleu.sh \ + wmt17 \ + en-de \ + data-bin/wmt18_en_de \ + data-bin/wmt18_en_de/code \ + $CHECKPOINT_DIR/checkpoint.avg10.pt +# BLEU+case.mixed+lang.en-de+numrefs.1+smooth.exp+test.wmt17+tok.13a+version.1.4.3 = 29.0 60.6/34.7/22.4/14.9 (BP = 1.000 ratio = 1.013 hyp_len = 62099 ref_len = 61287) +``` + + +#### Step 2. Back-translate monolingual German data + +Train a reverse model (German-English) to do the back-translation: +```bash +CHECKPOINT_DIR=checkpoints_de_en_parallel +fairseq-train --fp16 \ + data-bin/wmt18_en_de \ + --source-lang de --target-lang en \ + --arch transformer_wmt_en_de_big --share-all-embeddings \ + --dropout 0.3 --weight-decay 0.0 \ + --criterion label_smoothed_cross_entropy --label-smoothing 0.1 \ + --optimizer adam --adam-betas '(0.9, 0.98)' --clip-norm 0.0 \ + --lr 0.001 --lr-scheduler inverse_sqrt --warmup-updates 4000 \ + --max-tokens 3584 --update-freq 16 \ + --max-update 30000 \ + --save-dir $CHECKPOINT_DIR +# Note: the above command assumes 8 GPUs. Adjust `--update-freq` if you have a +# different number of GPUs. +``` + +Let's evaluate the back-translation (BT) model to make sure it is well trained: +```bash +bash examples/backtranslation/sacrebleu.sh \ + wmt17 \ + de-en \ + data-bin/wmt18_en_de \ + data-bin/wmt18_en_de/code \ + $CHECKPOINT_DIR/checkpoint_best.py +# BLEU+case.mixed+lang.de-en+numrefs.1+smooth.exp+test.wmt17+tok.13a+version.1.4.3 = 34.9 66.9/41.8/28.5/19.9 (BP = 0.983 ratio = 0.984 hyp_len = 63342 ref_len = 64399) +# compare to the best system from WMT'17 which scored 35.1: http://matrix.statmt.org/matrix/systems_list/1868 +``` + +Next prepare the monolingual data: +```bash +# Download and prepare the monolingual data +# By default the script samples 25M monolingual sentences, which after +# deduplication should be just over 24M sentences. These are split into 25 +# shards, each with 1M sentences (except for the last shard). +cd examples/backtranslation/ +bash prepare-de-monolingual.sh +cd ../.. + +# Binarize each shard of the monolingual data +TEXT=examples/backtranslation/wmt18_de_mono +for SHARD in $(seq -f "%02g" 0 24); do \ + fairseq-preprocess \ + --only-source \ + --source-lang de --target-lang en \ + --joined-dictionary \ + --srcdict data-bin/wmt18_en_de/dict.de.txt \ + --testpref $TEXT/bpe.monolingual.dedup.${SHARD} \ + --destdir data-bin/wmt18_de_mono/shard${SHARD} \ + --workers 20; \ + cp data-bin/wmt18_en_de/dict.en.txt data-bin/wmt18_de_mono/shard${SHARD}/; \ +done +``` + +Now we're ready to perform back-translation over the monolingual data. The +following command generates via sampling, but it's possible to use greedy +decoding (`--beam 1`), beam search (`--beam 5`), +top-k sampling (`--sampling --beam 1 --sampling-topk 10`), etc.: +```bash +mkdir backtranslation_output +for SHARD in $(seq -f "%02g" 0 24); do \ + fairseq-generate --fp16 \ + data-bin/wmt18_de_mono/shard${SHARD} \ + --path $CHECKPOINT_DIR/checkpoint_best.pt \ + --skip-invalid-size-inputs-valid-test \ + --max-tokens 4096 \ + --sampling --beam 1 \ + > backtranslation_output/sampling.shard${SHARD}.out; \ +done +``` + +After BT, use the `extract_bt_data.py` script to re-combine the shards, extract +the back-translations and apply length ratio filters: +```bash +python examples/backtranslation/extract_bt_data.py \ + --minlen 1 --maxlen 250 --ratio 1.5 \ + --output backtranslation_output/bt_data --srclang en --tgtlang de \ + backtranslation_output/sampling.shard*.out + +# Ensure lengths are the same: +# wc -l backtranslation_output/bt_data.{en,de} +# 21795614 backtranslation_output/bt_data.en +# 21795614 backtranslation_output/bt_data.de +# 43591228 total +``` + +Binarize the filtered BT data and combine it with the parallel data: +```bash +TEXT=backtranslation_output +fairseq-preprocess \ + --source-lang en --target-lang de \ + --joined-dictionary \ + --srcdict data-bin/wmt18_en_de/dict.en.txt \ + --trainpref $TEXT/bt_data \ + --destdir data-bin/wmt18_en_de_bt \ + --workers 20 + +# We want to train on the combined data, so we'll symlink the parallel + BT data +# in the wmt18_en_de_para_plus_bt directory. We link the parallel data as "train" +# and the BT data as "train1", so that fairseq will combine them automatically +# and so that we can use the `--upsample-primary` option to upsample the +# parallel data (if desired). +PARA_DATA=$(readlink -f data-bin/wmt18_en_de) +BT_DATA=$(readlink -f data-bin/wmt18_en_de_bt) +COMB_DATA=data-bin/wmt18_en_de_para_plus_bt +mkdir -p $COMB_DATA +for LANG in en de; do \ + ln -s ${PARA_DATA}/dict.$LANG.txt ${COMB_DATA}/dict.$LANG.txt; \ + for EXT in bin idx; do \ + ln -s ${PARA_DATA}/train.en-de.$LANG.$EXT ${COMB_DATA}/train.en-de.$LANG.$EXT; \ + ln -s ${BT_DATA}/train.en-de.$LANG.$EXT ${COMB_DATA}/train1.en-de.$LANG.$EXT; \ + ln -s ${PARA_DATA}/valid.en-de.$LANG.$EXT ${COMB_DATA}/valid.en-de.$LANG.$EXT; \ + ln -s ${PARA_DATA}/test.en-de.$LANG.$EXT ${COMB_DATA}/test.en-de.$LANG.$EXT; \ + done; \ +done +``` + + +#### 3. Train an English-German model over the combined parallel + BT data + +Finally we can train a model over the parallel + BT data: +```bash +CHECKPOINT_DIR=checkpoints_en_de_parallel_plus_bt +fairseq-train --fp16 \ + data-bin/wmt18_en_de_para_plus_bt \ + --upsample-primary 16 \ + --source-lang en --target-lang de \ + --arch transformer_wmt_en_de_big --share-all-embeddings \ + --dropout 0.3 --weight-decay 0.0 \ + --criterion label_smoothed_cross_entropy --label-smoothing 0.1 \ + --optimizer adam --adam-betas '(0.9, 0.98)' --clip-norm 0.0 \ + --lr 0.0007 --lr-scheduler inverse_sqrt --warmup-updates 4000 \ + --max-tokens 3584 --update-freq 16 \ + --max-update 100000 \ + --save-dir $CHECKPOINT_DIR +# Note: the above command assumes 8 GPUs. Adjust `--update-freq` if you have a +# different number of GPUs. +``` + +Average the last 10 checkpoints: +```bash +python scripts/average_checkpoints.py \ + --inputs $CHECKPOINT_DIR \ + --num-epoch-checkpoints 10 \ + --output $CHECKPOINT_DIR/checkpoint.avg10.pt +``` + +Evaluate BLEU: +```bash +# tokenized BLEU on newstest2017: +bash examples/backtranslation/tokenized_bleu.sh \ + wmt17 \ + en-de \ + data-bin/wmt18_en_de \ + data-bin/wmt18_en_de/code \ + $CHECKPOINT_DIR/checkpoint.avg10.pt +# BLEU4 = 32.35, 64.4/38.9/26.2/18.3 (BP=0.977, ratio=0.977, syslen=60729, reflen=62152) +# compare to 32.35 in Table 1, which is also for tokenized BLEU + +# generally it's better to report (detokenized) sacrebleu: +bash examples/backtranslation/sacrebleu.sh \ + wmt17 \ + en-de \ + data-bin/wmt18_en_de \ + data-bin/wmt18_en_de/code \ + $CHECKPOINT_DIR/checkpoint.avg10.pt +# BLEU+case.mixed+lang.en-de+numrefs.1+smooth.exp+test.wmt17+tok.13a+version.1.4.3 = 31.5 64.3/38.2/25.6/17.6 (BP = 0.971 ratio = 0.971 hyp_len = 59515 ref_len = 61287) +``` + + +## Citation +```bibtex +@inproceedings{edunov2018backtranslation, + title = {Understanding Back-Translation at Scale}, + author = {Edunov, Sergey and Ott, Myle and Auli, Michael and Grangier, David}, + booktitle = {Conference of the Association for Computational Linguistics (ACL)}, + year = 2018, +} +``` diff --git a/examples/backtranslation/deduplicate_lines.py b/examples/backtranslation/deduplicate_lines.py new file mode 100644 index 0000000000000000000000000000000000000000..50e458328c80b71c42a66d473381ca7e98d294da --- /dev/null +++ b/examples/backtranslation/deduplicate_lines.py @@ -0,0 +1,41 @@ +#!/usr/bin/python3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import fileinput +import hashlib +import sys +from multiprocessing import Pool + + +def get_hashes_and_lines(raw_line): + hash = hashlib.md5(raw_line).hexdigest() + return hash, raw_line + + +def main(): + parser = argparse.ArgumentParser() + parser.add_argument("--workers", type=int, default=10) + parser.add_argument("files", nargs="*", help="input files") + args = parser.parse_args() + + seen = set() + with fileinput.input(args.files, mode="rb") as h: + pool = Pool(args.workers) + results = pool.imap_unordered(get_hashes_and_lines, h, 1000) + for i, (hash, raw_line) in enumerate(results): + if hash not in seen: + seen.add(hash) + sys.stdout.buffer.write(raw_line) + if i % 1000000 == 0: + print(i, file=sys.stderr, end="", flush=True) + elif i % 100000 == 0: + print(".", file=sys.stderr, end="", flush=True) + print(file=sys.stderr, flush=True) + + +if __name__ == "__main__": + main() diff --git a/examples/backtranslation/extract_bt_data.py b/examples/backtranslation/extract_bt_data.py new file mode 100644 index 0000000000000000000000000000000000000000..e766391e873d0d9a9561d67d5864934b2fad0681 --- /dev/null +++ b/examples/backtranslation/extract_bt_data.py @@ -0,0 +1,72 @@ +#!/usr/bin/env python +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import fileinput + +from tqdm import tqdm + + +def main(): + parser = argparse.ArgumentParser( + description=( + "Extract back-translations from the stdout of fairseq-generate. " + "If there are multiply hypotheses for a source, we only keep the first one. " + ) + ) + parser.add_argument("--output", required=True, help="output prefix") + parser.add_argument( + "--srclang", required=True, help="source language (extracted from H-* lines)" + ) + parser.add_argument( + "--tgtlang", required=True, help="target language (extracted from S-* lines)" + ) + parser.add_argument("--minlen", type=int, help="min length filter") + parser.add_argument("--maxlen", type=int, help="max length filter") + parser.add_argument("--ratio", type=float, help="ratio filter") + parser.add_argument("files", nargs="*", help="input files") + args = parser.parse_args() + + def validate(src, tgt): + srclen = len(src.split(" ")) if src != "" else 0 + tgtlen = len(tgt.split(" ")) if tgt != "" else 0 + if ( + (args.minlen is not None and (srclen < args.minlen or tgtlen < args.minlen)) + or ( + args.maxlen is not None + and (srclen > args.maxlen or tgtlen > args.maxlen) + ) + or ( + args.ratio is not None + and (max(srclen, tgtlen) / float(min(srclen, tgtlen)) > args.ratio) + ) + ): + return False + return True + + def safe_index(toks, index, default): + try: + return toks[index] + except IndexError: + return default + + with open(args.output + "." + args.srclang, "w") as src_h, open( + args.output + "." + args.tgtlang, "w" + ) as tgt_h: + for line in tqdm(fileinput.input(args.files)): + if line.startswith("S-"): + tgt = safe_index(line.rstrip().split("\t"), 1, "") + elif line.startswith("H-"): + if tgt is not None: + src = safe_index(line.rstrip().split("\t"), 2, "") + if validate(src, tgt): + print(src, file=src_h) + print(tgt, file=tgt_h) + tgt = None + + +if __name__ == "__main__": + main() diff --git a/examples/backtranslation/prepare-de-monolingual.sh b/examples/backtranslation/prepare-de-monolingual.sh new file mode 100644 index 0000000000000000000000000000000000000000..5e67b2b3bcf27d3436031453e796e58a0ae79ec4 --- /dev/null +++ b/examples/backtranslation/prepare-de-monolingual.sh @@ -0,0 +1,98 @@ +#!/bin/bash + +SCRIPTS=mosesdecoder/scripts +TOKENIZER=$SCRIPTS/tokenizer/tokenizer.perl +NORM_PUNC=$SCRIPTS/tokenizer/normalize-punctuation.perl +REM_NON_PRINT_CHAR=$SCRIPTS/tokenizer/remove-non-printing-char.perl +BPEROOT=subword-nmt/subword_nmt + + +BPE_CODE=wmt18_en_de/code +SUBSAMPLE_SIZE=25000000 +LANG=de + + +OUTDIR=wmt18_${LANG}_mono +orig=orig +tmp=$OUTDIR/tmp +mkdir -p $OUTDIR $tmp + + +URLS=( + "http://www.statmt.org/wmt14/training-monolingual-news-crawl/news.2007.de.shuffled.gz" + "http://www.statmt.org/wmt14/training-monolingual-news-crawl/news.2008.de.shuffled.gz" + "http://www.statmt.org/wmt14/training-monolingual-news-crawl/news.2009.de.shuffled.gz" + "http://www.statmt.org/wmt14/training-monolingual-news-crawl/news.2010.de.shuffled.gz" + "http://www.statmt.org/wmt14/training-monolingual-news-crawl/news.2011.de.shuffled.gz" + "http://www.statmt.org/wmt14/training-monolingual-news-crawl/news.2012.de.shuffled.gz" + "http://www.statmt.org/wmt14/training-monolingual-news-crawl/news.2013.de.shuffled.gz" + "http://www.statmt.org/wmt15/training-monolingual-news-crawl-v2/news.2014.de.shuffled.v2.gz" + "http://data.statmt.org/wmt16/translation-task/news.2015.de.shuffled.gz" + "http://data.statmt.org/wmt17/translation-task/news.2016.de.shuffled.gz" + "http://data.statmt.org/wmt18/translation-task/news.2017.de.shuffled.deduped.gz" +) +FILES=( + "news.2007.de.shuffled.gz" + "news.2008.de.shuffled.gz" + "news.2009.de.shuffled.gz" + "news.2010.de.shuffled.gz" + "news.2011.de.shuffled.gz" + "news.2012.de.shuffled.gz" + "news.2013.de.shuffled.gz" + "news.2014.de.shuffled.v2.gz" + "news.2015.de.shuffled.gz" + "news.2016.de.shuffled.gz" + "news.2017.de.shuffled.deduped.gz" +) + + +cd $orig +for ((i=0;i<${#URLS[@]};++i)); do + file=${FILES[i]} + if [ -f $file ]; then + echo "$file already exists, skipping download" + else + url=${URLS[i]} + wget "$url" + fi +done +cd .. + + +if [ -f $tmp/monolingual.${SUBSAMPLE_SIZE}.${LANG} ]; then + echo "found monolingual sample, skipping shuffle/sample/tokenize" +else + gzip -c -d -k $(for FILE in "${FILES[@]}"; do echo $orig/$FILE; done) \ + | shuf -n $SUBSAMPLE_SIZE \ + | perl $NORM_PUNC $LANG \ + | perl $REM_NON_PRINT_CHAR \ + | perl $TOKENIZER -threads 8 -a -l $LANG \ + > $tmp/monolingual.${SUBSAMPLE_SIZE}.${LANG} +fi + + +if [ -f $tmp/bpe.monolingual.${SUBSAMPLE_SIZE}.${LANG} ]; then + echo "found BPE monolingual sample, skipping BPE step" +else + python $BPEROOT/apply_bpe.py -c $BPE_CODE \ + < $tmp/monolingual.${SUBSAMPLE_SIZE}.${LANG} \ + > $tmp/bpe.monolingual.${SUBSAMPLE_SIZE}.${LANG} +fi + + +if [ -f $tmp/bpe.monolingual.dedup.${SUBSAMPLE_SIZE}.${LANG} ]; then + echo "found deduplicated monolingual sample, skipping deduplication step" +else + python deduplicate_lines.py $tmp/bpe.monolingual.${SUBSAMPLE_SIZE}.${LANG} \ + > $tmp/bpe.monolingual.dedup.${SUBSAMPLE_SIZE}.${LANG} +fi + + +if [ -f $OUTDIR/bpe.monolingual.dedup.00.de ]; then + echo "found sharded data, skipping sharding step" +else + split --lines 1000000 --numeric-suffixes \ + --additional-suffix .${LANG} \ + $tmp/bpe.monolingual.dedup.${SUBSAMPLE_SIZE}.${LANG} \ + $OUTDIR/bpe.monolingual.dedup. +fi diff --git a/examples/backtranslation/prepare-wmt18en2de.sh b/examples/backtranslation/prepare-wmt18en2de.sh new file mode 100644 index 0000000000000000000000000000000000000000..f6fd275307db50ca84c299440ae02dce49064030 --- /dev/null +++ b/examples/backtranslation/prepare-wmt18en2de.sh @@ -0,0 +1,135 @@ +#!/bin/bash +# Adapted from https://github.com/facebookresearch/MIXER/blob/master/prepareData.sh + +echo 'Cloning Moses github repository (for tokenization scripts)...' +git clone https://github.com/moses-smt/mosesdecoder.git + +echo 'Cloning Subword NMT repository (for BPE pre-processing)...' +git clone https://github.com/rsennrich/subword-nmt.git + +SCRIPTS=mosesdecoder/scripts +TOKENIZER=$SCRIPTS/tokenizer/tokenizer.perl +CLEAN=$SCRIPTS/training/clean-corpus-n.perl +NORM_PUNC=$SCRIPTS/tokenizer/normalize-punctuation.perl +REM_NON_PRINT_CHAR=$SCRIPTS/tokenizer/remove-non-printing-char.perl +BPEROOT=subword-nmt/subword_nmt +BPE_TOKENS=32000 + +URLS=( + "http://statmt.org/wmt13/training-parallel-europarl-v7.tgz" + "http://statmt.org/wmt13/training-parallel-commoncrawl.tgz" + "http://data.statmt.org/wmt18/translation-task/training-parallel-nc-v13.tgz" + "http://data.statmt.org/wmt18/translation-task/rapid2016.tgz" + "http://data.statmt.org/wmt17/translation-task/dev.tgz" + "http://statmt.org/wmt14/test-full.tgz" +) +FILES=( + "training-parallel-europarl-v7.tgz" + "training-parallel-commoncrawl.tgz" + "training-parallel-nc-v13.tgz" + "rapid2016.tgz" + "dev.tgz" + "test-full.tgz" +) +CORPORA=( + "training/europarl-v7.de-en" + "commoncrawl.de-en" + "training-parallel-nc-v13/news-commentary-v13.de-en" + "rapid2016.de-en" +) + +if [ ! -d "$SCRIPTS" ]; then + echo "Please set SCRIPTS variable correctly to point to Moses scripts." + exit 1 +fi + +OUTDIR=wmt18_en_de + +src=en +tgt=de +lang=en-de +prep=$OUTDIR +tmp=$prep/tmp +orig=orig + +mkdir -p $orig $tmp $prep + +cd $orig + +for ((i=0;i<${#URLS[@]};++i)); do + file=${FILES[i]} + if [ -f $file ]; then + echo "$file already exists, skipping download" + else + url=${URLS[i]} + wget "$url" + if [ -f $file ]; then + echo "$url successfully downloaded." + else + echo "$url not successfully downloaded." + exit 1 + fi + if [ ${file: -4} == ".tgz" ]; then + tar zxvf $file + elif [ ${file: -4} == ".tar" ]; then + tar xvf $file + fi + fi +done +cd .. + +echo "pre-processing train data..." +for l in $src $tgt; do + rm $tmp/train.tags.$lang.tok.$l + for f in "${CORPORA[@]}"; do + cat $orig/$f.$l | \ + perl $NORM_PUNC $l | \ + perl $REM_NON_PRINT_CHAR | \ + perl $TOKENIZER -threads 8 -a -l $l >> $tmp/train.tags.$lang.tok.$l + done +done + +echo "pre-processing test data..." +for l in $src $tgt; do + if [ "$l" == "$src" ]; then + t="src" + else + t="ref" + fi + grep '\s*//g' | \ + sed -e 's/\s*<\/seg>\s*//g' | \ + sed -e "s/\’/\'/g" | \ + perl $TOKENIZER -threads 8 -a -l $l > $tmp/test.$l + echo "" +done + +echo "splitting train and valid..." +for l in $src $tgt; do + awk '{if (NR%100 == 0) print $0; }' $tmp/train.tags.$lang.tok.$l > $tmp/valid.$l + awk '{if (NR%100 != 0) print $0; }' $tmp/train.tags.$lang.tok.$l > $tmp/train.$l +done + +TRAIN=$tmp/train.de-en +BPE_CODE=$prep/code +rm -f $TRAIN +for l in $src $tgt; do + cat $tmp/train.$l >> $TRAIN +done + +echo "learn_bpe.py on ${TRAIN}..." +python $BPEROOT/learn_bpe.py -s $BPE_TOKENS < $TRAIN > $BPE_CODE + +for L in $src $tgt; do + for f in train.$L valid.$L test.$L; do + echo "apply_bpe.py to ${f}..." + python $BPEROOT/apply_bpe.py -c $BPE_CODE < $tmp/$f > $tmp/bpe.$f + done +done + +perl $CLEAN -ratio 1.5 $tmp/bpe.train $src $tgt $prep/train 1 250 +perl $CLEAN -ratio 1.5 $tmp/bpe.valid $src $tgt $prep/valid 1 250 + +for L in $src $tgt; do + cp $tmp/bpe.test.$L $prep/test.$L +done diff --git a/examples/backtranslation/sacrebleu.sh b/examples/backtranslation/sacrebleu.sh new file mode 100644 index 0000000000000000000000000000000000000000..a70da23f48e2699297799611412783d4560dc45a --- /dev/null +++ b/examples/backtranslation/sacrebleu.sh @@ -0,0 +1,37 @@ +#!/bin/bash + +if [ $# -ne 5 ]; then + echo "usage: $0 [dataset=wmt14/full] [langpair=en-de] [databin] [bpecode] [model]" + exit +fi + + +DATASET=$1 +LANGPAIR=$2 +DATABIN=$3 +BPECODE=$4 +MODEL=$5 + +SRCLANG=$(echo $LANGPAIR | cut -d '-' -f 1) +TGTLANG=$(echo $LANGPAIR | cut -d '-' -f 2) + + +BPEROOT=examples/backtranslation/subword-nmt/subword_nmt +if [ ! -e $BPEROOT ]; then + BPEROOT=subword-nmt/subword_nmt + if [ ! -e $BPEROOT ]; then + echo 'Cloning Subword NMT repository (for BPE pre-processing)...' + git clone https://github.com/rsennrich/subword-nmt.git + fi +fi + + +sacrebleu -t $DATASET -l $LANGPAIR --echo src \ +| sacremoses tokenize -a -l $SRCLANG -q \ +| python $BPEROOT/apply_bpe.py -c $BPECODE \ +| fairseq-interactive $DATABIN --path $MODEL \ + -s $SRCLANG -t $TGTLANG \ + --beam 5 --remove-bpe --buffer-size 1024 --max-tokens 8000 \ +| grep ^H- | cut -f 3- \ +| sacremoses detokenize -l $TGTLANG -q \ +| sacrebleu -t $DATASET -l $LANGPAIR diff --git a/examples/backtranslation/tokenized_bleu.sh b/examples/backtranslation/tokenized_bleu.sh new file mode 100644 index 0000000000000000000000000000000000000000..c6d6aaa193f6059299bc98909324fe4b9b060372 --- /dev/null +++ b/examples/backtranslation/tokenized_bleu.sh @@ -0,0 +1,46 @@ +#!/bin/bash + +if [ $# -ne 5 ]; then + echo "usage: $0 [dataset=wmt14/full] [langpair=en-de] [databin] [bpecode] [model]" + exit +fi + + +DATASET=$1 +LANGPAIR=$2 +DATABIN=$3 +BPECODE=$4 +MODEL=$5 + +SRCLANG=$(echo $LANGPAIR | cut -d '-' -f 1) +TGTLANG=$(echo $LANGPAIR | cut -d '-' -f 2) + + +BPEROOT=examples/backtranslation/subword-nmt/subword_nmt +if [ ! -e $BPEROOT ]; then + BPEROOT=subword-nmt/subword_nmt + if [ ! -e $BPEROOT ]; then + echo 'Cloning Subword NMT repository (for BPE pre-processing)...' + git clone https://github.com/rsennrich/subword-nmt.git + fi +fi + + +TMP_REF=$(mktemp) + +sacrebleu -t $DATASET -l $LANGPAIR --echo ref -q \ +| sacremoses normalize -l $TGTLANG -q \ +| sacremoses tokenize -a -l $TGTLANG -q \ +> $TMP_REF + +sacrebleu -t $DATASET -l $LANGPAIR --echo src -q \ +| sacremoses normalize -l $SRCLANG -q \ +| sacremoses tokenize -a -l $SRCLANG -q \ +| python $BPEROOT/apply_bpe.py -c $BPECODE \ +| fairseq-interactive $DATABIN --path $MODEL \ + -s $SRCLANG -t $TGTLANG \ + --beam 5 --remove-bpe --buffer-size 1024 --max-tokens 8000 \ +| grep ^H- | cut -f 3- \ +| fairseq-score --ref $TMP_REF + +rm -f $TMP_REF diff --git a/examples/bart/README.glue.md b/examples/bart/README.glue.md new file mode 100644 index 0000000000000000000000000000000000000000..a010934e1e6dec491eb1c704ec02ba7405760510 --- /dev/null +++ b/examples/bart/README.glue.md @@ -0,0 +1,99 @@ +# Fine-tuning BART on GLUE tasks + +### 1) Download the data from GLUE website (https://gluebenchmark.com/tasks) using following commands: +```bash +wget https://gist.githubusercontent.com/W4ngatang/60c2bdb54d156a41194446737ce03e2e/raw/17b8dd0d724281ed7c3b2aeeda662b92809aadd5/download_glue_data.py +python download_glue_data.py --data_dir glue_data --tasks all +``` + +### 2) Preprocess GLUE task data (same as RoBERTa): +```bash +./examples/roberta/preprocess_GLUE_tasks.sh glue_data +``` +`glue_task_name` is one of the following: +`{ALL, QQP, MNLI, QNLI, MRPC, RTE, STS-B, SST-2, CoLA}` +Use `ALL` for preprocessing all the glue tasks. + +### 3) Fine-tuning on GLUE task: +Example fine-tuning cmd for `RTE` task +```bash +TOTAL_NUM_UPDATES=2036 # 10 epochs through RTE for bsz 16 +WARMUP_UPDATES=61 # 6 percent of the number of updates +LR=1e-05 # Peak LR for polynomial LR scheduler. +NUM_CLASSES=2 +MAX_SENTENCES=16 # Batch size. +BART_PATH=/path/to/bart/model.pt + +CUDA_VISIBLE_DEVICES=0,1 fairseq-train RTE-bin/ \ + --restore-file $BART_PATH \ + --batch-size $MAX_SENTENCES \ + --max-tokens 4400 \ + --task sentence_prediction \ + --add-prev-output-tokens \ + --layernorm-embedding \ + --share-all-embeddings \ + --share-decoder-input-output-embed \ + --reset-optimizer --reset-dataloader --reset-meters \ + --required-batch-size-multiple 1 \ + --init-token 0 \ + --arch bart_large \ + --criterion sentence_prediction \ + --num-classes $NUM_CLASSES \ + --dropout 0.1 --attention-dropout 0.1 \ + --weight-decay 0.01 --optimizer adam --adam-betas "(0.9, 0.98)" --adam-eps 1e-08 \ + --clip-norm 0.0 \ + --lr-scheduler polynomial_decay --lr $LR --total-num-update $TOTAL_NUM_UPDATES --warmup-updates $WARMUP_UPDATES \ + --fp16 --fp16-init-scale 4 --threshold-loss-scale 1 --fp16-scale-window 128 \ + --max-epoch 10 \ + --find-unused-parameters \ + --best-checkpoint-metric accuracy --maximize-best-checkpoint-metric; +``` + +For each of the GLUE task, you will need to use following cmd-line arguments: + +Model | MNLI | QNLI | QQP | RTE | SST-2 | MRPC | CoLA | STS-B +---|---|---|---|---|---|---|---|--- +`--num-classes` | 3 | 2 | 2 | 2 | 2 | 2 | 2 | 1 +`--lr` | 5e-6 | 1e-5 | 1e-5 | 1e-5 | 5e-6 | 2e-5 | 2e-5 | 2e-5 +`bsz` | 128 | 32 | 32 | 32 | 128 | 64 | 64 | 32 +`--total-num-update` | 30968 | 33112 | 113272 | 1018 | 5233 | 1148 | 1334 | 1799 +`--warmup-updates` | 1858 | 1986 | 6796 | 61 | 314 | 68 | 80 | 107 + +For `STS-B` additionally add `--regression-target --best-checkpoint-metric loss` and remove `--maximize-best-checkpoint-metric`. + +**Note:** + +a) `--total-num-updates` is used by `--polynomial_decay` scheduler and is calculated for `--max-epoch=10` and `--batch-size=32/64/128` depending on the task. + +b) Above cmd-args and hyperparams are tested on Nvidia `V100` GPU with `32gb` of memory for each task. Depending on the GPU memory resources available to you, you can use increase `--update-freq` and reduce `--batch-size`. + +### Inference on GLUE task +After training the model as mentioned in previous step, you can perform inference with checkpoints in `checkpoints/` directory using following python code snippet: + +```python +from fairseq.models.bart import BARTModel + +bart = BARTModel.from_pretrained( + 'checkpoints/', + checkpoint_file='checkpoint_best.pt', + data_name_or_path='RTE-bin' +) + +label_fn = lambda label: bart.task.label_dictionary.string( + [label + bart.task.label_dictionary.nspecial] +) +ncorrect, nsamples = 0, 0 +bart.cuda() +bart.eval() +with open('glue_data/RTE/dev.tsv') as fin: + fin.readline() + for index, line in enumerate(fin): + tokens = line.strip().split('\t') + sent1, sent2, target = tokens[1], tokens[2], tokens[3] + tokens = bart.encode(sent1, sent2) + prediction = bart.predict('sentence_classification_head', tokens).argmax().item() + prediction_label = label_fn(prediction) + ncorrect += int(prediction_label == target) + nsamples += 1 +print('| Accuracy: ', float(ncorrect)/float(nsamples)) +``` diff --git a/examples/bart/README.md b/examples/bart/README.md new file mode 100644 index 0000000000000000000000000000000000000000..4050a724ee6a2f20c9998a95df48c58b64764ab1 --- /dev/null +++ b/examples/bart/README.md @@ -0,0 +1,228 @@ +# BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension + +[https://arxiv.org/abs/1910.13461](https://arxiv.org/abs/1910.13461) + +## Introduction + +BART is sequence-to-sequence model trained with denoising as pretraining objective. We show that this pretraining objective is more generic and show that we can match [RoBERTa](../roberta) results on SQuAD and GLUE and gain state-of-the-art results on summarization (XSum, CNN dataset), long form generative question answering (ELI5) and dialog response genration (ConvAI2). See the associated paper for more details. + +## Pre-trained models + +Model | Description | # params | Download +---|---|---|--- +`bart.base` | BART model with 6 encoder and decoder layers | 140M | [bart.base.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/bart.base.tar.gz) +`bart.large` | BART model with 12 encoder and decoder layers | 400M | [bart.large.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/bart.large.tar.gz) +`bart.large.mnli` | `bart.large` finetuned on `MNLI` | 400M | [bart.large.mnli.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/bart.large.mnli.tar.gz) +`bart.large.cnn` | `bart.large` finetuned on `CNN-DM` | 400M | [bart.large.cnn.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/bart.large.cnn.tar.gz) +`bart.large.xsum` | `bart.large` finetuned on `Xsum` | 400M | [bart.large.xsum.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/bart.large.xsum.tar.gz) + +## Results + +**[GLUE (Wang et al., 2019)](https://gluebenchmark.com/)** +_(dev set, single model, single-task finetuning)_ + +Model | MNLI | QNLI | QQP | RTE | SST-2 | MRPC | CoLA | STS-B +---|---|---|---|---|---|---|---|--- +`roberta.large` | 90.2 | 94.7 | 92.2 | 86.6 | 96.4 | 90.9 | 68.0 | 92.4 +`bart.large` | 89.9 | 94.9 | 92.5 | 87.0 | 96.6 | 90.4 | 62.8 | 91.2 + +**[SQuAD (Rajpurkar et al., 2018)](https://rajpurkar.github.io/SQuAD-explorer/)** +_(dev set, no additional data used)_ + +Model | SQuAD 1.1 EM/F1 | SQuAD 2.0 EM/F1 +---|---|--- +`roberta.large` | 88.9/94.6 | 86.5/89.4 +`bart.large` | 88.8/94.6 | 86.1/89.2 + +**[CNN/Daily Mail](http://nlpprogress.com/english/summarization.html)** +_(test set, no additional data used)_ + +Model | R1 | R2 | RL +---|---|---|--- +`BERTSUMEXTABS` | 42.13 | 19.60 | 39.18 +`bart.large` | 44.16 | 21.28 | 40.90 + +## Example usage + +##### Load BART from torch.hub (PyTorch >= 1.1): +```python +import torch +bart = torch.hub.load('pytorch/fairseq', 'bart.large') +bart.eval() # disable dropout (or leave in train mode to finetune) +``` + +##### Load BART (for PyTorch 1.0 or custom models): +```python +# Download bart.large model +wget https://dl.fbaipublicfiles.com/fairseq/models/bart.large.tar.gz +tar -xzvf bart.large.tar.gz + +# Load the model in fairseq +from fairseq.models.bart import BARTModel +bart = BARTModel.from_pretrained('/path/to/bart.large', checkpoint_file='model.pt') +bart.eval() # disable dropout (or leave in train mode to finetune) +``` + +##### Apply Byte-Pair Encoding (BPE) to input text: +```python +tokens = bart.encode('Hello world!') +assert tokens.tolist() == [0, 31414, 232, 328, 2] +bart.decode(tokens) # 'Hello world!' +``` + +##### Extract features from BART: +```python +# Extract the last layer's features +last_layer_features = bart.extract_features(tokens) +assert last_layer_features.size() == torch.Size([1, 5, 1024]) + +# Extract all layer's features from decoder (layer 0 is the embedding layer) +all_layers = bart.extract_features(tokens, return_all_hiddens=True) +assert len(all_layers) == 13 +assert torch.all(all_layers[-1] == last_layer_features) +``` + +##### Use BART for sentence-pair classification tasks: +```python +# Download BART already finetuned for MNLI +bart = torch.hub.load('pytorch/fairseq', 'bart.large.mnli') +bart.eval() # disable dropout for evaluation + +# Encode a pair of sentences and make a prediction +tokens = bart.encode('BART is a seq2seq model.', 'BART is not sequence to sequence.') +bart.predict('mnli', tokens).argmax() # 0: contradiction + +# Encode another pair of sentences +tokens = bart.encode('BART is denoising autoencoder.', 'BART is version of autoencoder.') +bart.predict('mnli', tokens).argmax() # 2: entailment +``` + +##### Register a new (randomly initialized) classification head: +```python +bart.register_classification_head('new_task', num_classes=3) +logprobs = bart.predict('new_task', tokens) +``` + +##### Batched prediction: +```python +import torch +from fairseq.data.data_utils import collate_tokens + +bart = torch.hub.load('pytorch/fairseq', 'bart.large.mnli') +bart.eval() + +batch_of_pairs = [ + ['BART is a seq2seq model.', 'BART is not sequence to sequence.'], + ['BART is denoising autoencoder.', 'BART is version of autoencoder.'], +] + +batch = collate_tokens( + [bart.encode(pair[0], pair[1]) for pair in batch_of_pairs], pad_idx=1 +) + +logprobs = bart.predict('mnli', batch) +print(logprobs.argmax(dim=1)) +# tensor([0, 2]) +``` + +##### Using the GPU: +```python +bart.cuda() +bart.predict('new_task', tokens) +``` + +#### Filling masks: + +BART can be used to fill multiple `` tokens in the input. +```python +bart = torch.hub.load('pytorch/fairseq', 'bart.base') +bart.eval() +bart.fill_mask(['The cat on the .'], topk=3, beam=10) +# [[('The cat was on the ground.', tensor(-0.6183)), ('The cat was on the floor.', tensor(-0.6798)), ('The cat sleeps on the couch.', tensor(-0.6830))]] +``` + +Note that by default we enforce the output length to match the input length. +This can be disabled by setting ``match_source_len=False``: +``` +bart.fill_mask(['The cat on the .'], topk=3, beam=10, match_source_len=False) +# [[('The cat was on the ground.', tensor(-0.6185)), ('The cat was asleep on the couch.', tensor(-0.6276)), ('The cat was on the floor.', tensor(-0.6800))]] +``` + +Example code to fill masks for a batch of sentences using GPU +``` +bart.cuda() +bart.fill_mask(['The cat on the .', 'The dog on the .'], topk=3, beam=10) +# [[('The cat was on the ground.', tensor(-0.6183)), ('The cat was on the floor.', tensor(-0.6798)), ('The cat sleeps on the couch.', tensor(-0.6830))], [('The dog was on the ground.', tensor(-0.6190)), ('The dog lay on the ground.', tensor(-0.6711)), +('The dog was asleep on the couch', tensor(-0.6796))]] +``` + +#### Evaluating the `bart.large.mnli` model: + +Example python code snippet to evaluate accuracy on the MNLI `dev_matched` set. +```python +label_map = {0: 'contradiction', 1: 'neutral', 2: 'entailment'} +ncorrect, nsamples = 0, 0 +bart.cuda() +bart.eval() +with open('glue_data/MNLI/dev_matched.tsv') as fin: + fin.readline() + for index, line in enumerate(fin): + tokens = line.strip().split('\t') + sent1, sent2, target = tokens[8], tokens[9], tokens[-1] + tokens = bart.encode(sent1, sent2) + prediction = bart.predict('mnli', tokens).argmax().item() + prediction_label = label_map[prediction] + ncorrect += int(prediction_label == target) + nsamples += 1 + print('| Accuracy: ', float(ncorrect)/float(nsamples)) +# Expected output: 0.9010 +``` + +#### Evaluating the `bart.large.cnn` model: +- Follow instructions [here](https://github.com/abisee/cnn-dailymail) to download and process into data-files such that `test.source` and `test.target` has one line for each non-tokenized sample. +- For simpler preprocessing, you can also `wget https://cdn-datasets.huggingface.co/summarization/cnn_dm_v2.tgz`, although there is no guarantee of identical scores +- `huggingface/transformers` has a simpler interface that supports [single-gpu](https://github.com/huggingface/transformers/blob/master/examples/legacy/seq2seq/run_eval.py) and [multi-gpu](https://github.com/huggingface/transformers/blob/master/examples/legacy/seq2seq/run_distributed_eval.py) beam search. + In `huggingface/transformers`, the BART models' paths are `facebook/bart-large-cnn` and `facebook/bart-large-xsum`. + +In `fairseq`, summaries can be generated using: + +```bash +cp data-bin/cnn_dm/dict.source.txt checkpoints/ +python examples/bart/summarize.py \ + --model-dir pytorch/fairseq \ + --model-file bart.large.cnn \ + --src cnn_dm/test.source \ + --out cnn_dm/test.hypo +``` + +For calculating rouge, install `files2rouge` from [here](https://github.com/pltrdy/files2rouge). + +```bash +export CLASSPATH=/path/to/stanford-corenlp-full-2016-10-31/stanford-corenlp-3.7.0.jar + +# Tokenize hypothesis and target files. +cat test.hypo | java edu.stanford.nlp.process.PTBTokenizer -ioFileList -preserveLines > test.hypo.tokenized +cat test.target | java edu.stanford.nlp.process.PTBTokenizer -ioFileList -preserveLines > test.hypo.target +files2rouge test.hypo.tokenized test.hypo.target +# Expected output: (ROUGE-2 Average_F: 0.21238) +``` + + +## Finetuning + +- [Finetuning on GLUE](README.glue.md) +- [Finetuning on CNN-DM](README.summarization.md) + +## Citation + +```bibtex +@article{lewis2019bart, + title = {BART: Denoising Sequence-to-Sequence Pre-training for Natural +Language Generation, Translation, and Comprehension}, + author = {Mike Lewis and Yinhan Liu and Naman Goyal and Marjan Ghazvininejad and + Abdelrahman Mohamed and Omer Levy and Veselin Stoyanov + and Luke Zettlemoyer }, + journal={arXiv preprint arXiv:1910.13461}, + year = {2019}, +} +``` diff --git a/examples/bart/README.summarization.md b/examples/bart/README.summarization.md new file mode 100644 index 0000000000000000000000000000000000000000..8727584f2b2bdd880c6cd3abbf39b75dfbf4a67c --- /dev/null +++ b/examples/bart/README.summarization.md @@ -0,0 +1,102 @@ +# Fine-tuning BART on CNN-Dailymail summarization task + +### 1) Download the CNN and Daily Mail data and preprocess it into data files with non-tokenized cased samples. + +Follow the instructions [here](https://github.com/abisee/cnn-dailymail) to download the original CNN and Daily Mail datasets. To preprocess the data, refer to the pointers in [this issue](https://github.com/pytorch/fairseq/issues/1391) or check out the code [here](https://github.com/artmatsak/cnn-dailymail). + +Follow the instructions [here](https://github.com/EdinburghNLP/XSum) to download the original Extreme Summarization datasets, or check out the code [here](https://github.com/EdinburghNLP/XSum/tree/master/XSum-Dataset), Please keep the raw dataset and make sure no tokenization nor BPE on the dataset. + +### 2) BPE preprocess: + +```bash +wget -N 'https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/encoder.json' +wget -N 'https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/vocab.bpe' +wget -N 'https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/dict.txt' + +TASK=cnn_dm +for SPLIT in train val +do + for LANG in source target + do + python -m examples.roberta.multiprocessing_bpe_encoder \ + --encoder-json encoder.json \ + --vocab-bpe vocab.bpe \ + --inputs "$TASK/$SPLIT.$LANG" \ + --outputs "$TASK/$SPLIT.bpe.$LANG" \ + --workers 60 \ + --keep-empty; + done +done +``` + +### 3) Binarize dataset: +```bash +fairseq-preprocess \ + --source-lang "source" \ + --target-lang "target" \ + --trainpref "${TASK}/train.bpe" \ + --validpref "${TASK}/val.bpe" \ + --destdir "${TASK}-bin/" \ + --workers 60 \ + --srcdict dict.txt \ + --tgtdict dict.txt; +``` + +### 4) Fine-tuning on CNN-DM summarization task: +Example fine-tuning CNN-DM +```bash +TOTAL_NUM_UPDATES=20000 +WARMUP_UPDATES=500 +LR=3e-05 +MAX_TOKENS=2048 +UPDATE_FREQ=4 +BART_PATH=/path/to/bart/model.pt + +CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 fairseq-train cnn_dm-bin \ + --restore-file $BART_PATH \ + --max-tokens $MAX_TOKENS \ + --task translation \ + --source-lang source --target-lang target \ + --truncate-source \ + --layernorm-embedding \ + --share-all-embeddings \ + --share-decoder-input-output-embed \ + --reset-optimizer --reset-dataloader --reset-meters \ + --required-batch-size-multiple 1 \ + --arch bart_large \ + --criterion label_smoothed_cross_entropy \ + --label-smoothing 0.1 \ + --dropout 0.1 --attention-dropout 0.1 \ + --weight-decay 0.01 --optimizer adam --adam-betas "(0.9, 0.999)" --adam-eps 1e-08 \ + --clip-norm 0.1 \ + --lr-scheduler polynomial_decay --lr $LR --total-num-update $TOTAL_NUM_UPDATES --warmup-updates $WARMUP_UPDATES \ + --fp16 --update-freq $UPDATE_FREQ \ + --skip-invalid-size-inputs-valid-test \ + --find-unused-parameters; +``` +Above is expected to run on `1` node with `8 32gb-V100`. +Expected training time is about `5 hours`. Training time can be reduced with distributed training on `4` nodes and `--update-freq 1`. + +Use TOTAL_NUM_UPDATES=15000 UPDATE_FREQ=2 for Xsum task + +### Inference for CNN-DM test data using above trained checkpoint. +After training the model as mentioned in previous step, you can perform inference with checkpoints in `checkpoints/` directory using `eval_cnn.py`, for example + +```bash +cp data-bin/cnn_dm/dict.source.txt checkpoints/ +python examples/bart/summarize.py \ + --model-dir checkpoints \ + --model-file checkpoint_best.pt \ + --src cnn_dm/test.source \ + --out cnn_dm/test.hypo +``` +For XSUM, which uses beam=6, lenpen=1.0, max_len_b=60, min_len=10: +```bash +cp data-bin/cnn_dm/dict.source.txt checkpoints/ +python examples/bart/summarize.py \ + --model-dir checkpoints \ + --model-file checkpoint_best.pt \ + --src cnn_dm/test.source \ + --out cnn_dm/test.hypo \ + --xsum-kwargs +``` diff --git a/examples/bart/summarize.py b/examples/bart/summarize.py new file mode 100644 index 0000000000000000000000000000000000000000..04435f80e39c2d9d894696dae7cba5b381e13da9 --- /dev/null +++ b/examples/bart/summarize.py @@ -0,0 +1,100 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +from fairseq.models.bart import BARTModel +import argparse + +XSUM_KWARGS = dict(beam=6, lenpen=1.0, max_len_b=60, min_len=10, no_repeat_ngram_size=3) +CNN_KWARGS = dict(beam=4, lenpen=2.0, max_len_b=140, min_len=55, no_repeat_ngram_size=3) + + +@torch.no_grad() +def generate(bart, infile, outfile="bart_hypo.txt", bsz=32, n_obs=None, **eval_kwargs): + count = 1 + + # if n_obs is not None: bsz = min(bsz, n_obs) + + with open(infile) as source, open(outfile, "w") as fout: + sline = source.readline().strip() + slines = [sline] + for sline in source: + if n_obs is not None and count > n_obs: + break + if count % bsz == 0: + hypotheses_batch = bart.sample(slines, **eval_kwargs) + for hypothesis in hypotheses_batch: + fout.write(hypothesis + "\n") + fout.flush() + slines = [] + + slines.append(sline.strip()) + count += 1 + + if slines != []: + hypotheses_batch = bart.sample(slines, **eval_kwargs) + for hypothesis in hypotheses_batch: + fout.write(hypothesis + "\n") + fout.flush() + + +def main(): + """ + Usage:: + + python examples/bart/summarize.py \ + --model-dir $HOME/bart.large.cnn \ + --model-file model.pt \ + --src $HOME/data-bin/cnn_dm/test.source + """ + parser = argparse.ArgumentParser() + parser.add_argument( + "--model-dir", + required=True, + type=str, + default="bart.large.cnn/", + help="path containing model file and src_dict.txt", + ) + parser.add_argument( + "--model-file", + default="checkpoint_best.pt", + help="where in model_dir are weights saved", + ) + parser.add_argument( + "--src", default="test.source", help="text to summarize", type=str + ) + parser.add_argument( + "--out", default="test.hypo", help="where to save summaries", type=str + ) + parser.add_argument("--bsz", default=32, help="where to save summaries", type=int) + parser.add_argument( + "--n", default=None, help="how many examples to summarize", type=int + ) + parser.add_argument( + "--xsum-kwargs", + action="store_true", + default=False, + help="if true use XSUM_KWARGS else CNN_KWARGS", + ) + args = parser.parse_args() + eval_kwargs = XSUM_KWARGS if args.xsum_kwargs else CNN_KWARGS + if args.model_dir == "pytorch/fairseq": + bart = torch.hub.load("pytorch/fairseq", args.model_file) + else: + bart = BARTModel.from_pretrained( + args.model_dir, + checkpoint_file=args.model_file, + data_name_or_path=args.model_dir, + ) + bart = bart.eval() + if torch.cuda.is_available(): + bart = bart.cuda().half() + generate( + bart, args.src, bsz=args.bsz, n_obs=args.n, outfile=args.out, **eval_kwargs + ) + + +if __name__ == "__main__": + main() diff --git a/examples/byte_level_bpe/README.md b/examples/byte_level_bpe/README.md new file mode 100644 index 0000000000000000000000000000000000000000..657092660eae42d20f67647417623b8b8cb7b66c --- /dev/null +++ b/examples/byte_level_bpe/README.md @@ -0,0 +1,88 @@ +# Neural Machine Translation with Byte-Level Subwords + +https://arxiv.org/abs/1909.03341 + +We provide an implementation of byte-level byte-pair encoding (BBPE), taking IWSLT 2017 Fr-En translation as +example. + +## Data +Get data and generate fairseq binary dataset: +```bash +bash ./get_data.sh +``` + +## Model Training +Train Transformer model with Bi-GRU embedding contextualization (implemented in `gru_transformer.py`): +```bash +# VOCAB=bytes +# VOCAB=chars +VOCAB=bbpe2048 +# VOCAB=bpe2048 +# VOCAB=bbpe4096 +# VOCAB=bpe4096 +# VOCAB=bpe16384 +``` +```bash +fairseq-train "data/bin_${VOCAB}" --task translation --user-dir examples/byte_level_bpe/gru_transformer \ + --arch gru_transformer --encoder-layers 2 --decoder-layers 2 --dropout 0.3 --share-all-embeddings \ + --optimizer adam --adam-betas '(0.9, 0.98)' \ + --lr 5e-4 --lr-scheduler inverse_sqrt --warmup-updates 4000 \ + --criterion label_smoothed_cross_entropy --label-smoothing 0.1 \ + --log-format 'simple' --log-interval 100 --save-dir "checkpoints/${VOCAB}" \ + --batch-size 100 --max-update 100000 --update-freq 2 +``` + +## Generation +`fairseq-generate` requires bytes (BBPE) decoder to convert byte-level representation back to characters: +```bash +# BPE=--bpe bytes +# BPE=--bpe characters +BPE=--bpe byte_bpe --sentencepiece-model-path data/spm_bbpe2048.model +# BPE=--bpe sentencepiece --sentencepiece-model data/spm_bpe2048.model +# BPE=--bpe byte_bpe --sentencepiece-model-path data/spm_bbpe4096.model +# BPE=--bpe sentencepiece --sentencepiece-model data/spm_bpe4096.model +# BPE=--bpe sentencepiece --sentencepiece-model data/spm_bpe16384.model +``` + +```bash +fairseq-generate "data/bin_${VOCAB}" --task translation --user-dir examples/byte_level_bpe/gru_transformer \ + --source-lang fr --gen-subset test --sacrebleu --path "checkpoints/${VOCAB}/checkpoint_last.pt" \ + --tokenizer moses --moses-target-lang en ${BPE} +``` +When using `fairseq-interactive`, bytes (BBPE) encoder/decoder is required to tokenize input data and detokenize model predictions: +```bash +fairseq-interactive "data/bin_${VOCAB}" --task translation --user-dir examples/byte_level_bpe/gru_transformer \ + --path "checkpoints/${VOCAB}/checkpoint_last.pt" --input data/test.fr --tokenizer moses --moses-source-lang fr \ + --moses-target-lang en ${BPE} --buffer-size 1000 --max-tokens 10000 +``` + +## Results +| Vocabulary | Model | BLEU | +|:-------------:|:-------------:|:-------------:| +| Joint BPE 16k ([Kudo, 2018](https://arxiv.org/abs/1804.10959)) | 512d LSTM 2+2 | 33.81 | +| Joint BPE 16k | Transformer base 2+2 (w/ GRU) | 36.64 (36.72) | +| Joint BPE 4k | Transformer base 2+2 (w/ GRU) | 35.49 (36.10) | +| Joint BBPE 4k | Transformer base 2+2 (w/ GRU) | 35.61 (35.82) | +| Joint BPE 2k | Transformer base 2+2 (w/ GRU) | 34.87 (36.13) | +| Joint BBPE 2k | Transformer base 2+2 (w/ GRU) | 34.98 (35.43) | +| Characters | Transformer base 2+2 (w/ GRU) | 31.78 (33.30) | +| Bytes | Transformer base 2+2 (w/ GRU) | 31.57 (33.62) | + + +## Citation +``` +@misc{wang2019neural, + title={Neural Machine Translation with Byte-Level Subwords}, + author={Changhan Wang and Kyunghyun Cho and Jiatao Gu}, + year={2019}, + eprint={1909.03341}, + archivePrefix={arXiv}, + primaryClass={cs.CL} +} +``` + + +## Contact +Changhan Wang ([changhan@fb.com](mailto:changhan@fb.com)), +Kyunghyun Cho ([kyunghyuncho@fb.com](mailto:kyunghyuncho@fb.com)), +Jiatao Gu ([jgu@fb.com](mailto:jgu@fb.com)) diff --git a/examples/byte_level_bpe/get_bitext.py b/examples/byte_level_bpe/get_bitext.py new file mode 100644 index 0000000000000000000000000000000000000000..6ac1eeec1e6167ec6bafd76b37173ee6987cae7e --- /dev/null +++ b/examples/byte_level_bpe/get_bitext.py @@ -0,0 +1,254 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +import argparse +import os +import os.path as op +from collections import namedtuple +from multiprocessing import cpu_count +from typing import List, Optional + +import sentencepiece as sp +from fairseq.data.encoders.byte_bpe import ByteBPE +from fairseq.data.encoders.byte_utils import byte_encode +from fairseq.data.encoders.bytes import Bytes +from fairseq.data.encoders.characters import Characters +from fairseq.data.encoders.moses_tokenizer import MosesTokenizer +from fairseq.data.encoders.sentencepiece_bpe import SentencepieceBPE + + +SPLITS = ["train", "valid", "test"] + + +def _convert_xml(in_path: str, out_path: str): + with open(in_path) as f, open(out_path, "w") as f_o: + for s in f: + ss = s.strip() + if not ss.startswith("", "").split('">') + assert len(ss) == 2 + f_o.write(ss[1].strip() + "\n") + + +def _convert_train(in_path: str, out_path: str): + with open(in_path) as f, open(out_path, "w") as f_o: + for s in f: + ss = s.strip() + if ss.startswith("<"): + continue + f_o.write(ss.strip() + "\n") + + +def _get_bytes(in_path: str, out_path: str): + with open(in_path) as f, open(out_path, "w") as f_o: + for s in f: + f_o.write(Bytes.encode(s.strip()) + "\n") + + +def _get_chars(in_path: str, out_path: str): + with open(in_path) as f, open(out_path, "w") as f_o: + for s in f: + f_o.write(Characters.encode(s.strip()) + "\n") + + +def pretokenize(in_path: str, out_path: str, src: str, tgt: str): + Args = namedtuple( + "Args", + [ + "moses_source_lang", + "moses_target_lang", + "moses_no_dash_splits", + "moses_no_escape", + ], + ) + args = Args( + moses_source_lang=src, + moses_target_lang=tgt, + moses_no_dash_splits=False, + moses_no_escape=False, + ) + pretokenizer = MosesTokenizer(args) + with open(in_path) as f, open(out_path, "w") as f_o: + for s in f: + f_o.write(pretokenizer.encode(s.strip()) + "\n") + + +def _convert_to_bchar(in_path_prefix: str, src: str, tgt: str, out_path: str): + with open(out_path, "w") as f_o: + for lang in [src, tgt]: + with open(f"{in_path_prefix}.{lang}") as f: + for s in f: + f_o.write(byte_encode(s.strip()) + "\n") + + +def _get_bpe(in_path: str, model_prefix: str, vocab_size: int): + arguments = [ + f"--input={in_path}", + f"--model_prefix={model_prefix}", + f"--model_type=bpe", + f"--vocab_size={vocab_size}", + "--character_coverage=1.0", + "--normalization_rule_name=identity", + f"--num_threads={cpu_count()}", + ] + sp.SentencePieceTrainer.Train(" ".join(arguments)) + + +def _apply_bbpe(model_path: str, in_path: str, out_path: str): + Args = namedtuple("Args", ["sentencepiece_model_path"]) + args = Args(sentencepiece_model_path=model_path) + tokenizer = ByteBPE(args) + with open(in_path) as f, open(out_path, "w") as f_o: + for s in f: + f_o.write(tokenizer.encode(s.strip()) + "\n") + + +def _apply_bpe(model_path: str, in_path: str, out_path: str): + Args = namedtuple("Args", ["sentencepiece_model"]) + args = Args(sentencepiece_model=model_path) + tokenizer = SentencepieceBPE(args) + with open(in_path) as f, open(out_path, "w") as f_o: + for s in f: + f_o.write(tokenizer.encode(s.strip()) + "\n") + + +def _concat_files(in_paths: List[str], out_path: str): + with open(out_path, "w") as f_o: + for p in in_paths: + with open(p) as f: + for r in f: + f_o.write(r) + + +def preprocess_iwslt17( + root: str, + src: str, + tgt: str, + bpe_size: Optional[int], + need_chars: bool, + bbpe_size: Optional[int], + need_bytes: bool, +): + # extract bitext + in_root = op.join(root, f"{src}-{tgt}") + for lang in [src, tgt]: + _convert_train( + op.join(in_root, f"train.tags.{src}-{tgt}.{lang}"), + op.join(root, f"train.{lang}"), + ) + _convert_xml( + op.join(in_root, f"IWSLT17.TED.dev2010.{src}-{tgt}.{lang}.xml"), + op.join(root, f"valid.{lang}"), + ) + _convert_xml( + op.join(in_root, f"IWSLT17.TED.tst2015.{src}-{tgt}.{lang}.xml"), + op.join(root, f"test.{lang}"), + ) + # pre-tokenize + for lang in [src, tgt]: + for split in SPLITS: + pretokenize( + op.join(root, f"{split}.{lang}"), + op.join(root, f"{split}.moses.{lang}"), + src, + tgt, + ) + # tokenize with BPE vocabulary + if bpe_size is not None: + # learn vocabulary + concated_train_path = op.join(root, "train.all") + _concat_files( + [op.join(root, "train.moses.fr"), op.join(root, "train.moses.en")], + concated_train_path, + ) + bpe_model_prefix = op.join(root, f"spm_bpe{bpe_size}") + _get_bpe(concated_train_path, bpe_model_prefix, bpe_size) + os.remove(concated_train_path) + # apply + for lang in [src, tgt]: + for split in SPLITS: + _apply_bpe( + bpe_model_prefix + ".model", + op.join(root, f"{split}.moses.{lang}"), + op.join(root, f"{split}.moses.bpe{bpe_size}.{lang}"), + ) + # tokenize with bytes vocabulary + if need_bytes: + for lang in [src, tgt]: + for split in SPLITS: + _get_bytes( + op.join(root, f"{split}.moses.{lang}"), + op.join(root, f"{split}.moses.bytes.{lang}"), + ) + # tokenize with characters vocabulary + if need_chars: + for lang in [src, tgt]: + for split in SPLITS: + _get_chars( + op.join(root, f"{split}.moses.{lang}"), + op.join(root, f"{split}.moses.chars.{lang}"), + ) + # tokenize with byte-level BPE vocabulary + if bbpe_size is not None: + # learn vocabulary + bchar_path = op.join(root, "train.bchar") + _convert_to_bchar(op.join(root, "train.moses"), src, tgt, bchar_path) + bbpe_model_prefix = op.join(root, f"spm_bbpe{bbpe_size}") + _get_bpe(bchar_path, bbpe_model_prefix, bbpe_size) + os.remove(bchar_path) + # apply + for lang in [src, tgt]: + for split in SPLITS: + _apply_bbpe( + bbpe_model_prefix + ".model", + op.join(root, f"{split}.moses.{lang}"), + op.join(root, f"{split}.moses.bbpe{bbpe_size}.{lang}"), + ) + + +def main(): + parser = argparse.ArgumentParser() + parser.add_argument("--root", type=str, default="data") + parser.add_argument( + "--bpe-vocab", + default=None, + type=int, + help="Generate tokenized bitext with BPE of size K." + "Default to None (disabled).", + ) + parser.add_argument( + "--bbpe-vocab", + default=None, + type=int, + help="Generate tokenized bitext with BBPE of size K." + "Default to None (disabled).", + ) + parser.add_argument( + "--byte-vocab", + action="store_true", + help="Generate tokenized bitext with bytes vocabulary", + ) + parser.add_argument( + "--char-vocab", + action="store_true", + help="Generate tokenized bitext with chars vocabulary", + ) + args = parser.parse_args() + + preprocess_iwslt17( + args.root, + "fr", + "en", + args.bpe_vocab, + args.char_vocab, + args.bbpe_vocab, + args.byte_vocab, + ) + + +if __name__ == "__main__": + main() diff --git a/examples/byte_level_bpe/get_data.sh b/examples/byte_level_bpe/get_data.sh new file mode 100644 index 0000000000000000000000000000000000000000..c3d55d4925a6e6e23d12d293f093c1ae14acf76e --- /dev/null +++ b/examples/byte_level_bpe/get_data.sh @@ -0,0 +1,47 @@ +#!/bin/bash + +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +PY_BIN_ROOT= + +# PyPI dependency +${PY_BIN_ROOT}pip install sentencepiece sacremoses + +# Get data +if [ ! -d "data" ]; then + mkdir data +fi + +if [ ! -f "data/fr-en.tgz" ]; then + wget https://wit3.fbk.eu/archive/2017-01-trnted/texts/fr/en/fr-en.tgz -P data + tar xvf data/fr-en.tgz -C data +fi +${PY_BIN_ROOT}python get_bitext.py --bpe-vocab 16384 --byte-vocab --char-vocab +for VOCAB_SIZE in 2048 4096; do + ${PY_BIN_ROOT}python get_bitext.py --bpe-vocab ${VOCAB_SIZE} --bbpe-vocab ${VOCAB_SIZE} +done +rm -r data/fr-en data/fr-en.tgz + +# Generate binary dataset +${PY_BIN_ROOT}/fairseq-preprocess --source-lang fr --target-lang en --destdir data/bin_bpe16384 --joined-dictionary \ + --workers "$(nproc)" --trainpref data/train.moses.bpe16384 --validpref data/valid.moses.bpe16384 \ + --testpref data/test.moses.bpe16384 + +${PY_BIN_ROOT}/fairseq-preprocess --source-lang fr --target-lang en --destdir data/bin_bytes --joined-dictionary \ + --workers "$(nproc)" --trainpref data/train.moses.bytes --validpref data/valid.moses.bytes \ + --testpref data/test.moses.bytes + +${PY_BIN_ROOT}/fairseq-preprocess --source-lang fr --target-lang en --destdir data/bin_chars --joined-dictionary \ + --workers "$(nproc)" --trainpref data/train.moses.chars --validpref data/valid.moses.chars \ + --testpref data/test.moses.chars + +for VOCAB_SIZE in 2048 4096; do + for TYPE in bbpe bpe; do + ${PY_BIN_ROOT}/fairseq-preprocess --source-lang fr --target-lang en --destdir "data/bin_${TYPE}${VOCAB_SIZE}" \ + --joined-dictionary --workers "$(nproc)" --trainpref "data/train.moses.${TYPE}${VOCAB_SIZE}" \ + --validpref "data/valid.moses.${TYPE}${VOCAB_SIZE}" --testpref "data/test.moses.${TYPE}${VOCAB_SIZE}" + done +done diff --git a/examples/byte_level_bpe/gru_transformer.py b/examples/byte_level_bpe/gru_transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..d4efa93a4d75da71c78e786d7f62101ef3266af4 --- /dev/null +++ b/examples/byte_level_bpe/gru_transformer.py @@ -0,0 +1,107 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch.nn as nn +import torch.nn.functional as F +from fairseq.models import register_model, register_model_architecture +from fairseq.models.transformer import TransformerEncoder, TransformerModel + + +@register_model("gru_transformer") +class GRUTransformerModel(TransformerModel): + @classmethod + def build_encoder(cls, args, src_dict, embed_tokens): + return GRUTransformerEncoder(args, src_dict, embed_tokens) + + +class GRUTransformerEncoder(TransformerEncoder): + def __init__(self, args, dictionary, embed_tokens): + super().__init__(args, dictionary, embed_tokens) + self.emb_ctx = nn.GRU( + input_size=embed_tokens.embedding_dim, + hidden_size=embed_tokens.embedding_dim // 2, + num_layers=1, + bidirectional=True, + ) + + def forward_embedding(self, src_tokens): + # embed tokens and positions + x = embed = self.embed_scale * self.embed_tokens(src_tokens) + if self.embed_positions is not None: + x = embed + self.embed_positions(src_tokens) + + # contextualize embeddings + x = x.transpose(0, 1) + x = self.dropout_module(x) + x, _ = self.emb_ctx.forward(x) + x = x.transpose(0, 1) + + if self.layernorm_embedding is not None: + x = self.layernorm_embedding(x) + x = self.dropout_module(x) + return x, embed + + +@register_model_architecture("gru_transformer", "gru_transformer") +def gru_transformer_base_architecture(args): + args.encoder_embed_path = getattr(args, "encoder_embed_path", None) + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 2048) + args.encoder_layers = getattr(args, "encoder_layers", 6) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 8) + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False) + args.encoder_learned_pos = getattr(args, "encoder_learned_pos", False) + args.decoder_embed_path = getattr(args, "decoder_embed_path", None) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", args.encoder_embed_dim) + args.decoder_ffn_embed_dim = getattr( + args, "decoder_ffn_embed_dim", args.encoder_ffn_embed_dim + ) + args.decoder_layers = getattr(args, "decoder_layers", 6) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 8) + args.decoder_normalize_before = getattr(args, "decoder_normalize_before", False) + args.decoder_learned_pos = getattr(args, "decoder_learned_pos", False) + args.attention_dropout = getattr(args, "attention_dropout", 0.0) + args.activation_dropout = getattr(args, "activation_dropout", 0.0) + args.activation_fn = getattr(args, "activation_fn", "relu") + args.dropout = getattr(args, "dropout", 0.1) + args.adaptive_softmax_cutoff = getattr(args, "adaptive_softmax_cutoff", None) + args.adaptive_softmax_dropout = getattr(args, "adaptive_softmax_dropout", 0) + args.share_decoder_input_output_embed = getattr( + args, "share_decoder_input_output_embed", False + ) + args.share_all_embeddings = getattr(args, "share_all_embeddings", False) + args.no_token_positional_embeddings = getattr( + args, "no_token_positional_embeddings", False + ) + args.adaptive_input = getattr(args, "adaptive_input", False) + args.no_cross_attention = getattr(args, "no_cross_attention", False) + args.cross_self_attention = getattr(args, "cross_self_attention", False) + args.layer_wise_attention = getattr(args, "layer_wise_attention", False) + + args.decoder_output_dim = getattr( + args, "decoder_output_dim", args.decoder_embed_dim + ) + args.decoder_input_dim = getattr(args, "decoder_input_dim", args.decoder_embed_dim) + + args.no_scale_embedding = getattr(args, "no_scale_embedding", False) + args.layernorm_embedding = getattr(args, "layernorm_embedding", False) + + +@register_model_architecture("gru_transformer", "gru_transformer_big") +def gru_transformer_big(args): + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 1024) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 4096) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 16) + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 1024) + args.decoder_ffn_embed_dim = getattr(args, "decoder_ffn_embed_dim", 4096) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 16) + args.dropout = getattr(args, "dropout", 0.3) + gru_transformer_base_architecture(args) diff --git a/examples/camembert/README.md b/examples/camembert/README.md new file mode 100644 index 0000000000000000000000000000000000000000..5ef4fe3f151bb468712f3be935ea5bb1b1360bf7 --- /dev/null +++ b/examples/camembert/README.md @@ -0,0 +1,75 @@ +# CamemBERT: a Tasty French Language Model + +## Introduction + +[CamemBERT](https://arxiv.org/abs/1911.03894) is a pretrained language model trained on 138GB of French text based on RoBERTa. + +Also available in [github.com/huggingface/transformers](https://github.com/huggingface/transformers/). + +## Pre-trained models + +| Model | #params | Download | Arch. | Training data | +|--------------------------------|---------|--------------------------------------------------------------------------------------------------------------------------|-------|-----------------------------------| +| `camembert` / `camembert-base` | 110M | [camembert-base.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/camembert-base.tar.gz) | Base | OSCAR (138 GB of text) | +| `camembert-large` | 335M | [camembert-large.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/camembert-large.tar.gz) | Large | CCNet (135 GB of text) | +| `camembert-base-ccnet` | 110M | [camembert-base-ccnet.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/camembert-base-ccnet.tar.gz) | Base | CCNet (135 GB of text) | +| `camembert-base-wikipedia-4gb` | 110M | [camembert-base-wikipedia-4gb.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/camembert-base-wikipedia-4gb.tar.gz) | Base | Wikipedia (4 GB of text) | +| `camembert-base-oscar-4gb` | 110M | [camembert-base-oscar-4gb.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/camembert-base-oscar-4gb.tar.gz) | Base | Subsample of OSCAR (4 GB of text) | +| `camembert-base-ccnet-4gb` | 110M | [camembert-base-ccnet-4gb.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/camembert-base-ccnet-4gb.tar.gz) | Base | Subsample of CCNet (4 GB of text) | + +## Example usage + +### fairseq +##### Load CamemBERT from torch.hub (PyTorch >= 1.1): +```python +import torch +camembert = torch.hub.load('pytorch/fairseq', 'camembert') +camembert.eval() # disable dropout (or leave in train mode to finetune) +``` + +##### Load CamemBERT (for PyTorch 1.0 or custom models): +```python +# Download camembert model +wget https://dl.fbaipublicfiles.com/fairseq/models/camembert-base.tar.gz +tar -xzvf camembert.tar.gz + +# Load the model in fairseq +from fairseq.models.roberta import CamembertModel +camembert = CamembertModel.from_pretrained('/path/to/camembert') +camembert.eval() # disable dropout (or leave in train mode to finetune) +``` + +##### Filling masks: +```python +masked_line = 'Le camembert est :)' +camembert.fill_mask(masked_line, topk=3) +# [('Le camembert est délicieux :)', 0.4909118115901947, ' délicieux'), +# ('Le camembert est excellent :)', 0.10556942224502563, ' excellent'), +# ('Le camembert est succulent :)', 0.03453322499990463, ' succulent')] +``` + +##### Extract features from Camembert: +```python +# Extract the last layer's features +line = "J'aime le camembert !" +tokens = camembert.encode(line) +last_layer_features = camembert.extract_features(tokens) +assert last_layer_features.size() == torch.Size([1, 10, 768]) + +# Extract all layer's features (layer 0 is the embedding layer) +all_layers = camembert.extract_features(tokens, return_all_hiddens=True) +assert len(all_layers) == 13 +assert torch.all(all_layers[-1] == last_layer_features) +``` + +## Citation +If you use our work, please cite: + +```bibtex +@inproceedings{martin2020camembert, + title={CamemBERT: a Tasty French Language Model}, + author={Martin, Louis and Muller, Benjamin and Su{\'a}rez, Pedro Javier Ortiz and Dupont, Yoann and Romary, Laurent and de la Clergerie, {\'E}ric Villemonte and Seddah, Djam{\'e} and Sagot, Beno{\^\i}t}, + booktitle={Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics}, + year={2020} +} +``` diff --git a/examples/constrained_decoding/README.md b/examples/constrained_decoding/README.md new file mode 100644 index 0000000000000000000000000000000000000000..cfca9c91fdb65e64b80af54f2d89f6b5f0db61d0 --- /dev/null +++ b/examples/constrained_decoding/README.md @@ -0,0 +1,123 @@ +# (Vectorized) Lexically constrained decoding with dynamic beam allocation + +This page provides instructions for how to use lexically constrained decoding in Fairseq. +Fairseq implements the code described in the following papers: + +* [Fast Lexically Constrained Decoding With Dynamic Beam Allocation](https://www.aclweb.org/anthology/N18-1119/) (Post & Vilar, 2018) +* [Improved Lexically Constrained Decoding for Translation and Monolingual Rewriting](https://www.aclweb.org/anthology/N19-1090/) (Hu et al., 2019) + +## Quick start + +Constrained search is enabled by adding the command-line argument `--constraints` to `fairseq-interactive`. +Constraints are appended to each line of input, separated by tabs. Each constraint (one or more tokens) +is a separate field. + +The following command, using [Fairseq's WMT19 German--English model](https://github.com/pytorch/fairseq/blob/master/examples/wmt19/README.md), +translates the sentence *Die maschinelle Übersetzung ist schwer zu kontrollieren.* with the constraints +"hard" and "to influence". + + echo -e "Die maschinelle Übersetzung ist schwer zu kontrollieren.\thard\ttoinfluence" \ + | normalize.py | tok.py \ + | fairseq-interactive /path/to/model \ + --path /path/to/model/model1.pt \ + --bpe fastbpe \ + --bpe-codes /path/to/model/bpecodes \ + --constraints \ + -s de -t en \ + --beam 10 + +(tok.py and normalize.py can be found in the same directory as this README; they are just shortcuts around Fairseq's WMT19 preprocessing). +This will generate the following output: + + [snip] + S-0 Die masch@@ in@@ elle Über@@ setzung ist schwer zu kontrollieren . + W-0 1.844 seconds + C-0 hard + C-0 influence + H-0 -1.5333266258239746 Mach@@ ine trans@@ lation is hard to influence . + D-0 -1.5333266258239746 Machine translation is hard to influence . + P-0 -0.5434 -0.1423 -0.1930 -0.1415 -0.2346 -1.8031 -0.1701 -11.7727 -0.1815 -0.1511 + +By default, constraints are generated in the order supplied, with any number (zero or more) of tokens generated +between constraints. If you wish for the decoder to order the constraints, then use `--constraints unordered`. +Note that you may want to use a larger beam. + +## Implementation details + +The heart of the implementation is in `fairseq/search.py`, which adds a `LexicallyConstrainedBeamSearch` instance. +This instance of beam search tracks the progress of each hypothesis in the beam through the set of constraints +provided for each input sentence. It does this using one of two classes, both found in `fairseq/token_generation_contstraints.py`: + +* OrderedConstraintState: assumes the `C` input constraints will be generated in the provided order +* UnorderedConstraintState: tries to apply `C` (phrasal) constraints in all `C!` orders + +## Differences from Sockeye + +There are a number of [differences from Sockeye's implementation](https://awslabs.github.io/sockeye/inference.html#lexical-constraints). + +* Generating constraints in the order supplied (the default option here) is not available in Sockeye. +* Due to an improved beam allocation method, there is no need to prune the beam. +* Again due to better allocation, beam sizes as low as 10 or even 5 are often sufficient. +* [The vector extensions described in Hu et al.](https://github.com/edwardjhu/sockeye/tree/trie_constraints) (NAACL 2019) were never merged + into the main Sockeye branch. + +## Citation + +The paper first describing lexical constraints for seq2seq decoding is: + +```bibtex +@inproceedings{hokamp-liu-2017-lexically, + title = "Lexically Constrained Decoding for Sequence Generation Using Grid Beam Search", + author = "Hokamp, Chris and + Liu, Qun", + booktitle = "Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)", + month = jul, + year = "2017", + address = "Vancouver, Canada", + publisher = "Association for Computational Linguistics", + url = "https://www.aclweb.org/anthology/P17-1141", + doi = "10.18653/v1/P17-1141", + pages = "1535--1546", +} +``` + +The fairseq implementation uses the extensions described in + +```bibtex +@inproceedings{post-vilar-2018-fast, + title = "Fast Lexically Constrained Decoding with Dynamic Beam Allocation for Neural Machine Translation", + author = "Post, Matt and + Vilar, David", + booktitle = "Proceedings of the 2018 Conference of the North {A}merican Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers)", + month = jun, + year = "2018", + address = "New Orleans, Louisiana", + publisher = "Association for Computational Linguistics", + url = "https://www.aclweb.org/anthology/N18-1119", + doi = "10.18653/v1/N18-1119", + pages = "1314--1324", +} +``` + +and + +```bibtex +@inproceedings{hu-etal-2019-improved, + title = "Improved Lexically Constrained Decoding for Translation and Monolingual Rewriting", + author = "Hu, J. Edward and + Khayrallah, Huda and + Culkin, Ryan and + Xia, Patrick and + Chen, Tongfei and + Post, Matt and + Van Durme, Benjamin", + booktitle = "Proceedings of the 2019 Conference of the North {A}merican Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers)", + month = jun, + year = "2019", + address = "Minneapolis, Minnesota", + publisher = "Association for Computational Linguistics", + url = "https://www.aclweb.org/anthology/N19-1090", + doi = "10.18653/v1/N19-1090", + pages = "839--850", +} +``` diff --git a/examples/constrained_decoding/normalize.py b/examples/constrained_decoding/normalize.py new file mode 100755 index 0000000000000000000000000000000000000000..4ae2b5111ba025acb9e1613865c92fdc339a58d5 --- /dev/null +++ b/examples/constrained_decoding/normalize.py @@ -0,0 +1,27 @@ +#!/usr/bin/env python3 +# +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import sys + +from sacremoses.normalize import MosesPunctNormalizer + + +def main(args): + normalizer = MosesPunctNormalizer(lang=args.lang, penn=args.penn) + for line in sys.stdin: + print(normalizer.normalize(line.rstrip()), flush=True) + + +if __name__ == "__main__": + import argparse + + parser = argparse.ArgumentParser() + parser.add_argument("--lang", "-l", default="en") + parser.add_argument("--penn", "-p", action="store_true") + args = parser.parse_args() + + main(args) diff --git a/examples/constrained_decoding/tok.py b/examples/constrained_decoding/tok.py new file mode 100755 index 0000000000000000000000000000000000000000..b1f888a8c0d1b8ec7174859476cc3222456e0d2c --- /dev/null +++ b/examples/constrained_decoding/tok.py @@ -0,0 +1,34 @@ +#!/usr/bin/env python3 +# +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import sys + +import sacremoses + + +def main(args): + """Tokenizes, preserving tabs""" + mt = sacremoses.MosesTokenizer(lang=args.lang) + + def tok(s): + return mt.tokenize(s, return_str=True) + + for line in sys.stdin: + parts = list(map(tok, line.split("\t"))) + print(*parts, sep="\t", flush=True) + + +if __name__ == "__main__": + import argparse + + parser = argparse.ArgumentParser() + parser.add_argument("--lang", "-l", default="en") + parser.add_argument("--penn", "-p", action="store_true") + parser.add_argument("--fields", "-f", help="fields to tokenize") + args = parser.parse_args() + + main(args) diff --git a/examples/conv_seq2seq/README.md b/examples/conv_seq2seq/README.md new file mode 100644 index 0000000000000000000000000000000000000000..95fe7e7909a77ee0e50fe31d4b8be38daa8f3be7 --- /dev/null +++ b/examples/conv_seq2seq/README.md @@ -0,0 +1,25 @@ +# Convolutional Sequence to Sequence Learning (Gehring et al., 2017) + +## Pre-trained models + +Description | Dataset | Model | Test set(s) +---|---|---|--- +Convolutional
([Gehring et al., 2017](https://arxiv.org/abs/1705.03122)) | [WMT14 English-French](http://statmt.org/wmt14/translation-task.html#Download) | [download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/models/wmt14.v2.en-fr.fconv-py.tar.bz2) | newstest2014:
[download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/data/wmt14.v2.en-fr.newstest2014.tar.bz2)
newstest2012/2013:
[download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/data/wmt14.v2.en-fr.ntst1213.tar.bz2) +Convolutional
([Gehring et al., 2017](https://arxiv.org/abs/1705.03122)) | [WMT14 English-German](http://statmt.org/wmt14/translation-task.html#Download) | [download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/models/wmt14.en-de.fconv-py.tar.bz2) | newstest2014:
[download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/data/wmt14.en-de.newstest2014.tar.bz2) +Convolutional
([Gehring et al., 2017](https://arxiv.org/abs/1705.03122)) | [WMT17 English-German](http://statmt.org/wmt17/translation-task.html#Download) | [download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/models/wmt17.v2.en-de.fconv-py.tar.bz2) | newstest2014:
[download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/data/wmt17.v2.en-de.newstest2014.tar.bz2) + +## Example usage + +See the [translation README](../translation/README.md) for instructions on reproducing results for WMT'14 En-De and +WMT'14 En-Fr using the `fconv_wmt_en_de` and `fconv_wmt_en_fr` model architectures. + +## Citation + +```bibtex +@inproceedings{gehring2017convs2s, + title = {Convolutional Sequence to Sequence Learning}, + author = {Gehring, Jonas, and Auli, Michael and Grangier, David and Yarats, Denis and Dauphin, Yann N}, + booktitle = {Proc. of ICML}, + year = 2017, +} +``` diff --git a/examples/criss/README.md b/examples/criss/README.md new file mode 100644 index 0000000000000000000000000000000000000000..4689ed7c10497a5100b28fe6d6801a7c089da569 --- /dev/null +++ b/examples/criss/README.md @@ -0,0 +1,61 @@ +# Cross-lingual Retrieval for Iterative Self-Supervised Training + +https://arxiv.org/pdf/2006.09526.pdf + +## Introduction + +CRISS is a multilingual sequence-to-sequnce pretraining method where mining and training processes are applied iteratively, improving cross-lingual alignment and translation ability at the same time. + +## Requirements: + +* faiss: https://github.com/facebookresearch/faiss +* mosesdecoder: https://github.com/moses-smt/mosesdecoder +* flores: https://github.com/facebookresearch/flores +* LASER: https://github.com/facebookresearch/LASER + +## Unsupervised Machine Translation +##### 1. Download and decompress CRISS checkpoints +``` +cd examples/criss +wget https://dl.fbaipublicfiles.com/criss/criss_3rd_checkpoints.tar.gz +tar -xf criss_checkpoints.tar.gz +``` +##### 2. Download and preprocess Flores test dataset +Make sure to run all scripts from examples/criss directory +``` +bash download_and_preprocess_flores_test.sh +``` + +##### 3. Run Evaluation on Sinhala-English +``` +bash unsupervised_mt/eval.sh +``` + +## Sentence Retrieval +##### 1. Download and preprocess Tatoeba dataset +``` +bash download_and_preprocess_tatoeba.sh +``` + +##### 2. Run Sentence Retrieval on Tatoeba Kazakh-English +``` +bash sentence_retrieval/sentence_retrieval_tatoeba.sh +``` + +## Mining +##### 1. Install faiss +Follow instructions on https://github.com/facebookresearch/faiss/blob/master/INSTALL.md +##### 2. Mine pseudo-parallel data between Kazakh and English +``` +bash mining/mine_example.sh +``` + +## Citation +```bibtex +@article{tran2020cross, + title={Cross-lingual retrieval for iterative self-supervised training}, + author={Tran, Chau and Tang, Yuqing and Li, Xian and Gu, Jiatao}, + journal={arXiv preprint arXiv:2006.09526}, + year={2020} +} +``` diff --git a/examples/criss/download_and_preprocess_flores_test.sh b/examples/criss/download_and_preprocess_flores_test.sh new file mode 100644 index 0000000000000000000000000000000000000000..ed4b390fbdee3991efeb298050e12065d7fe605b --- /dev/null +++ b/examples/criss/download_and_preprocess_flores_test.sh @@ -0,0 +1,64 @@ +#!/bin/bash +# Copyright (c) Facebook, Inc. and its affiliates. +# All rights reserved. +# +# This source code is licensed under the license found in the +# LICENSE file in the root directory of this source tree. + +SPM_ENCODE=flores/scripts/spm_encode.py +DATA=data_tmp +SPM_MODEL=criss_checkpoints/sentence.bpe.model +DICT=criss_checkpoints/dict.txt + +download_data() { + CORPORA=$1 + URL=$2 + + if [ -f $CORPORA ]; then + echo "$CORPORA already exists, skipping download" + else + echo "Downloading $URL" + wget $URL -O $CORPORA --no-check-certificate || rm -f $CORPORA + if [ -f $CORPORA ]; then + echo "$URL successfully downloaded." + else + echo "$URL not successfully downloaded." + rm -f $CORPORA + fi + fi +} + +if [[ -f flores ]]; then + echo "flores already cloned" +else + git clone https://github.com/facebookresearch/flores +fi + +mkdir -p $DATA +download_data $DATA/wikipedia_en_ne_si_test_sets.tgz "https://github.com/facebookresearch/flores/raw/master/data/wikipedia_en_ne_si_test_sets.tgz" +pushd $DATA +pwd +tar -vxf wikipedia_en_ne_si_test_sets.tgz +popd + + +for lang in ne_NP si_LK; do + datadir=$DATA/${lang}-en_XX-flores + rm -rf $datadir + mkdir -p $datadir + TEST_PREFIX=$DATA/wikipedia_en_ne_si_test_sets/wikipedia.test + python $SPM_ENCODE \ + --model ${SPM_MODEL} \ + --output_format=piece \ + --inputs ${TEST_PREFIX}.${lang:0:2}-en.${lang:0:2} ${TEST_PREFIX}.${lang:0:2}-en.en \ + --outputs $datadir/test.bpe.${lang}-en_XX.${lang} $datadir/test.bpe.${lang}-en_XX.en_XX + + # binarize data + fairseq-preprocess \ + --source-lang ${lang} --target-lang en_XX \ + --testpref $datadir/test.bpe.${lang}-en_XX \ + --destdir $datadir \ + --srcdict ${DICT} \ + --joined-dictionary \ + --workers 4 +done diff --git a/examples/criss/download_and_preprocess_tatoeba.sh b/examples/criss/download_and_preprocess_tatoeba.sh new file mode 100644 index 0000000000000000000000000000000000000000..7ed64f017d5e62695ba73745c840507b994abc0f --- /dev/null +++ b/examples/criss/download_and_preprocess_tatoeba.sh @@ -0,0 +1,46 @@ +#!/bin/bash +# Copyright (c) Facebook, Inc. and its affiliates. +# All rights reserved. +# +# This source code is licensed under the license found in the +# LICENSE file in the root directory of this source tree. + +SPM_ENCODE=flores/scripts/spm_encode.py +DATA=data_tmp +SPM_MODEL=criss_checkpoints/sentence.bpe.model +DICT=criss_checkpoints/dict.txt + +if [[ -f flores ]]; then + echo "flores already cloned" +else + git clone https://github.com/facebookresearch/flores +fi +if [[ -f LASER ]]; then + echo "LASER already cloned" +else + git clone https://github.com/facebookresearch/LASER +fi +mkdir -p data_tmp +declare -A lang_tatoeba_map=( ["ar_AR"]="ara" ["de_DE"]="deu" ["es_XX"]="spa" ["et_EE"]="est" ["fi_FI"]="fin" ["fr_XX"]="fra" ["hi_IN"]="hin" ["it_IT"]="ita" ["ja_XX"]="jpn" ["ko_KR"]="kor" ["kk_KZ"]="kaz" ["nl_XX"]="nld" ["ru_RU"]="rus" ["tr_TR"]="tur" ["vi_VN"]="vie" ["zh_CN"]="cmn") +for lang in ar_AR de_DE es_XX et_EE fi_FI fr_XX hi_IN it_IT ja_XX kk_KZ ko_KR nl_XX ru_RU tr_TR vi_VN zh_CN; do + lang_tatoeba=${lang_tatoeba_map[$lang]} + echo $lang_tatoeba + datadir=$DATA/${lang}-en_XX-tatoeba + rm -rf $datadir + mkdir -p $datadir + TEST_PREFIX=LASER/data/tatoeba/v1/tatoeba + python $SPM_ENCODE \ + --model ${SPM_MODEL} \ + --output_format=piece \ + --inputs ${TEST_PREFIX}.${lang_tatoeba}-eng.${lang_tatoeba} ${TEST_PREFIX}.${lang_tatoeba}-eng.eng \ + --outputs $datadir/test.bpe.${lang}-en_XX.${lang} $datadir/test.bpe.${lang}-en_XX.en_XX + + # binarize data + fairseq-preprocess \ + --source-lang ${lang} --target-lang en_XX \ + --testpref $datadir/test.bpe.${lang}-en_XX \ + --destdir $datadir \ + --srcdict ${DICT} \ + --joined-dictionary \ + --workers 4 +done diff --git a/examples/criss/mining/mine.py b/examples/criss/mining/mine.py new file mode 100644 index 0000000000000000000000000000000000000000..c872da196fe0df776622365748ad7963fee1f0a0 --- /dev/null +++ b/examples/criss/mining/mine.py @@ -0,0 +1,240 @@ +#!/usr/bin/env python3 -u +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +import argparse +import glob +from subprocess import check_call + +try: + import faiss + + has_faiss = True +except ImportError: + has_faiss = False +import numpy as np + + +GB = 1024 * 1024 * 1024 + + +def call(cmd): + print(cmd) + check_call(cmd, shell=True) + + +def get_batches(directory, lang, prefix="all_avg_pool"): + print(f"Finding in {directory}/{prefix}.{lang}*") + files = glob.glob(f"{directory}/{prefix}.{lang}*") + emb_files = [] + txt_files = [] + for emb_fi in files: + emb_files.append(emb_fi) + txt_fi = emb_fi.replace(prefix, "sentences") + txt_files.append(txt_fi) + return emb_files, txt_files + + +def load_batch(emb_file, dim): + embeddings = np.fromfile(emb_file, dtype=np.float32) + num_rows = int(embeddings.shape[0] / dim) + embeddings = embeddings.reshape((num_rows, dim)) + faiss.normalize_L2(embeddings) + return embeddings + + +def knnGPU_sharded(x_batches_f, y_batches_f, dim, k, direction="x2y"): + if not has_faiss: + raise ImportError("Please install Faiss") + sims = [] + inds = [] + xfrom = 0 + xto = 0 + for x_batch_f in x_batches_f: + yfrom = 0 + yto = 0 + x_batch = load_batch(x_batch_f, dim) + xto = xfrom + x_batch.shape[0] + bsims, binds = [], [] + for y_batch_f in y_batches_f: + y_batch = load_batch(y_batch_f, dim) + neighbor_size = min(k, y_batch.shape[0]) + yto = yfrom + y_batch.shape[0] + print("{}-{} -> {}-{}".format(xfrom, xto, yfrom, yto)) + idx = faiss.IndexFlatIP(dim) + idx = faiss.index_cpu_to_all_gpus(idx) + idx.add(y_batch) + bsim, bind = idx.search(x_batch, neighbor_size) + + bsims.append(bsim) + binds.append(bind + yfrom) + yfrom += y_batch.shape[0] + del idx + del y_batch + bsims = np.concatenate(bsims, axis=1) + binds = np.concatenate(binds, axis=1) + aux = np.argsort(-bsims, axis=1) + sim_batch = np.zeros((x_batch.shape[0], k), dtype=np.float32) + ind_batch = np.zeros((x_batch.shape[0], k), dtype=np.int64) + for i in range(x_batch.shape[0]): + for j in range(k): + sim_batch[i, j] = bsims[i, aux[i, j]] + ind_batch[i, j] = binds[i, aux[i, j]] + sims.append(sim_batch) + inds.append(ind_batch) + xfrom += x_batch.shape[0] + del x_batch + sim = np.concatenate(sims, axis=0) + ind = np.concatenate(inds, axis=0) + return sim, ind + + +def score(sim, fwd_mean, bwd_mean, margin): + return margin(sim, (fwd_mean + bwd_mean) / 2) + + +def score_candidates( + sim_mat, candidate_inds, fwd_mean, bwd_mean, margin, verbose=False +): + print(" - scoring {:d} candidates".format(sim_mat.shape[0])) + scores = np.zeros(candidate_inds.shape) + for i in range(scores.shape[0]): + for j in range(scores.shape[1]): + k = int(candidate_inds[i, j]) + scores[i, j] = score(sim_mat[i, j], fwd_mean[i], bwd_mean[k], margin) + return scores + + +def load_text(files): + all_sentences = [] + for fi in files: + with open(fi) as sentence_fi: + for line in sentence_fi: + all_sentences.append(line.strip()) + print(f"Read {len(all_sentences)} sentences") + return all_sentences + + +if __name__ == "__main__": + parser = argparse.ArgumentParser(description="Mine bitext") + parser.add_argument("--src-lang", help="Source language") + parser.add_argument("--tgt-lang", help="Target language") + parser.add_argument( + "--dict-path", help="Path to dictionary file", default="dict.txt" + ) + parser.add_argument( + "--spm-path", help="Path to SPM model file", default="sentence.bpe.model" + ) + parser.add_argument("--dim", type=int, default=1024, help="Embedding dimension") + parser.add_argument("--mem", type=int, default=5, help="Memory in GB") + parser.add_argument("--src-dir", help="Source directory") + parser.add_argument("--tgt-dir", help="Target directory") + parser.add_argument("--output", help="Output path") + parser.add_argument( + "--neighborhood", type=int, default=4, help="Embedding dimension" + ) + parser.add_argument( + "--threshold", type=float, default=1.06, help="Threshold on mined bitext" + ) + parser.add_argument( + "--valid-size", + type=int, + default=2000, + help="Number of sentences used for validation set", + ) + parser.add_argument( + "--min-count", + type=int, + default=50000, + help="Min num sentences used for each language", + ) + args = parser.parse_args() + + x_batches_f, x_sents_f = get_batches(args.src_dir, args.src_lang) + y_batches_f, y_sents_f = get_batches(args.tgt_dir, args.tgt_lang) + margin = lambda a, b: a / b + y2x_sim, y2x_ind = knnGPU_sharded( + y_batches_f, x_batches_f, args.dim, args.neighborhood, direction="y2x" + ) + x2y_sim, x2y_ind = knnGPU_sharded( + x_batches_f, y_batches_f, args.dim, args.neighborhood, direction="x2y" + ) + + x2y_mean = x2y_sim.mean(axis=1) + y2x_mean = y2x_sim.mean(axis=1) + fwd_scores = score_candidates(x2y_sim, x2y_ind, x2y_mean, y2x_mean, margin) + bwd_scores = score_candidates(y2x_sim, y2x_ind, y2x_mean, x2y_mean, margin) + fwd_best = x2y_ind[np.arange(x2y_sim.shape[0]), fwd_scores.argmax(axis=1)] + bwd_best = y2x_ind[np.arange(y2x_sim.shape[0]), bwd_scores.argmax(axis=1)] + indices = np.stack( + ( + np.concatenate((np.arange(x2y_ind.shape[0]), bwd_best)), + np.concatenate((fwd_best, np.arange(y2x_ind.shape[0]))), + ), + axis=1, + ) + scores = np.concatenate((fwd_scores.max(axis=1), bwd_scores.max(axis=1))) + + x_sentences = load_text(x_sents_f) + y_sentences = load_text(y_sents_f) + + threshold = args.threshold + min_count = args.min_count + seen_src, seen_trg = set(), set() + directory = args.output + call(f"mkdir -p {directory}") + src_out = open( + f"{directory}/all.{args.src_lang}", + mode="w", + encoding="utf-8", + errors="surrogateescape", + ) + tgt_out = open( + f"{directory}/all.{args.tgt_lang}", + mode="w", + encoding="utf-8", + errors="surrogateescape", + ) + scores_out = open( + f"{directory}/all.scores", mode="w", encoding="utf-8", errors="surrogateescape" + ) + count = 0 + for i in np.argsort(-scores): + src_ind, trg_ind = indices[i] + if src_ind not in seen_src and trg_ind not in seen_trg: + seen_src.add(src_ind) + seen_trg.add(trg_ind) + if scores[i] > threshold or count < min_count: + if x_sentences[src_ind]: + print(scores[i], file=scores_out) + print(x_sentences[src_ind], file=src_out) + print(y_sentences[trg_ind], file=tgt_out) + count += 1 + else: + print(f"Ignoring sentence: {x_sentences[src_ind]}") + src_out.close() + tgt_out.close() + scores_out.close() + + print(f"Found {count} pairs for threshold={threshold}") + with open(f"{directory}/all.{args.src_lang}") as all_s, open( + f"{directory}/all.{args.tgt_lang}" + ) as all_t, open(f"{directory}/valid.{args.src_lang}", "w") as valid_s, open( + f"{directory}/valid.{args.tgt_lang}", "w" + ) as valid_t, open( + f"{directory}/train.{args.src_lang}", "w" + ) as train_s, open( + f"{directory}/train.{args.tgt_lang}", "w" + ) as train_t: + count = 0 + for s_line, t_line in zip(all_s, all_t): + s_line = s_line.split("\t")[1] + t_line = t_line.split("\t")[1] + if count >= args.valid_size: + train_s.write(s_line) + train_t.write(t_line) + else: + valid_s.write(s_line) + valid_t.write(t_line) + count += 1 diff --git a/examples/criss/mining/mine_example.sh b/examples/criss/mining/mine_example.sh new file mode 100644 index 0000000000000000000000000000000000000000..ace995ac44665f99d904b6a89d7fbbce24103afe --- /dev/null +++ b/examples/criss/mining/mine_example.sh @@ -0,0 +1,103 @@ +#!/bin/bash +# Copyright (c) Facebook, Inc. and its affiliates. +# All rights reserved. +# +# This source code is licensed under the license found in the +# LICENSE file in the root directory of this source tree. +# +source_lang=kk_KZ +target_lang=en_XX +MODEL=criss_checkpoints/criss.3rd.pt +SPM=criss_checkpoints/sentence.bpe.model +SPLIT=test +LANG_DICT=criss_checkpoints/lang_dict.txt +SPM_ENCODE=flores/scripts/spm_encode.py +SAVE_ENCODER=save_encoder.py +ENCODER_SAVE_ROOT=sentence_embeddings/$MODEL +DICT=criss_checkpoints/dict.txt +THRESHOLD=1.02 +MIN_COUNT=500 + +DATA_DIR=data_tmp +SAVE_DIR=mining/${source_lang}_${target_lang}_mined +ENCODER_SAVE_DIR=${ENCODER_SAVE_ROOT}/${source_lang}-${target_lang} +INPUT_DIR=$DATA_DIR/${source_lang}-${target_lang}-tatoeba + +mkdir -p $ENCODER_SAVE_DIR/${target_lang} +mkdir -p $ENCODER_SAVE_DIR/${source_lang} +mkdir -p $SAVE_DIR + +## Save encoder outputs + +# Save encoder outputs for source sentences +python $SAVE_ENCODER \ + ${INPUT_DIR} \ + --path ${MODEL} \ + --task translation_multi_simple_epoch \ + --lang-pairs ${source_lang}-${target_lang} \ + --lang-dict ${LANG_DICT} \ + --gen-subset ${SPLIT} \ + --bpe 'sentencepiece' \ + -s ${source_lang} -t ${target_lang} \ + --sentencepiece-model ${SPM} \ + --remove-bpe 'sentencepiece' \ + --beam 1 \ + --lang-tok-style mbart \ + --encoder-save-dir ${ENCODER_SAVE_DIR}/${source_lang} + +## Save encoder outputs for target sentences +python $SAVE_ENCODER \ + ${INPUT_DIR} \ + --path ${MODEL} \ + --lang-pairs ${source_lang}-${target_lang} \ + --lang-dict ${LANG_DICT} \ + --task translation_multi_simple_epoch \ + --gen-subset ${SPLIT} \ + --bpe 'sentencepiece' \ + -t ${source_lang} -s ${target_lang} \ + --sentencepiece-model ${SPM} \ + --remove-bpe 'sentencepiece' \ + --beam 1 \ + --lang-tok-style mbart \ + --encoder-save-dir ${ENCODER_SAVE_DIR}/${target_lang} + +## Mining +python mining/mine.py \ + --src-lang ${source_lang} \ + --tgt-lang ${target_lang} \ + --dim 1024 \ + --mem 10 \ + --neighborhood 4 \ + --src-dir ${ENCODER_SAVE_DIR}/${source_lang} \ + --tgt-dir ${ENCODER_SAVE_DIR}/${target_lang} \ + --output $SAVE_DIR \ + --threshold ${THRESHOLD} \ + --min-count ${MIN_COUNT} \ + --valid-size 100 \ + --dict-path ${DICT} \ + --spm-path ${SPM} \ + + +## Process and binarize mined data +python $SPM_ENCODE \ + --model ${SPM} \ + --output_format=piece \ + --inputs mining/${source_lang}_${target_lang}_mined/train.${source_lang} mining/${source_lang}_${target_lang}_mined/train.${target_lang} \ + --outputs mining/${source_lang}_${target_lang}_mined/train.bpe.${source_lang} mining/${source_lang}_${target_lang}_mined/train.bpe.${target_lang} + +python $SPM_ENCODE \ + --model ${SPM} \ + --output_format=piece \ + --inputs mining/${source_lang}_${target_lang}_mined/valid.${source_lang} mining/${source_lang}_${target_lang}_mined/valid.${target_lang} \ + --outputs mining/${source_lang}_${target_lang}_mined/valid.bpe.${source_lang} mining/${source_lang}_${target_lang}_mined/valid.bpe.${target_lang} + + +fairseq-preprocess \ + --source-lang ${source_lang} \ + --target-lang ${target_lang} \ + --trainpref mining/${source_lang}_${target_lang}_mined/train.bpe \ + --validpref mining/${source_lang}_${target_lang}_mined/valid.bpe \ + --destdir mining/${source_lang}_${target_lang}_mined \ + --srcdict ${DICT} \ + --joined-dictionary \ + --workers 8 diff --git a/examples/criss/save_encoder.py b/examples/criss/save_encoder.py new file mode 100644 index 0000000000000000000000000000000000000000..d911d066e359f5ce64aa4292d812d6e52fd3cc9b --- /dev/null +++ b/examples/criss/save_encoder.py @@ -0,0 +1,213 @@ +#!/usr/bin/env python3 -u +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +Translate pre-processed data with a trained model. +""" + +import numpy as np +import torch +from fairseq import checkpoint_utils, options, progress_bar, tasks, utils +from fairseq.sequence_generator import EnsembleModel + + +def get_avg_pool( + models, sample, prefix_tokens, src_dict, remove_bpe, has_langtok=False +): + model = EnsembleModel(models) + + # model.forward normally channels prev_output_tokens into the decoder + # separately, but SequenceGenerator directly calls model.encoder + encoder_input = { + k: v for k, v in sample["net_input"].items() if k != "prev_output_tokens" + } + + # compute the encoder output for each beam + encoder_outs = model.forward_encoder(encoder_input) + np_encoder_outs = encoder_outs[0].encoder_out.cpu().numpy().astype(np.float32) + encoder_mask = 1 - encoder_outs[0].encoder_padding_mask.cpu().numpy().astype( + np.float32 + ) + encoder_mask = np.expand_dims(encoder_mask.T, axis=2) + if has_langtok: + encoder_mask = encoder_mask[1:, :, :] + np_encoder_outs = np_encoder_outs[1, :, :] + masked_encoder_outs = encoder_mask * np_encoder_outs + avg_pool = (masked_encoder_outs / encoder_mask.sum(axis=0)).sum(axis=0) + return avg_pool + + +def main(args): + assert args.path is not None, "--path required for generation!" + assert ( + not args.sampling or args.nbest == args.beam + ), "--sampling requires --nbest to be equal to --beam" + assert ( + args.replace_unk is None or args.raw_text + ), "--replace-unk requires a raw text dataset (--raw-text)" + + args.beam = 1 + utils.import_user_module(args) + + if args.max_tokens is None: + args.max_tokens = 12000 + print(args) + use_cuda = torch.cuda.is_available() and not args.cpu + + # Load dataset splits + task = tasks.setup_task(args) + task.load_dataset(args.gen_subset) + + # Set dictionaries + try: + src_dict = getattr(task, "source_dictionary", None) + except NotImplementedError: + src_dict = None + tgt_dict = task.target_dictionary + + # Load ensemble + print("| loading model(s) from {}".format(args.path)) + models, _model_args = checkpoint_utils.load_model_ensemble( + args.path.split(":"), + arg_overrides=eval(args.model_overrides), + task=task, + ) + + # Optimize ensemble for generation + for model in models: + model.make_generation_fast_( + beamable_mm_beam_size=None if args.no_beamable_mm else args.beam, + need_attn=args.print_alignment, + ) + if args.fp16: + model.half() + if use_cuda: + model.cuda() + + # Load alignment dictionary for unknown word replacement + # (None if no unknown word replacement, empty if no path to align dictionary) + align_dict = utils.load_align_dict(args.replace_unk) + + # Load dataset (possibly sharded) + itr = task.get_batch_iterator( + dataset=task.dataset(args.gen_subset), + max_tokens=args.max_tokens, + max_positions=utils.resolve_max_positions( + task.max_positions(), + ), + ignore_invalid_inputs=args.skip_invalid_size_inputs_valid_test, + required_batch_size_multiple=args.required_batch_size_multiple, + num_shards=args.num_shards, + shard_id=args.shard_id, + num_workers=args.num_workers, + ).next_epoch_itr(shuffle=False) + + num_sentences = 0 + source_sentences = [] + shard_id = 0 + all_avg_pool = None + encoder_has_langtok = ( + hasattr(task.args, "encoder_langtok") + and task.args.encoder_langtok is not None + and hasattr(task.args, "lang_tok_replacing_bos_eos") + and not task.args.lang_tok_replacing_bos_eos + ) + with progress_bar.build_progress_bar(args, itr) as t: + for sample in t: + if sample is None: + print("Skipping None") + continue + sample = utils.move_to_cuda(sample) if use_cuda else sample + if "net_input" not in sample: + continue + + prefix_tokens = None + if args.prefix_size > 0: + prefix_tokens = sample["target"][:, : args.prefix_size] + + with torch.no_grad(): + avg_pool = get_avg_pool( + models, + sample, + prefix_tokens, + src_dict, + args.post_process, + has_langtok=encoder_has_langtok, + ) + if all_avg_pool is not None: + all_avg_pool = np.concatenate((all_avg_pool, avg_pool)) + else: + all_avg_pool = avg_pool + + if not isinstance(sample["id"], list): + sample_ids = sample["id"].tolist() + else: + sample_ids = sample["id"] + for i, sample_id in enumerate(sample_ids): + # Remove padding + src_tokens = utils.strip_pad( + sample["net_input"]["src_tokens"][i, :], tgt_dict.pad() + ) + + # Either retrieve the original sentences or regenerate them from tokens. + if align_dict is not None: + src_str = task.dataset(args.gen_subset).src.get_original_text( + sample_id + ) + else: + if src_dict is not None: + src_str = src_dict.string(src_tokens, args.post_process) + else: + src_str = "" + + if not args.quiet: + if src_dict is not None: + print("S-{}\t{}".format(sample_id, src_str)) + + source_sentences.append(f"{sample_id}\t{src_str}") + + num_sentences += sample["nsentences"] + if all_avg_pool.shape[0] >= 1000000: + with open( + f"{args.encoder_save_dir}/all_avg_pool.{args.source_lang}.{shard_id}", + "w", + ) as avg_pool_file: + all_avg_pool.tofile(avg_pool_file) + with open( + f"{args.encoder_save_dir}/sentences.{args.source_lang}.{shard_id}", + "w", + ) as sentence_file: + sentence_file.writelines(f"{line}\n" for line in source_sentences) + all_avg_pool = None + source_sentences = [] + shard_id += 1 + + if all_avg_pool is not None: + with open( + f"{args.encoder_save_dir}/all_avg_pool.{args.source_lang}.{shard_id}", "w" + ) as avg_pool_file: + all_avg_pool.tofile(avg_pool_file) + with open( + f"{args.encoder_save_dir}/sentences.{args.source_lang}.{shard_id}", "w" + ) as sentence_file: + sentence_file.writelines(f"{line}\n" for line in source_sentences) + return None + + +def cli_main(): + parser = options.get_generation_parser() + parser.add_argument( + "--encoder-save-dir", + default="", + type=str, + metavar="N", + help="directory to save encoder outputs", + ) + args = options.parse_args_and_arch(parser) + main(args) + + +if __name__ == "__main__": + cli_main() diff --git a/examples/criss/sentence_retrieval/encoder_analysis.py b/examples/criss/sentence_retrieval/encoder_analysis.py new file mode 100644 index 0000000000000000000000000000000000000000..b41bfbe38789ba14e6a5ea938c75d761424c00ab --- /dev/null +++ b/examples/criss/sentence_retrieval/encoder_analysis.py @@ -0,0 +1,92 @@ +#!/usr/bin/env python3 -u +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +import argparse +import glob + +import numpy as np + + +DIM = 1024 + + +def compute_dist(source_embs, target_embs, k=5, return_sim_mat=False): + target_ids = [tid for tid in target_embs] + source_mat = np.stack(source_embs.values(), axis=0) + normalized_source_mat = source_mat / np.linalg.norm( + source_mat, axis=1, keepdims=True + ) + target_mat = np.stack(target_embs.values(), axis=0) + normalized_target_mat = target_mat / np.linalg.norm( + target_mat, axis=1, keepdims=True + ) + sim_mat = normalized_source_mat.dot(normalized_target_mat.T) + if return_sim_mat: + return sim_mat + neighbors_map = {} + for i, sentence_id in enumerate(source_embs): + idx = np.argsort(sim_mat[i, :])[::-1][:k] + neighbors_map[sentence_id] = [target_ids[tid] for tid in idx] + return neighbors_map + + +def load_embeddings(directory, LANGS): + sentence_embeddings = {} + sentence_texts = {} + for lang in LANGS: + sentence_embeddings[lang] = {} + sentence_texts[lang] = {} + lang_dir = f"{directory}/{lang}" + embedding_files = glob.glob(f"{lang_dir}/all_avg_pool.{lang}.*") + for embed_file in embedding_files: + shard_id = embed_file.split(".")[-1] + embeddings = np.fromfile(embed_file, dtype=np.float32) + num_rows = embeddings.shape[0] // DIM + embeddings = embeddings.reshape((num_rows, DIM)) + + with open(f"{lang_dir}/sentences.{lang}.{shard_id}") as sentence_file: + for idx, line in enumerate(sentence_file): + sentence_id, sentence = line.strip().split("\t") + sentence_texts[lang][sentence_id] = sentence + sentence_embeddings[lang][sentence_id] = embeddings[idx, :] + + return sentence_embeddings, sentence_texts + + +def compute_accuracy(directory, LANGS): + sentence_embeddings, sentence_texts = load_embeddings(directory, LANGS) + + top_1_accuracy = {} + + top1_str = " ".join(LANGS) + "\n" + for source_lang in LANGS: + top_1_accuracy[source_lang] = {} + top1_str += f"{source_lang} " + for target_lang in LANGS: + top1 = 0 + top5 = 0 + neighbors_map = compute_dist( + sentence_embeddings[source_lang], sentence_embeddings[target_lang] + ) + for sentence_id, neighbors in neighbors_map.items(): + if sentence_id == neighbors[0]: + top1 += 1 + if sentence_id in neighbors[:5]: + top5 += 1 + n = len(sentence_embeddings[target_lang]) + top1_str += f"{top1/n} " + top1_str += "\n" + + print(top1_str) + print(top1_str, file=open(f"{directory}/accuracy", "w")) + + +if __name__ == "__main__": + parser = argparse.ArgumentParser(description="Analyze encoder outputs") + parser.add_argument("directory", help="Source language corpus") + parser.add_argument("--langs", help="List of langs") + args = parser.parse_args() + langs = args.langs.split(",") + compute_accuracy(args.directory, langs) diff --git a/examples/criss/sentence_retrieval/sentence_retrieval_tatoeba.sh b/examples/criss/sentence_retrieval/sentence_retrieval_tatoeba.sh new file mode 100644 index 0000000000000000000000000000000000000000..0428d8bef9d426ac3e664cd281ce0b688f5f580f --- /dev/null +++ b/examples/criss/sentence_retrieval/sentence_retrieval_tatoeba.sh @@ -0,0 +1,59 @@ +#!/bin/bash +# Copyright (c) Facebook, Inc. and its affiliates. +# All rights reserved. +# +# This source code is licensed under the license found in the +# LICENSE file in the root directory of this source tree. +# +source_lang=kk_KZ +target_lang=en_XX +MODEL=criss_checkpoints/criss.3rd.pt +SPM=criss_checkpoints/sentence.bpe.model +SPLIT=test +LANG_DICT=criss_checkpoints/lang_dict.txt +ENCODER_ANALYSIS=sentence_retrieval/encoder_analysis.py +SAVE_ENCODER=save_encoder.py +ENCODER_SAVE_ROOT=sentence_embeddings/$MODEL + + + +DATA_DIR=data_tmp +INPUT_DIR=$DATA_DIR/${source_lang}-${target_lang}-tatoeba +ENCODER_SAVE_DIR=${ENCODER_SAVE_ROOT}/${source_lang}-${target_lang} +mkdir -p $ENCODER_SAVE_DIR/${target_lang} +mkdir -p $ENCODER_SAVE_DIR/${source_lang} + +# Save encoder outputs for source sentences +python $SAVE_ENCODER \ + ${INPUT_DIR} \ + --path ${MODEL} \ + --task translation_multi_simple_epoch \ + --lang-dict ${LANG_DICT} \ + --gen-subset ${SPLIT} \ + --bpe 'sentencepiece' \ + --lang-pairs ${source_lang}-${target_lang} \ + -s ${source_lang} -t ${target_lang} \ + --sentencepiece-model ${SPM} \ + --remove-bpe 'sentencepiece' \ + --beam 1 \ + --lang-tok-style mbart \ + --encoder-save-dir ${ENCODER_SAVE_DIR}/${source_lang} + +# Save encoder outputs for target sentences +python $SAVE_ENCODER \ + ${INPUT_DIR} \ + --path ${MODEL} \ + --lang-dict ${LANG_DICT} \ + --task translation_multi_simple_epoch \ + --gen-subset ${SPLIT} \ + --bpe 'sentencepiece' \ + --lang-pairs ${target_lang}-${source_lang} \ + -t ${source_lang} -s ${target_lang} \ + --sentencepiece-model ${SPM} \ + --remove-bpe 'sentencepiece' \ + --beam 1 \ + --lang-tok-style mbart \ + --encoder-save-dir ${ENCODER_SAVE_DIR}/${target_lang} + +# Analyze sentence retrieval accuracy +python $ENCODER_ANALYSIS --langs "${source_lang},${target_lang}" ${ENCODER_SAVE_DIR} diff --git a/examples/criss/unsupervised_mt/eval.sh b/examples/criss/unsupervised_mt/eval.sh new file mode 100644 index 0000000000000000000000000000000000000000..03b773ed5a522eb82186fea8ffbb6c557e14b6d3 --- /dev/null +++ b/examples/criss/unsupervised_mt/eval.sh @@ -0,0 +1,37 @@ +#!/bin/bash +# Copyright (c) Facebook, Inc. and its affiliates. +# All rights reserved. +# +# This source code is licensed under the license found in the +# LICENSE file in the root directory of this source tree. +# +SRC=si_LK +TGT=en_XX +MODEL=criss_checkpoints/criss.3rd.pt + +MULTIBLEU=mosesdecoder/scripts/generic/multi-bleu.perl +MOSES=mosesdecoder +REPLACE_UNICODE_PUNCT=$MOSES/scripts/tokenizer/replace-unicode-punctuation.perl +NORM_PUNC=$MOSES/scripts/tokenizer/normalize-punctuation.perl +REM_NON_PRINT_CHAR=$MOSES/scripts/tokenizer/remove-non-printing-char.perl +TOKENIZER=$MOSES/scripts/tokenizer/tokenizer.perl +GEN_TMP_DIR=gen_tmp +LANG_DICT=criss_checkpoints/lang_dict.txt + +if [ ! -d "mosesdecoder" ]; then + git clone https://github.com/moses-smt/mosesdecoder +fi +mkdir -p $GEN_TMP_DIR +fairseq-generate data_tmp/${SRC}-${TGT}-flores \ + --task translation_multi_simple_epoch \ + --max-tokens 2000 \ + --path ${MODEL} \ + --skip-invalid-size-inputs-valid-test \ + --beam 5 --lenpen 1.0 --gen-subset test \ + --remove-bpe=sentencepiece \ + --source-lang ${SRC} --target-lang ${TGT} \ + --decoder-langtok --lang-pairs 'en_XX-ar_AR,en_XX-de_DE,en_XX-es_XX,en_XX-fr_XX,en_XX-hi_IN,en_XX-it_IT,en_XX-ja_XX,en_XX-ko_KR,en_XX-nl_XX,en_XX-ru_RU,en_XX-zh_CN,en_XX-tr_TR,en_XX-vi_VN,en_XX-ro_RO,en_XX-my_MM,en_XX-ne_NP,en_XX-si_LK,en_XX-cs_CZ,en_XX-lt_LT,en_XX-kk_KZ,en_XX-gu_IN,en_XX-fi_FI,en_XX-et_EE,en_XX-lv_LV,ar_AR-en_XX,cs_CZ-en_XX,de_DE-en_XX,es_XX-en_XX,et_EE-en_XX,fi_FI-en_XX,fr_XX-en_XX,gu_IN-en_XX,hi_IN-en_XX,it_IT-en_XX,ja_XX-en_XX,kk_KZ-en_XX,ko_KR-en_XX,lt_LT-en_XX,lv_LV-en_XX,my_MM-en_XX,ne_NP-en_XX,nl_XX-en_XX,ro_RO-en_XX,ru_RU-en_XX,si_LK-en_XX,tr_TR-en_XX,vi_VN-en_XX,zh_CN-en_XX,ar_AR-es_XX,es_XX-ar_AR,ar_AR-hi_IN,hi_IN-ar_AR,ar_AR-zh_CN,zh_CN-ar_AR,cs_CZ-es_XX,es_XX-cs_CZ,cs_CZ-hi_IN,hi_IN-cs_CZ,cs_CZ-zh_CN,zh_CN-cs_CZ,de_DE-es_XX,es_XX-de_DE,de_DE-hi_IN,hi_IN-de_DE,de_DE-zh_CN,zh_CN-de_DE,es_XX-hi_IN,hi_IN-es_XX,es_XX-zh_CN,zh_CN-es_XX,et_EE-es_XX,es_XX-et_EE,et_EE-hi_IN,hi_IN-et_EE,et_EE-zh_CN,zh_CN-et_EE,fi_FI-es_XX,es_XX-fi_FI,fi_FI-hi_IN,hi_IN-fi_FI,fi_FI-zh_CN,zh_CN-fi_FI,fr_XX-es_XX,es_XX-fr_XX,fr_XX-hi_IN,hi_IN-fr_XX,fr_XX-zh_CN,zh_CN-fr_XX,gu_IN-es_XX,es_XX-gu_IN,gu_IN-hi_IN,hi_IN-gu_IN,gu_IN-zh_CN,zh_CN-gu_IN,hi_IN-zh_CN,zh_CN-hi_IN,it_IT-es_XX,es_XX-it_IT,it_IT-hi_IN,hi_IN-it_IT,it_IT-zh_CN,zh_CN-it_IT,ja_XX-es_XX,es_XX-ja_XX,ja_XX-hi_IN,hi_IN-ja_XX,ja_XX-zh_CN,zh_CN-ja_XX,kk_KZ-es_XX,es_XX-kk_KZ,kk_KZ-hi_IN,hi_IN-kk_KZ,kk_KZ-zh_CN,zh_CN-kk_KZ,ko_KR-es_XX,es_XX-ko_KR,ko_KR-hi_IN,hi_IN-ko_KR,ko_KR-zh_CN,zh_CN-ko_KR,lt_LT-es_XX,es_XX-lt_LT,lt_LT-hi_IN,hi_IN-lt_LT,lt_LT-zh_CN,zh_CN-lt_LT,lv_LV-es_XX,es_XX-lv_LV,lv_LV-hi_IN,hi_IN-lv_LV,lv_LV-zh_CN,zh_CN-lv_LV,my_MM-es_XX,es_XX-my_MM,my_MM-hi_IN,hi_IN-my_MM,my_MM-zh_CN,zh_CN-my_MM,ne_NP-es_XX,es_XX-ne_NP,ne_NP-hi_IN,hi_IN-ne_NP,ne_NP-zh_CN,zh_CN-ne_NP,nl_XX-es_XX,es_XX-nl_XX,nl_XX-hi_IN,hi_IN-nl_XX,nl_XX-zh_CN,zh_CN-nl_XX,ro_RO-es_XX,es_XX-ro_RO,ro_RO-hi_IN,hi_IN-ro_RO,ro_RO-zh_CN,zh_CN-ro_RO,ru_RU-es_XX,es_XX-ru_RU,ru_RU-hi_IN,hi_IN-ru_RU,ru_RU-zh_CN,zh_CN-ru_RU,si_LK-es_XX,es_XX-si_LK,si_LK-hi_IN,hi_IN-si_LK,si_LK-zh_CN,zh_CN-si_LK,tr_TR-es_XX,es_XX-tr_TR,tr_TR-hi_IN,hi_IN-tr_TR,tr_TR-zh_CN,zh_CN-tr_TR,vi_VN-es_XX,es_XX-vi_VN,vi_VN-hi_IN,hi_IN-vi_VN,vi_VN-zh_CN,zh_CN-vi_VN' \ + --lang-dict ${LANG_DICT} --lang-tok-style 'mbart' --sampling-method 'temperature' --sampling-temperature '1.0' > $GEN_TMP_DIR/${SRC}_${TGT}.gen +cat $GEN_TMP_DIR/${SRC}_${TGT}.gen | grep -P "^T-" | cut -f2 | $REPLACE_UNICODE_PUNCT | $NORM_PUNC -l ${TGT:0:2} | $REM_NON_PRINT_CHAR | $TOKENIZER -no-escape ${TGT:0:2} > $GEN_TMP_DIR/${SRC}_${TGT}.hyp +cat $GEN_TMP_DIR/${SRC}_${TGT}.gen | grep -P "^H-" | cut -f3 | $REPLACE_UNICODE_PUNCT | $NORM_PUNC -l ${TGT:0:2} | $REM_NON_PRINT_CHAR | $TOKENIZER -no-escape ${TGT:0:2} > $GEN_TMP_DIR/${SRC}_${TGT}.ref +${MULTIBLEU} $GEN_TMP_DIR/${SRC}_${TGT}.ref < $GEN_TMP_DIR/${SRC}_${TGT}.hyp diff --git a/examples/cross_lingual_language_model/README.md b/examples/cross_lingual_language_model/README.md new file mode 100644 index 0000000000000000000000000000000000000000..af9128e39e5925e9411d162c2f24a19e4532d618 --- /dev/null +++ b/examples/cross_lingual_language_model/README.md @@ -0,0 +1,77 @@ +# Cross-Lingual Language Model Pre-training + +Below are some details for training Cross-Lingual Language Models (XLM) - similar to the ones presented in [Lample & Conneau, 2019](https://arxiv.org/pdf/1901.07291.pdf) - in Fairseq. The current implementation only supports the Masked Language Model (MLM) from the paper above. + +## Downloading and Tokenizing Monolingual Data + +Pointers to the monolingual data from wikipedia, used for training the XLM-style MLM model as well as details on processing (tokenization and BPE) it can be found in the [XLM Github Repository](https://github.com/facebookresearch/XLM#download--preprocess-monolingual-data). + +Let's assume the following for the code snippets in later sections to work +- Processed data is in the folder: monolingual_data/processed +- Each language has 3 files for train, test and validation. For example we have the following files for English: + train.en, valid.en +- We are training a model for 5 languages: Arabic (ar), German (de), English (en), Hindi (hi) and French (fr) +- The vocabulary file is monolingual_data/processed/vocab_mlm + + +## Fairseq Pre-processing and Binarization + +Pre-process and binarize the data with the MaskedLMDictionary and cross_lingual_lm task + +```bash +# Ensure the output directory exists +DATA_DIR=monolingual_data/fairseq_processed +mkdir -p "$DATA_DIR" + +for lg in ar de en hi fr +do + + fairseq-preprocess \ + --task cross_lingual_lm \ + --srcdict monolingual_data/processed/vocab_mlm \ + --only-source \ + --trainpref monolingual_data/processed/train \ + --validpref monolingual_data/processed/valid \ + --testpref monolingual_data/processed/test \ + --destdir monolingual_data/fairseq_processed \ + --workers 20 \ + --source-lang $lg + + # Since we only have a source language, the output file has a None for the + # target language. Remove this + + for stage in train test valid + + sudo mv "$DATA_DIR/$stage.$lg-None.$lg.bin" "$stage.$lg.bin" + sudo mv "$DATA_DIR/$stage.$lg-None.$lg.idx" "$stage.$lg.idx" + + done + +done +``` + +## Train a Cross-lingual Language Model similar to the XLM MLM model + +Use the following command to train the model on 5 languages. + +``` +fairseq-train \ +--task cross_lingual_lm monolingual_data/fairseq_processed \ +--save-dir checkpoints/mlm \ +--max-update 2400000 --save-interval 1 --no-epoch-checkpoints \ +--arch xlm_base \ +--optimizer adam --lr-scheduler reduce_lr_on_plateau \ +--lr-shrink 0.5 --lr 0.0001 --stop-min-lr 1e-09 \ +--dropout 0.1 \ +--criterion legacy_masked_lm_loss \ +--max-tokens 2048 --tokens-per-sample 256 --attention-dropout 0.1 \ +--dataset-impl lazy --seed 0 \ +--masked-lm-only \ +--monolingual-langs 'ar,de,en,hi,fr' --num-segment 5 \ +--ddp-backend=legacy_ddp +``` + +Some Notes: +- Using tokens_per_sample greater than 256 can cause OOM (out-of-memory) issues. Usually since MLM packs in streams of text, this parameter doesn't need much tuning. +- The Evaluation workflow for computing MLM Perplexity on test data is in progress. +- Finetuning this model on a downstream task is something which is not currently available. diff --git a/examples/fast_noisy_channel/README.md b/examples/fast_noisy_channel/README.md new file mode 100644 index 0000000000000000000000000000000000000000..a04151a796e4e092fa3c803a1679ab521af96aeb --- /dev/null +++ b/examples/fast_noisy_channel/README.md @@ -0,0 +1,345 @@ +# Language Models not just for Pre-training: Fast Online Neural Noisy Channel Modeling + +## Introduction +- [Yee et al. (2019)](https://www.aclweb.org/anthology/D19-1571.pdf) introduce a simple and effective noisy channel modeling approach for neural machine translation. However, the noisy channel online decoding approach introduced in this paper is too slow to be practical. +- To address this, [Bhosale et al. (2020)](http://www.statmt.org/wmt20/pdf/2020.wmt-1.68.pdf) introduces 3 simple approximations to make this approach very fast and practical without much loss in accuracy. +- This README provides intructions on how to run online decoding or generation with the noisy channel modeling approach, including ways to make it very fast without much loss in accuracy. + +## Noisy Channel Modeling + +[Yee et al. (2019)](https://www.aclweb.org/anthology/D19-1571.pdf) applies the Bayes Rule to predict `P(y|x)`, the probability of the target `y` given the source `x`. +```P(y|x) = P(x|y) * P(y) / P(x)``` +- `P(x|y)` predicts the source `x` given the target `y` and is referred to as the **channel model** +- `P(y)` is a **language model** over the target `y` +- `P(x)` is generally not modeled since it is constant for all `y`. + +We use Transformer models to parameterize the direct model `P(y|x)`, the channel model `P(x|y)` and the language model `P(y)`. + +During online decoding with beam search, we generate the top `K2` candidates per beam and score them with the following linear combination of the channel model, the language model as well as the direct model scores. + +```(1 / t) * log(P(y|x) + (1 / s) * ( λ1 * log(P(x|y)) + λ2 * log(P(y) ) )``` +- `t` - Target Prefix Length +- `s` - Source Length +- `λ1` - Channel Model Weight +- `λ2` - Language Model Weight + +The top `beam_size` candidates based on the above combined scores are chosen to continue the beams in beam search. In beam search with a direct model alone, the scores from the direct model `P(y|x)` are used to choose the top candidates in beam search. + +This framework provides a great way to utlize strong target language models trained on large amounts of unlabeled data. Language models can prefer targets unrelated to the source, so we also need a channel model whose role is to ensure that the target preferred by the language model also translates back to the source. + +### Training Translation Models and Language Models + +For training Transformer models in fairseq for machine translation, refer to instructions [here](https://github.com/pytorch/fairseq/tree/master/examples/translation) + +For training Transformer models in fairseq for language modeling, refer to instructions [here](https://github.com/pytorch/fairseq/tree/master/examples/language_model) + +### Generation with Language Model for German-English translation with fairseq + +Here are instructions to generate using a direct model and a target-side language model. + +Note: +- Download and install fairseq as per instructions [here](https://github.com/pytorch/fairseq) +- Preprocess and binarize the dataset as per instructions in section [Test Data Preprocessing](#test-data-preprocessing) + +```sh +binarized_data=data_dir/binarized +direct_model=de_en_seed4.pt +lm_model=en_lm.pt +lm_data=lm_data +wget https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/direct_models/seed4.pt -O ${direct_model} +wget https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/lm_model/transformer_lm.pt -O ${lm_model} +mkdir -p ${lm_data} +wget https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/lm_model/lm_dict/dict.txt -O ${lm_data}/dict.txt + +k2=10 +lenpen=0.16 +lm_wt=0.14 +fairseq-generate ${binarized_data} \ + --user-dir examples/fast_noisy_channel \ + --beam 5 \ + --path ${direct_model} \ + --lm-model ${lm_model} \ + --lm-data ${lm_data} \ + --k2 ${k2} \ + --combine-method lm_only \ + --task noisy_channel_translation \ + --lenpen ${lenpen} \ + --lm-wt ${lm_wt} \ + --gen-subset valid \ + --remove-bpe \ + --fp16 \ + --batch-size 10 +``` +### Noisy Channel Generation for German-English translation with fairseq + +Here are instructions for noisy channel generation with a direct model, channel model and language model as explained in section [Noisy Channel Modeling](#noisy-channel-modeling). + +Note: +- Download and install fairseq as per instructions [here](https://github.com/pytorch/fairseq) +- Preprocess and binarize the dataset as per instructions in section [Test Data Preprocessing](#test-data-preprocessing) + +```sh +binarized_data=data_dir/binarized +direct_model=de_en_seed4.pt +lm_model=en_lm.pt +lm_data=lm_data +ch_model=en_de.big.seed4.pt +wget https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/direct_models/seed4.pt -O ${direct_model} +wget https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/lm_model/transformer_lm.pt -O ${lm_model} +mkdir -p ${lm_data} +wget https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/lm_model/lm_dict/dict.txt -O ${lm_data}/dict.txt +wget https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/big.seed4.pt -O ${ch_model} + +k2=10 +lenpen=0.21 +lm_wt=0.50 +bw_wt=0.30 +fairseq-generate ${binarized_data} \ + --user-dir examples/fast_noisy_channel \ + --beam 5 \ + --path ${direct_model} \ + --lm-model ${lm_model} \ + --lm-data ${lm_data} \ + --channel-model ${ch_model} \ + --k2 ${k2} \ + --combine-method noisy_channel \ + --task noisy_channel_translation \ + --lenpen ${lenpen} \ + --lm-wt ${lm_wt} \ + --ch-wt ${bw_wt} \ + --gen-subset test \ + --remove-bpe \ + --fp16 \ + --batch-size 1 +``` +## Fast Noisy Channel Modeling + +[Bhosale et al. (2020)](http://www.statmt.org/wmt20/pdf/2020.wmt-1.68.pdf) introduces 3 approximations that speed up online noisy channel decoding - +- Smaller channel models (`Tranformer Base` with 1 encoder and decoder layer each vs. `Transformer Big`) + - This involves training a channel model that is possibly smaller and less accurate in terms of BLEU than a channel model of the same size as the direct model. + - Since the role of the channel model is mainly to assign low scores to generations from the language model if they don't translate back to the source, we may not need the most accurate channel model for this purpose. +- Smaller output vocabulary size for the channel model (~30,000 -> ~1000) + - The channel model doesn't need to score the full output vocabulary, it just needs to score the source tokens, which are completely known. + - This is specified using the arguments `--channel-scoring-type src_vocab --top-k-vocab 500` + - This means that the output vocabulary for the channel model will be the source tokens for all examples in the batch and the top-K most frequent tokens in the vocabulary + - This reduces the memory consumption needed to store channel model scores significantly +- Smaller number of candidates (`k2`) scored per beam + - This is specified by reducing the argument `--k2` + + +### Fast Noisy Channel Generation for German-English translation with fairseq + +Here are instructions for **fast** noisy channel generation with a direct model, channel model and language model as explained in section [Fast Noisy Channel Modeling](#fast-noisy-channel-modeling). The main differences are that we use a smaller channel model, reduce `--k2`, set `--channel-scoring-type src_vocab --top-k-vocab 500` and increase the `--batch-size`. + +Note: +- Download and install fairseq as per instructions [here](https://github.com/pytorch/fairseq) +- Preprocess and binarize the dataset as per instructions in section [Test Data Preprocessing](#test-data-preprocessing) + +```sh +binarized_data=data_dir/binarized +direct_model=de_en_seed4.pt +lm_model=en_lm.pt +lm_data=lm_data +small_ch_model=en_de.base_1_1.seed4.pt +wget https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/direct_models/seed4.pt -O ${direct_model} +wget https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/lm_model/transformer_lm.pt -O ${lm_model} +mkdir -p ${lm_data} +wget https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/lm_model/lm_dict/dict.txt -O ${lm_data}/dict.txt +wget https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/base_1_1.seed4.pt -O ${small_ch_model} + +k2=3 +lenpen=0.23 +lm_wt=0.58 +bw_wt=0.26 +fairseq-generate ${binarized_data} \ + --user-dir examples/fast_noisy_channel \ + --beam 5 \ + --path ${direct_model} \ + --lm-model ${lm_model} \ + --lm-data ${lm_data} \ + --channel-model ${small_ch_model} \ + --k2 ${k2} \ + --combine-method noisy_channel \ + --task noisy_channel_translation \ + --lenpen ${lenpen} \ + --lm-wt ${lm_wt} \ + --ch-wt ${bw_wt} \ + --gen-subset test \ + --remove-bpe \ + --fp16 \ + --batch-size 50 \ + --channel-scoring-type src_vocab --top-k-vocab 500 +``` + +## Test Data Preprocessing + +For preprocessing and binarizing the test sets for Romanian-English and German-English translation, we use the following script - + +```sh +FAIRSEQ=/path/to/fairseq +cd $FAIRSEQ +SCRIPTS=$FAIRSEQ/mosesdecoder/scripts +if [ ! -d "${SCRIPTS}" ]; then + echo 'Cloning Moses github repository (for tokenization scripts)...' + git clone https://github.com/moses-smt/mosesdecoder.git +fi +TOKENIZER=$SCRIPTS/tokenizer/tokenizer.perl +NORMALIZE=$SCRIPTS/tokenizer/normalize-punctuation.perl + +s=de +t=en +test=wmt18 + +mkdir -p data_dir + +# Tokenization +if [ $s == "ro" ] ; then + # Note: Get normalise-romanian.py and remove-diacritics.py from + # https://github.com/rsennrich/wmt16-scripts/tree/master/preprocess + sacrebleu -t $test -l $s-$t --echo src | \ + $NORMALIZE -l $s | \ + python normalise-romanian.py | \ + python remove-diacritics.py | \ + $TOKENIZER -l $s -a -q > data_dir/$test.$s-$t.$s +else + sacrebleu -t $test -l $s-$t --echo src | perl $NORMALIZE -l $s | perl $TOKENIZER -threads 8 -a -l $s > data_dir/$test.$s-$t.$s +fi + +sacrebleu -t $test -l $s-$t --echo ref | perl $NORMALIZE -l $t | perl $TOKENIZER -threads 8 -a -l $t > data_dir/$test.$s-$t.$t + + +# Applying BPE +src_bpe_code=/path/to/source/language/bpe/code +tgt_bpe_code=/path/to/target/language/bpe/code +src_dict=/path/to/source/language/dict +tgt_dict=/path/to/target/language/dict + +FASTBPE=$FAIRSEQ/fastBPE +if [ ! -d "${FASTBPE}" ] ; then + git clone https://github.com/glample/fastBPE.git + # Follow compilation instructions at https://github.com/glample/fastBPE + g++ -std=c++11 -pthread -O3 fastBPE/main.cc -IfastBPE -o fast +fi + +${FASTBPE}/fast applybpe data_dir/bpe.$test.$s-$t.$s data_dir/$test.$s-$t.$s ${src_bpe_code} +${FASTBPE}/fast applybpe data_dir/bpe.$test.$s-$t.$s data_dir/$test.$s-$t.$s ${tgt_bpe_code} + +fairseq-preprocess -s $s -t $t \ + --testpref data_dir/bpe.$test.$s-$t \ + --destdir data_dir/binarized \ + --srcdict ${src_dict} \ + --tgtdict ${tgt_dict} +``` + +## Calculating BLEU + +```sh +DETOKENIZER=$SCRIPTS/tokenizer/detokenizer.perl +cat ${generation_output} | grep -P "^H" | sort -V | cut -f 3- | $DETOKENIZER -l $t -q -a | sacrebleu -t $test -l $s-$t +``` + + +## Romanian-English Translation + +The direct and channel models are trained using bitext data (WMT16) combined with backtranslated data (The monolingual data used for backtranslation comes from http://data.statmt.org/rsennrich/wmt16_backtranslations/ (Sennrich et al., 2016c)) + +The backtranslated data is generated using an ensemble of 3 English-Romanian models trained on bitext training data (WMT16) with unrestricted sampling. + +### BPE Codes and Dictionary + +We learn a joint BPE vocabulary of 18K types on the bitext training data which is used for both the source and target. +||Path| +|----------|------| +| BPE Code | [joint_bpe_18k](https://dl.fbaipublicfiles.com/fast_noisy_channel/ro_en/bpe_18k) | +| Dictionary | [dict](https://dl.fbaipublicfiles.com/fast_noisy_channel/ro_en/dict) | + +### Direct Models +For Ro-En with backtranslation, the direct and channel models use a Transformer-Big architecture. + +| Seed | Model | +|----|----| +| 2 | [ro_en_seed2.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/ro_en/direct_models/seed2.pt) +| 4 | [ro_en_seed4.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/ro_en/direct_models/seed4.pt) +| 6 | [ro_en_seed6.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/ro_en/direct_models/seed6.pt) + +### Channel Models +For channel models, we follow the same steps as for the direct models. But backtranslated data is generated in the opposite direction using [this Romanian monolingual data](http://data.statmt.org/rsennrich/wmt16_backtranslations/). +The best lenpen, LM weight and CH weight are obtained by sweeping over the validation set (wmt16/dev) using beam 5. +| Model Size | Lenpen | LM Weight | CH Weight | Seed 2 | Seed 4 | Seed 6 | +|----|----|----|----|----|----|----| +| `big` | 0.84 | 0.64 | 0.56 | [big.seed2.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/ro_en/channel_models/big.seed2.pt) | [big.seed2.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/ro_en/channel_models/big.seed2.pt) | [big.seed2.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/ro_en/channel_models/big.seed2.pt) | +| `base_1_1` | 0.63 | 0.40 | 0.37 | [base_1_1.seed2.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/ro_en/channel_models/base_1_1.seed2.pt) | [base_1_1.seed4.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/ro_en/channel_models/base_1_1.seed4.pt) | [base_1_1.seed6.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/ro_en/channel_models/base_1_1.seed6.pt) | + +### Language Model +The model is trained on de-duplicated English Newscrawl data from 2007-2018 comprising 186 million sentences or 4.5B words after normalization and tokenization. +| | Path | +|----|----| +| `--lm-model` | [transformer_en_lm](https://dl.fbaipublicfiles.com/fast_noisy_channel/ro_en/lm_model/transformer_lm.pt) | +| `--lm-data` | [lm_data](https://dl.fbaipublicfiles.com/fast_noisy_channel/ro_en/lm_model/lm_dict) + +## German-English Translation + +### BPE Codes and Dictionaries + +| | Path| +|----------|------| +| Source BPE Code | [de_bpe_code_24K](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/de_bpe_code_24K) | +| Target BPE Code | [en_bpe_code_24K](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/en_bpe_code_24K) +| Source Dictionary | [de_dict](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/de_dict) | +| Target Dictionary | [en_dict](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/en_dict) | + +### Direct Models +We train on WMT’19 training data. Following [Ng et al., 2019](http://statmt.org/wmt19/pdf/53/WMT33.pdf), we apply language identification filtering and remove sentences longer than 250 tokens as well as sentence pairs with a source/target length ratio exceeding 1.5. This results in 26.8M sentence pairs. +We use the Transformer-Big architecture for the direct model. + +| Seed | Model | +|:----:|----| +| 4 | [de_en_seed4.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/direct_models/seed4.pt) +| 5 | [de_en_seed5.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/direct_models/seed5.pt) +| 6 | [de_en_seed6.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/direct_models/seed6.pt) + +### Channel Models + +We train on WMT’19 training data. Following [Ng et al., 2019](http://statmt.org/wmt19/pdf/53/WMT33.pdf), we apply language identification filtering and remove sentences longer than 250 tokens as well as sentence pairs with a source/target length ratio exceeding 1.5. This results in 26.8M sentence pairs. + +| Model Size | Seed 4 | Seed 5 | Seed 6 | +|----|----|----|----| +| `big` | [big.seed4.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/big.seed4.pt) | [big.seed5.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/big.seed5.pt) | [big.seed6.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/big.seed6.pt) | +| `big_1_1` | [big_1_1.seed4.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/big_1_1.seed4.pt) | [big_1_1.seed5.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/big_1_1.seed5.pt) | [big_1_1.seed6.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/big_1_1.seed6.pt) | +| `base` | [base.seed4.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/base.seed4.pt) | [base.seed5.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/base.seed5.pt) | [base.seed6.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/base.seed6.pt) | +| `base_1_1` | [base_1_1.seed4.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/base_1_1.seed4.pt) | [base_1_1.seed5.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/base_1_1.seed5.pt) | [base_1_1.seed6.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/base_1_1.seed6.pt) | +| `half` | [half.seed4.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/half.seed4.pt) | [half.seed5.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/half.seed5.pt) | [half.seed6.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/half.seed6.pt) | +| `half_1_1` | [half_1_1.seed4.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/half_1_1.seed4.pt) | [half_1_1.seed5.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/half_1_1.seed5.pt) | [half_1_1.seed6.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/half_1_1.seed6.pt) | +| `quarter` | [quarter.seed4.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/quarter.seed4.pt) | [quarter.seed5.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/quarter.seed5.pt) | [quarter.seed6.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/quarter.seed6.pt) | +| `quarter_1_1` | [quarter_1_1.seed4.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/quarter_1_1.seed4.pt) | [quarter_1_1.seed5.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/quarter_1_1.seed5.pt) | [quarter_1_1.seed6.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/quarter_1_1.seed6.pt) | +| `8th` | [8th.seed4.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/8th.seed4.pt) | [8th.seed5.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/8th.seed5.pt) | [8th.seed6.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/8th.seed6.pt) | +| `8th_1_1` | [8th_1_1.seed4.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/8th_1_1.seed4.pt) | [8th_1_1.seed5.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/8th_1_1.seed5.pt) | [8th_1_1.seed6.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/8th_1_1.seed6.pt) | +| `16th` | [16th.seed4.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/16th.seed4.pt) | [16th.seed5.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/16th.seed5.pt) | [16th.seed6.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/16th.seed6.pt) | +| `16th_1_1` | [16th_1_1.seed4.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/16th_1_1.seed4.pt) | [16th_1_1.seed5.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/16th_1_1.seed5.pt) | [16th_1_1.seed6.pt](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/channel_models/16th_1_1.seed6.pt) | + +### Language Model +The model is trained on de-duplicated English Newscrawl data from 2007-2018 comprising 186 million sentences or 4.5B words after normalization and tokenization. +| | Path | +|----|----| +| `--lm-model` | [transformer_en_lm](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/lm_model/transformer_lm.pt) | +| `--lm-data` | [lm_data](https://dl.fbaipublicfiles.com/fast_noisy_channel/de_en/lm_model/lm_dict/) + + +## Citation + +```bibtex +@inproceedings{bhosale2020language, + title={Language Models not just for Pre-training: Fast Online Neural Noisy Channel Modeling}, + author={Shruti Bhosale and Kyra Yee and Sergey Edunov and Michael Auli}, + booktitle={Proceedings of the Fifth Conference on Machine Translation (WMT)}, + year={2020}, +} + +@inproceedings{yee2019simple, + title={Simple and Effective Noisy Channel Modeling for Neural Machine Translation}, + author={Yee, Kyra and Dauphin, Yann and Auli, Michael}, + booktitle={Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)}, + pages={5700--5705}, + year={2019} +} +``` diff --git a/examples/fast_noisy_channel/__init__.py b/examples/fast_noisy_channel/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..9b248c3a24e12ad3da885a7f328c714942de2e6b --- /dev/null +++ b/examples/fast_noisy_channel/__init__.py @@ -0,0 +1,8 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from . import noisy_channel_translation # noqa +from . import noisy_channel_sequence_generator # noqa +from . import noisy_channel_beam_search # noqa diff --git a/examples/fast_noisy_channel/noisy_channel_beam_search.py b/examples/fast_noisy_channel/noisy_channel_beam_search.py new file mode 100644 index 0000000000000000000000000000000000000000..23869ebcd0c438f36e310c8ccddd3b5c07a71182 --- /dev/null +++ b/examples/fast_noisy_channel/noisy_channel_beam_search.py @@ -0,0 +1,71 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +from fairseq.search import Search + + +class NoisyChannelBeamSearch(Search): + + def __init__(self, tgt_dict): + super().__init__(tgt_dict) + self.fw_scores_buf = None + self.lm_scores_buf = None + + def _init_buffers(self, t): + # super()._init_buffers(t) + if self.fw_scores_buf is None: + self.scores_buf = t.new() + self.indices_buf = torch.LongTensor().to(device=t.device) + self.beams_buf = torch.LongTensor().to(device=t.device) + self.fw_scores_buf = t.new() + self.lm_scores_buf = t.new() + + def combine_fw_bw(self, combine_method, fw_cum, bw, step): + if combine_method == "noisy_channel": + fw_norm = fw_cum.div(step + 1) + lprobs = bw + fw_norm + elif combine_method == "lm_only": + lprobs = bw + fw_cum + + return lprobs + + def step(self, step, fw_lprobs, scores, bw_lprobs, lm_lprobs, combine_method): + self._init_buffers(fw_lprobs) + bsz, beam_size, vocab_size = fw_lprobs.size() + + if step == 0: + # at the first step all hypotheses are equally likely, so use + # only the first beam + fw_lprobs = fw_lprobs[:, ::beam_size, :].contiguous() + bw_lprobs = bw_lprobs[:, ::beam_size, :].contiguous() + # nothing to add since we are at the first step + fw_lprobs_cum = fw_lprobs + + else: + # make probs contain cumulative scores for each hypothesis + raw_scores = (scores[:, :, step - 1].unsqueeze(-1)) + fw_lprobs_cum = (fw_lprobs.add(raw_scores)) + + combined_lprobs = self.combine_fw_bw(combine_method, fw_lprobs_cum, bw_lprobs, step) + + # choose the top k according to the combined noisy channel model score + torch.topk( + combined_lprobs.view(bsz, -1), + k=min( + # Take the best 2 x beam_size predictions. We'll choose the first + # beam_size of these which don't predict eos to continue with. + beam_size * 2, + combined_lprobs.view(bsz, -1).size(1) - 1, # -1 so we never select pad + ), + out=(self.scores_buf, self.indices_buf), + ) + # save corresponding fw and lm scores + self.fw_scores_buf = torch.gather(fw_lprobs_cum.view(bsz, -1), 1, self.indices_buf) + self.lm_scores_buf = torch.gather(lm_lprobs.view(bsz, -1), 1, self.indices_buf) + # Project back into relative indices and beams + self.beams_buf = self.indices_buf // vocab_size + self.indices_buf.fmod_(vocab_size) + return self.scores_buf, self.fw_scores_buf, self.lm_scores_buf, self.indices_buf, self.beams_buf diff --git a/examples/fast_noisy_channel/noisy_channel_sequence_generator.py b/examples/fast_noisy_channel/noisy_channel_sequence_generator.py new file mode 100644 index 0000000000000000000000000000000000000000..ea8fae98e87e9f3e69bc51987703a6429eb0c92a --- /dev/null +++ b/examples/fast_noisy_channel/noisy_channel_sequence_generator.py @@ -0,0 +1,842 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from typing import Dict, List, Optional + +import math +import numpy as np + +import torch +import torch.nn.functional as F +from torch import Tensor + +from .noisy_channel_beam_search import NoisyChannelBeamSearch +from fairseq.sequence_generator import EnsembleModel + + +class NoisyChannelSequenceGenerator(object): + def __init__( + self, + combine_method, + tgt_dict, + src_dict=None, + beam_size=1, + max_len_a=0, + max_len_b=200, + min_len=1, + len_penalty=1.0, + unk_penalty=0.0, + retain_dropout=False, + temperature=1.0, + match_source_len=False, + no_repeat_ngram_size=0, + normalize_scores=True, + channel_models=None, + k2=10, + ch_weight=1.0, + channel_scoring_type='log_norm', + top_k_vocab=0, + lm_models=None, + lm_dict=None, + lm_weight=1.0, + normalize_lm_scores_by_tgt_len=False, + ): + """Generates translations of a given source sentence, + using beam search with noisy channel decoding. + + Args: + combine_method (string, optional): Method to combine direct, LM and + channel model scores (default: None) + tgt_dict (~fairseq.data.Dictionary): target dictionary + src_dict (~fairseq.data.Dictionary): source dictionary + beam_size (int, optional): beam width (default: 1) + max_len_a/b (int, optional): generate sequences of maximum length + ax + b, where x is the source length + min_len (int, optional): the minimum length of the generated output + (not including end-of-sentence) + len_penalty (float, optional): length penalty, where <1.0 favors + shorter, >1.0 favors longer sentences (default: 1.0) + unk_penalty (float, optional): unknown word penalty, where <0 + produces more unks, >0 produces fewer (default: 0.0) + retain_dropout (bool, optional): use dropout when generating + (default: False) + temperature (float, optional): temperature, where values + >1.0 produce more uniform samples and values <1.0 produce + sharper samples (default: 1.0) + match_source_len (bool, optional): outputs should match the source + length (default: False) + no_repeat_ngram_size (int, optional): Size of n-grams that we avoid + repeating in the generation (default: 0) + normalize_scores (bool, optional): normalize scores by the length + of the output (default: True) + channel_models (List[~fairseq.models.FairseqModel]): ensemble of models + translating from the target to the source + k2 (int, optional): Top K2 candidates to score per beam at each step (default:10) + ch_weight (int, optional): Weight associated with the channel model score + assuming that the direct model score has weight 1.0 (default: 1.0) + channel_scoring_type (str, optional): String specifying how to score + the channel model (default: 'log_norm') + top_k_vocab (int, optional): If `channel_scoring_type` is `'src_vocab'` or + `'src_vocab_batched'`, then this parameter specifies the number of + most frequent tokens to include in the channel model output vocabulary, + in addition to the source tokens in the input batch (default: 0) + lm_models (List[~fairseq.models.FairseqModel]): ensemble of models + generating text in the target language + lm_dict (~fairseq.data.Dictionary): LM Model dictionary + lm_weight (int, optional): Weight associated with the LM model score + assuming that the direct model score has weight 1.0 (default: 1.0) + normalize_lm_scores_by_tgt_len (bool, optional): Should we normalize LM scores + by the target length? By default, we normalize the combination of + LM and channel model scores by the source length + """ + self.pad = tgt_dict.pad() + self.unk = tgt_dict.unk() + self.eos = tgt_dict.eos() + self.vocab_size = len(tgt_dict) + self.beam_size = beam_size + # the max beam size is the dictionary size - 1, since we never select pad + self.beam_size = min(beam_size, self.vocab_size - 1) + self.max_len_a = max_len_a + self.max_len_b = max_len_b + self.min_len = min_len + self.normalize_scores = normalize_scores + self.len_penalty = len_penalty + self.unk_penalty = unk_penalty + self.retain_dropout = retain_dropout + self.temperature = temperature + self.match_source_len = match_source_len + self.no_repeat_ngram_size = no_repeat_ngram_size + self.channel_models = channel_models + self.src_dict = src_dict + self.tgt_dict = tgt_dict + self.combine_method = combine_method + self.k2 = k2 + self.ch_weight = ch_weight + self.channel_scoring_type = channel_scoring_type + self.top_k_vocab = top_k_vocab + self.lm_models = lm_models + self.lm_dict = lm_dict + self.lm_weight = lm_weight + self.log_softmax_fn = torch.nn.LogSoftmax(dim=1) + self.normalize_lm_scores_by_tgt_len = normalize_lm_scores_by_tgt_len + + self.share_tgt_dict = (self.lm_dict == self.tgt_dict) + self.tgt_to_lm = make_dict2dict(tgt_dict, lm_dict) + + self.ch_scoring_bsz = 3072 + + assert temperature > 0, '--temperature must be greater than 0' + + self.search = NoisyChannelBeamSearch(tgt_dict) + + @torch.no_grad() + def generate( + self, + models, + sample, + prefix_tokens=None, + bos_token=None, + **kwargs + ): + """Generate a batch of translations. + Args: + models (List[~fairseq.models.FairseqModel]): ensemble of models + sample (dict): batch + prefix_tokens (torch.LongTensor, optional): force decoder to begin + with these tokens + """ + model = EnsembleModel(models) + incremental_states = torch.jit.annotate( + List[Dict[str, Dict[str, Optional[Tensor]]]], + [ + torch.jit.annotate(Dict[str, Dict[str, Optional[Tensor]]], {}) + for i in range(model.models_size) + ], + ) + if not self.retain_dropout: + model.eval() + + # model.forward normally channels prev_output_tokens into the decoder + # separately, but SequenceGenerator directly calls model.encoder + encoder_input = { + k: v for k, v in sample['net_input'].items() + if k != 'prev_output_tokens' + } + src_tokens = encoder_input['src_tokens'] + src_lengths_no_eos = (src_tokens.ne(self.eos) & src_tokens.ne(self.pad)).long().sum(dim=1) + input_size = src_tokens.size() + # batch dimension goes first followed by source lengths + bsz = input_size[0] + src_len = input_size[1] + beam_size = self.beam_size + + if self.match_source_len: + max_len = src_lengths_no_eos.max().item() + else: + max_len = min( + int(self.max_len_a * src_len + self.max_len_b), + # exclude the EOS marker + model.max_decoder_positions() - 1, + ) + + # compute the encoder output for each beam + encoder_outs = model.forward_encoder(encoder_input) + new_order = torch.arange(bsz).view(-1, 1).repeat(1, beam_size).view(-1) + new_order = new_order.to(src_tokens.device).long() + encoder_outs = model.reorder_encoder_out(encoder_outs, new_order) + + src_lengths = encoder_input['src_lengths'] + # initialize buffers + scores = src_tokens.new(bsz * beam_size, max_len + 1).float().fill_(0) + lm_prefix_scores = src_tokens.new(bsz * beam_size).float().fill_(0) + + scores_buf = scores.clone() + tokens = src_tokens.new(bsz * beam_size, max_len + 2).long().fill_(self.pad) + tokens_buf = tokens.clone() + tokens[:, 0] = self.eos if bos_token is None else bos_token + + # reorder source tokens so they may be used as a reference in generating P(S|T) + src_tokens = reorder_all_tokens(src_tokens, src_lengths, self.src_dict.eos_index) + + src_tokens = src_tokens.repeat(1, beam_size).view(-1, src_len) + src_lengths = src_lengths.view(bsz, -1).repeat(1, beam_size).view(bsz*beam_size, -1) + + attn, attn_buf = None, None + nonpad_idxs = None + + # The cands_to_ignore indicates candidates that should be ignored. + # For example, suppose we're sampling and have already finalized 2/5 + # samples. Then the cands_to_ignore would mark 2 positions as being ignored, + # so that we only finalize the remaining 3 samples. + cands_to_ignore = src_tokens.new_zeros(bsz, beam_size).eq(-1) # forward and backward-compatible False mask + + # list of completed sentences + finalized = [[] for i in range(bsz)] + finished = [False for i in range(bsz)] + num_remaining_sent = bsz + + # number of candidate hypos per step + cand_size = 2 * beam_size # 2 x beam size in case half are EOS + + # offset arrays for converting between different indexing schemes + bbsz_offsets = (torch.arange(0, bsz) * beam_size).unsqueeze(1).type_as(tokens) + cand_offsets = torch.arange(0, cand_size).type_as(tokens) + + # helper function for allocating buffers on the fly + buffers = {} + + def buffer(name, type_of=tokens): # noqa + if name not in buffers: + buffers[name] = type_of.new() + return buffers[name] + + def is_finished(sent, step, unfin_idx): + """ + Check whether we've finished generation for a given sentence, by + comparing the worst score among finalized hypotheses to the best + possible score among unfinalized hypotheses. + """ + assert len(finalized[sent]) <= beam_size + if len(finalized[sent]) == beam_size: + return True + return False + + def finalize_hypos(step, bbsz_idx, eos_scores, combined_noisy_channel_eos_scores): + """ + Finalize the given hypotheses at this step, while keeping the total + number of finalized hypotheses per sentence <= beam_size. + + Note: the input must be in the desired finalization order, so that + hypotheses that appear earlier in the input are preferred to those + that appear later. + + Args: + step: current time step + bbsz_idx: A vector of indices in the range [0, bsz*beam_size), + indicating which hypotheses to finalize + eos_scores: A vector of the same size as bbsz_idx containing + fw scores for each hypothesis + combined_noisy_channel_eos_scores: A vector of the same size as bbsz_idx containing + combined noisy channel scores for each hypothesis + """ + assert bbsz_idx.numel() == eos_scores.numel() + + # clone relevant token and attention tensors + tokens_clone = tokens.index_select(0, bbsz_idx) + tokens_clone = tokens_clone[:, 1:step + 2] # skip the first index, which is EOS + assert not tokens_clone.eq(self.eos).any() + tokens_clone[:, step] = self.eos + attn_clone = attn.index_select(0, bbsz_idx)[:, :, 1:step+2] if attn is not None else None + + # compute scores per token position + pos_scores = scores.index_select(0, bbsz_idx)[:, :step+1] + pos_scores[:, step] = eos_scores + # convert from cumulative to per-position scores + pos_scores[:, 1:] = pos_scores[:, 1:] - pos_scores[:, :-1] + + # normalize sentence-level scores + if self.normalize_scores: + combined_noisy_channel_eos_scores /= (step + 1) ** self.len_penalty + + cum_unfin = [] + prev = 0 + for f in finished: + if f: + prev += 1 + else: + cum_unfin.append(prev) + + sents_seen = set() + for i, (idx, score) in enumerate(zip(bbsz_idx.tolist(), combined_noisy_channel_eos_scores.tolist())): + unfin_idx = idx // beam_size + sent = unfin_idx + cum_unfin[unfin_idx] + + sents_seen.add((sent, unfin_idx)) + + if self.match_source_len and step > src_lengths_no_eos[unfin_idx]: + score = -math.inf + + def get_hypo(): + + if attn_clone is not None: + # remove padding tokens from attn scores + hypo_attn = attn_clone[i][nonpad_idxs[sent]] + _, alignment = hypo_attn.max(dim=0) + else: + hypo_attn = None + alignment = None + + return { + 'tokens': tokens_clone[i], + 'score': score, + 'attention': hypo_attn, # src_len x tgt_len + 'alignment': alignment, + 'positional_scores': pos_scores[i], + } + + if len(finalized[sent]) < beam_size: + finalized[sent].append(get_hypo()) + + newly_finished = [] + for sent, unfin_idx in sents_seen: + # check termination conditions for this sentence + if not finished[sent] and is_finished(sent, step, unfin_idx): + finished[sent] = True + newly_finished.append(unfin_idx) + return newly_finished + + def noisy_channel_rescoring(lprobs, beam_size, bsz, src_tokens, tokens, k): + """Rescore the top k hypothesis from each beam using noisy channel modeling + Returns: + new_fw_lprobs: the direct model probabilities after pruning the top k + new_ch_lm_lprobs: the combined channel and language model probabilities + new_lm_lprobs: the language model probabilities after pruning the top k + """ + with torch.no_grad(): + lprobs_size = lprobs.size() + if prefix_tokens is not None and step < prefix_tokens.size(1): + probs_slice = lprobs.view(bsz, -1, lprobs.size(-1))[:, 0, :] + cand_scores = torch.gather( + probs_slice, dim=1, + index=prefix_tokens[:, step].view(-1, 1).data + ).expand(-1, beam_size).contiguous().view(bsz*beam_size, 1) + cand_indices = prefix_tokens[:, step].view(-1, 1).expand(bsz, beam_size).data.contiguous().view(bsz*beam_size, 1) + + # need to calculate and save fw and lm probs for prefix tokens + fw_top_k = cand_scores + fw_top_k_idx = cand_indices + k = 1 + else: + # take the top k best words for every sentence in batch*beam + fw_top_k, fw_top_k_idx = torch.topk(lprobs.view(beam_size*bsz, -1), k=k) + eos_idx = torch.nonzero(fw_top_k_idx.view(bsz*beam_size*k, -1) == self.eos)[:, 0] + ch_scores = fw_top_k.new_full((beam_size*bsz*k, ), 0) + src_size = torch.sum(src_tokens[:, :] != self.src_dict.pad_index, dim=1, keepdim=True, dtype=fw_top_k.dtype) + + if self.combine_method != "lm_only": + temp_src_tokens_full = src_tokens[:, :].repeat(1, k).view(bsz*beam_size*k, -1) + not_padding = temp_src_tokens_full[:, 1:] != self.src_dict.pad_index + cur_tgt_size = step+2 + + # add eos to all candidate sentences except those that already end in eos + eos_tokens = tokens[:, 0].repeat(1, k).view(-1, 1) + eos_tokens[eos_idx] = self.tgt_dict.pad_index + + if step == 0: + channel_input = torch.cat((fw_top_k_idx.view(-1, 1), eos_tokens), 1) + else: + # move eos from beginning to end of target sentence + channel_input = torch.cat((tokens[:, 1:step + 1].repeat(1, k).view(-1, step), fw_top_k_idx.view(-1, 1), eos_tokens), 1) + + ch_input_lengths = torch.tensor(np.full(channel_input.size(0), cur_tgt_size)) + ch_input_lengths[eos_idx] = cur_tgt_size-1 + if self.channel_scoring_type == "unnormalized": + ch_encoder_output = channel_model.encoder(channel_input, src_lengths=ch_input_lengths) + ch_decoder_output, _ = channel_model.decoder(temp_src_tokens_full, encoder_out=ch_encoder_output, features_only=True) + del ch_encoder_output + ch_intermed_scores = channel_model.decoder.unnormalized_scores_given_target(ch_decoder_output, target_ids=temp_src_tokens_full[:, 1:]) + ch_intermed_scores = ch_intermed_scores.float() + ch_intermed_scores *= not_padding.float() + ch_scores = torch.sum(ch_intermed_scores, dim=1) + elif self.channel_scoring_type == "k2_separate": + for k_idx in range(k): + k_eos_tokens = eos_tokens[k_idx::k, :] + if step == 0: + k_ch_input = torch.cat((fw_top_k_idx[:, k_idx:k_idx+1], k_eos_tokens), 1) + else: + # move eos from beginning to end of target sentence + k_ch_input = torch.cat((tokens[:, 1:step + 1], fw_top_k_idx[:, k_idx:k_idx+1], k_eos_tokens), 1) + k_ch_input_lengths = ch_input_lengths[k_idx::k] + k_ch_output = channel_model(k_ch_input, k_ch_input_lengths, src_tokens) + k_ch_lprobs = channel_model.get_normalized_probs(k_ch_output, log_probs=True) + k_ch_intermed_scores = torch.gather(k_ch_lprobs[:, :-1, :], 2, src_tokens[:, 1:].unsqueeze(2)).squeeze(2) + k_ch_intermed_scores *= not_padding.float() + ch_scores[k_idx::k] = torch.sum(k_ch_intermed_scores, dim=1) + elif self.channel_scoring_type == "src_vocab": + ch_encoder_output = channel_model.encoder(channel_input, src_lengths=ch_input_lengths) + ch_decoder_output, _ = channel_model.decoder(temp_src_tokens_full, encoder_out=ch_encoder_output, features_only=True) + + del ch_encoder_output + ch_lprobs = normalized_scores_with_batch_vocab( + channel_model.decoder, + ch_decoder_output, src_tokens, k, bsz, beam_size, + self.src_dict.pad_index, top_k=self.top_k_vocab) + ch_scores = torch.sum(ch_lprobs, dim=1) + elif self.channel_scoring_type == "src_vocab_batched": + ch_bsz_size = temp_src_tokens_full.shape[0] + ch_lprobs_list = [None] * len(range(0, ch_bsz_size, self.ch_scoring_bsz)) + for i, start_idx in enumerate(range(0, ch_bsz_size, self.ch_scoring_bsz)): + end_idx = min(start_idx + self.ch_scoring_bsz, ch_bsz_size) + temp_src_tokens_full_batch = temp_src_tokens_full[start_idx:end_idx, :] + channel_input_batch = channel_input[start_idx:end_idx, :] + ch_input_lengths_batch = ch_input_lengths[start_idx:end_idx] + ch_encoder_output_batch = channel_model.encoder(channel_input_batch, src_lengths=ch_input_lengths_batch) + ch_decoder_output_batch, _ = channel_model.decoder(temp_src_tokens_full_batch, encoder_out=ch_encoder_output_batch, features_only=True) + ch_lprobs_list[i] = normalized_scores_with_batch_vocab( + channel_model.decoder, + ch_decoder_output_batch, src_tokens, k, bsz, beam_size, + self.src_dict.pad_index, top_k=self.top_k_vocab, + start_idx=start_idx, end_idx=end_idx) + ch_lprobs = torch.cat(ch_lprobs_list, dim=0) + ch_scores = torch.sum(ch_lprobs, dim=1) + else: + ch_output = channel_model(channel_input, ch_input_lengths, temp_src_tokens_full) + ch_lprobs = channel_model.get_normalized_probs(ch_output, log_probs=True) + ch_intermed_scores = torch.gather(ch_lprobs[:, :-1, :], 2, temp_src_tokens_full[:, 1:].unsqueeze(2)).squeeze().view(bsz*beam_size*k, -1) + ch_intermed_scores *= not_padding.float() + ch_scores = torch.sum(ch_intermed_scores, dim=1) + + else: + cur_tgt_size = 0 + ch_scores = ch_scores.view(bsz*beam_size, k) + expanded_lm_prefix_scores = lm_prefix_scores.unsqueeze(1).expand(-1, k).flatten() + + if self.share_tgt_dict: + lm_scores = get_lm_scores(lm, tokens[:, :step + 1].view(-1, step+1), lm_incremental_states, fw_top_k_idx.view(-1, 1), torch.tensor(np.full(tokens.size(0), step+1)), k) + else: + new_lm_input = dict2dict(tokens[:, :step + 1].view(-1, step+1), self.tgt_to_lm) + new_cands = dict2dict(fw_top_k_idx.view(-1, 1), self.tgt_to_lm) + lm_scores = get_lm_scores(lm, new_lm_input, lm_incremental_states, new_cands, torch.tensor(np.full(tokens.size(0), step+1)), k) + + lm_scores.add_(expanded_lm_prefix_scores) + ch_lm_scores = combine_ch_lm(self.combine_method, ch_scores, lm_scores, src_size, cur_tgt_size) + # initialize all as min value + new_fw_lprobs = ch_scores.new(lprobs_size).fill_(-1e17).view(bsz*beam_size, -1) + new_ch_lm_lprobs = ch_scores.new(lprobs_size).fill_(-1e17).view(bsz*beam_size, -1) + new_lm_lprobs = ch_scores.new(lprobs_size).fill_(-1e17).view(bsz*beam_size, -1) + new_fw_lprobs[:, self.pad] = -math.inf + new_ch_lm_lprobs[:, self.pad] = -math.inf + new_lm_lprobs[:, self.pad] = -math.inf + + new_fw_lprobs.scatter_(1, fw_top_k_idx, fw_top_k) + new_ch_lm_lprobs.scatter_(1, fw_top_k_idx, ch_lm_scores) + new_lm_lprobs.scatter_(1, fw_top_k_idx, lm_scores.view(-1, k)) + return new_fw_lprobs, new_ch_lm_lprobs, new_lm_lprobs + + def combine_ch_lm(combine_type, ch_scores, lm_scores1, src_size, tgt_size): + if self.channel_scoring_type == "unnormalized": + ch_scores = self.log_softmax_fn( + ch_scores.view(-1, self.beam_size * self.k2) + ).view(ch_scores.shape) + ch_scores = ch_scores * self.ch_weight + lm_scores1 = lm_scores1 * self.lm_weight + + if combine_type == "lm_only": + # log P(T|S) + log P(T) + ch_scores = lm_scores1.view(ch_scores.size()) + elif combine_type == "noisy_channel": + # 1/t log P(T|S) + 1/s log P(S|T) + 1/t log P(T) + if self.normalize_lm_scores_by_tgt_len: + ch_scores.div_(src_size) + lm_scores_norm = lm_scores1.view(ch_scores.size()).div(tgt_size) + ch_scores.add_(lm_scores_norm) + # 1/t log P(T|S) + 1/s log P(S|T) + 1/s log P(T) + else: + ch_scores.add_(lm_scores1.view(ch_scores.size())) + ch_scores.div_(src_size) + + return ch_scores + + if self.channel_models is not None: + channel_model = self.channel_models[0] # assume only one channel_model model + else: + channel_model = None + + lm = EnsembleModel(self.lm_models) + lm_incremental_states = torch.jit.annotate( + List[Dict[str, Dict[str, Optional[Tensor]]]], + [ + torch.jit.annotate(Dict[str, Dict[str, Optional[Tensor]]], {}) + for i in range(lm.models_size) + ], + ) + + reorder_state = None + batch_idxs = None + for step in range(max_len + 1): # one extra step for EOS marker + # reorder decoder internal states based on the prev choice of beams + if reorder_state is not None: + if batch_idxs is not None: + # update beam indices to take into account removed sentences + corr = batch_idxs - torch.arange(batch_idxs.numel()).type_as(batch_idxs) + reorder_state.view(-1, beam_size).add_(corr.unsqueeze(-1) * beam_size) + model.reorder_incremental_state(incremental_states, reorder_state) + encoder_outs = model.reorder_encoder_out(encoder_outs, reorder_state) + + lm.reorder_incremental_state(lm_incremental_states, reorder_state) + + fw_lprobs, avg_attn_scores = model.forward_decoder( + tokens[:, :step + 1], encoder_outs, incremental_states, temperature=self.temperature, + ) + + fw_lprobs[:, self.pad] = -math.inf # never select pad + fw_lprobs[:, self.unk] -= self.unk_penalty # apply unk penalty + fw_lprobs, ch_lm_lprobs, lm_lprobs = noisy_channel_rescoring(fw_lprobs, beam_size, bsz, src_tokens, tokens, self.k2) + + # handle min and max length constraints + if step >= max_len: + fw_lprobs[:, :self.eos] = -math.inf + fw_lprobs[:, self.eos + 1:] = -math.inf + elif step < self.min_len: + fw_lprobs[:, self.eos] = -math.inf + + # handle prefix tokens (possibly with different lengths) + if prefix_tokens is not None and step < prefix_tokens.size(1): + prefix_toks = prefix_tokens[:, step].unsqueeze(-1).repeat(1, beam_size).view(-1) + prefix_mask = prefix_toks.ne(self.pad) + + prefix_fw_lprobs = fw_lprobs.gather(-1, prefix_toks.unsqueeze(-1)) + fw_lprobs[prefix_mask] = -math.inf + fw_lprobs[prefix_mask] = fw_lprobs[prefix_mask].scatter_( + -1, prefix_toks[prefix_mask].unsqueeze(-1), prefix_fw_lprobs + ) + + prefix_ch_lm_lprobs = ch_lm_lprobs.gather(-1, prefix_toks.unsqueeze(-1)) + ch_lm_lprobs[prefix_mask] = -math.inf + ch_lm_lprobs[prefix_mask] = ch_lm_lprobs[prefix_mask].scatter_( + -1, prefix_toks[prefix_mask].unsqueeze(-1), prefix_ch_lm_lprobs + ) + + prefix_lm_lprobs = lm_lprobs.gather(-1, prefix_toks.unsqueeze(-1)) + lm_lprobs[prefix_mask] = -math.inf + lm_lprobs[prefix_mask] = lm_lprobs[prefix_mask].scatter_( + -1, prefix_toks[prefix_mask].unsqueeze(-1), prefix_lm_lprobs + ) + + # if prefix includes eos, then we should make sure tokens and + # scores are the same across all beams + eos_mask = prefix_toks.eq(self.eos) + if eos_mask.any(): + # validate that the first beam matches the prefix + first_beam = tokens[eos_mask].view(-1, beam_size, tokens.size(-1))[:, 0, 1:step + 1] + eos_mask_batch_dim = eos_mask.view(-1, beam_size)[:, 0] + target_prefix = prefix_tokens[eos_mask_batch_dim][:, :step] + assert (first_beam == target_prefix).all() + + def replicate_first_beam(tensor, mask): + tensor = tensor.view(-1, beam_size, tensor.size(-1)) + tensor[mask] = tensor[mask][:, :1, :] + return tensor.view(-1, tensor.size(-1)) + + # copy tokens, scores and lprobs from the first beam to all beams + tokens = replicate_first_beam(tokens, eos_mask_batch_dim) + scores = replicate_first_beam(scores, eos_mask_batch_dim) + + fw_lprobs = replicate_first_beam(fw_lprobs, eos_mask_batch_dim) + ch_lm_lprobs = replicate_first_beam(ch_lm_lprobs, eos_mask_batch_dim) + lm_lprobs = replicate_first_beam(lm_lprobs, eos_mask_batch_dim) + + if self.no_repeat_ngram_size > 0: + # for each beam and batch sentence, generate a list of previous ngrams + gen_ngrams = [{} for bbsz_idx in range(bsz * beam_size)] + for bbsz_idx in range(bsz * beam_size): + gen_tokens = tokens[bbsz_idx].tolist() + for ngram in zip(*[gen_tokens[i:] for i in range(self.no_repeat_ngram_size)]): + gen_ngrams[bbsz_idx][tuple(ngram[:-1])] = \ + gen_ngrams[bbsz_idx].get(tuple(ngram[:-1]), []) + [ngram[-1]] + + # Record attention scores + if avg_attn_scores is not None: + if attn is None: + attn = scores.new(bsz * beam_size, src_tokens.size(1), max_len + 2) + attn_buf = attn.clone() + nonpad_idxs = src_tokens.ne(self.pad) + attn[:, :, step + 1].copy_(avg_attn_scores) + + scores = scores.type_as(fw_lprobs) + scores_buf = scores_buf.type_as(fw_lprobs) + + self.search.set_src_lengths(src_lengths_no_eos) + + if self.no_repeat_ngram_size > 0: + def calculate_banned_tokens(bbsz_idx): + # before decoding the next token, prevent decoding of ngrams that have already appeared + ngram_index = tuple(tokens[bbsz_idx, step + 2 - self.no_repeat_ngram_size:step + 1].tolist()) + return gen_ngrams[bbsz_idx].get(ngram_index, []) + + if step + 2 - self.no_repeat_ngram_size >= 0: + # no banned tokens if we haven't generated no_repeat_ngram_size tokens yet + banned_tokens = [calculate_banned_tokens(bbsz_idx) for bbsz_idx in range(bsz * beam_size)] + else: + banned_tokens = [[] for bbsz_idx in range(bsz * beam_size)] + + for bbsz_idx in range(bsz * beam_size): + fw_lprobs[bbsz_idx, banned_tokens[bbsz_idx]] = -math.inf + + combined_noisy_channel_scores, fw_lprobs_top_k, lm_lprobs_top_k, cand_indices, cand_beams = self.search.step( + step, + fw_lprobs.view(bsz, -1, self.vocab_size), + scores.view(bsz, beam_size, -1)[:, :, :step], ch_lm_lprobs.view(bsz, -1, self.vocab_size), + lm_lprobs.view(bsz, -1, self.vocab_size), self.combine_method + ) + + # cand_bbsz_idx contains beam indices for the top candidate + # hypotheses, with a range of values: [0, bsz*beam_size), + # and dimensions: [bsz, cand_size] + cand_bbsz_idx = cand_beams.add(bbsz_offsets) + + # finalize hypotheses that end in eos (except for candidates to be ignored) + eos_mask = cand_indices.eq(self.eos) + eos_mask[:, :beam_size] &= ~cands_to_ignore + + # only consider eos when it's among the top beam_size indices + eos_bbsz_idx = torch.masked_select( + cand_bbsz_idx[:, :beam_size], mask=eos_mask[:, :beam_size] + ) + + finalized_sents = set() + if eos_bbsz_idx.numel() > 0: + eos_scores = torch.masked_select( + fw_lprobs_top_k[:, :beam_size], mask=eos_mask[:, :beam_size] + ) + combined_noisy_channel_eos_scores = torch.masked_select( + combined_noisy_channel_scores[:, :beam_size], + mask=eos_mask[:, :beam_size], + ) + + # finalize hypo using channel model score + finalized_sents = finalize_hypos( + step, eos_bbsz_idx, eos_scores, combined_noisy_channel_eos_scores) + + num_remaining_sent -= len(finalized_sents) + + assert num_remaining_sent >= 0 + if num_remaining_sent == 0: + break + + if len(finalized_sents) > 0: + new_bsz = bsz - len(finalized_sents) + + # construct batch_idxs which holds indices of batches to keep for the next pass + batch_mask = cand_indices.new_ones(bsz) + batch_mask[cand_indices.new(finalized_sents)] = 0 + batch_idxs = torch.nonzero(batch_mask).squeeze(-1) + + eos_mask = eos_mask[batch_idxs] + cand_beams = cand_beams[batch_idxs] + bbsz_offsets.resize_(new_bsz, 1) + cand_bbsz_idx = cand_beams.add(bbsz_offsets) + + lm_lprobs_top_k = lm_lprobs_top_k[batch_idxs] + + fw_lprobs_top_k = fw_lprobs_top_k[batch_idxs] + cand_indices = cand_indices[batch_idxs] + if prefix_tokens is not None: + prefix_tokens = prefix_tokens[batch_idxs] + src_lengths_no_eos = src_lengths_no_eos[batch_idxs] + cands_to_ignore = cands_to_ignore[batch_idxs] + + scores = scores.view(bsz, -1)[batch_idxs].view(new_bsz * beam_size, -1) + scores_buf.resize_as_(scores) + tokens = tokens.view(bsz, -1)[batch_idxs].view(new_bsz * beam_size, -1) + tokens_buf.resize_as_(tokens) + src_tokens = src_tokens.view(bsz, -1)[batch_idxs].view(new_bsz * beam_size, -1) + src_lengths = src_lengths.view(bsz, -1)[batch_idxs].view(new_bsz * beam_size, -1) + lm_prefix_scores = lm_prefix_scores.view(bsz, -1)[batch_idxs].view(new_bsz * beam_size, -1).squeeze() + + if attn is not None: + attn = attn.view(bsz, -1)[batch_idxs].view(new_bsz * beam_size, attn.size(1), -1) + attn_buf.resize_as_(attn) + bsz = new_bsz + else: + batch_idxs = None + + # Set active_mask so that values > cand_size indicate eos or + # ignored hypos and values < cand_size indicate candidate + # active hypos. After this, the min values per row are the top + # candidate active hypos. + eos_mask[:, :beam_size] |= cands_to_ignore + active_mask = torch.add( + eos_mask.type_as(cand_offsets) * cand_size, + cand_offsets[: eos_mask.size(1)], + ) + + # get the top beam_size active hypotheses, which are just the hypos + # with the smallest values in active_mask + active_hypos, new_cands_to_ignore = buffer('active_hypos'), buffer('new_cands_to_ignore') + torch.topk( + active_mask, k=beam_size, dim=1, largest=False, + out=(new_cands_to_ignore, active_hypos) + ) + + # update cands_to_ignore to ignore any finalized hypos + cands_to_ignore = new_cands_to_ignore.ge(cand_size)[:, :beam_size] + assert (~cands_to_ignore).any(dim=1).all() + + active_bbsz_idx = buffer('active_bbsz_idx') + torch.gather( + cand_bbsz_idx, dim=1, index=active_hypos, + out=active_bbsz_idx, + ) + active_scores = torch.gather( + fw_lprobs_top_k, dim=1, index=active_hypos, + out=scores[:, step].view(bsz, beam_size), + ) + + active_bbsz_idx = active_bbsz_idx.view(-1) + active_scores = active_scores.view(-1) + + # copy tokens and scores for active hypotheses + torch.index_select( + tokens[:, :step + 1], dim=0, index=active_bbsz_idx, + out=tokens_buf[:, :step + 1], + ) + torch.gather( + cand_indices, dim=1, index=active_hypos, + out=tokens_buf.view(bsz, beam_size, -1)[:, :, step + 1], + ) + if step > 0: + torch.index_select( + scores[:, :step], dim=0, index=active_bbsz_idx, + out=scores_buf[:, :step], + ) + torch.gather( + fw_lprobs_top_k, dim=1, index=active_hypos, + out=scores_buf.view(bsz, beam_size, -1)[:, :, step], + ) + torch.gather( + lm_lprobs_top_k, dim=1, index=active_hypos, + out=lm_prefix_scores.view(bsz, beam_size) + ) + + # copy attention for active hypotheses + if attn is not None: + torch.index_select( + attn[:, :, :step + 2], dim=0, index=active_bbsz_idx, + out=attn_buf[:, :, :step + 2], + ) + + # swap buffers + tokens, tokens_buf = tokens_buf, tokens + scores, scores_buf = scores_buf, scores + if attn is not None: + attn, attn_buf = attn_buf, attn + + # reorder incremental state in decoder + reorder_state = active_bbsz_idx + + # sort by score descending + for sent in range(len(finalized)): + finalized[sent] = sorted(finalized[sent], key=lambda r: r['score'], reverse=True) + + return finalized + + +def get_lm_scores(model, input_tokens, incremental_states, cand_tokens, input_len, k): + with torch.no_grad(): + lm_lprobs, avg_attn_scores = model.forward_decoder( + input_tokens, encoder_outs=None, incremental_states=incremental_states, + ) + + lm_lprobs_size = lm_lprobs.size(0) + probs_next_wrd = torch.gather(lm_lprobs.repeat(1, k).view(lm_lprobs_size*k, -1), 1, cand_tokens).squeeze().view(-1) + + return probs_next_wrd + + +def make_dict2dict(old_dict, new_dict): + dict2dict_map = {} + for sym in old_dict.symbols: + dict2dict_map[old_dict.index(sym)] = new_dict.index(sym) + return dict2dict_map + + +def dict2dict(tokens, dict2dict_map): + if tokens.device == torch.device('cpu'): + tokens_tmp = tokens + else: + tokens_tmp = tokens.cpu() + return tokens_tmp.map_( + tokens_tmp, + lambda _, val, dict2dict_map=dict2dict_map : dict2dict_map[float(val)] + ).to(tokens.device) + + +def reorder_tokens(tokens, lengths, eos): + # reorder source tokens so they may be used as reference for P(S|T) + return torch.cat((tokens.new([eos]), tokens[-lengths:-1], tokens[:-lengths]), 0) + + +def reorder_all_tokens(tokens, lengths, eos): + # used to reorder src tokens from [ .. ] to [ ...] + # so source tokens can be used to predict P(S|T) + return torch.stack([reorder_tokens(token, length, eos) for token, length in zip(tokens, lengths)]) + + +def normalized_scores_with_batch_vocab( + model_decoder, features, target_ids, k, bsz, beam_size, + pad_idx, top_k=0, vocab_size_meter=None, start_idx=None, + end_idx=None, **kwargs): + """ + Get normalized probabilities (or log probs) from a net's output + w.r.t. vocab consisting of target IDs in the batch + """ + if model_decoder.adaptive_softmax is None: + weight = model_decoder.output_projection.weight + vocab_ids = torch.unique( + torch.cat( + (torch.unique(target_ids), torch.arange(top_k, device=target_ids.device)) + ) + ) + id_map = dict(zip(vocab_ids.tolist(), range(len(vocab_ids)))) + mapped_target_ids = target_ids.cpu().apply_( + lambda x, id_map=id_map: id_map[x] + ).to(target_ids.device) + expanded_target_ids = mapped_target_ids[:, :].repeat(1, k).view(bsz*beam_size*k, -1) + if start_idx is not None and end_idx is not None: + expanded_target_ids = expanded_target_ids[start_idx:end_idx, :] + logits = F.linear(features, weight[vocab_ids, :]) + log_softmax = F.log_softmax(logits, dim=-1, dtype=torch.float32) + intermed_scores = torch.gather( + log_softmax[:, :-1, :], + 2, + expanded_target_ids[:, 1:].unsqueeze(2), + ).squeeze() + not_padding = expanded_target_ids[:, 1:] != pad_idx + intermed_scores *= not_padding.float() + return intermed_scores + else: + raise ValueError("adaptive softmax doesn't work with " + + "`normalized_scores_with_batch_vocab()`") diff --git a/examples/fast_noisy_channel/noisy_channel_translation.py b/examples/fast_noisy_channel/noisy_channel_translation.py new file mode 100644 index 0000000000000000000000000000000000000000..b74bdfd456f9b7c546ce528173c77431b4f57ac1 --- /dev/null +++ b/examples/fast_noisy_channel/noisy_channel_translation.py @@ -0,0 +1,127 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq.tasks.translation import TranslationTask +from fairseq.tasks.language_modeling import LanguageModelingTask +from fairseq import checkpoint_utils +import argparse +from fairseq.tasks import register_task +import torch + + +@register_task("noisy_channel_translation") +class NoisyChannelTranslation(TranslationTask): + """ + Rescore the top k candidates from each beam using noisy channel modeling + """ + + @staticmethod + def add_args(parser): + """Add task-specific arguments to the parser.""" + TranslationTask.add_args(parser) + # fmt: off + parser.add_argument('--channel-model', metavar='FILE', + help='path to P(S|T) model. P(S|T) and P(T|S) must share source and target dictionaries.') + parser.add_argument('--combine-method', default='lm_only', + choices=['lm_only', 'noisy_channel'], + help="""method for combining direct and channel model scores. + lm_only: decode with P(T|S)P(T) + noisy_channel: decode with 1/t P(T|S) + 1/s(P(S|T)P(T))""") + parser.add_argument('--normalize-lm-scores-by-tgt-len', action='store_true', default=False, + help='normalize lm score by target length instead of source length') + parser.add_argument('--channel-scoring-type', default='log_norm', choices=['unnormalized', 'log_norm', 'k2_separate', 'src_vocab', 'src_vocab_batched'], + help="Normalize bw scores with log softmax or return bw scores without log softmax") + parser.add_argument('--top-k-vocab', default=0, type=int, + help='top k vocab IDs to use with `src_vocab` in channel model scoring') + parser.add_argument('--k2', default=50, type=int, + help='the top k2 candidates to rescore with the noisy channel model for each beam') + parser.add_argument('--ch-wt', default=1, type=float, + help='weight for the channel model') + parser.add_argument('--lm-model', metavar='FILE', + help='path to lm model file, to model P(T). P(T) must share the same vocab as the direct model on the target side') + parser.add_argument('--lm-data', metavar='FILE', + help='path to lm model training data for target language, used to properly load LM with correct dictionary') + parser.add_argument('--lm-wt', default=1, type=float, + help='the weight of the lm in joint decoding') + # fmt: on + + def build_generator( + self, models, args, seq_gen_cls=None, extra_gen_cls_kwargs=None + ): + if getattr(args, "score_reference", False): + raise NotImplementedError() + else: + from .noisy_channel_sequence_generator import NoisyChannelSequenceGenerator + use_cuda = torch.cuda.is_available() and not self.args.cpu + assert self.args.lm_model is not None, '--lm-model required for noisy channel generation!' + assert self.args.lm_data is not None, '--lm-data required for noisy channel generation to map between LM and bitext vocabs' + if self.args.channel_model is not None: + import copy + ch_args_task = copy.deepcopy(self.args) + tmp = ch_args_task.source_lang + ch_args_task.source_lang = ch_args_task.target_lang + ch_args_task.target_lang = tmp + ch_args_task._name = 'translation' + channel_task = TranslationTask.setup_task(ch_args_task) + + arg_dict = {} + arg_dict['task'] = 'language_modeling' + arg_dict['sample_break_mode'] = 'eos' + arg_dict['data'] = self.args.lm_data + arg_dict['output_dictionary_size'] = -1 + lm_args = argparse.Namespace(**arg_dict) + lm_task = LanguageModelingTask.setup_task(lm_args) + lm_dict = lm_task.output_dictionary + + if self.args.channel_model is not None: + channel_models, _ = checkpoint_utils.load_model_ensemble(self.args.channel_model.split(':'), task=channel_task) + + for model in channel_models: + model.make_generation_fast_( + beamable_mm_beam_size=None if args.no_beamable_mm else args.beam, + need_attn=args.print_alignment, + ) + if self.args.fp16: + model.half() + if use_cuda: + model.cuda() + else: + channel_models = None + + lm_models, _ = checkpoint_utils.load_model_ensemble(self.args.lm_model.split(':'), task=lm_task) + + for model in lm_models: + model.make_generation_fast_( + beamable_mm_beam_size=None if args.no_beamable_mm else args.beam, + need_attn=args.print_alignment, + ) + if self.args.fp16: + model.half() + if use_cuda: + model.cuda() + return NoisyChannelSequenceGenerator( + combine_method=self.args.combine_method, + tgt_dict=self.target_dictionary, + src_dict=self.source_dictionary, + beam_size=getattr(args, 'beam', 5), + max_len_a=getattr(args, 'max_len_a', 0), + max_len_b=getattr(args, 'max_len_b', 200), + min_len=getattr(args, 'min_len', 1), + len_penalty=getattr(args, 'lenpen', 1), + unk_penalty=getattr(args, 'unkpen', 0), + temperature=getattr(args, 'temperature', 1.), + match_source_len=getattr(args, 'match_source_len', False), + no_repeat_ngram_size=getattr(args, 'no_repeat_ngram_size', 0), + normalize_scores=(not getattr(args, 'unnormalized', False)), + channel_models=channel_models, + k2=getattr(self.args, 'k2', 50), + ch_weight=getattr(self.args, 'ch_wt', 1), + channel_scoring_type=self.args.channel_scoring_type, + top_k_vocab=self.args.top_k_vocab, + lm_models=lm_models, + lm_dict=lm_dict, + lm_weight=getattr(self.args, 'lm_wt', 1), + normalize_lm_scores_by_tgt_len=getattr(self.args, 'normalize_lm_scores_by_tgt_len', False), + ) diff --git a/examples/flores101/README.md b/examples/flores101/README.md new file mode 100644 index 0000000000000000000000000000000000000000..635c13f40bd0ccab704735bc5c26ea0192ea98cd --- /dev/null +++ b/examples/flores101/README.md @@ -0,0 +1,223 @@ +

+ +

+ +# Flores101: Large-Scale Multilingual Machine Translation + +## Introduction + +Baseline pretrained models for small and large tracks of WMT 21 Large-Scale Multilingual Machine Translation competition. + +Flores Task at WMT 21: http://www.statmt.org/wmt21/large-scale-multilingual-translation-task.html + +Flores announement blog post: https://ai.facebook.com/blog/flores-researchers-kick-off-multilingual-translation-challenge-at-wmt-and-call-for-compute-grants/ + + + +## Pretrained models + +Model | Num layers | Embed dimension | FFN dimension| Vocab Size | #params | Download +---|---|---|---|---|---|--- +`flores101_mm100_615M` | 12 | 1024 | 4096 | 256,000 | 615M | https://dl.fbaipublicfiles.com/flores101/pretrained_models/flores101_mm100_615M.tar.gz +`flores101_mm100_175M` | 6 | 512 | 2048 | 256,000 | 175M | https://dl.fbaipublicfiles.com/flores101/pretrained_models/flores101_mm100_175M.tar.gz + + +These models are trained similar to [M2M-100](https://arxiv.org/abs/2010.11125) with additional support for the languages that are part of the WMT Large-Scale Multilingual Machine Translation track. Full list of languages can be found at the bottom. + + +## Example Generation code + +### Download model, sentencepiece vocab + +```bash +fairseq=/path/to/fairseq +cd $fairseq + +# Download 615M param model. +wget https://dl.fbaipublicfiles.com/flores101/pretrained_models/flores101_mm100_615M.tar.gz + +# Extract +tar -xvzf flores101_mm100_615M.tar.gz +``` + +### Encode using our SentencePiece Model +Note: Install SentencePiece from [here](https://github.com/google/sentencepiece) + + +```bash +fairseq=/path/to/fairseq +cd $fairseq + +# Download example dataset From German to French +sacrebleu --echo src -l de-fr -t wmt19 | head -n 20 > raw_input.de-fr.de +sacrebleu --echo ref -l de-fr -t wmt19 | head -n 20 > raw_input.de-fr.fr + +for lang in de fr ; do + python scripts/spm_encode.py \ + --model flores101_mm100_615M/sentencepiece.bpe.model \ + --output_format=piece \ + --inputs=raw_input.de-fr.${lang} \ + --outputs=spm.de-fr.${lang} +done +``` + +### Binarization + +```bash +fairseq-preprocess \ + --source-lang de --target-lang fr \ + --testpref spm.de-fr \ + --thresholdsrc 0 --thresholdtgt 0 \ + --destdir data_bin \ + --srcdict flores101_mm100_615M/dict.txt --tgtdict flores101_mm100_615M/dict.txt +``` + +### Generation + + +```bash +fairseq-generate \ + data_bin \ + --batch-size 1 \ + --path flores101_mm100_615M/model.pt \ + --fixed-dictionary flores101_mm100_615M/dict.txt \ + -s de -t fr \ + --remove-bpe 'sentencepiece' \ + --beam 5 \ + --task translation_multi_simple_epoch \ + --lang-pairs flores101_mm100_615M/language_pairs.txt \ + --decoder-langtok --encoder-langtok src \ + --gen-subset test \ + --fp16 \ + --dataset-impl mmap \ + --distributed-world-size 1 --distributed-no-spawn +``` + +### Supported Languages and lang code + +Language | lang code +---|--- +Akrikaans | af +Amharic | am +Arabic | ar +Assamese | as +Asturian | ast +Aymara | ay +Azerbaijani | az +Bashkir | ba +Belarusian | be +Bulgarian | bg +Bengali | bn +Breton | br +Bosnian | bs +Catalan | ca +Cebuano | ceb +Chokwe | cjk +Czech | cs +Welsh | cy +Danish | da +German | de +Dyula| dyu +Greek | el +English | en +Spanish | es +Estonian | et +Persian | fa +Fulah | ff +Finnish | fi +French | fr +Western Frisian | fy +Irish | ga +Scottish Gaelic | gd +Galician | gl +Gujarati | gu +Hausa | ha +Hebrew | he +Hindi | hi +Croatian | hr +Haitian Creole | ht +Hungarian | hu +Armenian | hy +Indonesian | id +Igbo | ig +Iloko | ilo +Icelandic | is +Italian | it +Japanese | ja +Javanese | jv +Georgian | ka +Kachin | kac +Kamba | kam +Kabuverdianu | kea +Kongo | kg +Kazakh | kk +Central Khmer | km +Kimbundu | kmb +Northern Kurdish | kmr +Kannada | kn +Korean | ko +Kurdish | ku +Kyrgyz | ky +Luxembourgish | lb +Ganda | lg +Lingala | ln +Lao | lo +Lithuanian | lt +Luo | luo +Latvian | lv +Malagasy | mg +Maori | mi +Macedonian | mk +Malayalam | ml +Mongolian | mn +Marathi | mr +Malay | ms +Maltese | mt +Burmese | my +Nepali | ne +Dutch | nl +Norwegian | no +Northern Sotho | ns +Nyanja | ny +Occitan | oc +Oromo | om +Oriya | or +Punjabi | pa +Polish | pl +Pashto | ps +Portuguese | pt +Quechua | qu +Romanian | ro +Russian | ru +Sindhi | sd +Shan | shn +Sinhala | si +Slovak | sk +Slovenian | sl +Shona | sn +Somali | so +Albanian | sq +Serbian | sr +Swati | ss +Sundanese | su +Swedish | sv +Swahili | sw +Tamil | ta +Telugu | te +Tajik | tg +Thai | th +Tigrinya | ti +Tagalog | tl +Tswana | tn +Turkish | tr +Ukrainian | uk +Umbundu | umb +Urdu | ur +Uzbek | uz +Vietnamese | vi +Wolof | wo +Xhosa | xh +Yiddish | yi +Yoruba | yo +Chinese| zh +Zulu | zu diff --git a/examples/flores101/flores_logo.png b/examples/flores101/flores_logo.png new file mode 100644 index 0000000000000000000000000000000000000000..d4d1455c6eab608ff5317ce885183cd213564273 Binary files /dev/null and b/examples/flores101/flores_logo.png differ diff --git a/examples/fully_sharded_data_parallel/README.md b/examples/fully_sharded_data_parallel/README.md new file mode 100644 index 0000000000000000000000000000000000000000..d620f0e4f1a1561c267140b9b6f4c705a38a8865 --- /dev/null +++ b/examples/fully_sharded_data_parallel/README.md @@ -0,0 +1,177 @@ +# Fully Sharded Data Parallel (FSDP) + +## Overview +Recent work by [Microsoft](https://arxiv.org/abs/1910.02054) and +[Google](https://arxiv.org/abs/2004.13336) has shown that data parallel +training can be made significantly more efficient by sharding the model +parameters and optimizer state across data parallel workers. These ideas are +encapsulated in the new **`FullyShardedDataParallel` (FSDP)** wrapper provided +by [fairscale](https://github.com/facebookresearch/fairscale/). + +Compared to PyTorch DDP: +* FSDP produces identical results as PyTorch DDP (it's still synchronous data parallel training) +* FSDP shards parameters (FP16 + FP32) and optimizer state across data parallel GPUs +* FSDP is faster than PyTorch DDP because the optimizer step is sharded, and the communication can be overlapped with the forward pass +* FSDP enables training 13B parameter models on 8 GPUs and 175B parameter models on 128 GPUs + +FSDP is fully supported in fairseq via the following new arguments: +* `--ddp-backend=fully_sharded`: enables full sharding via FSDP +* `--cpu-offload`: offloads the optimizer state and FP32 model copy to CPU (combine with `--optimizer=cpu_adam`) +* `--no-reshard-after-forward`: increases training speed for large models (1B+ params) and is similar to ZeRO stage 2 +* other popular options (`--fp16`, `--update-freq`, `--checkpoint-activations`, `--offload-activations`, etc.) continue to work as normal + +
Limitations

+ +FSDP currently has several limitations compared to fairseq's default DDP backend (PyTorch DDP): +* while FSDP is full compatible with pointwise Optimizers (e.g., Adam, AdamW, Adadelta, Adamax, SGD, etc.), it is not currently compatible with non-pointwise Optimizers (e.g., Adagrad, Adafactor, LAMB, etc.) +* FSDP depends on flattening the parameters, so models that currently require `--fp16-no-flatten-grads` may not be supported + +See the [fairscale docs](https://fairscale.readthedocs.io/en/latest/api/nn/fsdp_tips.html) for a more detailed +explanation of these and other limitations. + +

+ +
How it works

+ +Fully Sharded Data Parallel + +See the [fairscale docs](https://fairscale.readthedocs.io/en/latest/api/nn/fsdp_tips.html) for a more detailed +explanation of how FSDP works. + +

+ +## Example usage + +The following examples illustrate how to train a very large language model with +13 billion parameters on 1 GPU by offloading parameters and optimizer states to +CPU, or on 8 GPUs by fully sharding the params and optimizer states across GPUs. + +These examples use the WikiText-103 dataset for demonstration purposes, but +in practice a much larger dataset will be needed to achieve good results. +Follow the [instructions here](https://github.com/pytorch/fairseq/blob/master/examples/roberta/README.pretraining.md#1-preprocess-the-data) +to preprocess the WikiText-103 dataset using the GPT-2/RoBERTa vocabulary. + +### 13B params on 1 V100 GPU (with CPU offloading) + +The following command trains a 13B parameter GPT-3 model on a single V100 GPU +using the `--cpu-offload` feature to offload parameters and optimizer states to +CPU. In this setting, the optimizer step (Adam) happens on CPU. We also use the +`--checkpoint-activations` feature (sometimes called [gradient checkpointing](https://pytorch.org/docs/stable/checkpoint.html)), +which further saves memory in exchange for a small increase in computation. + +**Requirements:** +- Install the latest master version of fairscale: `pip install git+https://github.com/facebookresearch/fairscale.git@master` +- You'll need 32GB of GPU memory and ~256GB of system memory to train the 13B param model. +- If you have less system memory, the 6.7B param model can be trained with ~128GB of system memory, just set `--arch transformer_lm_gpt3_6_7` +- We use the CPU Adam optimizer from [DeepSpeed](https://github.com/microsoft/DeepSpeed), so you'll need to `pip install deepspeed` before running the command. + +**Notes:** +- The command will take ~5 minutes to start training, during which time it will appear to be hung, since randomly initializing 13B weights can be slow. +- The `--cpu-offload` feature requires training in mixed precision (`--fp16`). +- Tune the `OMP_NUM_THREADS` env variable for best performance with CPU offloading. +- The example command below stops training after 10 steps (`--max-update 10`) and does not save checkpoints (`--no-save`). + +```bash +OMP_NUM_THREADS=20 CUDA_VISIBLE_DEVICES=0 \ + fairseq-train data-bin/wikitext-103-roberta-bpe-bin \ + --ddp-backend fully_sharded --fp16 --fp16-init-scale 4 \ + --cpu-offload --checkpoint-activations \ + --task language_modeling --tokens-per-sample 2048 --batch-size 8 \ + --arch transformer_lm_gpt3_13 \ + --optimizer cpu_adam --adam-betas "(0.9,0.98)" \ + --lr 0.0001 --lr-scheduler polynomial_decay --warmup-updates 5 --total-num-update 10 \ + --max-update 10 --no-save --log-format json --log-interval 1 +``` + +
Example output

+ +``` +(...) +2021-03-08 12:29:51 | INFO | fairseq_cli.train | num. model params: 13,110,865,920 (num. trained: 13,110,865,920) +(...) +2021-03-08 12:29:51 | INFO | fairseq_cli.train | training on 1 devices (GPUs/TPUs) +2021-03-08 12:29:51 | INFO | fairseq_cli.train | max tokens per GPU = None and batch size per GPU = 8 +(...) +Adam Optimizer #0 is created with AVX2 arithmetic capability. +Config: alpha=0.000100, betas=(0.900000, 0.980000), weight_decay=0.000000, adam_w=1 +(...) +2021-03-08 12:31:36 | INFO | train_inner | {"epoch": 1, "update": 0.0, "loss": "16.475", "ppl": "91120.8", "wps": "0", "ups": "0", "wpb": "16384", "bsz": "8", "num_updates": "1", "lr": "2e-05", "gnorm": "20.751", "loss_scale": "4", "train_wall": "99", "gb_free": "9.3", "wall": "105"} +2021-03-08 12:32:33 | INFO | train_inner | {"epoch": 1, "update": 0.0, "loss": "16.446", "ppl": "89281.6", "wps": "288.7", "ups": "0.02", "wpb": "16384", "bsz": "8", "num_updates": "2", "lr": "4e-05", "gnorm": "19.777", "loss_scale": "4", "train_wall": "57", "gb_free": "9.3", "wall": "161"} +2021-03-08 12:33:12 | INFO | fairseq.trainer | NOTE: gradient overflow detected, ignoring gradient, setting loss scale to: 2.0 +2021-03-08 12:33:51 | INFO | fairseq.trainer | NOTE: gradient overflow detected, ignoring gradient, setting loss scale to: 1.0 +2021-03-08 12:34:45 | INFO | train_inner | {"epoch": 1, "update": 0.001, "loss": "25.22", "ppl": "3.90691e+07", "wps": "123.4", "ups": "0.01", "wpb": "16384", "bsz": "8", "num_updates": "3", "lr": "6e-05", "gnorm": "131.281", "loss_scale": "1", "train_wall": "133", "gb_free": "9.3", "wall": "294"} +2021-03-08 12:35:43 | INFO | train_inner | {"epoch": 1, "update": 0.001, "loss": "18.079", "ppl": "276809", "wps": "285.5", "ups": "0.02", "wpb": "16384", "bsz": "8", "num_updates": "4", "lr": "8e-05", "gnorm": "13.776", "loss_scale": "1", "train_wall": "57", "gb_free": "9.3", "wall": "351"} +2021-03-08 12:36:35 | INFO | train_inner | {"epoch": 1, "update": 0.001, "loss": "23.729", "ppl": "1.39088e+07", "wps": "316.7", "ups": "0.02", "wpb": "16384", "bsz": "8", "num_updates": "5", "lr": "0.0001", "gnorm": "72.774", "loss_scale": "1", "train_wall": "52", "gb_free": "9.3", "wall": "403"} +2021-03-08 12:37:28 | INFO | train_inner | {"epoch": 1, "update": 0.001, "loss": "20.429", "ppl": "1.41203e+06", "wps": "307.6", "ups": "0.02", "wpb": "16384", "bsz": "8", "num_updates": "6", "lr": "8e-05", "gnorm": "60.846", "loss_scale": "1", "train_wall": "53", "gb_free": "9.3", "wall": "456"} +2021-03-08 12:38:27 | INFO | train_inner | {"epoch": 1, "update": 0.001, "loss": "18.965", "ppl": "511684", "wps": "279.4", "ups": "0.02", "wpb": "16384", "bsz": "8", "num_updates": "7", "lr": "6e-05", "gnorm": "22.687", "loss_scale": "1", "train_wall": "59", "gb_free": "9.3", "wall": "515"} +2021-03-08 12:39:18 | INFO | train_inner | {"epoch": 1, "update": 0.001, "loss": "18.345", "ppl": "332887", "wps": "319.1", "ups": "0.02", "wpb": "16384", "bsz": "8", "num_updates": "8", "lr": "4e-05", "gnorm": "8.451", "loss_scale": "1", "train_wall": "51", "gb_free": "9.3", "wall": "566"} +2021-03-08 12:40:11 | INFO | train_inner | {"epoch": 1, "update": 0.002, "loss": "18.262", "ppl": "314336", "wps": "305.9", "ups": "0.02", "wpb": "16384", "bsz": "8", "num_updates": "9", "lr": "2e-05", "gnorm": "6.457", "loss_scale": "1", "train_wall": "54", "gb_free": "9.3", "wall": "620"} +2021-03-08 12:41:04 | INFO | train_inner | {"epoch": 1, "update": 0.002, "loss": "17.556", "ppl": "192686", "wps": "311.8", "ups": "0.02", "wpb": "16384", "bsz": "8", "num_updates": "10", "lr": "0", "gnorm": "5.796", "loss_scale": "1", "train_wall": "53", "gb_free": "9.3", "wall": "673"} +2021-03-08 12:41:04 | INFO | fairseq_cli.train | Stopping training due to num_updates: 10 >= max_update: 10 +2021-03-08 12:41:04 | INFO | fairseq_cli.train | begin validation on "valid" subset +2021-03-08 12:43:15 | INFO | valid | {"epoch": 1, "valid_loss": "17.953", "valid_ppl": "253807", "valid_wps": "1868.4", "valid_wpb": "15400.2", "valid_bsz": "7.6", "valid_num_updates": "10"} +2021-03-08 12:43:15 | INFO | fairseq_cli.train | end of epoch 1 (average epoch stats below) +2021-03-08 12:43:15 | INFO | train | {"epoch": 1, "train_loss": "19.351", "train_ppl": "668509", "train_wps": "210.9", "train_ups": "0.01", "train_wpb": "16384", "train_bsz": "8", "train_num_updates": "10", "train_lr": "0", "train_gnorm": "36.26", "train_loss_scale": "1", "train_train_wall": "667", "train_gb_free": "9.3", "train_wall": "804"} +2021-03-08 12:43:15 | INFO | fairseq_cli.train | done training in 798.6 seconds +``` + +

+ +### 13B params on 8 V100 GPUs (with full parameter + optimizer state sharding) + +FSDP can also shard the parameters and optimizer states across multiple GPUs, +reducing memory requirements significantly. On 8 x 32GB GPUs, sharding enables +training the same 13B parameter model *without offloading the parameters to +CPU*. However, without CPU offloading we'd only be able to fit a batch size of +1 per GPU, which would cause training speed to suffer. + +We obtain the best performance on 8 GPUs by combining full sharding and CPU +offloading. The following command trains the same 13B parameter GPT-3 model as +before on 8 x 32GB V100 GPUs; training speed increases superlinearly from ~310 +words per second to ~3200 words per second. + +```bash +OMP_NUM_THREADS=20 CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \ + fairseq-train data-bin/wikitext-103-roberta-bpe-bin \ + --ddp-backend fully_sharded --fp16 --fp16-init-scale 4 \ + --cpu-offload --checkpoint-activations \ + --task language_modeling --tokens-per-sample 2048 --batch-size 8 \ + --arch transformer_lm_gpt3_13 \ + --optimizer cpu_adam --adam-betas "(0.9,0.98)" \ + --lr 0.0001 --lr-scheduler polynomial_decay --warmup-updates 5 --total-num-update 10 \ + --max-update 10 --no-save --log-format json --log-interval 1 +``` + +
Example output

+ +``` +(...) +2021-03-08 18:04:09 | INFO | fairseq_cli.train | num. model params: 13,110,865,920 (num. trained: 13,110,865,920) +(...) +2021-03-08 18:04:09 | INFO | fairseq_cli.train | training on 8 devices (GPUs/TPUs) +2021-03-08 18:04:09 | INFO | fairseq_cli.train | max tokens per GPU = None and batch size per GPU = 8 +(...) +Adam Optimizer #0 is created with AVX2 arithmetic capability. +Config: alpha=0.000100, betas=(0.900000, 0.980000), weight_decay=0.000000, adam_w=1 +(...) +2021-03-08 18:05:06 | INFO | train_inner | {"epoch": 1, "update": 0.001, "loss": "16.408", "ppl": "86945.6", "wps": "0", "ups": "0", "wpb": "131072", "bsz": "64", "num_updates": "1", "lr": "2e-05", "gnorm": "18.27", "loss_scale": "4", "train_wall": "47", "gb_free": "9.3", "wall": "56"} +2021-03-08 18:05:45 | INFO | train_inner | {"epoch": 1, "update": 0.002, "loss": "16.352", "ppl": "83644.3", "wps": "3283.4", "ups": "0.03", "wpb": "131072", "bsz": "64", "num_updates": "2", "lr": "4e-05", "gnorm": "18.411", "loss_scale": "4", "train_wall": "40", "gb_free": "9.3", "wall": "96"} +2021-03-08 18:06:21 | INFO | fairseq.trainer | NOTE: gradient overflow detected, ignoring gradient, setting loss scale to: 2.0 +2021-03-08 18:06:56 | INFO | fairseq.trainer | NOTE: gradient overflow detected, ignoring gradient, setting loss scale to: 1.0 +2021-03-08 18:07:37 | INFO | train_inner | {"epoch": 1, "update": 0.006, "loss": "23.682", "ppl": "1.34537e+07", "wps": "1176.6", "ups": "0.01", "wpb": "131072", "bsz": "64", "num_updates": "3", "lr": "6e-05", "gnorm": "119.682", "loss_scale": "1", "train_wall": "111", "gb_free": "9.3", "wall": "208"} +2021-03-08 18:08:18 | INFO | train_inner | {"epoch": 1, "update": 0.007, "loss": "18.988", "ppl": "519921", "wps": "3189.1", "ups": "0.02", "wpb": "131072", "bsz": "64", "num_updates": "4", "lr": "8e-05", "gnorm": "14.934", "loss_scale": "1", "train_wall": "41", "gb_free": "9.3", "wall": "249"} +2021-03-08 18:08:59 | INFO | train_inner | {"epoch": 1, "update": 0.008, "loss": "20.08", "ppl": "1.10798e+06", "wps": "3223.1", "ups": "0.02", "wpb": "131072", "bsz": "64", "num_updates": "5", "lr": "0.0001", "gnorm": "59.92", "loss_scale": "1", "train_wall": "41", "gb_free": "9.3", "wall": "289"} +2021-03-08 18:09:39 | INFO | train_inner | {"epoch": 1, "update": 0.009, "loss": "18.323", "ppl": "327980", "wps": "3256.6", "ups": "0.02", "wpb": "131072", "bsz": "64", "num_updates": "6", "lr": "8e-05", "gnorm": "37.425", "loss_scale": "1", "train_wall": "40", "gb_free": "9.3", "wall": "330"} +2021-03-08 18:10:20 | INFO | train_inner | {"epoch": 1, "update": 0.01, "loss": "17.264", "ppl": "157354", "wps": "3188.7", "ups": "0.02", "wpb": "131072", "bsz": "64", "num_updates": "7", "lr": "6e-05", "gnorm": "10.824", "loss_scale": "1", "train_wall": "41", "gb_free": "9.3", "wall": "371"} +2021-03-08 18:11:01 | INFO | train_inner | {"epoch": 1, "update": 0.011, "loss": "16.794", "ppl": "113647", "wps": "3230", "ups": "0.02", "wpb": "131072", "bsz": "64", "num_updates": "8", "lr": "4e-05", "gnorm": "5.616", "loss_scale": "1", "train_wall": "41", "gb_free": "9.3", "wall": "411"} +2021-03-08 18:11:39 | INFO | train_inner | {"epoch": 1, "update": 0.012, "loss": "16.706", "ppl": "106938", "wps": "3384", "ups": "0.03", "wpb": "131072", "bsz": "64", "num_updates": "9", "lr": "2e-05", "gnorm": "5.318", "loss_scale": "1", "train_wall": "39", "gb_free": "9.3", "wall": "450"} +2021-03-08 18:12:19 | INFO | train_inner | {"epoch": 1, "update": 0.013, "loss": "16.548", "ppl": "95796.2", "wps": "3274.4", "ups": "0.02", "wpb": "131072", "bsz": "64", "num_updates": "10", "lr": "0", "gnorm": "5.22", "loss_scale": "1", "train_wall": "40", "gb_free": "9.3", "wall": "490"} +2021-03-08 18:12:19 | INFO | fairseq_cli.train | Stopping training due to num_updates: 10 >= max_update: 10 +2021-03-08 18:12:19 | INFO | fairseq_cli.train | begin validation on "valid" subset +2021-03-08 18:12:45 | INFO | valid | {"epoch": 1, "valid_loss": "16.624", "valid_ppl": "101000", "valid_wps": "10855.9", "valid_wpb": "123202", "valid_bsz": "60.5", "valid_num_updates": "10"} +2021-03-08 18:12:45 | INFO | fairseq_cli.train | end of epoch 1 (average epoch stats below) +2021-03-08 18:12:45 | INFO | train | {"epoch": 1, "train_loss": "18.114", "train_ppl": "283776", "train_wps": "2567.8", "train_ups": "0.02", "train_wpb": "131072", "train_bsz": "64", "train_num_updates": "10", "train_lr": "0", "train_gnorm": "29.562", "train_loss_scale": "1", "train_train_wall": "480", "train_gb_free": "9.3", "train_wall": "516"} +2021-03-08 18:12:45 | INFO | fairseq_cli.train | done training in 509.9 seconds +``` + +

diff --git a/examples/gottbert/README.md b/examples/gottbert/README.md new file mode 100644 index 0000000000000000000000000000000000000000..1d58feb279a4a50222290546c3bb285d3cea98e6 --- /dev/null +++ b/examples/gottbert/README.md @@ -0,0 +1,64 @@ +# GottBERT: a pure German language model + +## Introduction + +[GottBERT](http://arxiv.org/abs/2012.02110) is a pretrained language model trained on 145GB of German text based on RoBERTa. + +## Example usage + +### fairseq +##### Load GottBERT from torch.hub (PyTorch >= 1.1): +```python +import torch +gottbert = torch.hub.load('pytorch/fairseq', 'gottbert-base') +gottbert.eval() # disable dropout (or leave in train mode to finetune) +``` + +##### Load GottBERT (for PyTorch 1.0 or custom models): +```python +# Download gottbert model +wget https://dl.gottbert.de/fairseq/models/gottbert-base.tar.gz +tar -xzvf gottbert.tar.gz + +# Load the model in fairseq +from fairseq.models.roberta import GottbertModel +gottbert = GottbertModel.from_pretrained('/path/to/gottbert') +gottbert.eval() # disable dropout (or leave in train mode to finetune) +``` + +##### Filling masks: +```python +masked_line = 'Gott ist ! :)' +gottbert.fill_mask(masked_line, topk=3) +# [('Gott ist gut ! :)', 0.3642110526561737, ' gut'), +# ('Gott ist überall ! :)', 0.06009674072265625, ' überall'), +# ('Gott ist großartig ! :)', 0.0370681993663311, ' großartig')] +``` + +##### Extract features from GottBERT + +```python +# Extract the last layer's features +line = "Der erste Schluck aus dem Becher der Naturwissenschaft macht atheistisch , aber auf dem Grunde des Bechers wartet Gott !" +tokens = gottbert.encode(line) +last_layer_features = gottbert.extract_features(tokens) +assert last_layer_features.size() == torch.Size([1, 27, 768]) + +# Extract all layer's features (layer 0 is the embedding layer) +all_layers = gottbert.extract_features(tokens, return_all_hiddens=True) +assert len(all_layers) == 13 +assert torch.all(all_layers[-1] == last_layer_features) +``` +## Citation +If you use our work, please cite: + +```bibtex +@misc{scheible2020gottbert, + title={GottBERT: a pure German Language Model}, + author={Raphael Scheible and Fabian Thomczyk and Patric Tippmann and Victor Jaravine and Martin Boeker}, + year={2020}, + eprint={2012.02110}, + archivePrefix={arXiv}, + primaryClass={cs.CL} +} +``` diff --git a/examples/hubert/README.md b/examples/hubert/README.md new file mode 100644 index 0000000000000000000000000000000000000000..3254b754f0272d3bc02a94ee8c33341f7d4a4bdf --- /dev/null +++ b/examples/hubert/README.md @@ -0,0 +1,116 @@ +# HuBERT + +## Pre-trained and fine-tuned (ASR) models +Model | Pretraining Data | Finetuning Dataset | Model +|---|---|---|--- +HuBERT Base (~95M params) | [Librispeech](http://www.openslr.org/12) 960 hr | No finetuning (Pretrained Model) | [download](https://dl.fbaipublicfiles.com/hubert/hubert_base_ls960.pt) +HuBERT Large (~316M params) | [Libri-Light](https://github.com/facebookresearch/libri-light) 60k hr | No finetuning (Pretrained Model) | [download](https://dl.fbaipublicfiles.com/hubert/hubert_large_ll60k.pt) +HuBERT Extra Large (~1B params) | [Libri-Light](https://github.com/facebookresearch/libri-light) 60k hr | No finetuning (Pretrained Model) | [download](https://dl.fbaipublicfiles.com/hubert/hubert_xtralarge_ll60k.pt) +HuBERT Large | [Libri-Light](https://github.com/facebookresearch/libri-light) 60k hr | [Librispeech](http://www.openslr.org/12) 960 hr | [download](https://dl.fbaipublicfiles.com/hubert/hubert_large_ll60k_finetune_ls960.pt) +HuBERT Extra Large | [Libri-Light](https://github.com/facebookresearch/libri-light) 60k hr | [Librispeech](http://www.openslr.org/12) 960 hr | [download](https://dl.fbaipublicfiles.com/hubert/hubert_xtralarge_ll60k_finetune_ls960.pt) + +## Load a pretrained model +``` +ckpt_path = "/path/to/the/checkpoint.pt" +models, cfg, task = fairseq.checkpoint_utils.load_model_ensemble_and_task([ckpt_path], strict=False) +model = models[0] +``` +** We will follow-up with a patch such that you wouldn't need to pass `strict=False` for loading the checkpoint in future. + +## Train a new model + +### Data preparation + +Follow the steps in `./simple_kmeans` to create: +- `{train,valid}.tsv` waveform list files +- `{train,valid}.km` frame-aligned pseudo label files. +The `label_rate` is the same as the feature frame rate used for clustering, +which is 100Hz for MFCC features and 50Hz for HuBERT features by default. + +### Pre-train a HuBERT model + +Suppose `{train,valid}.tsv` are saved at `/path/to/data`, `{train,valid}.km` +are saved at `/path/to/labels`, and the label rate is 100Hz. + +To train a base model (12 layer transformer), run: +```sh +$ python fairseq_cli/hydra_train.py \ + --config-dir /path/to/fairseq-py/examples/hubert/config/pretrain \ + --config-name hubert_base_librispeech \ + task.data=/path/to/data task.label_dir=/path/to/labels model.label_rate=100 +``` + +### Fine-tune a HuBERT model with a CTC loss + +Suppose `{train,valid}.tsv` are saved at `/path/to/data`, and their +corresponding character transcripts `{train,valid}.ltr` are saved at +`/path/to/trans`. + +To fine-tune a pre-trained HuBERT model at `/path/to/checkpoint`, run +```sh +$ python fairseq_cli/hydra_train.py \ + --config-dir /path/to/fairseq-py/examples/hubert/config/finetune \ + --config-name base_10h \ + task.data=/path/to/data task.label_dir=/path/to/trans \ + model.w2v_path=/path/to/checkpoint +``` + +### Decode a HuBERT model + +Suppose the `test.tsv` and `test.ltr` are the waveform list and transcripts of +the split to be decoded, saved at `/path/to/data`, and the fine-tuned model is +saved at `/path/to/checkpoint`. We support three decoding modes: +- Viterbi decoding: greedy decoding without a language model +- KenLM decoding: decoding with an arpa-format KenLM n-gram language model +- Fairseq-LM deocding: decoding with a Fairseq neural language model + + +#### Viterbi decoding + +`task.normalize` needs to be consistent with the value used during fine-tuning. +Decoding results will be saved at +`/path/to/experiment/directory/decode/viterbi/test`. + +```sh +$ python examples/speech_recognition/new/infer.py \ + --config-dir /path/to/fairseq-py/examples/hubert/config/decode \ + --config-name infer_viterbi \ + task.data=/path/to/data \ + task.normalize=[true|false] \ + decoding.exp_dir=/path/to/experiment/directory \ + common_eval.path=/path/to/checkpoint + dataset.gen_subset=test \ +``` + +#### KenLM / Fairseq-LM decoding + +Suppose the pronunciation lexicon and the n-gram LM are saved at +`/path/to/lexicon` and `/path/to/arpa`, respectively. Decoding results will be +saved at `/path/to/experiment/directory/decode/kenlm/test`. + +```sh +$ python examples/speech_recognition/new/infer.py \ + --config-dir /path/to/fairseq-py/examples/hubert/config/decode \ + --config-name infer_kenlm \ + task.data=/path/to/data \ + task.normalize=[true|false] \ + decoding.exp_dir=/path/to/experiment/directory \ + common_eval.path=/path/to/checkpoint + dataset.gen_subset=test \ + decoding.decoder.lexicon=/path/to/lexicon \ + decoding.decoder.lmpath=/path/to/arpa +``` + +The command above uses the default decoding hyperparameter, which can be found +in `examples/speech_recognition/hydra/decoder.py`. These parameters can be +configured from the command line. For example, to search with a beam size of +500, we can append the command above with `decoding.decoder.beam=500`. +Important parameters include: +- decoding.decoder.beam +- decoding.decoder.beamthreshold +- decoding.decoder.lmweight +- decoding.decoder.wordscore +- decoding.decoder.silweight + +To decode with a Fairseq LM, use `--config-name infer_fsqlm` instead, and +change the path of lexicon and LM accordingly. diff --git a/examples/hubert/config/decode/ax_sweep/ngram.yaml b/examples/hubert/config/decode/ax_sweep/ngram.yaml new file mode 100644 index 0000000000000000000000000000000000000000..5a02df1f7da7eebfebe4018ef2758a716fbab646 --- /dev/null +++ b/examples/hubert/config/decode/ax_sweep/ngram.yaml @@ -0,0 +1,33 @@ +# @package _global_ + +common_eval: + results_path: ${decoding.exp_dir}/decode/${decoding.decoder.name}_ax/${dataset.gen_subset} + +hydra: + sweeper: + ax_config: + max_trials: 60 + early_stop: + minimize: true + max_epochs_without_improvement: 10 + epsilon: 0.025 + experiment: + name: ${dataset.gen_subset} + objective_name: wer + minimize: true + parameter_constraints: null + outcome_constraints: null + status_quo: null + client: + verbose_logging: false + random_seed: null + params: + decoding.decoder.lmweight: + type: range + bounds: [0.0, 8.0] + decoding.decoder.wordscore: + type: range + bounds: [-5.0, 5.0] + decoding.decoder.silweight: + type: range + bounds: [-10.0, 0.0] diff --git a/examples/hubert/config/decode/ax_sweep/transformer.yaml b/examples/hubert/config/decode/ax_sweep/transformer.yaml new file mode 100644 index 0000000000000000000000000000000000000000..85ed3bd1a5a44871260f572786044c28f441add6 --- /dev/null +++ b/examples/hubert/config/decode/ax_sweep/transformer.yaml @@ -0,0 +1,33 @@ +# @package _global_ + +common_eval: + results_path: ${decoding.exp_dir}/decode/${decoding.decoder.name}_ax/${dataset.gen_subset} + +hydra: + sweeper: + ax_config: + max_trials: 60 + early_stop: + minimize: true + max_epochs_without_improvement: 10 + epsilon: 0.025 + experiment: + name: ${dataset.gen_subset} + objective_name: wer + minimize: true + parameter_constraints: null + outcome_constraints: null + status_quo: null + client: + verbose_logging: false + random_seed: null + params: + decoding.decoder.lmweight: + type: range + bounds: [0.0, 4.0] + decoding.decoder.wordscore: + type: range + bounds: [-5.0, 5.0] + decoding.decoder.silweight: + type: range + bounds: [-8.0, 0.0] diff --git a/examples/hubert/config/decode/infer_fsqlm.yaml b/examples/hubert/config/decode/infer_fsqlm.yaml new file mode 100644 index 0000000000000000000000000000000000000000..bc77cab32e156f393a2c3eae336392d1796b8a95 --- /dev/null +++ b/examples/hubert/config/decode/infer_fsqlm.yaml @@ -0,0 +1,36 @@ +# @package _group_ + +defaults: + - model: null + +hydra: + run: + dir: ${common_eval.results_path}/beam${decoding.decoder.beam}_lmw${decoding.decoder.lmweight}_wrd${decoding.decoder.wordscore}_sil${decoding.decoder.silweight} + sweep: + dir: ${common_eval.results_path} + subdir: beam${decoding.decoder.beam}_th${decoding.decoder.beamthreshold}_lmw${decoding.decoder.lmweight}_wrd${decoding.decoder.wordscore}_sil${decoding.decoder.silweight} + +task: + _name: hubert_pretraining + single_target: true + data: ??? + normalize: ??? + +decoding: + type: fairseqlm + lexicon: ??? + lmpath: ??? + beamthreshold: 25 # 100 + beam: 500 + lmweight: 2 + wordscore: -1 + silweight: 0 + unique_wer_file: true + beam: 500 +common_eval: + results_path: ??? + path: ??? + post_process: letter +dataset: + max_tokens: 1100000 + gen_subset: ??? diff --git a/examples/hubert/config/decode/infer_kenlm.yaml b/examples/hubert/config/decode/infer_kenlm.yaml new file mode 100644 index 0000000000000000000000000000000000000000..26f5c48928f67af84609ce49e41b93905dafb3ec --- /dev/null +++ b/examples/hubert/config/decode/infer_kenlm.yaml @@ -0,0 +1,36 @@ +# @package _group_ + +defaults: + - model: null + +hydra: + run: + dir: ${common_eval.results_path}/beam${decoding.decoder.beam}_lmw${decoding.decoder.lmweight}_wrd${decoding.decoder.wordscore}_sil${decoding.decoder.silweight} + sweep: + dir: ${common_eval.results_path} + subdir: beam${decoding.decoder.beam}_th${decoding.decoder.beamthreshold}_lmw${decoding.decoder.lmweight}_wrd${decoding.decoder.wordscore}_sil${decoding.decoder.silweight} + +task: + _name: hubert_pretraining + single_target: true + data: ??? + normalize: ??? + +decoding: + type: kenlm + lexicon: ??? + lmpath: ??? + beamthreshold: 100 + beam: 500 + lmweight: 2 + wordscore: -1 + silweight: 0 + unique_wer_file: true + beam: 500 +common_eval: + results_path: ??? + path: ??? + post_process: letter +dataset: + max_tokens: 1100000 + gen_subset: ??? diff --git a/examples/hubert/config/decode/infer_viterbi.yaml b/examples/hubert/config/decode/infer_viterbi.yaml new file mode 100644 index 0000000000000000000000000000000000000000..935d7d1d013136090e5a0154d70a1266d230ee96 --- /dev/null +++ b/examples/hubert/config/decode/infer_viterbi.yaml @@ -0,0 +1,31 @@ +# @package _group_ + +defaults: + - model: null + +hydra: + run: + dir: ${common_eval.results_path}/beam${decoding.decoder.beam}_lmw${decoding.decoder.lmweight}_wrd${decoding.decoder.wordscore}_sil${decoding.decoder.silweight} + sweep: + dir: ${common_eval.results_path} + subdir: beam${decoding.decoder.beam}_th${decoding.decoder.beamthreshold}_lmw${decoding.decoder.lmweight}_wrd${decoding.decoder.wordscore}_sil${decoding.decoder.silweight} + +task: + _name: hubert_pretraining + single_target: true + data: ??? + normalize: ??? + +decoding: + type: viterbi + unique_wer_file: true +common_eval: + results_path: ??? + path: ??? + post_process: letter +generation: + nbest: 1 + beam: 500 +dataset: + max_tokens: 1100000 + gen_subset: ??? diff --git a/examples/hubert/config/decode/run/submitit_slurm.yaml b/examples/hubert/config/decode/run/submitit_slurm.yaml new file mode 100644 index 0000000000000000000000000000000000000000..0b8065832ecacf9dd4fe4e99c87941e00fb3ef7f --- /dev/null +++ b/examples/hubert/config/decode/run/submitit_slurm.yaml @@ -0,0 +1,17 @@ +# @package _global_ +hydra: + launcher: + cpus_per_task: ${distributed_training.distributed_world_size} + gpus_per_node: ${distributed_training.distributed_world_size} + tasks_per_node: ${hydra.launcher.gpus_per_node} + nodes: 1 + mem_gb: 200 + timeout_min: 4320 + max_num_timeout: 50 + name: ${hydra.job.config_name} + submitit_folder: ${hydra.sweep.dir}/submitit + +distributed_training: + distributed_world_size: 1 + distributed_no_spawn: true + distributed_port: 29761 diff --git a/examples/hubert/config/decode/run/submitit_slurm_8gpu.yaml b/examples/hubert/config/decode/run/submitit_slurm_8gpu.yaml new file mode 100644 index 0000000000000000000000000000000000000000..2f669f376312dbfe4611cc08f4996a314155fb87 --- /dev/null +++ b/examples/hubert/config/decode/run/submitit_slurm_8gpu.yaml @@ -0,0 +1,17 @@ +# @package _global_ +hydra: + launcher: + cpus_per_task: ${distributed_training.distributed_world_size} + gpus_per_node: ${distributed_training.distributed_world_size} + tasks_per_node: ${hydra.launcher.gpus_per_node} + nodes: 1 + mem_gb: 200 + timeout_min: 4320 + max_num_timeout: 50 + name: ${hydra.job.config_name} + submitit_folder: ${hydra.sweep.dir}/submitit + +distributed_training: + distributed_world_size: 8 + distributed_no_spawn: true + distributed_port: 29761 diff --git a/examples/hubert/config/finetune/base_10h.yaml b/examples/hubert/config/finetune/base_10h.yaml new file mode 100644 index 0000000000000000000000000000000000000000..a22c7c0347f792221f209bcfba7ba380a69f90a8 --- /dev/null +++ b/examples/hubert/config/finetune/base_10h.yaml @@ -0,0 +1,100 @@ +# @package _group_ + +common: + fp16: true + log_format: json + log_interval: 200 + tensorboard_logdir: tblog + seed: 1337 + +checkpoint: + save_interval: 5 + keep_interval_updates: 1 + no_epoch_checkpoints: true + best_checkpoint_metric: wer + +distributed_training: + ddp_backend: c10d + find_unused_parameters: true + distributed_world_size: 1 + distributed_port: 29671 + nprocs_per_node: 8 + +task: + _name: hubert_pretraining + data: ??? + fine_tuning: true + label_dir: ??? + normalize: false # must be consistent with pre-training + labels: ["ltr"] + single_target: true + +dataset: + num_workers: 0 + max_tokens: 3200000 + validate_after_updates: ${model.freeze_finetune_updates} + validate_interval: 5 + train_subset: train + valid_subset: valid + +criterion: + _name: ctc + zero_infinity: true + +optimization: + max_update: 25000 + lr: [2e-5] + sentence_avg: true + update_freq: [1] + +optimizer: + _name: adam + adam_betas: (0.9,0.98) + adam_eps: 1e-08 + +lr_scheduler: + _name: tri_stage + warmup_steps: 8000 + hold_steps: 0 + decay_steps: 72000 + final_lr_scale: 0.05 + +model: + _name: hubert_ctc + w2v_path: ??? + apply_mask: true + mask_selection: static + mask_length: 10 + mask_other: 0 + mask_prob: 0.75 + mask_channel_selection: static + mask_channel_length: 64 + mask_channel_other: 0 + mask_channel_prob: 0.5 + layerdrop: 0.1 + dropout: 0.0 + activation_dropout: 0.1 + attention_dropout: 0.0 + feature_grad_mult: 0.0 + freeze_finetune_updates: 10000 + +hydra: + job: + config: + override_dirname: + kv_sep: '-' + item_sep: '__' + exclude_keys: + - run + - task.data + - task.label_dir + - model.w2v_path + - dataset.train_subset + - dataset.valid_subset + - criterion.wer_kenlm_model + - criterion.wer_lexicon + run: + dir: ??? + sweep: + dir: ??? + subdir: ${hydra.job.config_name}__${hydra.job.override_dirname} diff --git a/examples/hubert/config/finetune/ckpt/it1.yaml b/examples/hubert/config/finetune/ckpt/it1.yaml new file mode 100644 index 0000000000000000000000000000000000000000..2af96b3f72746f85feb13e7efcbdab6602b293de --- /dev/null +++ b/examples/hubert/config/finetune/ckpt/it1.yaml @@ -0,0 +1,7 @@ +# @package _global_ + +task: + normalize: false + +model: + w2v_path: /checkpoint/wnhsu/w2v/hubert_final/iter1/hubert.km.randcrop.pmw1_0.puw0_0.grpnorm.ml10.mp0_8.untie.mxsz250000.ufreq1.maxtok1400000.MU400k.s1337.ngpu32/checkpoint_last.pt diff --git a/examples/hubert/config/finetune/lm/ls_4gram.yaml b/examples/hubert/config/finetune/lm/ls_4gram.yaml new file mode 100644 index 0000000000000000000000000000000000000000..8c7728ad29965d3cf18605808a893bc442afd56b --- /dev/null +++ b/examples/hubert/config/finetune/lm/ls_4gram.yaml @@ -0,0 +1,7 @@ +# @package _global_ + +criterion: + wer_kenlm_model: /checkpoint/abdo/old_checkpoint02/datasets/librispeech/4-gram.bin + wer_lexicon: /checkpoint/abdo/old_checkpoint02/datasets/librispeech/10h/raw/lexicon_ltr.lst + wer_lm_weight: 2.0 + wer_word_score: -1.0 diff --git a/examples/hubert/config/finetune/run/submitit_reg.yaml b/examples/hubert/config/finetune/run/submitit_reg.yaml new file mode 100644 index 0000000000000000000000000000000000000000..27509503e7b306c07742fbed2fc5726d001bb7df --- /dev/null +++ b/examples/hubert/config/finetune/run/submitit_reg.yaml @@ -0,0 +1,20 @@ +# @package _global_ + +hydra: + launcher: + cpus_per_task: 8 + gpus_per_node: 8 + tasks_per_node: ${hydra.launcher.gpus_per_node} + nodes: 1 + comment: null + mem_gb: 384 + timeout_min: 4320 + max_num_timeout: 100 + constraint: volta32gb + name: ${hydra.job.config_name}/${hydra.job.override_dirname} + submitit_folder: ${hydra.sweep.dir}/submitit/%j + +distributed_training: + distributed_world_size: 8 + distributed_port: 29671 + nprocs_per_node: 8 diff --git a/examples/hubert/config/pretrain/data/iter1.yaml b/examples/hubert/config/pretrain/data/iter1.yaml new file mode 100644 index 0000000000000000000000000000000000000000..0a1b65d802c83128c53f32b21807fa5e51da6cc9 --- /dev/null +++ b/examples/hubert/config/pretrain/data/iter1.yaml @@ -0,0 +1,8 @@ +# @package _global_ + +task: + label_dir: ??? + labels: ["km"] + +model: + label_rate: 100 diff --git a/examples/hubert/config/pretrain/data/iter2.yaml b/examples/hubert/config/pretrain/data/iter2.yaml new file mode 100644 index 0000000000000000000000000000000000000000..2d4bfe61cc638af9de48e92c58994e435fba2abf --- /dev/null +++ b/examples/hubert/config/pretrain/data/iter2.yaml @@ -0,0 +1,8 @@ +# @package _global_ + +task: + label_dir: ??? + labels: ["km"] + +model: + label_rate: 50 diff --git a/examples/hubert/config/pretrain/hubert_base_librispeech.yaml b/examples/hubert/config/pretrain/hubert_base_librispeech.yaml new file mode 100644 index 0000000000000000000000000000000000000000..bd84461a163866f622b01bf6d36b4de6215f3d97 --- /dev/null +++ b/examples/hubert/config/pretrain/hubert_base_librispeech.yaml @@ -0,0 +1,97 @@ +# @package _group_ + +common: + fp16: true + log_format: json + log_interval: 200 + seed: 1337 + tensorboard_logdir: tblog + +checkpoint: + save_interval_updates: 25000 + keep_interval_updates: 1 + no_epoch_checkpoints: true + + +distributed_training: + ddp_backend: no_c10d + distributed_backend: 'nccl' + distributed_world_size: 32 + distributed_port: 29671 + nprocs_per_node: 8 + find_unused_parameters: true + +task: + _name: hubert_pretraining + data: ??? + label_dir: ??? + labels: ??? + label_rate: ${model.label_rate} + sample_rate: 16000 + max_sample_size: 250000 + min_sample_size: 32000 + pad_audio: false + random_crop: true + normalize: false # must be consistent with extractor + +dataset: + num_workers: 6 + max_tokens: 1400000 + skip_invalid_size_inputs_valid_test: true + validate_interval: 5 + validate_interval_updates: 10000 + +criterion: + _name: hubert + pred_masked_weight: 1.0 + pred_nomask_weight: 0.0 + loss_weights: [10,] + +optimization: + max_update: 400000 + lr: [0.0005] + clip_norm: 10.0 + +optimizer: + _name: adam + adam_betas: (0.9,0.98) + adam_eps: 1e-06 + weight_decay: 0.01 + +lr_scheduler: + _name: polynomial_decay + warmup_updates: 32000 + +model: + _name: hubert + label_rate: ??? + skip_masked: false + skip_nomask: false + mask_prob: 0.80 + extractor_mode: default + conv_feature_layers: '[(512,10,5)] + [(512,3,2)] * 4 + [(512,2,2)] * 2' + final_dim: 256 + encoder_layerdrop: 0.05 + dropout_input: 0.1 + dropout_features: 0.1 + dropout: 0.1 + attention_dropout: 0.1 + feature_grad_mult: 0.1 + untie_final_proj: true + activation_dropout: 0.0 + +hydra: + job: + config: + override_dirname: + kv_sep: '-' + item_sep: '__' + exclude_keys: + - run + - task.data + - task.label_dir + run: + dir: ??? + sweep: + dir: ??? + subdir: ${hydra.job.config_name}__${hydra.job.override_dirname} diff --git a/examples/hubert/config/pretrain/hubert_large_librivox.yaml b/examples/hubert/config/pretrain/hubert_large_librivox.yaml new file mode 100644 index 0000000000000000000000000000000000000000..a5192b5f29b53aa8391a0ab67b6238c0d0b4985e --- /dev/null +++ b/examples/hubert/config/pretrain/hubert_large_librivox.yaml @@ -0,0 +1,101 @@ +# @package _group_ + +common: + fp16: true + log_format: json + log_interval: 200 + seed: 1337 + tensorboard_logdir: tblog + +checkpoint: + save_interval_updates: 25000 + keep_interval_updates: 1 + no_epoch_checkpoints: true + + +distributed_training: + ddp_backend: no_c10d + distributed_backend: 'nccl' + distributed_world_size: 128 + distributed_port: 29671 + nprocs_per_node: 8 + find_unused_parameters: true + +task: + _name: hubert_pretraining + data: ??? + label_dir: ??? + labels: ??? + label_rate: ${model.label_rate} + sample_rate: 16000 + max_sample_size: 250000 + min_sample_size: 32000 + pad_audio: false + random_crop: true + normalize: true # must be consistent with extractor + +dataset: + num_workers: 6 + max_tokens: 900000 + skip_invalid_size_inputs_valid_test: true + validate_interval: 5 + validate_interval_updates: 10000 + +criterion: + _name: hubert + pred_masked_weight: 1.0 + pred_nomask_weight: 0.0 + loss_weights: [10,] + +optimization: + max_update: 400000 + lr: [0.0015] + clip_norm: 1.0 + +optimizer: + _name: adam + adam_betas: (0.9,0.98) + adam_eps: 1e-06 + weight_decay: 0.01 + +lr_scheduler: + _name: polynomial_decay + warmup_updates: 32000 + +model: + _name: hubert + label_rate: ??? + encoder_layers: 24 + encoder_embed_dim: 1024 + encoder_ffn_embed_dim: 4096 + encoder_attention_heads: 16 + final_dim: 768 + skip_masked: false + skip_nomask: false + mask_prob: 0.80 + extractor_mode: layer_norm + conv_feature_layers: '[(512,10,5)] + [(512,3,2)] * 4 + [(512,2,2)] * 2' + encoder_layerdrop: 0.0 + dropout_input: 0.0 + dropout_features: 0.0 + dropout: 0.0 + attention_dropout: 0.0 + layer_norm_first: true + feature_grad_mult: 1.0 + untie_final_proj: true + activation_dropout: 0.0 + +hydra: + job: + config: + override_dirname: + kv_sep: '-' + item_sep: '__' + exclude_keys: + - run + - task.data + run: + dir: /checkpoint/wnhsu/w2v/hubert_final/hydra_pt + sweep: + dir: /checkpoint/wnhsu/w2v/hubert_final/hydra_pt + subdir: ${hydra.job.config_name}__${hydra.job.override_dirname} diff --git a/examples/hubert/config/pretrain/hubert_xlarge_librivox.yaml b/examples/hubert/config/pretrain/hubert_xlarge_librivox.yaml new file mode 100644 index 0000000000000000000000000000000000000000..34e8f2bfb93863db122f694785b80857713ceb05 --- /dev/null +++ b/examples/hubert/config/pretrain/hubert_xlarge_librivox.yaml @@ -0,0 +1,101 @@ +# @package _group_ + +common: + fp16: true + log_format: json + log_interval: 200 + seed: 1337 + tensorboard_logdir: tblog + +checkpoint: + save_interval_updates: 25000 + keep_interval_updates: 1 + no_epoch_checkpoints: true + + +distributed_training: + ddp_backend: no_c10d + distributed_backend: 'nccl' + distributed_world_size: 256 + distributed_port: 29671 + nprocs_per_node: 8 + find_unused_parameters: true + +task: + _name: hubert_pretraining + data: ??? + label_dir: ??? + labels: ??? + label_rate: ${model.label_rate} + sample_rate: 16000 + max_sample_size: 250000 + min_sample_size: 32000 + pad_audio: false + random_crop: true + normalize: true # must be consistent with extractor + +dataset: + num_workers: 6 + max_tokens: 360000 + skip_invalid_size_inputs_valid_test: true + validate_interval: 5 + validate_interval_updates: 10000 + +criterion: + _name: hubert + pred_masked_weight: 1.0 + pred_nomask_weight: 0.0 + loss_weights: [10,] + +optimization: + max_update: 400000 + lr: [0.003] + clip_norm: 1.0 + +optimizer: + _name: adam + adam_betas: (0.9,0.98) + adam_eps: 1e-06 + weight_decay: 0.01 + +lr_scheduler: + _name: polynomial_decay + warmup_updates: 32000 + +model: + _name: hubert + label_rate: ??? + encoder_layers: 48 + encoder_embed_dim: 1280 + encoder_ffn_embed_dim: 5120 + encoder_attention_heads: 16 + final_dim: 1024 + skip_masked: false + skip_nomask: false + mask_prob: 0.80 + extractor_mode: layer_norm + conv_feature_layers: '[(512,10,5)] + [(512,3,2)] * 4 + [(512,2,2)] * 2' + encoder_layerdrop: 0.0 + dropout_input: 0.0 + dropout_features: 0.0 + dropout: 0.0 + attention_dropout: 0.0 + layer_norm_first: true + feature_grad_mult: 1.0 + untie_final_proj: true + activation_dropout: 0.0 + +hydra: + job: + config: + override_dirname: + kv_sep: '-' + item_sep: '__' + exclude_keys: + - run + - task.data + run: + dir: /checkpoint/wnhsu/w2v/hubert_final/hydra_pt + sweep: + dir: /checkpoint/wnhsu/w2v/hubert_final/hydra_pt + subdir: ${hydra.job.config_name}__${hydra.job.override_dirname} diff --git a/examples/hubert/config/pretrain/run/submitit_reg.yaml b/examples/hubert/config/pretrain/run/submitit_reg.yaml new file mode 100644 index 0000000000000000000000000000000000000000..46c979cd2835fe026b0a532a54533904d1001e54 --- /dev/null +++ b/examples/hubert/config/pretrain/run/submitit_reg.yaml @@ -0,0 +1,20 @@ +# @package _global_ + +hydra: + launcher: + cpus_per_task: 8 + gpus_per_node: 8 + tasks_per_node: ${hydra.launcher.gpus_per_node} + nodes: 4 + comment: null + mem_gb: 384 + timeout_min: 4320 + max_num_timeout: 100 + constraint: volta32gb + name: ${hydra.job.config_name}/${hydra.job.override_dirname} + submitit_folder: ${hydra.sweep.dir}/submitit/%j + +distributed_training: + distributed_world_size: 32 + distributed_port: 29671 + nprocs_per_node: 8 diff --git a/examples/hubert/measure_teacher_quality.py b/examples/hubert/measure_teacher_quality.py new file mode 100644 index 0000000000000000000000000000000000000000..92279b2214bb2ba4a99aea92098907ef4f55821b --- /dev/null +++ b/examples/hubert/measure_teacher_quality.py @@ -0,0 +1,241 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import numpy as np +import os.path as op +import re +from tabulate import tabulate +from collections import Counter + + +def comp_purity(p_xy, axis): + max_p = p_xy.max(axis=axis) + marg_p = p_xy.sum(axis=axis) + indv_pur = max_p / marg_p + aggr_pur = max_p.sum() + return indv_pur, aggr_pur + + +def comp_entropy(p): + return (-p * np.log(p + 1e-8)).sum() + + +def comp_norm_mutual_info(p_xy): + p_x = p_xy.sum(axis=1, keepdims=True) + p_y = p_xy.sum(axis=0, keepdims=True) + pmi = np.log(p_xy / np.matmul(p_x, p_y) + 1e-8) + mi = (p_xy * pmi).sum() + h_x = comp_entropy(p_x) + h_y = comp_entropy(p_y) + return mi, mi / h_x, mi / h_y, h_x, h_y + + +def pad(labs, n): + if n == 0: + return np.array(labs) + return np.concatenate([[labs[0]] * n, labs, [labs[-1]] * n]) + + +def comp_avg_seg_dur(labs_list): + n_frms = 0 + n_segs = 0 + for labs in labs_list: + labs = np.array(labs) + edges = np.zeros(len(labs)).astype(bool) + edges[0] = True + edges[1:] = labs[1:] != labs[:-1] + n_frms += len(edges) + n_segs += edges.astype(int).sum() + return n_frms / n_segs + + +def comp_joint_prob(uid2refs, uid2hyps): + """ + Args: + pad: padding for spliced-feature derived labels + """ + cnts = Counter() + skipped = [] + abs_frmdiff = 0 + for uid in uid2refs: + if uid not in uid2hyps: + skipped.append(uid) + continue + refs = uid2refs[uid] + hyps = uid2hyps[uid] + abs_frmdiff += abs(len(refs) - len(hyps)) + min_len = min(len(refs), len(hyps)) + refs = refs[:min_len] + hyps = hyps[:min_len] + cnts.update(zip(refs, hyps)) + tot = sum(cnts.values()) + + ref_set = sorted({ref for ref, _ in cnts.keys()}) + hyp_set = sorted({hyp for _, hyp in cnts.keys()}) + ref2pid = dict(zip(ref_set, range(len(ref_set)))) + hyp2lid = dict(zip(hyp_set, range(len(hyp_set)))) + # print(hyp_set) + p_xy = np.zeros((len(ref2pid), len(hyp2lid)), dtype=float) + for (ref, hyp), cnt in cnts.items(): + p_xy[ref2pid[ref], hyp2lid[hyp]] = cnt + p_xy /= p_xy.sum() + return p_xy, ref2pid, hyp2lid, tot, abs_frmdiff, skipped + + +def read_phn(tsv_path, rm_stress=True): + uid2phns = {} + with open(tsv_path) as f: + for line in f: + uid, phns = line.rstrip().split("\t") + phns = phns.split(",") + if rm_stress: + phns = [re.sub("[0-9]", "", phn) for phn in phns] + uid2phns[uid] = phns + return uid2phns + + +def read_lab(tsv_path, lab_path, pad_len=0, upsample=1): + """ + tsv is needed to retrieve the uids for the labels + """ + with open(tsv_path) as f: + f.readline() + uids = [op.splitext(op.basename(line.rstrip().split()[0]))[0] for line in f] + with open(lab_path) as f: + labs_list = [pad(line.rstrip().split(), pad_len).repeat(upsample) for line in f] + assert len(uids) == len(labs_list) + return dict(zip(uids, labs_list)) + + +def main_lab_lab( + tsv_dir, + lab_dir, + lab_name, + lab_sets, + ref_dir, + ref_name, + pad_len=0, + upsample=1, + verbose=False, +): + # assume tsv_dir is the same for both the reference and the hypotheses + tsv_dir = lab_dir if tsv_dir is None else tsv_dir + + uid2refs = {} + for s in lab_sets: + uid2refs.update(read_lab(f"{tsv_dir}/{s}.tsv", f"{ref_dir}/{s}.{ref_name}")) + + uid2hyps = {} + for s in lab_sets: + uid2hyps.update( + read_lab( + f"{tsv_dir}/{s}.tsv", f"{lab_dir}/{s}.{lab_name}", pad_len, upsample + ) + ) + _main(uid2refs, uid2hyps, verbose) + + +def main_phn_lab( + tsv_dir, + lab_dir, + lab_name, + lab_sets, + phn_dir, + phn_sets, + pad_len=0, + upsample=1, + verbose=False, +): + uid2refs = {} + for s in phn_sets: + uid2refs.update(read_phn(f"{phn_dir}/{s}.tsv")) + + uid2hyps = {} + tsv_dir = lab_dir if tsv_dir is None else tsv_dir + for s in lab_sets: + uid2hyps.update( + read_lab( + f"{tsv_dir}/{s}.tsv", f"{lab_dir}/{s}.{lab_name}", pad_len, upsample + ) + ) + _main(uid2refs, uid2hyps, verbose) + + +def _main(uid2refs, uid2hyps, verbose): + (p_xy, ref2pid, hyp2lid, tot, frmdiff, skipped) = comp_joint_prob( + uid2refs, uid2hyps + ) + ref_pur_by_hyp, ref_pur = comp_purity(p_xy, axis=0) + hyp_pur_by_ref, hyp_pur = comp_purity(p_xy, axis=1) + (mi, mi_norm_by_ref, mi_norm_by_hyp, h_ref, h_hyp) = comp_norm_mutual_info(p_xy) + outputs = { + "ref pur": ref_pur, + "hyp pur": hyp_pur, + "H(ref)": h_ref, + "H(hyp)": h_hyp, + "MI": mi, + "MI/H(ref)": mi_norm_by_ref, + "ref segL": comp_avg_seg_dur(uid2refs.values()), + "hyp segL": comp_avg_seg_dur(uid2hyps.values()), + "p_xy shape": p_xy.shape, + "frm tot": tot, + "frm diff": frmdiff, + "utt tot": len(uid2refs), + "utt miss": len(skipped), + } + print(tabulate([outputs.values()], outputs.keys(), floatfmt=".4f")) + + +if __name__ == "__main__": + """ + compute quality of labels with respect to phone or another labels if set + """ + import argparse + + parser = argparse.ArgumentParser() + parser.add_argument("tsv_dir") + parser.add_argument("lab_dir") + parser.add_argument("lab_name") + parser.add_argument("--lab_sets", default=["valid"], type=str, nargs="+") + parser.add_argument( + "--phn_dir", + default="/checkpoint/wnhsu/data/librispeech/960h/fa/raw_phn/phone_frame_align_v1", + ) + parser.add_argument( + "--phn_sets", default=["dev-clean", "dev-other"], type=str, nargs="+" + ) + parser.add_argument("--pad_len", default=0, type=int, help="padding for hypotheses") + parser.add_argument( + "--upsample", default=1, type=int, help="upsample factor for hypotheses" + ) + parser.add_argument("--ref_lab_dir", default="") + parser.add_argument("--ref_lab_name", default="") + parser.add_argument("--verbose", action="store_true") + args = parser.parse_args() + + if args.ref_lab_dir and args.ref_lab_name: + main_lab_lab( + args.tsv_dir, + args.lab_dir, + args.lab_name, + args.lab_sets, + args.ref_lab_dir, + args.ref_lab_name, + args.pad_len, + args.upsample, + args.verbose, + ) + else: + main_phn_lab( + args.tsv_dir, + args.lab_dir, + args.lab_name, + args.lab_sets, + args.phn_dir, + args.phn_sets, + args.pad_len, + args.upsample, + args.verbose, + ) diff --git a/examples/hubert/simple_kmeans/README.md b/examples/hubert/simple_kmeans/README.md new file mode 100644 index 0000000000000000000000000000000000000000..cd17da3b3e6f3e39083f7a76a56ff46c3a63b929 --- /dev/null +++ b/examples/hubert/simple_kmeans/README.md @@ -0,0 +1,71 @@ +# Sharded Feature Extraction and K-means Application + +This folder contains scripts for preparing HUBERT labels from tsv files, the +steps are: +1. feature extraction +2. k-means clustering +3. k-means application + + +## Data preparation + +`*.tsv` files contains a list of audio, where each line is the root, and +following lines are the subpath for each audio: +``` + + + +... +``` + + +## Feature extraction + +### MFCC feature +Suppose the tsv file is at `${tsv_dir}/${split}.tsv`. To extract 39-D +mfcc+delta+ddelta features for the 1st iteration HUBERT training, run: +```sh +python dump_mfcc_feature.py ${tsv_dir} ${split} ${nshard} ${rank} ${feat_dir} +``` +This would shard the tsv file into `${nshard}` and extract features for the +`${rank}`-th shard, where rank is an integer in `[0, nshard-1]`. Features would +be saved at `${feat_dir}/${split}_${rank}_${nshard}.{npy,len}`. + + +### HUBERT feature +To extract features from the `${layer}`-th transformer layer of a trained +HUBERT model saved at `${ckpt_path}`, run: +```sh +python dump_hubert_feature.py ${tsv_dir} ${split} ${ckpt_path} ${layer} ${nshard} ${rank} ${feat_dir} +``` +Features would also be saved at `${feat_dir}/${split}_${rank}_${nshard}.{npy,len}`. + +- if out-of-memory, decrease the chunk size with `--max_chunk` + + +## K-means clustering +To fit a k-means model with `${n_clusters}` clusters on 10% of the `${split}` data, run +```sh +python learn_kmeans.py ${feat_dir} ${split} ${nshard} ${km_path} ${n_cluster} --percent 0.1 +``` +This saves the k-means model to `${km_path}`. + +- set `--precent -1` to use all data +- more kmeans options can be found with `-h` flag + + +## K-means application +To apply a trained k-means model `${km_path}` to obtain labels for `${split}`, run +```sh +python dump_km_label.py ${feat_dir} ${split} ${km_path} ${nshard} ${rank} ${lab_dir} +``` +This would extract labels for the `${rank}`-th shard out of `${nshard}` shards +and dump them to `${lab_dir}/${split}_${rank}_${shard}.km` + + +Finally, merge shards for `${split}` by running +```sh +for rank in $(seq 0 $((nshard - 1))); do + cat $lab_dir/${split}_${rank}_${nshard}.km +done > $lab_dir/${split}.km +``` diff --git a/examples/hubert/simple_kmeans/dump_hubert_feature.py b/examples/hubert/simple_kmeans/dump_hubert_feature.py new file mode 100644 index 0000000000000000000000000000000000000000..cd242890e531208e5a732842e53085bb1acc8664 --- /dev/null +++ b/examples/hubert/simple_kmeans/dump_hubert_feature.py @@ -0,0 +1,133 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import math +import os +import sys + +import fairseq +import soundfile as sf +import torch +import torch.nn.functional as F +import tqdm +from npy_append_array import NpyAppendArray + +logging.basicConfig( + format="%(asctime)s | %(levelname)s | %(name)s | %(message)s", + datefmt="%Y-%m-%d %H:%M:%S", + level=os.environ.get("LOGLEVEL", "INFO").upper(), + stream=sys.stdout, +) +logger = logging.getLogger("dump_hubert_feature") + + +class HubertFeatureReader(object): + def __init__(self, ckpt_path, layer, max_chunk=1600000): + ( + model, + cfg, + task, + ) = fairseq.checkpoint_utils.load_model_ensemble_and_task([ckpt_path]) + self.model = model[0].eval().cuda() + self.task = task + self.layer = layer + self.max_chunk = max_chunk + logger.info(f"TASK CONFIG:\n{self.task.cfg}") + logger.info(f" max_chunk = {self.max_chunk}") + + def read_audio(self, path, ref_len=None): + wav, sr = sf.read(path) + assert sr == self.task.cfg.sample_rate, sr + if wav.ndim == 2: + wav = wav.mean(-1) + assert wav.ndim == 1, wav.ndim + if ref_len is not None and abs(ref_len - len(wav)) > 160: + logging.warning(f"ref {ref_len} != read {len(wav)} ({path})") + return wav + + def get_feats(self, path, ref_len=None): + x = self.read_audio(path, ref_len) + with torch.no_grad(): + x = torch.from_numpy(x).float().cuda() + if self.task.cfg.normalize: + x = F.layer_norm(x, x.shape) + x = x.view(1, -1) + + feat = [] + for start in range(0, x.size(1), self.max_chunk): + x_chunk = x[:, start: start + self.max_chunk] + feat_chunk, _ = self.model.extract_features( + source=x_chunk, + padding_mask=None, + mask=False, + output_layer=self.layer, + ) + feat.append(feat_chunk) + return torch.cat(feat, 1).squeeze(0) + + +def get_path_iterator(tsv, nshard, rank): + with open(tsv, "r") as f: + root = f.readline().rstrip() + lines = [line.rstrip() for line in f] + tot = len(lines) + shard_size = math.ceil(tot / nshard) + start, end = rank * shard_size, min((rank + 1) * shard_size, tot) + assert start < end, "start={start}, end={end}" + logger.info( + f"rank {rank} of {nshard}, process {end-start} " + f"({start}-{end}) out of {tot}" + ) + + lines = lines[start:end] + + def iterate(): + for line in lines: + subpath, nsample = line.split("\t") + yield f"{root}/{subpath}", int(nsample) + + return iterate, len(lines) + + +def dump_feature( + tsv_dir, split, ckpt_path, layer, nshard, rank, feat_dir, max_chunk +): + reader = HubertFeatureReader(ckpt_path, layer, max_chunk) + generator, num = get_path_iterator(f"{tsv_dir}/{split}.tsv", nshard, rank) + iterator = generator() + + feat_path = f"{feat_dir}/{split}_{rank}_{nshard}.npy" + leng_path = f"{feat_dir}/{split}_{rank}_{nshard}.len" + + os.makedirs(feat_dir, exist_ok=True) + if os.path.exists(feat_path): + os.remove(feat_path) + + feat_f = NpyAppendArray(feat_path) + with open(leng_path, "w") as leng_f: + for path, nsample in tqdm.tqdm(iterator, total=num): + feat = reader.get_feats(path, nsample) + feat_f.append(feat.cpu().numpy()) + leng_f.write(f"{len(feat)}\n") + logger.info("finished successfully") + + +if __name__ == "__main__": + import argparse + + parser = argparse.ArgumentParser() + parser.add_argument("tsv_dir") + parser.add_argument("split") + parser.add_argument("ckpt_path") + parser.add_argument("layer", type=int) + parser.add_argument("nshard", type=int) + parser.add_argument("rank", type=int) + parser.add_argument("feat_dir") + parser.add_argument("--max_chunk", type=int, default=1600000) + args = parser.parse_args() + logger.info(args) + + dump_feature(**vars(args)) diff --git a/examples/hubert/simple_kmeans/dump_hubert_feature_s2t.py b/examples/hubert/simple_kmeans/dump_hubert_feature_s2t.py new file mode 100644 index 0000000000000000000000000000000000000000..7ec8a7311b4eee7fa6d42c0615fc06148262f63c --- /dev/null +++ b/examples/hubert/simple_kmeans/dump_hubert_feature_s2t.py @@ -0,0 +1,126 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import csv +import io +import logging +import math +import os +import os.path as op +import sys + +import tqdm +from dump_hubert_feature import HubertFeatureReader +from fairseq.data.audio.audio_utils import get_waveform +from fairseq.data.audio.speech_to_text_dataset import ( + read_from_uncompressed_zip, +) +from npy_append_array import NpyAppendArray + +logging.basicConfig( + format="%(asctime)s | %(levelname)s | %(name)s | %(message)s", + datefmt="%Y-%m-%d %H:%M:%S", + level=os.environ.get("LOGLEVEL", "INFO").upper(), + stream=sys.stdout, +) +logger = logging.getLogger("dump_hubert_feature_s2t") + + +class HubertFeatureReaderS2T(HubertFeatureReader): + def read_audio(self, path, ref_len=None): + path, *extra = path.split(":") + assert len(extra) == 2 + assert path.endswith(".zip") + + data = read_from_uncompressed_zip(path, int(extra[0]), int(extra[1])) + f = io.BytesIO(data) + wav, sr = get_waveform(f) + assert sr == self.task.cfg.sample_rate, sr + if wav.ndim == 2: + wav = wav.mean(-1) + assert wav.ndim == 1, wav.ndim + if ref_len is not None and abs(ref_len - len(wav)) > 160: + logging.warning(f"ref {ref_len} != read {len(wav)} ({path})") + return wav + + +def get_path_iterator(root, tsv, nshard, rank): + with open(tsv) as f: + reader = csv.DictReader( + f, + delimiter="\t", + quotechar=None, + doublequote=False, + lineterminator="\n", + quoting=csv.QUOTE_NONE, + ) + subpaths = [op.join(root, e["audio"]) for e in reader] + + tot = len(subpaths) + shard_size = math.ceil(tot / nshard) + start, end = rank * shard_size, min((rank + 1) * shard_size, tot) + assert start < end, "start={start}, end={end}" + logger.info( + f"rank {rank} of {nshard}, process {end-start} " + f"({start}-{end}) out of {tot}" + ) + + subpaths = subpaths[start:end] + + def iterate(): + for subpath in subpaths: + yield op.join(root, subpath) + + return iterate, len(subpaths) + + +def dump_feature( + root, + tsv_path, + ckpt_path, + layer, + nshard, + rank, + feat_dir, + feat_name, + max_chunk, +): + reader = HubertFeatureReaderS2T(ckpt_path, layer, max_chunk) + generator, num = get_path_iterator(root, tsv_path, nshard, rank) + iterator = generator() + + feat_path = f"{feat_dir}/{feat_name}_{rank}_{nshard}.npy" + leng_path = f"{feat_dir}/{feat_name}_{rank}_{nshard}.len" + + os.makedirs(feat_dir, exist_ok=True) + if op.exists(feat_path): + os.remove(feat_path) + + feat_f = NpyAppendArray(feat_path) + with open(leng_path, "w") as leng_f: + for path in tqdm.tqdm(iterator, total=num): + feat = reader.get_feats(path) + feat_f.append(feat.cpu().numpy()) + leng_f.write(f"{len(feat)}\n") + logger.info("finished successfully") + + +if __name__ == "__main__": + import argparse + + parser = argparse.ArgumentParser() + parser.add_argument("root") + parser.add_argument("tsv_path") + parser.add_argument("ckpt_path") + parser.add_argument("layer", type=int) + parser.add_argument("nshard", type=int) + parser.add_argument("rank", type=int) + parser.add_argument("feat_dir") + parser.add_argument("feat_name") + parser.add_argument("--max_chunk", type=int, default=1600000) + args = parser.parse_args() + logger.info(args) + + dump_feature(**vars(args)) diff --git a/examples/hubert/simple_kmeans/dump_km_label.py b/examples/hubert/simple_kmeans/dump_km_label.py new file mode 100644 index 0000000000000000000000000000000000000000..8871307804d3f1e5c7cc49061614c69df26ab1ee --- /dev/null +++ b/examples/hubert/simple_kmeans/dump_km_label.py @@ -0,0 +1,98 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import os +import sys + +import numpy as np + +import joblib +import torch +import tqdm + +logging.basicConfig( + format="%(asctime)s | %(levelname)s | %(name)s | %(message)s", + datefmt="%Y-%m-%d %H:%M:%S", + level=os.environ.get("LOGLEVEL", "INFO").upper(), + stream=sys.stdout, +) +logger = logging.getLogger("dump_km_label") + + +class ApplyKmeans(object): + def __init__(self, km_path): + self.km_model = joblib.load(km_path) + self.C_np = self.km_model.cluster_centers_.transpose() + self.Cnorm_np = (self.C_np ** 2).sum(0, keepdims=True) + + self.C = torch.from_numpy(self.C_np) + self.Cnorm = torch.from_numpy(self.Cnorm_np) + if torch.cuda.is_available(): + self.C = self.C.cuda() + self.Cnorm = self.Cnorm.cuda() + + def __call__(self, x): + if isinstance(x, torch.Tensor): + dist = ( + x.pow(2).sum(1, keepdim=True) + - 2 * torch.matmul(x, self.C) + + self.Cnorm + ) + return dist.argmin(dim=1).cpu().numpy() + else: + dist = ( + (x ** 2).sum(1, keepdims=True) + - 2 * np.matmul(x, self.C_np) + + self.Cnorm_np + ) + return np.argmin(dist, axis=1) + + +def get_feat_iterator(feat_dir, split, nshard, rank): + feat_path = f"{feat_dir}/{split}_{rank}_{nshard}.npy" + leng_path = f"{feat_dir}/{split}_{rank}_{nshard}.len" + with open(leng_path, "r") as f: + lengs = [int(line.rstrip()) for line in f] + offsets = [0] + np.cumsum(lengs[:-1]).tolist() + + def iterate(): + feat = np.load(feat_path, mmap_mode="r") + assert feat.shape[0] == (offsets[-1] + lengs[-1]) + for offset, leng in zip(offsets, lengs): + yield feat[offset: offset + leng] + + return iterate, len(lengs) + + +def dump_label(feat_dir, split, km_path, nshard, rank, lab_dir): + apply_kmeans = ApplyKmeans(km_path) + generator, num = get_feat_iterator(feat_dir, split, nshard, rank) + iterator = generator() + + lab_path = f"{lab_dir}/{split}_{rank}_{nshard}.km" + os.makedirs(lab_dir, exist_ok=True) + with open(lab_path, "w") as f: + for feat in tqdm.tqdm(iterator, total=num): + # feat = torch.from_numpy(feat).cuda() + lab = apply_kmeans(feat).tolist() + f.write(" ".join(map(str, lab)) + "\n") + logger.info("finished successfully") + + +if __name__ == "__main__": + import argparse + + parser = argparse.ArgumentParser() + parser.add_argument("feat_dir") + parser.add_argument("split") + parser.add_argument("km_path") + parser.add_argument("nshard", type=int) + parser.add_argument("rank", type=int) + parser.add_argument("lab_dir") + args = parser.parse_args() + logging.info(str(args)) + + dump_label(**vars(args)) diff --git a/examples/hubert/simple_kmeans/dump_mfcc_feature.py b/examples/hubert/simple_kmeans/dump_mfcc_feature.py new file mode 100644 index 0000000000000000000000000000000000000000..a36fa643bd28134aef56d0630d49efdf7969f876 --- /dev/null +++ b/examples/hubert/simple_kmeans/dump_mfcc_feature.py @@ -0,0 +1,116 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import math +import os +import sys + +import soundfile as sf +import torch +import torchaudio +import tqdm +from npy_append_array import NpyAppendArray + +logging.basicConfig( + format="%(asctime)s | %(levelname)s | %(name)s | %(message)s", + datefmt="%Y-%m-%d %H:%M:%S", + level=os.environ.get("LOGLEVEL", "INFO").upper(), + stream=sys.stdout, +) +logger = logging.getLogger("dump_mfcc_feature") + + +class MfccFeatureReader(object): + def __init__(self, sample_rate): + self.sample_rate = sample_rate + + def read_audio(self, path, ref_len=None): + wav, sr = sf.read(path) + assert sr == self.sample_rate, sr + if wav.ndim == 2: + wav = wav.mean(-1) + assert wav.ndim == 1, wav.ndim + if ref_len is not None and abs(ref_len - len(wav)) > 160: + logging.warning(f"ref {ref_len} != read {len(wav)} ({path})") + return wav + + def get_feats(self, path, ref_len=None): + x = self.read_audio(path, ref_len) + with torch.no_grad(): + x = torch.from_numpy(x).float() + x = x.view(1, -1) + + mfccs = torchaudio.compliance.kaldi.mfcc( + waveform=x, + sample_frequency=self.sample_rate, + use_energy=False, + ) # (time, freq) + mfccs = mfccs.transpose(0, 1) # (freq, time) + deltas = torchaudio.functional.compute_deltas(mfccs) + ddeltas = torchaudio.functional.compute_deltas(deltas) + concat = torch.cat([mfccs, deltas, ddeltas], dim=0) + concat = concat.transpose(0, 1).contiguous() # (freq, time) + return concat + + +def get_path_iterator(tsv, nshard, rank): + with open(tsv, "r") as f: + root = f.readline().rstrip() + lines = [line.rstrip() for line in f] + tot = len(lines) + shard_size = math.ceil(tot / nshard) + start, end = rank * shard_size, min((rank + 1) * shard_size, tot) + assert start < end, "start={start}, end={end}" + logger.info( + f"rank {rank} of {nshard}, process {end-start} " + f"({start}-{end}) out of {tot}" + ) + + lines = lines[start:end] + + def iterate(): + for line in lines: + subpath, nsample = line.split("\t") + yield f"{root}/{subpath}", int(nsample) + + return iterate, len(lines) + + +def dump_feature(tsv_dir, split, sample_rate, nshard, rank, feat_dir): + reader = MfccFeatureReader(sample_rate) + generator, num = get_path_iterator(f"{tsv_dir}/{split}.tsv", nshard, rank) + iterator = generator() + + feat_path = f"{feat_dir}/{split}_{rank}_{nshard}.npy" + leng_path = f"{feat_dir}/{split}_{rank}_{nshard}.len" + + os.makedirs(feat_dir, exist_ok=True) + if os.path.exists(feat_path): + os.remove(feat_path) + + feat_f = NpyAppendArray(feat_path) + with open(leng_path, "w") as leng_f: + for path, nsample in tqdm.tqdm(iterator, total=num): + feat = reader.get_feats(path, nsample) + feat_f.append(feat.cpu().numpy()) + leng_f.write(f"{len(feat)}\n") + logger.info("finished successfully") + + +if __name__ == "__main__": + import argparse + + parser = argparse.ArgumentParser() + parser.add_argument("tsv_dir") + parser.add_argument("split") + parser.add_argument("nshard", type=int) + parser.add_argument("rank", type=int) + parser.add_argument("feat_dir") + parser.add_argument("--sample_rate", type=int, default=16000) + args = parser.parse_args() + logger.info(args) + + dump_feature(**vars(args)) diff --git a/examples/hubert/simple_kmeans/learn_kmeans.py b/examples/hubert/simple_kmeans/learn_kmeans.py new file mode 100644 index 0000000000000000000000000000000000000000..113ac655b8c0a585fe43797e99674e445098edd0 --- /dev/null +++ b/examples/hubert/simple_kmeans/learn_kmeans.py @@ -0,0 +1,146 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import os +import sys + +import numpy as np +from sklearn.cluster import MiniBatchKMeans + +import joblib + +logging.basicConfig( + format="%(asctime)s | %(levelname)s | %(name)s | %(message)s", + datefmt="%Y-%m-%d %H:%M:%S", + level=os.environ.get("LOGLEVEL", "INFO").upper(), + stream=sys.stdout, +) +logger = logging.getLogger("learn_kmeans") + + +def get_km_model( + n_clusters, + init, + max_iter, + batch_size, + tol, + max_no_improvement, + n_init, + reassignment_ratio, +): + return MiniBatchKMeans( + n_clusters=n_clusters, + init=init, + max_iter=max_iter, + batch_size=batch_size, + verbose=1, + compute_labels=False, + tol=tol, + max_no_improvement=max_no_improvement, + init_size=None, + n_init=n_init, + reassignment_ratio=reassignment_ratio, + ) + + +def load_feature_shard(feat_dir, split, nshard, rank, percent): + feat_path = f"{feat_dir}/{split}_{rank}_{nshard}.npy" + leng_path = f"{feat_dir}/{split}_{rank}_{nshard}.len" + with open(leng_path, "r") as f: + lengs = [int(line.rstrip()) for line in f] + offsets = [0] + np.cumsum(lengs[:-1]).tolist() + + if percent < 0: + return np.load(feat_path, mmap_mode="r") + else: + nsample = int(np.ceil(len(lengs) * percent)) + indices = np.random.choice(len(lengs), nsample, replace=False) + feat = np.load(feat_path, mmap_mode="r") + sampled_feat = np.concatenate( + [feat[offsets[i]: offsets[i] + lengs[i]] for i in indices], axis=0 + ) + logger.info( + ( + f"sampled {nsample} utterances, {len(sampled_feat)} frames " + f"from shard {rank}/{nshard}" + ) + ) + return sampled_feat + + +def load_feature(feat_dir, split, nshard, seed, percent): + assert percent <= 1.0 + feat = np.concatenate( + [ + load_feature_shard(feat_dir, split, nshard, r, percent) + for r in range(nshard) + ], + axis=0, + ) + logging.info(f"loaded feature with dimension {feat.shape}") + return feat + + +def learn_kmeans( + feat_dir, + split, + nshard, + km_path, + n_clusters, + seed, + percent, + init, + max_iter, + batch_size, + tol, + n_init, + reassignment_ratio, + max_no_improvement, +): + np.random.seed(seed) + feat = load_feature(feat_dir, split, nshard, seed, percent) + km_model = get_km_model( + n_clusters, + init, + max_iter, + batch_size, + tol, + max_no_improvement, + n_init, + reassignment_ratio, + ) + km_model.fit(feat) + joblib.dump(km_model, km_path) + + inertia = -km_model.score(feat) / len(feat) + logger.info("total intertia: %.5f", inertia) + logger.info("finished successfully") + + +if __name__ == "__main__": + import argparse + + parser = argparse.ArgumentParser() + parser.add_argument("feat_dir", type=str) + parser.add_argument("split", type=str) + parser.add_argument("nshard", type=int) + parser.add_argument("km_path", type=str) + parser.add_argument("n_clusters", type=int) + parser.add_argument("--seed", default=0, type=int) + parser.add_argument( + "--percent", default=-1, type=float, help="sample a subset; -1 for all" + ) + parser.add_argument("--init", default="k-means++") + parser.add_argument("--max_iter", default=100, type=int) + parser.add_argument("--batch_size", default=10000, type=int) + parser.add_argument("--tol", default=0.0, type=float) + parser.add_argument("--max_no_improvement", default=100, type=int) + parser.add_argument("--n_init", default=20, type=int) + parser.add_argument("--reassignment_ratio", default=0.0, type=float) + args = parser.parse_args() + logging.info(str(args)) + + learn_kmeans(**vars(args)) diff --git a/examples/hubert/update_ckpt.py b/examples/hubert/update_ckpt.py new file mode 100644 index 0000000000000000000000000000000000000000..53c9e74ea613e30aa5c22614e658f2b7272bac0c --- /dev/null +++ b/examples/hubert/update_ckpt.py @@ -0,0 +1,22 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch + +src_ckpt = "/checkpoint/wnhsu/w2v/archived/hubert_base_ls960_it2.pt" +ref_ckpt = "/checkpoint/wnhsu/w2v/hubert_icassp_oss_v3/iter2_km100-400k-grp-L6/oss.km500_p0_1_s334.pmw1_0.puw0_0.grpnorm.ml10.mp0_8.untie.mxsz250000.ufreq1.maxtok1400000.MU100k.s1337.ngpu32/checkpoint_last.pt" +new_ckpt = "/checkpoint/wnhsu/w2v/archived/hubert_base_ls960_it2_updated.pt" + + +def update_state(state): + state["model"]["label_embs_concat"] = state["model"].pop("label_embs") + state["args"].task = "hubert_pretraining" + state["args"].labels = f"['{state['args'].labels}']" + return state + + +src_state = torch.load(src_ckpt) +src_state = update_state(src_state) +torch.save(src_state, new_ckpt) diff --git a/examples/joint_alignment_translation/README.md b/examples/joint_alignment_translation/README.md new file mode 100644 index 0000000000000000000000000000000000000000..cd9c0ea65f5292198296a8f427b42e01b584e2d9 --- /dev/null +++ b/examples/joint_alignment_translation/README.md @@ -0,0 +1,89 @@ +# Jointly Learning to Align and Translate with Transformer Models (Garg et al., 2019) + +This page includes instructions for training models described in [Jointly Learning to Align and Translate with Transformer Models (Garg et al., 2019)](https://arxiv.org/abs/1909.02074). + +## Training a joint alignment-translation model on WMT'18 En-De + +##### 1. Extract and preprocess the WMT'18 En-De data +```bash +./prepare-wmt18en2de_no_norm_no_escape_no_agressive.sh +``` + +##### 2. Generate alignments from statistical alignment toolkits e.g. Giza++/FastAlign. +In this example, we use FastAlign. +```bash +git clone git@github.com:clab/fast_align.git +pushd fast_align +mkdir build +cd build +cmake .. +make +popd +ALIGN=fast_align/build/fast_align +paste bpe.32k/train.en bpe.32k/train.de | awk -F '\t' '{print $1 " ||| " $2}' > bpe.32k/train.en-de +$ALIGN -i bpe.32k/train.en-de -d -o -v > bpe.32k/train.align +``` + +##### 3. Preprocess the dataset with the above generated alignments. +```bash +fairseq-preprocess \ + --source-lang en --target-lang de \ + --trainpref bpe.32k/train \ + --validpref bpe.32k/valid \ + --testpref bpe.32k/test \ + --align-suffix align \ + --destdir binarized/ \ + --joined-dictionary \ + --workers 32 +``` + +##### 4. Train a model +```bash +fairseq-train \ + binarized \ + --arch transformer_wmt_en_de_big_align --share-all-embeddings \ + --optimizer adam --adam-betas '(0.9, 0.98)' --clip-norm 0.0 --activation-fn relu\ + --lr 0.0002 --lr-scheduler inverse_sqrt --warmup-updates 4000 --warmup-init-lr 1e-07 \ + --dropout 0.3 --attention-dropout 0.1 --weight-decay 0.0 \ + --max-tokens 3500 --label-smoothing 0.1 \ + --save-dir ./checkpoints --log-interval 1000 --max-update 60000 \ + --keep-interval-updates -1 --save-interval-updates 0 \ + --load-alignments --criterion label_smoothed_cross_entropy_with_alignment \ + --fp16 +``` + +Note that the `--fp16` flag requires you have CUDA 9.1 or greater and a Volta GPU or newer. + +If you want to train the above model with big batches (assuming your machine has 8 GPUs): +- add `--update-freq 8` to simulate training on 8x8=64 GPUs +- increase the learning rate; 0.0007 works well for big batches + +##### 5. Evaluate and generate the alignments (BPE level) +```bash +fairseq-generate \ + binarized --gen-subset test --print-alignment \ + --source-lang en --target-lang de \ + --path checkpoints/checkpoint_best.pt --beam 5 --nbest 1 +``` + +##### 6. Other resources. +The code for: +1. preparing alignment test sets +2. converting BPE level alignments to token level alignments +3. symmetrizing bidirectional alignments +4. evaluating alignments using AER metric +can be found [here](https://github.com/lilt/alignment-scripts) + +## Citation + +```bibtex +@inproceedings{garg2019jointly, + title = {Jointly Learning to Align and Translate with Transformer Models}, + author = {Garg, Sarthak and Peitz, Stephan and Nallasamy, Udhyakumar and Paulik, Matthias}, + booktitle = {Conference on Empirical Methods in Natural Language Processing (EMNLP)}, + address = {Hong Kong}, + month = {November}, + url = {https://arxiv.org/abs/1909.02074}, + year = {2019}, +} +``` diff --git a/examples/joint_alignment_translation/prepare-wmt18en2de_no_norm_no_escape_no_agressive.sh b/examples/joint_alignment_translation/prepare-wmt18en2de_no_norm_no_escape_no_agressive.sh new file mode 100755 index 0000000000000000000000000000000000000000..e3efeb21d302ef8d9eae8f1d4b06434c593705f6 --- /dev/null +++ b/examples/joint_alignment_translation/prepare-wmt18en2de_no_norm_no_escape_no_agressive.sh @@ -0,0 +1,118 @@ +#!/bin/bash + +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +echo 'Cloning Moses github repository (for tokenization scripts)...' +git clone https://github.com/moses-smt/mosesdecoder.git + +SCRIPTS=mosesdecoder/scripts +TOKENIZER=$SCRIPTS/tokenizer/tokenizer.perl +CLEAN=$SCRIPTS/training/clean-corpus-n.perl +REM_NON_PRINT_CHAR=$SCRIPTS/tokenizer/remove-non-printing-char.perl + +URLS=( + "http://statmt.org/wmt13/training-parallel-europarl-v7.tgz" + "http://statmt.org/wmt13/training-parallel-commoncrawl.tgz" + "http://data.statmt.org/wmt18/translation-task/training-parallel-nc-v13.tgz" + "http://data.statmt.org/wmt18/translation-task/rapid2016.tgz" + "http://data.statmt.org/wmt17/translation-task/dev.tgz" + "http://statmt.org/wmt14/test-full.tgz" +) +CORPORA=( + "training/europarl-v7.de-en" + "commoncrawl.de-en" + "training-parallel-nc-v13/news-commentary-v13.de-en" + "rapid2016.de-en" +) + +if [ ! -d "$SCRIPTS" ]; then + echo "Please set SCRIPTS variable correctly to point to Moses scripts." + exit +fi + +src=en +tgt=de +lang=en-de +prep=wmt18_en_de +tmp=$prep/tmp +orig=orig +dev=dev/newstest2012 +codes=32000 +bpe=bpe.32k + +mkdir -p $orig $tmp $prep $bpe + +cd $orig + +for ((i=0;i<${#URLS[@]};++i)); do + url=${URLS[i]} + file=$(basename $url) + if [ -f $file ]; then + echo "$file already exists, skipping download" + else + wget "$url" + if [ -f $file ]; then + echo "$url successfully downloaded." + else + echo "$url not successfully downloaded." + exit 1 + fi + if [ ${file: -4} == ".tgz" ]; then + tar zxvf $file + elif [ ${file: -4} == ".tar" ]; then + tar xvf $file + fi + fi +done +cd .. + +echo "pre-processing train data..." +for l in $src $tgt; do + rm -rf $tmp/train.tags.$lang.tok.$l + for f in "${CORPORA[@]}"; do + cat $orig/$f.$l | \ + perl $REM_NON_PRINT_CHAR | \ + perl $TOKENIZER -threads 8 -l $l -no-escape >> $tmp/train.tags.$lang.tok.$l + done +done + +echo "pre-processing test data..." +for l in $src $tgt; do + if [ "$l" == "$src" ]; then + t="src" + else + t="ref" + fi + grep '\s*//g' | \ + sed -e 's/\s*<\/seg>\s*//g' | \ + sed -e "s/\’/\'/g" | \ + perl $TOKENIZER -threads 8 -l $l -no-escape > $tmp/test.$l + echo "" +done + +# apply length filtering before BPE +perl $CLEAN -ratio 1.5 $tmp/train.tags.$lang.tok $src $tgt $tmp/train 1 100 + +# use newstest2012 for valid +echo "pre-processing valid data..." +for l in $src $tgt; do + rm -rf $tmp/valid.$l + cat $orig/$dev.$l | \ + perl $REM_NON_PRINT_CHAR | \ + perl $TOKENIZER -threads 8 -l $l -no-escape >> $tmp/valid.$l +done + +mkdir output +mv $tmp/{train,valid,test}.{$src,$tgt} output + +#BPE +git clone https://github.com/glample/fastBPE.git +pushd fastBPE +g++ -std=c++11 -pthread -O3 fastBPE/main.cc -IfastBPE -o fast +popd +fastBPE/fast learnbpe $codes output/train.$src output/train.$tgt > $bpe/codes +for split in {train,valid,test}; do for lang in {en,de}; do fastBPE/fast applybpe $bpe/$split.$lang output/$split.$lang $bpe/codes; done; done diff --git a/examples/language_model/README.adaptive_inputs.md b/examples/language_model/README.adaptive_inputs.md new file mode 100644 index 0000000000000000000000000000000000000000..6650d58f37f320aa46402d59ce6494b2dd1c3faa --- /dev/null +++ b/examples/language_model/README.adaptive_inputs.md @@ -0,0 +1,39 @@ +# Adaptive Input Representations for Neural Language Modeling (Baevski and Auli, 2018) + +## Pre-trained models + +Description | Parameters | Dataset | Model and Test set(s) +---|---:|---|--- +Adaptive Inputs
([Baevski and Auli, 2018](https://arxiv.org/abs/1809.10853)) | 1026M | [Google Billion Words](https://github.com/ciprian-chelba/1-billion-word-language-modeling-benchmark) | [download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/models/lm/adaptive_lm_gbw_huge.tar.bz2) +Adaptive Inputs
([Baevski and Auli, 2018](https://arxiv.org/abs/1809.10853)) | 247M | [WikiText-103](https://blog.einstein.ai/the-wikitext-long-term-dependency-language-modeling-dataset/) | [download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/models/lm/adaptive_lm_wiki103.v2.tar.bz2) + +## Training an LM with adaptive inputs + +First, see the general [language modeling README](README.md) for instructions on +preprocessing the WikiText-103 data. + +Then use the following training command to train a model with adaptive inputs +using the `transformer_lm_wiki103` model architecture: +```bash +fairseq-train --task language_modeling \ + data-bin/wikitext-103 \ + --save-dir checkpoints/transformer_wikitext-103 \ + --arch transformer_lm_wiki103 \ + --max-update 286000 --lr 1.0 --t-mult 2 --lr-period-updates 270000 --lr-scheduler cosine --lr-shrink 0.75 \ + --warmup-updates 16000 --warmup-init-lr 1e-07 --stop-min-lr 1e-09 --optimizer nag --min-lr 0.0001 --clip-norm 0.1 \ + --criterion adaptive_loss --max-tokens 3072 --update-freq 3 --tokens-per-sample 3072 --seed 1 \ + --sample-break-mode none --skip-invalid-size-inputs-valid-test --ddp-backend=legacy_ddp +``` + +## Citation + +```bibtex +@inproceedings{ + baevski2018adaptive, + title={Adaptive Input Representations for Neural Language Modeling}, + author={Alexei Baevski and Michael Auli}, + booktitle={International Conference on Learning Representations}, + year={2019}, + url={https://openreview.net/forum?id=ByxZX20qFQ}, +} +``` diff --git a/examples/language_model/README.conv.md b/examples/language_model/README.conv.md new file mode 100644 index 0000000000000000000000000000000000000000..1ff8635906cf278208be4714e0ef805a6a6b4da1 --- /dev/null +++ b/examples/language_model/README.conv.md @@ -0,0 +1,40 @@ +# Language Modeling with Gated Convolutional Networks (Dauphin et al., 2017) + +## Example usage + +First download and preprocess the data following the main [language modeling README](README.md). + +Then to train a convolutional LM using the `fconv_lm_dauphin_wikitext103` +architecture: +```bash +fairseq-train --task language_modeling \ + data-bin/wikitext-103 \ + --save-dir checkpoints/fconv_wikitext-103 \ + --arch fconv_lm_dauphin_wikitext103 \ + --adaptive-softmax-cutoff 10000,20000,200000 \ + --dropout 0.2 \ + --criterion adaptive_loss \ + --optimizer nag --clip-norm 0.1 --weight-decay 5e-06 \ + --lr 1.0 --lr-scheduler reduce_lr_on_plateau --lr-shrink 0.5 \ + --max-tokens 1024 --tokens-per-sample 1024 \ + --ddp-backend legacy_ddp \ + --max-epoch 35 +``` + +And evaluate with: +```bash +fairseq-eval-lm data-bin/wikitext-103 --path checkpoints/fconv_wiki103/checkpoint_best.pt +``` + +## Citation + +```bibtex +@inproceedings{dauphin2017language, + title={Language Modeling with Gated Convolutional Networks}, + author={Dauphin, Yann N and Fan, Angela and Auli, Michael and Grangier, David}, + booktitle={Proceedings of the 34th International Conference on Machine Learning-Volume 70}, + pages={933--941}, + year={2017}, + organization={JMLR} +} +``` diff --git a/examples/language_model/README.md b/examples/language_model/README.md new file mode 100644 index 0000000000000000000000000000000000000000..e78ea48e08dc99b69751923762107a8f8a9a5e3e --- /dev/null +++ b/examples/language_model/README.md @@ -0,0 +1,123 @@ +# Neural Language Modeling + +## Pre-trained models + +Model | Description | Dataset | Download +---|---|---|--- +`transformer_lm.gbw.adaptive_huge` | Adaptive Inputs
([Baevski and Auli, 2018](https://arxiv.org/abs/1809.10853))
1026M params | [Google Billion Words](https://github.com/ciprian-chelba/1-billion-word-language-modeling-benchmark) | [download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/models/lm/adaptive_lm_gbw_huge.tar.bz2) +`transformer_lm.wiki103.adaptive` | Adaptive Inputs
([Baevski and Auli, 2018](https://arxiv.org/abs/1809.10853))
247M params | [WikiText-103](https://blog.einstein.ai/the-wikitext-long-term-dependency-language-modeling-dataset) | [download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/models/lm/adaptive_lm_wiki103.v2.tar.bz2) +`transformer_lm.wmt19.en` | English LM
([Ng et al., 2019](https://arxiv.org/abs/1907.06616)) | [WMT News Crawl](http://data.statmt.org/news-crawl/) | [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/lm/wmt19.en.tar.gz) +`transformer_lm.wmt19.de` | German LM
([Ng et al., 2019](https://arxiv.org/abs/1907.06616)) | [WMT News Crawl](http://data.statmt.org/news-crawl/) | [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/lm/wmt19.de.tar.gz) +`transformer_lm.wmt19.ru` | Russian LM
([Ng et al., 2019](https://arxiv.org/abs/1907.06616)) | [WMT News Crawl](http://data.statmt.org/news-crawl/) | [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/lm/wmt19.ru.tar.gz) + +## Example usage + +We require a few additional Python dependencies for preprocessing: +```bash +pip install fastBPE sacremoses +``` + +To sample from a language model using PyTorch Hub: +```python +import torch + +# List available models +torch.hub.list('pytorch/fairseq') # [..., 'transformer_lm.wmt19.en', ...] + +# Load an English LM trained on WMT'19 News Crawl data +en_lm = torch.hub.load('pytorch/fairseq', 'transformer_lm.wmt19.en', tokenizer='moses', bpe='fastbpe') +en_lm.eval() # disable dropout + +# Move model to GPU +en_lm.cuda() + +# Sample from the language model +en_lm.sample('Barack Obama', beam=1, sampling=True, sampling_topk=10, temperature=0.8) +# "Barack Obama is coming to Sydney and New Zealand (...)" + +# Compute perplexity for a sequence +en_lm.score('Barack Obama is coming to Sydney and New Zealand')['positional_scores'].mean().neg().exp() +# tensor(15.1474) + +# The same interface can be used with custom models as well +from fairseq.models.transformer_lm import TransformerLanguageModel +custom_lm = TransformerLanguageModel.from_pretrained('/path/to/model/dir', 'checkpoint100.pt', tokenizer='moses', bpe='fastbpe') +custom_lm.sample('Barack Obama', beam=5) +# "Barack Obama (...)" +``` + +## Training a transformer language model with the CLI tools + +### 1) Preprocess the data + +First download and prepare the [WikiText-103 dataset](https://www.salesforce.com/products/einstein/ai-research/the-wikitext-dependency-language-modeling-dataset/): +```bash +cd examples/language_model/ +bash prepare-wikitext-103.sh +cd ../.. +``` + +Next preprocess/binarize the data: +```bash +TEXT=examples/language_model/wikitext-103 +fairseq-preprocess \ + --only-source \ + --trainpref $TEXT/wiki.train.tokens \ + --validpref $TEXT/wiki.valid.tokens \ + --testpref $TEXT/wiki.test.tokens \ + --destdir data-bin/wikitext-103 \ + --workers 20 +``` + +### 2) Train a language model + +Next we'll train a basic transformer language model on wikitext-103. For more +advanced usage, see the [adaptive inputs README](README.adaptive_inputs.md). + +To train a basic LM (assumes 2 GPUs): +``` +$ fairseq-train --task language_modeling \ + data-bin/wikitext-103 \ + --save-dir checkpoints/transformer_wikitext-103 \ + --arch transformer_lm --share-decoder-input-output-embed \ + --dropout 0.1 \ + --optimizer adam --adam-betas '(0.9, 0.98)' --weight-decay 0.01 --clip-norm 0.0 \ + --lr 0.0005 --lr-scheduler inverse_sqrt --warmup-updates 4000 --warmup-init-lr 1e-07 \ + --tokens-per-sample 512 --sample-break-mode none \ + --max-tokens 2048 --update-freq 16 \ + --fp16 \ + --max-update 50000 +``` + +If you run out of memory, try reducing `--max-tokens` (max number of tokens per +batch) or `--tokens-per-sample` (max sequence length). You can also adjust +`--update-freq` to accumulate gradients and simulate training on a different +number of GPUs. + +### 3) Evaluate + +```bash +fairseq-eval-lm data-bin/wikitext-103 \ + --path checkpoints/transformer_wiki103/checkpoint_best.pt \ + --batch-size 2 \ + --tokens-per-sample 512 \ + --context-window 400 +# | Evaluated 245569 tokens in 56.1s (4379.02 tokens/s) +# | Loss: 3.4164, Perplexity: 30.46 +``` + +*Note:* The `--context-window` option controls how much context is provided to +each token when computing perplexity. When the window size is 0, the dataset is +chunked into segments of length 512 and perplexity is computed over each segment +normally. However, this results in worse (higher) perplexity since tokens that +appear earlier in each segment have less conditioning. When the maximum window +size is used (511 in this case), then we compute perplexity for each token +fully conditioned on 511 tokens of context. This slows down evaluation +significantly, since we must run a separate forward pass for every token in the +dataset, but results in better (lower) perplexity. + + +## Convolutional language models + +Please see the [convolutional LM README](README.conv.md) for instructions on +training convolutional language models. diff --git a/examples/language_model/prepare-wikitext-103.sh b/examples/language_model/prepare-wikitext-103.sh new file mode 100644 index 0000000000000000000000000000000000000000..751302156f0a6829af9c2ee5e0e2ca62c2cd4187 --- /dev/null +++ b/examples/language_model/prepare-wikitext-103.sh @@ -0,0 +1,33 @@ +#!/bin/bash +# Adapted from https://github.com/facebookresearch/MIXER/blob/master/prepareData.sh + +URLS=( + "https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-103-v1.zip" +) +FILES=( + "wikitext-103-v1.zip" +) + +for ((i=0;i<${#URLS[@]};++i)); do + file=${FILES[i]} + if [ -f $file ]; then + echo "$file already exists, skipping download" + else + url=${URLS[i]} + wget "$url" + if [ -f $file ]; then + echo "$url successfully downloaded." + else + echo "$url not successfully downloaded." + exit -1 + fi + if [ ${file: -4} == ".tgz" ]; then + tar zxvf $file + elif [ ${file: -4} == ".tar" ]; then + tar xvf $file + elif [ ${file: -4} == ".zip" ]; then + unzip $file + fi + fi +done +cd .. diff --git a/examples/laser/README.md b/examples/laser/README.md new file mode 100644 index 0000000000000000000000000000000000000000..66acada04f58fa235cd312753f144f6f1e5f4a33 --- /dev/null +++ b/examples/laser/README.md @@ -0,0 +1,144 @@ +# LASER Language-Agnostic SEntence Representations + +LASER is a library to calculate and use multilingual sentence embeddings. + +You can find more information about LASER and how to use it on the official [LASER repository](https://github.com/facebookresearch/LASER). + +This folder contains source code for training LASER embeddings. + + +## Prepare data and configuration file + +Binarize your data with fairseq, as described [here](https://fairseq.readthedocs.io/en/latest/getting_started.html#data-pre-processing). + +Create a json config file with this format: +``` +{ + "src_vocab": "/path/to/spm.src.cvocab", + "tgt_vocab": "/path/to/spm.tgt.cvocab", + "train": [ + { + "type": "translation", + "id": 0, + "src": "/path/to/srclang1-tgtlang0/train.srclang1", + "tgt": "/path/to/srclang1-tgtlang0/train.tgtlang0" + }, + { + "type": "translation", + "id": 1, + "src": "/path/to/srclang1-tgtlang1/train.srclang1", + "tgt": "/path/to/srclang1-tgtlang1/train.tgtlang1" + }, + { + "type": "translation", + "id": 0, + "src": "/path/to/srclang2-tgtlang0/train.srclang2", + "tgt": "/path/to/srclang2-tgtlang0/train.tgtlang0" + }, + { + "type": "translation", + "id": 1, + "src": "/path/to/srclang2-tgtlang1/train.srclang2", + "tgt": "/path/to/srclang2-tgtlang1/train.tgtlang1" + }, + ... + ], + "valid": [ + { + "type": "translation", + "id": 0, + "src": "/unused", + "tgt": "/unused" + } + ] +} +``` +where paths are paths to binarized indexed fairseq dataset files. +`id` represents the target language id. + + +## Training Command Line Example + +``` +fairseq-train \ + /path/to/configfile_described_above.json \ + --user-dir examples/laser/laser_src \ + --log-interval 100 --log-format simple \ + --task laser --arch laser_lstm \ + --save-dir . \ + --optimizer adam \ + --lr 0.001 \ + --lr-scheduler inverse_sqrt \ + --clip-norm 5 \ + --warmup-updates 90000 \ + --update-freq 2 \ + --dropout 0.0 \ + --encoder-dropout-out 0.1 \ + --max-tokens 2000 \ + --max-epoch 50 \ + --encoder-bidirectional \ + --encoder-layers 5 \ + --encoder-hidden-size 512 \ + --decoder-layers 1 \ + --decoder-hidden-size 2048 \ + --encoder-embed-dim 320 \ + --decoder-embed-dim 320 \ + --decoder-lang-embed-dim 32 \ + --warmup-init-lr 0.001 \ + --disable-validation +``` + + +## Applications + +We showcase several applications of multilingual sentence embeddings +with code to reproduce our results (in the directory "tasks"). + +* [**Cross-lingual document classification**](https://github.com/facebookresearch/LASER/tree/master/tasks/mldoc) using the + [*MLDoc*](https://github.com/facebookresearch/MLDoc) corpus [2,6] +* [**WikiMatrix**](https://github.com/facebookresearch/LASER/tree/master/tasks/WikiMatrix) + Mining 135M Parallel Sentences in 1620 Language Pairs from Wikipedia [7] +* [**Bitext mining**](https://github.com/facebookresearch/LASER/tree/master/tasks/bucc) using the + [*BUCC*](https://comparable.limsi.fr/bucc2018/bucc2018-task.html) corpus [3,5] +* [**Cross-lingual NLI**](https://github.com/facebookresearch/LASER/tree/master/tasks/xnli) + using the [*XNLI*](https://www.nyu.edu/projects/bowman/xnli/) corpus [4,5,6] +* [**Multilingual similarity search**](https://github.com/facebookresearch/LASER/tree/master/tasks/similarity) [1,6] +* [**Sentence embedding of text files**](https://github.com/facebookresearch/LASER/tree/master/tasks/embed) + example how to calculate sentence embeddings for arbitrary text files in any of the supported language. + +**For all tasks, we use exactly the same multilingual encoder, without any task specific optimization or fine-tuning.** + + + +## References + +[1] Holger Schwenk and Matthijs Douze, + [*Learning Joint Multilingual Sentence Representations with Neural Machine Translation*](https://aclanthology.info/papers/W17-2619/w17-2619), + ACL workshop on Representation Learning for NLP, 2017 + +[2] Holger Schwenk and Xian Li, + [*A Corpus for Multilingual Document Classification in Eight Languages*](http://www.lrec-conf.org/proceedings/lrec2018/pdf/658.pdf), + LREC, pages 3548-3551, 2018. + +[3] Holger Schwenk, + [*Filtering and Mining Parallel Data in a Joint Multilingual Space*](http://aclweb.org/anthology/P18-2037) + ACL, July 2018 + +[4] Alexis Conneau, Guillaume Lample, Ruty Rinott, Adina Williams, Samuel R. Bowman, Holger Schwenk and Veselin Stoyanov, + [*XNLI: Cross-lingual Sentence Understanding through Inference*](https://aclweb.org/anthology/D18-1269), + EMNLP, 2018. + +[5] Mikel Artetxe and Holger Schwenk, + [*Margin-based Parallel Corpus Mining with Multilingual Sentence Embeddings*](https://arxiv.org/abs/1811.01136) + arXiv, Nov 3 2018. + +[6] Mikel Artetxe and Holger Schwenk, + [*Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond*](https://arxiv.org/abs/1812.10464) + arXiv, Dec 26 2018. + +[7] Holger Schwenk, Vishrav Chaudhary, Shuo Sun, Hongyu Gong and Paco Guzman, + [*WikiMatrix: Mining 135M Parallel Sentences in 1620 Language Pairs from Wikipedia*](https://arxiv.org/abs/1907.05791) + arXiv, July 11 2019. + +[8] Holger Schwenk, Guillaume Wenzek, Sergey Edunov, Edouard Grave and Armand Joulin + [*CCMatrix: Mining Billions of High-Quality Parallel Sentences on the WEB*](https://arxiv.org/abs/1911.04944) diff --git a/examples/laser/laser_src/__init__.py b/examples/laser/laser_src/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..9ffbd656d8786e421008fb4cb0d1d8911dc8330c --- /dev/null +++ b/examples/laser/laser_src/__init__.py @@ -0,0 +1,8 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .laser_task import * # noqa +from .laser_lstm import * # noqa +from .laser_transformer import * # noqa diff --git a/examples/laser/laser_src/laser_lstm.py b/examples/laser/laser_src/laser_lstm.py new file mode 100644 index 0000000000000000000000000000000000000000..10df90e002d5a7dd74a571dbc3b328c130c57a0a --- /dev/null +++ b/examples/laser/laser_src/laser_lstm.py @@ -0,0 +1,585 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn as nn +import torch.nn.functional as F + +from fairseq import options, utils + +from fairseq.models import ( + FairseqEncoder, + FairseqIncrementalDecoder, + FairseqEncoderDecoderModel, + register_model, + register_model_architecture, +) + + +@register_model("laser_lstm") +class LSTMModel(FairseqEncoderDecoderModel): + def __init__(self, encoder, decoder): + super().__init__(encoder, decoder) + + def forward( + self, + src_tokens, + src_lengths, + prev_output_tokens=None, + tgt_tokens=None, + tgt_lengths=None, + target_language_id=None, + dataset_name="", + ): + assert target_language_id is not None + + src_encoder_out = self.encoder(src_tokens, src_lengths, dataset_name) + return self.decoder( + prev_output_tokens, src_encoder_out, lang_id=target_language_id + ) + + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + parser.add_argument( + "--dropout", + default=0.1, + type=float, + metavar="D", + help="dropout probability", + ) + parser.add_argument( + "--encoder-embed-dim", + type=int, + metavar="N", + help="encoder embedding dimension", + ) + parser.add_argument( + "--encoder-embed-path", + default=None, + type=str, + metavar="STR", + help="path to pre-trained encoder embedding", + ) + parser.add_argument( + "--encoder-hidden-size", type=int, metavar="N", help="encoder hidden size" + ) + parser.add_argument( + "--encoder-layers", type=int, metavar="N", help="number of encoder layers" + ) + parser.add_argument( + "--encoder-bidirectional", + action="store_true", + help="make all layers of encoder bidirectional", + ) + parser.add_argument( + "--decoder-embed-dim", + type=int, + metavar="N", + help="decoder embedding dimension", + ) + parser.add_argument( + "--decoder-embed-path", + default=None, + type=str, + metavar="STR", + help="path to pre-trained decoder embedding", + ) + parser.add_argument( + "--decoder-hidden-size", type=int, metavar="N", help="decoder hidden size" + ) + parser.add_argument( + "--decoder-layers", type=int, metavar="N", help="number of decoder layers" + ) + parser.add_argument( + "--decoder-out-embed-dim", + type=int, + metavar="N", + help="decoder output embedding dimension", + ) + parser.add_argument( + "--decoder-zero-init", + type=str, + metavar="BOOL", + help="initialize the decoder hidden/cell state to zero", + ) + parser.add_argument( + "--decoder-lang-embed-dim", + type=int, + metavar="N", + help="decoder language embedding dimension", + ) + parser.add_argument( + "--fixed-embeddings", + action="store_true", + help="keep embeddings fixed (ENCODER ONLY)", + ) # TODO Also apply to decoder embeddings? + + # Granular dropout settings (if not specified these default to --dropout) + parser.add_argument( + "--encoder-dropout-in", + type=float, + metavar="D", + help="dropout probability for encoder input embedding", + ) + parser.add_argument( + "--encoder-dropout-out", + type=float, + metavar="D", + help="dropout probability for encoder output", + ) + parser.add_argument( + "--decoder-dropout-in", + type=float, + metavar="D", + help="dropout probability for decoder input embedding", + ) + parser.add_argument( + "--decoder-dropout-out", + type=float, + metavar="D", + help="dropout probability for decoder output", + ) + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + # make sure that all args are properly defaulted (in case there are any new ones) + base_architecture(args) + + def load_pretrained_embedding_from_file(embed_path, dictionary, embed_dim): + num_embeddings = len(dictionary) + padding_idx = dictionary.pad() + embed_tokens = Embedding(num_embeddings, embed_dim, padding_idx) + embed_dict = utils.parse_embedding(embed_path) + utils.print_embed_overlap(embed_dict, dictionary) + return utils.load_embedding(embed_dict, dictionary, embed_tokens) + + pretrained_encoder_embed = None + if args.encoder_embed_path: + pretrained_encoder_embed = load_pretrained_embedding_from_file( + args.encoder_embed_path, task.source_dictionary, args.encoder_embed_dim + ) + pretrained_decoder_embed = None + if args.decoder_embed_path: + pretrained_decoder_embed = load_pretrained_embedding_from_file( + args.decoder_embed_path, task.target_dictionary, args.decoder_embed_dim + ) + + num_langs = task.num_tasks if hasattr(task, "num_tasks") else 0 + + encoder = LSTMEncoder( + dictionary=task.source_dictionary, + embed_dim=args.encoder_embed_dim, + hidden_size=args.encoder_hidden_size, + num_layers=args.encoder_layers, + dropout_in=args.encoder_dropout_in, + dropout_out=args.encoder_dropout_out, + bidirectional=args.encoder_bidirectional, + pretrained_embed=pretrained_encoder_embed, + fixed_embeddings=args.fixed_embeddings, + ) + decoder = LSTMDecoder( + dictionary=task.target_dictionary, + embed_dim=args.decoder_embed_dim, + hidden_size=args.decoder_hidden_size, + out_embed_dim=args.decoder_out_embed_dim, + num_layers=args.decoder_layers, + dropout_in=args.decoder_dropout_in, + dropout_out=args.decoder_dropout_out, + zero_init=options.eval_bool(args.decoder_zero_init), + encoder_embed_dim=args.encoder_embed_dim, + encoder_output_units=encoder.output_units, + pretrained_embed=pretrained_decoder_embed, + num_langs=num_langs, + lang_embed_dim=args.decoder_lang_embed_dim, + ) + return cls(encoder, decoder) + + +class LSTMEncoder(FairseqEncoder): + """LSTM encoder.""" + + def __init__( + self, + dictionary, + embed_dim=512, + hidden_size=512, + num_layers=1, + dropout_in=0.1, + dropout_out=0.1, + bidirectional=False, + left_pad=True, + pretrained_embed=None, + padding_value=0.0, + fixed_embeddings=False, + ): + super().__init__(dictionary) + self.num_layers = num_layers + self.dropout_in = dropout_in + self.dropout_out = dropout_out + self.bidirectional = bidirectional + self.hidden_size = hidden_size + + num_embeddings = len(dictionary) + self.padding_idx = dictionary.pad() + if pretrained_embed is None: + self.embed_tokens = Embedding(num_embeddings, embed_dim, self.padding_idx) + else: + self.embed_tokens = pretrained_embed + if fixed_embeddings: + self.embed_tokens.weight.requires_grad = False + + self.lstm = LSTM( + input_size=embed_dim, + hidden_size=hidden_size, + num_layers=num_layers, + dropout=self.dropout_out if num_layers > 1 else 0.0, + bidirectional=bidirectional, + ) + self.left_pad = left_pad + self.padding_value = padding_value + + self.output_units = hidden_size + if bidirectional: + self.output_units *= 2 + + def forward(self, src_tokens, src_lengths, dataset_name): + if self.left_pad: + # convert left-padding to right-padding + src_tokens = utils.convert_padding_direction( + src_tokens, + self.padding_idx, + left_to_right=True, + ) + + bsz, seqlen = src_tokens.size() + + # embed tokens + x = self.embed_tokens(src_tokens) + x = F.dropout(x, p=self.dropout_in, training=self.training) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + + # pack embedded source tokens into a PackedSequence + try: + packed_x = nn.utils.rnn.pack_padded_sequence(x, src_lengths.data.tolist()) + except BaseException: + raise Exception(f"Packing failed in dataset {dataset_name}") + + # apply LSTM + if self.bidirectional: + state_size = 2 * self.num_layers, bsz, self.hidden_size + else: + state_size = self.num_layers, bsz, self.hidden_size + h0 = x.data.new(*state_size).zero_() + c0 = x.data.new(*state_size).zero_() + packed_outs, (final_hiddens, final_cells) = self.lstm(packed_x, (h0, c0)) + + # unpack outputs and apply dropout + x, _ = nn.utils.rnn.pad_packed_sequence( + packed_outs, padding_value=self.padding_value + ) + x = F.dropout(x, p=self.dropout_out, training=self.training) + assert list(x.size()) == [seqlen, bsz, self.output_units] + + if self.bidirectional: + + def combine_bidir(outs): + return torch.cat( + [ + torch.cat([outs[2 * i], outs[2 * i + 1]], dim=0).view( + 1, bsz, self.output_units + ) + for i in range(self.num_layers) + ], + dim=0, + ) + + final_hiddens = combine_bidir(final_hiddens) + final_cells = combine_bidir(final_cells) + + encoder_padding_mask = src_tokens.eq(self.padding_idx).t() + + # Set padded outputs to -inf so they are not selected by max-pooling + padding_mask = src_tokens.eq(self.padding_idx).t().unsqueeze(-1) + if padding_mask.any(): + x = x.float().masked_fill_(padding_mask, float("-inf")).type_as(x) + + # Build the sentence embedding by max-pooling over the encoder outputs + sentemb = x.max(dim=0)[0] + + return { + "sentemb": sentemb, + "encoder_out": (x, final_hiddens, final_cells), + "encoder_padding_mask": encoder_padding_mask + if encoder_padding_mask.any() + else None, + } + + def reorder_encoder_out(self, encoder_out_dict, new_order): + encoder_out_dict["sentemb"] = encoder_out_dict["sentemb"].index_select( + 0, new_order + ) + encoder_out_dict["encoder_out"] = tuple( + eo.index_select(1, new_order) for eo in encoder_out_dict["encoder_out"] + ) + if encoder_out_dict["encoder_padding_mask"] is not None: + encoder_out_dict["encoder_padding_mask"] = encoder_out_dict[ + "encoder_padding_mask" + ].index_select(1, new_order) + return encoder_out_dict + + def max_positions(self): + """Maximum input length supported by the encoder.""" + return int(1e5) # an arbitrary large number + + +class LSTMDecoder(FairseqIncrementalDecoder): + """LSTM decoder.""" + + def __init__( + self, + dictionary, + embed_dim=512, + hidden_size=512, + out_embed_dim=512, + num_layers=1, + dropout_in=0.1, + dropout_out=0.1, + zero_init=False, + encoder_embed_dim=512, + encoder_output_units=512, + pretrained_embed=None, + num_langs=1, + lang_embed_dim=0, + ): + super().__init__(dictionary) + self.dropout_in = dropout_in + self.dropout_out = dropout_out + self.hidden_size = hidden_size + + num_embeddings = len(dictionary) + padding_idx = dictionary.pad() + if pretrained_embed is None: + self.embed_tokens = Embedding(num_embeddings, embed_dim, padding_idx) + else: + self.embed_tokens = pretrained_embed + + self.layers = nn.ModuleList( + [ + LSTMCell( + input_size=encoder_output_units + embed_dim + lang_embed_dim + if layer == 0 + else hidden_size, + hidden_size=hidden_size, + ) + for layer in range(num_layers) + ] + ) + if hidden_size != out_embed_dim: + self.additional_fc = Linear(hidden_size, out_embed_dim) + self.fc_out = Linear(out_embed_dim, num_embeddings, dropout=dropout_out) + + if zero_init: + self.sentemb2init = None + else: + self.sentemb2init = Linear( + encoder_output_units, 2 * num_layers * hidden_size + ) + + if lang_embed_dim == 0: + self.embed_lang = None + else: + self.embed_lang = nn.Embedding(num_langs, lang_embed_dim) + nn.init.uniform_(self.embed_lang.weight, -0.1, 0.1) + + def forward( + self, prev_output_tokens, encoder_out_dict, incremental_state=None, lang_id=0 + ): + sentemb = encoder_out_dict["sentemb"] + encoder_out = encoder_out_dict["encoder_out"] + + if incremental_state is not None: + prev_output_tokens = prev_output_tokens[:, -1:] + bsz, seqlen = prev_output_tokens.size() + + # get outputs from encoder + encoder_outs, _, _ = encoder_out[:3] + srclen = encoder_outs.size(0) + + # embed tokens + x = self.embed_tokens(prev_output_tokens) + x = F.dropout(x, p=self.dropout_in, training=self.training) + + # embed language identifier + if self.embed_lang is not None: + lang_ids = prev_output_tokens.data.new_full((bsz,), lang_id) + langemb = self.embed_lang(lang_ids) + # TODO Should we dropout here??? + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + + # initialize previous states (or get from cache during incremental generation) + cached_state = utils.get_incremental_state( + self, incremental_state, "cached_state" + ) + if cached_state is not None: + prev_hiddens, prev_cells, input_feed = cached_state + else: + num_layers = len(self.layers) + if self.sentemb2init is None: + prev_hiddens = [ + x.data.new(bsz, self.hidden_size).zero_() for i in range(num_layers) + ] + prev_cells = [ + x.data.new(bsz, self.hidden_size).zero_() for i in range(num_layers) + ] + else: + init = self.sentemb2init(sentemb) + prev_hiddens = [ + init[:, (2 * i) * self.hidden_size : (2 * i + 1) * self.hidden_size] + for i in range(num_layers) + ] + prev_cells = [ + init[ + :, + (2 * i + 1) * self.hidden_size : (2 * i + 2) * self.hidden_size, + ] + for i in range(num_layers) + ] + input_feed = x.data.new(bsz, self.hidden_size).zero_() + + attn_scores = x.data.new(srclen, seqlen, bsz).zero_() + outs = [] + for j in range(seqlen): + if self.embed_lang is None: + input = torch.cat((x[j, :, :], sentemb), dim=1) + else: + input = torch.cat((x[j, :, :], sentemb, langemb), dim=1) + + for i, rnn in enumerate(self.layers): + # recurrent cell + hidden, cell = rnn(input, (prev_hiddens[i], prev_cells[i])) + + # hidden state becomes the input to the next layer + input = F.dropout(hidden, p=self.dropout_out, training=self.training) + + # save state for next time step + prev_hiddens[i] = hidden + prev_cells[i] = cell + + out = hidden + out = F.dropout(out, p=self.dropout_out, training=self.training) + + # input feeding + input_feed = out + + # save final output + outs.append(out) + + # cache previous states (no-op except during incremental generation) + utils.set_incremental_state( + self, + incremental_state, + "cached_state", + (prev_hiddens, prev_cells, input_feed), + ) + + # collect outputs across time steps + x = torch.cat(outs, dim=0).view(seqlen, bsz, self.hidden_size) + + # T x B x C -> B x T x C + x = x.transpose(1, 0) + + # srclen x tgtlen x bsz -> bsz x tgtlen x srclen + attn_scores = attn_scores.transpose(0, 2) + + # project back to size of vocabulary + if hasattr(self, "additional_fc"): + x = self.additional_fc(x) + x = F.dropout(x, p=self.dropout_out, training=self.training) + x = self.fc_out(x) + + return x, attn_scores + + def reorder_incremental_state(self, incremental_state, new_order): + super().reorder_incremental_state(incremental_state, new_order) + cached_state = utils.get_incremental_state( + self, incremental_state, "cached_state" + ) + if cached_state is None: + return + + def reorder_state(state): + if isinstance(state, list): + return [reorder_state(state_i) for state_i in state] + return state.index_select(0, new_order) + + new_state = tuple(map(reorder_state, cached_state)) + utils.set_incremental_state(self, incremental_state, "cached_state", new_state) + + def max_positions(self): + """Maximum output length supported by the decoder.""" + return int(1e5) # an arbitrary large number + + +def Embedding(num_embeddings, embedding_dim, padding_idx): + m = nn.Embedding(num_embeddings, embedding_dim, padding_idx=padding_idx) + nn.init.uniform_(m.weight, -0.1, 0.1) + nn.init.constant_(m.weight[padding_idx], 0) + return m + + +def LSTM(input_size, hidden_size, **kwargs): + m = nn.LSTM(input_size, hidden_size, **kwargs) + for name, param in m.named_parameters(): + if "weight" in name or "bias" in name: + param.data.uniform_(-0.1, 0.1) + return m + + +def LSTMCell(input_size, hidden_size, **kwargs): + m = nn.LSTMCell(input_size, hidden_size, **kwargs) + for name, param in m.named_parameters(): + if "weight" in name or "bias" in name: + param.data.uniform_(-0.1, 0.1) + return m + + +def Linear(in_features, out_features, bias=True, dropout=0): + """Weight-normalized Linear layer (input: N x T x C)""" + m = nn.Linear(in_features, out_features, bias=bias) + m.weight.data.uniform_(-0.1, 0.1) + if bias: + m.bias.data.uniform_(-0.1, 0.1) + return m + + +@register_model_architecture("laser_lstm", "laser_lstm") +def base_architecture(args): + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512) + args.encoder_embed_path = getattr(args, "encoder_embed_path", None) + args.encoder_hidden_size = getattr( + args, "encoder_hidden_size", args.encoder_embed_dim + ) + args.encoder_layers = getattr(args, "encoder_layers", 1) + args.encoder_bidirectional = getattr(args, "encoder_bidirectional", False) + args.encoder_dropout_in = getattr(args, "encoder_dropout_in", args.dropout) + args.encoder_dropout_out = getattr(args, "encoder_dropout_out", args.dropout) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 512) + args.decoder_embed_path = getattr(args, "decoder_embed_path", None) + args.decoder_hidden_size = getattr( + args, "decoder_hidden_size", args.decoder_embed_dim + ) + args.decoder_layers = getattr(args, "decoder_layers", 1) + args.decoder_out_embed_dim = getattr(args, "decoder_out_embed_dim", 512) + args.decoder_dropout_in = getattr(args, "decoder_dropout_in", args.dropout) + args.decoder_dropout_out = getattr(args, "decoder_dropout_out", args.dropout) + args.decoder_zero_init = getattr(args, "decoder_zero_init", "0") + args.decoder_lang_embed_dim = getattr(args, "decoder_lang_embed_dim", 0) + args.fixed_embeddings = getattr(args, "fixed_embeddings", False) diff --git a/examples/laser/laser_src/laser_task.py b/examples/laser/laser_src/laser_task.py new file mode 100644 index 0000000000000000000000000000000000000000..c8ac805f540030802e36360abcfc036a9c6f5427 --- /dev/null +++ b/examples/laser/laser_src/laser_task.py @@ -0,0 +1,326 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +from collections import OrderedDict, defaultdict +import json +import os +import logging + +from fairseq import options, models +from fairseq.data import ( + data_utils, + Dictionary, + LanguagePairDataset, + IndexedDataset, + FairseqDataset, +) +from .multitask_data_utils import ( + MultitaskDatasetWrapper, + MultidatasetEpochBatchIterator, +) + + +from fairseq.tasks import LegacyFairseqTask, register_task + +logger = logging.getLogger(__name__) + + +@register_task("laser") +class LaserTask(LegacyFairseqTask): + @staticmethod + def add_args(parser): + """Add task-specific arguments to the parser.""" + parser.add_argument( + "configfile", metavar="PATH", help="dataset configuration file in json" + ) + parser.add_argument( + "--weighting-alpha", + type=float, + default=None, + help="alpha for automatic weighting", + ) + parser.add_argument( + "--raw-text", action="store_true", help="load raw text dataset" + ) + parser.add_argument( + "--left-pad-source", + default="True", + type=str, + metavar="BOOL", + help="pad the source on the left (default: True)", + ) + parser.add_argument( + "--left-pad-target", + default="False", + type=str, + metavar="BOOL", + help="pad the target on the left (default: False)", + ) + parser.add_argument( + "--max-source-positions", + default=1024, + type=int, + metavar="N", + help="max number of tokens in the source sequence", + ) + parser.add_argument( + "--max-target-positions", + default=1024, + type=int, + metavar="N", + help="max number of tokens in the target sequence", + ) + + def __init__(self, args, config, src_dictionary, tgt_dictionary, num_tasks): + super().__init__(args) + self.config = config + self.src_dictionary = src_dictionary + self.tgt_dictionary = tgt_dictionary + self.num_tasks = num_tasks + + @classmethod + def setup_task(cls, args, **kwargs): + with open(args.configfile, "r") as f: + config = json.load(f) + num_tasks = max(dataset["id"] for dataset in config["train"]) + 1 + + args.left_pad_source = options.eval_bool(args.left_pad_source) + args.left_pad_target = options.eval_bool(args.left_pad_target) + + src_dictionary = Dictionary.load(config["src_vocab"]) + tgt_dictionary = Dictionary.load(config["tgt_vocab"]) + + logger.info( + "| src Dictionary {} : {} types".format( + config["src_vocab"], len(src_dictionary) + ) + ) + logger.info( + "| tgt Dictionary {} : {} types".format( + config["tgt_vocab"], len(tgt_dictionary) + ) + ) + + return cls(args, config, src_dictionary, tgt_dictionary, num_tasks) + + # Experimental overriding for backtranslation + def build_model(self, args): + model = models.build_model(args, self) + return model + + def dataset(self, split): + if split not in self.datasets: + raise KeyError("Dataset not loaded: " + split) + return self.datasets[split] + + def load_dataset(self, split, epoch=1, **kwargs): + """Load a dataset split.""" + + def indexed_dataset(path, dictionary): + if self.args.raw_text: + raise Exception("Unable to handle raw text.") + dataset = IndexedDataset(path, fix_lua_indexing=True) + + return dataset + + pair_datasets = OrderedDict() + + if split == "valid": + self.datasets[split] = pair_datasets + return + + if split not in self.config: + raise FileNotFoundError( + "Dataset not found in config file: {}".format(split) + ) + + size_by_corpus = defaultdict(int) + size_sum = 0 + size_sum_with_subsampling = 0 + init_pair_datasets = {} + + for dataset_config in self.config[split]: + src_path = os.path.dirname(dataset_config["src"]) + corpus_name = src_path.split("/")[-2] + language_pair_name = src_path.split("/")[-1] + pair_datasets_key = corpus_name + "-" + language_pair_name + + logger.info(f"loading... {pair_datasets_key}") + if "src" in dataset_config: + src_dataset = indexed_dataset( + dataset_config["src"], self.src_dictionary + ) + else: + src_dataset = None + + if "tgt" in dataset_config: + tgt_dataset = indexed_dataset( + dataset_config["tgt"], self.tgt_dictionary + ) + else: + tgt_dataset = None + + dataset = LanguagePairDataset( + src_dataset, + src_dataset.sizes, + self.src_dictionary, + tgt_dataset, + tgt_dataset.sizes, + self.tgt_dictionary, + left_pad_source=self.args.left_pad_source, + left_pad_target=self.args.left_pad_target, + ) + + if pair_datasets_key in init_pair_datasets: + logger.warning( + f"Ignoring already added {pair_datasets_key}. " + f"Consider using `sample` key in order to upsample." + ) + else: + init_pair_datasets[pair_datasets_key] = { + "dataset": dataset, + "sample": dataset_config.get("sample", None), + "id": dataset_config.get("id", None), + "len": len(dataset), + } + + length_sum = 0 + weighted_freqs_sum = 0 + freq_per_dataset = {} + vmax = 0 + vmin = 1 + weighted_freq_per_dataset = {} + + if self.args.weighting_alpha: + for key in init_pair_datasets: + if init_pair_datasets[key]["sample"] is None: + length_sum += len(init_pair_datasets[key]["dataset"]) + + for key in init_pair_datasets: + if init_pair_datasets[key]["sample"] is None: + val = float(init_pair_datasets[key]["len"]) / length_sum + freq_per_dataset[key] = val + weighted_freqs_sum += val ** self.args.weighting_alpha + + for key in freq_per_dataset: + val = ( + freq_per_dataset[key] ** self.args.weighting_alpha + / weighted_freqs_sum + ) + vmin = min(vmin, val) + vmax = max(vmax, val) + weighted_freq_per_dataset[key] = val + + for pair_datasets_key in init_pair_datasets: + dataset_config = init_pair_datasets[pair_datasets_key] + dataset = dataset_config["dataset"] + sample = dataset_config["sample"] + if sample is None: + sample = 1.0 + + if pair_datasets_key in weighted_freq_per_dataset: + w = vmax / weighted_freq_per_dataset[pair_datasets_key] + sample = w + + sample = round(sample) + + initial_sample = sample + initial_pair_datasets_key = pair_datasets_key + + while sample >= 1.0: + assert ( + pair_datasets_key not in pair_datasets + ), f"{pair_datasets_key} already in" + size_sum_with_subsampling += len(dataset) + pair_datasets[pair_datasets_key] = MultitaskDatasetWrapper( + dataset, dataset_config.get("id", 0), 1.0, name=pair_datasets_key + ) + size_sum += len(dataset) + sample -= 1.0 + pair_datasets_key += "-up" + + assert sample < 1e-6, f"sample remains > 0 {pair_datasets_key}" + + logger.info( + f"added pair {initial_pair_datasets_key} length {len(dataset)} new_length = {len(dataset)*initial_sample}" + ) + size_by_corpus[corpus_name] += len(dataset) + + self.datasets[split] = pair_datasets + logger.info( + f"Datasets number = {len(self.datasets[split])} size = {size_sum} size_sum_with_subsampling = {size_sum_with_subsampling}" + ) + + @property + def source_dictionary(self): + return self.src_dictionary + + @property + def target_dictionary(self): + return self.tgt_dictionary + + def get_batch_iterator( + self, + dataset, + max_tokens=None, + max_sentences=None, + max_positions=None, + ignore_invalid_inputs=False, + required_batch_size_multiple=1, + seed=1, + num_shards=1, + shard_id=0, + num_workers=0, + epoch=1, + data_buffer_size=0, + disable_iterator_cache=False, + ): + + assert isinstance(dataset, OrderedDict) + assert len(dataset) + assert isinstance(dataset[next(iter(dataset))], FairseqDataset) + + # initialize the dataset with the correct starting epoch + for _, dt in dataset.items(): + dt.set_epoch(epoch) + + indices = OrderedDict() + batch_sampler = OrderedDict() + + with data_utils.numpy_seed(seed + epoch): + for key, dt in dataset.items(): + logger.info(f"\t ordered_indices {key}") + indices[key] = dt.ordered_indices() + + # filter examples that are too large + if max_positions is not None: + for key, dt in dataset.items(): + logger.info(f"\t filter_by_size {key}") + indices[key], ignored = dt.filter_indices_by_size( + indices[key], max_positions + ) + + for key, dt in dataset.items(): + logger.info(f"\t batch_by_size {key}") + batch_sampler[key] = data_utils.batch_by_size( + indices[key], + dt.num_tokens, + max_tokens=max_tokens, + max_sentences=max_sentences, + required_batch_size_multiple=required_batch_size_multiple, + ) + + epoch_iter = MultidatasetEpochBatchIterator( + dataset=dataset, + batch_sampler=batch_sampler, + seed=seed, + num_shards=num_shards, + shard_id=shard_id, + num_workers=num_workers, + epoch=epoch, + ) + + return epoch_iter diff --git a/examples/laser/laser_src/laser_transformer.py b/examples/laser/laser_src/laser_transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..0be030994ff87334ca0392302374693f7f2c61b3 --- /dev/null +++ b/examples/laser/laser_src/laser_transformer.py @@ -0,0 +1,354 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging + +from typing import Any, Dict, List, Optional +from torch import Tensor + +import torch +import torch.nn as nn + +from fairseq.models import ( + FairseqEncoderDecoderModel, + register_model, + register_model_architecture, +) +from fairseq.models.transformer import ( + base_architecture, + Embedding, + TransformerModel, + TransformerEncoder, + TransformerDecoder, +) +from fairseq.modules import ( + TransformerDecoderLayer, +) + +logger = logging.getLogger(__name__) + + +@register_model("laser_transformer") +class LaserTransformerModel(FairseqEncoderDecoderModel): + """Train Transformer for LASER task + + Requires --task laser + """ + + def __init__(self, encoder, decoder): + super().__init__(encoder, decoder) + + def forward( + self, + src_tokens, + src_lengths, + prev_output_tokens=None, + tgt_tokens=None, + tgt_lengths=None, + target_language_id=-1, + dataset_name="", + ): + laser_encoder_out = self.encoder(src_tokens, src_lengths) + return self.decoder( + prev_output_tokens, laser_encoder_out, lang_id=target_language_id + ) + + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + TransformerModel.add_args(parser) + parser.add_argument( + "--decoder-lang-embed-dim", + type=int, + metavar="N", + help="decoder language embedding dimension", + ) + + @classmethod + def build_model(cls, args, task): + base_laser_transformer_architecture(args) + + num_langs = task.num_tasks if hasattr(task, "num_tasks") else 0 + + def load_embed_tokens(dictionary, embed_dim): + num_embeddings = len(dictionary) + padding_idx = dictionary.pad() + + return Embedding(num_embeddings, embed_dim, padding_idx) + + encoder_embed_tokens = load_embed_tokens( + task.source_dictionary, args.encoder_embed_dim + ) + decoder_embed_tokens = load_embed_tokens( + task.target_dictionary, args.decoder_embed_dim + ) + num_langs = task.num_tasks if hasattr(task, "num_tasks") else 0 + + encoder = LaserTransformerEncoder( + args, task.source_dictionary, encoder_embed_tokens + ) + + decoder = LaserTransformerDecoder( + args, + task.target_dictionary, + decoder_embed_tokens, + num_langs=num_langs, + lang_embed_dim=args.decoder_lang_embed_dim, + ) + + return cls(encoder, decoder) + + +class LaserTransformerEncoder(TransformerEncoder): + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + + def forward(self, src_tokens, *args, **kwargs): + encoder_out = super().forward(src_tokens, *args, **kwargs) + + x = encoder_out["encoder_out"][0] # T x B x C + padding_mask = src_tokens.eq(self.padding_idx).t().unsqueeze(-1) + + if padding_mask.any(): + x = x.float().masked_fill_(padding_mask, float("-inf")).type_as(x) + + # Build the sentence embedding by max-pooling over the encoder outputs + sentemb = x.max(dim=0)[0] + + # The Pytorch Mobile lite interpreter does not supports returning NamedTuple in + # `foward` so we use a dictionary instead. + # TorchScript does not support mixed values so the values are all lists. + # The empty list is equivalent to None. + return {"sentemb": [sentemb]} # B x C + + @torch.jit.export + def reorder_encoder_out(self, encoder_out: Dict[str, List[Tensor]], new_order): + """ + Same as the one in transformer.py, with new_sentemb + """ + if len(encoder_out["sentemb"]) == 0: + new_sentemb = [] + else: + new_sentemb = [encoder_out["sentemb"][0].index_select(0, new_order)] + + return { + "sentemb": new_sentemb, # B x C + } + + +class LaserTransformerDecoder(TransformerDecoder): + def __init__(self, args, dictionary, *kargs, **kwargs): + self.num_langs = kwargs.get("num_langs", 1) + self.lang_embed_dim = kwargs.get("lang_embed_dim", 0) + kwargs.pop("num_langs", None) + kwargs.pop("lang_embed_dim", None) + + super().__init__(args, dictionary, *kargs, **kwargs, no_encoder_attn=True) + + if self.lang_embed_dim == 0: + self.embed_lang = None + else: + self.embed_lang = nn.Embedding(self.num_langs, self.lang_embed_dim) + nn.init.uniform_(self.embed_lang.weight, -0.1, 0.1) + + if self.output_projection is not None: + laser_output_embed_dim = ( + self.output_embed_dim + self.lang_embed_dim + args.encoder_embed_dim + ) + self.output_projection = nn.Linear( + laser_output_embed_dim, len(dictionary), bias=False + ) + nn.init.normal_( + self.output_projection.weight, + mean=0, + std=laser_output_embed_dim ** -0.5, + ) + + def build_decoder_layer(self, args, no_encoder_attn=False): + decoder_embed_dim = args.decoder_embed_dim + args.decoder_embed_dim = ( + decoder_embed_dim + self.lang_embed_dim + args.encoder_embed_dim + ) + res = TransformerDecoderLayer(args, no_encoder_attn=True) + args.decoder_embed_dim = decoder_embed_dim + + return res + + def extract_features( + self, + prev_output_tokens, + encoder_out: Optional[Dict[str, List[Tensor]]], + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + full_context_alignment: bool = False, + alignment_layer: Optional[int] = None, + alignment_heads: Optional[int] = None, + lang_id: Optional[int] = None, + ): + """ + Similar to *forward* but only return features. + + Includes several features from "Jointly Learning to Align and + Translate with Transformer Models" (Garg et al., EMNLP 2019). + + Args: + full_context_alignment (bool, optional): don't apply + auto-regressive mask to self-attention (default: False). + alignment_layer (int, optional): return mean alignment over + heads at this layer (default: last layer). + alignment_heads (int, optional): only average alignment over + this many heads (default: all heads). + + Returns: + tuple: + - the decoder's features of shape `(batch, tgt_len, embed_dim)` + - a dictionary with any model-specific outputs + """ + if alignment_layer is None: + alignment_layer = self.num_layers - 1 + + # embed positions + positions = ( + self.embed_positions( + prev_output_tokens, incremental_state=incremental_state + ) + if self.embed_positions is not None + else None + ) + + if incremental_state is not None: + prev_output_tokens = prev_output_tokens[:, -1:] + if positions is not None: + positions = positions[:, -1:] + + bsz, seqlen = prev_output_tokens.size() + + # embed tokens and positions + x = self.embed_scale * self.embed_tokens(prev_output_tokens) + + if self.quant_noise is not None: + x = self.quant_noise(x) + + if self.project_in_dim is not None: + x = self.project_in_dim(x) + + if positions is not None: + x += positions + + if self.layernorm_embedding is not None: + x = self.layernorm_embedding(x) + + x = self.dropout_module(x) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + + if self.embed_lang is not None: + lang_ids = prev_output_tokens.data.new_full((bsz,), lang_id) + langemb = self.embed_lang(lang_ids) + langemb = langemb.unsqueeze(0) + repeat_vals = [x.shape[0] // langemb.shape[0]] + [-1] * ( + len(langemb.shape) - 1 + ) + x = torch.cat((x, langemb.expand(*repeat_vals)), dim=-1) + + sentemb = encoder_out["sentemb"][0] + sentemb = sentemb.unsqueeze(0) + + repeat_vals = [x.shape[0] // sentemb.shape[0]] + [-1] * (len(sentemb.shape) - 1) + x = torch.cat((x, sentemb.expand(*repeat_vals)), dim=-1) + + self_attn_padding_mask: Optional[Tensor] = None + if self.cross_self_attention or prev_output_tokens.eq(self.padding_idx).any(): + self_attn_padding_mask = prev_output_tokens.eq(self.padding_idx) + + # decoder layers + attn: Optional[Tensor] = None + inner_states: List[Optional[Tensor]] = [x] + for idx, layer in enumerate(self.layers): + if incremental_state is None and not full_context_alignment: + self_attn_mask = self.buffered_future_mask(x) + else: + self_attn_mask = None + + x, layer_attn, _ = layer( + x, + None, + None, + incremental_state, + self_attn_mask=self_attn_mask, + self_attn_padding_mask=self_attn_padding_mask, + need_attn=bool((idx == alignment_layer)), + need_head_weights=bool((idx == alignment_layer)), + ) + inner_states.append(x) + if layer_attn is not None and idx == alignment_layer: + attn = layer_attn.float().to(x) + + if attn is not None: + if alignment_heads is not None: + attn = attn[:alignment_heads] + + # average probabilities over heads + attn = attn.mean(dim=0) + + if self.layer_norm is not None: + x = self.layer_norm(x) + + # T x B x C -> B x T x C + x = x.transpose(0, 1) + + if self.project_out_dim is not None: + x = self.project_out_dim(x) + + return x, {"attn": [attn], "inner_states": inner_states} + + def forward( + self, + prev_output_tokens, + encoder_out: Optional[Dict[str, List[Tensor]]] = None, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + features_only: bool = False, + alignment_layer: Optional[int] = None, + alignment_heads: Optional[int] = None, + src_lengths: Optional[Any] = None, + return_all_hiddens: bool = False, + lang_id: Optional[int] = None, + ): + """ + Args: + prev_output_tokens (LongTensor): previous decoder outputs of shape + `(batch, tgt_len)`, for teacher forcing + encoder_out (optional): output from the encoder, used for + encoder-side attention + incremental_state (dict): dictionary used for storing state during + :ref:`Incremental decoding` + features_only (bool, optional): only return features without + applying output layer (default: False). + + Returns: + tuple: + - the decoder's output of shape `(batch, tgt_len, vocab)` + - a dictionary with any model-specific outputs + """ + + assert lang_id is not None + + x, extra = self.extract_features( + prev_output_tokens, + encoder_out=encoder_out, + incremental_state=incremental_state, + alignment_layer=alignment_layer, + alignment_heads=alignment_heads, + lang_id=lang_id, + ) + if not features_only: + x = self.output_layer(x) + return x, extra + + +@register_model_architecture("laser_transformer", "laser_transformer") +def base_laser_transformer_architecture(args): + base_architecture(args) + args.decoder_lang_embed_dim = getattr(args, "decoder_lang_embed_dim", 0) diff --git a/examples/laser/laser_src/multitask_data_utils.py b/examples/laser/laser_src/multitask_data_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..b05caea26793bf5112a7abc29d76225f578f3ebe --- /dev/null +++ b/examples/laser/laser_src/multitask_data_utils.py @@ -0,0 +1,143 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from collections import OrderedDict + +import numpy as np + +from fairseq.data import BaseWrapperDataset, FairseqDataset, iterators + + +class MultiItr(object): + def __init__(self, itr): + self.itr = itr + self._counts = [0 for x in itr] + + def __len__(self): + return sum(len(itr) for itr in self.itr) + + def __iter__(self): + return self + + def __next__(self): + ratios = [count / len(itr) for count, itr in zip(self._counts, self.itr)] + idx = ratios.index(min(ratios)) + self._counts[idx] += 1 + return next(self.itr[idx]) + + +class MultidatasetEpochBatchIterator(iterators.EpochBatchIterating): + """A wrapper around multiple epoch batch iterators.""" + + def __init__( + self, + dataset, + batch_sampler, + seed=1, + num_shards=1, + shard_id=0, + num_workers=0, + epoch=1, + ): + + assert isinstance(dataset, OrderedDict) + assert len(dataset) + assert isinstance(dataset[next(iter(dataset))], FairseqDataset) + + self.iterators = [] + + self.epoch = epoch + for key, dt in dataset.items(): + epoch_iter = iterators.EpochBatchIterator( + dataset=dt, + collate_fn=dt.collater, + batch_sampler=batch_sampler[key], + seed=seed, + num_shards=num_shards, + shard_id=shard_id, + num_workers=0, + epoch=epoch, + ) + self.iterators.append(epoch_iter) + + def __len__(self): + return sum(len(itr) for itr in self.iterators) + + def next_epoch_itr(self, shuffle=True, fix_batches_to_gpus=False): + # `self.epoch += 1` should be handled by underlying `EpochBatchIterator`s. + return MultiItr( + [ + itr.next_epoch_itr( + shuffle=shuffle, fix_batches_to_gpus=fix_batches_to_gpus + ) + for itr in self.iterators + ] + ) + + def end_of_epoch(self): + return all(itr.end_of_epoch() for itr in self.iterators) + + @property + def next_epoch_idx(self): + """Return the epoch index after *next_epoch_itr* is called.""" + + epochs = [itr.next_epoch_idx for itr in self.iterators] + self.epoch = epochs[0] + assert all(epoch == self.epoch for epoch in epochs) + + return self.epoch + + @property + def iterations_in_epoch(self): + return sum(itr.iterations_in_epoch for itr in self.iterators) + + def state_dict(self): + return { + "iterators": [it.state_dict() for it in self.iterators], + "epoch": self.epoch, + } + + def load_state_dict(self, state_dict): + self.epoch = state_dict["epoch"] + for it, d in zip(self.iterators, state_dict["iterators"]): + it.load_state_dict(d) + + +class MultitaskDatasetWrapper(BaseWrapperDataset): + """A wrapper for a multitask dataset.""" + + def __init__(self, dataset, target_language_id, sample=1.0, name=""): + super().__init__(dataset) + self.target_language_id = target_language_id + self.sample = sample + self.name = name + + def collater(self, *args, **kwargs): + ans = self.dataset.collater(*args, **kwargs) + if "net_input" in ans: + ans["net_input"]["target_language_id"] = self.target_language_id + ans["net_input"]["dataset_name"] = self.name + return ans + + def num_tokens(self, *args, **kwargs): + return self.dataset.num_tokens(*args, **kwargs) + + def ordered_indices(self, *args, **kwargs): + indices = self.dataset.ordered_indices(*args, **kwargs) + # Hacky solution for sampling + size = int(self.sample * indices.shape[0]) + + return indices.take(np.sort(np.random.permutation(indices.shape[0])[:size])) + + def size(self, index: int): + return self.dataset.size(index) + + @property + def supports_prefetch(self): + """Whether this dataset supports prefetching.""" + return getattr(self.dataset, "supports_prefetch", False) + + def prefetch(self, indices): + return self.dataset.prefetch(indices) diff --git a/examples/latent_depth/README.md b/examples/latent_depth/README.md new file mode 100644 index 0000000000000000000000000000000000000000..7774c333053b95d15b180fdfc3ee3cd817790520 --- /dev/null +++ b/examples/latent_depth/README.md @@ -0,0 +1,77 @@ +# Deep Transformers with Latent Depth (Li et al., 2020) + +[https://arxiv.org/abs/2009.13102](https://arxiv.org/abs/2009.13102). + +## Introduction + +We present a probabilistic framework to automatically learn which layer(s) to use by learning the posterior distributions of layer selection. As an extension of this framework, we propose a novel method to train one shared Transformer network for multilingual machine translation with different layer selection posteriors for each language pair. + +## Training a multilingual model with latent depth + +Below is an example of training with latent depth in decoder for one-to-many (O2M) related languages. We use the same preprocessed (numberized and binarized) TED8 dataset as in [Balancing Training for Multilingual Neural Machine Translation (Wang et al., 2020)](https://github.com/cindyxinyiwang/multiDDS), which could be generated by [the script](https://github.com/cindyxinyiwang/multiDDS/blob/multiDDS/util_scripts/prepare_multilingual_data.sh) the author provided. +```bash +lang_pairs_str="eng-aze,eng-bel,eng-ces,eng-glg,eng-por,eng-rus,eng-slk,eng-tur" +databin_dir= + +fairseq-train ${databin_dir} \ + --user-dir examples/latent_depth/latent_depth_src \ + --lang-pairs "${lang_pairs_str}" \ + --arch multilingual_transformer_iwslt_de_en \ + --task multilingual_translation_latent_depth \ + --criterion label_smoothed_cross_entropy --label-smoothing 0.1 \ + --share-encoders \ + --share-decoders \ + --decoder-langtok \ + --share-decoder-input-output-embed \ + --dropout 0.3 --attention-dropout 0.3 \ + --optimizer adam --adam-eps 1e-06 --adam-betas '(0.9, 0.98)' \ + --lr-scheduler inverse_sqrt --stop-min-lr 1e-9 --warmup-init-lr 1e-7 --warmup-updates 8000 \ + --max-tokens 4096 --update-freq 1 \ + --lr 0.0015 \ + --clip-norm 1.0 \ + --seed 2 \ + --ddp-backend=legacy_ddp \ + --encoder-layers 12 \ + --decoder-layers 24 \ + --decoder-latent-layer \ + --sparsity-weight 0.1 \ + --anneal-updates 5000 \ + --soft-update 500 \ + --target-layers 12 \ + --share-weight 0.1 +``` +## Inference command + +```bash +lang_pairs_str="eng-aze,eng-bel,eng-ces,eng-glg,eng-por,eng-rus,eng-slk,eng-tur" +databin_dir= +model_path= +src_lang= +tgt_lang= +gen_data= + +fairseq-generate ${databin_dir} \ + --path ${model_path} \ + --task multilingual_translation_latent_depth \ + --decoder-latent-layer \ + --lang-pairs "${lang_pairs_str}" \ + -s ${src_lang} -t ${tgt_lang} \ + --gen-subset $gen_data \ + --scoring sacrebleu \ + --remove-bpe 'sentencepiece' \ + --lenpen 1.0 \ + --beam 5 \ + --decoder-langtok \ + --max-tokens 4096 +``` + + +## Citation +```bibtex +@article{li2020deep, + title={Deep Transformers with Latent Depth}, + author={Li, Xian and Stickland, Asa Cooper and Tang, Yuqing and Kong, Xiang}, + journal={arXiv preprint arXiv:2009.13102}, + year={2020} +} +``` diff --git a/examples/latent_depth/latent_depth_src/__init__.py b/examples/latent_depth/latent_depth_src/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..c5fa76039ff98c18d3c14b5f4a8f73ffe644de11 --- /dev/null +++ b/examples/latent_depth/latent_depth_src/__init__.py @@ -0,0 +1,9 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from . import multilingual_translation_latent_depth # noqa +from .loss import latent_depth # noqa +from .models import latent_multilingual_transformer # noqa +from .modules import latent_layers # noqa diff --git a/examples/latent_depth/latent_depth_src/loss/__init__.py b/examples/latent_depth/latent_depth_src/loss/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/examples/latent_depth/latent_depth_src/loss/latent_depth.py b/examples/latent_depth/latent_depth_src/loss/latent_depth.py new file mode 100644 index 0000000000000000000000000000000000000000..a3b9535ecac3ec403868681a8b50c1fbe1c90dfe --- /dev/null +++ b/examples/latent_depth/latent_depth_src/loss/latent_depth.py @@ -0,0 +1,99 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +import torch +from torch.nn.modules.loss import _Loss + + +class LatentLayersKLLoss(_Loss): + def __init__(self, args): + super().__init__() + self.args = args + + def forward(self, layer_samples, lang_idx, update_num, sample_size): + prior = self.args.prior + samples = layer_samples[lang_idx] + eps = 1e-7 + if prior == "uniform": + # uniform prior + kl_loss = (samples * (torch.log(samples + eps) - math.log(0.5))).sum(-1) + elif prior == "agged_posterior": + # aggregated posterior + y_t = torch.stack([x.detach() for x in layer_samples], dim=0) + agged_q = torch.sum(y_t, dim=0) + row_norm = agged_q.sum(-1) + normed_agg_q = agged_q / row_norm + kl_loss = ( + samples * (torch.log(samples + eps) - torch.log(normed_agg_q + eps)) + ).sum(-1) + else: + raise NotImplementedError("The specified prior is not implemented.") + + # normalized by number of layers + kl_loss /= layer_samples[0].size()[0] + kl_weight = min( + self.args.sparsity_weight, + (update_num - self.args.soft_update) + * self.args.sparsity_weight + / self.args.anneal_updates, + ) + kl_loss *= kl_weight * sample_size + return kl_loss + + +class LatentLayersSparsityLoss(_Loss): + def __init__(self, args): + super().__init__() + self.args = args + + def is_valid(self, update_num): + if self.args.target_layers <= 0: + return False + return update_num > (self.args.soft_update + self.args.anneal_updates) + + def forward(self, layer_samples_list, update_num, sample_size): + batch_loss = 0 + share_loss = 0 + global_sparsity_loss = 0 + layer_samples = torch.stack(layer_samples_list, dim=0) + if ( + self.args.target_layers > 0 or self.args.share_weight > 0 + ) and update_num > (self.args.soft_update + self.args.anneal_updates): + # anneal sparsity weight + if update_num < (self.args.anneal_updates + self.args.soft_update): + weight_anneal = 0 + elif update_num < (2 * self.args.anneal_updates + self.args.soft_update): + weight_anneal = ( + (update_num - self.args.soft_update - self.args.anneal_updates) + * self.args.share_weight + / self.args.anneal_updates + ) + else: + weight_anneal = 1 + # compute ratio among languages + layer_utilization = torch.sum(layer_samples, dim=0) + layer_utilization /= layer_samples.size()[0] + if self.args.share_weight > 0: + # encouraging sharing across languages + share_loss = sum( + -1.0 * v * math.log(v) for v in layer_utilization if v > 0 + ) + batch_loss += ( + weight_anneal * self.args.share_weight * sample_size * share_loss + ) + if self.args.target_layers > 0: + # computed expected number of layers selected + expeted_layers = sum(layer_utilization) + # compute l2 loss wrt target number of layers + global_sparsity_loss = (expeted_layers - self.args.target_layers) ** 2 + batch_loss += ( + weight_anneal + * self.args.share_weight + * sample_size + * global_sparsity_loss + ) + return batch_loss diff --git a/examples/latent_depth/latent_depth_src/models/__init__.py b/examples/latent_depth/latent_depth_src/models/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/examples/latent_depth/latent_depth_src/models/latent_multilingual_transformer.py b/examples/latent_depth/latent_depth_src/models/latent_multilingual_transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..12b7e67d0336e54be05f9fdec49df2b7d4c7ae29 --- /dev/null +++ b/examples/latent_depth/latent_depth_src/models/latent_multilingual_transformer.py @@ -0,0 +1,75 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq.models import register_model, register_model_architecture +from fairseq.models.multilingual_transformer import MultilingualTransformerModel +from fairseq.models.transformer import ( + TransformerDecoder, + TransformerEncoder, + base_architecture, +) + +from .latent_transformer import LatentTransformerDecoder, LatentTransformerEncoder + + +@register_model("latent_multilingual_transformer") +class LatentMultilingualTransformerModel(MultilingualTransformerModel): + """A variant of standard multilingual Transformer models which encoder and/or + decoders supports latent depth, as is in "Deep Transformer with Latent Depth" + (https://arxiv.org/abs/2009.13102). + """ + + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + MultilingualTransformerModel.add_args(parser) + parser.add_argument( + '--soft-select', + action='store_true', + help='use soft samples in training an inference', + ) + parser.add_argument( + '--sampling-tau', + type=float, + default=5., + help='sampling temperature', + ) + + @classmethod + def _get_module_class(cls, is_encoder, args, lang_dict, embed_tokens, langs): + if is_encoder: + if hasattr(args, "encoder_latent_layer") and args.encoder_latent_layer: + return LatentTransformerEncoder( + args, lang_dict, embed_tokens, num_logits=len(langs) + ) + else: + return TransformerEncoder(args, lang_dict, embed_tokens) + else: + if hasattr(args, "decoder_latent_layer") and args.decoder_latent_layer: + return LatentTransformerDecoder( + args, lang_dict, embed_tokens, num_logits=len(langs) + ) + else: + return TransformerDecoder(args, lang_dict, embed_tokens) + + +@register_model_architecture( + "latent_multilingual_transformer", "latent_multilingual_transformer" +) +def latent_multilingual_architecture(args): + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 1024) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 4) + args.encoder_layers = getattr(args, "encoder_layers", 12) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 512) + args.decoder_ffn_embed_dim = getattr(args, "decoder_ffn_embed_dim", 1024) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 4) + args.decoder_layers = getattr(args, "decoder_layers", 24) + args.share_encoders = getattr(args, "share_encoders", True) + args.share_decoders = getattr(args, "share_decoders", True) + args.share_encoder_embeddings = getattr(args, "share_encoder_embeddings", True) + args.share_decoder_embeddings = getattr(args, "share_decoder_embeddings", True) + + base_architecture(args) diff --git a/examples/latent_depth/latent_depth_src/models/latent_transformer.py b/examples/latent_depth/latent_depth_src/models/latent_transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..6a825301a452bd935deafdaf78fa2427ca9a469e --- /dev/null +++ b/examples/latent_depth/latent_depth_src/models/latent_transformer.py @@ -0,0 +1,156 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from typing import Any, Dict, Optional + +import torch.nn as nn +from fairseq.models.fairseq_encoder import EncoderOut +from fairseq.models.transformer import TransformerDecoder, TransformerEncoder +from fairseq.modules import TransformerDecoderLayer, TransformerEncoderLayer +from torch import Tensor + +from ..modules.latent_layers import LayerSelect + + +class LatentTransformerEncoder(TransformerEncoder): + """Latent depth (https://arxiv.org/abs/2009.13102) implemented in + TransformerEncoder. + """ + + def __init__(self, args, dictionary, embed_tokens, num_logits=1): + self.num_logits = num_logits + self.num_layers = args.encoder_layers + super().__init__(args, dictionary, embed_tokens) + self.layer_select = LayerSelect( + num_layers=self.num_layers, + num_logits=self.num_logits, + soft_select=getattr(args, "soft_select", False), + sampling_tau=getattr(args, "sampling_tau", 5.), + ) + self.lang_idx = None + self.layers = nn.ModuleList( + [self._build_encoder_layer(args, idx) for idx in range(args.encoder_layers)] + ) + + def set_lang_idx(self, lang_idx): + self.lang_idx = lang_idx + + def _build_encoder_layer(self, args, idx=None): + return LatentTransformerEncoderLayer(args, idx, layer_select=self.layer_select) + + def forward(self, src_tokens, src_lengths, return_all_hiddens: bool = False): + self.layer_select.sample(self.lang_idx) + return super().forward(src_tokens, src_lengths, return_all_hiddens) + + +class LatentTransformerEncoderLayer(TransformerEncoderLayer): + """Encoder layer with each (non_residual) block weighted by samples of Bernouli + or Gumbel Signmoid samples. + + Args: + args (argparse.Namespace): parsed command-line arguments from standard + TransformerEncoderLayer. + idx (int): layer index (used to retrieve samples). + layer_select (LayerSelect, optional): instance of LayerSelect module with logits + parameters and sampling method. + """ + + def __init__(self, args, idx, layer_select=None): + super().__init__(args) + self.idx = idx + self.layer_select = layer_select + + def residual_connection(self, x, residual): + return residual + x * self.layer_select(self.idx) + + +class LatentTransformerDecoder(TransformerDecoder): + """Latent depth (https://arxiv.org/abs/2009.13102) implemented in + TransformerDecoder. + """ + + def __init__( + self, args, dictionary, embed_tokens, no_encoder_attn=False, num_logits=1 + ): + self.num_logits = num_logits + self.num_layers = args.decoder_layers + super().__init__( + args, dictionary, embed_tokens, no_encoder_attn=no_encoder_attn + ) + self.layer_select = LayerSelect( + num_layers=self.num_layers, + num_logits=self.num_logits, + soft_select=getattr(args, "soft_select", False), + sampling_tau=getattr(args, "sampling_tau", 5.), + ) + self.lang_idx = None + self.layers = nn.ModuleList( + [ + self._build_decoder_layer(args, no_encoder_attn, idx) + for idx in range(args.decoder_layers) + ] + ) + + def set_lang_idx(self, lang_idx): + self.lang_idx = lang_idx + + def _build_decoder_layer(self, args, no_encoder_attn=False, idx=None): + return LatentTransformerDecoderLayer( + args, idx, layer_select=self.layer_select, no_encoder_attn=no_encoder_attn + ) + + def forward( + self, + prev_output_tokens, + encoder_out: Optional[EncoderOut] = None, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + features_only: bool = False, + alignment_layer: Optional[int] = None, + alignment_heads: Optional[int] = None, + src_lengths: Optional[Any] = None, + return_all_hiddens: bool = False, + ): + self.layer_select.sample(self.lang_idx) + return super().forward( + prev_output_tokens=prev_output_tokens, + encoder_out=encoder_out, + incremental_state=incremental_state, + features_only=features_only, + alignment_layer=alignment_layer, + src_lengths=src_lengths, + return_all_hiddens=return_all_hiddens, + ) + + +class LatentTransformerDecoderLayer(TransformerDecoderLayer): + """Decoder layer with each (non_residual) block weighted by samples of Bernouli + or Gumbel Signmoid samples. + + Args: + args (argparse.Namespace): parsed command-line arguments from standard + TransformerDecoderLayer. + idx (int): layer index (used to retrieve samples). + layer_select (LayerSelect, optional): instance of LayerSelect module with logits + parameters and sampling method. + no_encoder_attn (bool, optional): whether to attend to encoder outputs + (default: False). + + """ + + def __init__( + self, + args, + idx, + layer_select=None, + no_encoder_attn=False, + add_bias_kv=False, + add_zero_attn=False, + ): + super().__init__(args, no_encoder_attn, add_bias_kv, add_zero_attn) + self.idx = idx + self.layer_select = layer_select + + def residual_connection(self, x, residual): + return residual + x * self.layer_select(self.idx) diff --git a/examples/latent_depth/latent_depth_src/modules/__init__.py b/examples/latent_depth/latent_depth_src/modules/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/examples/latent_depth/latent_depth_src/modules/latent_layers.py b/examples/latent_depth/latent_depth_src/modules/latent_layers.py new file mode 100644 index 0000000000000000000000000000000000000000..2be05d5535cb05b16f61603a7356df2326bf2e23 --- /dev/null +++ b/examples/latent_depth/latent_depth_src/modules/latent_layers.py @@ -0,0 +1,75 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn as nn + + +class LayerSelect(nn.Module): + """Compute samples (from a Gumbel-Sigmoid distribution) which is used as + either (soft) weighting or (hard) selection of residual connection. + https://arxiv.org/abs/2009.13102 + """ + def __init__(self, num_layers, num_logits, soft_select=False, sampling_tau=5.): + super(LayerSelect, self).__init__() + self.layer_logits = torch.nn.Parameter( + torch.Tensor(num_logits, num_layers), + requires_grad=True, + ) + self.hard_select = not soft_select + self.tau = sampling_tau + self.detach_grad = False + self.layer_samples = [None] * num_logits + + def sample(self, logit_idx): + """To leverage the efficiency of distributed training, samples for all + layers are computed at once for each logit_idx. Logits are parameters + learnt independent of each other. + + Args: + logit_idx: The index of logit parameters used for sampling. + """ + assert logit_idx is not None + self.samples = self._gumbel_sigmoid( + self.layer_logits[logit_idx, :].detach() + if self.detach_grad + else self.layer_logits[logit_idx, :], + dim=-1, + tau=self.tau, + hard=self.hard_select, + ) + self.layer_samples[logit_idx] = self.samples + + def forward(self, i): + sample = self.samples[i] + return sample + + def _gumbel_sigmoid( + self, logits, tau=1, hard=False, eps=1e-10, dim=-1, threshold=0.5 + ): + # ~Gumbel(0,1) + gumbels1 = ( + -torch.empty_like(logits, memory_format=torch.legacy_contiguous_format) + .exponential_() + .log() + ) + gumbels2 = ( + -torch.empty_like(logits, memory_format=torch.legacy_contiguous_format) + .exponential_() + .log() + ) + # Difference of two gumbels because we apply a sigmoid + gumbels1 = (logits + gumbels1 - gumbels2) / tau + y_soft = gumbels1.sigmoid() + if hard: + # Straight through. + y_hard = torch.zeros_like( + logits, memory_format=torch.legacy_contiguous_format + ).masked_fill(y_soft > threshold, 1.0) + ret = y_hard - y_soft.detach() + y_soft + else: + # Reparametrization trick. + ret = y_soft + return ret diff --git a/examples/latent_depth/latent_depth_src/multilingual_translation_latent_depth.py b/examples/latent_depth/latent_depth_src/multilingual_translation_latent_depth.py new file mode 100644 index 0000000000000000000000000000000000000000..b5cd51a470bd56266d4198b6cd20004c53b04c70 --- /dev/null +++ b/examples/latent_depth/latent_depth_src/multilingual_translation_latent_depth.py @@ -0,0 +1,194 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq.tasks import register_task +from fairseq.tasks.multilingual_translation import MultilingualTranslationTask + +from .loss.latent_depth import LatentLayersKLLoss, LatentLayersSparsityLoss + + +@register_task("multilingual_translation_latent_depth") +class MultilingualTranslationTaskLatentDepth(MultilingualTranslationTask): + """A task for multiple translation with latent depth. + + See `"Deep Transformer with Latent Depth" + (Li et al., 2020) `_. + """ + + @staticmethod + def add_args(parser): + """Add task-specific arguments to the parser.""" + # fmt: off + MultilingualTranslationTask.add_args(parser) + parser.add_argument('--encoder-latent-layer', action='store_true', help='latent layer selection in encoder') + parser.add_argument('--decoder-latent-layer', action='store_true', help='latent layer selection in decoder') + parser.add_argument('--target-layers', default=-1, type=int, + help='number of effective layers to learn; -1 means no constraint') + parser.add_argument('--sparsity-weight', default=0.0, type=float, + help='weight for sparsity loss') + parser.add_argument('--share-weight', default=0.0, type=float, + help='weight for sharing loss') + parser.add_argument('--soft-update', default=1, type=int, + help='number of updates with soft sampling') + parser.add_argument('--anneal-updates', default=1, type=int, + help='number of updates to anneal the KL loss weight') + parser.add_argument('--prior', default="uniform", type=str, + help='prior used for computing KL loss') + # fmt: on + + def __init__(self, args, dicts, training): + super().__init__(args, dicts, training) + self.src_langs, self.tgt_langs = zip( + *[(lang.split("-")[0], lang.split("-")[1]) for lang in args.lang_pairs] + ) + if self.training and self.encoder_latent_layer: + assert self.args.share_encoders + if self.training and self.decoder_latent_layer: + assert self.args.share_decoders + if training or self.encoder_latent_layer or self.decoder_latent_layer: + self.lang_pairs = args.lang_pairs + else: + self.lang_pairs = ["{}-{}".format(args.source_lang, args.target_lang)] + self.eval_lang_pairs = self.lang_pairs + self.model_lang_pairs = self.lang_pairs + if self.training and (self.encoder_latent_layer or self.decoder_latent_layer): + self.kl_loss = LatentLayersKLLoss(self.args) + self.sparsity_loss = LatentLayersSparsityLoss(self.args) + + def _per_lang_pair_train_loss( + self, lang_pair, model, update_num, criterion, sample, optimizer, ignore_grad + ): + src, tgt = lang_pair.split("-") + if self.encoder_latent_layer: + src_lang_idx = self.src_lang_idx_dict[src] + model.models[lang_pair].encoder.set_lang_idx(src_lang_idx) + model.models[lang_pair].encoder.layer_select.hard_select = ( + update_num > self.args.soft_update + ) + if self.decoder_latent_layer: + tgt_lang_idx = self.tgt_lang_idx_dict[tgt] + model.models[lang_pair].decoder.set_lang_idx(tgt_lang_idx) + model.models[lang_pair].decoder.layer_select.hard_select = ( + update_num > self.args.soft_update + ) + + loss, sample_size, logging_output = criterion( + model.models[lang_pair], sample[lang_pair] + ) + if self.encoder_latent_layer: + none_samples = sum( + 1 if x is None else 0 + for x in model.models[lang_pair].encoder.layer_select.layer_samples + ) + if none_samples == 0 or self.args.prior != "agged_posterior": + loss += self.kl_loss( + model.models[lang_pair].encoder.layer_select.layer_samples, + src_lang_idx, + update_num, + sample_size, + ) + if self.decoder_latent_layer: + none_samples = sum( + 1 if x is None else 0 + for x in model.models[lang_pair].decoder.layer_select.layer_samples + ) + if none_samples == 0 or self.args.prior != "agged_posterior": + loss += self.kl_loss( + model.models[lang_pair].decoder.layer_select.layer_samples, + tgt_lang_idx, + update_num, + sample_size, + ) + if ignore_grad: + loss *= 0 + + if hasattr(self, "sparsity_loss") and self.sparsity_loss.is_valid(update_num): + # need to retain the graph if sparsity loss needs to be added + loss.backward(retain_graph=True) + else: + optimizer.backward(loss) + + return loss, sample_size, logging_output + + def train_step( + self, sample, model, criterion, optimizer, update_num, ignore_grad=False + ): + agg_loss, agg_sample_size, agg_logging_output = super().train_step( + sample, model, criterion, optimizer, update_num, ignore_grad + ) + # compute auxiliary loss from layere sparsity, based on all samples from all languages + if hasattr(self, "sparsity_loss") and self.sparsity_loss.is_valid(update_num): + sparsity_loss = 0 + if self.encoder_latent_layer: + sparsity_loss += self.sparsity_loss( + next( + iter(model.models.values()) + ).encoder.layer_select.layer_samples, + update_num, + agg_sample_size, + ) + if self.decoder_latent_layer: + sparsity_loss += self.sparsity_loss( + next( + iter(model.models.values()) + ).decoder.layer_select.layer_samples, + update_num, + agg_sample_size, + ) + if sparsity_loss > 0: + optimizer.backward(sparsity_loss) + return agg_loss, agg_sample_size, agg_logging_output + + def _per_lang_pair_valid_loss(self, lang_pair, model, criterion, sample): + src, tgt = lang_pair.split("-") + if self.encoder_latent_layer: + src_lang_idx = self.src_lang_idx_dict[src] + model.models[lang_pair].encoder.set_lang_idx(src_lang_idx) + if self.decoder_latent_layer: + tgt_lang_idx = self.tgt_lang_idx_dict[tgt] + model.models[lang_pair].decoder.set_lang_idx(tgt_lang_idx) + loss, sample_size, logging_output = criterion( + model.models[lang_pair], sample[lang_pair] + ) + return loss, sample_size, logging_output + + def inference_step( + self, generator, models, sample, prefix_tokens=None, constraints=None + ): + if self.encoder_latent_layer or self.decoder_latent_layer: + for model in models: + if self.encoder_latent_layer: + assert model.encoder.layer_select is not None + src_lang_idx = self.src_lang_idx_dict[self.args.source_lang] + model.encoder.set_lang_idx(src_lang_idx) + if self.decoder_latent_layer: + assert model.decoder.layer_select is not None + tgt_lang_idx = self.tgt_lang_idx_dict[self.args.target_lang] + model.decoder.set_lang_idx(tgt_lang_idx) + return super().inference_step( + generator, models, sample, prefix_tokens, constraints + ) + + @property + def encoder_latent_layer(self): + return ( + hasattr(self.args, "encoder_latent_layer") + and self.args.encoder_latent_layer + ) + + @property + def decoder_latent_layer(self): + return ( + hasattr(self.args, "decoder_latent_layer") + and self.args.decoder_latent_layer + ) + + @property + def src_lang_idx_dict(self): + return {lang: lang_idx for lang_idx, lang in enumerate(self.src_langs)} + + @property + def tgt_lang_idx_dict(self): + return {lang: lang_idx for lang_idx, lang in enumerate(self.tgt_langs)} diff --git a/examples/layerdrop/README.md b/examples/layerdrop/README.md new file mode 100644 index 0000000000000000000000000000000000000000..394e710b0f522981dbb073f28eaf550ee28760cf --- /dev/null +++ b/examples/layerdrop/README.md @@ -0,0 +1,154 @@ +# Reducing Transformer Depth on Demand with Structured Dropout (Fan et al., 2019) +This page contains information for how to train models with LayerDrop, based on this [paper](https://arxiv.org/abs/1909.11556). + +## Citation: +If you found this technique useful, please cite our paper: +```bibtex +@article{fan2019reducing, + title={Reducing Transformer Depth on Demand with Structured Dropout}, + author={Fan, Angela and Grave, Edouard and Joulin, Armand}, + journal={arXiv preprint arXiv:1909.11556}, + year={2019} +} +``` + +## Pre-trained models + +Model | Description | Download +---|---|--- +`layerdrop_wmt_en_de_12_6` | Transformer + LayerDrop 0.2 trained on WMT16 en-de with 12 encoder and 6 decoder layers | [layerdrop_wmt_en_de_12_6.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/layerdrop_wmt_en_de_12_6.tar.gz) +`roberta_layerdrop.base` | RoBERTa Base + LayerDrop 0.2 | [roberta_layerdrop.base.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/roberta_layerdrop.base.qnli.tar.gz) +`roberta_layerdrop.large` | RoBERTa Large + LayerDrop 0.2 | [roberta_layerdrop.large.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/roberta_layerdrop.large.tar.gz) +`roberta_layerdrop.large.mnli` | `roberta_layerdrop.large` finetuned on [MNLI](http://www.nyu.edu/projects/bowman/multinli) | [roberta_layerdrop.large.mnli.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/roberta_layerdrop.large.mnli.tar.gz) +`roberta_layerdrop.large.qnli` | `roberta_layerdrop.large` finetuned on [QNLI](https://arxiv.org/abs/1804.07461) | [roberta_layerdrop.large.mnli.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/roberta_layerdrop.large.qnli.tar.gz) + + +Evaluate performance of these pre-trained models: +```bash +# Example for Machine Translation +fairseq-generate /path/to/bped/wmt/data --path nmt_checkpoint.pt \ + --beam 8 --lenpen 0.4 \ + --batch-size 64 \ + --remove-bpe \ + --gen-subset test > wmt16_gen.txt +bash scripts/compound_split_bleu.sh wmt16_gen.txt +# prints BLEU4 = 30.17 +``` + +```python +# Example for RoBERTa + LayerDrop finetuned on MNLI: +from fairseq.models.roberta import RobertaModel + +roberta_layerdrop = RobertaModel.from_pretrained( + '/path/to/MNLI/model', + checkpoint_file='mnli_checkpoint.pt', + data_name_or_path='/path/to/MNLI/data/MNLI-bin' +) +label_map = {0: 'contradiction', 2: 'neutral', 1: 'entailment'} +ncorrect, nsamples = 0, 0 +roberta_layerdrop.cuda() +roberta_layerdrop.eval() +with open('/path/to/MNLI/data/dev_matched.tsv') as fin: + fin.readline() + for index, line in enumerate(fin): + tokens = line.strip().split('\t') + sent1, sent2, target = tokens[8], tokens[9], tokens[-1] + tokens = roberta_layerdrop.encode(sent1, sent2) + prediction = roberta_layerdrop.predict('sentence_classification_head', tokens).argmax().item() + prediction_label = label_map[prediction] + ncorrect += int(prediction_label == target) + nsamples += 1 +print('| Accuracy: ', float(ncorrect)/float(nsamples)) +# prints | Accuracy: 0.9026999490575649 + + +# Example for RoBERTa + LayerDrop finetuned on QNLI: +roberta = RobertaModel.from_pretrained( + '/path/to/QNLI/model', + checkpoint_file='qnli_checkpoint.pt', + data_name_or_path='/path/to/QNLI/data/QNLI-bin' +) + +label_fn = lambda label: roberta.task.label_dictionary.string( + [label + roberta.task.target_dictionary.nspecial] +) +ncorrect, nsamples = 0, 0 +roberta.cuda() +roberta.eval() +with open('/path/to/QNLI/data/dev.tsv') as fin: + fin.readline() + for index, line in enumerate(fin): + tokens = line.strip().split('\t') + sent1, sent2, target = tokens[1], tokens[2], tokens[3] + tokens = roberta.encode(sent1, sent2) + prediction = roberta.predict('sentence_classification_head', tokens).argmax().item() + prediction_label = label_fn(prediction) + ncorrect += int(prediction_label == target) + nsamples += 1 +print('| Accuracy: ', float(ncorrect)/float(nsamples)) +# prints | Accuracy: 0.9480139117700896 +``` + + +## Example usage + +To train a model with LayerDrop, add the following flags. We recommend 0.2, a value that worked well in our experiments. For Language Models that are decoder-only, you need only the decoder flag. For RoBERTa, an encoder, you need only the encoder flag. The encoder and decoder LayerDrop values can be set differently. +``` +--encoder-layerdrop 0.2 --decoder-layerdrop 0.2 +``` + +To prune a model that has been trained with LayerDrop, add the following flags followed by a comma separated list of which layers you would like to keep. +``` +--encoder-layers-to-keep 0,2,4,6,8,10,12,14 --decoder-layers-to-keep 0,2,4,6,8,10,12,14 +``` +Setting these flags should print a message such as: +``` +| Pruning model to specified layer configuration +``` +You should also see a smaller number of parameters in the model, for example the 16-Layer Transformer Language Model prints: +``` +num. model params: 246933504 +``` +while a model pruned to 8 Layers prints: +``` +num. model params: 146163712 +``` + +If you would like to pick up training with a model that has been pruned, simply adding these flags is sufficient. If you would like to use a script that only does evaluation (no training), you may need to pass an override command. A specific example would be for language modeling: +```bash +fairseq-eval-lm /path/to/wikitext-103 \ + --path /path/to/model/checkpoint.pt \ + --model-overrides "{'decoder_layers_to_keep':'0,2,4,6,8,10,12,14'}" +``` +This model override command overrides the training parameters and updates the model arguments so that the pruned model is run instead of the full model. + +## Reproduce Paper Results + +Looking to reproduce the results in the paper? + +1. For Translation on WMT16 en-de, we followed this setting [here](https://github.com/pytorch/fairseq/blob/master/examples/scaling_nmt/README.md) +2. To train RoBERTa, we followed this setting [here](https://github.com/pytorch/fairseq/tree/master/examples/roberta) +3. To train Language Models on Wikitext-103, we followed this setting [here](https://github.com/pytorch/fairseq/tree/master/examples/language_model) + + +## Tips + +1. If you would like to train large models with better performance, LayerDrop should be set to a smaller value such as 0.1 or 0.2. Too much LayerDrop will mean the model has too much regularization, so may not reach the best performance. Since LayerDrop adds regularization, you may achieve the best performance by slightly reducing the amount of standard dropout (for example, reduce by 0.1). + +2. If you would like to train large models to be pruned and made smaller, LayerDrop should be set to a larger value such as 0.5 if you want to prune very aggressively (such as removing half the network or more). If you would like to prune fewer layers away, LayerDrop can be set to a smaller value such as 0.2. Our experiments were conducted with low values of LayerDrop (such as 0.1 and 0.2), for reference. + +3. When pruning layers at inference time, it is best to spread out the layers remaining so they are evenly spaced throughout the network. For example, if you want to remove 50% of the network, keeping every other layer is good. + + +## FAQ + +1. How did the sharing layers experiment work? In an appendix (https://openreview.net/pdf?id=SylO2yStDr) we added an experiment on Wikitext-103 language modeling that combined LayerDrop with Weight Sharing. We shared chunks of 2 layers such that every other layer had shared weights. For example, if our network has layers 1 through 6, then layer 1 and 2 are shared, layer 3 and 4 are shared, and layer 5 and 6 are shared. + +2. LayerDrop hasn't been helping in my setting? During training time, LayerDrop can help regularize your network. This is most important if your network is already overfitting - if your network is underfitting, it is possible LayerDrop is adding too much regularization. We recommend using smaller values (such as 0.1 or 0.2) and also decreasing the quantity of standard dropout (for example, reduce by 0.1). + +3. Can you train a model without LayerDrop and finetune with LayerDrop (e.g. for BERT)? In our experiments, we did not see great performance. Models such as RoBERTa have trained for a long time in the pre-training setting, so only finetuning with LayerDrop for a few epochs on a downstream task such as MNLI does not achieve the robustness required for successful pruning. + + +## Having an issue or have a question? + +Please open an issue in this repository with the details of your question. Thanks! diff --git a/examples/linformer/README.md b/examples/linformer/README.md new file mode 100644 index 0000000000000000000000000000000000000000..f8b36bc691cb8f5bf82942e07b6d9c014387bdd8 --- /dev/null +++ b/examples/linformer/README.md @@ -0,0 +1,22 @@ +# Linformer: Self-Attention with Linear Complexity (Wang et al., 2020) + +This example contains code to train Linformer models as described in our paper +[Linformer: Self-Attention with Linear Complexity](https://arxiv.org/abs/2006.04768). + +## Training a new Linformer RoBERTa model + +You can mostly follow the [RoBERTa pretraining README](/examples/roberta/README.pretraining.md), +updating your training command with `--user-dir examples/linformer/linformer_src --arch linformer_roberta_base`. + +## Citation + +If you use our work, please cite: + +```bibtex +@article{wang2020linformer, + title={Linformer: Self-Attention with Linear Complexity}, + author={Wang, Sinong and Li, Belinda and Khabsa, Madian and Fang, Han and Ma, Hao}, + journal={arXiv preprint arXiv:2006.04768}, + year={2020} +} +``` diff --git a/examples/linformer/linformer_src/__init__.py b/examples/linformer/linformer_src/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..1c52f135ea6f99d0effe8ce1f7d77cbd66be3745 --- /dev/null +++ b/examples/linformer/linformer_src/__init__.py @@ -0,0 +1,6 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .models import linformer_roberta # noqa diff --git a/examples/linformer/linformer_src/models/__init__.py b/examples/linformer/linformer_src/models/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/examples/linformer/linformer_src/models/linformer_roberta.py b/examples/linformer/linformer_src/models/linformer_roberta.py new file mode 100644 index 0000000000000000000000000000000000000000..18ad44f079e691e7f46aa2745fe4f35d4466ca33 --- /dev/null +++ b/examples/linformer/linformer_src/models/linformer_roberta.py @@ -0,0 +1,119 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +Linformer: Self-Attention with Linear Complexity +""" + +import logging + +import torch +from fairseq import utils +from fairseq.models import register_model, register_model_architecture +from fairseq.models.roberta import ( + init_bert_params, + roberta_base_architecture, + roberta_large_architecture, + RobertaEncoder, + RobertaModel, +) + +from ..modules.linformer_sentence_encoder import LinformerTransformerEncoder + + +logger = logging.getLogger(__name__) + + +@register_model("linformer_roberta") +class LinformerModel(RobertaModel): + @staticmethod + def add_args(parser): + RobertaModel.add_args(parser) + + # add args for Linformer + parser.add_argument( + "--compressed", type=int, help="compressed ratio of sequence length" + ) + parser.add_argument( + "--shared-kv-compressed", + type=int, + help="share compressed matrix between k and v, in each layer", + ) + parser.add_argument( + "--shared-layer-kv-compressed", + type=int, + help="share compressed matrix between k and v and across all layers", + ) + parser.add_argument( + "--freeze-compress", + type=int, + help="freeze the parameters in compressed layer", + ) + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + + # make sure all arguments are present + base_architecture(args) + + if not hasattr(args, "max_positions"): + args.max_positions = args.tokens_per_sample + + encoder = LinformerEncoder(args, task.source_dictionary) + return cls(args, encoder) + + +class LinformerEncoder(RobertaEncoder): + """Linformer encoder.""" + + def __init__(self, args, dictionary): + super().__init__(args, dictionary) + self.register_buffer("version", torch.tensor(2)) + + def build_encoder(self, args, dictionary, embed_tokens): + encoder = LinformerTransformerEncoder(args, dictionary, embed_tokens) + encoder.apply(init_bert_params) + return encoder + + def upgrade_state_dict_named(self, state_dict, name): + super().upgrade_state_dict_named(state_dict, name) + prefix = name + "." if name != "" else "" + + # some old checkpoints had weight sharing implemented incorrectly + # (note: this was correct in the original paper code) + if utils.item(state_dict.get(f"{prefix}version", torch.tensor(1))) < 2: + state_dict[f"{prefix}version"] = torch.tensor(1) + # check if input embeddings and output embeddings were tied + if not torch.allclose( + state_dict[f"{prefix}sentence_encoder.embed_tokens.weight"], + state_dict[f"{prefix}lm_head.weight"], + ): + # they weren't tied, re-init the LM head without weight sharing + self.lm_head = self.build_lm_head( + embed_dim=self.args.encoder_embed_dim, + output_dim=len(self.dictionary), + activation_fn=self.args.activation_fn, + weight=None, # don't share weights + ) + + +@register_model_architecture("linformer_roberta", "linformer_roberta") +def base_architecture(args): + args.compressed = getattr(args, "compressed", 4) + args.shared_kv_compressed = getattr(args, "shared_kv_compressed", 0) + args.shared_layer_kv_compressed = getattr(args, "shared_layer_kv_compressed", 0) + args.freeze_compress = getattr(args, "freeze_compress", 0) + roberta_base_architecture(args) + + +@register_model_architecture("linformer_roberta", "linformer_roberta_base") +def linformer_roberta_base_architecture(args): + base_architecture(args) + + +@register_model_architecture("linformer_roberta", "linformer_roberta_large") +def linformer_roberta_large_architecture(args): + roberta_large_architecture(args) + base_architecture(args) diff --git a/examples/linformer/linformer_src/modules/__init__.py b/examples/linformer/linformer_src/modules/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/examples/linformer/linformer_src/modules/linformer_sentence_encoder.py b/examples/linformer/linformer_src/modules/linformer_sentence_encoder.py new file mode 100644 index 0000000000000000000000000000000000000000..44f7989bd863329f763aa62b78df2eb42b3084ea --- /dev/null +++ b/examples/linformer/linformer_src/modules/linformer_sentence_encoder.py @@ -0,0 +1,54 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +import torch.nn as nn +from fairseq.models.transformer import TransformerEncoder + +from .linformer_sentence_encoder_layer import LinformerTransformerEncoderLayer + + +class LinformerTransformerEncoder(TransformerEncoder): + """ + Implementation for a Bi-directional Linformer based Sentence Encoder used + in BERT/XLM style pre-trained models. + + This first computes the token embedding using the token embedding matrix, + position embeddings (if specified) and segment embeddings + (if specified). After applying the specified number of + LinformerEncoderLayers, it outputs all the internal states of the + encoder as well as the final representation associated with the first + token (usually CLS token). + + Input: + - tokens: B x T matrix representing sentences + - segment_labels: B x T matrix representing segment label for tokens + + Output: + - a tuple of the following: + - a list of internal model states used to compute the + predictions where each tensor has shape T x B x C + - sentence representation associated with first input token + in format B x C. + """ + + def __init__(self, args, dictionary, embed_tokens): + self.compress_layer = None + super().__init__(args, dictionary, embed_tokens) + + def build_encoder_layer(self, args): + if self.args.shared_layer_kv_compressed == 1 and self.compress_layer is None: + compress_layer = nn.Linear( + self.args.max_positions, + self.args.max_positions // self.args.compressed, + ) + # intialize parameters for compressed layer + nn.init.xavier_uniform_(compress_layer.weight, gain=1 / math.sqrt(2)) + if self.args.freeze_compress == 1: + compress_layer.weight.requires_grad = False + self.compress_layer = compress_layer + + return LinformerTransformerEncoderLayer(args, self.compress_layer) diff --git a/examples/linformer/linformer_src/modules/linformer_sentence_encoder_layer.py b/examples/linformer/linformer_src/modules/linformer_sentence_encoder_layer.py new file mode 100644 index 0000000000000000000000000000000000000000..7e2caa03400129ac0bb34ae35274cdf46f27a055 --- /dev/null +++ b/examples/linformer/linformer_src/modules/linformer_sentence_encoder_layer.py @@ -0,0 +1,65 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +from fairseq import utils +from fairseq.modules import TransformerEncoderLayer + +from .multihead_linear_attention import MultiheadLinearAttention + + +class LinformerTransformerEncoderLayer(TransformerEncoderLayer): + """ + Implements a Linformer Encoder Layer used in BERT/XLM style pre-trained + models. + """ + + def __init__(self, args, shared_compress_layer): + # wrap in a list so it's not automatically registered by PyTorch + self.shared_compress_layer = [shared_compress_layer] + + super().__init__(args) + + self.register_buffer("version", torch.tensor(2)) + + def build_self_attention(self, embed_dim, args): + return MultiheadLinearAttention( + embed_dim, + args.encoder_attention_heads, + dropout=args.dropout, + self_attention=True, + q_noise=args.quant_noise_pq, + qn_block_size=args.quant_noise_pq_block_size, + compressed=args.compressed, + max_seq_len=args.max_positions, + shared_kv_compressed=args.shared_kv_compressed, + shared_compress_layer=self.shared_compress_layer[0], + freeze_compress=args.freeze_compress, + ) + + def upgrade_state_dict_named(self, state_dict, name): + super().upgrade_state_dict_named(state_dict, name) + prefix = name + "." if name != "" else "" + + # some old checkpoints had weight sharing implemented incorrectly + # (note: this was correct in the original paper code) + if utils.item(state_dict.get(f"{prefix}version", torch.tensor(1))) < 2: + state_dict[f"{prefix}version"] = torch.tensor(1) + # check compression layer sharing + if f"{prefix}shared_compress_layer.weight" in state_dict: + # reinitialize block without sharing compression layer to match + # old behavior + self.shared_compress_layer = [ + torch.nn.Linear( + self.shared_compress_layer[0].weight.size(1), + self.shared_compress_layer[0].weight.size(0), + ) + ] + self.self_attn = self.build_self_attention(self.embed_dim, self.args) + # delete shared_compress_layer, since it's already copied to + # self_attn.compress_k.weight + del state_dict[f"{prefix}shared_compress_layer.weight"] + if f"{prefix}shared_compress_layer.bias" in state_dict: + del state_dict[f"{prefix}shared_compress_layer.bias"] diff --git a/examples/linformer/linformer_src/modules/multihead_linear_attention.py b/examples/linformer/linformer_src/modules/multihead_linear_attention.py new file mode 100644 index 0000000000000000000000000000000000000000..6be1007279217c5de644e8b054f5d14a19f06c55 --- /dev/null +++ b/examples/linformer/linformer_src/modules/multihead_linear_attention.py @@ -0,0 +1,481 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +from typing import Dict, Optional, Tuple + +import torch +import torch.nn.functional as F +from fairseq import utils +from fairseq.incremental_decoding_utils import with_incremental_state +from fairseq.modules.quant_noise import quant_noise +from torch import Tensor, nn +from torch.nn import Parameter + + +@with_incremental_state +class MultiheadLinearAttention(nn.Module): + """Multi-headed linformer attention. + + Projects the key and values down to the compressed dimension, before computing self-attention. + + See "Linformer: Self-Attention with Linear Complexity" for more details. + """ + + def __init__( + self, + embed_dim, + num_heads, + kdim=None, + vdim=None, + dropout=0.0, + bias=True, + add_bias_kv=False, + add_zero_attn=False, + self_attention=False, + encoder_decoder_attention=False, + q_noise=0.0, + qn_block_size=8, + compressed=1, + max_seq_len=256, + shared_kv_compressed=0, + shared_compress_layer=None, + freeze_compress=0, + ): + super().__init__() + self.embed_dim = embed_dim + self.kdim = kdim if kdim is not None else embed_dim + self.vdim = vdim if vdim is not None else embed_dim + self.qkv_same_dim = self.kdim == embed_dim and self.vdim == embed_dim + + self.num_heads = num_heads + self.dropout = dropout + self.head_dim = embed_dim // num_heads + assert ( + self.head_dim * num_heads == self.embed_dim + ), "embed_dim must be divisible by num_heads" + self.scaling = self.head_dim ** -0.5 + + self.self_attention = self_attention + self.encoder_decoder_attention = encoder_decoder_attention + + assert not self.self_attention or self.qkv_same_dim, ( + "Self-attention requires query, key and " "value to be of the same size" + ) + + self.k_proj = quant_noise( + nn.Linear(self.kdim, embed_dim, bias=bias), q_noise, qn_block_size + ) + self.v_proj = quant_noise( + nn.Linear(self.vdim, embed_dim, bias=bias), q_noise, qn_block_size + ) + self.q_proj = quant_noise( + nn.Linear(embed_dim, embed_dim, bias=bias), q_noise, qn_block_size + ) + + # used for compress sequence to subsequence + if shared_compress_layer is None: + self.compress_seq_len = max_seq_len // compressed + self.compress_k = nn.Linear(max_seq_len, self.compress_seq_len, bias=False) + if shared_kv_compressed == 0: + self.compress_v = nn.Linear( + max_seq_len, self.compress_seq_len, bias=False + ) + self.layerwise_sharing = False + else: + self.compress_k = shared_compress_layer + if shared_kv_compressed == 0: + self.compress_v = shared_compress_layer + self.layerwise_sharing = True + self.shared_kv_compressed = shared_kv_compressed + + self.out_proj = quant_noise( + nn.Linear(embed_dim, embed_dim, bias=bias), q_noise, qn_block_size + ) + + if add_bias_kv: + self.bias_k = Parameter(torch.Tensor(1, 1, embed_dim)) + self.bias_v = Parameter(torch.Tensor(1, 1, embed_dim)) + else: + self.bias_k = self.bias_v = None + + self.add_zero_attn = add_zero_attn + + self.reset_parameters() + + if freeze_compress == 1: + self.compress_k.weight.requires_grad = False + if shared_kv_compressed == 0: + self.compress_v.weight.requires_grad = False + + self.onnx_trace = False + + def prepare_for_onnx_export_(self): + self.onnx_trace = True + + def reset_parameters(self): + if self.qkv_same_dim: + # Empirically observed the convergence to be much better with + # the scaled initialization + nn.init.xavier_uniform_(self.k_proj.weight, gain=1 / math.sqrt(2)) + nn.init.xavier_uniform_(self.v_proj.weight, gain=1 / math.sqrt(2)) + nn.init.xavier_uniform_(self.q_proj.weight, gain=1 / math.sqrt(2)) + if ( + not self.layerwise_sharing + ): # otherwise, we already initialize the parameters + nn.init.xavier_uniform_(self.compress_k.weight, gain=1 / math.sqrt(2)) + if self.shared_kv_compressed == 0: + nn.init.xavier_uniform_( + self.compress_v.weight, gain=1 / math.sqrt(2) + ) + else: + nn.init.xavier_uniform_(self.k_proj.weight) + nn.init.xavier_uniform_(self.v_proj.weight) + nn.init.xavier_uniform_(self.q_proj.weight) + if ( + not self.layerwise_sharing + ): # otherwise, we already initialize the parameters + nn.init.xavier_uniform_(self.compress_k.weight) + if self.shared_kv_compressed == 0: + nn.init.xavier_uniform_(self.compress_v.weight) + + nn.init.xavier_uniform_(self.out_proj.weight) + if self.out_proj.bias is not None: + nn.init.constant_(self.out_proj.bias, 0.0) + if self.bias_k is not None: + nn.init.xavier_normal_(self.bias_k) + if self.bias_v is not None: + nn.init.xavier_normal_(self.bias_v) + + def forward( + self, + query, + key: Optional[Tensor], + value: Optional[Tensor], + key_padding_mask: Optional[Tensor] = None, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + need_weights: bool = True, + static_kv: bool = False, + attn_mask: Optional[Tensor] = None, + before_softmax: bool = False, + need_head_weights: bool = False, + ) -> Tuple[Tensor, Optional[Tensor]]: + """Input shape: Time x Batch x Channel + + Args: + key_padding_mask (ByteTensor, optional): mask to exclude + keys that are pads, of shape `(batch, src_len)`, where + padding elements are indicated by 1s. + need_weights (bool, optional): return the attention weights, + averaged over heads (default: False). + attn_mask (ByteTensor, optional): typically used to + implement causal attention, where the mask prevents the + attention from looking forward in time (default: None). + before_softmax (bool, optional): return the raw attention + weights and values before the attention softmax. + need_head_weights (bool, optional): return the attention + weights for each head. Implies *need_weights*. Default: + return the average attention weights over all heads. + """ + if need_head_weights: + need_weights = True + + tgt_len, bsz, embed_dim = query.size() + assert embed_dim == self.embed_dim + assert list(query.size()) == [tgt_len, bsz, embed_dim] + + if incremental_state is not None: + saved_state = self._get_input_buffer(incremental_state) + if saved_state is not None and "prev_key" in saved_state: + # previous time steps are cached - no need to recompute + # key and value if they are static + if static_kv: + assert self.encoder_decoder_attention and not self.self_attention + key = value = None + else: + saved_state = None + + if self.self_attention: + q = self.q_proj(query) + + k_input = query.permute(1, 2, 0).contiguous() # B * C * T + k_input = ( + F.linear(k_input, self.compress_k.weight[:, 0:tgt_len]) + .permute(2, 0, 1) + .contiguous() + ) + k = self.k_proj(k_input) + + v_input = query.permute(1, 2, 0).contiguous() # B * C * T + if self.shared_kv_compressed == 0: + v_input = ( + F.linear(v_input, self.compress_v.weight[:, 0:tgt_len]) + .permute(2, 0, 1) + .contiguous() + ) + if self.shared_kv_compressed == 1: # use shared kv compressed linear layer + v_input = ( + F.linear(v_input, self.compress_k.weight[:, 0:tgt_len]) + .permute(2, 0, 1) + .contiguous() + ) + v = self.v_proj(v_input) + elif self.encoder_decoder_attention: + # encoder-decoder attention + q = self.q_proj(query) + if key is None: + assert value is None + k = v = None + else: + k = self.k_proj(key) + v = self.v_proj(key) + + else: + assert key is not None and value is not None + q = self.q_proj(query) + k = self.k_proj(key) + v = self.v_proj(value) + q *= self.scaling + + if self.bias_k is not None: + assert self.bias_v is not None + k = torch.cat([k, self.bias_k.repeat(1, bsz, 1)]) + v = torch.cat([v, self.bias_v.repeat(1, bsz, 1)]) + if attn_mask is not None: + attn_mask = torch.cat( + [attn_mask, attn_mask.new_zeros(attn_mask.size(0), 1)], dim=1 + ) + if key_padding_mask is not None: + key_padding_mask = torch.cat( + [ + key_padding_mask, + key_padding_mask.new_zeros(key_padding_mask.size(0), 1), + ], + dim=1, + ) + + q = ( + q.contiguous() + .view(tgt_len, bsz * self.num_heads, self.head_dim) + .transpose(0, 1) + ) + if k is not None: + k = ( + k.contiguous() + .view(-1, bsz * self.num_heads, self.head_dim) + .transpose(0, 1) + ) + if v is not None: + v = ( + v.contiguous() + .view(-1, bsz * self.num_heads, self.head_dim) + .transpose(0, 1) + ) + + if saved_state is not None: + # saved states are stored with shape (bsz, num_heads, seq_len, head_dim) + if "prev_key" in saved_state: + _prev_key = saved_state["prev_key"] + assert _prev_key is not None + prev_key = _prev_key.view(bsz * self.num_heads, -1, self.head_dim) + if static_kv: + k = prev_key + else: + assert k is not None + k = torch.cat([prev_key, k], dim=1) + if "prev_value" in saved_state: + _prev_value = saved_state["prev_value"] + assert _prev_value is not None + prev_value = _prev_value.view(bsz * self.num_heads, -1, self.head_dim) + if static_kv: + v = prev_value + else: + assert v is not None + v = torch.cat([prev_value, v], dim=1) + prev_key_padding_mask: Optional[Tensor] = None + if "prev_key_padding_mask" in saved_state: + prev_key_padding_mask = saved_state["prev_key_padding_mask"] + assert k is not None and v is not None + key_padding_mask = MultiheadLinearAttention._append_prev_key_padding_mask( + key_padding_mask=key_padding_mask, + prev_key_padding_mask=prev_key_padding_mask, + batch_size=bsz, + src_len=k.size(1), + static_kv=static_kv, + ) + + saved_state["prev_key"] = k.view(bsz, self.num_heads, -1, self.head_dim) + saved_state["prev_value"] = v.view(bsz, self.num_heads, -1, self.head_dim) + saved_state["prev_key_padding_mask"] = key_padding_mask + # In this branch incremental_state is never None + assert incremental_state is not None + incremental_state = self._set_input_buffer(incremental_state, saved_state) + assert k is not None + src_len = k.size(1) + + if self.add_zero_attn: + assert v is not None + src_len += 1 + k = torch.cat([k, k.new_zeros((k.size(0), 1) + k.size()[2:])], dim=1) + v = torch.cat([v, v.new_zeros((v.size(0), 1) + v.size()[2:])], dim=1) + if attn_mask is not None: + attn_mask = torch.cat( + [attn_mask, attn_mask.new_zeros(attn_mask.size(0), 1)], dim=1 + ) + + attn_weights = torch.bmm(q, k.transpose(1, 2)) + attn_weights = MultiheadLinearAttention.apply_sparse_mask( + attn_weights, tgt_len, src_len, bsz + ) + + assert list(attn_weights.size()) == [bsz * self.num_heads, tgt_len, src_len] + + if attn_mask is not None: + attn_mask = attn_mask.unsqueeze(0) + if self.onnx_trace: + attn_mask = attn_mask.repeat(attn_weights.size(0), 1, 1) + attn_weights += attn_mask + + if before_softmax: + return attn_weights, v + + attn_weights_float = utils.softmax( + attn_weights, dim=-1, onnx_trace=self.onnx_trace + ) + attn_weights = attn_weights_float.type_as(attn_weights) + attn_probs = F.dropout( + attn_weights, + p=self.dropout, + training=self.training, + ) + assert v is not None + attn = torch.bmm(attn_probs, v) + assert list(attn.size()) == [bsz * self.num_heads, tgt_len, self.head_dim] + if self.onnx_trace and attn.size(1) == 1: + # when ONNX tracing a single decoder step (sequence length == 1) + # the transpose is a no-op copy before view, thus unnecessary + attn = attn.contiguous().view(tgt_len, bsz, embed_dim) + else: + attn = attn.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim) + attn = self.out_proj(attn) + attn_weights: Optional[Tensor] = None + if need_weights: + attn_weights = attn_weights_float.view( + bsz, self.num_heads, tgt_len, src_len + ).transpose(1, 0) + if not need_head_weights: + # average attention weights over heads + attn_weights = attn_weights.mean(dim=0) + + return attn, attn_weights + + @staticmethod + def _append_prev_key_padding_mask( + key_padding_mask: Optional[Tensor], + prev_key_padding_mask: Optional[Tensor], + batch_size: int, + src_len: int, + static_kv: bool, + ) -> Optional[Tensor]: + # saved key padding masks have shape (bsz, seq_len) + if prev_key_padding_mask is not None and static_kv: + new_key_padding_mask = prev_key_padding_mask + elif prev_key_padding_mask is not None and key_padding_mask is not None: + new_key_padding_mask = torch.cat( + [prev_key_padding_mask.float(), key_padding_mask.float()], dim=1 + ) + # During incremental decoding, as the padding token enters and + # leaves the frame, there will be a time when prev or current + # is None + elif prev_key_padding_mask is not None: + filler = torch.zeros( + (batch_size, src_len - prev_key_padding_mask.size(1)), + device=prev_key_padding_mask.device, + ) + new_key_padding_mask = torch.cat( + [prev_key_padding_mask.float(), filler.float()], dim=1 + ) + elif key_padding_mask is not None: + filler = torch.zeros( + (batch_size, src_len - key_padding_mask.size(1)), + device=key_padding_mask.device, + ) + new_key_padding_mask = torch.cat( + [filler.float(), key_padding_mask.float()], dim=1 + ) + else: + new_key_padding_mask = prev_key_padding_mask + return new_key_padding_mask + + @torch.jit.export + def reorder_incremental_state( + self, + incremental_state: Dict[str, Dict[str, Optional[Tensor]]], + new_order: Tensor, + ): + """Reorder buffered internal state (for incremental generation).""" + input_buffer = self._get_input_buffer(incremental_state) + if input_buffer is not None: + for k in input_buffer.keys(): + input_buffer_k = input_buffer[k] + if input_buffer_k is not None: + if self.encoder_decoder_attention and input_buffer_k.size( + 0 + ) == new_order.size(0): + break + input_buffer[k] = input_buffer_k.index_select(0, new_order) + incremental_state = self._set_input_buffer(incremental_state, input_buffer) + return incremental_state + + def _get_input_buffer( + self, incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] + ) -> Dict[str, Optional[Tensor]]: + result = self.get_incremental_state(incremental_state, "attn_state") + if result is not None: + return result + else: + empty_result: Dict[str, Optional[Tensor]] = {} + return empty_result + + def _set_input_buffer( + self, + incremental_state: Dict[str, Dict[str, Optional[Tensor]]], + buffer: Dict[str, Optional[Tensor]], + ): + return self.set_incremental_state(incremental_state, "attn_state", buffer) + + def apply_sparse_mask(attn_weights, tgt_len: int, src_len: int, bsz: int): + return attn_weights + + def upgrade_state_dict_named(self, state_dict, name): + prefix = name + "." if name != "" else "" + items_to_add = {} + keys_to_remove = [] + for k in state_dict.keys(): + if k.endswith(prefix + "in_proj_weight"): + # in_proj_weight used to be q + k + v with same dimensions + dim = int(state_dict[k].shape[0] / 3) + items_to_add[prefix + "q_proj.weight"] = state_dict[k][:dim] + items_to_add[prefix + "k_proj.weight"] = state_dict[k][dim : 2 * dim] + items_to_add[prefix + "v_proj.weight"] = state_dict[k][2 * dim :] + + keys_to_remove.append(k) + + k_bias = prefix + "in_proj_bias" + if k_bias in state_dict.keys(): + dim = int(state_dict[k].shape[0] / 3) + items_to_add[prefix + "q_proj.bias"] = state_dict[k_bias][:dim] + items_to_add[prefix + "k_proj.bias"] = state_dict[k_bias][ + dim : 2 * dim + ] + items_to_add[prefix + "v_proj.bias"] = state_dict[k_bias][2 * dim :] + + keys_to_remove.append(prefix + "in_proj_bias") + + for k in keys_to_remove: + del state_dict[k] + + for key, value in items_to_add.items(): + state_dict[key] = value diff --git a/examples/m2m_100/README.md b/examples/m2m_100/README.md new file mode 100644 index 0000000000000000000000000000000000000000..05801584d61afef979bf43802a167ca9da4c7a8c --- /dev/null +++ b/examples/m2m_100/README.md @@ -0,0 +1,241 @@ +# Beyond English-Centric Multilingual Machine Translation + +## Introduction +In this work, we create a true Many-to-Many multilingual translation model that can translate directly between any pair of 100 languages. Our focus on non-English-Centric models brings gains of more than 10 BLEU when directly translating between non-English directions while performing competitively with the best single systems of WMT. + +If you are new to using fairseq, read the following walkthrough. Otherwise, skip to the sections below. + +0. **Generation Data** + +To download the generation data, follow the below commands. Note that all datasets need to be detokenized *before* applying SPM in the data preprocessing step. If you use these evaluation datasets, please cite their associated papers. +```bash +# WMT - use sacrebleu, example here: +sacrebleu -t wmt14 -l fr-en --echo src > wmt.test.fr-en.fr +sacrebleu -t wmt14 -l fr-en --echo ref > wmt.test.fr-en.en + +# WAT +wget http://lotus.kuee.kyoto-u.ac.jp/WAT/my-en-data/wat2020.my-en.zip +unzip wat2020.my-en.zip + +# FLORES +# download from: https://github.com/facebookresearch/flores + +# TED - need to detokenize with Moses! +# from: https://github.com/neulab/word-embeddings-for-nmt +wget http://phontron.com/data/ted_talks.tar.gz + +# Autshumato +# request to download: https://repo.sadilar.org/handle/20.500.12185/397 + +# Tatoeba Challenge +# available here: https://github.com/Helsinki-NLP/Tatoeba-Challenge +``` + +1. **Training Data** + +To produce the training data, we use a combination of [CCMatrix](https://arxiv.org/abs/1911.04944) and [CCAligned](https://arxiv.org/abs/1911.06154). Check out the instructions [here](https://github.com/facebookresearch/LASER/tree/master/tasks/CCMatrix) to download the raw data. + +2. **Preprocess Data** + +After downloading raw data, you will need to postprocess the data, then apply SPM, then binarize. Note that it is very important you run the postprocessing script, because this removes any instance of the evaluation data in the mined training data. + +```bash +# preprocess data + +# remove sentences with more than 50% punctuation +python /path/to/fairseq/examples/m2m_100/process_data/remove_too_much_punc.py + +# deduplicate training data +paste /path/to/datadir/train.$src /path/to/datadir/train.$tgt | awk '!x[$0]++' > /path/to/datadir/train.dedup +echo "keeping $(wc -l /path/to/datadir/train.dedup) bitext out of $(wc -l /path/to/datadir/train.$src)" +cut -f1 /path/to/datadir/train.dedup > /path/to/datadir/train.$src +cut -f2 /path/to/datadir/train.dedup > /path/to/datadir/train.$tgt + +# remove all instances of evaluation data from the training data +python /path/to/fairseq/examples/m2m_100/process_data/dedup_data.py + +# frequency cleaning +wget https://dl.fbaipublicfiles.com/m2m_100/histograms.tar.gz +tar -xvzf histograms.tar.gz +python /path/to/fairseq/examples/m2m_100/process_data/clean_histogram.py --src $src --tgt $tgt --src-file /path/to/source/file --tgt-file /path/to/output/file --src-output-file source_output.$src --tgt-output-file target_output.$tgt --histograms /path/to/histograms + +# apply SPM +wget https://dl.fbaipublicfiles.com/m2m_100/spm.128k.model +python /path/to/fairseq/scripts/spm_encode.py \ + --model spm.128k.model \ + --output_format=piece \ + --inputs=/path/to/input/file/here \ + --outputs=/path/to/output/file/here + +# length ratio cleaning +perl mosesdecoder/scripts/training/clean-corpus-n.perl --ratio 3 /path/to/training/data/train.spm.$src-$tgt $src $tgt /path/to/output/directory/train.spm.$src-$tgt 1 250 + +# binarize data +wget https://dl.fbaipublicfiles.com/m2m_100/data_dict.128k.txt +fairseq-preprocess \ + --source-lang $src --target-lang $tgt \ + --testpref spm.$src.$tgt \ + --thresholdsrc 0 --thresholdtgt 0 \ + --destdir data_bin \ + --srcdict data_dict.128k.txt --tgtdict data_dict.128k.txt +``` + +3. **Training Scripts** + +To reproduce the training of our models, we train with fairseq-py's multilingual translation [task](https://github.com/pytorch/fairseq/tree/master/examples/multilingual). If you are interested in model parallel training, also check out [fairscale](https://github.com/facebookresearch/fairscale). + +4. **Generation** + +To generate from our models, follow the the commands in the generation section below. + + +If you use any of the resources listed here, please cite: +```bibtex +@article{fan2020beyond, + title={Beyond English-Centric Multilingual Machine Translation}, + author={Fan, Angela and Bhosale, Shruti and Schwenk, Holger and Ma, Zhiyi and El-Kishky, Ahmed and Goyal, Siddharth and Baines, Mandeep and Celebi, Onur and Wenzek, Guillaume and Chaudhary, Vishrav and Goyal, Naman and Birch, Tom and Liptchinsky, Vitaliy and Edunov, Sergey and Grave, Edouard and Auli, Michael and Joulin, Armand}, + journal={arXiv preprint}, + year={2020} +} + +@article{schwenk2019ccmatrix, + title={Ccmatrix: Mining billions of high-quality parallel sentences on the web}, + author={Schwenk, Holger and Wenzek, Guillaume and Edunov, Sergey and Grave, Edouard and Joulin, Armand}, + journal={arXiv preprint arXiv:1911.04944}, + year={2019} +} + +@article{el2019massive, + title={A Massive Collection of Cross-Lingual Web-Document Pairs}, + author={El-Kishky, Ahmed and Chaudhary, Vishrav and Guzman, Francisco and Koehn, Philipp}, + journal={arXiv preprint arXiv:1911.06154}, + year={2019} +} +``` + + +## Trained Models + +### 418M and 1.2B Model +We include the last checkpoint for both of these models. + +```bash +wget https://dl.fbaipublicfiles.com/m2m_100/model_dict.128k.txt +wget https://dl.fbaipublicfiles.com/m2m_100/language_pairs_small_models.txt + +# 418M parameter model +wget https://dl.fbaipublicfiles.com/m2m_100/418M_last_checkpoint.pt + +# 1.2B parameter model +wget https://dl.fbaipublicfiles.com/m2m_100/1.2B_last_checkpoint.pt + +# Generation: +fairseq-generate $binarized_data_path --batch-size 32 --path $path_to_model --fixed-dictionary model_dict.128k.txt -s en -t fr --remove-bpe 'sentencepiece' --beam 5 --task translation_multi_simple_epoch --lang-pairs language_pairs_small_models.txt --decoder-langtok --encoder-langtok src --gen-subset test > gen_out +``` + +### 12B Model +12B parameter model trained on many-to-many training data for 100 languages. We include the last checkpoint, average of last 5 checkpoints, average of last 10 checkpoints. There isn't a universally best choice out of these three, but all three versions are pretty close in accuracy. You can either sweep over the 3 checkpoints on a dev test and use the best performing checkpoint for final testing. Or the last checkpoint can be a good default choice. + +**Model Download Links** +Configuration | 2 32GB GPUs | 4 16GB GPUs | 6 12GB GPUs | 8 8GB GPUs +:--|:--|:--|:--|:-- +Last Checkpoint | [12b_last_chk_2_gpus.pt](https://dl.fbaipublicfiles.com/m2m_100/12b_last_chk_2_gpus.pt) | [12b_last_chk_4_gpus.pt](https://dl.fbaipublicfiles.com/m2m_100/12b_last_chk_4_gpus.pt) | [12b_last_chk_6_gpus.pt](https://dl.fbaipublicfiles.com/m2m_100/12b_last_chk_6_gpus.pt) | [12b_last_chk_8_gpus.pt](https://dl.fbaipublicfiles.com/m2m_100/12b_last_chk_8_gpus.pt) +Average of last 5 checkpoints | [12b_avg5_chk_2_gpus.pt](https://dl.fbaipublicfiles.com/m2m_100/12b_avg5_chk_2_gpus.pt) | [12b_avg5_chk_4_gpus.pt](https://dl.fbaipublicfiles.com/m2m_100/12b_avg5_chk_4_gpus.pt) | [12b_avg5_chk_6_gpus.pt](https://dl.fbaipublicfiles.com/m2m_100/12b_avg5_chk_6_gpus.pt) | [12b_avg5_chk_8_gpus.pt](https://dl.fbaipublicfiles.com/m2m_100/12b_avg5_chk_8_gpus.pt) +Average of last 10 checkpoints | [12b_avg10_chk_2_gpus.pt](https://dl.fbaipublicfiles.com/m2m_100/12b_avg10_chk_2_gpus.pt) | [12b_avg10_chk_4_gpus.pt](https://dl.fbaipublicfiles.com/m2m_100/12b_avg10_chk_4_gpus.pt) | [12b_avg10_chk_6_gpus.pt](https://dl.fbaipublicfiles.com/m2m_100/12b_avg10_chk_6_gpus.pt) | [12b_avg10_chk_8_gpus.pt](https://dl.fbaipublicfiles.com/m2m_100/12b_avg10_chk_8_gpus.pt) + +**Generation Arguments** +Configuration | 2 32GB GPUs | 4 16GB GPUs | 6 12GB GPUs | 8 8GB GPUs +:--|:--|:--|:--|:-- +`--pipeline-encoder-balance` | `[26]` | `[1,15,10]` | `[1,9,9,7]` | `[1,6,6,6,7]` +`--pipeline-encoder-devices` | `[0]` | `[0,1,0]` | `[0,1,2,0]` | `[0,4,5,1,0]` +`--pipeline-decoder-balance` | `[3,22,1]` | `[3,11,11,1]` | `[3,7,7,8,1]` | `[1,6,6,6,6,1]` +`--pipeline-decoder-devices` | `[0,1,0]` | `[0,2,3,0]` | `[0,3,4,5,0]` | `[0,2,6,7,3,0]` + + +## SentencePiece Model + +```bash +wget https://dl.fbaipublicfiles.com/m2m_100/spm.128k.model +``` + +## Generation with M2M-100 + +### Encode using our SentencePiece Model + +Note: Install SentencePiece from [here](https://github.com/google/sentencepiece) + +```bash +fairseq=/path/to/fairseq +cd $fairseq +sacrebleu --echo src -l de-fr -t wmt19 | head -n 20 > raw_input.de-fr.de +sacrebleu --echo ref -l de-fr -t wmt19 | head -n 20 > raw_input.de-fr.fr +wget https://dl.fbaipublicfiles.com/m2m_100/spm.128k.model +for lang in de fr ; do + python scripts/spm_encode.py \ + --model spm.128k.model \ + --output_format=piece \ + --inputs=raw_input.de-fr.${lang} \ + --outputs=spm.de-fr.${lang} +done +``` + +### Binarization + +```bash +wget https://dl.fbaipublicfiles.com/m2m_100/data_dict.128k.txt +fairseq-preprocess \ + --source-lang de --target-lang fr \ + --testpref spm.de-fr \ + --thresholdsrc 0 --thresholdtgt 0 \ + --destdir data_bin \ + --srcdict data_dict.128k.txt --tgtdict data_dict.128k.txt +``` + +### Generation for the 12B model + +Note that generation can currently be run using 2 32GB / 4 16GB / 6 12GB / 8 8GB GPUs, and the corresponding model checkpoints and pipeline arguments can be found in the [12B Model Section](#12b-model). +Generation on CPUs will be added in the future. + +```bash +wget https://dl.fbaipublicfiles.com/m2m_100/model_dict.128k.txt +wget https://dl.fbaipublicfiles.com/m2m_100/language_pairs.txt +wget https://dl.fbaipublicfiles.com/m2m_100/12b_last_chk_4_gpus.pt +fairseq-generate \ + data_bin \ + --batch-size 1 \ + --path 12b_last_chk_4_gpus.pt \ + --fixed-dictionary model_dict.128k.txt \ + -s de -t fr \ + --remove-bpe 'sentencepiece' \ + --beam 5 \ + --task translation_multi_simple_epoch \ + --lang-pairs language_pairs.txt \ + --decoder-langtok --encoder-langtok src \ + --gen-subset test \ + --fp16 \ + --dataset-impl mmap \ + --distributed-world-size 1 --distributed-no-spawn \ + --pipeline-model-parallel \ + --pipeline-chunks 1 \ + --pipeline-encoder-balance '[1,15,10]' \ + --pipeline-encoder-devices '[0,1,0]' \ + --pipeline-decoder-balance '[3,11,11,1]' \ + --pipeline-decoder-devices '[0,2,3,0]' > gen_out +``` +## Evaluation with M2M-100 + +### Tokenization + +Note: Refer to tokenizers/README.md for more details on tokenization. + +```bash +cd ${fairseq}/examples/m2m_100 +cat ${fairseq}/gen_out | grep -P "^H" | sort -V | cut -f 3- | sh tok.sh fr > hyp +cat ${fairseq}/raw_input.de-fr.fr | sh tok.sh fr > ref +``` + +### BLEU + +```bash +sacrebleu -tok 'none' ref < hyp +``` diff --git a/examples/m2m_100/install_dependecies.sh b/examples/m2m_100/install_dependecies.sh new file mode 100755 index 0000000000000000000000000000000000000000..82a1054745264a56fbec4a8eb593884f8a42bd08 --- /dev/null +++ b/examples/m2m_100/install_dependecies.sh @@ -0,0 +1,78 @@ +#!/usr/bin/env bash +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +CWD=`pwd` +INSTALL_PATH=$CWD/tokenizers/thirdparty + +MOSES=$INSTALL_PATH/mosesdecoder +if [ ! -d $MOSES ]; then + echo 'Cloning Moses github repository (for tokenization scripts)...' + git clone https://github.com/moses-smt/mosesdecoder.git $MOSES + cd $MOSES + # To deal with differences in handling ' vs " + git checkout 03578921cc1a03402 + cd - +fi + +WMT16_SCRIPTS=$INSTALL_PATH/wmt16-scripts +if [ ! -d $WMT16_SCRIPTS ]; then + echo 'Cloning Romanian tokenization scripts' + git clone https://github.com/rsennrich/wmt16-scripts.git $WMT16_SCRIPTS +fi + +KYTEA=$INSTALL_PATH/kytea +if [ ! -f $KYTEA/bin/kytea ]; then + git clone https://github.com/neubig/kytea.git $KYTEA + cd $KYTEA + autoreconf -i + ./configure --prefix=`pwd` + make + make install + cd .. +fi + +export MECAB=$INSTALL_PATH/mecab-0.996-ko-0.9.2 +if [ ! -f $MECAB/bin/mecab ]; then + cd $INSTALL_PATH + curl -LO https://bitbucket.org/eunjeon/mecab-ko/downloads/mecab-0.996-ko-0.9.2.tar.gz + tar zxfv mecab-0.996-ko-0.9.2.tar.gz + cd mecab-0.996-ko-0.9.2/ + ./configure --prefix=`pwd` + make + make install + + cd .. + curl -LO https://bitbucket.org/eunjeon/mecab-ko-dic/downloads/mecab-ko-dic-2.1.1-20180720.tar.gz + tar zxfv mecab-ko-dic-2.1.1-20180720.tar.gz + cd mecab-ko-dic-2.1.1-20180720/ + ./autogen.sh + ./configure --prefix=`pwd` --with-dicdir=$MECAB/lib/mecab/dic/mecab-ko-dic --with-mecab-config=$MECAB/bin/mecab-config + make + sh -c 'echo "dicdir=$MECAB/lib/mecab/dic/mecab-ko-dic" > $MECAB/etc/mecabrc' + make install + cd $CWD +fi + +INDIC_RESOURCES_PATH=$INSTALL_PATH/indic_nlp_resources +if [ ! -d $INDIC_RESOURCES_PATH ]; then + echo 'Cloning indic_nlp_resources' + git clone https://github.com/anoopkunchukuttan/indic_nlp_resources.git $INDIC_RESOURCES_PATH +fi + + +if [ ! -f $INSTALL_PATH/seg_my.py ]; then + cd $INSTALL_PATH + wget http://lotus.kuee.kyoto-u.ac.jp/WAT/my-en-data/wat2020.my-en.zip + unzip wat2020.my-en.zip + # switch to python3 + cat wat2020.my-en/myseg.py |sed 's/^sys.std/###sys.std/g' | sed 's/### sys/sys/g' | sed 's/unichr/chr/g' > seg_my.py + cd $CWD +fi + + +pip install pythainlp sacrebleu indic-nlp-library + diff --git a/examples/m2m_100/process_data/clean_histogram.py b/examples/m2m_100/process_data/clean_histogram.py new file mode 100644 index 0000000000000000000000000000000000000000..e24e073dc0eb43c76e2ce717f52bb848c5b026b8 --- /dev/null +++ b/examples/m2m_100/process_data/clean_histogram.py @@ -0,0 +1,52 @@ +import argparse + +parser = argparse.ArgumentParser() +parser.add_argument('--src', type=str, help='Source language') +parser.add_argument('--tgt', type=str, help='Target language') +parser.add_argument('--src-file', type=str, help='Input source file') +parser.add_argument('--tgt-file', type=str, help='Input target file') +parser.add_argument('--src-output-file', type=str, help='Output source file') +parser.add_argument('--tgt-output-file', type=str, help='Output target file') +parser.add_argument('--threshold', type=float, default=0.5, help='Threshold') +parser.add_argument('--threshold-character', type=str, default=']', help='Threshold character') +parser.add_argument('--histograms', type=str, help='Path to histograms') + +args = parser.parse_args() + + +def read_hist(f): + ch = [] + for line in f: + c = line[0] + if c == args.threshold_character: + break + ch.append(c) + return ch + + +with(open("{}/{}".format(args.histograms, args.src), 'r', encoding='utf8')) as f: + ch1 = read_hist(f) + +with(open("{}/{}".format(args.histograms, args.tgt), 'r', encoding='utf8')) as f: + ch2 = read_hist(f) + +print("Accepted characters for {}: {}".format(args.src, ch1)) +print("Accepted characters for {}: {}".format(args.tgt, ch2)) + +with open(args.src_file, 'r', encoding='utf8') as fs1, open(args.tgt_file, 'r', encoding='utf8') as fs2, open(args.src_output_file, 'w', encoding='utf8') as fos1, open(args.tgt_output_file, 'w', encoding='utf8') as fos2: + ls1 = fs1.readline() + ls2 = fs2.readline() + + while ls1 or ls2: + cnt1 = len([c for c in ls1.strip() if c in ch1]) + cnt2 = len([c for c in ls2.strip() if c in ch2]) + + if cnt1 / len(ls1) > args.threshold and cnt2 / len(ls2) > args.threshold: + fos1.write(ls1) + fos2.write(ls2) + else: + print("{} {} {} \n{} {} {}".format(args.src, cnt1 / len(ls1), ls1.strip(), args.tgt, cnt2 / len(ls2), ls2.strip())) + + ls1 = fs1.readline() + ls2 = fs2.readline() + \ No newline at end of file diff --git a/examples/m2m_100/process_data/dedup_data.py b/examples/m2m_100/process_data/dedup_data.py new file mode 100644 index 0000000000000000000000000000000000000000..58d9ed1cd17b3ba70772a6d9adab709785495fd9 --- /dev/null +++ b/examples/m2m_100/process_data/dedup_data.py @@ -0,0 +1,91 @@ +import argparse +from collections import namedtuple +import os + +DATADIR = "/path/to/train_data" +DEDUP_FROM_DIR = "/path/to/eval/data" +OUTPUT_DIR = "/path/to/output/data" + + +def main(args): + languages = set() + for language_directory in os.listdir(DATADIR): + if "_" in language_directory: + src, tgt = language_directory.split("_") + languages.add(LanguagePair(src=src, tgt=tgt)) + + data = existing_data() + train_languages = sorted(languages) + for language_pair in train_languages[args.start_index:args.start_index + args.size]: + print(language_pair) + dedup(language_pair, data) + + +LanguagePair = namedtuple("LanguagePair", ["src", "tgt"]) + + +def existing_data(): + data = set() + for file in os.listdir(DEDUP_FROM_DIR): + with open(os.path.join(DEDUP_FROM_DIR, file)) as f: + data |= set(f.readlines()) + return data + +def dedup(language_pair, data, verbose=True, output=True): + train_filenames = LanguagePair( + src=f"{DATADIR}/{language_pair.src}_{language_pair.tgt}/train.{language_pair.src}", + tgt=f"{DATADIR}/{language_pair.src}_{language_pair.tgt}/train.{language_pair.tgt}", + ) + + output_filenames = LanguagePair( + src=f"{OUTPUT_DIR}/train.dedup.{language_pair.src}-{language_pair.tgt}.{language_pair.src}", + tgt=f"{OUTPUT_DIR}/train.dedup.{language_pair.src}-{language_pair.tgt}.{language_pair.tgt}" + ) + + # If output exists, skip this pair. It has already been done. + if (os.path.exists(output_filenames.src) and + os.path.exists(output_filenames.tgt)): + if verbose: + print(f"{language_pair.src}-{language_pair.tgt} already done.") + return + + if verbose: + print(f"{language_pair.src}-{language_pair.tgt} ready, will check dups.") + + # If there is no output, no need to actually do the loop. + if not output: + return + + if os.path.exists(train_filenames.src) and os.path.exists(train_filenames.tgt): + with open(train_filenames.src) as f: + train_source = f.readlines() + + with open(train_filenames.tgt) as f: + train_target = f.readlines() + + # do dedup + new_train_source = [] + new_train_target = [] + for i, train_line in enumerate(train_source): + if train_line not in data and train_target[i] not in data: + new_train_source.append(train_line) + new_train_target.append(train_target[i]) + + assert len(train_source) == len(train_target) + assert len(new_train_source) == len(new_train_target) + assert len(new_train_source) <= len(train_source) + + with open(output_filenames.src, "w") as o: + for line in new_train_source: + o.write(line) + + with open(output_filenames.tgt, "w") as o: + for line in new_train_target: + o.write(line) + + +if __name__ == '__main__': + parser = argparse.ArgumentParser() + parser.add_argument("-s", "--start-index", required=True, type=int) + parser.add_argument("-n", "--size", required=True, type=int) + main(parser.parse_args()) diff --git a/examples/m2m_100/process_data/remove_too_much_punc.py b/examples/m2m_100/process_data/remove_too_much_punc.py new file mode 100644 index 0000000000000000000000000000000000000000..6c280de2403daffab477ac88e2008a68b9e61ff0 --- /dev/null +++ b/examples/m2m_100/process_data/remove_too_much_punc.py @@ -0,0 +1,36 @@ +import gzip +import argparse +from string import punctuation + +def len_no_punc(s, punc): + return len([ch for ch in s if ch in punc]) + +def filter_overpunc(len_npunc, len_sen): + return len_npunc < 0.5*len_sen + +def main(args): + punc = punctuation + "—|–" + print('Processing file {}'.format(args.input)) + with gzip.open(args.input, 'rt', encoding=args.encoding) as tsv: + with open(args.bitext + '.' + args.src_lang, 'wt', encoding=args.encoding) as fsrc: + with open(args.bitext + '.' + args.tgt_lang, 'wt', encoding=args.encoding) as ftgt: + line = tsv.readline() + fields = line.split('\t') + + src, tgt = fields[1], fields[2] + + nchar_npunc_src = len_no_punc(src, punc) + nchar_npunc_tgt = len_no_punc(tgt, punc) + + if filter_overpunc(nchar_npunc_src, len(src)) and filter_overpunc(nchar_npunc_tgt, len(tgt)): + fsrc.write(src.strip() + '\n') + ftgt.write(tgt.strip() + '\n') + +if __name__ == '__main__': + parser = argparse.ArgumentParser() + parser.add_argument("--input", required=True, type=str) + parser.add_argument('--encoding', default='utf-8', help='character encoding for input/output') + parser.add_argument('--bitext', type=str, required=True, help='language direction') + parser.add_argument('--src-lang', type=str, required=True, help='Source language') + parser.add_argument('--tgt-lang', type=str, required=True, help='Target language') + main(parser.parse_args()) diff --git a/examples/m2m_100/tok.sh b/examples/m2m_100/tok.sh new file mode 100755 index 0000000000000000000000000000000000000000..ba2ec5a2f3f4794d2e528d3a6574bf05abe1d043 --- /dev/null +++ b/examples/m2m_100/tok.sh @@ -0,0 +1,83 @@ +#!/usr/bin/env bash +# Copyright (c) 2019-present, Facebook, Inc. +# All rights reserved. +# +# This source code is licensed under the license found in the +# LICENSE file in the root directory of this source tree. +# + +set -e + +TOKENIZERS_SCRIPTS=tokenizers +INSTALL_PATH=$TOKENIZERS_SCRIPTS/thirdparty + +N_THREADS=8 + +lg=$1 + +MOSES=$INSTALL_PATH/mosesdecoder +REPLACE_UNICODE_PUNCT=$MOSES/scripts/tokenizer/replace-unicode-punctuation.perl +NORM_PUNC=$MOSES/scripts/tokenizer/normalize-punctuation.perl +REM_NON_PRINT_CHAR=$MOSES/scripts/tokenizer/remove-non-printing-char.perl +TOKENIZER=$MOSES/scripts/tokenizer/tokenizer.perl + +# special tokenization for Romanian +WMT16_SCRIPTS=$INSTALL_PATH/wmt16-scripts + +NORMALIZE_ROMANIAN=$WMT16_SCRIPTS/preprocess/normalise-romanian.py +REMOVE_DIACRITICS=$WMT16_SCRIPTS/preprocess/remove-diacritics.py + +# Burmese +MY_SEGMENT=$INSTALL_PATH/seg_my.py + +# Arabic +AR_TOKENIZER=$TOKENIZERS_SCRIPTS/tokenizer_ar.sh + +# Korean +KO_SEGMENT=$TOKENIZERS_SCRIPTS/seg_ko.sh + +# Japanese +JA_SEGMENT=$TOKENIZERS_SCRIPTS/seg_ja.sh + +# Indic +IN_TOKENIZER=$TOKENIZERS_SCRIPTS/tokenize_indic.py +INDIC_RESOURCES_PATH=$INSTALL_PATH/indic_nlp_resources + +# Thai +THAI_TOKENIZER=$TOKENIZERS_SCRIPTS/tokenize_thai.py + +# Chinese +CHINESE_TOKENIZER=$TOKENIZERS_SCRIPTS/tokenize_zh.py + +# Chinese +if [ "$lg" = "zh" ]; then + cat - | $REPLACE_UNICODE_PUNCT | $NORM_PUNC -l $lg | $REM_NON_PRINT_CHAR | python $CHINESE_TOKENIZER +# Thai +elif [ "$lg" = "th" ]; then + cat - | python $THAI_TOKENIZER +# Japanese +elif [ "$lg" = "ja" ]; then + cat - | $REPLACE_UNICODE_PUNCT | $NORM_PUNC -l $lg | $REM_NON_PRINT_CHAR | ${JA_SEGMENT} +# Korean +elif [ "$lg" = "ko" ]; then + cat - | $REM_NON_PRINT_CHAR | ${KO_SEGMENT} +# Romanian +elif [ "$lg" = "ro" ]; then + cat - | $REPLACE_UNICODE_PUNCT | $NORM_PUNC -l $lg | $REM_NON_PRINT_CHAR | $NORMALIZE_ROMANIAN | $REMOVE_DIACRITICS | $TOKENIZER -no-escape -threads $N_THREADS -l $lg +# Burmese +elif [ "$lg" = "my" ]; then + cat - | python ${MY_SEGMENT} +# Arabic +elif [ "$lg" = "ar" ]; then + cat - | ${AR_TOKENIZER} +# Indic +elif [ "$lg" = "ne" ]; then + cat - | python ${IN_TOKENIZER} $lg +elif [ "$lg" = "si" ]; then + cat - | python ${IN_TOKENIZER} $lg +elif [ "$lg" = "hi" ]; then + cat - | python ${IN_TOKENIZER} $lg +# other languages +else + cat - | $REPLACE_UNICODE_PUNCT | $NORM_PUNC -l $lg | $REM_NON_PRINT_CHAR | $TOKENIZER -no-escape -threads $N_THREADS -l $lg +fi diff --git a/examples/m2m_100/tokenizers/README.md b/examples/m2m_100/tokenizers/README.md new file mode 100644 index 0000000000000000000000000000000000000000..e116932bc80572f221cff6472a7b1eea7032925d --- /dev/null +++ b/examples/m2m_100/tokenizers/README.md @@ -0,0 +1,18 @@ +# M2M-100 Tokenization + +We apply different tokenization strategies for different languages following the existing literature. Here we provide tok.sh a tokenizer that can be used to reproduce our results. + +To reproduce the results, follow these steps: + +``` +tgt_lang=... +reference_translation=... +cat generation_output | grep -P "^H" | sort -V | cut -f 3- | sh tok.sh $tgt_lang > hyp +cat $reference_translation |sh tok.sh $tgt_lang > ref +sacrebleu -tok 'none' ref < hyp +``` + +## Installation + +Tools needed for all the languages except Arabic can be installed by running install_dependencies.sh +If you want to evaluate Arabic models, please follow the instructions provided here: http://alt.qcri.org/tools/arabic-normalizer/ to install diff --git a/examples/m2m_100/tokenizers/seg_ja.sh b/examples/m2m_100/tokenizers/seg_ja.sh new file mode 100755 index 0000000000000000000000000000000000000000..be6f5ca5fe4ac8e8c786a439caaed1d1314f1aef --- /dev/null +++ b/examples/m2m_100/tokenizers/seg_ja.sh @@ -0,0 +1,11 @@ +#!/usr/bin/env bash +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +SCRIPT=`realpath $0` +KYTEA=`dirname $SCRIPT`/thirdparty/kytea +export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$KYTEA/lib:/usr/local/lib +export PATH=$PATH:"$KYTEA/bin" + +cat - | tr -d "[:blank:]" | kytea -notags diff --git a/examples/m2m_100/tokenizers/seg_ko.sh b/examples/m2m_100/tokenizers/seg_ko.sh new file mode 100755 index 0000000000000000000000000000000000000000..c523d92634d9b61b97bbcdbfd17dfc33465bfc09 --- /dev/null +++ b/examples/m2m_100/tokenizers/seg_ko.sh @@ -0,0 +1,12 @@ +#!/usr/bin/env bash +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +SCRIPT=`realpath $0` +MECAB=`dirname $SCRIPT`/thirdparty/mecab-0.996-ko-0.9.2 + +export PATH=$PATH:"$MECAB/bin":"$MECAB/lib" +export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:"$MECAB/lib" + +cat - | mecab -O wakati diff --git a/examples/m2m_100/tokenizers/thirdparty/.gitignore b/examples/m2m_100/tokenizers/thirdparty/.gitignore new file mode 100644 index 0000000000000000000000000000000000000000..19eb6a9dd705ac583f22ecb60d9b744987e27ff1 --- /dev/null +++ b/examples/m2m_100/tokenizers/thirdparty/.gitignore @@ -0,0 +1,12 @@ +seg_my.py +indic_nlp_library/ +indic_nlp_resources/ +kytea/ +mecab-0.996-ko-0.9.2.tar.gz +mecab-0.996-ko-0.9.2/ +mosesdecoder/ +wat2020.my-en.zip +wat2020.my-en/ +wmt16-scripts/ +mecab-ko-dic-2.1.1-20180720/ +mecab-ko-dic-2.1.1-20180720.tar.gz \ No newline at end of file diff --git a/examples/m2m_100/tokenizers/tokenize_indic.py b/examples/m2m_100/tokenizers/tokenize_indic.py new file mode 100644 index 0000000000000000000000000000000000000000..a44fad07f7c718f99cccd445f33c62b0e3c562f4 --- /dev/null +++ b/examples/m2m_100/tokenizers/tokenize_indic.py @@ -0,0 +1,23 @@ +#!/usr/bin/env python3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +# Use: echo {text} | python tokenize_indic.py {language} + +import sys + +from indicnlp.normalize.indic_normalize import IndicNormalizerFactory +from indicnlp.tokenize.indic_tokenize import trivial_tokenize + + +factory = IndicNormalizerFactory() +normalizer = factory.get_normalizer( + sys.argv[1], remove_nuktas=False, nasals_mode="do_nothing" +) + +for line in sys.stdin: + normalized_line = normalizer.normalize(line.strip()) + tokenized_line = " ".join(trivial_tokenize(normalized_line, sys.argv[1])) + print(tokenized_line) diff --git a/examples/m2m_100/tokenizers/tokenize_thai.py b/examples/m2m_100/tokenizers/tokenize_thai.py new file mode 100644 index 0000000000000000000000000000000000000000..9c72cb89056f6fc92a8963415e5f3a1e61b33a5b --- /dev/null +++ b/examples/m2m_100/tokenizers/tokenize_thai.py @@ -0,0 +1,13 @@ +#!/usr/bin/env python3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import sys + +from pythainlp import word_tokenize + + +for line in sys.stdin: + print(" ".join(word_tokenize(line.strip()))) diff --git a/examples/m2m_100/tokenizers/tokenize_zh.py b/examples/m2m_100/tokenizers/tokenize_zh.py new file mode 100644 index 0000000000000000000000000000000000000000..674b5849cba829cf4f07a69369e9cc6eed376d4c --- /dev/null +++ b/examples/m2m_100/tokenizers/tokenize_zh.py @@ -0,0 +1,14 @@ +#!/usr/bin/env python3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +import fileinput + +import sacrebleu + + +for line in fileinput.input(): + print(sacrebleu.tokenize_zh(line)) diff --git a/examples/m2m_100/tokenizers/tokenizer_ar.sh b/examples/m2m_100/tokenizers/tokenizer_ar.sh new file mode 100755 index 0000000000000000000000000000000000000000..ad35d7adf28dc9b23d13a6a3fec0b12cb760e855 --- /dev/null +++ b/examples/m2m_100/tokenizers/tokenizer_ar.sh @@ -0,0 +1,27 @@ +#!/usr/bin/env sh +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +# +# Please follow the instructions here http://alt.qcri.org/tools/arabic-normalizer/ +# to install tools needed for Arabic + +echo "Please install Arabic tools: http://alt.qcri.org/tools/arabic-normalizer/" +echo "Then update environment variables in tokenizer_ar.sh" +exit 1 + +SVMTOOL=... +GOMOSESGO=... +QCRI_ARABIC_NORMALIZER=... + +export PERL5LIB="$SVMTOOL/lib":"$GOMOSESGO/bin/MADA-3.2":$PERL5LIB + + +tempfile=$(mktemp) +cat - > $tempfile + +cd $QCRI_ARABIC_NORMALIZER + +bash qcri_normalizer_mada3.2_aramorph1.2.1.sh $tempfile +cat $tempfile.mada_norm-aramorph.europarl_tok diff --git a/examples/mbart/README.md b/examples/mbart/README.md new file mode 100644 index 0000000000000000000000000000000000000000..a45e37243c2c5d4027f79cf71498ca58bbac7d98 --- /dev/null +++ b/examples/mbart/README.md @@ -0,0 +1,123 @@ +# MBART: Multilingual Denoising Pre-training for Neural Machine Translation +[https://arxiv.org/abs/2001.08210] + +## Introduction + +MBART is a sequence-to-sequence denoising auto-encoder pre-trained on large-scale monolingual corpora in many languages using the BART objective. mBART is one of the first methods for pre-training a complete sequence-to-sequence model by denoising full texts in multiple languages, while previous approaches have focused only on the encoder, decoder, or reconstructing parts of the text. + +## Pre-trained models + +Model | Description | # params | Download +---|---|---|--- +`mbart.CC25` | mBART model with 12 encoder and decoder layers trained on 25 languages' monolingual corpus | 610M | [mbart.CC25.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/mbart/mbart.cc25.v2.tar.gz) +`mbart.ft.ro_en` | finetune mBART cc25 model on ro-en language pairs | 610M | [mbart.cc25.ft.enro.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/mbart/mbart.cc25.ft.enro.tar.gz) + +## Results + +**[WMT16 EN-RO](https://www.statmt.org/wmt16/translation-task.html)** + +_(test set, no additional data used)_ + +Model | en-ro | ro-en +---|---|--- +`Random` | 34.3 | 34.0 +`mbart.cc25` | 37.7 | 37.8 +`mbart.enro.bilingual` | 38.5 | 38.5 + +## BPE data +# download model +wget https://dl.fbaipublicfiles.com/fairseq/models/mbart/mbart.cc25.v2.tar.gz +tar -xzvf mbart.CC25.tar.gz +# bpe data +install SPM [here](https://github.com/google/sentencepiece) +```bash +SPM=/path/to/sentencepiece/build/src/spm_encode +MODEL=sentence.bpe.model +${SPM} --model=${MODEL} < ${DATA}/${TRAIN}.${SRC} > ${DATA}/${TRAIN}.spm.${SRC} & +${SPM} --model=${MODEL} < ${DATA}/${TRAIN}.${TGT} > ${DATA}/${TRAIN}.spm.${TGT} & +${SPM} --model=${MODEL} < ${DATA}/${VALID}.${SRC} > ${DATA}/${VALID}.spm.${SRC} & +${SPM} --model=${MODEL} < ${DATA}/${VALID}.${TGT} > ${DATA}/${VALID}.spm.${TGT} & +${SPM} --model=${MODEL} < ${DATA}/${TEST}.${SRC} > ${DATA}/${TEST}.spm.${SRC} & +${SPM} --model=${MODEL} < ${DATA}/${TEST}.${TGT} > ${DATA}/${TEST}.spm.${TGT} & +``` + +## Preprocess data + +```bash +DICT=dict.txt +fairseq-preprocess \ + --source-lang ${SRC} \ + --target-lang ${TGT} \ + --trainpref ${DATA}/${TRAIN}.spm \ + --validpref ${DATA}/${VALID}.spm \ + --testpref ${DATA}/${TEST}.spm \ + --destdir ${DEST}/${NAME} \ + --thresholdtgt 0 \ + --thresholdsrc 0 \ + --srcdict ${DICT} \ + --tgtdict ${DICT} \ + --workers 70 +``` + +## Finetune on EN-RO +Finetune on mbart CC25 + +```bash +PRETRAIN=mbart.cc25 # fix if you moved the downloaded checkpoint +langs=ar_AR,cs_CZ,de_DE,en_XX,es_XX,et_EE,fi_FI,fr_XX,gu_IN,hi_IN,it_IT,ja_XX,kk_KZ,ko_KR,lt_LT,lv_LV,my_MM,ne_NP,nl_XX,ro_RO,ru_RU,si_LK,tr_TR,vi_VN,zh_CN + +fairseq-train path_2_data \ + --encoder-normalize-before --decoder-normalize-before \ + --arch mbart_large --layernorm-embedding \ + --task translation_from_pretrained_bart \ + --source-lang en_XX --target-lang ro_RO \ + --criterion label_smoothed_cross_entropy --label-smoothing 0.2 \ + --optimizer adam --adam-eps 1e-06 --adam-betas '(0.9, 0.98)' \ + --lr-scheduler polynomial_decay --lr 3e-05 --warmup-updates 2500 --total-num-update 40000 \ + --dropout 0.3 --attention-dropout 0.1 --weight-decay 0.0 \ + --max-tokens 1024 --update-freq 2 \ + --save-interval 1 --save-interval-updates 5000 --keep-interval-updates 10 --no-epoch-checkpoints \ + --seed 222 --log-format simple --log-interval 2 \ + --restore-file $PRETRAIN \ + --reset-optimizer --reset-meters --reset-dataloader --reset-lr-scheduler \ + --langs $langs \ + --ddp-backend legacy_ddp +``` +## Generate on EN-RO +Get sacrebleu on finetuned en-ro model + +get tokenizer [here](https://github.com/rsennrich/wmt16-scripts) +```bash +wget https://dl.fbaipublicfiles.com/fairseq/models/mbart/mbart.cc25.ft.enro.tar.gz +tar -xzvf mbart.cc25.ft.enro.tar.gz +``` + +```bash +model_dir=MBART_finetuned_enro # fix if you moved the checkpoint + +fairseq-generate path_2_data \ + --path $model_dir/model.pt \ + --task translation_from_pretrained_bart \ + --gen-subset test \ + -t ro_RO -s en_XX \ + --bpe 'sentencepiece' --sentencepiece-model $model_dir/sentence.bpe.model \ + --sacrebleu --remove-bpe 'sentencepiece' \ + --batch-size 32 --langs $langs > en_ro + +cat en_ro | grep -P "^H" |sort -V |cut -f 3- | sed 's/\[ro_RO\]//g' |$TOKENIZER ro > en_ro.hyp +cat en_ro | grep -P "^T" |sort -V |cut -f 2- | sed 's/\[ro_RO\]//g' |$TOKENIZER ro > en_ro.ref +sacrebleu -tok 'none' -s 'none' en_ro.ref < en_ro.hyp +``` + +## Citation + +```bibtex +@article{liu2020multilingual, + title={Multilingual Denoising Pre-training for Neural Machine Translation}, + author={Yinhan Liu and Jiatao Gu and Naman Goyal and Xian Li and Sergey Edunov and Marjan Ghazvininejad and Mike Lewis and Luke Zettlemoyer}, + year={2020}, + eprint={2001.08210}, + archivePrefix={arXiv}, + primaryClass={cs.CL} +} +``` diff --git a/examples/megatron_11b/README.md b/examples/megatron_11b/README.md new file mode 100644 index 0000000000000000000000000000000000000000..945c96c91e2e2d93466abc28d90bc25a1e7dd471 --- /dev/null +++ b/examples/megatron_11b/README.md @@ -0,0 +1,161 @@ +# Megatron-11b + +Megatron-11b is a unidirectional language model with `11B` parameters based on [Megatron-LM](https://arxiv.org/pdf/1909.08053.pdf). Following the original Megatron work, we trained the model using intra-layer model parallelism with each layer's parameters split across 8 GPUs. + +Megatron-11b is trained on the same data and uses the same byte-pair encoding (BPE) as [RoBERTa](https://arxiv.org/pdf/1907.11692.pdf). + +## Pre-trained models + +Model | Description | # params | # filesize | Download +---|---|---|---|--- +`megatron_11b` | megatron_11b unidirectional language model | 11B | 19Gb | [megatron_11b.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/model_parallel/megatron_11b.tar.gz) + +#### Architecture: + +Param | Value +---|--- +embed_dim | 3072 +ffn_dim | 3072 * 6 +layers | 72 +attention heads | 32 + +#### Training details: + +Param | value +---|--- +bsz | 512 +num_updates | 300,000 +peak_lr | 1.5e-04 +lr scheduler | inverse_sqrt +clip norm | 0.0 + + +## Example training command (model parallel) + +Megatron-11b contains too many parameters to train on a single GPU. Following +the original Megatron work, we adopt an intra-layer model parallel training +approach in which each layer's parameters are split across multiple GPUs and +activations and gradients are communicated during the forward/backward pass, +respectively. We similarly split the loss computation using the +`vocab_parallel_cross_entropy` criterion. + +The following training command illustrates how to do model parallel training in +fairseq. We assume that each machine (node) has 8 GPUs among which to split the +model parameters (`--model-parallel-size 8`). If you have access to multiple +nodes, you may combine this with data parallel training by increasing +`--distributed-world-size`. + +To train Megatron-11b on a single node: + + +```bash +fairseq-train \ + --distributed-world-size 8 \ + --memory-efficient-fp16 \ + --num-workers 2 \ + --model-parallel-size 8 \ + --criterion vocab_parallel_cross_entropy \ + --task language_modeling \ + --sample-break-mode none \ + --tokens-per-sample 1024 \ + --arch transformer_lm_megatron_11b \ + --share-decoder-input-output-embed \ + --optimizer adam --adam-betas "(0.9, 0.98)" --adam-eps 1e-08 --clip-norm 0.0 \ + --lr-scheduler inverse_sqrt --lr 0.00015 \ + --warmup-updates 3000 --weight-decay 0.01 \ + --dropout 0.1 --attention-dropout 0.1 \ + --batch-size 2 \ + --max-update 300000; +``` + +Note: Above was tested on `DGX-1` box, with `8xV100-32Gb` GPUs. + +## Results + +**[Wikitext103](https://blog.einstein.ai/the-wikitext-long-term-dependency-language-modeling-dataset/)** + +Model | Valid perplexity | Test perplexity +---|---|--- +`megatron_11b` | 10.64 | 10.54 + + +## Evaluating `megatron_11b` on Wikitext-103 + +#### 1. Downloading Megatron-11b +```bash +# WARNING: this file is 19GB +wget https://dl.fbaipublicfiles.com/fairseq/models/model_parallel/megatron_11b.tar.gz +tar -xzvf megatron_11b.tar.gz +``` + +#### 2. Download Wikitext-103 +```bash +wget https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-103-raw-v1.zip +unzip wikitext-103-raw-v1.zip +``` + +#### 3. Detokenize test tokens +Megatron-11b uses a byte-level BPE that expects raw (untokenized) input. Since +the wikitext-103 dataset comes tokenized, we apply a simple detokenization +process to restore the untokenized test set: + +```bash +python -m examples.megatron_11b.detok wikitext-103-raw/wiki.test.raw > wikitext-103-raw/wiki.test.detok +``` + +#### 4. BPE encoding +```bash +wget -N 'https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/encoder.json' +wget -N 'https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/vocab.bpe' + +python -m examples.roberta.multiprocessing_bpe_encoder \ + --encoder-json encoder.json \ + --vocab-bpe vocab.bpe \ + --inputs "wikitext-103-raw/wiki.test.detok" \ + --outputs "wikitext-103-raw/wiki.test.bpe" \ + --workers 60; +``` + +#### 5. Fairseq binarize +```bash +fairseq-preprocess \ + --only-source \ + --testpref wikitext-103-raw/wiki.test.bpe \ + --srcdict megatron_11b/dict.txt \ + --destdir wikitext103-bin; +``` + +#### 6. Evaluating perplexity. +We can now evaluate perplexity on the test set. Note that because we've modified +the test set (via detokenization and BPE), the perplexity reported by +`fairseq-eval-lm` needs to be renormalized. + +Compute unnormalized perplexity: + +```bash +DATA_PATH=wikitext103-bin/ +fairseq-eval-lm \ + $DATA_PATH \ + --path megatron_11b/model.pt \ + --task language_modeling \ + --gen-subset test \ + --batch-size 8 \ + --criterion cross_entropy \ + --context-window 992 \ + --distributed-world-size 8 \ + --model-parallel-size 8; +# Expected PPL (unnormalized_ppl): [8.46] +# Note: the eval command needs to run on 8 GPUs for the released model +``` +Renormalizing formula: `2 ^ ( log_2(unnormalized_PPL) * (270847 / 245566))`. +PPL After normalization: `10.54` + +To renormalize the perplexity, we must account for the change in token count +after detokenizing and appling BPE. The formula for this is: +`2 ^ ( log_2(unnormalized_PPL) * (new_token_cnt / orig_token_cnt))` + +For the wikitext-103 test set, the original token count is `245566` and the +token count after detokenization and applying BPE is `270847`. + +The perplexity after renormalization is: +`2 ^ ( log_2(8.46) * (270847 / 245566)) = 10.54` diff --git a/examples/megatron_11b/detok.py b/examples/megatron_11b/detok.py new file mode 100644 index 0000000000000000000000000000000000000000..49921b28a1f35c6216b5ed85729453524e7a049d --- /dev/null +++ b/examples/megatron_11b/detok.py @@ -0,0 +1,32 @@ +#!/usr/bin/env python3 -u +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import fileinput + +import sacremoses + + +def main(): + parser = argparse.ArgumentParser(description="") + parser.add_argument("files", nargs="*", help="input files") + args = parser.parse_args() + + detok = sacremoses.MosesDetokenizer() + + for line in fileinput.input(args.files, openhook=fileinput.hook_compressed): + print( + detok.detokenize(line.strip().split(" ")) + .replace(" @", "") + .replace("@ ", "") + .replace(" =", "=") + .replace("= ", "=") + .replace(" – ", "–") + ) + + +if __name__ == "__main__": + main() diff --git a/examples/multilingual/ML50_langs.txt b/examples/multilingual/ML50_langs.txt new file mode 100644 index 0000000000000000000000000000000000000000..558abbc785072629de8000e343fc02a32c0afb97 --- /dev/null +++ b/examples/multilingual/ML50_langs.txt @@ -0,0 +1,52 @@ +ar_AR +cs_CZ +de_DE +en_XX +es_XX +et_EE +fi_FI +fr_XX +gu_IN +hi_IN +it_IT +ja_XX +kk_KZ +ko_KR +lt_LT +lv_LV +my_MM +ne_NP +nl_XX +ro_RO +ru_RU +si_LK +tr_TR +vi_VN +zh_CN +af_ZA +az_AZ +bn_IN +fa_IR +he_IL +hr_HR +id_ID +ka_GE +km_KH +mk_MK +ml_IN +mn_MN +mr_IN +pl_PL +ps_AF +pt_XX +sv_SE +sw_KE +ta_IN +te_IN +th_TH +tl_XX +uk_UA +ur_PK +xh_ZA +gl_ES +sl_SI \ No newline at end of file diff --git a/examples/multilingual/README.md b/examples/multilingual/README.md new file mode 100644 index 0000000000000000000000000000000000000000..0076f5e8f0ab5c2c8dfd32b3eef02c556dddb88a --- /dev/null +++ b/examples/multilingual/README.md @@ -0,0 +1,158 @@ +# Multilingual Translation + +[[Multilingual Translation with Extensible Multilingual Pretraining and Finetuning, https://arxiv.org/abs/2008.00401]](https://arxiv.org/abs/2008.00401) + +## Introduction + +This work is for training multilingual translation models with multiple bitext datasets. This multilingual translation framework supports (see [[training section]](#Training) and [[finetuning section]](#Finetuning) for examples) + +* temperature based sampling over unbalancing datasets of different translation directions + - --sampling-method' with + choices=['uniform', 'temperature', 'concat'] + - --sampling-temperature +* configurable to automatically add source and/or target language tokens to source/target sentences using data which are prepared in the same way as bilignual training + - --encoder-langtok with choices=['src', 'tgt', None] to specify whether to add source or target language tokens to the source sentences + - --decoder-langtok (binary option) to specify whether to add target language tokens to the target sentences or not +* finetuning mBART pretrained models for multilingual translation + - --finetune-from-model to specify the path from which to load the pretrained model + +## Preprocessing data +Multilingual training requires a joint BPE vocab. Please follow [mBART's preprocessing steps](https://github.com/pytorch/fairseq/tree/master/examples/mbart#bpe-data) to reuse our pretrained sentence-piece model. + +You can also train a joint BPE model on your own dataset and then follow the steps in [[link]](https://github.com/pytorch/fairseq/tree/master/examples/translation#multilingual-translation). + +## Training + + +```bash +lang_pairs= +path_2_data= +lang_list= + +fairseq-train $path_2_data \ + --encoder-normalize-before --decoder-normalize-before \ + --arch transformer --layernorm-embedding \ + --task translation_multi_simple_epoch \ + --sampling-method "temperature" \ + --sampling-temperature 1.5 \ + --encoder-langtok "src" \ + --decoder-langtok \ + --lang-dict "$lang_list" \ + --lang-pairs "$lang_pairs" \ + --criterion label_smoothed_cross_entropy --label-smoothing 0.2 \ + --optimizer adam --adam-eps 1e-06 --adam-betas '(0.9, 0.98)' \ + --lr-scheduler inverse_sqrt --lr 3e-05 --warmup-updates 2500 --max-update 40000 \ + --dropout 0.3 --attention-dropout 0.1 --weight-decay 0.0 \ + --max-tokens 1024 --update-freq 2 \ + --save-interval 1 --save-interval-updates 5000 --keep-interval-updates 10 --no-epoch-checkpoints \ + --seed 222 --log-format simple --log-interval 2 +``` + +## Finetuning +We can also finetune multilingual models from a monolingual pretrained models, e.g. [mMBART](https://github.com/pytorch/fairseq/tree/master/examples/mbart). +```bash +lang_pairs= +path_2_data= +lang_list= +pretrained_model= + +fairseq-train $path_2_data \ + --finetune-from-model $pretrained_model \ + --encoder-normalize-before --decoder-normalize-before \ + --arch transformer --layernorm-embedding \ + --task translation_multi_simple_epoch \ + --sampling-method "temperature" \ + --sampling-temperature 1.5 \ + --encoder-langtok "src" \ + --decoder-langtok \ + --lang-dict "$lang_list" \ + --lang-pairs "$lang_pairs" \ + --criterion label_smoothed_cross_entropy --label-smoothing 0.2 \ + --optimizer adam --adam-eps 1e-06 --adam-betas '(0.9, 0.98)' \ + --lr-scheduler inverse_sqrt --lr 3e-05 --warmup-updates 2500 --max-update 40000 \ + --dropout 0.3 --attention-dropout 0.1 --weight-decay 0.0 \ + --max-tokens 1024 --update-freq 2 \ + --save-interval 1 --save-interval-updates 5000 --keep-interval-updates 10 --no-epoch-checkpoints \ + --seed 222 --log-format simple --log-interval 2 +``` +## Generate +The following command uses the multilingual task (translation_multi_simple_epoch) to generate translation from $source_lang to $target_lang on the test dataset. During generaton, the source language tokens are added to source sentences and the target language tokens are added as the starting token to decode target sentences. Options --lang-dict and --lang-pairs are needed to tell the generation process the ordered list of languages and translation directions that the trained model are awared of; they will need to be consistent with the training. + +```bash +model= +source_lang= +target_lang= + +fairseq-generate $path_2_data \ + --path $model \ + --task translation_multi_simple_epoch \ + --gen-subset test \ + --source-lang $source_lang \ + --target-lang $target_lang + --sacrebleu --remove-bpe 'sentencepiece'\ + --batch-size 32 \ + --encoder-langtok "src" \ + --decoder-langtok \ + --lang-dict "$lang_list" \ + --lang-pairs "$lang_pairs" > ${source_lang}_${target_lang}.txt +``` +Fairseq will generate translation into a file {source_lang}_${target_lang}.txt with sacreblue at the end. + +You can also use costomized tokenizer to compare the performance with the literature. For example, you get a tokenizer [here](https://github.com/rsennrich/wmt16-scripts) and do the following: +```bash +TOKENIZER= +TOK_CMD=<"$TOKENIZER $target_lang" or cat for sacrebleu> + +cat {source_lang}_${target_lang}.txt | grep -P "^H" |sort -V |cut -f 3- |$TOK_CMD > ${source_lang}_${target_lang}.hyp +cat {source_lang}_${target_lang}.txt | grep -P "^T" |sort -V |cut -f 2- |$TOK_CMD > ${source_lang}_${target_lang}.ref +sacrebleu -tok 'none' -s 'none' ${source_lang}_${target_lang}.ref < ${source_lang}_${target_lang}.hyp +``` + +# mBART50 models + +* [mMBART 50 pretrained model](https://dl.fbaipublicfiles.com/fairseq/models/mbart50/mbart50.pretrained.tar.gz). +* [mMBART 50 finetuned many-to-one](https://dl.fbaipublicfiles.com/fairseq/models/mbart50/mbart50.ft.n1.tar.gz). +* [mMBART 50 finetuned one-to-many](https://dl.fbaipublicfiles.com/fairseq/models/mbart50/mbart50.ft.1n.tar.gz). +* [mMBART 50 finetuned many-to-many](https://dl.fbaipublicfiles.com/fairseq/models/mbart50/mbart50.ft.nn.tar.gz). + +Please download and extract from the above tarballs. Each tarball contains +* The fairseq model checkpoint: model.pt +* The list of supported languages: ML50_langs.txt +* Sentence piece model: sentence.bpe.model +* Fairseq dictionary of each language: dict.{lang}.txt (please replace lang with a language specified in ML50_langs.txt) + +To use the trained models, +* use the tool [binarize.py](./data_scripts/binarize.py) to binarize your data using sentence.bpe.model and dict.{lang}.txt, and copy the dictionaries to your data path +* then run the generation command: +```bash +path_2_data= +model=/model.pt +lang_list=/ML50_langs.txt +source_lang= +target_lang= + +fairseq-generate $path_2_data \ + --path $model \ + --task translation_multi_simple_epoch \ + --gen-subset test \ + --source-lang $source_lang \ + --target-lang $target_lang + --sacrebleu --remove-bpe 'sentencepiece'\ + --batch-size 32 \ + --encoder-langtok "src" \ + --decoder-langtok \ + --lang-dict "$lang_list" +``` + +## Citation + +```bibtex +@article{tang2020multilingual, + title={Multilingual Translation with Extensible Multilingual Pretraining and Finetuning}, + author={Yuqing Tang and Chau Tran and Xian Li and Peng-Jen Chen and Naman Goyal and Vishrav Chaudhary and Jiatao Gu and Angela Fan}, + year={2020}, + eprint={2008.00401}, + archivePrefix={arXiv}, + primaryClass={cs.CL} +} +``` diff --git a/examples/multilingual/data_scripts/README.md b/examples/multilingual/data_scripts/README.md new file mode 100644 index 0000000000000000000000000000000000000000..cc610c0c9e936a5ae4659ceda691c6db6d387296 --- /dev/null +++ b/examples/multilingual/data_scripts/README.md @@ -0,0 +1,24 @@ + +# Install dependency +```bash +pip install -r requirement.txt +``` + +# Download the data set +```bash +export WORKDIR_ROOT= + +``` +The downloaded data will be at $WORKDIR_ROOT/ML50 + +# preprocess the data +Install SPM [here](https://github.com/google/sentencepiece) +```bash +export WORKDIR_ROOT= +export SPM_PATH= +``` +* $WORKDIR_ROOT/ML50/raw: extracted raw data +* $WORKDIR_ROOT/ML50/dedup: dedup data +* $WORKDIR_ROOT/ML50/clean: data with valid and test sentences removed from the dedup data + + diff --git a/examples/multilingual/data_scripts/binarize.py b/examples/multilingual/data_scripts/binarize.py new file mode 100755 index 0000000000000000000000000000000000000000..ee54c6aabf021ca526743f8f1f67b91889e1e335 --- /dev/null +++ b/examples/multilingual/data_scripts/binarize.py @@ -0,0 +1,200 @@ +import shutil +import os, sys +from subprocess import check_call, check_output +import glob +import argparse +import shutil +import pathlib +import itertools + +def call_output(cmd): + print(f"Executing: {cmd}") + ret = check_output(cmd, shell=True) + print(ret) + return ret + +def call(cmd): + print(cmd) + check_call(cmd, shell=True) + + +WORKDIR_ROOT = os.environ.get('WORKDIR_ROOT', None) + +if WORKDIR_ROOT is None or not WORKDIR_ROOT.strip(): + print('please specify your working directory root in OS environment variable WORKDIR_ROOT. Exitting..."') + sys.exit(-1) + +SPM_PATH = os.environ.get('SPM_PATH', None) + +if SPM_PATH is None or not SPM_PATH.strip(): + print("Please install sentence piecence from https://github.com/google/sentencepiece and set SPM_PATH pointing to the installed spm_encode.py. Exitting...") + sys.exit(-1) + + +SPM_MODEL = f'{WORKDIR_ROOT}/sentence.bpe.model' +SPM_VOCAB = f'{WORKDIR_ROOT}/dict_250k.txt' + +SPM_ENCODE = f'{SPM_PATH}' + +if not os.path.exists(SPM_MODEL): + call(f"wget https://dl.fbaipublicfiles.com/fairseq/models/mbart50/sentence.bpe.model -O {SPM_MODEL}") + + +if not os.path.exists(SPM_VOCAB): + call(f"wget https://dl.fbaipublicfiles.com/fairseq/models/mbart50/dict_250k.txt -O {SPM_VOCAB}") + + + +def get_data_size(raw): + cmd = f'wc -l {raw}' + ret = call_output(cmd) + return int(ret.split()[0]) + +def encode_spm(model, direction, prefix='', splits=['train', 'test', 'valid'], pairs_per_shard=None): + src, tgt = direction.split('-') + + for split in splits: + src_raw, tgt_raw = f'{RAW_DIR}/{split}{prefix}.{direction}.{src}', f'{RAW_DIR}/{split}{prefix}.{direction}.{tgt}' + if os.path.exists(src_raw) and os.path.exists(tgt_raw): + cmd = f"""python {SPM_ENCODE} \ + --model {model}\ + --output_format=piece \ + --inputs {src_raw} {tgt_raw} \ + --outputs {BPE_DIR}/{direction}{prefix}/{split}.bpe.{src} {BPE_DIR}/{direction}{prefix}/{split}.bpe.{tgt} """ + print(cmd) + call(cmd) + + +def binarize_( + bpe_dir, + databin_dir, + direction, spm_vocab=SPM_VOCAB, + splits=['train', 'test', 'valid'], +): + src, tgt = direction.split('-') + + try: + shutil.rmtree(f'{databin_dir}', ignore_errors=True) + os.mkdir(f'{databin_dir}') + except OSError as error: + print(error) + cmds = [ + "fairseq-preprocess", + f"--source-lang {src} --target-lang {tgt}", + f"--destdir {databin_dir}/", + f"--workers 8", + ] + if isinstance(spm_vocab, tuple): + src_vocab, tgt_vocab = spm_vocab + cmds.extend( + [ + f"--srcdict {src_vocab}", + f"--tgtdict {tgt_vocab}", + ] + ) + else: + cmds.extend( + [ + f"--joined-dictionary", + f"--srcdict {spm_vocab}", + ] + ) + input_options = [] + if 'train' in splits and glob.glob(f"{bpe_dir}/train.bpe*"): + input_options.append( + f"--trainpref {bpe_dir}/train.bpe", + ) + if 'valid' in splits and glob.glob(f"{bpe_dir}/valid.bpe*"): + input_options.append(f"--validpref {bpe_dir}/valid.bpe") + if 'test' in splits and glob.glob(f"{bpe_dir}/test.bpe*"): + input_options.append(f"--testpref {bpe_dir}/test.bpe") + if len(input_options) > 0: + cmd = " ".join(cmds + input_options) + print(cmd) + call(cmd) + + +def binarize( + databin_dir, + direction, spm_vocab=SPM_VOCAB, prefix='', + splits=['train', 'test', 'valid'], + pairs_per_shard=None, +): + def move_databin_files(from_folder, to_folder): + for bin_file in glob.glob(f"{from_folder}/*.bin") \ + + glob.glob(f"{from_folder}/*.idx") \ + + glob.glob(f"{from_folder}/dict*"): + try: + shutil.move(bin_file, to_folder) + except OSError as error: + print(error) + bpe_databin_dir = f"{BPE_DIR}/{direction}{prefix}_databin" + bpe_dir = f"{BPE_DIR}/{direction}{prefix}" + if pairs_per_shard is None: + binarize_(bpe_dir, bpe_databin_dir, direction, spm_vocab=spm_vocab, splits=splits) + move_databin_files(bpe_databin_dir, databin_dir) + else: + # binarize valid and test which will not be sharded + binarize_( + bpe_dir, bpe_databin_dir, direction, + spm_vocab=spm_vocab, splits=[s for s in splits if s != "train"]) + for shard_bpe_dir in glob.glob(f"{bpe_dir}/shard*"): + path_strs = os.path.split(shard_bpe_dir) + shard_str = path_strs[-1] + shard_folder = f"{bpe_databin_dir}/{shard_str}" + databin_shard_folder = f"{databin_dir}/{shard_str}" + print(f'working from {shard_folder} to {databin_shard_folder}') + os.makedirs(databin_shard_folder, exist_ok=True) + binarize_( + shard_bpe_dir, shard_folder, direction, + spm_vocab=spm_vocab, splits=["train"]) + + for test_data in glob.glob(f"{bpe_databin_dir}/valid.*") + glob.glob(f"{bpe_databin_dir}/test.*"): + filename = os.path.split(test_data)[-1] + try: + os.symlink(test_data, f"{databin_shard_folder}/{filename}") + except OSError as error: + print(error) + move_databin_files(shard_folder, databin_shard_folder) + + +def load_langs(path): + with open(path) as fr: + langs = [l.strip() for l in fr] + return langs + +if __name__ == '__main__': + parser = argparse.ArgumentParser() + parser.add_argument("--data_root", default=f"{WORKDIR_ROOT}/ML50") + parser.add_argument("--raw-folder", default='raw') + parser.add_argument("--bpe-folder", default='bpe') + parser.add_argument("--databin-folder", default='databin') + + args = parser.parse_args() + + DATA_PATH = args.data_root #'/private/home/yuqtang/public_data/ML50' + RAW_DIR = f'{DATA_PATH}/{args.raw_folder}' + BPE_DIR = f'{DATA_PATH}/{args.bpe_folder}' + DATABIN_DIR = f'{DATA_PATH}/{args.databin_folder}' + os.makedirs(BPE_DIR, exist_ok=True) + + raw_files = itertools.chain( + glob.glob(f'{RAW_DIR}/train*'), + glob.glob(f'{RAW_DIR}/valid*'), + glob.glob(f'{RAW_DIR}/test*'), + ) + + directions = [os.path.split(file_path)[-1].split('.')[1] for file_path in raw_files] + + for direction in directions: + prefix = "" + splits = ['train', 'valid', 'test'] + try: + shutil.rmtree(f'{BPE_DIR}/{direction}{prefix}', ignore_errors=True) + os.mkdir(f'{BPE_DIR}/{direction}{prefix}') + os.makedirs(DATABIN_DIR, exist_ok=True) + except OSError as error: + print(error) + spm_model, spm_vocab = SPM_MODEL, SPM_VOCAB + encode_spm(spm_model, direction=direction, splits=splits) + binarize(DATABIN_DIR, direction, spm_vocab=spm_vocab, splits=splits) diff --git a/examples/multilingual/data_scripts/check_iswlt_test_data.py b/examples/multilingual/data_scripts/check_iswlt_test_data.py new file mode 100644 index 0000000000000000000000000000000000000000..f8e2eb0f15699f1b458a8445d0c1dd6229a21f77 --- /dev/null +++ b/examples/multilingual/data_scripts/check_iswlt_test_data.py @@ -0,0 +1,67 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +import os, sys +import subprocess +import re +from subprocess import check_call, check_output + +WORKDIR_ROOT = os.environ.get('WORKDIR_ROOT', None) + +if WORKDIR_ROOT is None or not WORKDIR_ROOT.strip(): + print('please specify your working directory root in OS environment variable WORKDIR_ROOT. Exitting..."') + sys.exit(-1) + + +BLEU_REGEX = re.compile("^BLEU\\S* = (\\S+) ") +def run_eval_bleu(cmd): + output = check_output(cmd, shell=True, stderr=subprocess.STDOUT).decode("utf-8").strip() + print(output) + bleu = -1.0 + for line in output.strip().split('\n'): + m = BLEU_REGEX.search(line) + if m is not None: + bleu = m.groups()[0] + bleu = float(bleu) + break + return bleu + +def check_data_test_bleu(raw_folder, data_lang_pairs): + not_matchings = [] + for sacrebleu_set, src_tgts in data_lang_pairs: + for src_tgt in src_tgts: + print(f'checking test bleus for: {src_tgt} at {sacrebleu_set}') + src, tgt = src_tgt.split('-') + ssrc, stgt = src[:2], tgt[:2] + if os.path.exists(f'{raw_folder}/test.{tgt}-{src}.{src}'): + # reversed direction may have different test set + test_src = f'{raw_folder}/test.{tgt}-{src}.{src}' + else: + test_src = f'{raw_folder}/test.{src}-{tgt}.{src}' + cmd1 = f'cat {test_src} | sacrebleu -t "{sacrebleu_set}" -l {stgt}-{ssrc}; [ $? -eq 0 ] || echo ""' + test_tgt = f'{raw_folder}/test.{src}-{tgt}.{tgt}' + cmd2 = f'cat {test_tgt} | sacrebleu -t "{sacrebleu_set}" -l {ssrc}-{stgt}; [ $? -eq 0 ] || echo ""' + bleu1 = run_eval_bleu(cmd1) + if bleu1 != 100.0: + not_matchings.append(f'{sacrebleu_set}:{src_tgt} source side not matching: {test_src}') + bleu2 = run_eval_bleu(cmd2) + if bleu2 != 100.0: + not_matchings.append(f'{sacrebleu_set}:{src_tgt} target side not matching: {test_tgt}') + return not_matchings + +if __name__ == "__main__": + to_data_path = f'{WORKDIR_ROOT}/iwsltv2' + not_matching = check_data_test_bleu( + f'{to_data_path}/raw', + [ + ('iwslt17', ['en_XX-ar_AR', 'en_XX-ko_KR', 'ar_AR-en_XX', 'ko_KR-en_XX']), + ('iwslt17', ['en_XX-it_IT', 'en_XX-nl_XX', 'it_IT-en_XX', 'nl_XX-en_XX']), + ('iwslt17/tst2015', ['en_XX-vi_VN', "vi_VN-en_XX"]), + ] + ) + if len(not_matching) > 0: + print('the following datasets do not have matching test datasets:\n\t', '\n\t'.join(not_matching)) + diff --git a/examples/multilingual/data_scripts/check_self_overlaps.py b/examples/multilingual/data_scripts/check_self_overlaps.py new file mode 100644 index 0000000000000000000000000000000000000000..07b338dcfd2d7f10317608274631d0edd93ba889 --- /dev/null +++ b/examples/multilingual/data_scripts/check_self_overlaps.py @@ -0,0 +1,103 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +import os +import glob +import argparse +from utils.dedup import deup +import sys + +WORKDIR_ROOT = os.environ.get('WORKDIR_ROOT', None) + +if WORKDIR_ROOT is None or not WORKDIR_ROOT.strip(): + print('please specify your working directory root in OS environment variable WORKDIR_ROOT. Exitting..."') + sys.exit(-1) + +def get_directions(folder): + raw_files = glob.glob(f'{folder}/train*') + directions = [os.path.split(file_path)[-1].split('.')[1] for file_path in raw_files] + return directions + +def diff_list(lhs, rhs): + return set(lhs).difference(set(rhs)) + +def check_diff( + from_src_file, from_tgt_file, + to_src_file, to_tgt_file, +): + seen_in_from = set() + seen_src_in_from = set() + seen_tgt_in_from = set() + from_count = 0 + with open(from_src_file, encoding='utf-8') as fsrc, \ + open(from_tgt_file, encoding='utf-8') as ftgt: + for s, t in zip(fsrc, ftgt): + seen_in_from.add((s, t)) + seen_src_in_from.add(s) + seen_tgt_in_from.add(t) + from_count += 1 + common = 0 + common_src = 0 + common_tgt = 0 + to_count = 0 + seen = set() + + with open(to_src_file, encoding='utf-8') as fsrc, \ + open(to_tgt_file, encoding='utf-8') as ftgt: + for s, t in zip(fsrc, ftgt): + to_count += 1 + if (s, t) not in seen: + if (s, t) in seen_in_from: + common += 1 + if s in seen_src_in_from: + common_src += 1 + seen_src_in_from.remove(s) + if t in seen_tgt_in_from: + common_tgt += 1 + seen_tgt_in_from.remove(t) + seen.add((s, t)) + return common, common_src, common_tgt, from_count, to_count + +def main(): + parser = argparse.ArgumentParser() + parser.add_argument("--folder", type=str, required=True, + help="the data folder ") + parser.add_argument("--split", type=str, default='test', + help="split (valid, test) to check against training data") + parser.add_argument('--directions', type=str, default=None, required=False) + + args = parser.parse_args() + + if args.directions is None: + directions = set(get_directions(args.folder)) + directions = sorted(directions) + else: + directions = args.directions.split(',') + directions = sorted(set(directions)) + + results = [] + print(f'checking where {args.split} split data are in training') + print(f'direction\tcommon_count\tsrc common\ttgt common\tfrom_size\tto_size') + + for direction in directions: + src, tgt = direction.split('-') + from_src_file = f'{args.folder}/{args.split}.{src}-{tgt}.{src}' + from_tgt_file = f'{args.folder}/{args.split}.{src}-{tgt}.{tgt}' + if not os.path.exists(from_src_file): + # some test/valid data might in reverse directinos: + from_src_file = f'{args.folder}/{args.split}.{tgt}-{src}.{src}' + from_tgt_file = f'{args.folder}/{args.split}.{tgt}-{src}.{tgt}' + to_src_file = f'{args.folder}/train.{src}-{tgt}.{src}' + to_tgt_file = f'{args.folder}/train.{src}-{tgt}.{tgt}' + if not os.path.exists(to_src_file) or not os.path.exists(from_src_file): + continue + r = check_diff(from_src_file, from_tgt_file, to_src_file, to_tgt_file) + results.append(r) + print(f'{direction}\t', '\t'.join(map(str, r))) + + +if __name__ == "__main__": + main() diff --git a/examples/multilingual/data_scripts/check_valid_test_overlaps.py b/examples/multilingual/data_scripts/check_valid_test_overlaps.py new file mode 100644 index 0000000000000000000000000000000000000000..40fa9aecdf9108e095feb3661236453c0f7ed7c4 --- /dev/null +++ b/examples/multilingual/data_scripts/check_valid_test_overlaps.py @@ -0,0 +1,124 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +import os +import argparse +import pandas as pd +import sys + + +WORKDIR_ROOT = os.environ.get('WORKDIR_ROOT', None) + +if WORKDIR_ROOT is None or not WORKDIR_ROOT.strip(): + print('please specify your working directory root in OS environment variable WORKDIR_ROOT. Exitting..."') + sys.exit(-1) + +def load_langs(path): + with open(path) as fr: + langs = [l.strip() for l in fr] + return langs + + + +def load_sentences(raw_data, split, direction): + src, tgt = direction.split('-') + src_path = f"{raw_data}/{split}.{direction}.{src}" + tgt_path = f"{raw_data}/{split}.{direction}.{tgt}" + if os.path.exists(src_path) and os.path.exists(tgt_path): + return [(src, open(src_path).read().splitlines()), (tgt, open(tgt_path).read().splitlines())] + else: + return [] + +def swap_direction(d): + src, tgt = d.split('-') + return f'{tgt}-{src}' + +def get_all_test_data(raw_data, directions, split='test'): + test_data = [ + x + for dd in directions + for d in [dd, swap_direction(dd)] + for x in load_sentences(raw_data, split, d) + ] + # all_test_data = {s for _, d in test_data for s in d} + all_test_data = {} + for lang, d in test_data: + for s in d: + s = s.strip() + lgs = all_test_data.get(s, set()) + lgs.add(lang) + all_test_data[s] = lgs + return all_test_data, test_data + + +def check_train_sentences(src_path, tgt_path, direction, all_test_data, mess_up_train={}): + # src, tgt = direction.split('-') + print(f'check training data for {direction} in {src_path} and {tgt_path}') + size = 0 + overlapped_size_counted_dup = 0 + if not os.path.exists(tgt_path) or not os.path.exists(src_path): + return mess_up_train, size, overlapped_size_counted_dup + + with open(src_path) as f, open(tgt_path) as g: + for src_line, tgt_line in zip(f, g): + s = src_line.strip() + t = tgt_line.strip() + size += 1 + if s in all_test_data: + langs = mess_up_train.get(s, set()) + langs.add(direction) + mess_up_train[s] = langs + overlapped_size_counted_dup += 1 + if t in all_test_data: + langs = mess_up_train.get(t, set()) + langs.add(direction) + mess_up_train[t] = langs + overlapped_size_counted_dup += 1 + print(f'{direction}: size={size}, overlapped={overlapped_size_counted_dup}') + return mess_up_train, size, overlapped_size_counted_dup + +def check_train_all(raw_data, directions, all_test_data): + mess_up_train = {} + data_sizes = {} + # raw_data = '~chau/data-bin/MineBART/multilingual_mined_100M/en_XX/et_EE-en_XX/all.{en_XX, et_EE}' + print(f'checking training data againsts # {len(all_test_data)} sentences') + print(f'example test data: ', [s for i, s in enumerate(all_test_data.keys()) if i < 10]) + for direction in directions: + src, tgt = direction.split('-') + path = f'{raw_data}/en_XX/{direction}/all' + src_path = f'{path}.{src}' + tgt_path = f'{path}.{tgt}' + print(f'checking {src_path} {tgt_path}') + _, size, overlapped_size_counted_dup = check_train_sentences(src_path, tgt_path, direction, all_test_data, mess_up_train) + data_sizes[direction] = (size, overlapped_size_counted_dup) + return mess_up_train, data_sizes + + + + +def main(): + parser = argparse.ArgumentParser() + parser.add_argument("--folder", type=str, required=True, + help="the data folder ") + parser.add_argument("--test-data", type=str, required=True, + help="the test data folder ") + parser.add_argument('--directions', type=str, default=None, required=False) + + args = parser.parse_args() + directions = args.directions.split(',') + directions = sorted(set(directions)) + + results = [] + # print(f'checking where {args.split} split data are in training') + # print(f'direction\tcommon_count\tsrc common\ttgt common\tfrom_size\tto_size') + raw_data = args.folder + all_test_data, test_data = get_all_test_data(args.test_data, directions, split='test') + mess_up_train, data_sizes = check_train_all(raw_data, directions, all_test_data) + print(data_sizes) + + +if __name__ == "__main__": + main() diff --git a/examples/multilingual/data_scripts/dedup_all.py b/examples/multilingual/data_scripts/dedup_all.py new file mode 100644 index 0000000000000000000000000000000000000000..ef39c05ee606aaeda1d9e94970932d2241a8b281 --- /dev/null +++ b/examples/multilingual/data_scripts/dedup_all.py @@ -0,0 +1,52 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + + +import os +import glob +import argparse +from utils.dedup import deup + +import sys +WORKDIR_ROOT = os.environ.get('WORKDIR_ROOT', None) + +if WORKDIR_ROOT is None or not WORKDIR_ROOT.strip(): + print('please specify your working directory root in OS environment variable WORKDIR_ROOT. Exitting..."') + sys.exit(-1) + + +def main(): + parser = argparse.ArgumentParser() + parser.add_argument("--from-folder", type=str, required=True, + help="the data folder to be dedup") + parser.add_argument("--to-folder", type=str, required=True, + help="the data folder to save deduped data") + parser.add_argument('--directions', type=str, default=None, required=False) + + args = parser.parse_args() + + if args.directions is None: + raw_files = glob.glob(f'{args.from_folder}/train*') + + directions = [os.path.split(file_path)[-1].split('.')[1] for file_path in raw_files] + else: + directions = args.directions.split(',') + directions = sorted(set(directions)) + + for direction in directions: + src, tgt = direction.split('-') + src_file = f'{args.from_folder}/train.{src}-{tgt}.{src}' + tgt_file = f'{args.from_folder}/train.{src}-{tgt}.{tgt}' + src_file_out = f'{args.to_folder}/train.{src}-{tgt}.{src}' + tgt_file_out = f'{args.to_folder}/train.{src}-{tgt}.{tgt}' + assert src_file != src_file_out + assert tgt_file != tgt_file_out + print(f'deduping {src_file}, {tgt_file}') + deup(src_file, tgt_file, src_file_out, tgt_file_out) + + +if __name__ == "__main__": + main() diff --git a/examples/multilingual/data_scripts/download_ML50_v1.sh b/examples/multilingual/data_scripts/download_ML50_v1.sh new file mode 100644 index 0000000000000000000000000000000000000000..99fbc75920836a4b4bbdbd6b523749843288e450 --- /dev/null +++ b/examples/multilingual/data_scripts/download_ML50_v1.sh @@ -0,0 +1,30 @@ +#!/bin/bash +# Copyright (c) Facebook, Inc. and its affiliates. +# All rights reserved. +# +# This source code is licensed under the license found in the +# LICENSE file in the root directory of this source tree. + +if [ -z $WORKDIR_ROOT ] ; +then + echo "please specify your working directory root in environment variable WORKDIR_ROOT. Exitting..." + exit +fi + +# first run download_wmt20.sh; it will install a few useful tools for other scripts +# TODO: need to print out instructions on downloading a few files which requires manually authentication from the websites +bash ./download_wmt20.sh + +python ./download_wmt19_and_before.py +bash ./download_wat19_my.sh +python ./download_ted_and_extract.py +bash ./download_lotus.sh +bash ./download_iitb.sh +bash ./download_af_xh.sh + + +# IWSLT downloading URLs have changed in between; TODO: fix them: +bash ./download_iwslt_and_extract.sh + +# TODO: globalvoices URLs changed; need to be fixed +bash ./download_flores_data.sh diff --git a/examples/multilingual/data_scripts/download_af_xh.sh b/examples/multilingual/data_scripts/download_af_xh.sh new file mode 100644 index 0000000000000000000000000000000000000000..a78fbbbbccb6f6ae005a1f03b97f083a2d958ebe --- /dev/null +++ b/examples/multilingual/data_scripts/download_af_xh.sh @@ -0,0 +1,164 @@ +#!/bin/bash +# Copyright (c) Facebook, Inc. and its affiliates. +# All rights reserved. +# +# This source code is licensed under the license found in the +# LICENSE file in the root directory of this source tree. + +# set -x -e + +if [ -z $WORKDIR_ROOT ] ; +then + echo "please specify your working directory root in environment variable WORKDIR_ROOT. Exitting..." + exit +fi + + +# put intermediate files +TMP_DIR=$WORKDIR_ROOT/temp/af_xhv2 +# output {train,valid,test} files to dest +DEST=${WORKDIR_ROOT}/ML50/raw + + + +ROOT=${WORKDIR_ROOT} +UTILS=$PWD/utils +TMX2CORPUS="${UTILS}/tmx2corpus" +TMX_TOOL="python ${TMX2CORPUS}/tmx2corpus.py" + +mkdir -p $TMP_DIR +mkdir -p $DEST +mkdir -p $UTILS + +function download_opus(){ + src=$1 + tgt=$2 + subset=$3 + ulr=$4 + + mkdir extract_$subset.$src-$tgt + pushd extract_$subset.$src-$tgt + if [ ! -f "$subset.$src-$tgt.tmx.gz" ]; then + wget $url -O "$subset.$src-$tgt.tmx.gz" + gzip -d "$subset.$src-$tgt.tmx.gz" + f=$subset.$src-$tgt.tmx + $TMX_TOOL $f + mv bitext.$src ../$subset.$src-$tgt.$src + mv bitext.$tgt ../$subset.$src-$tgt.$tgt + fi + popd +} + +function concat_subsets(){ + src=$1 + tgt=$2 + subsets=$3 + src_train=raw_train.$src-$tgt.$src + tgt_train=raw_train.$src-$tgt.$tgt + > $src_train + > $tgt_train + for subset in $subsets; do + cat $subset.$src-$tgt.$src >> $src_train + cat $subset.$src-$tgt.$tgt >> $tgt_train + done +} + + + +function get_seeded_random() +{ + seed="$1" + openssl enc -aes-256-ctr -pass pass:"$seed" -nosalt \ + /dev/null +} + +function split_train_valid(){ + src=$1 + tgt=$2 + raw_src_train=raw_train.$src-$tgt.$src + raw_tgt_train=raw_train.$src-$tgt.$tgt + + shuf --random-source=<(get_seeded_random 43) $raw_src_train > shuffled.$src-$tgt.$src + shuf --random-source=<(get_seeded_random 43) $raw_tgt_train > shuffled.$src-$tgt.$tgt + + head -n 1500 shuffled.$src-$tgt.$src > valid.$src-$tgt.$src + head -n 1500 shuffled.$src-$tgt.$tgt > valid.$src-$tgt.$tgt + + tail +1501 shuffled.$src-$tgt.$src > train.$src-$tgt.$src + tail +1501 shuffled.$src-$tgt.$tgt > train.$src-$tgt.$tgt +} + +function copy2dst(){ + lsrc=$1 + ltgt=$2 + src=${lsrc:0:2} + tgt=${ltgt:0:2} + + + cp valid.$src-$tgt.$src $DEST/valid.$lsrc-$ltgt.$lsrc + cp valid.$src-$tgt.$tgt $DEST/valid.$lsrc-$ltgt.$ltgt + + cp train.$src-$tgt.$src $DEST/train.$lsrc-$ltgt.$lsrc + cp train.$src-$tgt.$tgt $DEST/train.$lsrc-$ltgt.$ltgt +} + + + + +#for xh-en +declare -A xh_en_urls +xh_en_urls=( + [Tatoeba]=https://object.pouta.csc.fi/OPUS-Tatoeba/v20190709/tmx/en-xh.tmx.gz + [wikimedia]=https://object.pouta.csc.fi/OPUS-wikimedia/v20190628/tmx/en-xh.tmx.gz + [memat]=https://object.pouta.csc.fi/OPUS-memat/v1/tmx/en-xh.tmx.gz + [uedin]=https://object.pouta.csc.fi/OPUS-bible-uedin/v1/tmx/en-xh.tmx.gz + [GNOME]=https://object.pouta.csc.fi/OPUS-GNOME/v1/tmx/en-xh.tmx.gz + [XhosaNavy]=https://object.pouta.csc.fi/OPUS-XhosaNavy/v1/tmx/en-xh.tmx.gz + [KDE4]=https://object.pouta.csc.fi/OPUS-KDE4/v2/tmx/en-xh.tmx.gz + [Ubuntu]=https://object.pouta.csc.fi/OPUS-Ubuntu/v14.10/tmx/en-xh.tmx.gz +) + +mkdir $TMP_DIR/xh-en +pushd $TMP_DIR/xh-en +for k in "${!xh_en_urls[@]}" +do + name=$k + url=${xh_en_urls[$k]} + echo "$name: $url" + download_opus xh en $name $ulr +done +concat_subsets xh en "${!xh_en_urls[@]}" +split_train_valid xh en +copy2dst xh_ZA en_XX +popd + + +## +#for af-en +declare -A af_en_urls +af_en_urls=( + [Tatoeba]=https://object.pouta.csc.fi/OPUS-Tatoeba/v20190709/tmx/af-en.tmx.gz + [uedin]=https://object.pouta.csc.fi/OPUS-bible-uedin/v1/tmx/af-en.tmx.gz + [GNOME]=https://object.pouta.csc.fi/OPUS-GNOME/v1/tmx/af-en.tmx.gz + [QED]=https://object.pouta.csc.fi/OPUS-QED/v2.0a/tmx/af-en.tmx.gz + [KDE4]=https://object.pouta.csc.fi/OPUS-KDE4/v2/tmx/af-en.tmx.gz + [OpenSubtitles]=https://object.pouta.csc.fi/OPUS-OpenSubtitles/v2018/tmx/af-en.tmx.gz + [SPC]=https://object.pouta.csc.fi/OPUS-SPC/v1/tmx/af-en.tmx.gz + [Ubuntu]=https://object.pouta.csc.fi/OPUS-Ubuntu/v14.10/tmx/af-en.tmx.gz +) + +mkdir $TMP_DIR/af-en +pushd $TMP_DIR/af-en +for k in "${!af_en_urls[@]}" +do + name=$k + url=${af_en_urls[$k]} + echo "$name: $url" + download_opus af en $name $ulr +done +concat_subsets af en "${!af_en_urls[@]}" +split_train_valid af en +copy2dst af_ZA en_XX +popd + + diff --git a/examples/multilingual/data_scripts/download_flores_data.sh b/examples/multilingual/data_scripts/download_flores_data.sh new file mode 100644 index 0000000000000000000000000000000000000000..e6175ce0c38b06a1ebddaeca808f71b47f77f500 --- /dev/null +++ b/examples/multilingual/data_scripts/download_flores_data.sh @@ -0,0 +1,246 @@ +#!/bin/bash + +# Copyright (c) Facebook, Inc. and its affiliates. +# All rights reserved. +# +# This source code is licensed under the license found in the +# LICENSE file in the root directory of this source tree. +# + +if [ -z $WORKDIR_ROOT ] ; +then + echo "please specify your working directory root in environment variable WORKDIR_ROOT. Exitting..." + exit +fi + + +set -e +set -o pipefail + +SRC=en +SI_TGT=si +NE_TGT=ne + +DESTDIR=${WORKDIR_ROOT}/ML50/raw/ + +ROOT=${WORKDIR_ROOT}/tmp +mkdir -p $ROOT +DATA=$ROOT/data +NE_ROOT=$DATA/all-clean-ne +SI_ROOT=$DATA/all-clean-si + +mkdir -p $DATA $NE_ROOT $SI_ROOT + +SI_OPUS_DATASETS=( + "$SI_ROOT/GNOME.en-si" + "$SI_ROOT/Ubuntu.en-si" + "$SI_ROOT/KDE4.en-si" + "$SI_ROOT/OpenSubtitles.en-si" +) + +SI_OPUS_URLS=( + "https://object.pouta.csc.fi/OPUS-GNOME/v1/moses/en-si.txt.zip" + "https://object.pouta.csc.fi/OPUS-Ubuntu/v14.10/moses/en-si.txt.zip" + "https://object.pouta.csc.fi/OPUS-KDE4/v2/moses/en-si.txt.zip" + "https://object.pouta.csc.fi/OPUS-OpenSubtitles/v2018/moses/en-si.txt.zip" +) + +NE_OPUS_DATASETS=( + "$NE_ROOT/GNOME.en-ne" + "$NE_ROOT/Ubuntu.en-ne" + "$NE_ROOT/KDE4.en-ne" +) + +NE_OPUS_URLS=( + "https://object.pouta.csc.fi/OPUS-GNOME/v1/moses/en-ne.txt.zip" + "https://object.pouta.csc.fi/OPUS-Ubuntu/v14.10/moses/en-ne.txt.zip" + "https://object.pouta.csc.fi/OPUS-KDE4/v2/moses/en-ne.txt.zip" +) + +REMOVE_FILE_PATHS=() + +# Download data +download_data() { + CORPORA=$1 + URL=$2 + + if [ -f $CORPORA ]; then + echo "$CORPORA already exists, skipping download" + else + echo "Downloading $URL" + wget $URL -O $CORPORA --no-check-certificate || rm -f $CORPORA + if [ -f $CORPORA ]; then + echo "$URL successfully downloaded." + else + echo "$URL not successfully downloaded." + rm -f $CORPORA + exit -1 + fi + fi +} + +# Example: download_opus_data $LANG_ROOT $TGT +download_opus_data() { + LANG_ROOT=$1 + TGT=$2 + + if [ "$TGT" = "si" ]; then + URLS=("${SI_OPUS_URLS[@]}") + DATASETS=("${SI_OPUS_DATASETS[@]}") + else + URLS=("${NE_OPUS_URLS[@]}") + DATASETS=("${NE_OPUS_DATASETS[@]}") + fi + + # Download and extract data + for ((i=0;i<${#URLS[@]};++i)); do + URL=${URLS[i]} + CORPORA=${DATASETS[i]} + + download_data $CORPORA $URL + unzip -o $CORPORA -d $LANG_ROOT + REMOVE_FILE_PATHS+=( $CORPORA $CORPORA.xml $CORPORA.ids $LANG_ROOT/README $LANG_ROOT/LICENSE ) + done + + cat ${DATASETS[0]}.$SRC ${DATASETS[1]}.$SRC ${DATASETS[2]}.$SRC > $LANG_ROOT/GNOMEKDEUbuntu.$SRC-$TGT.$SRC + cat ${DATASETS[0]}.$TGT ${DATASETS[1]}.$TGT ${DATASETS[2]}.$TGT > $LANG_ROOT/GNOMEKDEUbuntu.$SRC-$TGT.$TGT + + REMOVE_FILE_PATHS+=( ${DATASETS[0]}.$SRC ${DATASETS[1]}.$SRC ${DATASETS[2]}.$SRC ) + REMOVE_FILE_PATHS+=( ${DATASETS[0]}.$TGT ${DATASETS[1]}.$TGT ${DATASETS[2]}.$TGT ) +} + +download_opus_data $SI_ROOT $SI_TGT +cp ${SI_OPUS_DATASETS[3]}.$SRC $SI_ROOT/OpenSubtitles2018.$SRC-$SI_TGT.$SRC +cp ${SI_OPUS_DATASETS[3]}.$SI_TGT $SI_ROOT/OpenSubtitles2018.$SRC-$SI_TGT.$SI_TGT +REMOVE_FILE_PATHS+=( ${SI_OPUS_DATASETS[3]}.$SRC ${SI_OPUS_DATASETS[3]}.$SI_TGT ) + +download_opus_data $NE_ROOT $NE_TGT + + +# Download and extract Global Voices data +GLOBAL_VOICES="$NE_ROOT/globalvoices.2018q4.ne-en" +GLOBAL_VOICES_URL="http://www.casmacat.eu/corpus/global-voices/globalvoices.ne-en.xliff.gz" + +download_data $GLOBAL_VOICES.gz $GLOBAL_VOICES_URL +gunzip -Nf $GLOBAL_VOICES.gz + +sed -ne 's?.*\(.*\).*?\1?p' $GLOBAL_VOICES > $GLOBAL_VOICES.$NE_TGT +sed -ne 's?.*]*>\(.*\).*?\1?p' $GLOBAL_VOICES > $GLOBAL_VOICES.$SRC + +REMOVE_FILE_PATHS+=( $GLOBAL_VOICES ) + +# Download and extract the bible dataset +BIBLE_TOOLS=bible-corpus-tools +XML_BIBLES=XML_Bibles +XML_BIBLES_DUP=XML_Bibles_dup + +if [ ! -e $BIBLE_TOOLS ]; then + echo "Cloning bible-corpus-tools repository..." + git clone https://github.com/christos-c/bible-corpus-tools.git +fi + +mkdir -p $BIBLE_TOOLS/bin $XML_BIBLES $XML_BIBLES_DUP +javac -cp "$BIBLE_TOOLS/lib/*" -d $BIBLE_TOOLS/bin $BIBLE_TOOLS/src/bible/readers/*.java $BIBLE_TOOLS/src/bible/*.java + +download_data bible.tar.gz "https://github.com/christos-c/bible-corpus/archive/v1.2.1.tar.gz" +tar xvzf bible.tar.gz + +cp bible-corpus-1.2.1/bibles/{Greek.xml,English.xml,Nepali.xml} $XML_BIBLES/ +cp bible-corpus-1.2.1/bibles/{Greek.xml,English-WEB.xml,Nepali.xml} $XML_BIBLES_DUP/ + +java -cp $BIBLE_TOOLS/lib/*:$BIBLE_TOOLS/bin bible.CreateMLBooks $XML_BIBLES +java -cp $BIBLE_TOOLS/lib/*:$BIBLE_TOOLS/bin bible.CreateMLBooks $XML_BIBLES_DUP +java -cp $BIBLE_TOOLS/lib/*:$BIBLE_TOOLS/bin bible.CreateVerseAlignedBooks $XML_BIBLES +java -cp $BIBLE_TOOLS/lib/*:$BIBLE_TOOLS/bin bible.CreateVerseAlignedBooks $XML_BIBLES_DUP + +cat $XML_BIBLES/aligned/*/English.txt > $NE_ROOT/bible.$SRC-$NE_TGT.$SRC +cat $XML_BIBLES/aligned/*/Nepali.txt > $NE_ROOT/bible.$SRC-$NE_TGT.$NE_TGT +cat $XML_BIBLES_DUP/aligned/*/English-WEB.txt > $NE_ROOT/bible_dup.$SRC-$NE_TGT.$SRC +cat $XML_BIBLES_DUP/aligned/*/Nepali.txt > $NE_ROOT/bible_dup.$SRC-$NE_TGT.$NE_TGT +REMOVE_FILE_PATHS+=( bible-corpus-1.2.1 bible.tar.gz $BIBLE_TOOLS $XML_BIBLES $XML_BIBLES_DUP ) + +# Download and extract the Penn Treebank dataset +NE_TAGGED=$ROOT/new_submissions_parallel_corpus_project_Nepal +NE_TAGGED_URL="http://www.cle.org.pk/Downloads/ling_resources/parallelcorpus/NepaliTaggedCorpus.zip" +EN_TAGGED_PATCH_URL="https://dl.fbaipublicfiles.com/fairseq/data/nepali-penn-treebank.en.patch" +NE_TAGGED_PATCH_URL="https://dl.fbaipublicfiles.com/fairseq/data/nepali-penn-treebank.ne.patch" +MOSES=mosesdecoder +MOSES_TOK=$MOSES/scripts/tokenizer +EN_PATCH_REGEX="{s:\\\/:\/:g;s/\*\T\*\-\n+//g;s/\-LCB\-/\{/g;s/\-RCB\-/\}/g; s/\-LSB\-/\[/g; s/\-RSB\-/\]/g;s/\-LRB\-/\(/g; s/\-RRB\-/\)/g; s/\'\'/\"/g; s/\`\`/\"/g; s/\ +\'s\ +/\'s /g; s/\ +\'re\ +/\'re /g; s/\"\ +/\"/g; s/\ +\"/\"/g; s/\ n't([\ \.\"])/n't\1/g; s/\r+(.)/\1/g;}" +NE_PATCH_REGEX="{s:\p{Cf}::g;s:\\\/:\/:g;s/\*\T\*\-\n+//g;s/\-LCB\-/\{/g;s/\-RCB\-/\}/g; s/\-LSB\-/\[/g; s/\-RSB\-/\]/g;s/\-LRB\-/\(/g; s/\-RRB\-/\)/g; s/\'\'/\"/g; s/\`\`/\"/g; s/\ +\'s\ +/\'s /g; s/\ +\'re\ +/\'re /g; s/\"\ +/\"/g; s/\ +\"/\"/g; s/\ n't([\ \.\"])/n't\1/g; s/\r+(.)/\1/g;}" + +download_data $DATA/nepali-penn-treebank.$SRC.patch $EN_TAGGED_PATCH_URL +download_data $DATA/nepali-penn-treebank.$NE_TGT.patch $NE_TAGGED_PATCH_URL +download_data original.zip $NE_TAGGED_URL +unzip -o original.zip -d $ROOT + +cat $NE_TAGGED/00.txt $NE_TAGGED/01.txt $NE_TAGGED/02.txt > $NE_TAGGED/nepali-penn-treebank.$SRC +cat $NE_TAGGED/00ne_revised.txt $NE_TAGGED/01ne_revised.txt $NE_TAGGED/02ne_revised.txt > $NE_TAGGED/nepali-penn-treebank.$NE_TGT + +patch $NE_TAGGED/nepali-penn-treebank.$SRC -i $DATA/nepali-penn-treebank.$SRC.patch -o $NE_TAGGED/nepali-penn-treebank-patched.$SRC +patch $NE_TAGGED/nepali-penn-treebank.$NE_TGT -i $DATA/nepali-penn-treebank.$NE_TGT.patch -o $NE_TAGGED/nepali-penn-treebank-patched.$NE_TGT + +if [ ! -e $MOSES ]; then + echo "Cloning moses repository..." + git clone https://github.com/moses-smt/mosesdecoder.git +fi + +cat $NE_TAGGED/nepali-penn-treebank-patched.$SRC | \ + perl -anpe "$EN_PATCH_REGEX" | \ + $MOSES_TOK/tokenizer.perl -l $SRC | \ + $MOSES_TOK/detokenizer.perl -l $SRC > $NE_ROOT/nepali-penn-treebank.$SRC + +cat $NE_TAGGED/nepali-penn-treebank-patched.$NE_TGT | \ + perl -CIO -anpe "$NE_PATCH_REGEX" | \ + $MOSES_TOK/detokenizer.perl -l $SRC > $NE_ROOT/nepali-penn-treebank.$NE_TGT + + +# Download nepali dictionary data +NE_DICT=$NE_ROOT/dictionaries +download_data $NE_DICT "http://www.seas.upenn.edu/~nlp/resources/TACL-data-release/dictionaries.tar.gz" +tar xvzf $NE_DICT +cp dictionaries/dict.ne $NE_ROOT/dictionary.$NE_TGT-$SRC +REMOVE_FILE_PATHS+=( $NE_DICT dictionaries ) + +REMOVE_FILE_PATHS+=( $MOSES $NE_TAGGED original.zip $DATA/nepali-penn-treebank.$SRC.patch $DATA/nepali-penn-treebank.$NE_TGT.patch ) + + +# Remove the temporary files +for ((i=0;i<${#REMOVE_FILE_PATHS[@]};++i)); do + rm -rf ${REMOVE_FILE_PATHS[i]} +done + +# Copy the training data +si=si_LK +ne=ne_NP +en=en_XX +cat $SI_ROOT/GNOMEKDEUbuntu.en-si.si $SI_ROOT/OpenSubtitles2018.en-si.si > $DESTDIR/train.$si-$en.$si +cat $SI_ROOT/GNOMEKDEUbuntu.en-si.en $SI_ROOT/OpenSubtitles2018.en-si.en > $DESTDIR/train.$si-$en.$en + +cat $NE_ROOT/bible_dup.en-ne.ne $NE_ROOT/bible.en-ne.ne $NE_ROOT/globalvoices.2018q4.ne-en.ne $NE_ROOT/GNOMEKDEUbuntu.en-ne.ne $NE_ROOT/nepali-penn-treebank.ne > $DESTDIR/train.$ne-$en.$ne +cat $NE_ROOT/bible_dup.en-ne.en $NE_ROOT/bible.en-ne.en $NE_ROOT/globalvoices.2018q4.ne-en.en $NE_ROOT/GNOMEKDEUbuntu.en-ne.en $NE_ROOT/nepali-penn-treebank.en > $DESTDIR/train.$ne-$en.$en + + +#Download the test sets +wget https://github.com/facebookresearch/flores/raw/master/data/wikipedia_en_ne_si_test_sets.tgz +tar -xvzf wikipedia_en_ne_si_test_sets.tgz + +cp wikipedia_en_ne_si_test_sets/wikipedia.dev.ne-en.ne $DESTDIR/valid.$ne-$en.$ne +cp wikipedia_en_ne_si_test_sets/wikipedia.dev.ne-en.en $DESTDIR/valid.$ne-$en.$en + +cp wikipedia_en_ne_si_test_sets/wikipedia.dev.si-en.si $DESTDIR/valid.$si-$en.$si +cp wikipedia_en_ne_si_test_sets/wikipedia.dev.si-en.en $DESTDIR/valid.$si-$en.$en + +cp wikipedia_en_ne_si_test_sets/wikipedia.devtest.ne-en.ne $DESTDIR/devtest.$ne-$en.$ne +cp wikipedia_en_ne_si_test_sets/wikipedia.devtest.ne-en.en $DESTDIR/devtest.$ne-$en.$en + +cp wikipedia_en_ne_si_test_sets/wikipedia.devtest.si-en.si $DESTDIR/devtest.$si-$en.$si +cp wikipedia_en_ne_si_test_sets/wikipedia.devtest.si-en.en $DESTDIR/devtest.$si-$en.$en + +cp wikipedia_en_ne_si_test_sets/wikipedia.test.ne-en.ne $DESTDIR/test.$ne-$en.$ne +cp wikipedia_en_ne_si_test_sets/wikipedia.test.ne-en.en $DESTDIR/test.$ne-$en.$en + +cp wikipedia_en_ne_si_test_sets/wikipedia.test.si-en.si $DESTDIR/test.$si-$en.$si +cp wikipedia_en_ne_si_test_sets/wikipedia.test.si-en.en $DESTDIR/test.$si-$en.$en + +rm -rf wikipedia_en_ne_si_test_sets.tgz wikipedia_en_ne_si_test_sets diff --git a/examples/multilingual/data_scripts/download_iitb.sh b/examples/multilingual/data_scripts/download_iitb.sh new file mode 100644 index 0000000000000000000000000000000000000000..a884e20839e2a41a57405cb6af362e37bd16ab6f --- /dev/null +++ b/examples/multilingual/data_scripts/download_iitb.sh @@ -0,0 +1,35 @@ +#!/bin/bash +# Copyright (c) Facebook, Inc. and its affiliates. +# All rights reserved. +# +# This source code is licensed under the license found in the +# LICENSE file in the root directory of this source tree. + + +if [ -z $WORKDIR_ROOT ] ; +then + echo "please specify your working directory root in environment variable WORKDIR_ROOT. Exitting..." + exit +fi + +IITB=$WORKDIR_ROOT/IITB +mkdir -p $IITB +pushd $IITB + +wget http://www.cfilt.iitb.ac.in/~moses/iitb_en_hi_parallel/iitb_corpus_download/parallel.tgz +tar -xvzf parallel.tgz + +wget http://www.cfilt.iitb.ac.in/~moses/iitb_en_hi_parallel/iitb_corpus_download/dev_test.tgz +tar -xvzf dev_test.tgz + +DESTDIR=${WORKDIR_ROOT}/ML50/raw/ + +cp parallel/IITB.en-hi.en $DESTDIR/train.hi_IN-en_XX.en_XX +cp parallel/IITB.en-hi.hi $DESTDIR/train.hi_IN-en_XX.hi_IN + +cp dev_test/dev.en $DESTDIR/valid.hi_IN-en_XX.en_XX +cp dev_test/dev.hi $DESTDIR/valid.hi_IN-en_XX.hi_IN + +cp dev_test/test.en $DESTDIR/test.hi_IN-en_XX.en_XX +cp dev_test/test.hi $DESTDIR/test.hi_IN-en_XX.hi_IN +popd \ No newline at end of file diff --git a/examples/multilingual/data_scripts/download_iwslt_and_extract.sh b/examples/multilingual/data_scripts/download_iwslt_and_extract.sh new file mode 100644 index 0000000000000000000000000000000000000000..ca3591b3db1715f136773d62e4b9b9ede97d436c --- /dev/null +++ b/examples/multilingual/data_scripts/download_iwslt_and_extract.sh @@ -0,0 +1,225 @@ +#!/bin/bash +# Copyright (c) Facebook, Inc. and its affiliates. +# All rights reserved. +# +# This source code is licensed under the license found in the +# LICENSE file in the root directory of this source tree. + +#echo 'Cloning Moses github repository (for tokenization scripts)...' +#git clone https://github.com/moses-smt/mosesdecoder.git + +if [ -z $WORKDIR_ROOT ] ; +then + echo "please specify your working directory root in environment variable WORKDIR_ROOT. Exitting..." + exit +fi + + + +data_root=${WORKDIR_ROOT}/iwsltv2 +DESTDIR=${WORKDIR_ROOT}/ML50/raw + + +langs="ar_AR it_IT nl_XX ko_KR vi_VN" +echo "data_root: $data_root" + +download_path=${data_root}/downloads +raw=${DESTDIR} +tmp=${data_root}/tmp +orig=${data_root}/orig + +mkdir -p $download_path $orig $raw $tmp +####################### +download_iwslt(){ + iwslt_key=$1 + src=$2 + tgt=$3 + save_prefix=$4 + pushd ${download_path} + if [[ ! -f ${save_prefix}$src-$tgt.tgz ]]; then + wget https://wit3.fbk.eu/archive/${iwslt_key}/texts/$src/$tgt/$src-$tgt.tgz -O ${save_prefix}$src-$tgt.tgz + [ $? -eq 0 ] && return 0 + fi + popd +} + +extract_iwslt(){ + src=$1 + tgt=$2 + prefix=$3 + pushd $orig + tar zxvf ${download_path}/${prefix}$src-${tgt}.tgz + popd +} + +generate_train(){ + lsrc=$1 + ltgt=$2 + src=${lsrc:0:2} + tgt=${ltgt:0:2} + for ll in $lsrc $ltgt; do + l=${ll:0:2} + f="$orig/*/train.tags.$src-$tgt.$l" + f_raw=$raw/train.$lsrc-$ltgt.$ll + cat $f \ + | grep -v '' \ + | grep -v '' \ + | grep -v '' \ + | grep -v '' \ + | grep -v '' \ + | sed -e 's///g' \ + | sed -e 's/<\/title>//g' \ + | sed -e 's/<description>//g' \ + | sed -e 's/<\/description>//g' \ + | sed 's/^\s*//g' \ + | sed 's/\s*$//g' \ + > $f_raw + [ $? -eq 0 ] && echo "extracted $f to $f_raw" + done + return 0 +} + +convert_valid_test(){ + src=$1 + tgt=$2 + for l in $src $tgt; do + echo "lang: ${l}" + for o in `ls $orig/*/IWSLT*.TED*.$src-$tgt.$l.xml`; do + fname=${o##*/} + f=$tmp/${fname%.*} + echo "$o => $f" + grep '<seg id' $o \ + | sed -e 's/<seg id="[0-9]*">\s*//g' \ + | sed -e 's/\s*<\/seg>\s*//g' \ + | sed -e "s/\’/\'/g" \ + > $f + echo "" + done + done +} + +generate_subset(){ + lsrc=$1 + ltgt=$2 + src=${lsrc:0:2} + tgt=${ltgt:0:2} + subset=$3 + prefix=$4 + for ll in $lsrc $ltgt; do + l=${ll:0:2} + f=$tmp/$prefix.${src}-${tgt}.$l + if [[ -f $f ]]; then + cp $f $raw/$subset.${lsrc}-$ltgt.${ll} + fi + done +} +################# + +echo "downloading iwslt training and dev data" +# using multilingual for it, nl +download_iwslt "2017-01-trnmted" DeEnItNlRo DeEnItNlRo +download_iwslt "2017-01-trnted" ar en +download_iwslt "2017-01-trnted" en ar +download_iwslt "2017-01-trnted" ko en +download_iwslt "2017-01-trnted" en ko +download_iwslt "2015-01" vi en +download_iwslt "2015-01" en vi + +echo "donwloading iwslt test data" +download_iwslt "2017-01-mted-test" it en "test." +download_iwslt "2017-01-mted-test" en it "test." +download_iwslt "2017-01-mted-test" nl en "test." +download_iwslt "2017-01-mted-test" en nl "test." + +download_iwslt "2017-01-ted-test" ar en "test." +download_iwslt "2017-01-ted-test" en ar "test." +download_iwslt "2017-01-ted-test" ko en "test." +download_iwslt "2017-01-ted-test" en ko "test." +download_iwslt "2015-01-test" vi en "test." +download_iwslt "2015-01-test" en vi "test." + +echo "extract training data tar balls" +extract_iwslt DeEnItNlRo DeEnItNlRo +extract_iwslt ar en +extract_iwslt en ar +extract_iwslt ko en +extract_iwslt en ko +extract_iwslt vi en +extract_iwslt en vi + + +echo "extracting iwslt test data" +for lang in $langs; do + l=${lang:0:2} + extract_iwslt $l en "test." + extract_iwslt en $l "test." +done + +echo "convert dev and test data" +for lang in $langs; do + s_lang=${lang:0:2} + convert_valid_test $s_lang en + convert_valid_test en $s_lang +done + + + +echo "creating training data into $raw" +for lang in $langs; do + generate_train $lang en_XX + generate_train en_XX $lang +done + +echo "creating iwslt dev data into raw" +generate_subset en_XX vi_VN valid "IWSLT15.TED.tst2013" +generate_subset vi_VN en_XX valid "IWSLT15.TED.tst2013" + +generate_subset en_XX ar_AR valid "IWSLT17.TED.tst2016" +generate_subset ar_AR en_XX valid "IWSLT17.TED.tst2016" +generate_subset en_XX ko_KR valid "IWSLT17.TED.tst2016" +generate_subset ko_KR en_XX valid "IWSLT17.TED.tst2016" + + +generate_subset en_XX it_IT valid "IWSLT17.TED.tst2010" +generate_subset it_IT en_XX valid "IWSLT17.TED.tst2010" +generate_subset en_XX nl_XX valid "IWSLT17.TED.tst2010" +generate_subset nl_XX en_XX valid "IWSLT17.TED.tst2010" + +echo "creating iswslt test data into raw" +generate_subset en_XX vi_VN test "IWSLT15.TED.tst2015" +generate_subset vi_VN en_XX test "IWSLT15.TED.tst2015" + +generate_subset en_XX ar_AR test "IWSLT17.TED.tst2017" +generate_subset ar_AR en_XX test "IWSLT17.TED.tst2017" +generate_subset en_XX ko_KR test "IWSLT17.TED.tst2017" +generate_subset ko_KR en_XX test "IWSLT17.TED.tst2017" + +generate_subset en_XX it_IT test "IWSLT17.TED.tst2017.mltlng" +generate_subset it_IT en_XX test "IWSLT17.TED.tst2017.mltlng" +generate_subset en_XX nl_XX test "IWSLT17.TED.tst2017.mltlng" +generate_subset nl_XX en_XX test "IWSLT17.TED.tst2017.mltlng" + +# normalze iwslt directions into x-en +pushd $raw +for lang in $langs; do + for split in test valid; do + x_en_f1=$split.$lang-en_XX.en_XX + x_en_f2=$split.$lang-en_XX.${lang} + + en_x_f1=$split.en_XX-$lang.en_XX + en_x_f2=$split.en_XX-$lang.${lang} + + if [ -f $en_x_f1 ] && [ ! -f $x_en_f1 ]; then + echo "cp $en_x_f1 $x_en_f1" + cp $en_x_f1 $x_en_f1 + fi + if [ -f $x_en_f2 ] && [ ! -f $x_en_f2 ]; then + echo "cp $en_x_f2 $x_en_f2" + cp $en_x_f2 $x_en_f2 + fi + done +done +popd \ No newline at end of file diff --git a/examples/multilingual/data_scripts/download_lotus.sh b/examples/multilingual/data_scripts/download_lotus.sh new file mode 100644 index 0000000000000000000000000000000000000000..c08c701314a8e575637deff78381ab02c2ef6728 --- /dev/null +++ b/examples/multilingual/data_scripts/download_lotus.sh @@ -0,0 +1,46 @@ +#!/bin/bash +# Copyright (c) Facebook, Inc. and its affiliates. +# All rights reserved. +# +# This source code is licensed under the license found in the +# LICENSE file in the root directory of this source tree. + + +if [ -z $WORKDIR_ROOT ] ; +then + echo "please specify your working directory root in environment variable WORKDIR_ROOT. Exitting..." + exit +fi + + +SRCDIR=$WORKDIR_ROOT/indic_languages_corpus +DESTDIR=${WORKDIR_ROOT}/ML50/raw/ +mkdir -p $SRCDIR +mkdir -p $DESTDIR + +cd $SRCDIR +wget http://lotus.kuee.kyoto-u.ac.jp/WAT/indic-multilingual/indic_languages_corpus.tar.gz +tar -xvzf indic_languages_corpus.tar.gz + +SRC_EXTRACT_DIR=$SRCDIR/indic_languages_corpus/bilingual + +cp $SRC_EXTRACT_DIR/ml-en/train.ml $DESTDIR/train.ml_IN-en_XX.ml_IN +cp $SRC_EXTRACT_DIR/ml-en/train.en $DESTDIR/train.ml_IN-en_XX.en_XX +cp $SRC_EXTRACT_DIR/ml-en/dev.ml $DESTDIR/valid.ml_IN-en_XX.ml_IN +cp $SRC_EXTRACT_DIR/ml-en/dev.en $DESTDIR/valid.ml_IN-en_XX.en_XX +cp $SRC_EXTRACT_DIR/ml-en/test.ml $DESTDIR/test.ml_IN-en_XX.ml_IN +cp $SRC_EXTRACT_DIR/ml-en/test.en $DESTDIR/test.ml_IN-en_XX.en_XX + +cp $SRC_EXTRACT_DIR/ur-en/train.ur $DESTDIR/train.ur_PK-en_XX.ur_PK +cp $SRC_EXTRACT_DIR/ur-en/train.en $DESTDIR/train.ur_PK-en_XX.en_XX +cp $SRC_EXTRACT_DIR/ur-en/dev.ur $DESTDIR/valid.ur_PK-en_XX.ur_PK +cp $SRC_EXTRACT_DIR/ur-en/dev.en $DESTDIR/valid.ur_PK-en_XX.en_XX +cp $SRC_EXTRACT_DIR/ur-en/test.ur $DESTDIR/test.ur_PK-en_XX.ur_PK +cp $SRC_EXTRACT_DIR/ur-en/test.en $DESTDIR/test.ur_PK-en_XX.en_XX + +cp $SRC_EXTRACT_DIR/te-en/train.te $DESTDIR/train.te_IN-en_XX.te_IN +cp $SRC_EXTRACT_DIR/te-en/train.en $DESTDIR/train.te_IN-en_XX.en_XX +cp $SRC_EXTRACT_DIR/te-en/dev.te $DESTDIR/valid.te_IN-en_XX.te_IN +cp $SRC_EXTRACT_DIR/te-en/dev.en $DESTDIR/valid.te_IN-en_XX.en_XX +cp $SRC_EXTRACT_DIR/te-en/test.te $DESTDIR/test.te_IN-en_XX.te_IN +cp $SRC_EXTRACT_DIR/te-en/test.en $DESTDIR/test.te_IN-en_XX.en_XX diff --git a/examples/multilingual/data_scripts/download_ted_and_extract.py b/examples/multilingual/data_scripts/download_ted_and_extract.py new file mode 100644 index 0000000000000000000000000000000000000000..eb756680fa7dc31a14ba45c216776a6d60c16b60 --- /dev/null +++ b/examples/multilingual/data_scripts/download_ted_and_extract.py @@ -0,0 +1,338 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +import itertools +import os +import csv +from collections import defaultdict +from six.moves import zip +import io +import wget +import sys + +from subprocess import check_call, check_output + +# scripts and data locations +CWD = os.getcwd() +UTILS = f"{CWD}/utils" + +MOSES = f"{UTILS}/mosesdecoder" + +WORKDIR_ROOT = os.environ.get('WORKDIR_ROOT', None) + +if WORKDIR_ROOT is None or not WORKDIR_ROOT.strip(): + print('please specify your working directory root in OS environment variable WORKDIR_ROOT. Exitting..."') + sys.exit(-1) + + +# please donwload mosesdecoder here: +detok_cmd = f'{MOSES}/scripts/tokenizer/detokenizer.perl' + + +def call(cmd): + print(f"Executing: {cmd}") + check_call(cmd, shell=True) + +class MultiLingualAlignedCorpusReader(object): + """A class to read TED talk dataset + """ + + def __init__(self, corpus_path, delimiter='\t', + target_token=True, bilingual=True, corpus_type='file', + lang_dict={'source': ['fr'], 'target': ['en']}, + eval_lang_dict=None, zero_shot=False, + detok=True, + ): + + self.empty_line_flag = 'NULL' + self.corpus_path = corpus_path + self.delimiter = delimiter + self.bilingual = bilingual + self.lang_dict = lang_dict + self.lang_set = set() + self.target_token = target_token + self.zero_shot = zero_shot + self.eval_lang_dict = eval_lang_dict + self.corpus_type = corpus_type + self.detok = detok + + for list_ in self.lang_dict.values(): + for lang in list_: + self.lang_set.add(lang) + + self.data = dict() + self.data['train'] = self.read_aligned_corpus(split_type='train') + self.data['test'] = self.read_aligned_corpus(split_type='test') + self.data['dev'] = self.read_aligned_corpus(split_type='dev') + + def read_data(self, file_loc_): + data_list = list() + with io.open(file_loc_, 'r', encoding='utf8') as fp: + for line in fp: + try: + text = line.strip() + except IndexError: + text = self.empty_line_flag + data_list.append(text) + return data_list + + def filter_text(self, dict_): + if self.target_token: + field_index = 1 + else: + field_index = 0 + data_dict = defaultdict(list) + list1 = dict_['source'] + list2 = dict_['target'] + for sent1, sent2 in zip(list1, list2): + try: + src_sent = ' '.join(sent1.split()[field_index: ]) + except IndexError: + src_sent = 'NULL' + + if src_sent.find(self.empty_line_flag) != -1 or len(src_sent) == 0: + continue + + elif sent2.find(self.empty_line_flag) != -1 or len(sent2) == 0: + continue + + else: + data_dict['source'].append(sent1) + data_dict['target'].append(sent2) + return data_dict + + def read_file(self, split_type, data_type): + return self.data[split_type][data_type] + + def save_file(self, path_, split_type, data_type, lang): + tok_file = tok_file_name(path_, lang) + with io.open(tok_file, 'w', encoding='utf8') as fp: + for line in self.data[split_type][data_type]: + fp.write(line + '\n') + if self.detok: + de_tok(tok_file, lang) + + def add_target_token(self, list_, lang_id): + new_list = list() + token = '__' + lang_id + '__' + for sent in list_: + new_list.append(token + ' ' + sent) + return new_list + + def read_from_single_file(self, path_, s_lang, t_lang): + data_dict = defaultdict(list) + with io.open(path_, 'r', encoding='utf8') as fp: + reader = csv.DictReader(fp, delimiter='\t', quoting=csv.QUOTE_NONE) + for row in reader: + data_dict['source'].append(row[s_lang]) + data_dict['target'].append(row[t_lang]) + + if self.target_token: + text = self.add_target_token(data_dict['source'], t_lang) + data_dict['source'] = text + + return data_dict['source'], data_dict['target'] + + def read_aligned_corpus(self, split_type='train'): + data_dict = defaultdict(list) + iterable = [] + s_list = [] + t_list = [] + + if self.zero_shot: + if split_type == "train": + iterable = zip(self.lang_dict['source'], self.lang_dict['target']) + else: + iterable = zip(self.eval_lang_dict['source'], self.eval_lang_dict['target']) + + elif self.bilingual: + iterable = itertools.product(self.lang_dict['source'], self.lang_dict['target']) + + for s_lang, t_lang in iterable: + if s_lang == t_lang: + continue + if self.corpus_type == 'file': + split_type_file_path = os.path.join(self.corpus_path, + "all_talks_{}.tsv".format(split_type)) + s_list, t_list = self.read_from_single_file(split_type_file_path, + s_lang=s_lang, + t_lang=t_lang) + data_dict['source'] += s_list + data_dict['target'] += t_list + new_data_dict = self.filter_text(data_dict) + return new_data_dict + + +def read_langs(corpus_path): + split_type_file_path = os.path.join(corpus_path, 'extracted', + "all_talks_dev.tsv") + with io.open(split_type_file_path, 'r', encoding='utf8') as fp: + reader = csv.DictReader(fp, delimiter='\t', quoting=csv.QUOTE_NONE) + header = next(reader) + return [k for k in header.keys() if k != 'talk_name'] + +def extra_english(corpus_path, split): + split_type_file_path = os.path.join(corpus_path, + f"all_talks_{split}.tsv") + output_split_type_file_path = os.path.join(corpus_path, + f"all_talks_{split}.en") + with io.open(split_type_file_path, 'r', encoding='utf8') as fp, io.open(output_split_type_file_path, 'w', encoding='utf8') as fw: + reader = csv.DictReader(fp, delimiter='\t', quoting=csv.QUOTE_NONE) + for row in reader: + line = row['en'] + fw.write(line + '\n') + de_tok(output_split_type_file_path, 'en') + + + +def tok_file_name(filename, lang): + seps = filename.split('.') + seps.insert(-1, 'tok') + tok_file = '.'.join(seps) + return tok_file + +def de_tok(tok_file, lang): + # seps = tok_file.split('.') + # seps.insert(-1, 'detok') + # de_tok_file = '.'.join(seps) + de_tok_file = tok_file.replace('.tok.', '.') + cmd = 'perl {detok_cmd} -l {lang} < {tok_file} > {de_tok_file}'.format( + detok_cmd=detok_cmd, tok_file=tok_file, + de_tok_file=de_tok_file, lang=lang[:2]) + call(cmd) + +def extra_bitex( + ted_data_path, + lsrc_lang, + ltrg_lang, + target_token, + output_data_path, +): + def get_ted_lang(lang): + long_langs = ['pt-br', 'zh-cn', 'zh-tw', 'fr-ca'] + if lang[:5] in long_langs: + return lang[:5] + elif lang[:4] =='calv': + return lang[:5] + elif lang in ['pt_BR', 'zh_CN', 'zh_TW', 'fr_CA']: + return lang.lower().replace('_', '-') + return lang[:2] + src_lang = get_ted_lang(lsrc_lang) + trg_lang = get_ted_lang(ltrg_lang) + train_lang_dict={'source': [src_lang], 'target': [trg_lang]} + eval_lang_dict = {'source': [src_lang], 'target': [trg_lang]} + + obj = MultiLingualAlignedCorpusReader(corpus_path=ted_data_path, + lang_dict=train_lang_dict, + target_token=target_token, + corpus_type='file', + eval_lang_dict=eval_lang_dict, + zero_shot=False, + bilingual=True) + + os.makedirs(output_data_path, exist_ok=True) + lsrc_lang = lsrc_lang.replace('-', '_') + ltrg_lang = ltrg_lang.replace('-', '_') + obj.save_file(output_data_path + f"/train.{lsrc_lang}-{ltrg_lang}.{lsrc_lang}", + split_type='train', data_type='source', lang=src_lang) + obj.save_file(output_data_path + f"/train.{lsrc_lang}-{ltrg_lang}.{ltrg_lang}", + split_type='train', data_type='target', lang=trg_lang) + + obj.save_file(output_data_path + f"/test.{lsrc_lang}-{ltrg_lang}.{lsrc_lang}", + split_type='test', data_type='source', lang=src_lang) + obj.save_file(output_data_path + f"/test.{lsrc_lang}-{ltrg_lang}.{ltrg_lang}", + split_type='test', data_type='target', lang=trg_lang) + + obj.save_file(output_data_path + f"/valid.{lsrc_lang}-{ltrg_lang}.{lsrc_lang}", + split_type='dev', data_type='source', lang=src_lang) + obj.save_file(output_data_path + f"/valid.{lsrc_lang}-{ltrg_lang}.{ltrg_lang}", + split_type='dev', data_type='target', lang=trg_lang) + + +def bar_custom(current, total, width=80): + print("Downloading: %d%% [%d / %d] Ks" % (current / total * 100, current / 1000, total / 1000), end='\r') + + +def download_and_extract(download_to, extract_to): + url = 'http://phontron.com/data/ted_talks.tar.gz' + filename = f"{download_to}/ted_talks.tar.gz" + if os.path.exists(filename): + print(f'{filename} has already been downloaded so skip') + else: + filename = wget.download(url, filename, bar=bar_custom) + if os.path.exists(f'{extract_to}/all_talks_train.tsv'): + print(f'Already extracted so skip') + else: + extract_cmd = f'tar xzfv "{filename}" -C "{extract_to}"' + call(extract_cmd) + + +if __name__ == "__main__": + import argparse + parser = argparse.ArgumentParser() + parser.add_argument('--ted_data_path', type=str, default=WORKDIR_ROOT, required=False) + parser.add_argument( + '--direction-list', + type=str, + # default=None, + #for ML50 + default=( + "bn_IN-en_XX,he_IL-en_XX,fa_IR-en_XX,id_ID-en_XX,sv_SE-en_XX,pt_XX-en_XX,ka_GE-en_XX,ka_GE-en_XX,th_TH-en_XX," + "mr_IN-en_XX,hr_HR-en_XX,uk_UA-en_XX,az_AZ-en_XX,mk_MK-en_XX,gl_ES-en_XX,sl_SI-en_XX,mn_MN-en_XX," + #non-english directions + # "fr_XX-de_DE," # replaced with wmt20 + # "ja_XX-ko_KR,es_XX-pt_XX,ru_RU-sv_SE,hi_IN-bn_IN,id_ID-ar_AR,cs_CZ-pl_PL,ar_AR-tr_TR" + ), + required=False) + parser.add_argument('--target-token', action='store_true', default=False) + parser.add_argument('--extract-all-english', action='store_true', default=False) + + args = parser.parse_args() + + import sys + import json + + # TED Talks data directory + ted_data_path = args.ted_data_path + + download_to = f'{ted_data_path}/downloads' + extract_to = f'{ted_data_path}/extracted' + + #DESTDIR=${WORKDIR_ROOT}/ML50/raw/ + output_path = f'{ted_data_path}/ML50/raw' + os.makedirs(download_to, exist_ok=True) + os.makedirs(extract_to, exist_ok=True) + os.makedirs(output_path, exist_ok=True) + download_and_extract(download_to, extract_to) + + + if args.extract_all_english: + for split in ['train', 'dev', 'test']: + extra_english(ted_data_path, split) + exit(0) + if args.direction_list is not None: + directions = args.direction_list.strip().split(',') + directions = [tuple(d.strip().split('-', 1)) for d in directions if d] + else: + langs = read_langs(ted_data_path) + # directions = [ + # '{}.{}'.format(src, tgt) + # for src in langs + # for tgt in langs + # if src < tgt + # ] + directions = [('en', tgt) for tgt in langs if tgt != 'en'] + print(f'num directions={len(directions)}: {directions}') + + for src_lang, trg_lang in directions: + print('--working on {}-{}'.format(src_lang, trg_lang)) + extra_bitex( + extract_to, + src_lang, + trg_lang, + target_token=args.target_token, + output_data_path=output_path + ) diff --git a/examples/multilingual/data_scripts/download_wat19_my.sh b/examples/multilingual/data_scripts/download_wat19_my.sh new file mode 100644 index 0000000000000000000000000000000000000000..c1e2d47287a29af4576e7a63641e8152ecb63c44 --- /dev/null +++ b/examples/multilingual/data_scripts/download_wat19_my.sh @@ -0,0 +1,36 @@ +#!/bin/bash +# Copyright (c) Facebook, Inc. and its affiliates. +# All rights reserved. +# +# This source code is licensed under the license found in the +# LICENSE file in the root directory of this source tree. + + +if [ -z $WORKDIR_ROOT ] ; +then + echo "please specify your working directory root in environment variable WORKDIR_ROOT. Exitting..." + exit +fi + + +SRCDIR=$WORKDIR_ROOT/indic_languages_corpus +DESTDIR=$WORKDIR_ROOT/ML50/raw +mkdir -p $SRCDIR +mkdir -p $DESTDIR + +WAT_MY_EN=wat2020.my-en.zip +cd $SRCDIR +# please refer to http://lotus.kuee.kyoto-u.ac.jp/WAT/my-en-data/ for latest URL if the following url expired +#- The data used for WAT2020 are identical to those used in WAT2019. +wget http://lotus.kuee.kyoto-u.ac.jp/WAT/my-en-data/$WAT_MY_EN +unzip $WAT_MY_EN + + +SRC_EXTRACT_DIR=$SRCDIR/wat2020.my-en/alt + +cp $SRC_EXTRACT_DIR/train.alt.en $DESTDIR/train.my_MM-en_XX.en_XX +cp $SRC_EXTRACT_DIR/train.alt.my $DESTDIR/train.my_MM-en_XX.my_MM +cp $SRC_EXTRACT_DIR/dev.alt.en $DESTDIR/valid.my_MM-en_XX.en_XX +cp $SRC_EXTRACT_DIR/dev.alt.my $DESTDIR/valid.my_MM-en_XX.my_MM +cp $SRC_EXTRACT_DIR/test.alt.en $DESTDIR/test.my_MM-en_XX.en_XX +cp $SRC_EXTRACT_DIR/test.alt.my $DESTDIR/test.my_MM-en_XX.my_MM diff --git a/examples/multilingual/data_scripts/download_wmt19_and_before.py b/examples/multilingual/data_scripts/download_wmt19_and_before.py new file mode 100644 index 0000000000000000000000000000000000000000..3465731eb3e55047c44d1b336a97e99cb3a89a53 --- /dev/null +++ b/examples/multilingual/data_scripts/download_wmt19_and_before.py @@ -0,0 +1,899 @@ +from typing import NamedTuple, List +from urllib.parse import urlparse +import os, sys +import subprocess +from subprocess import check_call, check_output +import glob +import wget +import re +import multiprocessing as mp +from functools import partial +import pathlib +from collections import OrderedDict + +WORKDIR_ROOT = os.environ.get('WORKDIR_ROOT', None) + +if WORKDIR_ROOT is None or not WORKDIR_ROOT.strip(): + print('please specify your working directory root in OS environment variable WORKDIR_ROOT. Exitting..."') + sys.exit(-1) + +# scripts and data locations +CWD = os.getcwd() +UTILS = f"{CWD}/utils" + +MOSES = f"{UTILS}/mosesdecoder" +SGM_TOOL = f'{MOSES}/scripts/ems/support/input-from-sgm.perl' + +TMX2CORPUS = f"{UTILS}/tmx2corpus" +TMX_TOOL = f'python {TMX2CORPUS}/tmx2corpus.py' + +to_data_path = f'{WORKDIR_ROOT}/wmt' +download_to = f'{to_data_path}/downloads' +manually_downloads = f'{to_data_path}/downloads' +extract_to = f'{to_data_path}/extracted' +#DESTDIR=${WORKDIR_ROOT}/ML50/raw/ +raw_data = f'{WORKDIR_ROOT}/ML50/raw' +#### + +class DLDataset(NamedTuple): + name: str + train_urls: List[str] + valid_urls: List[str] + test_urls: List[str] + train_files_patterns: List[str] = [] + valid_files_patterns: List[str] = [] + test_files_patterns: List[str] = [] + + + +def bar_custom(current, total, width=80): + print("Downloading: %d%% [%d / %d] Ks" % (current / total * 100, current / 1000, total / 1000), end='\r') + +def get_downloaded_file(dl_folder, url): + if isinstance(url, tuple): + url, f = url + else: + url_f = urlparse(url) + # f = os.path.split(url_f.path)[-1] + f = '_'.join(url_f.path.split('/')[1:]) + return url, f"{dl_folder}/{f}" + +def download_parts_and_combine(dl_folder, urls, filename): + parts = [] + for url_record in urls: + url, part_file = get_downloaded_file(dl_folder, url_record) + if os.path.exists(part_file): + print(f'{part_file} has already been downloaded so skip') + else: + part_file = wget.download(url, part_file, bar=bar_custom) + parts.append(part_file) + + def get_combine_cmd(parts): + #default as tar.gz.?? + return f'cat {" ".join(parts)} > {filename}' + + combine_cmd = get_combine_cmd(parts) + call(combine_cmd, debug=True) + return filename + +def download_a_url(dl_folder, url): + url, filename = get_downloaded_file(dl_folder, url) + if os.path.exists(filename): + print(f'{filename} has already been downloaded so skip') + return filename + + print(f'downloading {url} to {filename}') + if isinstance(url, list) or isinstance(url, tuple): + download_parts_and_combine(dl_folder, url, filename) + else: + wget.download(url, filename, bar=bar_custom) + print(f'dowloaded: {filename}') + return filename + +def download_files(dl_folder, urls, completed_urls={}): + for url_record in urls: + url, _ = get_downloaded_file(dl_folder, url_record) + filename = download_a_url(dl_folder, url_record) + completed_urls[str(url)] = filename + return completed_urls + +def check_need_manual_downalod(dl_folder, to_manually_download_urls): + to_be_manually_dowloaded = [] + manually_completed_urls = {} + for url_record, instruction in to_manually_download_urls: + url, filename = get_downloaded_file(dl_folder, url_record) + if not os.path.exists(filename): + print(f'{url} need to be download manually, please download it manually following {instruction}; and copy it to {filename}') + to_be_manually_dowloaded.append((url, filename)) + else: + manually_completed_urls[url] = filename + # if len(to_be_manually_dowloaded) > 0: + # raise ValueError('Missing files that need to be downloaded manually; stop the process now.') + return to_be_manually_dowloaded + +def download_dataset(to_folder, dl_dataset, completed_urls={}): + download_files(to_folder, dl_dataset.train_urls, completed_urls) + download_files(to_folder, dl_dataset.valid_urls, completed_urls) + download_files(to_folder, dl_dataset.test_urls, completed_urls) + print('completed downloading') + return completed_urls + +def call(cmd, debug=False): + if debug: + print(cmd) + check_call(cmd, shell=True) + + +def get_extract_name(file_path): + path = os.path.split(file_path) + return path[-1] + '_extract' #.split('.')[0] + +def extract_file(downloaded_file, extract_folder, get_extract_name=get_extract_name, debug=False): + extract_name = get_extract_name(downloaded_file) + extract_to = f'{extract_folder}/{extract_name}' + os.makedirs(extract_to, exist_ok=True) + if os.path.exists(f'{extract_to}/DONE'): + print(f'{downloaded_file} has already been extracted to {extract_to} so skip') + return extract_to + def get_extract_cmd(filename): + if filename.endswith('.tgz') or filename.endswith('tar.gz'): + return f'tar xzfv {filename} -C {extract_to}' + elif filename.endswith('.gz.tar'): + return f'tar xfv {filename} -C {extract_to}; (cd {extract_to}; gzip -d *.gz; [ $? -eq 0 ] || gzip -d */*.gz)' + elif filename.endswith('.tar'): + return f'tar xfv {filename} -C {extract_to}' + elif filename.endswith('.gz'): + return f'cp {filename} {extract_to}; (cd {extract_to}; gzip -d *.gz)' + elif filename.endswith('.zip'): + return f'unzip {filename} -d {extract_to}' + extract_cmd = get_extract_cmd(downloaded_file) + print(f'extracting {downloaded_file}') + if isinstance(extract_cmd, list): + for c in extract_cmd: + call(c, debug=debug) + else: + call(extract_cmd, debug=debug) + call(f'echo DONE > {extract_to}/DONE') + return extract_to + + +def extract_all_files( + completed_urls, extract_folder, + get_extract_name=get_extract_name, + completed_extraction={}, + debug=False): + extracted_folders = OrderedDict() + for url, downloaded_file in set(completed_urls.items()): + if downloaded_file in completed_extraction: + print(f'{downloaded_file} is already extracted; so skip') + continue + folder = extract_file(downloaded_file, extract_folder, get_extract_name, debug) + extracted_folders[url] = folder + return extracted_folders + + +def my_glob(folder): + for p in [f'{folder}/*', f'{folder}/*/*', f'{folder}/*/*/*']: + for f in glob.glob(p): + yield f + + +def sgm2raw(sgm, debug): + to_file = sgm[0:len(sgm) - len('.sgm')] + if os.path.exists(to_file): + debug and print(f'{sgm} already converted to {to_file}; so skip') + return to_file + cmd = f'{SGM_TOOL} < {sgm} > {to_file}' + call(cmd, debug) + return to_file + +def tmx2raw(tmx, debug): + to_file = tmx[0:len(tmx) - len('.tmx')] + to_folder = os.path.join(*os.path.split(tmx)[:-1]) + if os.path.exists(f'{to_folder}/bitext.en'): + debug and print(f'{tmx} already extracted to {to_file}; so skip') + return to_file + cmd = f'(cd {to_folder}; {TMX_TOOL} {tmx})' + call(cmd, debug) + return to_file + +CZENG16_REGEX = re.compile(r'.*?data.plaintext-format/0[0-9]train$') +WMT19_WIKITITLES_REGEX = re.compile(r'.*?wikititles-v1.(\w\w)-en.tsv.gz') +TSV_REGEX = re.compile(r'.*?(\w\w)-(\w\w).tsv$') + + + +def cut_wikitles(wiki_file, debug): + # different languages have different file names: + if wiki_file.endswith('wiki/fi-en/titles.fi-en'): + to_file1 = f'{wiki_file}.fi' + to_file2 = f'{wiki_file}.en' + BACKSLASH = '\\' + cmd1 = f"cat {wiki_file} | sed 's/|||/{BACKSLASH}t/g' |cut -f1 |awk '{{$1=$1}};1' > {to_file1}" + cmd2 = f"cat {wiki_file} | sed 's/|||/{BACKSLASH}t/g' |cut -f2 |awk '{{$1=$1}};1' > {to_file2}" +# elif WMT19_WIKITITLES_REGEX.match(wiki_file): +# src = WMT19_WIKITITLES_REGEX.match(wiki_file).groups()[0] +# to_file1 = f'{wiki_file}.{src}' +# to_file2 = f'{wiki_file}.en' +# cmd1 = f"cat {wiki_file} | cut -f1 |awk '{{$1=$1}};1' > {to_file1}" +# cmd2 = f"cat {wiki_file} | cut -f2 |awk '{{$1=$1}};1' > {to_file2}" + else: + return None + if os.path.exists(to_file1) and os.path.exists(to_file2): + debug and print(f'{wiki_file} already processed to {to_file1} and {to_file2}; so skip') + return wiki_file + + call(cmd1, debug=debug) + call(cmd2, debug=debug) + return wiki_file + +def cut_tsv(file, debug): + m = TSV_REGEX.match(file) + if m is None: + raise ValueError(f'{file} is not matching tsv pattern') + src = m.groups()[0] + tgt = m.groups()[1] + + to_file1 = f'{file}.{src}' + to_file2 = f'{file}.{tgt}' + cmd1 = f"cat {file} | cut -f1 |awk '{{$1=$1}};1' > {to_file1}" + cmd2 = f"cat {file} | cut -f2 |awk '{{$1=$1}};1' > {to_file2}" + if os.path.exists(to_file1) and os.path.exists(to_file2): + debug and print(f'{file} already processed to {to_file1} and {to_file2}; so skip') + return file + + call(cmd1, debug=debug) + call(cmd2, debug=debug) + return file + + +def convert_file_if_needed(file, debug): + if file.endswith('.sgm'): + return sgm2raw(file, debug) + elif file.endswith('.tmx'): + return tmx2raw(file, debug) + elif file.endswith('wiki/fi-en/titles.fi-en'): + return cut_wikitles(file, debug) +# elif WMT19_WIKITITLES_REGEX.match(file): +# return cut_wikitles(file, debug) + elif file.endswith('.tsv'): + return cut_tsv(file, debug) + elif CZENG16_REGEX.match(file): + return convert2czeng17(file, debug) + else: + return file + + +def convert_files_if_needed(extracted_foldrs, my_glob=my_glob, debug=False): + return { + url: list(sorted(set(convert_file_if_needed(f, debug)) for f in sorted(set(my_glob(folder))))) + for url, folder in extracted_foldrs.items() + } + +def match_patt(file_path, file_pattern, src, tgt, lang): + return file_pattern.format(src=src, tgt=tgt, lang=lang) in file_path + +def match_patts(file_path, file_patterns, src, tgt, lang): + for file_pattern in file_patterns: + params = { k: v for k, v in [('src', src), ('tgt', tgt), ('lang', lang)] if k in file_pattern} + matching = file_pattern.format(**params) + + if isinstance(file_pattern, tuple): + pattern, directions = file_pattern + if f'{src}-{tgt}' in directions and matching in file_path: + return True + else: + if matching in file_path: + return True + return False + +def extracted_glob(extracted_folder, file_patterns, src, tgt, lang): + def get_matching_pattern(file_pattern): + params = { + k: v + for k, v in [('src', src), ('tgt', tgt), ('lang', lang)] + if '{' + k + '}' in file_pattern + } + file_pattern = re.sub(r'{src:(.*?)}', r'\1' if lang == src else '', file_pattern) + file_pattern = re.sub(r'{tgt:(.*?)}', r'\1' if lang == tgt else '', file_pattern) + file_pattern = file_pattern.format(**params) + return file_pattern + for file_pattern in file_patterns: + if isinstance(file_pattern, tuple): + file_pattern, lang_pairs = file_pattern + if f'{src}-{tgt}' not in lang_pairs: + continue +# print('working on pattern: ', file_pattern, lang_pairs ) + matching_pattern = get_matching_pattern(file_pattern) + if matching_pattern is None: + continue + glob_patterns = f'{extracted_folder}/{matching_pattern}' +# print('glob_patterns: ', glob_patterns) + for f in glob.glob(glob_patterns): + yield f + +# for debug usage +def all_extracted_files(split, src, tgt, extracted_folders, split_urls): + def get_url(url): + if isinstance(url, tuple): + url, downloaded_file = url + return url + return [ + f + for url in split_urls + for f in my_glob(extracted_folders[str(get_url(url))]) + ] + +def concat_files(split, src, tgt, extracted_folders, split_urls, path_patterns, to_folder, debug=False): +# if debug: +# print('extracted files to be filtered by patterns: ', +# '\n\t'.join(sorted(all_extracted_files(split, src, tgt, extracted_folders, split_urls)))) + for lang in [src, tgt]: + to_file = f'{to_folder}/{split}.{src}-{tgt}.{lang}' + s_src, s_tgt, s_lang = src.split('_')[0], tgt.split('_')[0], lang.split('_')[0] + files = [] + for url in split_urls: + if isinstance(url, tuple): + url, downloaded_file = url + if str(url) not in extracted_folders: + print(f'warning: {url} not in extracted files') + for extracted_file in set( + extracted_glob( + extracted_folders[str(url)], path_patterns, + s_src, s_tgt, s_lang)): + files.append(extracted_file) + if len(files) == 0: + print('warning: ', f'No files found for split {to_file}') + continue + files = sorted(set(files)) + print(f'concating {len(files)} files into {to_file}') + cmd = ['cat'] + [f'"{f}"' for f in files] + [f'>{to_file}'] + cmd = " ".join(cmd) + call(cmd, debug=debug) + +UTILS = os.path.join(pathlib.Path(__file__).parent, 'utils') +LID_MODEL = f'{download_to}/lid.176.bin' +LID_MULTI = f'{UTILS}/fasttext_multi_filter.py' + +def lid_filter(split, src, tgt, from_folder, to_folder, debug=False): + if not os.path.exists(LID_MODEL): + call(f'wget -nc https://dl.fbaipublicfiles.com/fasttext/supervised-models/lid.176.bin -O {LID_MODEL}') + from_prefix = f'{from_folder}/{split}.{src}-{tgt}' + to_prefix = f'{to_folder}/{split}.{src}-{tgt}' + if os.path.exists(f'{from_prefix}.{src}') and os.path.exists(f'{from_prefix}.{tgt}'): + s_src, s_tgt = src.split('_')[0], tgt.split('_')[0] + cmd = ( + f'python {LID_MULTI} --model {LID_MODEL} --inputs {from_prefix}.{src} {from_prefix}.{tgt} ' + f'--langs {s_src} {s_tgt} --outputs {to_prefix}.{src} {to_prefix}.{tgt}' + ) + print(f'filtering {from_prefix}') + call(cmd, debug=debug) + +def concat_into_splits(dl_dataset, src, tgt, extracted_folders, to_folder, debug): + to_folder_tmp = f"{to_folder}_tmp" + os.makedirs(to_folder_tmp, exist_ok=True) + concat_files('train', src, tgt, + extracted_folders, + split_urls=dl_dataset.train_urls, + path_patterns=dl_dataset.train_files_patterns, + to_folder=to_folder_tmp, debug=debug) + lid_filter('train', src, tgt, to_folder_tmp, to_folder, debug) + + concat_files('valid', src, tgt, + extracted_folders, + split_urls=dl_dataset.valid_urls, + path_patterns=dl_dataset.valid_files_patterns, + to_folder=to_folder, debug=debug) + concat_files('test', src, tgt, + extracted_folders, + split_urls=dl_dataset.test_urls, + path_patterns=dl_dataset.test_files_patterns, + to_folder=to_folder, debug=debug) + + +def download_multi(dl_folder, extract_folder, urls, num_processes=8, debug=False): + pool = mp.Pool(processes=num_processes) + download_f = partial(download_a_url, dl_folder) + downloaded_files = pool.imap_unordered(download_f, urls) + pool.close() + pool.join() + +BLEU_REGEX = re.compile("^BLEU\\S* = (\\S+) ") +def run_eval_bleu(cmd): + output = check_output(cmd, shell=True, stderr=subprocess.STDOUT).decode("utf-8").strip() + print(output) + bleu = -1.0 + for line in output.strip().split('\n'): + m = BLEU_REGEX.search(line) + if m is not None: + bleu = m.groups()[0] + bleu = float(bleu) + break + return bleu + +def check_wmt_test_bleu(raw_folder, wmt_lang_pairs): + not_matchings = [] + for wmt, src_tgts in wmt_lang_pairs: + for src_tgt in src_tgts: + print(f'checking test bleus for: {src_tgt} at {wmt}') + src, tgt = src_tgt.split('-') + ssrc, stgt = src[:2], tgt[:2] + if os.path.exists(f'{raw_folder}/test.{tgt}-{src}.{src}'): + # reversed direction may have different test set + test_src = f'{raw_folder}/test.{tgt}-{src}.{src}' + else: + test_src = f'{raw_folder}/test.{src}-{tgt}.{src}' + cmd1 = f'cat {test_src} | sacrebleu -t "{wmt}" -l {stgt}-{ssrc}; [ $? -eq 0 ] || echo ""' + test_tgt = f'{raw_folder}/test.{src}-{tgt}.{tgt}' + cmd2 = f'cat {test_tgt} | sacrebleu -t "{wmt}" -l {ssrc}-{stgt}; [ $? -eq 0 ] || echo ""' + bleu1 = run_eval_bleu(cmd1) + if bleu1 != 100.0: + not_matchings.append(f'{wmt}:{src_tgt} source side not matching: {test_src}') + bleu2 = run_eval_bleu(cmd2) + if bleu2 != 100.0: + not_matchings.append(f'{wmt}:{src_tgt} target side not matching: {test_tgt}') + return not_matchings + +def download_and_extract( + to_folder, lang_pairs, dl_dataset, + to_manually_download_urls, + completed_urls={}, completed_extraction={}, + debug=False): + + dl_folder = f'{to_folder}/downloads' + extract_folder = f'{to_folder}/extracted' + raw_folder = f'{to_folder}/raw' + lid_filtered = f'{to_folder}/lid_filtered' + + os.makedirs(extract_folder, exist_ok=True) + os.makedirs(raw_folder, exist_ok=True) + os.makedirs(lid_filtered, exist_ok=True) + + + to_be_manually_dowloaded = check_need_manual_downalod(dl_folder, to_manually_download_urls) + + completed_urls = download_dataset( + dl_folder, dl_dataset, completed_urls) + if debug: + print('completed urls: ', completed_urls) + + + extracted_folders = extract_all_files( + completed_urls, + extract_folder=extract_folder, + completed_extraction=completed_extraction, + debug=debug) + if debug: + print('download files have been extracted to folders: ', extracted_folders) + + converted_files = convert_files_if_needed(extracted_folders, debug=False) + for src_tgt in lang_pairs: + print(f'working on {dl_dataset.name}: {src_tgt}') + src, tgt = src_tgt.split('-') + concat_into_splits(dl_dataset, + src=src, tgt=tgt, + extracted_folders=extracted_folders, + to_folder=raw_folder, debug=debug) + print('completed data into: ', raw_folder) + +def download_czang16(download_to, username=None): + wgets = [ + f'wget --user={username} --password=czeng -P {download_to} http://ufallab.ms.mff.cuni.cz/~bojar/czeng16-data/data-plaintext-format.{i}.tar' + for i in range(10)] + cmds = [] + for i, cmd in enumerate(wgets): + filename = f'{download_to}/data-plaintext-format.{i}.tar' + if os.path.exists(filename): + print(f'{filename} has already been downloaded; so skip') + continue + cmds.append(cmd) + if cmds and username is None: + raise ValueError('No czeng username is given; please register at http://ufal.mff.cuni.cz/czeng/czeng16 to obtain username to download') + for cmd in cmds: + call(cmd) + print('done with downloading czeng1.6') + +def download_czeng17_script(download_to, extract_folder, debug=False): + url = 'http://ufal.mff.cuni.cz/czeng/download.php?f=convert_czeng16_to_17.pl.zip' + filename = f'{download_to}/convert_czeng16_to_17.pl.zip' + extract_to = f'{extract_folder}/{get_extract_name(filename)}' + script_path = f'{extract_to}/convert_czeng16_to_17.pl' + + if not os.path.exists(script_path): + wget.download(url, filename, bar=bar_custom) + extract_to = extract_file(f'{download_to}/convert_czeng16_to_17.pl.zip', extract_folder, get_extract_name=get_extract_name, debug=debug) + return script_path + +czeng17_script_path = "" +def convert2czeng17(file, debug): + en_file = f'{file}.en' + cs_file = f'{file}.cs' + + if not os.path.exists(en_file) or not os.path.exists(cs_file): + cs_cmd = f'cat {file} | perl {czeng17_script_path} | cut -f3 > {cs_file}' + en_cmd = f'cat {file} | perl {czeng17_script_path} | cut -f4 > {en_file}' + call(cs_cmd, debug) + call(en_cmd, debug) + else: + print(f'already extracted: {en_file} and {cs_file}') + return file + +def extract_czeng17(extract_folder, debug=False): + url = 'http://ufal.mff.cuni.cz/czeng/download.php?f=convert_czeng16_to_17.pl.zip' + filename = f'{download_to}/convert_czeng16_to_17.pl.zip' + extract_to = f'{extract_folder}/{get_extract_name(filename)}' + script_path = f'{extract_to}/convert_czeng16_to_17.pl' + + if not os.path.exists(script_path): + wget.download(url, filename, bar=bar_custom) + extract_to = extract_file(f'{download_to}/convert_czeng16_to_17.pl.zip', extract_folder, get_extract_name=get_extract_name, debug=debug) + return script_path + +######### +# definitions of wmt data sources +# for es-en +# Punctuation in the official test sets will be encoded with ASCII characters (not complex Unicode characters) as much as possible. You may want to normalize your system's output before submission. You are able able to use a rawer version of the test sets that does not have this normalization. +# script to normalize punctuation: http://www.statmt.org/wmt11/normalize-punctuation.perl +wmt13_es_en = DLDataset( + name='wmt13_es-en', + train_urls=[ + 'http://www.statmt.org/wmt13/training-parallel-europarl-v7.tgz', + 'http://www.statmt.org/wmt13/training-parallel-commoncrawl.tgz', + 'http://www.statmt.org/wmt13/training-parallel-un.tgz', + 'http://www.statmt.org/wmt13/training-parallel-nc-v8.tgz', + ], + valid_urls=[ + ('http://www.statmt.org/wmt13/dev.tgz', 'wmt13_dev.tgz') + ], + test_urls=[ + ('http://www.statmt.org/wmt13/test.tgz', 'wmt13_test.tgz') + ], + train_files_patterns=[ + ('*/europarl-v7.{src}-{tgt}.{lang}', ['es-en']), + ('*commoncrawl.{src}-{tgt}.{lang}', ['es-en']), + ('*/news-commentary-v8.{src}-{tgt}.{lang}', ['es-en']), + ('un/*undoc.2000.{src}-{tgt}.{lang}', ['es-en']), + ] , + valid_files_patterns=[ + ('dev/newstest2012.{lang}', ['es-en']) + ], + test_files_patterns=[ + ('test/newstest*.{lang}', ['es-en']) + ], +) + +wmt14_de_fr_en = DLDataset( + name='wmt14_de_fr_en', + train_urls=[ + 'http://www.statmt.org/wmt13/training-parallel-europarl-v7.tgz', + 'http://www.statmt.org/wmt13/training-parallel-commoncrawl.tgz', + 'http://www.statmt.org/wmt13/training-parallel-un.tgz', + 'http://www.statmt.org/wmt14/training-parallel-nc-v9.tgz', + ('http://www.statmt.org/wmt10/training-giga-fren.tar', 'training-giga-fren.gz.tar'), #it is actuall a gz.tar + ], + valid_urls=[ + ('http://www.statmt.org/wmt14/dev.tgz', 'wmt14_dev.tgz'), + ], + test_urls=[ + ('http://www.statmt.org/wmt14/test-full.tgz', 'wmt14_test_full.tgz'), # cleaned test sets + ], + train_files_patterns=[ + ('*/europarl-v7.{src}-{tgt}.{lang}', ['fr-en', 'de-en']), + ('*commoncrawl.{src}-{tgt}.{lang}', ['fr-en', 'de-en']), + ('*/*news-commentary-v9.{src}-{tgt}.{lang}', ['fr-en', 'de-en']), + ('un/undoc.2000.{src}-{tgt}.{lang}', ['fr-en']), + ('*giga-{src}{tgt}*{lang}', ['fr-en']) + ], + valid_files_patterns=[ + ('dev/newstest2013.{lang}', ['fr-en', 'de-en']) + ], + test_files_patterns=[ + ('test-full/newstest*{src}{tgt}-{src:src}{tgt:ref}.{lang}', ['en-de', 'de-en', 'fr-en', 'en-fr']), + ], +) + +# pip install git+https://github.com/amake/tmx2corpus.git +wmt16_ro_en = DLDataset( + name='wmt16_ro-en', + train_urls=[ + ('http://data.statmt.org/wmt16/translation-task/training-parallel-ep-v8.tgz', 'wmt16_training-parallel-ep-v8.tgz'), + ('http://opus.nlpl.eu/download.php?f=SETIMES/v2/tmx/en-ro.tmx.gz', 'en-ro.tmx.gz'), + ], + valid_urls=[ + ('http://data.statmt.org/wmt16/translation-task/dev-romanian-updated.tgz', 'wmt16_dev.tgz') + ], + test_urls=[ + ('http://data.statmt.org/wmt16/translation-task/test.tgz', 'wmt16_test.tgz') + ], + train_files_patterns=[ + ('*/*europarl-v8.{src}-{tgt}.{lang}', ['ro-en']), + ('bitext.{lang}', ['ro-en']) #setimes from tmux + ] , + valid_files_patterns=[ + ('dev/newsdev2016*{src}{tgt}*.{lang}', ['ro-en', 'ro-en']) + ], + test_files_patterns=[ + ('test/newstest*{src}{tgt}*.{lang}', ['ro-en', 'en-ro']) + ], +) + +cwmt_wmt_instruction = 'cwmt download instruction at: http://nlp.nju.edu.cn/cwmt-wmt' +wmt17_fi_lv_tr_zh_en_manual_downloads = [ + # fake urls to have unique keys for the data + ( ('http://nlp.nju.edu.cn/cwmt-wmt/CASIA2015.zip', 'CASIA2015.zip'), cwmt_wmt_instruction), + ( ('http://nlp.nju.edu.cn/cwmt-wmt/CASICT2011.zip', 'CASICT2011.zip'), cwmt_wmt_instruction), + ( ('http://nlp.nju.edu.cn/cwmt-wmt/CASICT2015.zip', 'CASICT2015.zip'), cwmt_wmt_instruction), + ( ('http://nlp.nju.edu.cn/cwmt-wmt/Datum2015.zip', 'Datum2015.zip'), cwmt_wmt_instruction), + ( ('http://nlp.nju.edu.cn/cwmt-wmt/Datum2017.zip', 'Datum2017.zip'), cwmt_wmt_instruction), + ( ('http://nlp.nju.edu.cn/cwmt-wmt/NEU2017.zip', 'NEU2017.zip'), cwmt_wmt_instruction), +] +wmt17_fi_lv_tr_zh_en = DLDataset( + name='wmt17_fi_lv_tr_zh_en', + train_urls=[ + ('http://data.statmt.org/wmt17/translation-task/training-parallel-ep-v8.tgz', 'wmt17_training-parallel-ep-v8.tgz'), + 'http://data.statmt.org/wmt17/translation-task/training-parallel-nc-v12.tgz', + 'http://www.statmt.org/wmt15/wiki-titles.tgz', + ('http://opus.nlpl.eu/download.php?f=SETIMES/v2/tmx/en-tr.tmx.gz', 'en-tr.tmx.gz'), + ('http://data.statmt.org/wmt17/translation-task/rapid2016.tgz', 'wmt17_rapid2016.tgz'), + 'http://data.statmt.org/wmt17/translation-task/leta.v1.tgz', + 'http://data.statmt.org/wmt17/translation-task/dcep.lv-en.v1.tgz', + 'http://data.statmt.org/wmt17/translation-task/books.lv-en.v1.tgz', + (('https://stuncorpusprod.blob.core.windows.net/corpusfiles/UNv1.0.en-zh.tar.gz.00', + 'https://stuncorpusprod.blob.core.windows.net/corpusfiles/UNv1.0.en-zh.tar.gz.01',), 'UNv1.0.en-zh.tar.gz'), + #manually download files: + ('http://nlp.nju.edu.cn/cwmt-wmt/CASIA2015.zip', 'CASIA2015.zip'), + ('http://nlp.nju.edu.cn/cwmt-wmt/CASICT2011.zip', 'CASICT2011.zip'), + ('http://nlp.nju.edu.cn/cwmt-wmt/CASICT2015.zip', 'CASICT2015.zip'), + ('http://nlp.nju.edu.cn/cwmt-wmt/Datum2015.zip', 'Datum2015.zip'), + ('http://nlp.nju.edu.cn/cwmt-wmt/Datum2017.zip', 'Datum2017.zip'), + ('http://nlp.nju.edu.cn/cwmt-wmt/NEU2017.zip', 'NEU2017.zip'), + ], + valid_urls=[ + ('http://data.statmt.org/wmt17/translation-task/dev.tgz', 'wmt17_dev.tgz'), + ], + test_urls=[ + #NEW: Improved translations for zh test sets + ('http://data.statmt.org/wmt17/translation-task/test-update-1.tgz', 'wmt17_test_zh_en.tgz'), + ('http://data.statmt.org/wmt17/translation-task/test.tgz', 'wmt17_test_others.tgz') + ], + train_files_patterns=[ + ('casict*/cas*{src:ch}{tgt:en}.txt', ['zh-en', 'zh-en'] ), + ('casia*/cas*{src:ch}{tgt:en}.txt', ['zh-en', 'zh-en'] ), + ('dataum*/Book*{src:cn}{tgt:en}.txt', ['zh-en', 'zh-en']), + ('neu*/NEU*{src:cn}{tgt:en}.txt', ['zh-en', 'zh-en'] ), + ('*/*UNv1.0.en-zh.{src:zh}{tgt:en}', ['zh-en']), + ('training/*news-commentary-v12.{src}-{tgt}.{lang}', ['zh-en', ]), + + ('*/*europarl-v8.{src}-{tgt}.{lang}', ['fi-en', 'lv-en']), + ('wiki/fi-en/titles.{src}-{tgt}.{lang}', ['fi-en', ]), + ('rapid2016.{tgt}-{src}.{lang}', ['fi-en', 'lv-en']), + ('*/leta.{lang}', ['lv-en']), + ('*/dcep.{lang}', ['lv-en']), + ('*/farewell.{lang}', ['lv-en']), + ('bitext.{lang}', ['tr-en']), + ] , + valid_files_patterns=[ + ('dev/newsdev2017*{src}{tgt}-{src:src}{tgt:ref}.{lang}', + [ + 'fi-en', 'lv-en', 'tr-en', 'zh-en', + 'en-fi', 'en-lv', 'en-tr', 'en-zh' + ]), + ('dev/newstest2016*{src}{tgt}-{src:src}{tgt:ref}.{lang}', + [ + 'fi-en', 'tr-en', + 'en-fi', 'en-tr', + ]), + ], + test_files_patterns=[ + ('test/newstest2017-{src}{tgt}-{src:src}{tgt:ref}.{lang}', + [ + 'fi-en', 'lv-en', 'tr-en', + 'en-fi', 'en-lv', 'en-tr', + ]), + ('newstest2017-{src}{tgt}-{src:src}{tgt:ref}.{lang}', + [ + 'zh-en', + 'en-zh' + ]), + ], +) + +czeng_instruction = 'download instruction at: http://ufal.mff.cuni.cz/czeng/czeng16' +#alternative: use the prepared data but detokenize it? +wmt18_cs_et_en_manual_downloads = [ +#for cs, need to register and download; Register and download CzEng 1.6. +#Better results can be obtained by using a subset of sentences, released under a new version name CzEng 1.7. + # ((f'http://ufallab.ms.mff.cuni.cz/~bojar/czeng16-data/data-plaintext-format.{i}.tar', + # f'data-plaintext-format.{i}.tar'), czeng_instruction) + # for i in range(10) +] + +wmt18_cs_et_en = DLDataset( + name='wmt18_cs_et_en', + train_urls=[ + 'http://www.statmt.org/wmt13/training-parallel-europarl-v7.tgz', + 'http://data.statmt.org/wmt18/translation-task/training-parallel-ep-v8.tgz', + 'https://s3.amazonaws.com/web-language-models/paracrawl/release1/paracrawl-release1.en-cs.zipporah0-dedup-clean.tgz', + 'https://s3.amazonaws.com/web-language-models/paracrawl/release1/paracrawl-release1.en-et.zipporah0-dedup-clean.tgz', + 'http://www.statmt.org/wmt13/training-parallel-commoncrawl.tgz', + 'http://data.statmt.org/wmt18/translation-task/training-parallel-nc-v13.tgz', + ('http://data.statmt.org/wmt18/translation-task/rapid2016.tgz', 'wmt18_rapid2016.tgz'), + # (tuple( + # (f'http://ufallab.ms.mff.cuni.cz/~bojar/czeng16-data/data-plaintext-format.{i}.tar', + # f'data-plaintext-format.{i}.tar') + # for i in range(10) + # ), + # 'czeng16_data_plaintext.gz.tar'), + ], + valid_urls=[ + ('http://data.statmt.org/wmt18/translation-task/dev.tgz', 'wmt18_dev.tgz'), + ], + test_urls=[ + ('http://data.statmt.org/wmt18/translation-task/test.tgz', 'wmt18_test.tgz'), + ], + train_files_patterns=[ + # ('*/*europarl-v7.{src}-{tgt}.{lang}', ['cs-en']), + ('*/*europarl-v8.{src}-{tgt}.{lang}', ['et-en']), + # ('*paracrawl-release1.{tgt}-{src}.zipporah0-dedup-clean.{lang}', ['cs-en', 'et-en']), + ('*paracrawl-release1.{tgt}-{src}.zipporah0-dedup-clean.{lang}', ['et-en']), + # ('*commoncrawl.{src}-{tgt}.{lang}', ['cs-en']), + # ('*/news-commentary-v13.{src}-{tgt}.{lang}', ['cs-en']), + # ('data.plaintext-format/*train.{lang}', ['cs-en']), + ('rapid2016.{tgt}-{src}.{lang}', ['et-en']), + ] , + valid_files_patterns=[ + ('dev/newsdev2018*{src}{tgt}-{src:src}{tgt:ref}.{lang}', ['et-en']), + # ('dev/newstest2017*{src}{tgt}-{src:src}{tgt:ref}.{lang}', ['cs-en']) + ], + test_files_patterns=[ + ('test/newstest2018-{src}{tgt}-{src:src}{tgt:ref}.{lang}', + # ['cs-en', 'et-en']), + ['et-en']), + ] +) + +ru_en_yandex_instruction = 'Yandex Corpus download instruction at: https://translate.yandex.ru/corpus?lang=en' +wmt19_ru_gu_kk_lt_manual_downloads = [ + (('https://translate.yandex.ru/corpus?lang=en', 'wmt19_1mcorpus.zip'), ru_en_yandex_instruction) +] +wmt19_ru_gu_kk_lt = DLDataset( + name='wmt19_ru_gu_kk_lt', + train_urls=[ + 'http://www.statmt.org/europarl/v9/training/europarl-v9.lt-en.tsv.gz', + 'https://s3.amazonaws.com/web-language-models/paracrawl/release3/en-lt.bicleaner07.tmx.gz', + 'https://s3.amazonaws.com/web-language-models/paracrawl/release1/paracrawl-release1.en-ru.zipporah0-dedup-clean.tgz', + 'http://www.statmt.org/wmt13/training-parallel-commoncrawl.tgz', + 'http://data.statmt.org/news-commentary/v14/training/news-commentary-v14-wmt19.en-kk.tsv.gz', + 'http://data.statmt.org/news-commentary/v14/training/news-commentary-v14.en-ru.tsv.gz', + 'http://data.statmt.org/wikititles/v1/wikititles-v1.kk-en.tsv.gz', + 'http://data.statmt.org/wikititles/v1/wikititles-v1.ru-en.tsv.gz', + 'http://data.statmt.org/wikititles/v1/wikititles-v1.kk-en.tsv.gz', + 'http://data.statmt.org/wikititles/v1/wikititles-v1.lt-en.tsv.gz', + 'http://data.statmt.org/wikititles/v1/wikititles-v1.gu-en.tsv.gz', + (('https://stuncorpusprod.blob.core.windows.net/corpusfiles/UNv1.0.en-ru.tar.gz.00', + 'https://stuncorpusprod.blob.core.windows.net/corpusfiles/UNv1.0.en-ru.tar.gz.01', + 'https://stuncorpusprod.blob.core.windows.net/corpusfiles/UNv1.0.en-ru.tar.gz.02',), + 'wmt19_UNv1.0.en-ru.tar.gz'), + 'https://tilde-model.s3-eu-west-1.amazonaws.com/rapid2016.en-lt.tmx.zip', + ('https://translate.yandex.ru/corpus?lang=en', 'wmt19_1mcorpus.zip'), + ], + valid_urls=[ + ('http://data.statmt.org/wmt19/translation-task/dev.tgz', 'wmt19_dev.tgz'), + ], + test_urls=[ + ('http://data.statmt.org/wmt19/translation-task/test.tgz', 'wmt19_test.tgz'), + ], + train_files_patterns=[ + ('*europarl-v9.{src}-{tgt}.tsv.{lang}', ['lt-en']), + #paracrawl + ('*paracrawl-release1.{tgt}-{src}.zipporah0-dedup-clean.{lang}', ['ru-en']), + ('bitext.{lang}', ['lt-en',]), + ('*commoncrawl.{src}-{tgt}.{lang}', ['ru-en',]), + ('*news-commentary-v14-wmt19.{tgt}-{src}.tsv.{lang}', ['kk-en', ]), + ('*news-commentary-v14.{tgt}-{src}.tsv.{lang}', ['ru-en']), + #yandex + ('corpus.{tgt}_{src}.1m.{lang}', ['ru-en']), + ('wikititles_v1_wikititles-v1.{src}-{tgt}.tsv.{lang}', ['ru-en', 'kk-en', 'lt-en', 'gu-en']), + ('*/UNv1.0.{tgt}-{src}.{lang}', ['ru-en']), + #rapid + ('bitext.{lang}', ['lt-en']) + ], + valid_files_patterns=[ + ('dev/newsdev2019*{src}{tgt}-{src:src}{tgt:ref}.{lang}', ['gu-en', 'kk-en', 'lt-en']), + ('dev/newstest2018*{src}{tgt}-{src:src}{tgt:ref}.{lang}', ['ru-en']), + ], + test_files_patterns=[ + ('sgm/newstest2019-{src}{tgt}-{src:src}{tgt:ref}.{lang}', + ['ru-en', 'gu-en', 'kk-en', 'lt-en', 'en-ru', 'en-gu', 'en-kk', 'en-lt']), + ] +) + + +######### + +if __name__ == "__main__": + # speed up the downloads with multiple processing + dl_folder = f'{to_data_path}/downloads' + extract_folder = f'{to_data_path}/extracted' + + urls = [ + url + for dataset in [wmt13_es_en, wmt14_de_fr_en, wmt16_ro_en, wmt18_cs_et_en, wmt19_ru_gu_kk_lt] + for urls in [dataset.train_urls, dataset.valid_urls, dataset.test_urls] + for url in urls + ] + urls = set(urls) + download_multi(dl_folder, extract_folder, urls, num_processes=8, debug=True) + + # check manually downlaods + to_manually_download_urls = ( + wmt17_fi_lv_tr_zh_en_manual_downloads + wmt18_cs_et_en_manual_downloads + wmt19_ru_gu_kk_lt_manual_downloads + ) + to_be_manually_dowloaded = check_need_manual_downalod(dl_folder, to_manually_download_urls) + if len(to_be_manually_dowloaded) > 0: + print('Missing files that need to be downloaded manually; stop the process now.') + exit(-1) + + completed_urls = {} + completed_extraction = {} + def work_on_wmt(directions, wmt_data): + download_and_extract( + to_data_path, + directions, + wmt_data, + to_manually_download_urls=to_manually_download_urls, + completed_urls=completed_urls, completed_extraction=completed_extraction, debug=True) + + work_on_wmt( + ['es_XX-en_XX'], + wmt13_es_en,) + work_on_wmt( + [ + 'fr_XX-en_XX', 'en_XX-fr_XX', + # 'en_XX-de_DE', 'de_DE-en_XX', + ], + wmt14_de_fr_en,) + work_on_wmt( + ['ro_RO-en_XX', 'en_XX-ro_XX'], + wmt16_ro_en,) + work_on_wmt( + [ + # 'zh_CN-en_XX', + 'lv_LV-en_XX', 'fi_FI-en_XX', 'tr_TR-en_XX', + #in case the reversed directions have different train/valid/test data + # 'en_XX-zh_CN', + 'en_XX-lv_LV', 'en_XX-fi_FI', 'en_XX-tr_TR', + ], + wmt17_fi_lv_tr_zh_en, ) + # czeng17_script_path = download_czeng17_script(download_to, extract_to, debug=False) + # cz_username = None + work_on_wmt( + [ + # 'cs_CZ-en_XX', + 'et_EE-en_XX'], + wmt18_cs_et_en,) + work_on_wmt( + [ + # 'ru_RU-en_XX', 'en_XX-ru_RU', + 'gu_IN-en_XX', 'kk_KZ-en_XX', 'lt_LT-en_XX', + #in case the reversed directions have different train/valid/test data + 'en_XX-gu_IN', 'en_XX-kk_KZ', 'en_XX-lt_LT' + ], + wmt19_ru_gu_kk_lt,) + + not_matching = check_wmt_test_bleu( + f'{to_data_path}/raw', + [ + ('wmt13', ['es_XX-en_XX']), + ('wmt14/full', ['fr_XX-en_XX',]), + ('wmt16', ['ro_RO-en_XX',]), + # ('wmt17/improved', ['zh_CN-en_XX']), + ('wmt17', [ 'lv_LV-en_XX', 'fi_FI-en_XX', 'tr_TR-en_XX']), + ('wmt18', ['cs_CZ-en_XX', 'et_EE-en_XX']), + ('wmt19', ['gu_IN-en_XX', 'kk_KZ-en_XX', 'lt_LT-en_XX']), + #'ru_RU-en_XX', + ] + ) + if len(not_matching) > 0: + print('the following datasets do not have matching test datasets:\n\t', '\n\t'.join(not_matching)) + diff --git a/examples/multilingual/data_scripts/download_wmt20.sh b/examples/multilingual/data_scripts/download_wmt20.sh new file mode 100644 index 0000000000000000000000000000000000000000..31cd5c76b75081331ae03c5ea70ea7ddebaa06e1 --- /dev/null +++ b/examples/multilingual/data_scripts/download_wmt20.sh @@ -0,0 +1,547 @@ +#!/bin/bash +# Copyright (c) Facebook, Inc. and its affiliates. +# All rights reserved. +# +# This source code is licensed under the license found in the +# LICENSE file in the root directory of this source tree. + +if [ -z $WORKDIR_ROOT ] ; +then + echo "please specify your working directory root in environment variable WORKDIR_ROOT. Exitting..." + exit +fi + + + +set -x -e + +# TODO update the workdir and dest dir name +# put fasttext model +WORKDIR=$WORKDIR_ROOT +# put intermediate files +TMP_DIR=$WORKDIR_ROOT/tmp/tmp_wmt20_lowres_download +# output {train,valid,test} files to dest +DEST=$WORKDIR_ROOT/ML50/raw + +UTILS=$PWD/utils + +# per dataset locations +COMMONCRAWL_DIR=$TMP_DIR/commoncrawl +YANDEX_CORPUS=$WORKDIR_ROOT/wmt20/official/ru/yandex/1mcorpus.zip +# unzipped +CZENG_CORPUS=$WORKDIR_ROOT/wmt20/official/cs/czeng/czeng20-train +CCMT_DIR=$WORKDIR_ROOT/wmt20/official/zh/ccmt/parallel + +download_and_select() { + SUBFOLDER=$1 + URL=$2 + UNCOMPRESS_CMD=$3 + LANG=$4 + INPUT_FILEPATH=$5 + if [[ $# -gt 5 ]]; then + LANG_COL=$6 + EN_COL=$7 + fi + + mkdir -p $SUBFOLDER + cd $SUBFOLDER + wget -nc --content-disposition $URL + $UNCOMPRESS_CMD + + if [[ $# -gt 5 ]]; then + cut -f$LANG_COL $INPUT_FILEPATH > $INPUT_FILEPATH.$LANG + cut -f$EN_COL $INPUT_FILEPATH > $INPUT_FILEPATH.en + fi + cd .. + + ln -sf $SUBFOLDER/$INPUT_FILEPATH.$LANG $SUBFOLDER.$LANG + ln -sf $SUBFOLDER/$INPUT_FILEPATH.en $SUBFOLDER.en +} + +prepare_lid() { + pip install fasttext + + # TODO specify global workdir + MODEL=$WORKDIR/fasttext/lid.176.bin + LID_MULTI=$UTILS/fasttext_multi_filter.py + + if [ ! -f "$MODEL" ]; then + echo "downloading fasttext lid model..." + mkdir -p $WORKDIR/fasttext + wget -nc https://dl.fbaipublicfiles.com/fasttext/supervised-models/lid.176.bin -O $MODEL + fi +} + +prepare_moses() { + pushd $UTILS + echo 'Cloning Moses github repository (for tokenization scripts)...' + git clone https://github.com/moses-smt/mosesdecoder.git + popd +} + +lid_filter() { + # TODO specify global workdir + MODEL=$WORKDIR/fasttext/lid.176.bin + LID_MULTI=$UTILS/fasttext_multi_filter.py + + prepare_lid + + SRC=$1 + SRC_FILE=$2 + SRC_OUTPUT=$3 + TGT=$4 + TGT_FILE=$5 + TGT_OUTPUT=$6 + python $LID_MULTI --model $MODEL --inputs $SRC_FILE $TGT_FILE --langs $SRC $TGT --outputs $SRC_OUTPUT $TGT_OUTPUT +} + +prepare_ja_ted() { + mkdir -p ted + cd ted + + wget -nc https://wit3.fbk.eu/archive/2017-01-trnted//texts/en/ja/en-ja.tgz + tar -zxvf en-ja.tgz + cat en-ja/train.tags.en-ja.en | grep -v -P "^[ ]*\<" | sed 's/^[ \t]*//g' | sed 's/[ \t]*$//g' > en-ja/train.en-ja.en + cat en-ja/train.tags.en-ja.ja | grep -v -P "^[ ]*\<" | sed 's/^[ \t]*//g' | sed 's/[ \t]*$//g' > en-ja/train.en-ja.ja + + cd .. + ln -sf ted/en-ja/train.en-ja.ja ted.ja + ln -sf ted/en-ja/train.en-ja.en ted.en +} + +prepare_ja() { + OUTPUT_DIR=$TMP_DIR/ja + mkdir -p $OUTPUT_DIR + cd $OUTPUT_DIR + + download_and_select paracrawl "http://www.kecl.ntt.co.jp/icl/lirg/jparacrawl/release/2.0/bitext/en-ja.tar.gz" "tar -zxvf en-ja.tar.gz" ja en-ja/en-ja.bicleaner05.txt 4 3 & + download_and_select newscommentary "http://data.statmt.org/news-commentary/v15/training/news-commentary-v15.en-ja.tsv.gz" "gunzip -f news-commentary-v15.en-ja.tsv.gz" ja news-commentary-v15.en-ja.tsv 2 1 & + download_and_select wikititles "http://data.statmt.org/wikititles/v2/wikititles-v2.ja-en.tsv.gz" "gunzip -f wikititles-v2.ja-en.tsv.gz" ja wikititles-v2.ja-en.tsv 1 2 & + download_and_select wikimatrix "http://data.statmt.org/wmt20/translation-task/WikiMatrix/WikiMatrix.v1.en-ja.langid.tsv.gz" "gunzip -f WikiMatrix.v1.en-ja.langid.tsv.gz" ja WikiMatrix.v1.en-ja.langid.tsv 3 2 & + download_and_select subtitle "https://nlp.stanford.edu/projects/jesc/data/split.tar.gz" "tar -zxvf split.tar.gz" ja split/train 2 1 & + download_and_select kftt "http://www.phontron.com/kftt/download/kftt-data-1.0.tar.gz" "tar -zxvf kftt-data-1.0.tar.gz" ja kftt-data-1.0/data/orig/kyoto-train & + + prepare_ja_ted & + + # ted data needs to + + wait + + # remove previous results + rm -f all.?? + find ./ -maxdepth 1 -name "*.ja" | sort -V | xargs cat > all.ja + find ./ -maxdepth 1 -name "*.en" | sort -V | xargs cat > all.en + lid_filter ja all.ja $DEST/train.ja_XX-en_XX.ja_XX en all.en $DEST/train.ja_XX-en_XX.en_XX +} + +prepare_ta() { + OUTPUT_DIR=$TMP_DIR/ta + mkdir -p $OUTPUT_DIR + cd $OUTPUT_DIR + + download_and_select wikititles "http://data.statmt.org/wikititles/v2/wikititles-v2.ta-en.tsv.gz" "gunzip -f wikititles-v2.ta-en.tsv.gz" ta wikititles-v2.ta-en.tsv 1 2 & + download_and_select wikimatrix "http://data.statmt.org/wmt20/translation-task/WikiMatrix/WikiMatrix.v1.en-ta.langid.tsv.gz" "gunzip -f WikiMatrix.v1.en-ta.langid.tsv.gz" ta WikiMatrix.v1.en-ta.langid.tsv 3 2 & + download_and_select pmindia "http://data.statmt.org/pmindia/v1/parallel/pmindia.v1.ta-en.tsv" "" ta pmindia.v1.ta-en.tsv 2 1 & + download_and_select tanzil "https://object.pouta.csc.fi/OPUS-Tanzil/v1/moses/en-ta.txt.zip" "unzip en-ta.txt.zip" ta Tanzil.en-ta & + download_and_select pib "http://preon.iiit.ac.in/~jerin/resources/datasets/pib-v0.tar" "tar -xvf pib-v0.tar" ta pib/en-ta/train & + download_and_select mkb "http://preon.iiit.ac.in/~jerin/resources/datasets/mkb-v0.tar" "tar -xvf mkb-v0.tar" ta mkb/en-ta/mkb & + download_and_select ufal "http://ufal.mff.cuni.cz/~ramasamy/parallel/data/v2/en-ta-parallel-v2.tar.gz" "tar -zxvf en-ta-parallel-v2.tar.gz" ta en-ta-parallel-v2/corpus.bcn.train & + + wait + + # need special handling for nlpc + mkdir -p nlpc + cd nlpc + wget -nc https://raw.githubusercontent.com/nlpc-uom/English-Tamil-Parallel-Corpus/master/En-Ta%20Corpus/En-Ta%20English.txt + wget -nc https://github.com/nlpc-uom/English-Tamil-Parallel-Corpus/raw/master/En-Ta%20Corpus/En-Ta%20Tamil.txt + tail -n +4 "En-Ta English.txt" > en-ta.en + tail -n +4 "En-Ta Tamil.txt" > en-ta.ta + cd .. + ln -sf nlpc/en-ta.en nlpc.en + ln -sf nlpc/en-ta.ta nlpc.ta + + # remove previous results + rm -f all.?? + find ./ -maxdepth 1 -name "*.ta" | sort -V | xargs cat > all.ta + find ./ -maxdepth 1 -name "*.en" | sort -V | xargs cat > all.en + lid_filter ta all.ta $DEST/train.ta_IN-en_XX.ta_IN en all.en $DEST/train.ta_IN-en_XX.en_XX +} + +prepare_iu() { + OUTPUT_DIR=$TMP_DIR/iu + mkdir -p $OUTPUT_DIR + cd $OUTPUT_DIR + + download_and_select nh "https://nrc-digital-repository.canada.ca/eng/view/dataset/?id=c7e34fa7-7629-43c2-bd6d-19b32bf64f60" "tar -zxvf Nunavut-Hansard-Inuktitut-English-Parallel-Corpus-3.0.1.tgz" iu Nunavut-Hansard-Inuktitut-English-Parallel-Corpus-3.0/NunavutHansard > /dev/null & + download_and_select wikititles "http://data.statmt.org/wikititles/v2/wikititles-v2.iu-en.tsv.gz" "gunzip -f wikititles-v2.iu-en.tsv.gz" iu wikititles-v2.iu-en.tsv 1 2 & + + wait + + # remove previous results + rm -f all.?? + find ./ -maxdepth 1 -name "*.iu" | sort -V | xargs cat | nh/Nunavut-Hansard-Inuktitut-English-Parallel-Corpus-3.0/scripts/normalize-iu-spelling.pl > all.iu + find ./ -maxdepth 1 -name "*.en" | sort -V | xargs cat > all.en + paste all.iu all.en | awk -F $'\t' '$1!=""&&$2!=""' > all.iuen + cut -f1 all.iuen > $DEST/train.iu_CA-en_XX.iu_CA + cut -f2 all.iuen > $DEST/train.iu_CA-en_XX.en_XX +} + +prepare_km() { + OUTPUT_DIR=$TMP_DIR/km + mkdir -p $OUTPUT_DIR + cd $OUTPUT_DIR + + download_and_select paracrawl "http://data.statmt.org/wmt20/translation-task/ps-km/wmt20-sent.en-km.xz" "unxz wmt20-sent.en-km.zx" km wmt20-sent.en-km 2 1 & + + # km-parallel has multiple sets, concat all of them together + mkdir -p opus + cd opus + wget -nc "http://data.statmt.org/wmt20/translation-task/ps-km/km-parallel.tgz" + tar -zxvf km-parallel.tgz + find ./km-parallel -maxdepth 1 -name "*.km" | sort -V | xargs cat > opus.km + find ./km-parallel -maxdepth 1 -name "*.en" | sort -V | xargs cat > opus.en + cd .. + ln -sf opus/opus.km . + ln -sf opus/opus.en . + + wait + + # remove previous results + rm -f all.?? + find ./ -maxdepth 1 -name "*.km" | sort -V | xargs cat > all.km + find ./ -maxdepth 1 -name "*.en" | sort -V | xargs cat > all.en + lid_filter km all.km $DEST/train.km_KH-en_XX.km_KH en all.en $DEST/train.km_KH-en_XX.en_XX +} + +prepare_ps() { + OUTPUT_DIR=$TMP_DIR/ps + mkdir -p $OUTPUT_DIR + cd $OUTPUT_DIR + + download_and_select paracrawl "http://data.statmt.org/wmt20/translation-task/ps-km/wmt20-sent.en-ps.xz" "unxz wmt20-sent.en-ps.xz" ps wmt20-sent.en-ps 2 1 & + download_and_select wikititles "http://data.statmt.org/wikititles/v2/wikititles-v2.ps-en.tsv.gz" "gunzip -f wikititles-v2.ps-en.tsv.gz" ps wikititles-v2.ps-en.tsv 1 2 & + # ps-parallel has multiple sets, concat all of them together + mkdir -p opus + cd opus + wget -nc "http://data.statmt.org/wmt20/translation-task/ps-km/ps-parallel.tgz" + tar -zxvf ps-parallel.tgz + find ./ps-parallel -maxdepth 1 -name "*.ps" | sort -V | xargs cat > opus.ps + find ./ps-parallel -maxdepth 1 -name "*.en" | sort -V | xargs cat > opus.en + cd .. + ln -sf opus/opus.ps opus.ps + ln -sf opus/opus.en opus.en + + wait + + # remove previous results + rm -f all.?? + find ./ -maxdepth 1 -name "*.ps" | sort -V | xargs cat > all.ps + find ./ -maxdepth 1 -name "*.en" | sort -V | xargs cat > all.en + lid_filter ps all.ps $DEST/train.ps_AF-en_XX.ps_AF en all.en $DEST/train.ps_AF-en_XX.en_XX +} + +download_commoncrawl() { + mkdir -p $COMMONCRAWL_DIR + cd $COMMONCRAWL_DIR + + wget -nc "http://www.statmt.org/wmt13/training-parallel-commoncrawl.tgz" + tar -zxvf training-parallel-commoncrawl.tgz +} +link_commoncrawl() { + LANG=$1 + ln -sf $COMMONCRAWL_DIR/commoncrawl.$LANG-en.en commoncrawl.en + ln -sf $COMMONCRAWL_DIR/commoncrawl.$LANG-en.$LANG commoncrawl.$LANG +} + +strip_xlf() { + INPUT_FILE=$1 + SRC=$2 + TGT=$3 + grep '<source xml:lang=' $INPUT_FILE | sed 's/^<[^<>]*>//g' | sed 's/<[^<>]*>$//g' > $INPUT_FILE.$SRC + grep '<target xml:lang=' $INPUT_FILE | sed 's/^<[^<>]*>//g' | sed 's/<[^<>]*>$//g' > $INPUT_FILE.$TGT +} + +download_and_process_tilde() { + URL=$1 + UNCOMPRESS_CMD=$2 + FILENAME=$3 + LANG=$4 + PROCESS_CMD=$5 + + mkdir -p tilde + cd tilde + wget -nc $URL + $UNCOMPRESS_CMD + echo "executing cmd" + echo $PROCESS_CMD + $PROCESS_CMD + cd .. + ln -sf tilde/$FILENAME.$LANG tilde.$LANG + ln -sf tilde/$FILENAME.en tilde.en +} + +prepare_cs() { + OUTPUT_DIR=$TMP_DIR/cs + mkdir -p $OUTPUT_DIR + cd $OUTPUT_DIR + + #download_and_select europarl "http://www.statmt.org/europarl/v10/training/europarl-v10.cs-en.tsv.gz" "gunzip europarl-v10.cs-en.tsv.gz" cs europarl-v10.cs-en.tsv 1 2 & + #download_and_select paracrawl "https://s3.amazonaws.com/web-language-models/paracrawl/release5.1/en-cs.txt.gz" "gunzip en-cs.txt.gz" cs en-cs.txt 2 1 & + #link_commoncrawl cs + #download_and_select newscommentary "http://data.statmt.org/news-commentary/v15/training/news-commentary-v15.cs-en.tsv.gz" "gunzip news-commentary-v15.cs-en.tsv.gz" cs news-commentary-v15.cs-en.tsv 1 2 & + #download_and_select wikititles "http://data.statmt.org/wikititles/v2/wikititles-v2.cs-en.tsv.gz" "gunzip wikititles-v2.cs-en.tsv.gz" cs wikititles-v2.cs-en.tsv 1 2 & + #download_and_process_tilde "http://data.statmt.org/wmt20/translation-task/rapid/RAPID_2019.cs-en.xlf.gz" "gunzip RAPID_2019.cs-en.xlf.gz" RAPID_2019.cs-en.xlf cs "strip_xlf RAPID_2019.cs-en.xlf cs en" & + #download_and_select wikimatrix "http://data.statmt.org/wmt20/translation-task/WikiMatrix/WikiMatrix.v1.cs-en.langid.tsv.gz" "gunzip WikiMatrix.v1.cs-en.langid.tsv.gz" cs WikiMatrix.v1.cs-en.langid.tsv 2 3 & + + #wait + + # remove previous results + #rm -f all.?? + #find ./ -maxdepth 1 -name "*.cs" | sort -V | xargs cat > all.cs + #find ./ -maxdepth 1 -name "*.en" | sort -V | xargs cat > all.en + if [ -z $CZENG_CORPUS ] ; + then + echo "Please download CZENG_CORPUS manually and place them at $CZENG_CORPUS. Exitting..." + exit + fi + cat $CZENG_CORPUS | sed '/^$/d' | cut -f5 > all.cs + cat $CZENG_CORPUS | sed '/^$/d' | cut -f6 > all.en + + lid_filter cs all.cs $DEST/train.cs_CZ-en_XX.cs_CZ en all.en $DEST/train.cs_CZ-en_XX.en_XX +} + +prepare_de() { + OUTPUT_DIR=$TMP_DIR/de + mkdir -p $OUTPUT_DIR + cd $OUTPUT_DIR + + download_and_select europarl "http://www.statmt.org/europarl/v10/training/europarl-v10.de-en.tsv.gz" "gunzip europarl-v10.de-en.tsv.gz" de europarl-v10.de-en.tsv 1 2 & + download_and_select paracrawl "https://s3.amazonaws.com/web-language-models/paracrawl/release5.1/en-de.txt.gz" "gunzip en-de.txt.gz" de en-de.txt 2 1 & + link_commoncrawl de + download_and_select newscommentary "http://data.statmt.org/news-commentary/v15/training/news-commentary-v15.de-en.tsv.gz" "gunzip news-commentary-v15.de-en.tsv.gz" de news-commentary-v15.de-en.tsv 1 2 & + download_and_select wikititles "http://data.statmt.org/wikititles/v2/wikititles-v2.de-en.tsv.gz" "gunzip wikititles-v2.de-en.tsv.gz" de wikititles-v2.de-en.tsv 1 2 & + download_and_process_tilde "http://data.statmt.org/wmt20/translation-task/rapid/RAPID_2019.de-en.xlf.gz" "gunzip RAPID_2019.de-en.xlf.gz" RAPID_2019.de-en.xlf de "strip_xlf RAPID_2019.de-en.xlf de en" & + download_and_select wikimatrix "http://data.statmt.org/wmt20/translation-task/WikiMatrix/WikiMatrix.v1.de-en.langid.tsv.gz" "gunzip WikiMatrix.v1.de-en.langid.tsv.gz" de WikiMatrix.v1.de-en.langid.tsv 2 3 & + + wait + + # remove previous results + rm -f all.?? + find ./ -maxdepth 1 -name "*.de" | sort -V | xargs cat > all.de + find ./ -maxdepth 1 -name "*.en" | sort -V | xargs cat > all.en + lid_filter de all.de $DEST/train.de_DE-en_XX.de_DE en all.en $DEST/train.de_DE-en_XX.en_XX +} + +prepare_tmx() { + TMX_FILE=$1 + git clone https://github.com/amake/TMX2Corpus $UTILS/tmx2corpus + pip install tinysegmenter + + python $UTILS/tmx2corpus/tmx2corpus.py $TMX_FILE +} + +prepare_pl() { + OUTPUT_DIR=$TMP_DIR/pl + mkdir -p $OUTPUT_DIR + cd $OUTPUT_DIR + + # download_and_select europarl "http://www.statmt.org/europarl/v10/training/europarl-v10.pl-en.tsv.gz" "gunzip europarl-v10.pl-en.tsv.gz" pl europarl-v10.pl-en.tsv 1 2 & + # download_and_select paracrawl "https://s3.amazonaws.com/web-language-models/paracrawl/release5.1/en-pl.txt.gz" "gunzip en-pl.txt.gz" pl en-pl.txt 2 1 & + # download_and_select wikititles "http://data.statmt.org/wikititles/v2/wikititles-v2.pl-en.tsv.gz" "gunzip wikititles-v2.pl-en.tsv.gz" pl wikititles-v2.pl-en.tsv 1 2 & + download_and_select tilde "https://tilde-model.s3-eu-west-1.amazonaws.com/rapid2019.en-pl.tmx.zip" "gunzip rapid2019.en-pl.tmx.zip" bitext pl "prepare_tmx RAPID_2019.UNIQUE.en-pl.tmx" & + # download_and_select wikimatrix "http://data.statmt.org/wmt20/translation-task/WikiMatrix/WikiMatrix.v1.en-pl.langid.tsv.gz" "gunzip WikiMatrix.v1.en-pl.langid.tsv.gz" pl WikiMatrix.v1.en-pl.langid.tsv 3 2 & + + wait + + # remove previous results + rm -f all.?? + find ./ -maxdepth 1 -name "*.pl" | sort -V | xargs cat > all.pl + find ./ -maxdepth 1 -name "*.en" | sort -V | xargs cat > all.en + lid_filter pl all.pl $DEST/train.pl_PL-en_XX.pl_PL en all.en $DEST/train.pl_PL-en_XX.en_XX +} + +prepare_uncorpus() { + $URLS=$1 + $FILES=$2 + + mkdir -p uncorpus + cd uncorpus + + for URL in $URLS; do + wget -nc $URL + done + cat $FILES > uncorpus.tar.gz + tar -zxvf uncorpus.tar.gz + + cd .. + ln -sf uncorpus/en-$LANG/UNv1.0.en-$LANG.$LANG uncorpus.$LANG + ln -sf uncorpus/en-$LANG/UNv1.0.en-$LANG.en uncorpus.en +} + +prepare_yandex() { + mkdir -p yandex + cd yandex + unzip $YANDEX_CORPUS ./ + cd .. + ln -s yandex/corpus.en_ru.1m.en yandex.en + ln -s yandex/corpus.en_ru.1m.ru yandex.ru +} + +prepare_ru() { + OUTPUT_DIR=$TMP_DIR/ru + mkdir -p $OUTPUT_DIR + cd $OUTPUT_DIR + + download_and_select paracrawl "https://s3.amazonaws.com/web-language-models/paracrawl/release1/paracrawl-release1.en-ru.zipporah0-dedup-clean.tgz" "tar -zxvf paracrawl-release1.en-ru.zipporah0-dedup-clean.tgz" ru paracrawl-release1.en-ru.zipporah0-dedup-clean & + link_commoncrawl ru + download_and_select newscommentary "http://data.statmt.org/news-commentary/v15/training/news-commentary-v15.en-ru.tsv.gz" "gunzip news-commentary-v15.en-ru.tsv.gz" ru news-commentary-v15.en-ru.tsv 2 1 & + prepare_yandex & + download_and_select wikititles "http://data.statmt.org/wikititles/v2/wikititles-v2.ru-en.tsv.gz" "gunzip wikititles-v2.ru-en.tsv.gz" ru wikititles-v2.ru-en.tsv 1 2 & + prepare_uncorpus "https://stuncorpusprod.blob.core.windows.net/corpusfiles/UNv1.0.en-ru.tar.gz.00 https://stuncorpusprod.blob.core.windows.net/corpusfiles/UNv1.0.en-ru.tar.gz.01 https://stuncorpusprod.blob.core.windows.net/corpusfiles/UNv1.0.en-ru.tar.gz.02" "UNv1.0.en-ru.tar.gz.00 UNv1.0.en-ru.tar.gz.01 UNv1.0.en-ru.tar.gz.02" & + download_and_select wikimatrix "http://data.statmt.org/wmt20/translation-task/WikiMatrix/WikiMatrix.v1.en-ru.langid.tsv.gz" "gunzip WikiMatrix.v1.en-ru.langid.tsv.gz" ru WikiMatrix.v1.en-ru.langid.tsv 3 2 & + + wait + + # remove previous results + rm -f all.?? + find ./ -maxdepth 1 -name "*.ru" | sort -V | xargs cat > all.ru + find ./ -maxdepth 1 -name "*.en" | sort -V | xargs cat > all.en + lid_filter ru all.ru $DEST/train.ru_RU-en_XX.ru_RU en all.en $DEST/train.ru_RU-en_XX.en_XX +} + +prepare_ccmt() { + mkdir -p ccmt + cd ccmt + # assume ccmt data is already unzipped under CCMT_DIR folder + cat $CCMT_DIR/datum2017/Book*_cn.txt | sed 's/ //g' > datum2017.detok.zh + cat $CCMT_DIR/datum2017/Book*_en.txt > datum2017.detok.en + cat $CCMT_DIR/casict2011/casict-A_ch.txt $CCMT_DIR/casict2011/casict-B_ch.txt $CCMT_DIR/casict2015/casict2015_ch.txt $CCMT_DIR/datum2015/datum_ch.txt $CCMT_DIR/neu2017/NEU_cn.txt datum2017.detok.zh > ccmt.zh + cat $CCMT_DIR/casict2011/casict-A_en.txt $CCMT_DIR/casict2011/casict-B_en.txt $CCMT_DIR/casict2015/casict2015_en.txt $CCMT_DIR/datum2015/datum_en.txt $CCMT_DIR/neu2017/NEU_en.txt datum2017.detok.en > ccmt.en + cd .. + ln -sf ccmt/ccmt.zh ccmt.zh + ln -sf ccmt/ccmt.en ccmt.en +} + +prepare_zh() { + OUTPUT_DIR=$TMP_DIR/zh + mkdir -p $OUTPUT_DIR + cd $OUTPUT_DIR + + download_and_select newscommentary "http://data.statmt.org/news-commentary/v15/training/news-commentary-v15.en-zh.tsv.gz" "gunzip news-commentary-v15.en-zh.tsv.gz" zh news-commentary-v15.en-zh.tsv 2 1 & + download_and_select wikititles "http://data.statmt.org/wikititles/v2/wikititles-v2.zh-en.tsv.gz" "gunzip wikititles-v2.zh-en.tsv.gz" zh wikititles-v2.zh-en.tsv 1 2 & + prepare_uncorpus "https://stuncorpusprod.blob.core.windows.net/corpusfiles/UNv1.0.en-zh.tar.gz.00 https://stuncorpusprod.blob.core.windows.net/corpusfiles/UNv1.0.en-zh.tar.gz.01" "UNv1.0.en-zh.tar.gz.00 UNv1.0.en-zh.tar.gz.01" & + prepare_ccmt & + download_and_select wikimatrix "http://data.statmt.org/wmt20/translation-task/WikiMatrix/WikiMatrix.v1.en-zh.langid.tsv.gz" "gunzip WikiMatrix.v1.en-zh.langid.tsv.gz" zh WikiMatrix.v1.en-zh.langid.tsv 3 2 & + + wait + + # remove previous results + rm -f all.?? + find ./ -maxdepth 1 -name "*.zh" | sort -V | xargs cat > all.zh + find ./ -maxdepth 1 -name "*.en" | sort -V | xargs cat > all.en + lid_filter zh all.zh $DEST/train.zh_CN-en_XX.zh_CN en all.en $DEST/train.zh_CN-en_XX.en_XX +} + +prepare_tests() { + OUTPUT_DIR=$TMP_DIR + mkdir -p $OUTPUT_DIR + cd $OUTPUT_DIR + wget -nc http://data.statmt.org/wmt20/translation-task/dev.tgz + tar -zxvf dev.tgz + cd dev + + cat newsdev2020-jaen-src.ja.sgm | $UTILS/strip_sgm.sh > newsdev2020-jaen.ja + cat newsdev2020-jaen-ref.en.sgm | $UTILS/strip_sgm.sh > newsdev2020-jaen.en + split newsdev2020-jaen.ja -a 0 -n r/1/2 > $DEST/valid.ja_XX-en_XX.ja_XX + split newsdev2020-jaen.en -a 0 -n r/1/2 > $DEST/valid.ja_XX-en_XX.en_XX + split newsdev2020-jaen.ja -a 0 -n r/2/2 > $DEST/test.ja_XX-en_XX.ja_XX + split newsdev2020-jaen.en -a 0 -n r/2/2 > $DEST/test.ja_XX-en_XX.en_XX + + cat newsdev2020-iuen-src.iu.sgm | strip_sgm.sh > newsdev2020-iuen.iu + cat newsdev2020-iuen-ref.en.sgm | strip_sgm.sh > newsdev2020-iuen.en + split newsdev2020-iuen.iu -a 0 -n r/1/2 > $DEST/valid.iu_CA-en_XX.iu_CA + split newsdev2020-iuen.en -a 0 -n r/1/2 > $DEST/valid.iu_CA-en_XX.en_XX + split newsdev2020-iuen.iu -a 0 -n r/2/2 > $DEST/test.iu_CA-en_XX.iu_CA + split newsdev2020-iuen.en -a 0 -n r/2/2 > $DEST/test.iu_CA-en_XX.en_XX + + cat newsdev2020-taen-src.ta.sgm | strip_sgm.sh > newsdev2020-taen.ta + cat newsdev2020-taen-ref.en.sgm | strip_sgm.sh > newsdev2020-taen.en + split newsdev2020-taen.ta -a 0 -n r/1/2 > $DEST/valid.ta_IN-en_XX.ta_IN + split newsdev2020-taen.en -a 0 -n r/1/2 > $DEST/valid.ta_IN-en_XX.en_XX + split newsdev2020-taen.ta -a 0 -n r/2/2 > $DEST/test.ta_IN-en_XX.ta_IN + split newsdev2020-taen.en -a 0 -n r/2/2 > $DEST/test.ta_IN-en_XX.en_XX + + cp wikipedia.dev.km-en.km $DEST/valid.km_KH-en_XX.km_KH + cp wikipedia.dev.km-en.en $DEST/valid.km_KH-en_XX.en_XX + cp wikipedia.devtest.km-en.km $DEST/test.km_KH-en_XX.km_KH + cp wikipedia.devtest.km-en.en $DEST/test.km_KH-en_XX.en_XX + + cp wikipedia.dev.ps-en.ps $DEST/valid.ps_AF-en_XX.ps_AF + cp wikipedia.dev.ps-en.en $DEST/valid.ps_AF-en_XX.en_XX + cp wikipedia.devtest.ps-en.ps $DEST/test.ps_AF-en_XX.ps_AF + cp wikipedia.devtest.ps-en.en $DEST/test.ps_AF-en_XX.en_XX + + cat newsdev2020-plen-src.pl.sgm | strip_sgm.sh > newsdev2020-plen.pl + cat newsdev2020-plen-ref.en.sgm | strip_sgm.sh > newsdev2020-plen.en + split newsdev2020-plen.pl -a 0 -n r/1/2 > $DEST/valid.pl_PL-en_XX.pl_PL + split newsdev2020-plen.en -a 0 -n r/1/2 > $DEST/valid.pl_PL-en_XX.en_XX + split newsdev2020-plen.pl -a 0 -n r/2/2 > $DEST/test.pl_PL-en_XX.pl_PL + split newsdev2020-plen.en -a 0 -n r/2/2 > $DEST/test.pl_PL-en_XX.en_XX + + cat newstest2018-encs-src.en.sgm | strip_sgm.sh > $DEST/valid.en_XX-cs_CZ.en_XX + cat newstest2018-encs-ref.cs.sgm | strip_sgm.sh > $DEST/valid.en_XX-cs_CZ.cs_CZ + cat newstest2019-encs-src.en.sgm | strip_sgm.sh > $DEST/test.en_XX-cs_CZ.en_XX + cat newstest2019-encs-ref.cs.sgm | strip_sgm.sh > $DEST/test.en_XX-cs_CZ.cs_CZ + + cat newstest2018-deen-src.de.sgm | strip_sgm.sh > $DEST/valid.de_DE-en_XX.de_DE + cat newstest2018-deen-ref.en.sgm | strip_sgm.sh > $DEST/valid.de_DE-en_XX.en_XX + cat newstest2018-ende-src.en.sgm | strip_sgm.sh > $DEST/valid.en_XX-de_DE.en_XX + cat newstest2018-ende-ref.de.sgm | strip_sgm.sh > $DEST/valid.en_XX-de_DE.de_DE + cat newstest2019-deen-src.de.sgm | strip_sgm.sh > $DEST/test.de_DE-en_XX.de_DE + cat newstest2019-deen-ref.en.sgm | strip_sgm.sh > $DEST/test.de_DE-en_XX.en_XX + cat newstest2019-ende-src.en.sgm | strip_sgm.sh > $DEST/test.en_XX-de_DE.en_XX + cat newstest2019-ende-ref.de.sgm | strip_sgm.sh > $DEST/test.en_XX-de_DE.de_DE + + cat newstest2018-ruen-src.ru.sgm | strip_sgm.sh > $DEST/valid.ru_RU-en_XX.ru_RU + cat newstest2018-ruen-ref.en.sgm | strip_sgm.sh > $DEST/valid.ru_RU-en_XX.en_XX + cat newstest2018-enru-src.en.sgm | strip_sgm.sh > $DEST/valid.en_XX-ru_RU.en_XX + cat newstest2018-enru-ref.ru.sgm | strip_sgm.sh > $DEST/valid.en_XX-ru_RU.ru_RU + cat newstest2019-ruen-src.ru.sgm | strip_sgm.sh > $DEST/test.ru_RU-en_XX.ru_RU + cat newstest2019-ruen-ref.en.sgm | strip_sgm.sh > $DEST/test.ru_RU-en_XX.en_XX + cat newstest2019-enru-src.en.sgm | strip_sgm.sh > $DEST/test.en_XX-ru_RU.en_XX + cat newstest2019-enru-ref.ru.sgm | strip_sgm.sh > $DEST/test.en_XX-ru_RU.ru_RU + + cat newstest2018-zhen-src.zh.sgm | strip_sgm.sh > $DEST/valid.zh_CN-en_XX.zh_CN + cat newstest2018-zhen-ref.en.sgm | strip_sgm.sh > $DEST/valid.zh_CN-en_XX.en_XX + cat newstest2018-enzh-src.en.sgm | strip_sgm.sh > $DEST/valid.en_XX-zh_CN.en_XX + cat newstest2018-enzh-ref.zh.sgm | strip_sgm.sh > $DEST/valid.en_XX-zh_CN.zh_CN + cat newstest2019-zhen-src.zh.sgm | strip_sgm.sh > $DEST/test.zh_CN-en_XX.zh_CN + cat newstest2019-zhen-ref.en.sgm | strip_sgm.sh > $DEST/test.zh_CN-en_XX.en_XX + cat newstest2019-enzh-src.en.sgm | strip_sgm.sh > $DEST/test.en_XX-zh_CN.en_XX + cat newstest2019-enzh-ref.zh.sgm | strip_sgm.sh > $DEST/test.en_XX-zh_CN.zh_CN +} + +mkdir -p $DEST + +prepare_lid +prepare_moses +download_commoncrawl + +prepare_ja & +prepare_ta & +prepare_km & +prepare_ps & +prepare_iu & +prepare_cs & +prepare_de & +prepare_pl & +prepare_ru & +prepare_zh & + +# prepare valid/test set +prepare_tests & + +# wait + +# TODO remove intermediate files +# rm -rf $TMP_DIR diff --git a/examples/multilingual/data_scripts/preprocess_ML50_v1.sh b/examples/multilingual/data_scripts/preprocess_ML50_v1.sh new file mode 100644 index 0000000000000000000000000000000000000000..4655936149cab212b3cfa14f306d71153729f9d7 --- /dev/null +++ b/examples/multilingual/data_scripts/preprocess_ML50_v1.sh @@ -0,0 +1,27 @@ +#!/bin/bash +# Copyright (c) Facebook, Inc. and its affiliates. +# All rights reserved. +# +# This source code is licensed under the license found in the +# LICENSE file in the root directory of this source tree. + +if [ -z $WORKDIR_ROOT ] ; +then + echo "please specify your working directory root in environment variable WORKDIR_ROOT. Exitting..." + exit +fi + +if [ -z $SPM_PATH ] ; +then + echo "Please install sentence piecence from https://github.com/google/sentencepiece and set SPM_PATH pointing to the installed spm_encode.py. Exitting..." + exit +fi + +ML50=${WORKDIR_ROOT}/ML50 + +mkdir -p $ML50/dedup +mkdir -p $ML50/cleaned_dedup + +python ./dedup_all.py --from-folder $ML50/raw --to-folder $ML50/dedup +python ./remove_valid_test_in_train.py --from-folder $ML50/dedup --to-folder $ML50/clean +python ./binarize.py --raw-folder $ML50/clean \ No newline at end of file diff --git a/examples/multilingual/data_scripts/remove_valid_test_in_train.py b/examples/multilingual/data_scripts/remove_valid_test_in_train.py new file mode 100755 index 0000000000000000000000000000000000000000..ef618adef7c7d010f8de38fb5ebeb5a35d2d3cac --- /dev/null +++ b/examples/multilingual/data_scripts/remove_valid_test_in_train.py @@ -0,0 +1,290 @@ +import os, sys +import glob, itertools +import pandas as pd + +WORKDIR_ROOT = os.environ.get('WORKDIR_ROOT', None) + +if WORKDIR_ROOT is None or not WORKDIR_ROOT.strip(): + print('please specify your working directory root in OS environment variable WORKDIR_ROOT. Exitting..."') + sys.exit(-1) + + +def load_langs(path): + with open(path) as fr: + langs = [l.strip() for l in fr] + return langs + + + +def load_sentences(raw_data, split, direction): + src, tgt = direction.split('-') + src_path = f"{raw_data}/{split}.{direction}.{src}" + tgt_path = f"{raw_data}/{split}.{direction}.{tgt}" + if os.path.exists(src_path) and os.path.exists(tgt_path): + return [(src, open(src_path).read().splitlines()), (tgt, open(tgt_path).read().splitlines())] + else: + return [] + +def swap_direction(d): + src, tgt = d.split('-') + return f'{tgt}-{src}' + +def get_all_test_data(raw_data, directions, split='test'): + test_data = [ + x + for dd in directions + for d in [dd, swap_direction(dd)] + for x in load_sentences(raw_data, split, d) + ] + # all_test_data = {s for _, d in test_data for s in d} + all_test_data = {} + for lang, d in test_data: + for s in d: + s = s.strip() + lgs = all_test_data.get(s, set()) + lgs.add(lang) + all_test_data[s] = lgs + return all_test_data, test_data + +def check_train_sentences(raw_data, direction, all_test_data, mess_up_train={}): + src, tgt = direction.split('-') + tgt_path = f"{raw_data}/train.{direction}.{tgt}" + src_path = f"{raw_data}/train.{direction}.{src}" + print(f'check training data in {raw_data}/train.{direction}') + size = 0 + if not os.path.exists(tgt_path) or not os.path.exists(src_path): + return mess_up_train, size + with open(src_path) as f, open(tgt_path) as g: + for src_line, tgt_line in zip(f, g): + s = src_line.strip() + t = tgt_line.strip() + size += 1 + if s in all_test_data: + langs = mess_up_train.get(s, set()) + langs.add(direction) + mess_up_train[s] = langs + if t in all_test_data: + langs = mess_up_train.get(t, set()) + langs.add(direction) + mess_up_train[t] = langs + return mess_up_train, size + +def check_train_all(raw_data, directions, all_test_data): + mess_up_train = {} + data_sizes = {} + for direction in directions: + _, size = check_train_sentences(raw_data, direction, all_test_data, mess_up_train) + data_sizes[direction] = size + return mess_up_train, data_sizes + +def count_train_in_other_set(mess_up_train): + train_in_others = [(direction, s) for s, directions in mess_up_train.items() for direction in directions] + counts = {} + for direction, s in train_in_others: + counts[direction] = counts.get(direction, 0) + 1 + return counts + +def train_size_if_remove_in_otherset(data_sizes, mess_up_train): + counts_in_other = count_train_in_other_set(mess_up_train) + remain_sizes = [] + for direction, count in counts_in_other.items(): + remain_sizes.append((direction, data_sizes[direction] - count, data_sizes[direction], count, 100 * count / data_sizes[direction] )) + return remain_sizes + + +def remove_messed_up_sentences(raw_data, direction, mess_up_train, mess_up_train_pairs, corrected_langs): + split = 'train' + src_lang, tgt_lang = direction.split('-') + + tgt = f"{raw_data}/{split}.{direction}.{tgt_lang}" + src = f"{raw_data}/{split}.{direction}.{src_lang}" + print(f'working on {direction}: ', src, tgt) + if not os.path.exists(tgt) or not os.path.exists(src) : + return + + corrected_tgt = f"{to_folder}/{split}.{direction}.{tgt_lang}" + corrected_src = f"{to_folder}/{split}.{direction}.{src_lang}" + line_num = 0 + keep_num = 0 + with open(src, encoding='utf8',) as fsrc, \ + open(tgt, encoding='utf8',) as ftgt, \ + open(corrected_src, 'w', encoding='utf8') as fsrc_corrected, \ + open(corrected_tgt, 'w', encoding='utf8') as ftgt_corrected: + for s, t in zip(fsrc, ftgt): + s = s.strip() + t = t.strip() + if t not in mess_up_train \ + and s not in mess_up_train \ + and (s, t) not in mess_up_train_pairs \ + and (t, s) not in mess_up_train_pairs: + corrected_langs.add(direction) + print(s, file=fsrc_corrected) + print(t, file=ftgt_corrected) + keep_num += 1 + line_num += 1 + if line_num % 1000 == 0: + print(f'completed {line_num} lines', end='\r') + return line_num, keep_num + +########## + + +def merge_valid_test_messup(mess_up_train_valid, mess_up_train_test): + merged_mess = [] + for s in set(list(mess_up_train_valid.keys()) + list(mess_up_train_test.keys())): + if not s: + continue + valid = mess_up_train_valid.get(s, set()) + test = mess_up_train_test.get(s, set()) + merged_mess.append((s, valid | test)) + return dict(merged_mess) + + + +######### +def check_train_pairs(raw_data, direction, all_test_data, mess_up_train={}): + src, tgt = direction.split('-') + #a hack; TODO: check the reversed directions + path1 = f"{raw_data}/train.{src}-{tgt}.{src}" + path2 = f"{raw_data}/train.{src}-{tgt}.{tgt}" + if not os.path.exists(path1) or not os.path.exists(path2) : + return + + with open(path1) as f1, open(path2) as f2: + for src_line, tgt_line in zip(f1, f2): + s = src_line.strip() + t = tgt_line.strip() + if (s, t) in all_test_data or (t, s) in all_test_data: + langs = mess_up_train.get( (s, t), set()) + langs.add(src) + langs.add(tgt) + mess_up_train[(s, t)] = langs + + +def load_pairs(raw_data, split, direction): + src, tgt = direction.split('-') + src_f = f"{raw_data}/{split}.{direction}.{src}" + tgt_f = f"{raw_data}/{split}.{direction}.{tgt}" + if tgt != 'en_XX': + src_f, tgt_f = tgt_f, src_f + if os.path.exists(src_f) and os.path.exists(tgt_f): + return list(zip(open(src_f).read().splitlines(), + open(tgt_f).read().splitlines(), + )) + else: + return [] + +# skip_langs = ['cs_CZ', 'en_XX', 'tl_XX', 'tr_TR'] +def get_messed_up_test_pairs(split, directions): + test_pairs = [ + (d, load_pairs(raw_data, split, d)) + for d in directions + ] + # all_test_data = {s for _, d in test_data for s in d} + all_test_pairs = {} + for direction, d in test_pairs: + src, tgt = direction.split('-') + for s in d: + langs = all_test_pairs.get(s, set()) + langs.add(src) + langs.add(tgt) + all_test_pairs[s] = langs + mess_up_train_pairs = {} + for direction in directions: + check_train_pairs(raw_data, direction, all_test_pairs, mess_up_train_pairs) + return all_test_pairs, mess_up_train_pairs + + + +if __name__ == "__main__": + ####### + import argparse + parser = argparse.ArgumentParser() + parser.add_argument( + '--from-folder', + required=True, + type=str) + parser.add_argument( + '--to-folder', + required=True, + type=str) + parser.add_argument( + '--directions', + default=None, + type=str) + + + args = parser.parse_args() + raw_data = args.from_folder + to_folder = args.to_folder + os.makedirs(to_folder, exist_ok=True) + + if args.directions: + directions = args.directions.split(',') + else: + raw_files = itertools.chain( + glob.glob(f'{raw_data}/train*'), + glob.glob(f'{raw_data}/valid*'), + glob.glob(f'{raw_data}/test*'), + ) + directions = [os.path.split(file_path)[-1].split('.')[1] for file_path in raw_files] + print('working on directions: ', directions) + + ########## + + + + all_test_data, test_data = get_all_test_data(raw_data, directions, 'test') + print('==loaded test data==') + all_valid_data, valid_data = get_all_test_data(raw_data, directions, 'valid') + print('==loaded valid data==') + all_valid_test_data = merge_valid_test_messup(all_test_data, all_valid_data) + mess_up_train, data_sizes = check_train_all(raw_data, directions, all_valid_test_data) + print('training messing up with valid, test data:', len(mess_up_train)) + data_situation = train_size_if_remove_in_otherset(data_sizes, mess_up_train) + df = pd.DataFrame(data_situation, columns=['direction', 'train_size_after_remove', 'orig_size', 'num_to_remove', 'remove_percent']) + df.sort_values('remove_percent', ascending=False) + df.to_csv(f'{raw_data}/clean_summary.tsv', sep='\t') + print(f'projected data clean summary in: {raw_data}/clean_summary.tsv') + + # correct the dataset: + all_test_pairs, mess_up_test_train_pairs = get_messed_up_test_pairs('test', directions) + all_valid_pairs, mess_up_valid_train_pairs = get_messed_up_test_pairs('valid', directions) + + all_messed_pairs = set(mess_up_test_train_pairs.keys()).union(set(mess_up_valid_train_pairs.keys())) + corrected_directions = set() + + real_data_situation = [] + for direction in directions: + org_size, new_size = remove_messed_up_sentences(raw_data, direction, mess_up_train, all_messed_pairs, corrected_directions) + if org_size == 0: + print(f"{direction} has size 0") + continue + real_data_situation.append( + (direction, new_size, org_size, org_size - new_size, (org_size - new_size) / org_size * 100) + ) + print('corrected directions: ', corrected_directions) + df = pd.DataFrame(real_data_situation, columns=['direction', 'train_size_after_remove', 'orig_size', 'num_to_remove', 'remove_percent']) + df.sort_values('remove_percent', ascending=False) + df.to_csv(f'{raw_data}/actual_clean_summary.tsv', sep='\t') + print(f'actual data clean summary (which can be different from the projected one because of duplications) in: {raw_data}/actual_clean_summary.tsv') + + import shutil + for direction in directions: + src_lang, tgt_lang = direction.split('-') + for split in ['train', 'valid', 'test']: + # copying valid, test and uncorrected train + if direction in corrected_directions and split == 'train': + continue + tgt = f"{raw_data}/{split}.{direction}.{tgt_lang}" + src = f"{raw_data}/{split}.{direction}.{src_lang}" + if not (os.path.exists(src) and os.path.exists(tgt)): + continue + corrected_tgt = f"{to_folder}/{split}.{direction}.{tgt_lang}" + corrected_src = f"{to_folder}/{split}.{direction}.{src_lang}" + print(f'copying {src} to {corrected_src}') + shutil.copyfile(src, corrected_src) + print(f'copying {tgt} to {corrected_tgt}') + shutil.copyfile(tgt, corrected_tgt) + + print('completed') \ No newline at end of file diff --git a/examples/multilingual/data_scripts/requirement.txt b/examples/multilingual/data_scripts/requirement.txt new file mode 100644 index 0000000000000000000000000000000000000000..e85d7d540e08a1407f92dfb2311972a1a5a30123 --- /dev/null +++ b/examples/multilingual/data_scripts/requirement.txt @@ -0,0 +1,2 @@ +wget +pandas \ No newline at end of file diff --git a/examples/multilingual/data_scripts/utils/dedup.py b/examples/multilingual/data_scripts/utils/dedup.py new file mode 100644 index 0000000000000000000000000000000000000000..d6fed8c695cf218d3502d6ed8d23015520c0e179 --- /dev/null +++ b/examples/multilingual/data_scripts/utils/dedup.py @@ -0,0 +1,41 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +import argparse + +def deup(src_file, tgt_file, src_file_out, tgt_file_out): + seen = set() + dup_count = 0 + with open(src_file, encoding='utf-8') as fsrc, \ + open(tgt_file, encoding='utf-8') as ftgt, \ + open(src_file_out, 'w', encoding='utf-8') as fsrc_out, \ + open(tgt_file_out, 'w', encoding='utf-8') as ftgt_out: + for s, t in zip(fsrc, ftgt): + if (s, t) not in seen: + fsrc_out.write(s) + ftgt_out.write(t) + seen.add((s, t)) + else: + dup_count += 1 + print(f'number of duplication: {dup_count}') + + +def main(): + parser = argparse.ArgumentParser() + parser.add_argument("--src-file", type=str, required=True, + help="src file") + parser.add_argument("--tgt-file", type=str, required=True, + help="tgt file") + parser.add_argument("--src-file-out", type=str, required=True, + help="src ouptut file") + parser.add_argument("--tgt-file-out", type=str, required=True, + help="tgt ouput file") + args = parser.parse_args() + deup(args.src_file, args.tgt_file, args.src_file_out, args.tgt_file_out) + + +if __name__ == "__main__": + main() diff --git a/examples/multilingual/data_scripts/utils/fasttext_multi_filter.py b/examples/multilingual/data_scripts/utils/fasttext_multi_filter.py new file mode 100644 index 0000000000000000000000000000000000000000..41b38ba5bef20cb043921ac61820db8689189a5a --- /dev/null +++ b/examples/multilingual/data_scripts/utils/fasttext_multi_filter.py @@ -0,0 +1,63 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +#!/bin/python + +import fasttext +from multiprocessing import Pool +import contextlib +import sys +import argparse +from functools import partial +import io + +model = None +def init(model_path): + global model + model = fasttext.load_model(model_path) + +def pred(lines): + return lines, [model.predict(line.strip())[0][0][9:] for line in lines] + +def main(): + parser = argparse.ArgumentParser() + parser.add_argument("--model", type=str, required=True, + help="model to load") + parser.add_argument("--inputs", nargs="+", default=['-'], + help="input files to filter") + parser.add_argument("--langs", nargs="+", required=True, + help="lang ids of each input file") + parser.add_argument("--outputs", nargs="+", default=['-'], + help="path to save lid filtered outputs") + parser.add_argument("--num-workers", type=int, metavar="N", default=10, + help="number of processes in parallel") + args = parser.parse_args() + + assert len(args.inputs) == len(args.langs) and len(args.inputs) == len(args.outputs) + + with contextlib.ExitStack() as stack: + inputs = [ + stack.enter_context(open(input, "r", encoding="utf-8", newline="\n", errors="replace")) + if input != "-" else io.TextIOWrapper(sys.stdin.buffer, encoding='utf-8', errors="replace") + for input in args.inputs + ] + outputs = [ + stack.enter_context(open(output, "w", encoding="utf-8", newline="\n")) + if output != "-" else sys.stdout + for output in args.outputs + ] + with Pool(args.num_workers, initializer=partial(init, args.model)) as p: + skip_cnt = 0 + for lines, preds in p.imap(pred, list(zip(*inputs)), chunksize=500): + if not all(a == b for a, b in zip(preds, args.langs)): + skip_cnt += 1 + continue + for line, output_h in zip(lines, outputs): + print(line.strip(), file=output_h) + print(f"Skipped {skip_cnt} lines.") + +if __name__ == "__main__": + main() diff --git a/examples/multilingual/data_scripts/utils/strip_sgm.sh b/examples/multilingual/data_scripts/utils/strip_sgm.sh new file mode 100755 index 0000000000000000000000000000000000000000..7f4f61d7b1a46f51a1221de6b336cb70b5a0b8b3 --- /dev/null +++ b/examples/multilingual/data_scripts/utils/strip_sgm.sh @@ -0,0 +1 @@ +grep "seg id" | sed 's/<seg id="[0-9]\+">//g' | sed 's/<\/seg>//g' diff --git a/examples/multilingual/finetune_multilingual_model.sh b/examples/multilingual/finetune_multilingual_model.sh new file mode 100644 index 0000000000000000000000000000000000000000..25960c5dc8a02e5580b61837099770a082b4dd83 --- /dev/null +++ b/examples/multilingual/finetune_multilingual_model.sh @@ -0,0 +1,32 @@ +#!/bin/bash +# Copyright (c) Facebook, Inc. and its affiliates. +# All rights reserved. +# +# This source code is licensed under the license found in the +# LICENSE file in the root directory of this source tree. + +path_2_data=$1 # <path to data> which contains binarized data for each directions +lang_list=$2 # <path to a file which contains a list of languages separted by new lines> +lang_pairs=$3 #a list language pairs to train multilingual models, e.g. "en-fr,en-cs,fr-en,cs-en" +# pretrained can be an mBART pretrained model as well +pretrained_model=$4 #<path to a pretrained model> + + +fairseq-train "$path_2_data" \ + --encoder-normalize-before --decoder-normalize-before \ + --arch transformer --layernorm-embedding \ + --task translation_multi_simple_epoch \ + --finetune-from-model "$pretrained_model" \ + --sampling-method "temperature" \ + --sampling-temperature "1.5" \ + --encoder-langtok "src" \ + --decoder-langtok \ + --lang-dict "$lang_list" \ + --lang-pairs "$lang_pairs" \ + --criterion label_smoothed_cross_entropy --label-smoothing 0.2 \ + --optimizer adam --adam-eps 1e-06 --adam-betas '(0.9, 0.98)' \ + --lr-scheduler inverse_sqrt --lr 3e-05 --warmup-updates 2500 --max-update 40000 \ + --dropout 0.3 --attention-dropout 0.1 --weight-decay 0.0 \ + --max-tokens 1024 --update-freq 2 \ + --save-interval 1 --save-interval-updates 5000 --keep-interval-updates 10 --no-epoch-checkpoints \ + --seed 222 --log-format simple --log-interval 2 diff --git a/examples/multilingual/multilingual_fairseq_gen.sh b/examples/multilingual/multilingual_fairseq_gen.sh new file mode 100644 index 0000000000000000000000000000000000000000..65aa322d7daaa428015de98abe4664a6a4164bfd --- /dev/null +++ b/examples/multilingual/multilingual_fairseq_gen.sh @@ -0,0 +1,26 @@ +#!/bin/bash +# Copyright (c) Facebook, Inc. and its affiliates. +# All rights reserved. +# +# This source code is licensed under the license found in the +# LICENSE file in the root directory of this source tree. + +lang_pairs="en-fr,en-cs,fr-en,cs-en" +path_2_data=$1 # <path to data> +lang_list=$2 # <path to a file which contains list of languages separted by new lines> +model=$3 # <path to a trained model> +source_lang=cs +target_lang=en + +fairseq-generate "$path_2_data" \ + --path "$model" \ + --task translation_multi_simple_epoch \ + --gen-subset test \ + --source-lang "$source_lang" \ + --target-lang "$target_lang" \ + --sacrebleu --remove-bpe 'sentencepiece'\ + --batch-size 32 \ + --encoder-langtok "src" \ + --decoder-langtok \ + --lang-dict "$lang_list" \ + --lang-pairs "$lang_pairs" diff --git a/examples/multilingual/train_multilingual_model.sh b/examples/multilingual/train_multilingual_model.sh new file mode 100644 index 0000000000000000000000000000000000000000..cc050bd3f02de8a2f303737f187442d2eb80e4ef --- /dev/null +++ b/examples/multilingual/train_multilingual_model.sh @@ -0,0 +1,28 @@ +#!/bin/bash +# Copyright (c) Facebook, Inc. and its affiliates. +# All rights reserved. +# +# This source code is licensed under the license found in the +# LICENSE file in the root directory of this source tree. + +path_2_data=$1 # <path to data> which contains binarized data for each directions +lang_list=$2 # <path to a file which contains a list of languages separted by new lines> +lang_pairs=$3 #a list language pairs to train multilingual models, e.g. "en-fr,en-cs,fr-en,cs-en" + +fairseq-train "$path_2_data" \ + --encoder-normalize-before --decoder-normalize-before \ + --arch transformer --layernorm-embedding \ + --task translation_multi_simple_epoch \ + --sampling-method "temperature" \ + --sampling-temperature 1.5 \ + --encoder-langtok "src" \ + --decoder-langtok \ + --lang-dict "$lang_list" \ + --lang-pairs "$lang_pairs" \ + --criterion label_smoothed_cross_entropy --label-smoothing 0.2 \ + --optimizer adam --adam-eps 1e-06 --adam-betas '(0.9, 0.98)' \ + --lr-scheduler inverse_sqrt --lr 3e-05 --warmup-updates 2500 --max-update 40000 \ + --dropout 0.3 --attention-dropout 0.1 --weight-decay 0.0 \ + --max-tokens 1024 --update-freq 2 \ + --save-interval 1 --save-interval-updates 5000 --keep-interval-updates 10 --no-epoch-checkpoints \ + --seed 222 --log-format simple --log-interval 2 diff --git a/examples/noisychannel/README.md b/examples/noisychannel/README.md new file mode 100644 index 0000000000000000000000000000000000000000..9d101aa874ec36ff3bb5c1166169a4c4f38ffe2b --- /dev/null +++ b/examples/noisychannel/README.md @@ -0,0 +1,72 @@ +# Simple and Effective Noisy Channel Modeling for Neural Machine Translation (Yee et al., 2019) +This page contains pointers to pre-trained models as well as instructions on how to run the reranking scripts. + +## Citation: +```bibtex +@inproceedings{yee2019simple, + title = {Simple and Effective Noisy Channel Modeling for Neural Machine Translation}, + author = {Kyra Yee and Yann Dauphin and Michael Auli}, + booktitle = {Conference on Empirical Methods in Natural Language Processing}, + year = {2019}, +} +``` + +## Pre-trained Models: + +Model | Description | Download +---|---|--- +`transformer.noisychannel.de-en` | De->En Forward Model | [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/noisychannel/forward_de2en.tar.bz2) +`transformer.noisychannel.en-de` | En->De Channel Model | [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/noisychannel/backward_en2de.tar.bz2) +`transformer_lm.noisychannel.en` | En Language model | [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/noisychannel/reranking_en_lm.tar.bz2) + +Test Data: [newstest_wmt17](https://dl.fbaipublicfiles.com/fairseq/models/noisychannel/wmt17test.tar.bz2) + +## Example usage + +``` +mkdir rerank_example +curl https://dl.fbaipublicfiles.com/fairseq/models/noisychannel/forward_de2en.tar.bz2 | tar xvjf - -C rerank_example +curl https://dl.fbaipublicfiles.com/fairseq/models/noisychannel/backward_en2de.tar.bz2 | tar xvjf - -C rerank_example +curl https://dl.fbaipublicfiles.com/fairseq/models/noisychannel/reranking_en_lm.tar.bz2 | tar xvjf - -C rerank_example +curl https://dl.fbaipublicfiles.com/fairseq/models/noisychannel/wmt17test.tar.bz2 | tar xvjf - -C rerank_example + +beam=50 +num_trials=1000 +fw_name=fw_model_ex +bw_name=bw_model_ex +lm_name=lm_ex +data_dir=rerank_example/hyphen-splitting-mixed-case-wmt17test-wmt14bpe +data_dir_name=wmt17 +lm=rerank_example/lm/checkpoint_best.pt +lm_bpe_code=rerank_example/lm/bpe32k.code +lm_dict=rerank_example/lm/dict.txt +batch_size=32 +bw=rerank_example/backward_en2de.pt +fw=rerank_example/forward_de2en.pt + +# reranking with P(T|S) P(S|T) and P(T) +python examples/noisychannel/rerank_tune.py $data_dir --tune-param lenpen weight1 weight3 \ + --lower-bound 0 0 0 --upper-bound 3 3 3 --data-dir-name $data_dir_name \ + --num-trials $num_trials --source-lang de --target-lang en --gen-model $fw \ + -n $beam --batch-size $batch_size --score-model2 $fw --score-model1 $bw \ + --backwards1 --weight2 1 \ + -lm $lm --lm-dict $lm_dict --lm-name en_newscrawl --lm-bpe-code $lm_bpe_code \ + --model2-name $fw_name --model1-name $bw_name --gen-model-name $fw_name + +# reranking with P(T|S) and P(T) +python examples/noisychannel/rerank_tune.py $data_dir --tune-param lenpen weight3 \ + --lower-bound 0 0 --upper-bound 3 3 --data-dir-name $data_dir_name \ + --num-trials $num_trials --source-lang de --target-lang en --gen-model $fw \ + -n $beam --batch-size $batch_size --score-model1 $fw \ + -lm $lm --lm-dict $lm_dict --lm-name en_newscrawl --lm-bpe-code $lm_bpe_code \ + --model1-name $fw_name --gen-model-name $fw_name + +# to run with a preconfigured set of hyperparameters for the lenpen and model weights, using rerank.py instead. +python examples/noisychannel/rerank.py $data_dir \ + --lenpen 0.269 --weight1 1 --weight2 0.929 --weight3 0.831 \ + --data-dir-name $data_dir_name --source-lang de --target-lang en --gen-model $fw \ + -n $beam --batch-size $batch_size --score-model2 $fw --score-model1 $bw --backwards1 \ + -lm $lm --lm-dict $lm_dict --lm-name en_newscrawl --lm-bpe-code $lm_bpe_code \ + --model2-name $fw_name --model1-name $bw_name --gen-model-name $fw_name +``` + diff --git a/examples/noisychannel/__init__.py b/examples/noisychannel/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..89f1aef4f6328d25425e0bcabb42dfffd2ed35f0 --- /dev/null +++ b/examples/noisychannel/__init__.py @@ -0,0 +1,6 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .rerank_options import * # noqa diff --git a/examples/noisychannel/rerank.py b/examples/noisychannel/rerank.py new file mode 100644 index 0000000000000000000000000000000000000000..bb80d11a67cd75764a89f6f41915b0348ae96e92 --- /dev/null +++ b/examples/noisychannel/rerank.py @@ -0,0 +1,428 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +from multiprocessing import Pool + +import numpy as np +from fairseq import options +from fairseq.data import dictionary +from fairseq.scoring import bleu + +from examples.noisychannel import ( + rerank_generate, + rerank_options, + rerank_score_bw, + rerank_score_lm, + rerank_utils, +) + + +def score_target_hypo( + args, a, b, c, lenpen, target_outfile, hypo_outfile, write_hypos, normalize +): + + print("lenpen", lenpen, "weight1", a, "weight2", b, "weight3", c) + gen_output_lst, bitext1_lst, bitext2_lst, lm_res_lst = load_score_files(args) + dict = dictionary.Dictionary() + scorer = scorer = bleu.Scorer( + bleu.BleuConfig( + pad=dict.pad(), + eos=dict.eos(), + unk=dict.unk(), + ) + ) + + ordered_hypos = {} + ordered_targets = {} + + for shard_id in range(len(bitext1_lst)): + bitext1 = bitext1_lst[shard_id] + bitext2 = bitext2_lst[shard_id] + gen_output = gen_output_lst[shard_id] + lm_res = lm_res_lst[shard_id] + + total = len(bitext1.rescore_source.keys()) + source_lst = [] + hypo_lst = [] + score_lst = [] + reference_lst = [] + j = 1 + best_score = -math.inf + + for i in range(total): + # length is measured in terms of words, not bpe tokens, since models may not share the same bpe + target_len = len(bitext1.rescore_hypo[i].split()) + + if lm_res is not None: + lm_score = lm_res.score[i] + else: + lm_score = 0 + + if bitext2 is not None: + bitext2_score = bitext2.rescore_score[i] + bitext2_backwards = bitext2.backwards + else: + bitext2_score = None + bitext2_backwards = None + + score = rerank_utils.get_score( + a, + b, + c, + target_len, + bitext1.rescore_score[i], + bitext2_score, + lm_score=lm_score, + lenpen=lenpen, + src_len=bitext1.source_lengths[i], + tgt_len=bitext1.target_lengths[i], + bitext1_backwards=bitext1.backwards, + bitext2_backwards=bitext2_backwards, + normalize=normalize, + ) + + if score > best_score: + best_score = score + best_hypo = bitext1.rescore_hypo[i] + + if j == gen_output.num_hypos[i] or j == args.num_rescore: + j = 1 + hypo_lst.append(best_hypo) + score_lst.append(best_score) + source_lst.append(bitext1.rescore_source[i]) + reference_lst.append(bitext1.rescore_target[i]) + + best_score = -math.inf + best_hypo = "" + else: + j += 1 + + gen_keys = list(sorted(gen_output.no_bpe_target.keys())) + + for key in range(len(gen_keys)): + if args.prefix_len is None: + assert hypo_lst[key] in gen_output.no_bpe_hypo[gen_keys[key]], ( + "pred and rescore hypo mismatch: i: " + + str(key) + + ", " + + str(hypo_lst[key]) + + str(gen_keys[key]) + + str(gen_output.no_bpe_hypo[key]) + ) + sys_tok = dict.encode_line(hypo_lst[key]) + ref_tok = dict.encode_line(gen_output.no_bpe_target[gen_keys[key]]) + scorer.add(ref_tok, sys_tok) + + else: + full_hypo = rerank_utils.get_full_from_prefix( + hypo_lst[key], gen_output.no_bpe_hypo[gen_keys[key]] + ) + sys_tok = dict.encode_line(full_hypo) + ref_tok = dict.encode_line(gen_output.no_bpe_target[gen_keys[key]]) + scorer.add(ref_tok, sys_tok) + + # if only one set of hyper parameters is provided, write the predictions to a file + if write_hypos: + # recover the orinal ids from n best list generation + for key in range(len(gen_output.no_bpe_target)): + if args.prefix_len is None: + assert hypo_lst[key] in gen_output.no_bpe_hypo[gen_keys[key]], ( + "pred and rescore hypo mismatch:" + + "i:" + + str(key) + + str(hypo_lst[key]) + + str(gen_output.no_bpe_hypo[key]) + ) + ordered_hypos[gen_keys[key]] = hypo_lst[key] + ordered_targets[gen_keys[key]] = gen_output.no_bpe_target[ + gen_keys[key] + ] + + else: + full_hypo = rerank_utils.get_full_from_prefix( + hypo_lst[key], gen_output.no_bpe_hypo[gen_keys[key]] + ) + ordered_hypos[gen_keys[key]] = full_hypo + ordered_targets[gen_keys[key]] = gen_output.no_bpe_target[ + gen_keys[key] + ] + + # write the hypos in the original order from nbest list generation + if args.num_shards == (len(bitext1_lst)): + with open(target_outfile, "w") as t: + with open(hypo_outfile, "w") as h: + for key in range(len(ordered_hypos)): + t.write(ordered_targets[key]) + h.write(ordered_hypos[key]) + + res = scorer.result_string(4) + if write_hypos: + print(res) + score = rerank_utils.parse_bleu_scoring(res) + return score + + +def match_target_hypo(args, target_outfile, hypo_outfile): + """combine scores from the LM and bitext models, and write the top scoring hypothesis to a file""" + if len(args.weight1) == 1: + res = score_target_hypo( + args, + args.weight1[0], + args.weight2[0], + args.weight3[0], + args.lenpen[0], + target_outfile, + hypo_outfile, + True, + args.normalize, + ) + rerank_scores = [res] + else: + print("launching pool") + with Pool(32) as p: + rerank_scores = p.starmap( + score_target_hypo, + [ + ( + args, + args.weight1[i], + args.weight2[i], + args.weight3[i], + args.lenpen[i], + target_outfile, + hypo_outfile, + False, + args.normalize, + ) + for i in range(len(args.weight1)) + ], + ) + + if len(rerank_scores) > 1: + best_index = np.argmax(rerank_scores) + best_score = rerank_scores[best_index] + print("best score", best_score) + print("best lenpen", args.lenpen[best_index]) + print("best weight1", args.weight1[best_index]) + print("best weight2", args.weight2[best_index]) + print("best weight3", args.weight3[best_index]) + return ( + args.lenpen[best_index], + args.weight1[best_index], + args.weight2[best_index], + args.weight3[best_index], + best_score, + ) + + else: + return ( + args.lenpen[0], + args.weight1[0], + args.weight2[0], + args.weight3[0], + rerank_scores[0], + ) + + +def load_score_files(args): + if args.all_shards: + shard_ids = list(range(args.num_shards)) + else: + shard_ids = [args.shard_id] + + gen_output_lst = [] + bitext1_lst = [] + bitext2_lst = [] + lm_res1_lst = [] + + for shard_id in shard_ids: + using_nbest = args.nbest_list is not None + ( + pre_gen, + left_to_right_preprocessed_dir, + right_to_left_preprocessed_dir, + backwards_preprocessed_dir, + lm_preprocessed_dir, + ) = rerank_utils.get_directories( + args.data_dir_name, + args.num_rescore, + args.gen_subset, + args.gen_model_name, + shard_id, + args.num_shards, + args.sampling, + args.prefix_len, + args.target_prefix_frac, + args.source_prefix_frac, + ) + + rerank1_is_gen = ( + args.gen_model == args.score_model1 and args.source_prefix_frac is None + ) + rerank2_is_gen = ( + args.gen_model == args.score_model2 and args.source_prefix_frac is None + ) + + score1_file = rerank_utils.rescore_file_name( + pre_gen, + args.prefix_len, + args.model1_name, + target_prefix_frac=args.target_prefix_frac, + source_prefix_frac=args.source_prefix_frac, + backwards=args.backwards1, + ) + if args.score_model2 is not None: + score2_file = rerank_utils.rescore_file_name( + pre_gen, + args.prefix_len, + args.model2_name, + target_prefix_frac=args.target_prefix_frac, + source_prefix_frac=args.source_prefix_frac, + backwards=args.backwards2, + ) + if args.language_model is not None: + lm_score_file = rerank_utils.rescore_file_name( + pre_gen, args.prefix_len, args.lm_name, lm_file=True + ) + + # get gen output + predictions_bpe_file = pre_gen + "/generate_output_bpe.txt" + if using_nbest: + print("Using predefined n-best list from interactive.py") + predictions_bpe_file = args.nbest_list + gen_output = rerank_utils.BitextOutputFromGen( + predictions_bpe_file, + bpe_symbol=args.post_process, + nbest=using_nbest, + prefix_len=args.prefix_len, + target_prefix_frac=args.target_prefix_frac, + ) + + if rerank1_is_gen: + bitext1 = gen_output + else: + bitext1 = rerank_utils.BitextOutput( + score1_file, + args.backwards1, + args.right_to_left1, + args.post_process, + args.prefix_len, + args.target_prefix_frac, + args.source_prefix_frac, + ) + + if args.score_model2 is not None or args.nbest_list is not None: + if rerank2_is_gen: + bitext2 = gen_output + else: + bitext2 = rerank_utils.BitextOutput( + score2_file, + args.backwards2, + args.right_to_left2, + args.post_process, + args.prefix_len, + args.target_prefix_frac, + args.source_prefix_frac, + ) + + assert ( + bitext2.source_lengths == bitext1.source_lengths + ), "source lengths for rescoring models do not match" + assert ( + bitext2.target_lengths == bitext1.target_lengths + ), "target lengths for rescoring models do not match" + else: + if args.diff_bpe: + assert args.score_model2 is None + bitext2 = gen_output + else: + bitext2 = None + + if args.language_model is not None: + lm_res1 = rerank_utils.LMOutput( + lm_score_file, + args.lm_dict, + args.prefix_len, + args.post_process, + args.target_prefix_frac, + ) + else: + lm_res1 = None + + gen_output_lst.append(gen_output) + bitext1_lst.append(bitext1) + bitext2_lst.append(bitext2) + lm_res1_lst.append(lm_res1) + return gen_output_lst, bitext1_lst, bitext2_lst, lm_res1_lst + + +def rerank(args): + if type(args.lenpen) is not list: + args.lenpen = [args.lenpen] + if type(args.weight1) is not list: + args.weight1 = [args.weight1] + if type(args.weight2) is not list: + args.weight2 = [args.weight2] + if type(args.weight3) is not list: + args.weight3 = [args.weight3] + if args.all_shards: + shard_ids = list(range(args.num_shards)) + else: + shard_ids = [args.shard_id] + + for shard_id in shard_ids: + ( + pre_gen, + left_to_right_preprocessed_dir, + right_to_left_preprocessed_dir, + backwards_preprocessed_dir, + lm_preprocessed_dir, + ) = rerank_utils.get_directories( + args.data_dir_name, + args.num_rescore, + args.gen_subset, + args.gen_model_name, + shard_id, + args.num_shards, + args.sampling, + args.prefix_len, + args.target_prefix_frac, + args.source_prefix_frac, + ) + rerank_generate.gen_and_reprocess_nbest(args) + rerank_score_bw.score_bw(args) + rerank_score_lm.score_lm(args) + + if args.write_hypos is None: + write_targets = pre_gen + "/matched_targets" + write_hypos = pre_gen + "/matched_hypos" + else: + write_targets = args.write_hypos + "_targets" + args.gen_subset + write_hypos = args.write_hypos + "_hypos" + args.gen_subset + + if args.all_shards: + write_targets += "_all_shards" + write_hypos += "_all_shards" + + ( + best_lenpen, + best_weight1, + best_weight2, + best_weight3, + best_score, + ) = match_target_hypo(args, write_targets, write_hypos) + + return best_lenpen, best_weight1, best_weight2, best_weight3, best_score + + +def cli_main(): + parser = rerank_options.get_reranking_parser() + args = options.parse_args_and_arch(parser) + rerank(args) + + +if __name__ == "__main__": + cli_main() diff --git a/examples/noisychannel/rerank_generate.py b/examples/noisychannel/rerank_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..daeeae059a677a9fcd7c370be087f1f5c189bc52 --- /dev/null +++ b/examples/noisychannel/rerank_generate.py @@ -0,0 +1,397 @@ +#!/usr/bin/env python3 -u +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +""" +Generate n-best translations using a trained model. +""" + +import os +import subprocess +from contextlib import redirect_stdout + +from fairseq import options +from fairseq_cli import generate, preprocess + +from examples.noisychannel import rerank_options, rerank_utils + + +def gen_and_reprocess_nbest(args): + if args.score_dict_dir is None: + args.score_dict_dir = args.data + if args.prefix_len is not None: + assert ( + args.right_to_left1 is False + ), "prefix length not compatible with right to left models" + assert ( + args.right_to_left2 is False + ), "prefix length not compatible with right to left models" + + if args.nbest_list is not None: + assert args.score_model2 is None + + if args.backwards1: + scorer1_src = args.target_lang + scorer1_tgt = args.source_lang + else: + scorer1_src = args.source_lang + scorer1_tgt = args.target_lang + + store_data = ( + os.path.join(os.path.dirname(__file__)) + "/rerank_data/" + args.data_dir_name + ) + if not os.path.exists(store_data): + os.makedirs(store_data) + + ( + pre_gen, + left_to_right_preprocessed_dir, + right_to_left_preprocessed_dir, + backwards_preprocessed_dir, + lm_preprocessed_dir, + ) = rerank_utils.get_directories( + args.data_dir_name, + args.num_rescore, + args.gen_subset, + args.gen_model_name, + args.shard_id, + args.num_shards, + args.sampling, + args.prefix_len, + args.target_prefix_frac, + args.source_prefix_frac, + ) + assert not ( + args.right_to_left1 and args.backwards1 + ), "backwards right to left not supported" + assert not ( + args.right_to_left2 and args.backwards2 + ), "backwards right to left not supported" + assert not ( + args.prefix_len is not None and args.target_prefix_frac is not None + ), "target prefix frac and target prefix len incompatible" + + # make directory to store generation results + if not os.path.exists(pre_gen): + os.makedirs(pre_gen) + + rerank1_is_gen = ( + args.gen_model == args.score_model1 and args.source_prefix_frac is None + ) + rerank2_is_gen = ( + args.gen_model == args.score_model2 and args.source_prefix_frac is None + ) + + if args.nbest_list is not None: + rerank2_is_gen = True + + # make directories to store preprossed nbest list for reranking + if not os.path.exists(left_to_right_preprocessed_dir): + os.makedirs(left_to_right_preprocessed_dir) + if not os.path.exists(right_to_left_preprocessed_dir): + os.makedirs(right_to_left_preprocessed_dir) + if not os.path.exists(lm_preprocessed_dir): + os.makedirs(lm_preprocessed_dir) + if not os.path.exists(backwards_preprocessed_dir): + os.makedirs(backwards_preprocessed_dir) + + score1_file = rerank_utils.rescore_file_name( + pre_gen, + args.prefix_len, + args.model1_name, + target_prefix_frac=args.target_prefix_frac, + source_prefix_frac=args.source_prefix_frac, + backwards=args.backwards1, + ) + if args.score_model2 is not None: + score2_file = rerank_utils.rescore_file_name( + pre_gen, + args.prefix_len, + args.model2_name, + target_prefix_frac=args.target_prefix_frac, + source_prefix_frac=args.source_prefix_frac, + backwards=args.backwards2, + ) + + predictions_bpe_file = pre_gen + "/generate_output_bpe.txt" + + using_nbest = args.nbest_list is not None + + if using_nbest: + print("Using predefined n-best list from interactive.py") + predictions_bpe_file = args.nbest_list + + else: + if not os.path.isfile(predictions_bpe_file): + print("STEP 1: generate predictions using the p(T|S) model with bpe") + print(args.data) + param1 = [ + args.data, + "--path", + args.gen_model, + "--shard-id", + str(args.shard_id), + "--num-shards", + str(args.num_shards), + "--nbest", + str(args.num_rescore), + "--batch-size", + str(args.batch_size), + "--beam", + str(args.num_rescore), + "--batch-size", + str(args.num_rescore), + "--gen-subset", + args.gen_subset, + "--source-lang", + args.source_lang, + "--target-lang", + args.target_lang, + ] + if args.sampling: + param1 += ["--sampling"] + + gen_parser = options.get_generation_parser() + input_args = options.parse_args_and_arch(gen_parser, param1) + + print(input_args) + with open(predictions_bpe_file, "w") as f: + with redirect_stdout(f): + generate.main(input_args) + + gen_output = rerank_utils.BitextOutputFromGen( + predictions_bpe_file, + bpe_symbol=args.post_process, + nbest=using_nbest, + prefix_len=args.prefix_len, + target_prefix_frac=args.target_prefix_frac, + ) + + if args.diff_bpe: + rerank_utils.write_reprocessed( + gen_output.no_bpe_source, + gen_output.no_bpe_hypo, + gen_output.no_bpe_target, + pre_gen + "/source_gen_bpe." + args.source_lang, + pre_gen + "/target_gen_bpe." + args.target_lang, + pre_gen + "/reference_gen_bpe." + args.target_lang, + ) + bitext_bpe = args.rescore_bpe_code + bpe_src_param = [ + "-c", + bitext_bpe, + "--input", + pre_gen + "/source_gen_bpe." + args.source_lang, + "--output", + pre_gen + "/rescore_data." + args.source_lang, + ] + bpe_tgt_param = [ + "-c", + bitext_bpe, + "--input", + pre_gen + "/target_gen_bpe." + args.target_lang, + "--output", + pre_gen + "/rescore_data." + args.target_lang, + ] + + subprocess.call( + [ + "python", + os.path.join( + os.path.dirname(__file__), "subword-nmt/subword_nmt/apply_bpe.py" + ), + ] + + bpe_src_param, + shell=False, + ) + + subprocess.call( + [ + "python", + os.path.join( + os.path.dirname(__file__), "subword-nmt/subword_nmt/apply_bpe.py" + ), + ] + + bpe_tgt_param, + shell=False, + ) + + if (not os.path.isfile(score1_file) and not rerank1_is_gen) or ( + args.score_model2 is not None + and not os.path.isfile(score2_file) + and not rerank2_is_gen + ): + print( + "STEP 2: process the output of generate.py so we have clean text files with the translations" + ) + + rescore_file = "/rescore_data" + if args.prefix_len is not None: + prefix_len_rescore_file = rescore_file + "prefix" + str(args.prefix_len) + if args.target_prefix_frac is not None: + target_prefix_frac_rescore_file = ( + rescore_file + "target_prefix_frac" + str(args.target_prefix_frac) + ) + if args.source_prefix_frac is not None: + source_prefix_frac_rescore_file = ( + rescore_file + "source_prefix_frac" + str(args.source_prefix_frac) + ) + + if not args.right_to_left1 or not args.right_to_left2: + if not args.diff_bpe: + rerank_utils.write_reprocessed( + gen_output.source, + gen_output.hypo, + gen_output.target, + pre_gen + rescore_file + "." + args.source_lang, + pre_gen + rescore_file + "." + args.target_lang, + pre_gen + "/reference_file", + bpe_symbol=args.post_process, + ) + if args.prefix_len is not None: + bw_rescore_file = prefix_len_rescore_file + rerank_utils.write_reprocessed( + gen_output.source, + gen_output.hypo, + gen_output.target, + pre_gen + prefix_len_rescore_file + "." + args.source_lang, + pre_gen + prefix_len_rescore_file + "." + args.target_lang, + pre_gen + "/reference_file", + prefix_len=args.prefix_len, + bpe_symbol=args.post_process, + ) + elif args.target_prefix_frac is not None: + bw_rescore_file = target_prefix_frac_rescore_file + rerank_utils.write_reprocessed( + gen_output.source, + gen_output.hypo, + gen_output.target, + pre_gen + + target_prefix_frac_rescore_file + + "." + + args.source_lang, + pre_gen + + target_prefix_frac_rescore_file + + "." + + args.target_lang, + pre_gen + "/reference_file", + bpe_symbol=args.post_process, + target_prefix_frac=args.target_prefix_frac, + ) + else: + bw_rescore_file = rescore_file + + if args.source_prefix_frac is not None: + fw_rescore_file = source_prefix_frac_rescore_file + rerank_utils.write_reprocessed( + gen_output.source, + gen_output.hypo, + gen_output.target, + pre_gen + + source_prefix_frac_rescore_file + + "." + + args.source_lang, + pre_gen + + source_prefix_frac_rescore_file + + "." + + args.target_lang, + pre_gen + "/reference_file", + bpe_symbol=args.post_process, + source_prefix_frac=args.source_prefix_frac, + ) + else: + fw_rescore_file = rescore_file + + if args.right_to_left1 or args.right_to_left2: + rerank_utils.write_reprocessed( + gen_output.source, + gen_output.hypo, + gen_output.target, + pre_gen + "/right_to_left_rescore_data." + args.source_lang, + pre_gen + "/right_to_left_rescore_data." + args.target_lang, + pre_gen + "/right_to_left_reference_file", + right_to_left=True, + bpe_symbol=args.post_process, + ) + + print("STEP 3: binarize the translations") + if ( + not args.right_to_left1 + or args.score_model2 is not None + and not args.right_to_left2 + or not rerank1_is_gen + ): + + if args.backwards1 or args.backwards2: + if args.backwards_score_dict_dir is not None: + bw_dict = args.backwards_score_dict_dir + else: + bw_dict = args.score_dict_dir + bw_preprocess_param = [ + "--source-lang", + scorer1_src, + "--target-lang", + scorer1_tgt, + "--trainpref", + pre_gen + bw_rescore_file, + "--srcdict", + bw_dict + "/dict." + scorer1_src + ".txt", + "--tgtdict", + bw_dict + "/dict." + scorer1_tgt + ".txt", + "--destdir", + backwards_preprocessed_dir, + ] + preprocess_parser = options.get_preprocessing_parser() + input_args = preprocess_parser.parse_args(bw_preprocess_param) + preprocess.main(input_args) + + preprocess_param = [ + "--source-lang", + scorer1_src, + "--target-lang", + scorer1_tgt, + "--trainpref", + pre_gen + fw_rescore_file, + "--srcdict", + args.score_dict_dir + "/dict." + scorer1_src + ".txt", + "--tgtdict", + args.score_dict_dir + "/dict." + scorer1_tgt + ".txt", + "--destdir", + left_to_right_preprocessed_dir, + ] + preprocess_parser = options.get_preprocessing_parser() + input_args = preprocess_parser.parse_args(preprocess_param) + preprocess.main(input_args) + + if args.right_to_left1 or args.right_to_left2: + preprocess_param = [ + "--source-lang", + scorer1_src, + "--target-lang", + scorer1_tgt, + "--trainpref", + pre_gen + "/right_to_left_rescore_data", + "--srcdict", + args.score_dict_dir + "/dict." + scorer1_src + ".txt", + "--tgtdict", + args.score_dict_dir + "/dict." + scorer1_tgt + ".txt", + "--destdir", + right_to_left_preprocessed_dir, + ] + preprocess_parser = options.get_preprocessing_parser() + input_args = preprocess_parser.parse_args(preprocess_param) + preprocess.main(input_args) + + return gen_output + + +def cli_main(): + parser = rerank_options.get_reranking_parser() + args = options.parse_args_and_arch(parser) + gen_and_reprocess_nbest(args) + + +if __name__ == "__main__": + cli_main() diff --git a/examples/noisychannel/rerank_options.py b/examples/noisychannel/rerank_options.py new file mode 100644 index 0000000000000000000000000000000000000000..de91939e6635bdf33c9dc330116be07d9e8be6a2 --- /dev/null +++ b/examples/noisychannel/rerank_options.py @@ -0,0 +1,149 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq import options + + +def get_reranking_parser(default_task="translation"): + parser = options.get_parser("Generation and reranking", default_task) + add_reranking_args(parser) + return parser + + +def get_tuning_parser(default_task="translation"): + parser = options.get_parser("Reranking tuning", default_task) + add_reranking_args(parser) + add_tuning_args(parser) + return parser + + +def add_reranking_args(parser): + group = parser.add_argument_group("Reranking") + # fmt: off + group.add_argument('--score-model1', '-s1', type=str, metavar='FILE', required=True, + help='path to first model or ensemble of models for rescoring') + group.add_argument('--score-model2', '-s2', type=str, metavar='FILE', required=False, + help='path to second model or ensemble of models for rescoring') + group.add_argument('--num-rescore', '-n', type=int, metavar='N', default=10, + help='the number of candidate hypothesis to rescore') + group.add_argument('-bz', '--batch-size', type=int, metavar='N', default=128, + help='batch size for generating the nbest list') + group.add_argument('--gen-subset', default='test', metavar='SET', choices=['test', 'train', 'valid'], + help='data subset to generate (train, valid, test)') + group.add_argument('--gen-model', default=None, metavar='FILE', + help='the model to generate translations') + group.add_argument('-b1', '--backwards1', action='store_true', + help='whether or not the first model group is backwards') + group.add_argument('-b2', '--backwards2', action='store_true', + help='whether or not the second model group is backwards') + group.add_argument('-a', '--weight1', default=1, nargs='+', type=float, + help='the weight(s) of the first model') + group.add_argument('-b', '--weight2', default=1, nargs='+', type=float, + help='the weight(s) of the second model, or the gen model if using nbest from interactive.py') + group.add_argument('-c', '--weight3', default=1, nargs='+', type=float, + help='the weight(s) of the third model') + + # lm arguments + group.add_argument('-lm', '--language-model', default=None, metavar='FILE', + help='language model for target language to rescore translations') + group.add_argument('--lm-dict', default=None, metavar='FILE', + help='the dict of the language model for the target language') + group.add_argument('--lm-name', default=None, + help='the name of the language model for the target language') + group.add_argument('--lm-bpe-code', default=None, metavar='FILE', + help='the bpe code for the language model for the target language') + group.add_argument('--data-dir-name', default=None, + help='name of data directory') + group.add_argument('--lenpen', default=1, nargs='+', type=float, + help='length penalty: <1.0 favors shorter, >1.0 favors longer sentences') + group.add_argument('--score-dict-dir', default=None, + help='the directory with dictionaries for the scoring models') + group.add_argument('--right-to-left1', action='store_true', + help='whether the first model group is a right to left model') + group.add_argument('--right-to-left2', action='store_true', + help='whether the second model group is a right to left model') + group.add_argument('--post-process', '--remove-bpe', default='@@ ', + help='the bpe symbol, used for the bitext and LM') + group.add_argument('--prefix-len', default=None, type=int, + help='the length of the target prefix to use in rescoring (in terms of words wo bpe)') + group.add_argument('--sampling', action='store_true', + help='use sampling instead of beam search for generating n best list') + group.add_argument('--diff-bpe', action='store_true', + help='bpe for rescoring and nbest list not the same') + group.add_argument('--rescore-bpe-code', default=None, + help='bpe code for rescoring models') + group.add_argument('--nbest-list', default=None, + help='use predefined nbest list in interactive.py format') + group.add_argument('--write-hypos', default=None, + help='filename prefix to write hypos to') + group.add_argument('--ref-translation', default=None, + help='reference translation to use with nbest list from interactive.py') + group.add_argument('--backwards-score-dict-dir', default=None, + help='the directory with dictionaries for the backwards model,' + 'if None then it is assumed the fw and backwards models share dictionaries') + + # extra scaling args + group.add_argument('--gen-model-name', default=None, + help='the name of the models that generated the nbest list') + group.add_argument('--model1-name', default=None, + help='the name of the set for model1 group ') + group.add_argument('--model2-name', default=None, + help='the name of the set for model2 group') + group.add_argument('--shard-id', default=0, type=int, + help='the id of the shard to generate') + group.add_argument('--num-shards', default=1, type=int, + help='the number of shards to generate across') + group.add_argument('--all-shards', action='store_true', + help='use all shards') + group.add_argument('--target-prefix-frac', default=None, type=float, + help='the fraction of the target prefix to use in rescoring (in terms of words wo bpe)') + group.add_argument('--source-prefix-frac', default=None, type=float, + help='the fraction of the source prefix to use in rescoring (in terms of words wo bpe)') + group.add_argument('--normalize', action='store_true', + help='whether to normalize by src and target len') + # fmt: on + return group + + +def add_tuning_args(parser): + group = parser.add_argument_group("Tuning") + + group.add_argument( + "--lower-bound", + default=[-0.7], + nargs="+", + type=float, + help="lower bound of search space", + ) + group.add_argument( + "--upper-bound", + default=[3], + nargs="+", + type=float, + help="upper bound of search space", + ) + group.add_argument( + "--tune-param", + default=["lenpen"], + nargs="+", + choices=["lenpen", "weight1", "weight2", "weight3"], + help="the parameter(s) to tune", + ) + group.add_argument( + "--tune-subset", + default="valid", + choices=["valid", "test", "train"], + help="the subset to tune on ", + ) + group.add_argument( + "--num-trials", + default=1000, + type=int, + help="number of trials to do for random search", + ) + group.add_argument( + "--share-weights", action="store_true", help="share weight2 and weight 3" + ) + return group diff --git a/examples/noisychannel/rerank_score_bw.py b/examples/noisychannel/rerank_score_bw.py new file mode 100644 index 0000000000000000000000000000000000000000..b0bc913651bd76667e25c214acb70f2bca19e185 --- /dev/null +++ b/examples/noisychannel/rerank_score_bw.py @@ -0,0 +1,143 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import os +from contextlib import redirect_stdout + +from fairseq import options +from fairseq_cli import generate + +from examples.noisychannel import rerank_options, rerank_utils + + +def score_bw(args): + if args.backwards1: + scorer1_src = args.target_lang + scorer1_tgt = args.source_lang + else: + scorer1_src = args.source_lang + scorer1_tgt = args.target_lang + + if args.score_model2 is not None: + if args.backwards2: + scorer2_src = args.target_lang + scorer2_tgt = args.source_lang + else: + scorer2_src = args.source_lang + scorer2_tgt = args.target_lang + + rerank1_is_gen = ( + args.gen_model == args.score_model1 and args.source_prefix_frac is None + ) + rerank2_is_gen = ( + args.gen_model == args.score_model2 and args.source_prefix_frac is None + ) + + ( + pre_gen, + left_to_right_preprocessed_dir, + right_to_left_preprocessed_dir, + backwards_preprocessed_dir, + lm_preprocessed_dir, + ) = rerank_utils.get_directories( + args.data_dir_name, + args.num_rescore, + args.gen_subset, + args.gen_model_name, + args.shard_id, + args.num_shards, + args.sampling, + args.prefix_len, + args.target_prefix_frac, + args.source_prefix_frac, + ) + + score1_file = rerank_utils.rescore_file_name( + pre_gen, + args.prefix_len, + args.model1_name, + target_prefix_frac=args.target_prefix_frac, + source_prefix_frac=args.source_prefix_frac, + backwards=args.backwards1, + ) + + if args.score_model2 is not None: + score2_file = rerank_utils.rescore_file_name( + pre_gen, + args.prefix_len, + args.model2_name, + target_prefix_frac=args.target_prefix_frac, + source_prefix_frac=args.source_prefix_frac, + backwards=args.backwards2, + ) + + if args.right_to_left1: + rerank_data1 = right_to_left_preprocessed_dir + elif args.backwards1: + rerank_data1 = backwards_preprocessed_dir + else: + rerank_data1 = left_to_right_preprocessed_dir + + gen_param = ["--batch-size", str(128), "--score-reference", "--gen-subset", "train"] + if not rerank1_is_gen and not os.path.isfile(score1_file): + print("STEP 4: score the translations for model 1") + + model_param1 = [ + "--path", + args.score_model1, + "--source-lang", + scorer1_src, + "--target-lang", + scorer1_tgt, + ] + gen_model1_param = [rerank_data1] + gen_param + model_param1 + + gen_parser = options.get_generation_parser() + input_args = options.parse_args_and_arch(gen_parser, gen_model1_param) + + with open(score1_file, "w") as f: + with redirect_stdout(f): + generate.main(input_args) + + if ( + args.score_model2 is not None + and not os.path.isfile(score2_file) + and not rerank2_is_gen + ): + print("STEP 4: score the translations for model 2") + + if args.right_to_left2: + rerank_data2 = right_to_left_preprocessed_dir + elif args.backwards2: + rerank_data2 = backwards_preprocessed_dir + else: + rerank_data2 = left_to_right_preprocessed_dir + + model_param2 = [ + "--path", + args.score_model2, + "--source-lang", + scorer2_src, + "--target-lang", + scorer2_tgt, + ] + gen_model2_param = [rerank_data2] + gen_param + model_param2 + + gen_parser = options.get_generation_parser() + input_args = options.parse_args_and_arch(gen_parser, gen_model2_param) + + with open(score2_file, "w") as f: + with redirect_stdout(f): + generate.main(input_args) + + +def cli_main(): + parser = rerank_options.get_reranking_parser() + args = options.parse_args_and_arch(parser) + score_bw(args) + + +if __name__ == "__main__": + cli_main() diff --git a/examples/noisychannel/rerank_score_lm.py b/examples/noisychannel/rerank_score_lm.py new file mode 100644 index 0000000000000000000000000000000000000000..e80948d78b02561cbd09d72c319222105f41f6bb --- /dev/null +++ b/examples/noisychannel/rerank_score_lm.py @@ -0,0 +1,81 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import os + +from fairseq import options + +from examples.noisychannel import rerank_options, rerank_utils + + +def score_lm(args): + using_nbest = args.nbest_list is not None + ( + pre_gen, + left_to_right_preprocessed_dir, + right_to_left_preprocessed_dir, + backwards_preprocessed_dir, + lm_preprocessed_dir, + ) = rerank_utils.get_directories( + args.data_dir_name, + args.num_rescore, + args.gen_subset, + args.gen_model_name, + args.shard_id, + args.num_shards, + args.sampling, + args.prefix_len, + args.target_prefix_frac, + args.source_prefix_frac, + ) + + predictions_bpe_file = pre_gen + "/generate_output_bpe.txt" + if using_nbest: + print("Using predefined n-best list from interactive.py") + predictions_bpe_file = args.nbest_list + + gen_output = rerank_utils.BitextOutputFromGen( + predictions_bpe_file, bpe_symbol=args.post_process, nbest=using_nbest + ) + + if args.language_model is not None: + lm_score_file = rerank_utils.rescore_file_name( + pre_gen, args.prefix_len, args.lm_name, lm_file=True + ) + + if args.language_model is not None and not os.path.isfile(lm_score_file): + print("STEP 4.5: language modeling for P(T)") + if args.lm_bpe_code is None: + bpe_status = "no bpe" + elif args.lm_bpe_code == "shared": + bpe_status = "shared" + else: + bpe_status = "different" + + rerank_utils.lm_scoring( + lm_preprocessed_dir, + bpe_status, + gen_output, + pre_gen, + args.lm_dict, + args.lm_name, + args.language_model, + args.lm_bpe_code, + 128, + lm_score_file, + args.target_lang, + args.source_lang, + prefix_len=args.prefix_len, + ) + + +def cli_main(): + parser = rerank_options.get_reranking_parser() + args = options.parse_args_and_arch(parser) + score_lm(args) + + +if __name__ == "__main__": + cli_main() diff --git a/examples/noisychannel/rerank_tune.py b/examples/noisychannel/rerank_tune.py new file mode 100644 index 0000000000000000000000000000000000000000..b2e8b7594a370b2462f77252d54d7ef80e290f7c --- /dev/null +++ b/examples/noisychannel/rerank_tune.py @@ -0,0 +1,102 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import random + +import numpy as np +from fairseq import options + +from examples.noisychannel import rerank, rerank_options + + +def random_search(args): + param_values = [] + tuneable_parameters = ["lenpen", "weight1", "weight2", "weight3"] + initial_params = [args.lenpen, args.weight1, args.weight2, args.weight3] + for i, elem in enumerate(initial_params): + if type(elem) is not list: + initial_params[i] = [elem] + else: + initial_params[i] = elem + + tune_parameters = args.tune_param.copy() + for i in range(len(args.tune_param)): + assert args.upper_bound[i] >= args.lower_bound[i] + index = tuneable_parameters.index(args.tune_param[i]) + del tuneable_parameters[index] + del initial_params[index] + + tune_parameters += tuneable_parameters + param_values += initial_params + random.seed(args.seed) + + random_params = np.array( + [ + [ + random.uniform(args.lower_bound[i], args.upper_bound[i]) + for i in range(len(args.tune_param)) + ] + for k in range(args.num_trials) + ] + ) + set_params = np.array( + [ + [initial_params[i][0] for i in range(len(tuneable_parameters))] + for k in range(args.num_trials) + ] + ) + random_params = np.concatenate((random_params, set_params), 1) + + rerank_args = vars(args).copy() + if args.nbest_list: + rerank_args["gen_subset"] = "test" + else: + rerank_args["gen_subset"] = args.tune_subset + + for k in range(len(tune_parameters)): + rerank_args[tune_parameters[k]] = list(random_params[:, k]) + + if args.share_weights: + k = tune_parameters.index("weight2") + rerank_args["weight3"] = list(random_params[:, k]) + + rerank_args = argparse.Namespace(**rerank_args) + best_lenpen, best_weight1, best_weight2, best_weight3, best_score = rerank.rerank( + rerank_args + ) + rerank_args = vars(args).copy() + rerank_args["lenpen"] = [best_lenpen] + rerank_args["weight1"] = [best_weight1] + rerank_args["weight2"] = [best_weight2] + rerank_args["weight3"] = [best_weight3] + + # write the hypothesis from the valid set from the best trial + + if args.gen_subset != "valid": + rerank_args["gen_subset"] = "valid" + rerank_args = argparse.Namespace(**rerank_args) + rerank.rerank(rerank_args) + + # test with the best hyperparameters on gen subset + rerank_args = vars(args).copy() + rerank_args["gen_subset"] = args.gen_subset + rerank_args["lenpen"] = [best_lenpen] + rerank_args["weight1"] = [best_weight1] + rerank_args["weight2"] = [best_weight2] + rerank_args["weight3"] = [best_weight3] + rerank_args = argparse.Namespace(**rerank_args) + rerank.rerank(rerank_args) + + +def cli_main(): + parser = rerank_options.get_tuning_parser() + args = options.parse_args_and_arch(parser) + + random_search(args) + + +if __name__ == "__main__": + cli_main() diff --git a/examples/noisychannel/rerank_utils.py b/examples/noisychannel/rerank_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..2c6bf1b1afbb089cf5e84f720eb7a067479fbcbc --- /dev/null +++ b/examples/noisychannel/rerank_utils.py @@ -0,0 +1,850 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +import os +import re +import subprocess +from contextlib import redirect_stdout + +from fairseq import options +from fairseq_cli import eval_lm, preprocess + + +def reprocess(fle): + # takes in a file of generate.py translation generate_output + # returns a source dict and hypothesis dict, where keys are the ID num (as a string) + # and values and the corresponding source and translation. There may be several translations + # per source, so the values for hypothesis_dict are lists. + # parses output of generate.py + + with open(fle, "r") as f: + txt = f.read() + + """reprocess generate.py output""" + p = re.compile(r"[STHP][-]\d+\s*") + hp = re.compile(r"(\s*[-]?\d+[.]?\d+\s*)|(\s*(-inf)\s*)") + source_dict = {} + hypothesis_dict = {} + score_dict = {} + target_dict = {} + pos_score_dict = {} + lines = txt.split("\n") + + for line in lines: + line += "\n" + prefix = re.search(p, line) + if prefix is not None: + assert len(prefix.group()) > 2, "prefix id not found" + _, j = prefix.span() + id_num = prefix.group()[2:] + id_num = int(id_num) + line_type = prefix.group()[0] + if line_type == "H": + h_txt = line[j:] + hypo = re.search(hp, h_txt) + assert ( + hypo is not None + ), "regular expression failed to find the hypothesis scoring" + _, i = hypo.span() + score = hypo.group() + if id_num in hypothesis_dict: + hypothesis_dict[id_num].append(h_txt[i:]) + score_dict[id_num].append(float(score)) + else: + hypothesis_dict[id_num] = [h_txt[i:]] + score_dict[id_num] = [float(score)] + + elif line_type == "S": + source_dict[id_num] = line[j:] + elif line_type == "T": + target_dict[id_num] = line[j:] + elif line_type == "P": + pos_scores = (line[j:]).split() + pos_scores = [float(x) for x in pos_scores] + if id_num in pos_score_dict: + pos_score_dict[id_num].append(pos_scores) + else: + pos_score_dict[id_num] = [pos_scores] + + return source_dict, hypothesis_dict, score_dict, target_dict, pos_score_dict + + +def reprocess_nbest(fle): + """reprocess interactive.py output""" + with open(fle, "r") as f: + txt = f.read() + + source_dict = {} + hypothesis_dict = {} + score_dict = {} + target_dict = {} + pos_score_dict = {} + lines = txt.split("\n") + + hp = re.compile(r"[-]?\d+[.]?\d+") + j = -1 + + for _i, line in enumerate(lines): + line += "\n" + line_type = line[0] + + if line_type == "H": + hypo = re.search(hp, line) + _, start_index = hypo.span() + score = hypo.group() + if j in score_dict: + score_dict[j].append(float(score)) + hypothesis_dict[j].append(line[start_index:].strip("\t")) + else: + score_dict[j] = [float(score)] + hypothesis_dict[j] = [line[start_index:].strip("\t")] + elif line_type == "O": + j += 1 + source_dict[j] = line[2:] + # we don't have the targets for interactive.py + target_dict[j] = "filler" + + elif line_type == "P": + pos_scores = [float(pos_score) for pos_score in line.split()[1:]] + if j in pos_score_dict: + pos_score_dict[j].append(pos_scores) + else: + pos_score_dict[j] = [pos_scores] + + assert source_dict.keys() == hypothesis_dict.keys() + assert source_dict.keys() == pos_score_dict.keys() + assert source_dict.keys() == score_dict.keys() + + return source_dict, hypothesis_dict, score_dict, target_dict, pos_score_dict + + +def write_reprocessed( + sources, + hypos, + targets, + source_outfile, + hypo_outfile, + target_outfile, + right_to_left=False, + prefix_len=None, + bpe_symbol=None, + target_prefix_frac=None, + source_prefix_frac=None, +): + + """writes nbest hypothesis for rescoring""" + assert not ( + prefix_len is not None and target_prefix_frac is not None + ), "in writing reprocessed, only one type of prefix may be used" + assert not ( + prefix_len is not None and source_prefix_frac is not None + ), "in writing reprocessed, only one type of prefix may be used" + assert not ( + target_prefix_frac is not None and source_prefix_frac is not None + ), "in writing reprocessed, only one type of prefix may be used" + + with open(source_outfile, "w") as source_file, open( + hypo_outfile, "w" + ) as hypo_file, open(target_outfile, "w") as target_file: + + assert len(sources) == len(hypos), "sources and hypos list length mismatch" + if right_to_left: + for i in range(len(sources)): + for j in range(len(hypos[i])): + if prefix_len is None: + hypo_file.write(make_right_to_left(hypos[i][j]) + "\n") + else: + raise NotImplementedError() + source_file.write(make_right_to_left(sources[i]) + "\n") + target_file.write(make_right_to_left(targets[i]) + "\n") + else: + for i in sorted(sources.keys()): + for j in range(len(hypos[i])): + if prefix_len is not None: + shortened = ( + get_prefix_no_bpe(hypos[i][j], bpe_symbol, prefix_len) + + "\n" + ) + hypo_file.write(shortened) + source_file.write(sources[i]) + target_file.write(targets[i]) + elif target_prefix_frac is not None: + num_words, shortened, num_bpe_tokens = calc_length_from_frac( + hypos[i][j], target_prefix_frac, bpe_symbol + ) + shortened += "\n" + hypo_file.write(shortened) + source_file.write(sources[i]) + target_file.write(targets[i]) + elif source_prefix_frac is not None: + num_words, shortened, num_bpe_tokensn = calc_length_from_frac( + sources[i], source_prefix_frac, bpe_symbol + ) + shortened += "\n" + hypo_file.write(hypos[i][j]) + source_file.write(shortened) + target_file.write(targets[i]) + else: + hypo_file.write(hypos[i][j]) + source_file.write(sources[i]) + target_file.write(targets[i]) + + +def calc_length_from_frac(bpe_sentence, prefix_frac, bpe_symbol): + # return number of words, (not bpe tokens) that we want + no_bpe_sen = remove_bpe(bpe_sentence, bpe_symbol) + len_sen = len(no_bpe_sen.split()) + + num_words = math.ceil(len_sen * prefix_frac) + prefix = get_prefix_no_bpe(bpe_sentence, bpe_symbol, num_words) + num_bpe_tokens = len(prefix.split()) + return num_words, prefix, num_bpe_tokens + + +def get_prefix(sentence, prefix_len): + """assuming no bpe, gets the prefix of the sentence with prefix_len words""" + tokens = sentence.strip("\n").split() + if prefix_len >= len(tokens): + return sentence.strip("\n") + else: + return " ".join(tokens[:prefix_len]) + + +def get_prefix_no_bpe(sentence, bpe_symbol, prefix_len): + if bpe_symbol is None: + return get_prefix(sentence, prefix_len) + else: + return " ".join(get_prefix_from_len(sentence.split(), bpe_symbol, prefix_len)) + + +def get_prefix_from_len(sentence, bpe_symbol, prefix_len): + """get the prefix of sentence with bpe, with prefix len in terms of words, not bpe tokens""" + bpe_count = sum([bpe_symbol.strip(" ") in t for t in sentence[:prefix_len]]) + if bpe_count == 0: + return sentence[:prefix_len] + else: + return sentence[:prefix_len] + get_prefix_from_len( + sentence[prefix_len:], bpe_symbol, bpe_count + ) + + +def get_num_bpe_tokens_from_len(sentence, bpe_symbol, prefix_len): + """given a prefix length in terms of words, return the number of bpe tokens""" + prefix = get_prefix_no_bpe(sentence, bpe_symbol, prefix_len) + assert len(remove_bpe(prefix, bpe_symbol).split()) <= prefix_len + return len(prefix.split(" ")) + + +def make_right_to_left(line): + tokens = line.split() + tokens.reverse() + new_line = " ".join(tokens) + return new_line + + +def remove_bpe(line, bpe_symbol): + line = line.replace("\n", "") + line = (line + " ").replace(bpe_symbol, "").rstrip() + return line + ("\n") + + +def remove_bpe_dict(pred_dict, bpe_symbol): + new_dict = {} + for i in pred_dict: + if type(pred_dict[i]) == list: + new_list = [remove_bpe(elem, bpe_symbol) for elem in pred_dict[i]] + new_dict[i] = new_list + else: + new_dict[i] = remove_bpe(pred_dict[i], bpe_symbol) + return new_dict + + +def parse_bleu_scoring(line): + p = re.compile(r"(BLEU4 = )\d+[.]\d+") + res = re.search(p, line) + assert res is not None, line + return float(res.group()[8:]) + + +def get_full_from_prefix(hypo_prefix, hypos): + """given a hypo prefix, recover the first hypo from the list of complete hypos beginning with that prefix""" + for hypo in hypos: + hypo_prefix = hypo_prefix.strip("\n") + len_prefix = len(hypo_prefix) + if hypo[:len_prefix] == hypo_prefix: + return hypo + # no match found + raise Exception() + + +def get_score( + a, + b, + c, + target_len, + bitext_score1, + bitext_score2=None, + lm_score=None, + lenpen=None, + src_len=None, + tgt_len=None, + bitext1_backwards=False, + bitext2_backwards=False, + normalize=False, +): + if bitext1_backwards: + bitext1_norm = src_len + else: + bitext1_norm = tgt_len + if bitext_score2 is not None: + if bitext2_backwards: + bitext2_norm = src_len + else: + bitext2_norm = tgt_len + else: + bitext2_norm = 1 + bitext_score2 = 0 + if normalize: + score = ( + a * bitext_score1 / bitext1_norm + + b * bitext_score2 / bitext2_norm + + c * lm_score / src_len + ) + else: + score = a * bitext_score1 + b * bitext_score2 + c * lm_score + + if lenpen is not None: + score /= (target_len) ** float(lenpen) + + return score + + +class BitextOutput(object): + def __init__( + self, + output_file, + backwards, + right_to_left, + bpe_symbol, + prefix_len=None, + target_prefix_frac=None, + source_prefix_frac=None, + ): + """process output from rescoring""" + source, hypo, score, target, pos_score = reprocess(output_file) + if backwards: + self.hypo_fracs = source_prefix_frac + else: + self.hypo_fracs = target_prefix_frac + + # remove length penalty so we can use raw scores + score, num_bpe_tokens = get_score_from_pos( + pos_score, prefix_len, hypo, bpe_symbol, self.hypo_fracs, backwards + ) + source_lengths = {} + target_lengths = {} + + assert hypo.keys() == source.keys(), "key mismatch" + if backwards: + tmp = hypo + hypo = source + source = tmp + for i in source: + # since we are reranking, there should only be one hypo per source sentence + if backwards: + len_src = len(source[i][0].split()) + # record length without <eos> + if len_src == num_bpe_tokens[i][0] - 1: + source_lengths[i] = num_bpe_tokens[i][0] - 1 + else: + source_lengths[i] = num_bpe_tokens[i][0] + + target_lengths[i] = len(hypo[i].split()) + + source[i] = remove_bpe(source[i][0], bpe_symbol) + target[i] = remove_bpe(target[i], bpe_symbol) + hypo[i] = remove_bpe(hypo[i], bpe_symbol) + + score[i] = float(score[i][0]) + pos_score[i] = pos_score[i][0] + + else: + len_tgt = len(hypo[i][0].split()) + # record length without <eos> + if len_tgt == num_bpe_tokens[i][0] - 1: + target_lengths[i] = num_bpe_tokens[i][0] - 1 + else: + target_lengths[i] = num_bpe_tokens[i][0] + + source_lengths[i] = len(source[i].split()) + + if right_to_left: + source[i] = remove_bpe(make_right_to_left(source[i]), bpe_symbol) + target[i] = remove_bpe(make_right_to_left(target[i]), bpe_symbol) + hypo[i] = remove_bpe(make_right_to_left(hypo[i][0]), bpe_symbol) + score[i] = float(score[i][0]) + pos_score[i] = pos_score[i][0] + else: + assert ( + len(hypo[i]) == 1 + ), "expected only one hypothesis per source sentence" + source[i] = remove_bpe(source[i], bpe_symbol) + target[i] = remove_bpe(target[i], bpe_symbol) + hypo[i] = remove_bpe(hypo[i][0], bpe_symbol) + score[i] = float(score[i][0]) + pos_score[i] = pos_score[i][0] + + self.rescore_source = source + self.rescore_hypo = hypo + self.rescore_score = score + self.rescore_target = target + self.rescore_pos_score = pos_score + self.backwards = backwards + self.right_to_left = right_to_left + self.target_lengths = target_lengths + self.source_lengths = source_lengths + + +class BitextOutputFromGen(object): + def __init__( + self, + predictions_bpe_file, + bpe_symbol=None, + nbest=False, + prefix_len=None, + target_prefix_frac=None, + ): + if nbest: + ( + pred_source, + pred_hypo, + pred_score, + pred_target, + pred_pos_score, + ) = reprocess_nbest(predictions_bpe_file) + else: + pred_source, pred_hypo, pred_score, pred_target, pred_pos_score = reprocess( + predictions_bpe_file + ) + + assert len(pred_source) == len(pred_hypo) + assert len(pred_source) == len(pred_score) + assert len(pred_source) == len(pred_target) + assert len(pred_source) == len(pred_pos_score) + + # remove length penalty so we can use raw scores + pred_score, num_bpe_tokens = get_score_from_pos( + pred_pos_score, prefix_len, pred_hypo, bpe_symbol, target_prefix_frac, False + ) + + self.source = pred_source + self.target = pred_target + self.score = pred_score + self.pos_score = pred_pos_score + self.hypo = pred_hypo + self.target_lengths = {} + self.source_lengths = {} + + self.no_bpe_source = remove_bpe_dict(pred_source.copy(), bpe_symbol) + self.no_bpe_hypo = remove_bpe_dict(pred_hypo.copy(), bpe_symbol) + self.no_bpe_target = remove_bpe_dict(pred_target.copy(), bpe_symbol) + + # indexes to match those from the rescoring models + self.rescore_source = {} + self.rescore_target = {} + self.rescore_pos_score = {} + self.rescore_hypo = {} + self.rescore_score = {} + self.num_hypos = {} + self.backwards = False + self.right_to_left = False + + index = 0 + + for i in sorted(pred_source.keys()): + for j in range(len(pred_hypo[i])): + + self.target_lengths[index] = len(self.hypo[i][j].split()) + self.source_lengths[index] = len(self.source[i].split()) + + self.rescore_source[index] = self.no_bpe_source[i] + self.rescore_target[index] = self.no_bpe_target[i] + self.rescore_hypo[index] = self.no_bpe_hypo[i][j] + self.rescore_score[index] = float(pred_score[i][j]) + self.rescore_pos_score[index] = pred_pos_score[i][j] + self.num_hypos[index] = len(pred_hypo[i]) + index += 1 + + +def get_score_from_pos( + pos_score_dict, prefix_len, hypo_dict, bpe_symbol, hypo_frac, backwards +): + score_dict = {} + num_bpe_tokens_dict = {} + assert prefix_len is None or hypo_frac is None + for key in pos_score_dict: + score_dict[key] = [] + num_bpe_tokens_dict[key] = [] + for i in range(len(pos_score_dict[key])): + if prefix_len is not None and not backwards: + num_bpe_tokens = get_num_bpe_tokens_from_len( + hypo_dict[key][i], bpe_symbol, prefix_len + ) + score_dict[key].append(sum(pos_score_dict[key][i][:num_bpe_tokens])) + num_bpe_tokens_dict[key].append(num_bpe_tokens) + elif hypo_frac is not None: + num_words, shortened, hypo_prefix_len = calc_length_from_frac( + hypo_dict[key][i], hypo_frac, bpe_symbol + ) + score_dict[key].append(sum(pos_score_dict[key][i][:hypo_prefix_len])) + num_bpe_tokens_dict[key].append(hypo_prefix_len) + else: + score_dict[key].append(sum(pos_score_dict[key][i])) + num_bpe_tokens_dict[key].append(len(pos_score_dict[key][i])) + return score_dict, num_bpe_tokens_dict + + +class LMOutput(object): + def __init__( + self, + lm_score_file, + lm_dict=None, + prefix_len=None, + bpe_symbol=None, + target_prefix_frac=None, + ): + ( + lm_sentences, + lm_sen_scores, + lm_sen_pos_scores, + lm_no_bpe_sentences, + lm_bpe_tokens, + ) = parse_lm( + lm_score_file, + prefix_len=prefix_len, + bpe_symbol=bpe_symbol, + target_prefix_frac=target_prefix_frac, + ) + + self.sentences = lm_sentences + self.score = lm_sen_scores + self.pos_score = lm_sen_pos_scores + self.lm_dict = lm_dict + self.no_bpe_sentences = lm_no_bpe_sentences + self.bpe_tokens = lm_bpe_tokens + + +def parse_lm(input_file, prefix_len=None, bpe_symbol=None, target_prefix_frac=None): + """parse output of eval_lm""" + with open(input_file, "r") as f: + text = f.readlines() + text = text[7:] + cleaned_text = text[:-2] + + sentences = {} + sen_scores = {} + sen_pos_scores = {} + no_bpe_sentences = {} + num_bpe_tokens_dict = {} + for _i, line in enumerate(cleaned_text): + tokens = line.split() + if tokens[0].isdigit(): + line_id = int(tokens[0]) + scores = [float(x[1:-1]) for x in tokens[2::2]] + sentences[line_id] = " ".join(tokens[1::2][:-1]) + "\n" + if bpe_symbol is not None: + # exclude <eos> symbol to match output from generate.py + bpe_sen = " ".join(tokens[1::2][:-1]) + "\n" + no_bpe_sen = remove_bpe(bpe_sen, bpe_symbol) + no_bpe_sentences[line_id] = no_bpe_sen + + if prefix_len is not None: + num_bpe_tokens = get_num_bpe_tokens_from_len( + bpe_sen, bpe_symbol, prefix_len + ) + sen_scores[line_id] = sum(scores[:num_bpe_tokens]) + num_bpe_tokens_dict[line_id] = num_bpe_tokens + elif target_prefix_frac is not None: + num_words, shortened, target_prefix_len = calc_length_from_frac( + bpe_sen, target_prefix_frac, bpe_symbol + ) + sen_scores[line_id] = sum(scores[:target_prefix_len]) + num_bpe_tokens_dict[line_id] = target_prefix_len + else: + sen_scores[line_id] = sum(scores) + num_bpe_tokens_dict[line_id] = len(scores) + + sen_pos_scores[line_id] = scores + + return sentences, sen_scores, sen_pos_scores, no_bpe_sentences, num_bpe_tokens_dict + + +def get_directories( + data_dir_name, + num_rescore, + gen_subset, + fw_name, + shard_id, + num_shards, + sampling=False, + prefix_len=None, + target_prefix_frac=None, + source_prefix_frac=None, +): + nbest_file_id = ( + "nbest_" + + str(num_rescore) + + "_subset_" + + gen_subset + + "_fw_name_" + + fw_name + + "_shard_" + + str(shard_id) + + "_of_" + + str(num_shards) + ) + + if sampling: + nbest_file_id += "_sampling" + + # the directory containing all information for this nbest list + pre_gen = ( + os.path.join(os.path.dirname(__file__)) + + "/rerank_data/" + + data_dir_name + + "/" + + nbest_file_id + ) + # the directory to store the preprocessed nbest list, for left to right rescoring + left_to_right_preprocessed_dir = pre_gen + "/left_to_right_preprocessed" + if source_prefix_frac is not None: + left_to_right_preprocessed_dir = ( + left_to_right_preprocessed_dir + "/prefix_frac" + str(source_prefix_frac) + ) + # the directory to store the preprocessed nbest list, for right to left rescoring + right_to_left_preprocessed_dir = pre_gen + "/right_to_left_preprocessed" + # the directory to store the preprocessed nbest list, for backwards rescoring + backwards_preprocessed_dir = pre_gen + "/backwards" + if target_prefix_frac is not None: + backwards_preprocessed_dir = ( + backwards_preprocessed_dir + "/prefix_frac" + str(target_prefix_frac) + ) + elif prefix_len is not None: + backwards_preprocessed_dir = ( + backwards_preprocessed_dir + "/prefix_" + str(prefix_len) + ) + + # the directory to store the preprocessed nbest list, for rescoring with P(T) + lm_preprocessed_dir = pre_gen + "/lm_preprocessed" + + return ( + pre_gen, + left_to_right_preprocessed_dir, + right_to_left_preprocessed_dir, + backwards_preprocessed_dir, + lm_preprocessed_dir, + ) + + +def lm_scoring( + preprocess_directory, + bpe_status, + gen_output, + pre_gen, + cur_lm_dict, + cur_lm_name, + cur_language_model, + cur_lm_bpe_code, + batch_size, + lm_score_file, + target_lang, + source_lang, + prefix_len=None, +): + if prefix_len is not None: + assert ( + bpe_status == "different" + ), "bpe status must be different to use prefix len" + if bpe_status == "no bpe": + # run lm on output without bpe + write_reprocessed( + gen_output.no_bpe_source, + gen_output.no_bpe_hypo, + gen_output.no_bpe_target, + pre_gen + "/rescore_data_no_bpe.de", + pre_gen + "/rescore_data_no_bpe.en", + pre_gen + "/reference_file_no_bpe", + ) + + preprocess_lm_param = [ + "--only-source", + "--trainpref", + pre_gen + "/rescore_data_no_bpe." + target_lang, + "--srcdict", + cur_lm_dict, + "--destdir", + preprocess_directory, + ] + preprocess_parser = options.get_preprocessing_parser() + input_args = preprocess_parser.parse_args(preprocess_lm_param) + preprocess.main(input_args) + + eval_lm_param = [ + preprocess_directory, + "--path", + cur_language_model, + "--output-word-probs", + "--batch-size", + str(batch_size), + "--max-tokens", + "1024", + "--sample-break-mode", + "eos", + "--gen-subset", + "train", + ] + + eval_lm_parser = options.get_eval_lm_parser() + input_args = options.parse_args_and_arch(eval_lm_parser, eval_lm_param) + + with open(lm_score_file, "w") as f: + with redirect_stdout(f): + eval_lm.main(input_args) + + elif bpe_status == "shared": + preprocess_lm_param = [ + "--only-source", + "--trainpref", + pre_gen + "/rescore_data." + target_lang, + "--srcdict", + cur_lm_dict, + "--destdir", + preprocess_directory, + ] + preprocess_parser = options.get_preprocessing_parser() + input_args = preprocess_parser.parse_args(preprocess_lm_param) + preprocess.main(input_args) + + eval_lm_param = [ + preprocess_directory, + "--path", + cur_language_model, + "--output-word-probs", + "--batch-size", + str(batch_size), + "--sample-break-mode", + "eos", + "--gen-subset", + "train", + ] + + eval_lm_parser = options.get_eval_lm_parser() + input_args = options.parse_args_and_arch(eval_lm_parser, eval_lm_param) + + with open(lm_score_file, "w") as f: + with redirect_stdout(f): + eval_lm.main(input_args) + + elif bpe_status == "different": + rescore_file = pre_gen + "/rescore_data_no_bpe" + rescore_bpe = pre_gen + "/rescore_data_new_bpe" + + rescore_file += "." + rescore_bpe += "." + + write_reprocessed( + gen_output.no_bpe_source, + gen_output.no_bpe_hypo, + gen_output.no_bpe_target, + rescore_file + source_lang, + rescore_file + target_lang, + pre_gen + "/reference_file_no_bpe", + bpe_symbol=None, + ) + + # apply LM bpe to nbest list + bpe_src_param = [ + "-c", + cur_lm_bpe_code, + "--input", + rescore_file + target_lang, + "--output", + rescore_bpe + target_lang, + ] + subprocess.call( + [ + "python", + os.path.join( + os.path.dirname(__file__), "subword-nmt/subword_nmt/apply_bpe.py" + ), + ] + + bpe_src_param, + shell=False, + ) + # uncomment to use fastbpe instead of subword-nmt bpe + # bpe_src_param = [rescore_bpe+target_lang, rescore_file+target_lang, cur_lm_bpe_code] + # subprocess.call(["/private/home/edunov/fastBPE/fast", "applybpe"] + bpe_src_param, shell=False) + + preprocess_dir = preprocess_directory + + preprocess_lm_param = [ + "--only-source", + "--trainpref", + rescore_bpe + target_lang, + "--srcdict", + cur_lm_dict, + "--destdir", + preprocess_dir, + ] + preprocess_parser = options.get_preprocessing_parser() + input_args = preprocess_parser.parse_args(preprocess_lm_param) + preprocess.main(input_args) + + eval_lm_param = [ + preprocess_dir, + "--path", + cur_language_model, + "--output-word-probs", + "--batch-size", + str(batch_size), + "--max-tokens", + "1024", + "--sample-break-mode", + "eos", + "--gen-subset", + "train", + ] + + eval_lm_parser = options.get_eval_lm_parser() + input_args = options.parse_args_and_arch(eval_lm_parser, eval_lm_param) + + with open(lm_score_file, "w") as f: + with redirect_stdout(f): + eval_lm.main(input_args) + + +def rescore_file_name( + nbest_dir, + prefix_len, + scorer_name, + lm_file=False, + target_prefix_frac=None, + source_prefix_frac=None, + backwards=None, +): + if lm_file: + score_file = nbest_dir + "/lm_score_translations_model_" + scorer_name + ".txt" + else: + score_file = nbest_dir + "/" + scorer_name + "_score_translations.txt" + if backwards: + if prefix_len is not None: + score_file += "prefix_len" + str(prefix_len) + elif target_prefix_frac is not None: + score_file += "target_prefix_frac" + str(target_prefix_frac) + else: + if source_prefix_frac is not None: + score_file += "source_prefix_frac" + str(source_prefix_frac) + return score_file diff --git a/examples/nonautoregressive_translation/README.md b/examples/nonautoregressive_translation/README.md new file mode 100644 index 0000000000000000000000000000000000000000..8793e225c99732c42c9c19e22075cde37c73341d --- /dev/null +++ b/examples/nonautoregressive_translation/README.md @@ -0,0 +1,146 @@ +# Non-autoregressive Neural Machine Translation (NAT) + +This page mainly includes instructions for reproducing results from the following papers +* [Levenshtein Transformer (Gu et al., 2019)](https://arxiv.org/abs/1905.11006). +* [Understanding Knowledge Distillation in Non-autoregressive Machine Translation (Zhou et al., 2019)](https://arxiv.org/abs/1911.02727). + +We also provided our own implementations for several popular non-autoregressive-based models as reference:<br> +* [Non-Autoregressive Neural Machine Translation (Gu et al., 2017)](https://arxiv.org/abs/1711.02281)<br> +* [Deterministic Non-Autoregressive Neural Sequence Modeling by Iterative Refinement (Lee et al., 2018)](https://arxiv.org/abs/1802.06901)<br> +* [Insertion Transformer: Flexible Sequence Generation via Insertion Operations (Stern et al., 2019)](https://arxiv.org/abs/1902.03249)<br> +* [Mask-Predict: Parallel Decoding of Conditional Masked Language Models (Ghazvininejad et al., 2019)](https://arxiv.org/abs/1904.09324v2)<br> +* [Fast Structured Decoding for Sequence Models (Sun et al., 2019)](https://arxiv.org/abs/1910.11555) + +## Dataset + +First, follow the [instructions to download and preprocess the WMT'14 En-De dataset](../translation#wmt14-english-to-german-convolutional). +Make sure to learn a joint vocabulary by passing the `--joined-dictionary` option to `fairseq-preprocess`. + +### Knowledge Distillation +Following [Gu et al. 2019](https://arxiv.org/abs/1905.11006), [knowledge distillation](https://arxiv.org/abs/1606.07947) from an autoregressive model can effectively simplify the training data distribution, which is sometimes essential for NAT-based models to learn good translations. +The easiest way of performing distillation is to follow the [instructions of training a standard transformer model](../translation) on the same data, and then decode the training set to produce a distillation dataset for NAT. + +### Download +We also provided the preprocessed [original](http://dl.fbaipublicfiles.com/nat/original_dataset.zip) and [distillation](http://dl.fbaipublicfiles.com/nat/distill_dataset.zip) datasets. Please build the binarized dataset on your own. + + +## Train a model + +Then we can train a nonautoregressive model using the `translation_lev` task and a new criterion `nat_loss`. +Use the `--noise` flag to specify the input noise used on the target sentences. +In default, we run the task for *Levenshtein Transformer*, with `--noise='random_delete'`. Full scripts to run other models can also be found [here](./scripts.md). + +The following command will train a *Levenshtein Transformer* on the binarized dataset. + +```bash +fairseq-train \ + data-bin/wmt14_en_de_distill \ + --save-dir checkpoints \ + --ddp-backend=legacy_ddp \ + --task translation_lev \ + --criterion nat_loss \ + --arch levenshtein_transformer \ + --noise random_delete \ + --share-all-embeddings \ + --optimizer adam --adam-betas '(0.9,0.98)' \ + --lr 0.0005 --lr-scheduler inverse_sqrt \ + --stop-min-lr '1e-09' --warmup-updates 10000 \ + --warmup-init-lr '1e-07' --label-smoothing 0.1 \ + --dropout 0.3 --weight-decay 0.01 \ + --decoder-learned-pos \ + --encoder-learned-pos \ + --apply-bert-init \ + --log-format 'simple' --log-interval 100 \ + --fixed-validation-seed 7 \ + --max-tokens 8000 \ + --save-interval-updates 10000 \ + --max-update 300000 +``` + +## Translate + +Once a model is trained, we can generate translations using an `iterative_refinement_generator` which will based on the model's initial output and iteratively read and greedily refine the translation until (1) the model predicts the same translations for two consecutive iterations; or (2) the generator reaches the maximum iterations (`--iter-decode-max-iter`). Use `--print-step` to check the actual # of iteration for each sentence. + +For *Levenshtein Transformer*, it sometimes helps to apply a `--iter-decode-eos-penalty` (typically, 0~3) to penalize the model finishing generation too early and generating too short translations. + +For example, to generate with `--iter-decode-max-iter=9`: +```bash +fairseq-generate \ + data-bin/wmt14_en_de_distill \ + --gen-subset test \ + --task translation_lev \ + --path checkpoints/checkpoint_best.pt \ + --iter-decode-max-iter 9 \ + --iter-decode-eos-penalty 0 \ + --beam 1 --remove-bpe \ + --print-step \ + --batch-size 400 +``` +In the end of the generation, we can see the tokenized BLEU score for the translation. + +## Advanced Decoding Methods +### Ensemble +The NAT models use special implementations of [ensembling](https://github.com/fairinternal/fairseq-py/blob/b98d88da52f2f21f1b169bab8c70c1c4ca19a768/fairseq/sequence_generator.py#L522) to support iterative refinement and a variety of parallel operations in different models, while it shares the same API as standard autoregressive models as follows: +```bash +fairseq-generate \ + data-bin/wmt14_en_de_distill \ + --gen-subset test \ + --task translation_lev \ + --path checkpoint_1.pt:checkpoint_2.pt:checkpoint_3.pt \ + --iter-decode-max-iter 9 \ + --iter-decode-eos-penalty 0 \ + --beam 1 --remove-bpe \ + --print-step \ + --batch-size 400 +``` +We use ``:`` to split multiple models. Note that, not all NAT models support ensembling for now. + + +### Length-beam +For models that predict lengths before decoding (e.g. the vanilla NAT, Mask-Predict, etc), it is possible to improve the translation quality by varying the target lengths around the predicted value, and translating the same example multiple times in parallel. We can select the best translation with the highest scores defined by your model's output. + +Note that, not all models support length beams. For models which dynamically change the lengths (e.g. *Insertion Transformer*, *Levenshtein Transformer*), the same trick does not apply. + +### Re-ranking +If the model generates multiple translations with length beam, we can also introduce an autoregressive model to rerank the translations considering scoring from an autoregressive model is much faster than decoding from that. + +For example, to generate translations with length beam and reranking, +```bash +fairseq-generate \ + data-bin/wmt14_en_de_distill \ + --gen-subset test \ + --task translation_lev \ + --path checkpoints/checkpoint_best.pt:at_checkpoints/checkpoint_best.pt \ + --iter-decode-max-iter 9 \ + --iter-decode-eos-penalty 0 \ + --iter-decode-with-beam 9 \ + --iter-decode-with-external-reranker \ + --beam 1 --remove-bpe \ + --print-step \ + --batch-size 100 +``` +Note that we need to make sure the autoregressive model shares the same vocabulary as our target non-autoregressive model. + + +## Citation + +```bibtex +@incollection{NIPS2019_9297, + title = {Levenshtein Transformer}, + author = {Gu, Jiatao and Wang, Changhan and Zhao, Junbo}, + booktitle = {Advances in Neural Information Processing Systems 32}, + editor = {H. Wallach and H. Larochelle and A. Beygelzimer and F. d\textquotesingle Alch\'{e}-Buc and E. Fox and R. Garnett}, + pages = {11179--11189}, + year = {2019}, + publisher = {Curran Associates, Inc.}, + url = {http://papers.nips.cc/paper/9297-levenshtein-transformer.pdf} +} +``` +```bibtex +@article{zhou2019understanding, + title={Understanding Knowledge Distillation in Non-autoregressive Machine Translation}, + author={Zhou, Chunting and Neubig, Graham and Gu, Jiatao}, + journal={arXiv preprint arXiv:1911.02727}, + year={2019} +} +``` diff --git a/examples/nonautoregressive_translation/scripts.md b/examples/nonautoregressive_translation/scripts.md new file mode 100644 index 0000000000000000000000000000000000000000..9d3d7b67dc08440b5f4d1c5a7ffcd4bd6e76c14f --- /dev/null +++ b/examples/nonautoregressive_translation/scripts.md @@ -0,0 +1,179 @@ +# Examples of Training scripts for Non-autoregressive Machine Translation models + +### Non-autoregressive Transformer (NAT, Gu et al., 2017) +Note that we need to have an additional module to perform "length prediction" (`--length-loss-factor`) before generating the whole sequence. +```bash +fairseq-train \ + data-bin/wmt14_en_de_distill \ + --save-dir checkpoints \ + --ddp-backend=legacy_ddp \ + --task translation_lev \ + --criterion nat_loss \ + --arch nonautoregressive_transformer \ + --noise full_mask \ + --share-all-embeddings \ + --optimizer adam --adam-betas '(0.9,0.98)' \ + --lr 0.0005 --lr-scheduler inverse_sqrt \ + --stop-min-lr '1e-09' --warmup-updates 10000 \ + --warmup-init-lr '1e-07' --label-smoothing 0.1 \ + --dropout 0.3 --weight-decay 0.01 \ + --decoder-learned-pos \ + --encoder-learned-pos \ + --pred-length-offset \ + --length-loss-factor 0.1 \ + --apply-bert-init \ + --log-format 'simple' --log-interval 100 \ + --fixed-validation-seed 7 \ + --max-tokens 8000 \ + --save-interval-updates 10000 \ + --max-update 300000 +``` + +### Fast Structured Decoding for Sequence Models (NAT-CRF, Sun et al., 2019) +Note that we implemented a low-rank appromixated CRF model by setting `--crf-lowrank-approx=32` and `--crf-beam-approx=64` as discribed in the original paper. All other settings are the same as the vanilla NAT model. +```bash +fairseq-train \ + data-bin/wmt14_en_de_distill \ + --save-dir checkpoints \ + --ddp-backend=legacy_ddp \ + --task translation_lev \ + --criterion nat_loss \ + --arch nacrf_transformer \ + --noise full_mask \ + --share-all-embeddings \ + --optimizer adam --adam-betas '(0.9,0.98)' \ + --lr 0.0005 --lr-scheduler inverse_sqrt \ + --stop-min-lr '1e-09' --warmup-updates 10000 \ + --warmup-init-lr '1e-07' --label-smoothing 0.1 \ + --dropout 0.3 --weight-decay 0.01 \ + --decoder-learned-pos \ + --encoder-learned-pos \ + --pred-length-offset \ + --length-loss-factor 0.1 \ + --word-ins-loss-factor 0.5 \ + --crf-lowrank-approx 32 \ + --crf-beam-approx 64 \ + --apply-bert-init \ + --log-format 'simple' --log-interval 100 \ + --fixed-validation-seed 7 \ + --max-tokens 8000 \ + --save-interval-updates 10000 \ + --max-update 300000 +``` + + +### Non-autoregressive Transformer with Iterative Refinement (iNAT, Lee et al., 2018) +Note that `--train-step` means how many iterations of refinement we used during training, and `--dae-ratio` controls the ratio of denoising auto-encoder training described in the original paper. +```bash +fairseq-train \ + data-bin/wmt14_en_de_distill \ + --save-dir checkpoints \ + --ddp-backend=legacy_ddp \ + --task translation_lev \ + --criterion nat_loss \ + --arch iterative_nonautoregressive_transformer \ + --noise full_mask \ + --share-all-embeddings \ + --optimizer adam --adam-betas '(0.9,0.98)' \ + --lr 0.0005 --lr-scheduler inverse_sqrt \ + --stop-min-lr '1e-09' --warmup-updates 10000 \ + --warmup-init-lr '1e-07' --label-smoothing 0.1 \ + --dropout 0.3 --weight-decay 0.01 \ + --decoder-learned-pos \ + --encoder-learned-pos \ + --pred-length-offset \ + --length-loss-factor 0.1 \ + --train-step 4 \ + --dae-ratio 0.5 \ + --stochastic-approx \ + --apply-bert-init \ + --log-format 'simple' --log-interval 100 \ + --fixed-validation-seed 7 \ + --max-tokens 8000 \ + --save-interval-updates 10000 \ + --max-update 300000 +``` + +### Insertion Transformer (InsT, Stern et al., 2019) +Note that we need to specify the "slot-loss" (uniform or balanced tree) described in the original paper. Here we use `--label-tau` to control the temperature. + +```bash +fairseq-train \ + data-bin/wmt14_en_de_distill \ + --save-dir checkpoints \ + --ddp-backend=legacy_ddp \ + --task translation_lev \ + --criterion nat_loss \ + --arch insertion_transformer \ + --noise random_delete \ + --share-all-embeddings \ + --optimizer adam --adam-betas '(0.9,0.98)' \ + --lr 0.0005 --lr-scheduler inverse_sqrt \ + --stop-min-lr '1e-09' --warmup-updates 10000 \ + --warmup-init-lr '1e-07' --label-smoothing 0.1 \ + --dropout 0.3 --weight-decay 0.01 \ + --decoder-learned-pos \ + --encoder-learned-pos \ + --apply-bert-init \ + --log-format 'simple' --log-interval 100 \ + --fixed-validation-seed 7 \ + --max-tokens 8000 \ + --save-interval-updates 10000 \ + --max-update 300000 +``` + + +### Mask Predict (CMLM, Ghazvininejad et al., 2019) +```bash +fairseq-train \ + data-bin/wmt14_en_de_distill \ + --save-dir checkpoints \ + --ddp-backend=legacy_ddp \ + --task translation_lev \ + --criterion nat_loss \ + --arch cmlm_transformer \ + --noise random_mask \ + --share-all-embeddings \ + --optimizer adam --adam-betas '(0.9,0.98)' \ + --lr 0.0005 --lr-scheduler inverse_sqrt \ + --stop-min-lr '1e-09' --warmup-updates 10000 \ + --warmup-init-lr '1e-07' --label-smoothing 0.1 \ + --dropout 0.3 --weight-decay 0.01 \ + --decoder-learned-pos \ + --encoder-learned-pos \ + --apply-bert-init \ + --log-format 'simple' --log-interval 100 \ + --fixed-validation-seed 7 \ + --max-tokens 8000 \ + --save-interval-updates 10000 \ + --max-update 300000 +``` + + + + +### Levenshtein Transformer (LevT, Gu et al., 2019) +```bash +fairseq-train \ + data-bin/wmt14_en_de_distill \ + --save-dir checkpoints \ + --ddp-backend=legacy_ddp \ + --task translation_lev \ + --criterion nat_loss \ + --arch levenshtein_transformer \ + --noise random_delete \ + --share-all-embeddings \ + --optimizer adam --adam-betas '(0.9,0.98)' \ + --lr 0.0005 --lr-scheduler inverse_sqrt \ + --stop-min-lr '1e-09' --warmup-updates 10000 \ + --warmup-init-lr '1e-07' --label-smoothing 0.1 \ + --dropout 0.3 --weight-decay 0.01 \ + --decoder-learned-pos \ + --encoder-learned-pos \ + --apply-bert-init \ + --log-format 'simple' --log-interval 100 \ + --fixed-validation-seed 7 \ + --max-tokens 8000 \ + --save-interval-updates 10000 \ + --max-update 300000 +``` diff --git a/examples/paraphraser/README.md b/examples/paraphraser/README.md new file mode 100644 index 0000000000000000000000000000000000000000..3810311f30f99f0a07fd8e5d3723bffeba9948c3 --- /dev/null +++ b/examples/paraphraser/README.md @@ -0,0 +1,46 @@ +# Paraphrasing with round-trip translation and mixture of experts + +Machine translation models can be used to paraphrase text by translating it to +an intermediate language and back (round-trip translation). + +This example shows how to paraphrase text by first passing it to an +English-French translation model, followed by a French-English [mixture of +experts translation model](/examples/translation_moe). + +##### 0. Setup + +Clone fairseq from source and install necessary dependencies: +```bash +git clone https://github.com/pytorch/fairseq.git +cd fairseq +pip install --editable . +pip install sacremoses sentencepiece +``` + +##### 1. Download models +```bash +wget https://dl.fbaipublicfiles.com/fairseq/models/paraphraser.en-fr.tar.gz +wget https://dl.fbaipublicfiles.com/fairseq/models/paraphraser.fr-en.hMoEup.tar.gz +tar -xzvf paraphraser.en-fr.tar.gz +tar -xzvf paraphraser.fr-en.hMoEup.tar.gz +``` + +##### 2. Paraphrase +```bash +python examples/paraphraser/paraphrase.py \ + --en2fr paraphraser.en-fr \ + --fr2en paraphraser.fr-en.hMoEup +# Example input: +# The new date for the Games, postponed for a year in response to the coronavirus pandemic, gives athletes time to recalibrate their training schedules. +# Example outputs: +# Delayed one year in response to the coronavirus pandemic, the new date of the Games gives athletes time to rebalance their training schedule. +# The new date of the Games, which was rescheduled one year in response to the coronavirus (CV) pandemic, gives athletes time to rebalance their training schedule. +# The new date of the Games, postponed one year in response to the coronavirus pandemic, provides athletes with time to rebalance their training schedule. +# The Games' new date, postponed one year in response to the coronavirus pandemic, gives athletes time to rebalance their training schedule. +# The new Games date, postponed one year in response to the coronavirus pandemic, gives the athletes time to rebalance their training schedule. +# The new date of the Games, which was postponed one year in response to the coronavirus pandemic, gives the athletes time to rebalance their training schedule. +# The new date of the Games, postponed one year in response to the coronavirus pandemic, gives athletes time to rebalance their training schedule. +# The new date of the Games, postponed one year in response to the coronavirus pandemic, gives athletes time to re-balance their training schedule. +# The new date of the Games, postponed one year in response to the coronavirus pandemic, gives the athletes time to rebalance their schedule of training. +# The new date of the Games, postponed one year in response to the pandemic of coronavirus, gives the athletes time to rebalance their training schedule. +``` diff --git a/examples/paraphraser/paraphrase.py b/examples/paraphraser/paraphrase.py new file mode 100644 index 0000000000000000000000000000000000000000..d3422fb3db9a381b73a854d2379df214ebe544a2 --- /dev/null +++ b/examples/paraphraser/paraphrase.py @@ -0,0 +1,85 @@ +#!/usr/bin/env python3 -u + +import argparse +import fileinput +import logging +import os +import sys + +from fairseq.models.transformer import TransformerModel + + +logging.getLogger().setLevel(logging.INFO) + + +def main(): + parser = argparse.ArgumentParser(description="") + parser.add_argument("--en2fr", required=True, help="path to en2fr model") + parser.add_argument( + "--fr2en", required=True, help="path to fr2en mixture of experts model" + ) + parser.add_argument( + "--user-dir", help="path to fairseq examples/translation_moe/src directory" + ) + parser.add_argument( + "--num-experts", + type=int, + default=10, + help="(keep at 10 unless using a different model)", + ) + parser.add_argument( + "files", + nargs="*", + default=["-"], + help='input files to paraphrase; "-" for stdin', + ) + args = parser.parse_args() + + if args.user_dir is None: + args.user_dir = os.path.join( + os.path.dirname(os.path.dirname(os.path.abspath(__file__))), # examples/ + "translation_moe", + "src", + ) + if os.path.exists(args.user_dir): + logging.info("found user_dir:" + args.user_dir) + else: + raise RuntimeError( + "cannot find fairseq examples/translation_moe/src " + "(tried looking here: {})".format(args.user_dir) + ) + + logging.info("loading en2fr model from:" + args.en2fr) + en2fr = TransformerModel.from_pretrained( + model_name_or_path=args.en2fr, + tokenizer="moses", + bpe="sentencepiece", + ).eval() + + logging.info("loading fr2en model from:" + args.fr2en) + fr2en = TransformerModel.from_pretrained( + model_name_or_path=args.fr2en, + tokenizer="moses", + bpe="sentencepiece", + user_dir=args.user_dir, + task="translation_moe", + ).eval() + + def gen_paraphrases(en): + fr = en2fr.translate(en) + return [ + fr2en.translate(fr, inference_step_args={"expert": i}) + for i in range(args.num_experts) + ] + + logging.info("Type the input sentence and press return:") + for line in fileinput.input(args.files): + line = line.strip() + if len(line) == 0: + continue + for paraphrase in gen_paraphrases(line): + print(paraphrase) + + +if __name__ == "__main__": + main() diff --git a/examples/pay_less_attention_paper/README.md b/examples/pay_less_attention_paper/README.md new file mode 100644 index 0000000000000000000000000000000000000000..5adab11f4dc3461f9e7126ac391b04e703616e6b --- /dev/null +++ b/examples/pay_less_attention_paper/README.md @@ -0,0 +1,176 @@ +# Pay Less Attention with Lightweight and Dynamic Convolutions (Wu et al., 2019) + +This page contains pointers to pre-trained models as well as instructions on how to train new models for [our paper](https://arxiv.org/abs/1901.10430). + +## Citation: +```bibtex +@inproceedings{wu2018pay, + title = {Pay Less Attention with Lightweight and Dynamic Convolutions}, + author = {Felix Wu and Angela Fan and Alexei Baevski and Yann Dauphin and Michael Auli}, + booktitle = {International Conference on Learning Representations}, + year = {2019}, + url = {https://arxiv.org/abs/1901.10430}, +} +``` + +## Translation + +### Pre-trained models +For some datasets we release models without GLUs which are faster at inference. + +Model | Description | Dataset | Download +---|---|---|--- +`lightconv.no_glu.iwslt14.de-en` | LightConv (without GLUs) | [IWSLT14 German-English](https://wit3.fbk.eu/archive/2014-01/texts/de/en/de-en.tgz) | model: <br> [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/dynamicconv/iwslt14.de-en.lightconv.tar.gz) <br> IWSLT14 test: <br> [download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/data/iwslt14.de-en.test.tar.bz2) +`dynamicconv.no_glu.iwslt14.de-en` | DynamicConv (without GLUs) | [IWSLT14 German-English](https://wit3.fbk.eu/archive/2014-01/texts/de/en/de-en.tgz) | model: <br> [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/dynamicconv/iwslt14.de-en.dynamicconv.tar.gz) <br> IWSLT14 test: <br> [download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/data/iwslt14.de-en.test.tar.bz2) +`lightconv.no_glu.wmt16.en-de` | LightConv (without GLUs) | [WMT16 English-German](https://drive.google.com/uc?export=download&id=0B_bZck-ksdkpM25jRUN2X2UxMm8) | model: <br> [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/dynamicconv/wmt16.en-de.joined-dict.lightconv.tar.gz) <br> newstest2014 (shared vocab): <br> [download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/data/wmt16.en-de.joined-dict.newstest2014.tar.bz2) +`dynamicconv.no_glu.wmt16.en-de` | DynamicConv (without GLUs) | [WMT16 English-German](https://drive.google.com/uc?export=download&id=0B_bZck-ksdkpM25jRUN2X2UxMm8) | model: <br> [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/dynamicconv/wmt16.en-de.joined-dict.dynamicconv.tar.gz) <br> newstest2014 (shared vocab): <br> [download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/data/wmt16.en-de.joined-dict.newstest2014.tar.bz2) +`lightconv.glu.wmt16.en-de` | LightConv | [WMT16 English-German](https://drive.google.com/uc?export=download&id=0B_bZck-ksdkpM25jRUN2X2UxMm8) | model: <br> [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/dynamicconv/wmt16.en-de.joined-dict.lightconv-glu.tar.gz) <br> newstest2014 (shared vocab): <br> [download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/data/wmt16.en-de.joined-dict.newstest2014.tar.bz2) +`dynamicconv.glu.wmt16.en-de` | DynamicConv | [WMT16 English-German](https://drive.google.com/uc?export=download&id=0B_bZck-ksdkpM25jRUN2X2UxMm8) | model: <br> [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/dynamicconv/wmt16.en-de.joined-dict.dynamicconv-glu.tar.gz) <br> newstest2014 (shared vocab): <br> [download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/data/wmt16.en-de.joined-dict.newstest2014.tar.bz2) +`lightconv.glu.wmt14.en-fr` | LightConv | [WMT14 English-French](http://statmt.org/wmt14/translation-task.html#Download) | model: <br> [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/dynamicconv/wmt14.en-fr.joined-dict.lightconv-glu.tar.gz) <br> newstest2014: <br> [download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/data/wmt14.en-fr.joined-dict.newstest2014.tar.bz2) +`dynamicconv.glu.wmt14.en-fr` | DynamicConv | [WMT14 English-French](http://statmt.org/wmt14/translation-task.html#Download) | model: <br> [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/dynamicconv/wmt14.en-fr.joined-dict.dynamicconv-glu.tar.gz) <br> newstest2014: <br> [download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/data/wmt14.en-fr.joined-dict.newstest2014.tar.bz2) +`lightconv.glu.wmt17.zh-en` | LightConv | [WMT17 Chinese-English](http://statmt.org/wmt17/translation-task.html#Download) | model: <br> [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/dynamicconv/wmt17.zh-en.lightconv-glu.tar.gz) <br> newstest2017: <br> [download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/data/wmt17.zh-en.newstest2017.tar.bz2) +`dynamicconv.glu.wmt17.zh-en` | DynamicConv | [WMT17 Chinese-English](http://statmt.org/wmt17/translation-task.html#Download) | model: <br> [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/dynamicconv/wmt17.zh-en.dynamicconv-glu.tar.gz) <br> newstest2017: <br> [download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/data/wmt17.zh-en.newstest2017.tar.bz2) + +### Memory-Efficient CUDA Kernels + +Since the PyTorch implementations of Light/Dynamic conv are quite memory intensive, we have developed CUDA kernels that implement the light and dynamic convolution operator in a memory-efficient and performant manner. For large sequence lengths, these kernels save about 50% memory compared to the PyTorch equivalent. + +To install the kernels, use the commands below. Once installed, they will automatically be used in place of the PyTorch implementations whenever a light or dynamic convolution is used. + +```sh +# to install lightconv +cd fairseq/modules/lightconv_layer +python cuda_function_gen.py +python setup.py install + +# to install dynamicconv +cd fairseq/modules/dynamicconv_layer +python cuda_function_gen.py +python setup.py install +``` + +### Example usage (torch.hub) + +We require a few additional Python dependencies for preprocessing: +```bash +pip install sacremoses subword_nmt +``` + +Interactive translation via PyTorch Hub: +```python +import torch + +# List available models +torch.hub.list('pytorch/fairseq') # [..., 'lightconv.glu.wmt17.zh-en', ... ] + +# Load a transformer trained on WMT'16 En-De +zh2en = torch.hub.load('pytorch/fairseq', 'lightconv.glu.wmt17.zh-en', tokenizer='moses', bpe='subword_nmt') + +# The underlying model is available under the *models* attribute +assert isinstance(zh2en.models[0], fairseq.models.lightconv.LightConvModel) + +# Translate a sentence +zh2en.translate('你好 世界') +# 'Hello World' +``` + +Loading custom models: +```python +from fairseq.models.lightconv import LightConvModel +en2fr = LightConvModel.from_pretrained( + '/path/to/checkpoints', + checkpoint_file='checkpoint_best.pt', + data_name_or_path='data-bin/wmt14_en_fr', + bpe='subword_nmt', + bpe_codes='data-bin/wmt14_en_fr/en.code' +) +en2fr.translate('Hello world!') +# 'Bonjour le monde' +``` + +### Preprocessing the training datasets + +Please follow the instructions in [`examples/translation/README.md`](../translation/README.md) to preprocess the data. + +### Training and evaluation options: +To use the model without GLU, please set `--encoder-glu 0 --decoder-glu 0`. +For LightConv, please use `--encoder-conv-type lightweight --decoder-conv-type lightweight`, otherwise the default is DynamicConv. +For best BLEU results, lenpen may need to be manually tuned. + +To use the CUDA kernels, first install the PyTorch modules using the commands +above. Once the CUDA modules are installed, they will automatically be used +instead of the PyTorch modules. + +### IWSLT14 De-En +Training and evaluating DynamicConv (without GLU) on a GPU: +```sh +# Training +SAVE="save/dynamic_conv_iwslt" +mkdir -p $SAVE +CUDA_VISIBLE_DEVICES=0 $(which fairseq-train) data-bin/iwslt14.tokenized.de-en \ + --clip-norm 0 --optimizer adam --lr 0.0005 \ + --source-lang de --target-lang en --max-tokens 4000 --no-progress-bar \ + --log-interval 100 --stop-min-lr '1e-09' --weight-decay 0.0001 \ + --criterion label_smoothed_cross_entropy --label-smoothing 0.1 \ + --lr-scheduler inverse_sqrt \ + --ddp-backend=legacy_ddp \ + --max-update 50000 --warmup-updates 4000 --warmup-init-lr '1e-07' \ + --adam-betas '(0.9, 0.98)' --keep-last-epochs 10 \ + -a lightconv_iwslt_de_en --save-dir $SAVE \ + --dropout 0.3 --attention-dropout 0.1 --weight-dropout 0.1 \ + --encoder-glu 0 --decoder-glu 0 +python scripts/average_checkpoints.py --inputs $SAVE \ + --num-epoch-checkpoints 10 --output "${SAVE}/checkpoint_last10_avg.pt" + +# Evaluation +CUDA_VISIBLE_DEVICES=0 fairseq-generate data-bin/iwslt14.tokenized.de-en --path "${SAVE}/checkpoint_last10_avg.pt" --batch-size 128 --beam 4 --remove-bpe --lenpen 1 --gen-subset test --quiet +``` + +### WMT16 En-De +Training and evaluating DynamicConv (with GLU) on WMT16 En-De using cosine scheduler on one machine with 8 V100 GPUs: +```sh +# Training +SAVE="save/dynamic_conv_wmt16en2de" +mkdir -p $SAVE +python -m torch.distributed.launch --nproc_per_node 8 $(which fairseq-train) \ + data-bin/wmt16_en_de_bpe32k --fp16 --log-interval 100 --no-progress-bar \ + --max-update 30000 --share-all-embeddings --optimizer adam \ + --adam-betas '(0.9, 0.98)' --clip-norm 0.0 --weight-decay 0.0 \ + --criterion label_smoothed_cross_entropy --label-smoothing 0.1 \ + --stop-min-lr 1e-09 --update-freq 16 --attention-dropout 0.1 --keep-last-epochs 10 \ + --ddp-backend=legacy_ddp --max-tokens 3584 \ + --lr-scheduler cosine --warmup-init-lr 1e-7 --warmup-updates 10000 \ + --lr-shrink 1 --lr 0.001 --min-lr 1e-7 --warmup-init-lr 1e-07 \ + --t-mult 1 --lr-period-updates 20000 \ + --arch lightconv_wmt_en_de_big --save-dir $SAVE \ + --dropout 0.3 --attention-dropout 0.1 --weight-dropout 0.1 \ + --encoder-glu 1 --decoder-glu 1 + +# Evaluation +CUDA_VISIBLE_DEVICES=0 fairseq-generate data-bin/wmt16.en-de.joined-dict.newstest2014 --path "${SAVE}/checkpoint_best.pt" --batch-size 128 --beam 5 --remove-bpe --lenpen 0.5 --gen-subset test > wmt16_gen.txt +bash scripts/compound_split_bleu.sh wmt16_gen.txt +``` + +### WMT14 En-Fr +Training DynamicConv (with GLU) on WMT14 En-Fr using cosine scheduler on one machine with 8 V100 GPUs: +```sh +# Training +SAVE="save/dynamic_conv_wmt14en2fr" +mkdir -p $SAVE +python -m torch.distributed.launch --nproc_per_node 8 $(which fairseq-train) \ + data-bin/wmt14_en_fr --fp16 --log-interval 100 --no-progress-bar \ + --max-update 30000 --share-all-embeddings --optimizer adam \ + --adam-betas '(0.9, 0.98)' --clip-norm 0.0 --weight-decay 0.0 \ + --criterion label_smoothed_cross_entropy --label-smoothing 0.1 \ + --stop-min-lr 1e-09 --update-freq 16 --attention-dropout 0.1 --keep-last-epochs 10 \ + --ddp-backend=legacy_ddp --max-tokens 3584 \ + --lr-scheduler cosine --warmup-init-lr 1e-7 --warmup-updates 10000 \ + --lr-shrink 1 --lr 0.001 --min-lr 1e-7 --warmup-init-lr 1e-07 \ + --t-mult 1 --lr-period-updates 70000 \ + --arch lightconv_wmt_en_fr_big --save-dir $SAVE \ + --dropout 0.1 --attention-dropout 0.1 --weight-dropout 0.1 \ + --encoder-glu 1 --decoder-glu 1 + +# Evaluation +CUDA_VISIBLE_DEVICES=0 fairseq-generate data-bin/wmt14.en-fr.joined-dict.newstest2014 --path "${SAVE}/checkpoint_best.pt" --batch-size 128 --beam 5 --remove-bpe --lenpen 0.9 --gen-subset test +``` diff --git a/examples/pointer_generator/README.md b/examples/pointer_generator/README.md new file mode 100644 index 0000000000000000000000000000000000000000..60965708254aae2174812ea6686a9807825b7fb6 --- /dev/null +++ b/examples/pointer_generator/README.md @@ -0,0 +1,82 @@ +# Transformer with Pointer-Generator Network + +This page describes the `transformer_pointer_generator` model that incorporates +a pointing mechanism in the Transformer model that facilitates copying of input +words to the output. This architecture is described in [Enarvi et al. (2020)](https://www.aclweb.org/anthology/2020.nlpmc-1.4/). + +## Background + +The pointer-generator network was introduced in [See et al. (2017)](https://arxiv.org/abs/1704.04368) +for RNN encoder-decoder attention models. A similar mechanism can be +incorporated in a Transformer model by reusing one of the many attention +distributions for pointing. The attention distribution over the input words is +interpolated with the normal output distribution over the vocabulary words. This +allows the model to generate words that appear in the input, even if they don't +appear in the vocabulary, helping especially with small vocabularies. + +## Implementation + +The mechanism for copying out-of-vocabulary words from the input has been +implemented differently to See et al. In their [implementation](https://github.com/abisee/pointer-generator) +they convey the word identities through the model in order to be able to produce +words that appear in the input sequence but not in the vocabulary. A different +approach was taken in the Fairseq implementation to keep it self-contained in +the model file, avoiding any changes to the rest of the code base. Copying +out-of-vocabulary words is possible by pre-processing the input and +post-processing the output. This is described in detail in the next section. + +## Usage + +The training and evaluation procedure is outlined below. You can also find a +more detailed example for the XSum dataset on [this page](README.xsum.md). + +##### 1. Create a vocabulary and extend it with source position markers + +The pointing mechanism is especially helpful with small vocabularies, if we are +able to recover the identities of any out-of-vocabulary words that are copied +from the input. For this purpose, the model allows extending the vocabulary with +special tokens that can be used in place of `<unk>` tokens to identify different +input positions. For example, the user may add `<unk-0>`, `<unk-1>`, `<unk-2>`, +etc. to the end of the vocabulary, after the normal words. Below is an example +of how to create a vocabulary of 10000 most common words and add 1000 input +position markers. + +```bash +vocab_size=10000 +position_markers=1000 +export LC_ALL=C +cat train.src train.tgt | + tr -s '[:space:]' '\n' | + sort | + uniq -c | + sort -k1,1bnr -k2 | + head -n "$((vocab_size - 4))" | + awk '{ print $2 " " $1 }' >dict.pg.txt +python3 -c "[print('<unk-{}> 0'.format(n)) for n in range($position_markers)]" >>dict.pg.txt +``` + +##### 2. Preprocess the text data + +The idea is that any `<unk>` tokens in the text are replaced with `<unk-0>` if +it appears in the first input position, `<unk-1>` if it appears in the second +input position, and so on. This can be achieved using the `preprocess.py` script +that is provided in this directory. + +##### 3. Train a model + +The number of these special tokens is given to the model with the +`--source-position-markers` argument—the model simply maps all of these to the +same word embedding as `<unk>`. + +The attention distribution that is used for pointing is selected using the +`--alignment-heads` and `--alignment-layer` command-line arguments in the same +way as with the `transformer_align` model. + +##### 4. Generate text and postprocess it + +When using the model to generate text, you want to preprocess the input text in +the same way that training data was processed, replacing out-of-vocabulary words +with `<unk-N>` tokens. If any of these tokens are copied to the output, the +actual words can be retrieved from the unprocessed input text. Any `<unk-N>` +token should be replaced with the word at position N in the original input +sequence. This can be achieved using the `postprocess.py` script. diff --git a/examples/pointer_generator/README.xsum.md b/examples/pointer_generator/README.xsum.md new file mode 100644 index 0000000000000000000000000000000000000000..ac3a8c3ddc96cd9810b45d49f6b361e43de1e9fb --- /dev/null +++ b/examples/pointer_generator/README.xsum.md @@ -0,0 +1,180 @@ +## Training a pointer-generator model on the Extreme Summarization dataset + +##### 1. Download the Extreme Summarization data and preprocess it + +Follow the instructions [here](https://github.com/EdinburghNLP/XSum) to obtain +the original Extreme Summarization dataset. You should have six files, +{train,validation,test}.{document,summary}. + +##### 2. Create a vocabulary and extend it with source position markers + +```bash +vocab_size=10000 +position_markers=1000 +export LC_ALL=C +cat train.document train.summary | + tr -s '[:space:]' '\n' | + sort | + uniq -c | + sort -k1,1bnr -k2 | + head -n "$((vocab_size - 4))" | + awk '{ print $2 " " $1 }' >dict.pg.txt +python3 -c "[print('<unk-{}> 0'.format(n)) for n in range($position_markers)]" >>dict.pg.txt +``` + +This creates the file dict.pg.txt that contains the 10k most frequent words, +followed by 1k source position markers: + +``` +the 4954867 +. 4157552 +, 3439668 +to 2212159 +a 1916857 +of 1916820 +and 1823350 +... +<unk-0> 0 +<unk-1> 0 +<unk-2> 0 +<unk-3> 0 +<unk-4> 0 +... +``` + +##### 2. Preprocess the text data + +```bash +./preprocess.py --source train.document --target train.summary --vocab <(cut -d' ' -f1 dict.pg.txt) --source-out train.pg.src --target-out train.pg.tgt +./preprocess.py --source validation.document --target validation.summary --vocab <(cut -d' ' -f1 dict.pg.txt) --source-out valid.pg.src --target-out valid.pg.tgt +./preprocess.py --source test.document --vocab <(cut -d' ' -f1 dict.pg.txt) --source-out test.pg.src +``` + +The data should now contain `<unk-N>` tokens in place of out-of-vocabulary words. + +##### 3. Binarize the dataset: + +```bash +fairseq-preprocess \ + --source-lang src \ + --target-lang tgt \ + --trainpref train.pg \ + --validpref valid.pg \ + --destdir bin \ + --workers 60 \ + --srcdict dict.pg.txt \ + --joined-dictionary +``` + +##### 3. Train a model + +```bash +total_updates=20000 +warmup_updates=500 +lr=0.001 +max_tokens=4096 +update_freq=4 +pointer_layer=-2 + +CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 fairseq-train bin \ + --user-dir examples/pointer_generator/pointer_generator_src \ + --max-tokens "$max_tokens" \ + --task translation \ + --source-lang src --target-lang tgt \ + --truncate-source \ + --layernorm-embedding \ + --share-all-embeddings \ + --encoder-normalize-before \ + --decoder-normalize-before \ + --required-batch-size-multiple 1 \ + --arch transformer_pointer_generator \ + --alignment-layer "$pointer_layer" \ + --alignment-heads 1 \ + --source-position-markers 1000 \ + --criterion label_smoothed_cross_entropy \ + --label-smoothing 0.1 \ + --dropout 0.1 --attention-dropout 0.1 \ + --weight-decay 0.01 --optimizer adam --adam-betas "(0.9, 0.999)" --adam-eps 1e-08 \ + --clip-norm 0.1 \ + --lr-scheduler inverse_sqrt --lr "$lr" --max-update "$total_updates" --warmup-updates "$warmup_updates" \ + --update-freq "$update_freq" \ + --skip-invalid-size-inputs-valid-test +``` + +Above we specify that our dictionary contains 1000 source position markers, and +that we want to use one attention head from the penultimate decoder layer for +pointing. It should run in 5.5 hours on one node with eight 32GB V100 GPUs. The +logged messages confirm that dictionary indices above 10000 will be mapped to +the `<unk>` embedding: + +``` +2020-09-24 20:43:53 | INFO | fairseq.tasks.translation | [src] dictionary: 11000 types +2020-09-24 20:43:53 | INFO | fairseq.tasks.translation | [tgt] dictionary: 11000 types +2020-09-24 20:43:53 | INFO | fairseq.data.data_utils | loaded 11332 examples from: bin/valid.src-tgt.src +2020-09-24 20:43:53 | INFO | fairseq.data.data_utils | loaded 11332 examples from: bin/valid.src-tgt.tgt +2020-09-24 20:43:53 | INFO | fairseq.tasks.translation | bin valid src-tgt 11332 examples +2020-09-24 20:43:53 | INFO | fairseq.models.transformer_pg | dictionary indices from 10000 to 10999 will be mapped to 3 +``` + +##### 4. Summarize the test sequences + +```bash +batch_size=32 +beam_size=6 +max_length=60 +length_penalty=1.0 + +fairseq-interactive bin \ + --user-dir examples/pointer_generator/pointer_generator_src \ + --batch-size "$batch_size" \ + --task translation \ + --source-lang src --target-lang tgt \ + --path checkpoints/checkpoint_last.pt \ + --input test.pg.src \ + --buffer-size 200 \ + --max-len-a 0 \ + --max-len-b "$max_length" \ + --lenpen "$length_penalty" \ + --beam "$beam_size" \ + --skip-invalid-size-inputs-valid-test | + tee generate.out +grep ^H generate.out | cut -f 3- >generate.hyp +``` + +Now you should have the generated sequences in `generate.hyp`. They contain +`<unk-N>` tokens that the model has copied from the source sequence. In order to +retrieve the original words, we need the unprocessed source sequences from +`test.document`. + +##### 5. Process the generated output + +Since we skipped too long inputs when producing `generate.hyp`, we also have to +skip too long sequences now that we read `test.document`. + +```bash +./postprocess.py \ + --source <(awk 'NF<1024' test.document) \ + --target generate.hyp \ + --target-out generate.hyp.processed +``` + +Now you'll find the final sequences from `generate.hyp.processed`, with +`<unk-N>` replaced with the original word from the source sequence. + +##### An example of a summarized sequence + +The original source document in `test.document`: + +> de roon moved to teesside in june 2016 for an initial # 8.8 m fee and played 33 premier league games last term . the netherlands international , 26 , scored five goals in 36 league and cup games during his spell at boro . meanwhile , manager garry monk confirmed the championship club 's interest in signing chelsea midfielder lewis baker . `` he 's a target and one of many that we 've had throughout the summer months , '' said monk . find all the latest football transfers on our dedicated page . + +The preprocessed source document in `test.src.pg`: + +> de \<unk-1> moved to \<unk-4> in june 2016 for an initial # \<unk-12> m fee and played 33 premier league games last term . the netherlands international , 26 , scored five goals in 36 league and cup games during his spell at boro . meanwhile , manager garry monk confirmed the championship club 's interest in signing chelsea midfielder lewis baker . `` he 's a target and one of many that we 've had throughout the summer months , '' said monk . find all the latest football transfers on our dedicated page . + +The generated summary in `generate.hyp`: + +> middlesbrough striker \<unk> de \<unk-1> has joined spanish side \<unk> on a season-long loan . + +The generated summary after postprocessing in `generate.hyp.processed`: + +> middlesbrough striker \<unk> de roon has joined spanish side \<unk> on a season-long loan . diff --git a/examples/pointer_generator/pointer_generator_src/__init__.py b/examples/pointer_generator/pointer_generator_src/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..c361ff6bd616512fe2521387665de1ad1aff66d0 --- /dev/null +++ b/examples/pointer_generator/pointer_generator_src/__init__.py @@ -0,0 +1,6 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from . import transformer_pg # noqa diff --git a/examples/pointer_generator/pointer_generator_src/transformer_pg.py b/examples/pointer_generator/pointer_generator_src/transformer_pg.py new file mode 100644 index 0000000000000000000000000000000000000000..4ccf30f4eb154f8fab1e285934fb973a2d1166cb --- /dev/null +++ b/examples/pointer_generator/pointer_generator_src/transformer_pg.py @@ -0,0 +1,518 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +from typing import Any, Dict, Optional, List, Tuple + +import torch +import torch.nn as nn +from fairseq import utils +from fairseq.models import register_model, register_model_architecture +from fairseq.models.transformer import ( + DEFAULT_MAX_SOURCE_POSITIONS, + DEFAULT_MAX_TARGET_POSITIONS, + TransformerDecoder, + TransformerEncoder, + TransformerModel, + base_architecture, +) +from torch import Tensor + + +logger = logging.getLogger(__name__) + + +@register_model("transformer_pointer_generator") +class TransformerPointerGeneratorModel(TransformerModel): + """ + Transformer model from `"Attention Is All You Need" (Vaswani et al, 2017) + <https://arxiv.org/abs/1706.03762>`_, augmented with a pointer-generator + network from `"Get To The Point: Summarization with Pointer-Generator + Networks" (See et al, 2017) <https://arxiv.org/abs/1704.04368>`_. + + Args: + encoder (TransformerPointerGeneratorEncoder): the encoder + decoder (TransformerPointerGeneratorDecoder): the decoder + + The Transformer pointer-generator model provides the following named + architectures and command-line arguments: + + .. argparse:: + :ref: fairseq.models.transformer_pointer_generator_parser + :prog: + """ + + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + # fmt: off + TransformerModel.add_args(parser) + parser.add_argument('--alignment-heads', type=int, metavar='N', + help='number of attention heads to be used for ' + 'pointing') + parser.add_argument('--alignment-layer', type=int, metavar='I', + help='layer number to be used for pointing (0 ' + 'corresponding to the bottommost layer)') + parser.add_argument('--source-position-markers', type=int, metavar='N', + help='dictionary includes N additional items that ' + 'represent an OOV token at a particular input ' + 'position') + parser.add_argument('--force-generation', type=float, metavar='P', + default=None, + help='set the vocabulary distribution weight to P, ' + 'instead of predicting it from the input (1.0 ' + 'corresponding to generation, 0.0 to pointing)') + # fmt: on + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + + # make sure all arguments are present in older models + base_architecture(args) + + if args.encoder_layers_to_keep: + args.encoder_layers = len(args.encoder_layers_to_keep.split(",")) + if args.decoder_layers_to_keep: + args.decoder_layers = len(args.decoder_layers_to_keep.split(",")) + + if getattr(args, "max_source_positions", None) is None: + args.max_source_positions = DEFAULT_MAX_SOURCE_POSITIONS + if getattr(args, "max_target_positions", None) is None: + args.max_target_positions = DEFAULT_MAX_TARGET_POSITIONS + if getattr(args, "source_position_markers", None) is None: + args.source_position_markers = args.max_source_positions + + src_dict, tgt_dict = task.source_dictionary, task.target_dictionary + if src_dict != tgt_dict: + raise ValueError("Pointer-generator requires a joined dictionary") + + def build_embedding(dictionary, embed_dim, path=None): + # The dictionary may include additional items that can be used in + # place of the normal OOV token and that all map to the same + # embedding. Using a different token for each input position allows + # one to restore the word identities from the original source text. + num_embeddings = len(dictionary) - args.source_position_markers + padding_idx = dictionary.pad() + unk_idx = dictionary.unk() + logger.info( + "dictionary indices from {0} to {1} will be mapped to {2}".format( + num_embeddings, len(dictionary) - 1, unk_idx + ) + ) + emb = Embedding(num_embeddings, embed_dim, padding_idx, unk_idx) + # if provided, load from preloaded dictionaries + if path: + embed_dict = utils.parse_embedding(path) + utils.load_embedding(embed_dict, dictionary, emb) + return emb + + if args.share_all_embeddings: + if args.encoder_embed_dim != args.decoder_embed_dim: + raise ValueError( + "--share-all-embeddings requires --encoder-embed-dim to match --decoder-embed-dim" + ) + if args.decoder_embed_path and ( + args.decoder_embed_path != args.encoder_embed_path + ): + raise ValueError( + "--share-all-embeddings not compatible with --decoder-embed-path" + ) + encoder_embed_tokens = build_embedding( + src_dict, args.encoder_embed_dim, args.encoder_embed_path + ) + decoder_embed_tokens = encoder_embed_tokens + args.share_decoder_input_output_embed = True + else: + encoder_embed_tokens = build_embedding( + src_dict, args.encoder_embed_dim, args.encoder_embed_path + ) + decoder_embed_tokens = build_embedding( + tgt_dict, args.decoder_embed_dim, args.decoder_embed_path + ) + + encoder = cls.build_encoder(args, src_dict, encoder_embed_tokens) + decoder = cls.build_decoder(args, tgt_dict, decoder_embed_tokens) + return cls(args, encoder, decoder) + + @classmethod + def build_encoder(cls, args, src_dict, embed_tokens): + return TransformerPointerGeneratorEncoder(args, src_dict, embed_tokens) + + @classmethod + def build_decoder(cls, args, tgt_dict, embed_tokens): + return TransformerPointerGeneratorDecoder(args, tgt_dict, embed_tokens) + + +class TransformerPointerGeneratorEncoder(TransformerEncoder): + """ + Transformer encoder consisting of *args.encoder_layers* layers. Each layer + is a :class:`TransformerEncoderLayer`. The pointer-generator variant adds + the source tokens to the encoder output as these are otherwise not passed + to the decoder. + """ + + def forward( + self, + src_tokens, + src_lengths: Optional[Tensor] = None, + return_all_hiddens: bool = False, + token_embeddings: Optional[Tensor] = None + ): + """ + Runs the `forward()` method of the parent Transformer class. Then adds + the source tokens into the encoder output tuple. + + While it might be more elegant that the model would pass the source + tokens to the `forward()` method of the decoder too, this would require + changes to `SequenceGenerator`. + + Args: + src_tokens (torch.LongTensor): tokens in the source language of + shape `(batch, src_len)` + src_lengths (torch.LongTensor): lengths of each source sentence of + shape `(batch)` + return_all_hiddens (bool, optional): also return all of the + intermediate hidden states (default: False). + token_embeddings (torch.Tensor, optional): precomputed embeddings + default `None` will recompute embeddings + + Returns: + namedtuple: + - **encoder_out** (Tensor): the last encoder layer's output of + shape `(src_len, batch, embed_dim)` + - **encoder_padding_mask** (ByteTensor): the positions of + padding elements of shape `(batch, src_len)` + - **encoder_embedding** (Tensor): the (scaled) embedding lookup + of shape `(batch, src_len, embed_dim)` + - **encoder_states** (List[Tensor]): all intermediate + hidden states of shape `(src_len, batch, embed_dim)`. + Only populated if *return_all_hiddens* is True. + - **src_tokens** (Tensor): input token ids of shape + `(batch, src_len)` + """ + encoder_out = self.forward_scriptable(src_tokens, + src_lengths, + return_all_hiddens, + token_embeddings) + + # The Pytorch Mobile lite interpreter does not supports returning NamedTuple in + # `forward` so we use a dictionary instead. + # TorchScript does not support mixed values so the values are all lists. + # The empty list is equivalent to None. + return { + "encoder_out": encoder_out["encoder_out"], # T x B x C + "encoder_padding_mask": encoder_out["encoder_padding_mask"], # B x T + "encoder_embedding": encoder_out["encoder_embedding"], # B x T x C + "encoder_states": encoder_out["encoder_states"], # List[T x B x C] + "src_tokens": [src_tokens], # B x T + "src_lengths": [], + } + + +class TransformerPointerGeneratorDecoder(TransformerDecoder): + """ + Transformer decoder consisting of *args.decoder_layers* layers. Each layer + is a :class:`TransformerDecoderLayer`. The pointer-generator variant mixes + the output probabilities with an attention distribution in the output layer. + + Args: + args (argparse.Namespace): parsed command-line arguments + dictionary (~fairseq.data.Dictionary): decoding dictionary + embed_tokens (torch.nn.Embedding): output embedding + """ + + def __init__(self, args, dictionary, embed_tokens): + super().__init__(args, dictionary, embed_tokens, no_encoder_attn=False) + + # In the pointer-generator model these arguments define the decoder + # layer and the number of attention heads that will be averaged to + # create the alignment for pointing. + self.alignment_heads = args.alignment_heads + self.alignment_layer = args.alignment_layer + + input_embed_dim = embed_tokens.embedding_dim + + # Generation probabilities / interpolation coefficients are predicted + # from the current decoder input embedding and the decoder output, which + # is the size of output_embed_dim. + p_gen_input_size = input_embed_dim + self.output_embed_dim + self.project_p_gens = nn.Linear(p_gen_input_size, 1) + nn.init.zeros_(self.project_p_gens.bias) + + # The dictionary may include a separate entry for an OOV token in each + # input position, so that their identity can be restored from the + # original source text. + self.num_types = len(dictionary) + self.num_oov_types = args.source_position_markers + self.num_embeddings = self.num_types - self.num_oov_types + self.force_p_gen = args.force_generation + + def forward( + self, + prev_output_tokens, + encoder_out: Optional[Dict[str, List[Tensor]]] = None, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + features_only: bool = False, + alignment_layer: Optional[int] = 0, + alignment_heads: Optional[int] = 1, + src_lengths: Optional[Any] = None, + return_all_hiddens: bool = False, + ): + """ + Args: + prev_output_tokens (LongTensor): previous decoder outputs of shape + `(batch, tgt_len)`, for teacher forcing + encoder_out (optional): output from the encoder, used for + encoder-side attention + incremental_state (dict, optional): dictionary used for storing + state during :ref:`Incremental decoding` + features_only (bool, optional): only return features without + applying output layer (default: False) + alignment_layer (int, optional): 0-based index of the layer to be + used for pointing (default: 0) + alignment_heads (int, optional): number of attention heads to be + used for pointing (default: 1) + + Returns: + tuple: + - the decoder's output of shape `(batch, tgt_len, vocab)` + - a dictionary with any model-specific outputs + """ + # The normal Transformer model doesn't pass the alignment_layer and + # alignment_heads parameters correctly. We use our local variables. + x, extra = self.extract_features( + prev_output_tokens, + encoder_out=encoder_out, + incremental_state=incremental_state, + alignment_layer=self.alignment_layer, + alignment_heads=self.alignment_heads, + ) + if not features_only: + # Embedding the tokens again for generation probability prediction, + # so that we don't have to reimplement the whole extract_features() + # method. + if incremental_state is not None: + prev_output_tokens = prev_output_tokens[:, -1:] + prev_output_embed = self.embed_tokens(prev_output_tokens) + prev_output_embed *= self.embed_scale + predictors = torch.cat((prev_output_embed, x), 2) + p_gens = self.project_p_gens(predictors) + p_gens = torch.sigmoid(p_gens.float()) + # Torchscript complains if encoder_out or attn are None because + # `output_layer()` signature expects tensors instead + attn: Optional[Tensor] = extra["attn"][0] + assert encoder_out is not None + assert attn is not None + x = self.output_layer(x, attn, encoder_out["src_tokens"][0], p_gens) + return x, extra + + def output_layer( + self, + features: Tensor, + attn: Tensor, + src_tokens: Tensor, + p_gens: Tensor + ) -> Tensor: + """ + Project features to the vocabulary size and mix with the attention + distributions. + """ + if self.force_p_gen is not None: + p_gens = self.force_p_gen + + # project back to size of vocabulary + if self.adaptive_softmax is None: + logits = self.output_projection(features) + else: + logits = features + + batch_size = logits.shape[0] + output_length = logits.shape[1] + assert logits.shape[2] == self.num_embeddings + assert src_tokens.shape[0] == batch_size + src_length = src_tokens.shape[1] + + # The final output distribution will be a mixture of the normal output + # distribution (softmax of logits) and attention weights. + gen_dists = self.get_normalized_probs_scriptable( + (logits, None), log_probs=False, sample=None + ) + gen_dists = torch.mul(gen_dists, p_gens) + padding_size = (batch_size, output_length, self.num_oov_types) + padding = gen_dists.new_zeros(padding_size) + gen_dists = torch.cat((gen_dists, padding), 2) + assert gen_dists.shape[2] == self.num_types + + # Scatter attention distributions to distributions over the extended + # vocabulary in a tensor of shape [batch_size, output_length, + # vocab_size]. Each attention weight will be written into a location + # that is for other dimensions the same as in the index tensor, but for + # the third dimension it's the value of the index tensor (the token ID). + attn = torch.mul(attn.float(), 1 - p_gens) + index = src_tokens[:, None, :] + index = index.expand(batch_size, output_length, src_length) + attn_dists_size = (batch_size, output_length, self.num_types) + attn_dists = attn.new_zeros(attn_dists_size) + attn_dists.scatter_add_(2, index, attn.float()) + + # Final distributions, [batch_size, output_length, num_types]. + return gen_dists + attn_dists + + def get_normalized_probs( + self, + net_output: Tuple[Tensor, Optional[Dict[str, List[Optional[Tensor]]]]], + log_probs: bool, + sample: Optional[Dict[str, Tensor]] = None, + ): + """ + Get normalized probabilities (or log probs) from a net's output. + Pointer-generator network output is already normalized. + """ + probs = net_output[0] + # Make sure the probabilities are greater than zero when returning log + # probabilities. + return probs.clamp(1e-10, 1.0).log() if log_probs else probs + + +class Embedding(nn.Embedding): + r"""A simple lookup table that stores embeddings of a fixed dictionary and size. + This module is often used to store word embeddings and retrieve them using indices. + The input to the module is a list of indices, and the output is the corresponding + word embeddings. This subclass differs from the standard PyTorch Embedding class by + allowing additional vocabulary entries that will be mapped to the unknown token + embedding. + Args: + num_embeddings (int): size of the dictionary of embeddings + embedding_dim (int): the size of each embedding vector + padding_idx (int): Pads the output with the embedding vector at :attr:`padding_idx` + (initialized to zeros) whenever it encounters the index. + unk_idx (int): Maps all token indices that are greater than or equal to + num_embeddings to this index. + Attributes: + weight (Tensor): the learnable weights of the module of shape (num_embeddings, embedding_dim) + initialized from :math:`\mathcal{N}(0, 1)` + Shape: + - Input: :math:`(*)`, LongTensor of arbitrary shape containing the indices to extract + - Output: :math:`(*, H)`, where `*` is the input shape and :math:`H=\text{embedding\_dim}` + .. note:: + Keep in mind that only a limited number of optimizers support + sparse gradients: currently it's :class:`optim.SGD` (`CUDA` and `CPU`), + :class:`optim.SparseAdam` (`CUDA` and `CPU`) and :class:`optim.Adagrad` (`CPU`) + .. note:: + With :attr:`padding_idx` set, the embedding vector at + :attr:`padding_idx` is initialized to all zeros. However, note that this + vector can be modified afterwards, e.g., using a customized + initialization method, and thus changing the vector used to pad the + output. The gradient for this vector from :class:`~torch.nn.Embedding` + is always zero. + """ + __constants__ = ["unk_idx"] + + # Torchscript: Inheriting from Embedding class produces an error when exporting to Torchscript + # -> RuntimeError: Unable to cast Python instance to C++ type (compile in debug mode for details + # It's happening because max_norm attribute from nn.Embedding is None by default and it cannot be + # cast to a C++ type + def __init__( + self, + num_embeddings: int, + embedding_dim: int, + padding_idx: Optional[int], + unk_idx: int, + max_norm: Optional[float] = float("inf"), + ): + super().__init__(num_embeddings, embedding_dim, padding_idx=padding_idx, max_norm=max_norm) + self.unk_idx = unk_idx + nn.init.normal_(self.weight, mean=0, std=embedding_dim ** -0.5) + nn.init.constant_(self.weight[padding_idx], 0) + + def forward(self, input): + input = torch.where( + input >= self.num_embeddings, torch.ones_like(input) * self.unk_idx, input + ) + return nn.functional.embedding( + input, self.weight, self.padding_idx, self.max_norm, + self.norm_type, self.scale_grad_by_freq, self.sparse + ) + + +@register_model_architecture( + "transformer_pointer_generator", "transformer_pointer_generator" +) +def transformer_pointer_generator(args): + args.alignment_heads = getattr(args, "alignment_heads", 1) + args.alignment_layer = getattr(args, "alignment_layer", -1) + base_architecture(args) + if args.alignment_layer < 0: + args.alignment_layer = args.decoder_layers + args.alignment_layer + + +@register_model_architecture( + "transformer_pointer_generator", "transformer_pointer_generator_iwslt_de_en" +) +def transformer_pointer_generator_iwslt_de_en(args): + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 1024) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 4) + args.encoder_layers = getattr(args, "encoder_layers", 6) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 512) + args.decoder_ffn_embed_dim = getattr(args, "decoder_ffn_embed_dim", 1024) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 4) + args.decoder_layers = getattr(args, "decoder_layers", 6) + transformer_pointer_generator(args) + + +@register_model_architecture( + "transformer_pointer_generator", "transformer_pointer_generator_wmt_en_de" +) +def transformer_pointer_generator_wmt_en_de(args): + transformer_pointer_generator(args) + + +# Transformer pointer-generator with the base Transformer parameters as used in +# the "Attention Is All You Need" paper (Vaswani et al., 2017) +@register_model_architecture( + "transformer_pointer_generator", + "transformer_pointer_generator_vaswani_wmt_en_de_big", +) +def transformer_pointer_generator_vaswani_wmt_en_de_big(args): + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 1024) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 4096) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 16) + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 1024) + args.decoder_ffn_embed_dim = getattr(args, "decoder_ffn_embed_dim", 4096) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 16) + args.dropout = getattr(args, "dropout", 0.3) + transformer_pointer_generator(args) + + +@register_model_architecture( + "transformer_pointer_generator", + "transformer_pointer_generator_vaswani_wmt_en_fr_big", +) +def transformer_pointer_generator_vaswani_wmt_en_fr_big(args): + args.dropout = getattr(args, "dropout", 0.1) + transformer_pointer_generator_vaswani_wmt_en_de_big(args) + + +@register_model_architecture( + "transformer_pointer_generator", "transformer_pointer_generator_wmt_en_de_big" +) +def transformer_pointer_generator_wmt_en_de_big(args): + args.attention_dropout = getattr(args, "attention_dropout", 0.1) + transformer_pointer_generator_vaswani_wmt_en_de_big(args) + + +# default parameters used in tensor2tensor implementation +@register_model_architecture( + "transformer_pointer_generator", "transformer_pointer_generator_wmt_en_de_big_t2t" +) +def transformer_pointer_generator_wmt_en_de_big_t2t(args): + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", True) + args.decoder_normalize_before = getattr(args, "decoder_normalize_before", True) + args.attention_dropout = getattr(args, "attention_dropout", 0.1) + args.activation_dropout = getattr(args, "activation_dropout", 0.1) + transformer_pointer_generator_vaswani_wmt_en_de_big(args) diff --git a/examples/pointer_generator/postprocess.py b/examples/pointer_generator/postprocess.py new file mode 100755 index 0000000000000000000000000000000000000000..b213aed80fd1e3d86f975256fcb7d9d4c16ca857 --- /dev/null +++ b/examples/pointer_generator/postprocess.py @@ -0,0 +1,96 @@ +#!/usr/bin/env python3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import re +import sys + + +class OOVIndexError(IndexError): + def __init__(self, pos, source_seq, target_seq): + super(OOVIndexError, self).__init__( + "A <unk-N> tag in the target sequence refers to a position that is " + "outside the source sequence. Most likely there was a mismatch in " + "provided source and target sequences. Otherwise this would mean that " + "the pointing mechanism somehow attended to a position that is past " + "the actual sequence end." + ) + self.source_pos = pos + self.source_seq = source_seq + self.target_seq = target_seq + + +def replace_oovs(source_in, target_in, target_out): + """Replaces <unk-N> tokens in the target text with the corresponding word in + the source text. + """ + + oov_re = re.compile("^<unk-([0-9]+)>$") + + for source_seq, target_seq in zip(source_in, target_in): + target_seq_out = [] + + pos_to_word = source_seq.strip().split() + for token in target_seq.strip().split(): + m = oov_re.match(token) + if m: + pos = int(m.group(1)) + if pos >= len(pos_to_word): + raise OOVIndexError(pos, source_seq, target_seq) + token_out = pos_to_word[pos] + else: + token_out = token + target_seq_out.append(token_out) + target_out.write(" ".join(target_seq_out) + "\n") + + +def main(): + parser = argparse.ArgumentParser( + description="Replaces <unk-N> tokens in target sequences with words from " + "the corresponding position in the source sequence." + ) + parser.add_argument( + "--source", type=str, help="text file with source sequences", required=True + ) + parser.add_argument( + "--target", type=str, help="text file with target sequences", required=True + ) + parser.add_argument( + "--target-out", + type=str, + help="where to write target sequences without <unk-N> " "entries", + required=True, + ) + args = parser.parse_args() + + target_in = ( + open(args.target, "r", encoding="utf-8") if args.target is not None else None + ) + target_out = ( + open(args.target_out, "w", encoding="utf-8") + if args.target_out is not None + else None + ) + with open(args.source, "r", encoding="utf-8") as source_in, open( + args.target, "r", encoding="utf-8" + ) as target_in, open(args.target_out, "w", encoding="utf-8") as target_out: + replace_oovs(source_in, target_in, target_out) + + +if __name__ == "__main__": + try: + main() + except OOVIndexError as e: + print(e, file=sys.stderr) + print("Source sequence:", e.source_seq.strip(), file=sys.stderr) + print("Target sequence:", e.target_seq.strip(), file=sys.stderr) + print( + "Source sequence length:", + len(e.source_seq.strip().split()), + file=sys.stderr, + ) + print("The offending tag points to:", e.source_pos) + sys.exit(2) diff --git a/examples/pointer_generator/preprocess.py b/examples/pointer_generator/preprocess.py new file mode 100755 index 0000000000000000000000000000000000000000..f72ca7d3d97e12ab7b405dcff314bdb6c0a78755 --- /dev/null +++ b/examples/pointer_generator/preprocess.py @@ -0,0 +1,102 @@ +#!/usr/bin/env python3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +from itertools import zip_longest + + +def replace_oovs(source_in, target_in, vocabulary, source_out, target_out): + """Replaces out-of-vocabulary words in source and target text with <unk-N>, + where N in is the position of the word in the source sequence. + """ + + def format_unk(pos): + return "<unk-{}>".format(pos) + + if target_in is None: + target_in = [] + + for seq_num, (source_seq, target_seq) in enumerate( + zip_longest(source_in, target_in) + ): + source_seq_out = [] + target_seq_out = [] + + word_to_pos = dict() + for position, token in enumerate(source_seq.strip().split()): + if token in vocabulary: + token_out = token + else: + if token in word_to_pos: + oov_pos = word_to_pos[token] + else: + word_to_pos[token] = position + oov_pos = position + token_out = format_unk(oov_pos) + source_seq_out.append(token_out) + source_out.write(" ".join(source_seq_out) + "\n") + + if target_seq is not None: + for token in target_seq.strip().split(): + if token in word_to_pos: + token_out = format_unk(word_to_pos[token]) + else: + token_out = token + target_seq_out.append(token_out) + if target_out is not None: + target_out.write(" ".join(target_seq_out) + "\n") + + +def main(): + parser = argparse.ArgumentParser( + description="Replaces out-of-vocabulary words in both source and target " + "sequences with tokens that indicate the position of the word " + "in the source sequence." + ) + parser.add_argument( + "--source", type=str, help="text file with source sequences", required=True + ) + parser.add_argument( + "--target", type=str, help="text file with target sequences", default=None + ) + parser.add_argument("--vocab", type=str, help="vocabulary file", required=True) + parser.add_argument( + "--source-out", + type=str, + help="where to write source sequences with <unk-N> entries", + required=True, + ) + parser.add_argument( + "--target-out", + type=str, + help="where to write target sequences with <unk-N> entries", + default=None, + ) + args = parser.parse_args() + + with open(args.vocab, encoding="utf-8") as vocab: + vocabulary = vocab.read().splitlines() + + target_in = ( + open(args.target, "r", encoding="utf-8") if args.target is not None else None + ) + target_out = ( + open(args.target_out, "w", encoding="utf-8") + if args.target_out is not None + else None + ) + with open(args.source, "r", encoding="utf-8") as source_in, open( + args.source_out, "w", encoding="utf-8" + ) as source_out: + replace_oovs(source_in, target_in, vocabulary, source_out, target_out) + if target_in is not None: + target_in.close() + if target_out is not None: + target_out.close() + + +if __name__ == "__main__": + main() diff --git a/examples/quant_noise/README.md b/examples/quant_noise/README.md new file mode 100644 index 0000000000000000000000000000000000000000..539c3d5af906d353e264a1c44612229255428dba --- /dev/null +++ b/examples/quant_noise/README.md @@ -0,0 +1,298 @@ +# Training with Quantization Noise for Extreme Model Compression ({Fan\*, Stock\*} *et al.*, 2020) +This page contains information for how to train and quantize models with Quantization Noise, for both scalar quantization like `int8` and Iterative Product Quantization. +Check out our paper [here](https://arxiv.org/abs/2004.07320). + +Looking for pretrained models? They will be added shortly. +Looking for code to train vision models? We are working on open sourcing our code as part of ClassyVision. Please check back, but note that both the Scalar and Iterative Product Quantization counterparts of the `nn.Conv2d` module are already included in this release. + +**Contents**: +- [Walk through of code](#walk-through-the-code) +- [Reproduce NLP Results](#looking-to-reproduce-the-nlp-results-in-the-paper) +- [Reproduce Vision Results](#looking-to-reproduce-the-vision-results-in-the-paper) + + +## Citation +```bibtex +@article{fan2020training, + title={Training with Quantization Noise for Extreme Model Compression}, + author={Angela Fan* and Pierre Stock* and and Benjamin Graham and Edouard Grave and Remi Gribonval and Herve Jegou and Armand Joulin}, + year={2020}, + eprint={2004.07320}, + archivePrefix={arXiv}, + primaryClass={cs.ML} +} +``` + +## Walk through the code + +Training a model with Quant-Noise improves the performance in subsequent inference-time quantization by training models to be robust to quantization. This technique is useful for both scalar and product quantization methods, as well as multiple domains. We detail below our approach to train, quantize models and integrate our code to quantize your favorite models. + +### Scalar Quantization + +Unlike the section [Iterative Product Quantization](#iterative-product-quantization) which gives state-of-the-art compression, this section showcases the usefulness of our approach for simple scalar quantization baselines such as int8 using on-GPU Fake Quantization. + +#### Training + +Scalar quantization with Quant-Noise consists in randomly quantizing a proportion `p` of the weights during training. Scalar quantization is implemented [here](https://github.com/pytorch/fairseq/tree/master/fairseq/modules/quantization/scalar) under the form of Fake Quantization, meaning that we emulate int8 on GPU by quantizing and de-quantizing both the weights and the activations. We rely on PyTorch's [quantization primitives](https://github.com/pytorch/pytorch/tree/master/torch/quantization). + +To train a model with Quant-Noise, add the following flag: +``` +--quant-noise-scalar 0.5 +``` +Large values of noise make the network easier to quantize but may result in higher non-quantized test and validation perplexities. + +#### Quantization + +When evaluating a network, all quantized modules and activation hooks automatically switch to `p=1` so the validation accuracy reported by Fairseq is actually the quantized one, nothing more to do. + + +#### Integration with your own code + +Looking to quantize your own models with Quant-Noise + Scalar Quantization? +- Use the function `quantize_model_` implemented [here](https://github.com/pytorch/fairseq/tree/master/fairseq/modules/quantization/scalar/utils.py) to (1) replace all your modules by their quantized counterparts and (2) add hooks to those modules to quantize the activations. +- Then, perform your training as usual. Note that in `eval()` mode, the network is always fully quantized (weights and activations) by default (`p=1`). + + + +### Iterative Product Quantization + + +Iterative Product Quantization with Quant-Noise proceeds in two steps. First, a model must be trained uncompressed with Quant-Noise. Second, the model must be quantized with iPQ. Note that we implement here the simplest form of noise, which consists in randomly dropping a proportion `p` of blocks, and that worked as well as assigning those blocks to their current centroid. + +#### Training + +To train a model with Quant-Noise, add the following flags: +``` +--quant-noise-pq 0.1 --quant-noise-pq-block-size 8 +``` +`quant-noise-pq` controls how much dropout is applied to the blocks of the weight matrix. `quant-noise-pq-block-size` controls the size of the weight matrix blocks. +We recommend training with 0.05 to 0.2 Quant-Noise, a value that worked well in our experiments. For the block-size, we recommend training with block-size of 8. Note that the block size must be a multiple of `input_features`, see the size checks [here](https://github.com/pytorch/fairseq/tree/master/fairseq/modules/quant_noise.py). Large block sizes result in higher compression ratio but may induce a loss in accuracy. + +We currently support training Transformer based models, such as sequence-to-sequence, language models, and BERT architectures. The `quant_noise` function [here](https://github.com/pytorch/fairseq/tree/master/fairseq/modules/quant_noise.py) wraps a module. It splits a weight matrix into blocks and applies random dropout to these blocks. +In the Transformer architectures, quant-noise is applied to the input and output embeddings, the attention, and the FFN. + +Quant-Noise can also be combined with **LayerDrop** (see [here](https://github.com/pytorch/fairseq/tree/master/examples/layerdrop)) to add its pruning effect to the quantized model and make the model even smaller. We recommend training with LayerDrop 0.1 or 0.2. + +#### Quantization + +We implement an improved version of product quantization from Stock et al, **iPQ**, described [here](https://arxiv.org/abs/1907.05686), see code with old API [here](https://github.com/facebookresearch/kill-the-bits). Note that we improved the iPQ API in terms of both compute speed and usability as described below. + +For the particular case of PQ, quantization is made sequentially. We recommend first quantizing the FFNs, then the EMBs, and finally the ATTNs. Quantization is done in two sub-steps: +- First, perform `n` steps of Product Quantization (generally `n=20` is enough). +- Then, finetune the obtained centroids. + +#### Integration with your own code + +Looking to quantize your own models with Quant-Noise + iPQ? +- First wrap your modules with the `quant_noise` function [here](https://github.com/pytorch/fairseq/tree/master/fairseq/modules/quant_noise.py), which is module-agnostic and train your favorite model. +- Then, quantize your trained model using the code [here](https://github.com/pytorch/fairseq/tree/master/fairseq/modules/quantization/pq). This can be done *without any changes to your training loop*. Below is an example code for integration. +Note that we tried our approach only on Transformers and various Convolutional Models such as EfficientNets. + +```python +from fairseq.modules.quantization.pq import quantize_model_, SizeTracker + +# get configuration parameters +n_centroids_config = config["n_centroids"] +block_sizes_config = config["block_sizes"] +layers_to_quantize = config["layers_to_quantize"] + +# size tracker for keeping track of assignments, centroids and non-compressed sizes +size_tracker = SizeTracker(model) + +# Quantize model by stages +for step in range(len(layers_to_quantize)): + + # quantize model in-place + quantized_layers = quantize_model_( + model, + size_tracker, + layers_to_quantize, + block_sizes_config, + n_centroids_config, + step=step, + ) + logger.info(f"Finetuning stage {step}, quantized layers: {quantized_layers}") + logger.info(f"{size_tracker}") + + # Don't forget to re-create/update trainer/optimizer since model parameters have changed + optimizer = ... + + # Finetune the centroids with your usual training loop for a few epochs + trainer.train_epoch() +``` + + +## Looking to reproduce the NLP results in the paper? + +We detail below how to reproduce the state-of-the-art results in reported in the paper for Quant-Noise + Iterative Product Quantization. + +### Training with Quant-Noise + +To **train** RoBERTa + QuantNoise, we followed this setting [here](https://github.com/pytorch/fairseq/tree/master/examples/roberta). +The following command can be used to train a RoBERTa Base + QuantNoise model: + +```bash +TOTAL_UPDATES=125000 +WARMUP_UPDATES=10000 +PEAK_LR=0.0005 +TOKENS_PER_SAMPLE=512 +MAX_POSITIONS=512 +MAX_SENTENCES=16 +UPDATE_FREQ=2 +DATA_DIR=/path/to/data/here + +fairseq-train $DATA_DIR \ + --task masked_lm --criterion masked_lm --arch roberta_base \ + --sample-break-mode complete \ + --tokens-per-sample $TOKENS_PER_SAMPLE --max-positions $MAX_POSITIONS \ + --optimizer adam --adam-betas '(0.9, 0.98)' --adam-eps 1e-6 \ + --clip-norm 0.0 \ + --lr-scheduler polynomial_decay --lr $PEAK_LR \ + --warmup-updates $WARMUP_UPDATES --total-num-update $TOTAL_UPDATES \ + --dropout 0.1 --attention-dropout 0.1 \ + --weight-decay 0.01 \ + --batch-size $MAX_SENTENCES \ + --update-freq $UPDATE_FREQ --max-update $TOTAL_UPDATES \ + --save-dir checkpoint/roberta \ + --ddp-backend legacy_ddp --encoder-layerdrop 0.2 \ + --quant-noise-pq 0.2 --quant-noise-pq-block-size 8 --untie-weights-roberta +``` + +To **finetune** RoBERTa + QuantNoise, we followed this setting [here](https://github.com/pytorch/fairseq/blob/master/examples/roberta/README.glue.md). +The following command can be used to finetune a RoBERTa Base + QuantNoise model on the RTE dataset: + +```bash +TOTAL_NUM_UPDATES=2036 +WARMUP_UPDATES=122 +LR=2e-05 +NUM_CLASSES=2 +MAX_SENTENCES=16 +ROBERTA_PATH=/path/to/roberta_quantnoise/model.pt + +fairseq-train /path/to/rte/data/ \ + --restore-file $ROBERTA_PATH \ + --max-positions 512 \ + --batch-size $MAX_SENTENCES \ + --max-tokens 4400 \ + --task sentence_prediction \ + --reset-optimizer --reset-dataloader --reset-meters \ + --required-batch-size-multiple 1 \ + --init-token 0 --separator-token 2 \ + --arch roberta_large \ + --criterion sentence_prediction \ + --num-classes $NUM_CLASSES \ + --dropout 0.1 --attention-dropout 0.1 \ + --weight-decay 0.1 --optimizer adam --adam-betas "(0.9, 0.98)" --adam-eps 1e-06 \ + --clip-norm 0.0 \ + --lr-scheduler polynomial_decay --lr $LR --total-num-update $TOTAL_NUM_UPDATES --warmup-updates $WARMUP_UPDATES \ + --fp16 --fp16-init-scale 4 --threshold-loss-scale 1 --fp16-scale-window 128 \ + --max-epoch 10 \ + --find-unused-parameters \ + --best-checkpoint-metric accuracy --maximize-best-checkpoint-metric \ + --ddp-backend legacy_ddp \ + --quant-noise-pq 0.2 --quant-noise-pq-block-size 8 +``` + +To **train** Language Models on Wikitext-103, we followed this setting [here](https://github.com/pytorch/fairseq/tree/master/examples/language_model). +The following command can be used to train a Transformer + QuantNoise model on Wikitext-103: + +```bash +fairseq-train --task language_modeling /path/to/wikitext-103/data \ + --save-dir checkpoints/transformer_wikitext-103 \ + --adaptive-input --adaptive-input-cutoff 20000,60000 --adaptive-input-factor 4 \ + --adaptive-softmax-cutoff 20000,60000 --adaptive-softmax-dropout 0.2 --adaptive-softmax-factor 4.0 \ + --tie-adaptive-proj --tie-adaptive-weights \ + --arch transformer_lm_gbw \ + --attention-dropout 0.1 --dropout 0.2 --relu-dropout 0.1 \ + --clip-norm 0.1 --criterion adaptive_loss \ + --ddp-backend legacy_ddp \ + --decoder-attention-heads 8 --decoder-embed-dim 1024 --decoder-ffn-embed-dim 4096 --decoder-input-dim 1024 \ + --decoder-layers 16 --decoder-normalize-before --decoder-output-dim 1024 \ + --min-lr 0.0001 --lr-period-updates 270000 --lr-scheduler cosine --lr-shrink 0.75 --lr 1.0 --t-mult 2.0 \ + --max-tokens 3072 --tokens-per-sample 3072 --momentum 0.99 --optimizer nag \ + --sample-break-mode none --update-freq 3 \ + --warmup-init-lr 1e-07 --warmup-updates 16000 \ + --weight-decay 0 --seed 1 --stop-min-lr 1e-09 \ + --quant-noise-pq 0.05 --quant-noise-pq-block-size 8 +``` + +To **evaluate** this model, note you need to use the `eval.py` script. The following command can be used to evaluate: + +```bash +fairseq-eval-lm /path/to/wikitext-103/data --path /path/to/model/checkpoint \ + --sample-break-mode complete \ + --max-tokens 3072 \ + --context-window 2560 \ + --softmax-batch 1024 \ + --gen-subset valid +``` +and change the `--gen-subset` to `test` if you would like to evaluate on the test set instead. + + +### Iterative Product Quantization + +To quantize the finetuned RoBERTa model, we use this command on 1 GPU. This should run in a day. +```bash +TOTAL_NUM_UPDATES=6108 # 2036 updates for each iteration +WARMUP_UPDATES=122 +LR=2e-05 +NUM_CLASSES=2 +MAX_SENTENCES=16 +fairseq-train --task sentence_prediction /path/to/data/ \ + --restore-file $ROBERTA_PATH \ + --save-dir checkpoints/roberta_finetuned \ + --max-positions 512 \ + --batch-size $MAX_SENTENCES \ + --max-tokens 4400 \ + --init-token 0 --separator-token 2 \ + --arch roberta_large \ + --criterion sentence_prediction \ + --num-classes $NUM_CLASSES \ + --dropout 0.1 --attention-dropout 0.1 \ + --weight-decay 0.1 --optimizer adam --adam-betas "(0.9, 0.98)" --adam-eps 1e-06 \ + --clip-norm 0.0 --lr-scheduler polynomial_decay \ + --fp16 --fp16-init-scale 4 --threshold-loss-scale 1 --fp16-scale-window 128 \ + --no-progress-bar --skip-invalid-size-inputs-valid-test --ddp-backend legacy_ddp \ + --quantization-config-path /path/to/config/yaml +``` + +To quantize the trained Language Model, we use this command on 8 V100 23GB GPUs. This should run in a couple of hours. +```bash +fairseq-train --task language_modeling /path/to/wikitext-103/data \ + --save-dir checkpoints/transformer_wikitext-103 \ + --adaptive-input --adaptive-input-cutoff 20000,60000 --adaptive-input-factor 4 \ + --adaptive-softmax-cutoff 20000,60000 --adaptive-softmax-dropout 0.2 --adaptive-softmax-factor 4.0 \ + --arch transformer_lm_gbw \ + --attention-dropout 0.1 --dropout 0.2 --relu-dropout 0.1 \ + --bucket-cap-mb 25 --char-embedder-highway-layers 2 --character-embedding-dim 4 \ + --clip-norm 0.1 --criterion adaptive_loss \ + --ddp-backend legacy_ddp \ + --decoder-attention-heads 8 --decoder-embed-dim 1024 --decoder-ffn-embed-dim 4096 --decoder-input-dim 1024 --decoder-layers 16 --decoder-normalize-before --decoder-output-dim 1024 \ + --fp16 --keep-last-epochs -1 \ + --min-lr 0.0001 --lr-period-updates 270000 --lr-scheduler cosine --lr-shrink 0.75 --lr 0.05 --stop-min-lr 1e-09 \ + --max-tokens 2944 --tokens-per-sample 2944\ + --momentum 0.99 --no-epoch-checkpoints --no-progress-bar --optimizer nag --required-batch-size-multiple 8 \ + --sample-break-mode none --t-mult 2.0 --skip-invalid-size-inputs-valid-test \ + --tie-adaptive-proj --tie-adaptive-weights --update-freq 3 --weight-decay 0 --seed 1 \ + --log-interval 100 --no-progress-bar --skip-invalid-size-inputs-valid-test \ + --restore-file path/to/trained/lm/with/quant/noise \ + --max-update 13500 --quantization-config-path /path/to/config/yaml +``` +If you have less capacity or if your distributed training freezes, try reducing `--max-tokens` and `--tokens-per-sample` (this may reduce the quantized accuracy a bit). + +### Remarks + +We try to keep the open-sourced code as readable and as easy-to-plug as possible. Therefore, we did not test it for the following cases: +- Scalar quantization with RoBERTa. +- Quantization with iPQ and `int8` combined. + +If you have trouble adapting it, we will be more than happy to help! + +## Looking to reproduce the Vision results in the paper? + +We are working on open sourcing our code as part of ClassyVision. Please check back. + + +## Having an issue or have a question? + +Please open an issue in this repository with the details of your question. Thanks! diff --git a/examples/quant_noise/transformer_quantization_config.yaml b/examples/quant_noise/transformer_quantization_config.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d4be14a93a3593f8e6dc66c3b05061bfdde3e0e0 --- /dev/null +++ b/examples/quant_noise/transformer_quantization_config.yaml @@ -0,0 +1,33 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +# This file defines example configuration arguments for quantizing +# a transformer model with product quantization + +# Number of Centroids for Product Quantization, by default 256 (byte-aligned) +n_centroids: + Linear: + key: in_features + value: {"*": 256} + Embedding: + key: embedding_dim + value: {"*": 256} + +# Block Sizes for Product Quantization +# We suggest: 8 for FFN, 4 for ATTN, 4 for embedding projections, 8 for embeddings +block_sizes: + Linear: + key: fuzzy_name + value: {fc: 8, attn: 4, emb: 4} + Embedding: + key: fuzzy_name + value: {emb: 8} + +# Layers to Quantize Sequentially +# We suggest: first FFN, then EMB, then ATTN +layers_to_quantize: + - decoder\\.layers\\.\d+\\.fc[12] + - decoder\\.embed_tokens\\.embeddings\\.[012]\\.[01] + - decoder\\.layers\\.\d+\\.self_attn\\.(k_proj|v_proj|q_proj|out_proj) diff --git a/examples/roberta/README.custom_classification.md b/examples/roberta/README.custom_classification.md new file mode 100644 index 0000000000000000000000000000000000000000..7254bb7d178760ef5b847901bbcac3711af33ca2 --- /dev/null +++ b/examples/roberta/README.custom_classification.md @@ -0,0 +1,168 @@ +# Finetuning RoBERTa on a custom classification task + +This example shows how to finetune RoBERTa on the IMDB dataset, but should illustrate the process for most classification tasks. + +### 1) Get the data + +```bash +wget http://ai.stanford.edu/~amaas/data/sentiment/aclImdb_v1.tar.gz +tar zxvf aclImdb_v1.tar.gz +``` + + +### 2) Format data + +`IMDB` data has one data-sample in each file, below python code-snippet converts it one file for train and valid each for ease of processing. +```python +import argparse +import os +import random +from glob import glob + +random.seed(0) + +def main(args): + for split in ['train', 'test']: + samples = [] + for class_label in ['pos', 'neg']: + fnames = glob(os.path.join(args.datadir, split, class_label) + '/*.txt') + for fname in fnames: + with open(fname) as fin: + line = fin.readline() + samples.append((line, 1 if class_label == 'pos' else 0)) + random.shuffle(samples) + out_fname = 'train' if split == 'train' else 'dev' + f1 = open(os.path.join(args.datadir, out_fname + '.input0'), 'w') + f2 = open(os.path.join(args.datadir, out_fname + '.label'), 'w') + for sample in samples: + f1.write(sample[0] + '\n') + f2.write(str(sample[1]) + '\n') + f1.close() + f2.close() + +if __name__ == '__main__': + parser = argparse.ArgumentParser() + parser.add_argument('--datadir', default='aclImdb') + args = parser.parse_args() + main(args) +``` + + +### 3) BPE encode + +Run `multiprocessing_bpe_encoder`, you can also do this in previous step for each sample but that might be slower. +```bash +# Download encoder.json and vocab.bpe +wget -N 'https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/encoder.json' +wget -N 'https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/vocab.bpe' + +for SPLIT in train dev; do + python -m examples.roberta.multiprocessing_bpe_encoder \ + --encoder-json encoder.json \ + --vocab-bpe vocab.bpe \ + --inputs "aclImdb/$SPLIT.input0" \ + --outputs "aclImdb/$SPLIT.input0.bpe" \ + --workers 60 \ + --keep-empty +done +``` + + +### 4) Preprocess data + +```bash +# Download fairseq dictionary. +wget -N 'https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/dict.txt' + +fairseq-preprocess \ + --only-source \ + --trainpref "aclImdb/train.input0.bpe" \ + --validpref "aclImdb/dev.input0.bpe" \ + --destdir "IMDB-bin/input0" \ + --workers 60 \ + --srcdict dict.txt + +fairseq-preprocess \ + --only-source \ + --trainpref "aclImdb/train.label" \ + --validpref "aclImdb/dev.label" \ + --destdir "IMDB-bin/label" \ + --workers 60 + +``` + + +### 5) Run training + +```bash +TOTAL_NUM_UPDATES=7812 # 10 epochs through IMDB for bsz 32 +WARMUP_UPDATES=469 # 6 percent of the number of updates +LR=1e-05 # Peak LR for polynomial LR scheduler. +HEAD_NAME=imdb_head # Custom name for the classification head. +NUM_CLASSES=2 # Number of classes for the classification task. +MAX_SENTENCES=8 # Batch size. +ROBERTA_PATH=/path/to/roberta.large/model.pt + +CUDA_VISIBLE_DEVICES=0 fairseq-train IMDB-bin/ \ + --restore-file $ROBERTA_PATH \ + --max-positions 512 \ + --batch-size $MAX_SENTENCES \ + --max-tokens 4400 \ + --task sentence_prediction \ + --reset-optimizer --reset-dataloader --reset-meters \ + --required-batch-size-multiple 1 \ + --init-token 0 --separator-token 2 \ + --arch roberta_large \ + --criterion sentence_prediction \ + --classification-head-name $HEAD_NAME \ + --num-classes $NUM_CLASSES \ + --dropout 0.1 --attention-dropout 0.1 \ + --weight-decay 0.1 --optimizer adam --adam-betas "(0.9, 0.98)" --adam-eps 1e-06 \ + --clip-norm 0.0 \ + --lr-scheduler polynomial_decay --lr $LR --total-num-update $TOTAL_NUM_UPDATES --warmup-updates $WARMUP_UPDATES \ + --fp16 --fp16-init-scale 4 --threshold-loss-scale 1 --fp16-scale-window 128 \ + --max-epoch 10 \ + --best-checkpoint-metric accuracy --maximize-best-checkpoint-metric \ + --shorten-method "truncate" \ + --find-unused-parameters \ + --update-freq 4 +``` + +The above command will finetune RoBERTa-large with an effective batch-size of 32 +sentences (`--batch-size=8 --update-freq=4`). The expected +`best-validation-accuracy` after 10 epochs is ~96.5%. + +If you run out of GPU memory, try decreasing `--batch-size` and increase +`--update-freq` to compensate. + + +### 6) Load model using hub interface + +Now we can load the trained model checkpoint using the RoBERTa hub interface. + +Assuming your checkpoints are stored in `checkpoints/`: +```python +from fairseq.models.roberta import RobertaModel +roberta = RobertaModel.from_pretrained( + 'checkpoints', + checkpoint_file='checkpoint_best.pt', + data_name_or_path='IMDB-bin' +) +roberta.eval() # disable dropout +``` + +Finally you can make predictions using the `imdb_head` (or whatever you set +`--classification-head-name` to during training): +```python +label_fn = lambda label: roberta.task.label_dictionary.string( + [label + roberta.task.label_dictionary.nspecial] +) + +tokens = roberta.encode('Best movie this year') +pred = label_fn(roberta.predict('imdb_head', tokens).argmax().item()) +assert pred == '1' # positive + +tokens = roberta.encode('Worst movie ever') +pred = label_fn(roberta.predict('imdb_head', tokens).argmax().item()) +assert pred == '0' # negative +``` diff --git a/examples/roberta/README.glue.md b/examples/roberta/README.glue.md new file mode 100644 index 0000000000000000000000000000000000000000..77015d2e2f76fb7d62fe20c504d14b0c817f19c9 --- /dev/null +++ b/examples/roberta/README.glue.md @@ -0,0 +1,99 @@ +# Finetuning RoBERTa on GLUE tasks + +### 1) Download the data from GLUE website (https://gluebenchmark.com/tasks) using following commands: +```bash +wget https://gist.githubusercontent.com/W4ngatang/60c2bdb54d156a41194446737ce03e2e/raw/17b8dd0d724281ed7c3b2aeeda662b92809aadd5/download_glue_data.py +python download_glue_data.py --data_dir glue_data --tasks all +``` + +### 2) Preprocess GLUE task data: +```bash +./examples/roberta/preprocess_GLUE_tasks.sh glue_data <glue_task_name> +``` +`glue_task_name` is one of the following: +`{ALL, QQP, MNLI, QNLI, MRPC, RTE, STS-B, SST-2, CoLA}` +Use `ALL` for preprocessing all the glue tasks. + +### 3) Fine-tuning on GLUE task: +Example fine-tuning cmd for `RTE` task +```bash +TOTAL_NUM_UPDATES=2036 # 10 epochs through RTE for bsz 16 +WARMUP_UPDATES=122 # 6 percent of the number of updates +LR=2e-05 # Peak LR for polynomial LR scheduler. +NUM_CLASSES=2 +MAX_SENTENCES=16 # Batch size. +ROBERTA_PATH=/path/to/roberta/model.pt + +CUDA_VISIBLE_DEVICES=0 fairseq-train RTE-bin/ \ + --restore-file $ROBERTA_PATH \ + --max-positions 512 \ + --batch-size $MAX_SENTENCES \ + --max-tokens 4400 \ + --task sentence_prediction \ + --reset-optimizer --reset-dataloader --reset-meters \ + --required-batch-size-multiple 1 \ + --init-token 0 --separator-token 2 \ + --arch roberta_large \ + --criterion sentence_prediction \ + --num-classes $NUM_CLASSES \ + --dropout 0.1 --attention-dropout 0.1 \ + --weight-decay 0.1 --optimizer adam --adam-betas "(0.9, 0.98)" --adam-eps 1e-06 \ + --clip-norm 0.0 \ + --lr-scheduler polynomial_decay --lr $LR --total-num-update $TOTAL_NUM_UPDATES --warmup-updates $WARMUP_UPDATES \ + --fp16 --fp16-init-scale 4 --threshold-loss-scale 1 --fp16-scale-window 128 \ + --max-epoch 10 \ + --find-unused-parameters \ + --best-checkpoint-metric accuracy --maximize-best-checkpoint-metric; +``` + +For each of the GLUE task, you will need to use following cmd-line arguments: + +Model | MNLI | QNLI | QQP | RTE | SST-2 | MRPC | CoLA | STS-B +---|---|---|---|---|---|---|---|--- +`--num-classes` | 3 | 2 | 2 | 2 | 2 | 2 | 2 | 1 +`--lr` | 1e-5 | 1e-5 | 1e-5 | 2e-5 | 1e-5 | 1e-5 | 1e-5 | 2e-5 +`--batch-size` | 32 | 32 | 32 | 16 | 32 | 16 | 16 | 16 +`--total-num-update` | 123873 | 33112 | 113272 | 2036 | 20935 | 2296 | 5336 | 3598 +`--warmup-updates` | 7432 | 1986 | 28318 | 122 | 1256 | 137 | 320 | 214 + +For `STS-B` additionally add `--regression-target --best-checkpoint-metric loss` and remove `--maximize-best-checkpoint-metric`. + +**Note:** + +a) `--total-num-updates` is used by `--polynomial_decay` scheduler and is calculated for `--max-epoch=10` and `--batch-size=16/32` depending on the task. + +b) Above cmd-args and hyperparams are tested on one Nvidia `V100` GPU with `32gb` of memory for each task. Depending on the GPU memory resources available to you, you can use increase `--update-freq` and reduce `--batch-size`. + +c) All the settings in above table are suggested settings based on our hyperparam search within a fixed search space (for careful comparison across models). You might be able to find better metrics with wider hyperparam search. + +### Inference on GLUE task +After training the model as mentioned in previous step, you can perform inference with checkpoints in `checkpoints/` directory using following python code snippet: + +```python +from fairseq.models.roberta import RobertaModel + +roberta = RobertaModel.from_pretrained( + 'checkpoints/', + checkpoint_file='checkpoint_best.pt', + data_name_or_path='RTE-bin' +) + +label_fn = lambda label: roberta.task.label_dictionary.string( + [label + roberta.task.label_dictionary.nspecial] +) +ncorrect, nsamples = 0, 0 +roberta.cuda() +roberta.eval() +with open('glue_data/RTE/dev.tsv') as fin: + fin.readline() + for index, line in enumerate(fin): + tokens = line.strip().split('\t') + sent1, sent2, target = tokens[1], tokens[2], tokens[3] + tokens = roberta.encode(sent1, sent2) + prediction = roberta.predict('sentence_classification_head', tokens).argmax().item() + prediction_label = label_fn(prediction) + ncorrect += int(prediction_label == target) + nsamples += 1 +print('| Accuracy: ', float(ncorrect)/float(nsamples)) + +``` diff --git a/examples/roberta/README.md b/examples/roberta/README.md new file mode 100644 index 0000000000000000000000000000000000000000..58091b2c7d7949e10fe963c7e85d0c727a006b5e --- /dev/null +++ b/examples/roberta/README.md @@ -0,0 +1,296 @@ +# RoBERTa: A Robustly Optimized BERT Pretraining Approach + +https://arxiv.org/abs/1907.11692 + +## Introduction + +RoBERTa iterates on BERT's pretraining procedure, including training the model longer, with bigger batches over more data; removing the next sentence prediction objective; training on longer sequences; and dynamically changing the masking pattern applied to the training data. See the associated paper for more details. + +### What's New: + +- December 2020: German model (GottBERT) is available: [GottBERT](https://github.com/pytorch/fairseq/tree/master/examples/gottbert). +- January 2020: Italian model (UmBERTo) is available from Musixmatch Research: [UmBERTo](https://github.com/musixmatchresearch/umberto). +- November 2019: French model (CamemBERT) is available: [CamemBERT](https://github.com/pytorch/fairseq/tree/master/examples/camembert). +- November 2019: Multilingual encoder (XLM-RoBERTa) is available: [XLM-R](https://github.com/pytorch/fairseq/tree/master/examples/xlmr). +- September 2019: TensorFlow and TPU support via the [transformers library](https://github.com/huggingface/transformers). +- August 2019: RoBERTa is now supported in the [pytorch-transformers library](https://github.com/huggingface/pytorch-transformers). +- August 2019: Added [tutorial for finetuning on WinoGrande](https://github.com/pytorch/fairseq/tree/master/examples/roberta/wsc#roberta-training-on-winogrande-dataset). +- August 2019: Added [tutorial for pretraining RoBERTa using your own data](README.pretraining.md). + +## Pre-trained models + +Model | Description | # params | Download +---|---|---|--- +`roberta.base` | RoBERTa using the BERT-base architecture | 125M | [roberta.base.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/roberta.base.tar.gz) +`roberta.large` | RoBERTa using the BERT-large architecture | 355M | [roberta.large.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/roberta.large.tar.gz) +`roberta.large.mnli` | `roberta.large` finetuned on [MNLI](http://www.nyu.edu/projects/bowman/multinli) | 355M | [roberta.large.mnli.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/roberta.large.mnli.tar.gz) +`roberta.large.wsc` | `roberta.large` finetuned on [WSC](wsc/README.md) | 355M | [roberta.large.wsc.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/roberta.large.wsc.tar.gz) + +## Results + +**[GLUE (Wang et al., 2019)](https://gluebenchmark.com/)** +_(dev set, single model, single-task finetuning)_ + +Model | MNLI | QNLI | QQP | RTE | SST-2 | MRPC | CoLA | STS-B +---|---|---|---|---|---|---|---|--- +`roberta.base` | 87.6 | 92.8 | 91.9 | 78.7 | 94.8 | 90.2 | 63.6 | 91.2 +`roberta.large` | 90.2 | 94.7 | 92.2 | 86.6 | 96.4 | 90.9 | 68.0 | 92.4 +`roberta.large.mnli` | 90.2 | - | - | - | - | - | - | - + +**[SuperGLUE (Wang et al., 2019)](https://super.gluebenchmark.com/)** +_(dev set, single model, single-task finetuning)_ + +Model | BoolQ | CB | COPA | MultiRC | RTE | WiC | WSC +---|---|---|---|---|---|---|--- +`roberta.large` | 86.9 | 98.2 | 94.0 | 85.7 | 89.5 | 75.6 | - +`roberta.large.wsc` | - | - | - | - | - | - | 91.3 + +**[SQuAD (Rajpurkar et al., 2018)](https://rajpurkar.github.io/SQuAD-explorer/)** +_(dev set, no additional data used)_ + +Model | SQuAD 1.1 EM/F1 | SQuAD 2.0 EM/F1 +---|---|--- +`roberta.large` | 88.9/94.6 | 86.5/89.4 + +**[RACE (Lai et al., 2017)](http://www.qizhexie.com/data/RACE_leaderboard.html)** +_(test set)_ + +Model | Accuracy | Middle | High +---|---|---|--- +`roberta.large` | 83.2 | 86.5 | 81.3 + +**[HellaSwag (Zellers et al., 2019)](https://rowanzellers.com/hellaswag/)** +_(test set)_ + +Model | Overall | In-domain | Zero-shot | ActivityNet | WikiHow +---|---|---|---|---|--- +`roberta.large` | 85.2 | 87.3 | 83.1 | 74.6 | 90.9 + +**[Commonsense QA (Talmor et al., 2019)](https://www.tau-nlp.org/commonsenseqa)** +_(test set)_ + +Model | Accuracy +---|--- +`roberta.large` (single model) | 72.1 +`roberta.large` (ensemble) | 72.5 + +**[Winogrande (Sakaguchi et al., 2019)](https://arxiv.org/abs/1907.10641)** +_(test set)_ + +Model | Accuracy +---|--- +`roberta.large` | 78.1 + +**[XNLI (Conneau et al., 2018)](https://arxiv.org/abs/1809.05053)** +_(TRANSLATE-TEST)_ + +Model | en | fr | es | de | el | bg | ru | tr | ar | vi | th | zh | hi | sw | ur +---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|--- +`roberta.large.mnli` | 91.3 | 82.91 | 84.27 | 81.24 | 81.74 | 83.13 | 78.28 | 76.79 | 76.64 | 74.17 | 74.05 | 77.5 | 70.9 | 66.65 | 66.81 + +## Example usage + +##### Load RoBERTa from torch.hub (PyTorch >= 1.1): +```python +import torch +roberta = torch.hub.load('pytorch/fairseq', 'roberta.large') +roberta.eval() # disable dropout (or leave in train mode to finetune) +``` + +##### Load RoBERTa (for PyTorch 1.0 or custom models): +```python +# Download roberta.large model +wget https://dl.fbaipublicfiles.com/fairseq/models/roberta.large.tar.gz +tar -xzvf roberta.large.tar.gz + +# Load the model in fairseq +from fairseq.models.roberta import RobertaModel +roberta = RobertaModel.from_pretrained('/path/to/roberta.large', checkpoint_file='model.pt') +roberta.eval() # disable dropout (or leave in train mode to finetune) +``` + +##### Apply Byte-Pair Encoding (BPE) to input text: +```python +tokens = roberta.encode('Hello world!') +assert tokens.tolist() == [0, 31414, 232, 328, 2] +roberta.decode(tokens) # 'Hello world!' +``` + +##### Extract features from RoBERTa: +```python +# Extract the last layer's features +last_layer_features = roberta.extract_features(tokens) +assert last_layer_features.size() == torch.Size([1, 5, 1024]) + +# Extract all layer's features (layer 0 is the embedding layer) +all_layers = roberta.extract_features(tokens, return_all_hiddens=True) +assert len(all_layers) == 25 +assert torch.all(all_layers[-1] == last_layer_features) +``` + +##### Use RoBERTa for sentence-pair classification tasks: +```python +# Download RoBERTa already finetuned for MNLI +roberta = torch.hub.load('pytorch/fairseq', 'roberta.large.mnli') +roberta.eval() # disable dropout for evaluation + +# Encode a pair of sentences and make a prediction +tokens = roberta.encode('Roberta is a heavily optimized version of BERT.', 'Roberta is not very optimized.') +roberta.predict('mnli', tokens).argmax() # 0: contradiction + +# Encode another pair of sentences +tokens = roberta.encode('Roberta is a heavily optimized version of BERT.', 'Roberta is based on BERT.') +roberta.predict('mnli', tokens).argmax() # 2: entailment +``` + +##### Register a new (randomly initialized) classification head: +```python +roberta.register_classification_head('new_task', num_classes=3) +logprobs = roberta.predict('new_task', tokens) # tensor([[-1.1050, -1.0672, -1.1245]], grad_fn=<LogSoftmaxBackward>) +``` + +##### Batched prediction: +```python +import torch +from fairseq.data.data_utils import collate_tokens + +roberta = torch.hub.load('pytorch/fairseq', 'roberta.large.mnli') +roberta.eval() + +batch_of_pairs = [ + ['Roberta is a heavily optimized version of BERT.', 'Roberta is not very optimized.'], + ['Roberta is a heavily optimized version of BERT.', 'Roberta is based on BERT.'], + ['potatoes are awesome.', 'I like to run.'], + ['Mars is very far from earth.', 'Mars is very close.'], +] + +batch = collate_tokens( + [roberta.encode(pair[0], pair[1]) for pair in batch_of_pairs], pad_idx=1 +) + +logprobs = roberta.predict('mnli', batch) +print(logprobs.argmax(dim=1)) +# tensor([0, 2, 1, 0]) +``` + +##### Using the GPU: +```python +roberta.cuda() +roberta.predict('new_task', tokens) # tensor([[-1.1050, -1.0672, -1.1245]], device='cuda:0', grad_fn=<LogSoftmaxBackward>) +``` + +## Advanced usage + +#### Filling masks: + +RoBERTa can be used to fill `<mask>` tokens in the input. Some examples from the +[Natural Questions dataset](https://ai.google.com/research/NaturalQuestions/): +```python +roberta.fill_mask('The first Star wars movie came out in <mask>', topk=3) +# [('The first Star wars movie came out in 1977', 0.9504708051681519, ' 1977'), ('The first Star wars movie came out in 1978', 0.009986862540245056, ' 1978'), ('The first Star wars movie came out in 1979', 0.009574787691235542, ' 1979')] + +roberta.fill_mask('Vikram samvat calender is official in <mask>', topk=3) +# [('Vikram samvat calender is official in India', 0.21878819167613983, ' India'), ('Vikram samvat calender is official in Delhi', 0.08547237515449524, ' Delhi'), ('Vikram samvat calender is official in Gujarat', 0.07556215673685074, ' Gujarat')] + +roberta.fill_mask('<mask> is the common currency of the European Union', topk=3) +# [('Euro is the common currency of the European Union', 0.9456493854522705, 'Euro'), ('euro is the common currency of the European Union', 0.025748178362846375, 'euro'), ('€ is the common currency of the European Union', 0.011183084920048714, '€')] +``` + +#### Pronoun disambiguation (Winograd Schema Challenge): + +RoBERTa can be used to disambiguate pronouns. First install spaCy and download the English-language model: +```bash +pip install spacy +python -m spacy download en_core_web_lg +``` + +Next load the `roberta.large.wsc` model and call the `disambiguate_pronoun` +function. The pronoun should be surrounded by square brackets (`[]`) and the +query referent surrounded by underscores (`_`), or left blank to return the +predicted candidate text directly: +```python +roberta = torch.hub.load('pytorch/fairseq', 'roberta.large.wsc', user_dir='examples/roberta/wsc') +roberta.cuda() # use the GPU (optional) + +roberta.disambiguate_pronoun('The _trophy_ would not fit in the brown suitcase because [it] was too big.') +# True +roberta.disambiguate_pronoun('The trophy would not fit in the brown _suitcase_ because [it] was too big.') +# False + +roberta.disambiguate_pronoun('The city councilmen refused the demonstrators a permit because [they] feared violence.') +# 'The city councilmen' +roberta.disambiguate_pronoun('The city councilmen refused the demonstrators a permit because [they] advocated violence.') +# 'demonstrators' +``` + +See the [RoBERTA Winograd Schema Challenge (WSC) README](wsc/README.md) for more details on how to train this model. + +#### Extract features aligned to words: + +By default RoBERTa outputs one feature vector per BPE token. You can instead +realign the features to match [spaCy's word-level tokenization](https://spacy.io/usage/linguistic-features#tokenization) +with the `extract_features_aligned_to_words` method. This will compute a +weighted average of the BPE-level features for each word and expose them in +spaCy's `Token.vector` attribute: +```python +doc = roberta.extract_features_aligned_to_words('I said, "hello RoBERTa."') +assert len(doc) == 10 +for tok in doc: + print('{:10}{} (...)'.format(str(tok), tok.vector[:5])) +# <s> tensor([-0.1316, -0.0386, -0.0832, -0.0477, 0.1943], grad_fn=<SliceBackward>) (...) +# I tensor([ 0.0559, 0.1541, -0.4832, 0.0880, 0.0120], grad_fn=<SliceBackward>) (...) +# said tensor([-0.1565, -0.0069, -0.8915, 0.0501, -0.0647], grad_fn=<SliceBackward>) (...) +# , tensor([-0.1318, -0.0387, -0.0834, -0.0477, 0.1944], grad_fn=<SliceBackward>) (...) +# " tensor([-0.0486, 0.1818, -0.3946, -0.0553, 0.0981], grad_fn=<SliceBackward>) (...) +# hello tensor([ 0.0079, 0.1799, -0.6204, -0.0777, -0.0923], grad_fn=<SliceBackward>) (...) +# RoBERTa tensor([-0.2339, -0.1184, -0.7343, -0.0492, 0.5829], grad_fn=<SliceBackward>) (...) +# . tensor([-0.1341, -0.1203, -0.1012, -0.0621, 0.1892], grad_fn=<SliceBackward>) (...) +# " tensor([-0.1341, -0.1203, -0.1012, -0.0621, 0.1892], grad_fn=<SliceBackward>) (...) +# </s> tensor([-0.0930, -0.0392, -0.0821, 0.0158, 0.0649], grad_fn=<SliceBackward>) (...) +``` + +#### Evaluating the `roberta.large.mnli` model: + +Example python code snippet to evaluate accuracy on the MNLI `dev_matched` set. +```python +label_map = {0: 'contradiction', 1: 'neutral', 2: 'entailment'} +ncorrect, nsamples = 0, 0 +roberta.cuda() +roberta.eval() +with open('glue_data/MNLI/dev_matched.tsv') as fin: + fin.readline() + for index, line in enumerate(fin): + tokens = line.strip().split('\t') + sent1, sent2, target = tokens[8], tokens[9], tokens[-1] + tokens = roberta.encode(sent1, sent2) + prediction = roberta.predict('mnli', tokens).argmax().item() + prediction_label = label_map[prediction] + ncorrect += int(prediction_label == target) + nsamples += 1 +print('| Accuracy: ', float(ncorrect)/float(nsamples)) +# Expected output: 0.9060 +``` + +## Finetuning + +- [Finetuning on GLUE](README.glue.md) +- [Finetuning on custom classification tasks (e.g., IMDB)](README.custom_classification.md) +- [Finetuning on Winograd Schema Challenge (WSC)](wsc/README.md) +- [Finetuning on Commonsense QA (CQA)](commonsense_qa/README.md) + +## Pretraining using your own data + +See the [tutorial for pretraining RoBERTa using your own data](README.pretraining.md). + +## Citation + +```bibtex +@article{liu2019roberta, + title = {RoBERTa: A Robustly Optimized BERT Pretraining Approach}, + author = {Yinhan Liu and Myle Ott and Naman Goyal and Jingfei Du and + Mandar Joshi and Danqi Chen and Omer Levy and Mike Lewis and + Luke Zettlemoyer and Veselin Stoyanov}, + journal={arXiv preprint arXiv:1907.11692}, + year = {2019}, +} +``` diff --git a/examples/roberta/README.pretraining.md b/examples/roberta/README.pretraining.md new file mode 100644 index 0000000000000000000000000000000000000000..8b6e10c08c14713e7f3f7ee37c44e9b6a662df06 --- /dev/null +++ b/examples/roberta/README.pretraining.md @@ -0,0 +1,98 @@ +# Pretraining RoBERTa using your own data + +This tutorial will walk you through pretraining RoBERTa over your own data. + +### 1) Preprocess the data + +Data should be preprocessed following the [language modeling format](/examples/language_model), i.e. each document should be separated by an empty line (only useful with `--sample-break-mode complete_doc`). Lines will be concatenated as a 1D text stream during training. + +We'll use the [WikiText-103 dataset](https://www.salesforce.com/products/einstein/ai-research/the-wikitext-dependency-language-modeling-dataset/) +to demonstrate how to preprocess raw text data with the GPT-2 BPE. Of course +this dataset is quite small, so the resulting pretrained model will perform +poorly, but it gives the general idea. + +First download the dataset: +```bash +wget https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-103-raw-v1.zip +unzip wikitext-103-raw-v1.zip +``` + +Next encode it with the GPT-2 BPE: +```bash +mkdir -p gpt2_bpe +wget -O gpt2_bpe/encoder.json https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/encoder.json +wget -O gpt2_bpe/vocab.bpe https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/vocab.bpe +for SPLIT in train valid test; do \ + python -m examples.roberta.multiprocessing_bpe_encoder \ + --encoder-json gpt2_bpe/encoder.json \ + --vocab-bpe gpt2_bpe/vocab.bpe \ + --inputs wikitext-103-raw/wiki.${SPLIT}.raw \ + --outputs wikitext-103-raw/wiki.${SPLIT}.bpe \ + --keep-empty \ + --workers 60; \ +done +``` + +Finally preprocess/binarize the data using the GPT-2 fairseq dictionary: +```bash +wget -O gpt2_bpe/dict.txt https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/dict.txt +fairseq-preprocess \ + --only-source \ + --srcdict gpt2_bpe/dict.txt \ + --trainpref wikitext-103-raw/wiki.train.bpe \ + --validpref wikitext-103-raw/wiki.valid.bpe \ + --testpref wikitext-103-raw/wiki.test.bpe \ + --destdir data-bin/wikitext-103 \ + --workers 60 +``` + +### 2) Train RoBERTa base +```bash +TOTAL_UPDATES=125000 # Total number of training steps +WARMUP_UPDATES=10000 # Warmup the learning rate over this many updates +PEAK_LR=0.0005 # Peak learning rate, adjust as needed +TOKENS_PER_SAMPLE=512 # Max sequence length +MAX_POSITIONS=512 # Num. positional embeddings (usually same as above) +MAX_SENTENCES=16 # Number of sequences per batch (batch size) +UPDATE_FREQ=16 # Increase the batch size 16x + +DATA_DIR=data-bin/wikitext-103 + +fairseq-train --fp16 $DATA_DIR \ + --task masked_lm --criterion masked_lm \ + --arch roberta_base --sample-break-mode complete --tokens-per-sample $TOKENS_PER_SAMPLE \ + --optimizer adam --adam-betas '(0.9,0.98)' --adam-eps 1e-6 --clip-norm 0.0 \ + --lr-scheduler polynomial_decay --lr $PEAK_LR --warmup-updates $WARMUP_UPDATES --total-num-update $TOTAL_UPDATES \ + --dropout 0.1 --attention-dropout 0.1 --weight-decay 0.01 \ + --batch-size $MAX_SENTENCES --update-freq $UPDATE_FREQ \ + --max-update $TOTAL_UPDATES --log-format simple --log-interval 1 +``` + +**Note:** You can optionally resume training the released RoBERTa base model by +adding `--restore-file /path/to/roberta.base/model.pt`. + +**Note:** The above command assumes training on 8x32GB V100 GPUs. Each GPU uses +a batch size of 16 sequences (`$MAX_SENTENCES`) and accumulates gradients to +further increase the batch size by 16x (`$UPDATE_FREQ`), for a total batch size +of 2048 sequences. If you have fewer GPUs or GPUs with less memory you may need +to reduce `$MAX_SENTENCES` and increase `$UPDATE_FREQ` to compensate. +Alternatively if you have more GPUs you can decrease `$UPDATE_FREQ` accordingly +to increase training speed. + +**Note:** The learning rate and batch size are tightly connected and need to be +adjusted together. We generally recommend increasing the learning rate as you +increase the batch size according to the following table (although it's also +dataset dependent, so don't rely on the following values too closely): + +batch size | peak learning rate +---|--- +256 | 0.0001 +2048 | 0.0005 +8192 | 0.0007 + +### 3) Load your pretrained model +```python +from fairseq.models.roberta import RobertaModel +roberta = RobertaModel.from_pretrained('checkpoints', 'checkpoint_best.pt', 'path/to/data') +assert isinstance(roberta.model, torch.nn.Module) +``` diff --git a/examples/roberta/README.race.md b/examples/roberta/README.race.md new file mode 100644 index 0000000000000000000000000000000000000000..13c917e8eca6621e91dce541c7e41436b38cbdc1 --- /dev/null +++ b/examples/roberta/README.race.md @@ -0,0 +1,68 @@ +# Finetuning RoBERTa on RACE tasks + +### 1) Download the data from RACE website (http://www.cs.cmu.edu/~glai1/data/race/) + +### 2) Preprocess RACE data: +```bash +python ./examples/roberta/preprocess_RACE.py --input-dir <input-dir> --output-dir <extracted-data-dir> +./examples/roberta/preprocess_RACE.sh <extracted-data-dir> <output-dir> +``` + +### 3) Fine-tuning on RACE: + +```bash +MAX_EPOCH=5 # Number of training epochs. +LR=1e-05 # Peak LR for fixed LR scheduler. +NUM_CLASSES=4 +MAX_SENTENCES=1 # Batch size per GPU. +UPDATE_FREQ=8 # Accumulate gradients to simulate training on 8 GPUs. +DATA_DIR=/path/to/race-output-dir +ROBERTA_PATH=/path/to/roberta/model.pt + +CUDA_VISIBLE_DEVICES=0,1 fairseq-train $DATA_DIR --ddp-backend=legacy_ddp \ + --restore-file $ROBERTA_PATH \ + --reset-optimizer --reset-dataloader --reset-meters \ + --best-checkpoint-metric accuracy --maximize-best-checkpoint-metric \ + --task sentence_ranking \ + --num-classes $NUM_CLASSES \ + --init-token 0 --separator-token 2 \ + --max-option-length 128 \ + --max-positions 512 \ + --shorten-method "truncate" \ + --arch roberta_large \ + --dropout 0.1 --attention-dropout 0.1 --weight-decay 0.01 \ + --criterion sentence_ranking \ + --optimizer adam --adam-betas '(0.9, 0.98)' --adam-eps 1e-06 \ + --clip-norm 0.0 \ + --lr-scheduler fixed --lr $LR \ + --fp16 --fp16-init-scale 4 --threshold-loss-scale 1 --fp16-scale-window 128 \ + --batch-size $MAX_SENTENCES \ + --required-batch-size-multiple 1 \ + --update-freq $UPDATE_FREQ \ + --max-epoch $MAX_EPOCH +``` + +**Note:** + +a) As contexts in RACE are relatively long, we are using smaller batch size per GPU while increasing update-freq to achieve larger effective batch size. + +b) Above cmd-args and hyperparams are tested on one Nvidia `V100` GPU with `32gb` of memory for each task. Depending on the GPU memory resources available to you, you can use increase `--update-freq` and reduce `--batch-size`. + +c) The setting in above command is based on our hyperparam search within a fixed search space (for careful comparison across models). You might be able to find better metrics with wider hyperparam search. + +### 4) Evaluation: + +``` +DATA_DIR=/path/to/race-output-dir # data directory used during training +MODEL_PATH=/path/to/checkpoint_best.pt # path to the finetuned model checkpoint +PREDS_OUT=preds.tsv # output file path to save prediction +TEST_SPLIT=test # can be test (Middle) or test1 (High) +fairseq-validate \ + $DATA_DIR \ + --valid-subset $TEST_SPLIT \ + --path $MODEL_PATH \ + --batch-size 1 \ + --task sentence_ranking \ + --criterion sentence_ranking \ + --save-predictions $PREDS_OUT +``` diff --git a/examples/roberta/commonsense_qa/README.md b/examples/roberta/commonsense_qa/README.md new file mode 100644 index 0000000000000000000000000000000000000000..05c6f841a8966d2b74a8d3fe73bca22694fe9a8a --- /dev/null +++ b/examples/roberta/commonsense_qa/README.md @@ -0,0 +1,99 @@ +# Finetuning RoBERTa on Commonsense QA + +We follow a similar approach to [finetuning RACE](../README.race.md). Specifically +for each question we construct five inputs, one for each of the five candidate +answer choices. Each input is constructed by concatenating the question and +candidate answer. We then encode each input and pass the resulting "[CLS]" +representations through a fully-connected layer to predict the correct answer. +We train with a standard cross-entropy loss. + +We also found it helpful to prepend a prefix of `Q:` to the question and `A:` to +the answer. The complete input format is: +``` +<s> Q: Where would I not want a fox? </s> A: hen house </s> +``` + +Our final submission is based on a hyperparameter search over the learning rate +(1e-5, 2e-5, 3e-5), batch size (8, 16), number of training steps (2000, 3000, +4000) and random seed. We selected the model with the best performance on the +development set after 100 trials. + +### 1) Download data from the Commonsense QA website (https://www.tau-nlp.org/commonsenseqa) +```bash +bash examples/roberta/commonsense_qa/download_cqa_data.sh +``` + +### 2) Finetune + +```bash +MAX_UPDATES=3000 # Number of training steps. +WARMUP_UPDATES=150 # Linearly increase LR over this many steps. +LR=1e-05 # Peak LR for polynomial LR scheduler. +MAX_SENTENCES=16 # Batch size. +SEED=1 # Random seed. +ROBERTA_PATH=/path/to/roberta/model.pt +DATA_DIR=data/CommonsenseQA + +# we use the --user-dir option to load the task from +# the examples/roberta/commonsense_qa directory: +FAIRSEQ_PATH=/path/to/fairseq +FAIRSEQ_USER_DIR=${FAIRSEQ_PATH}/examples/roberta/commonsense_qa + +CUDA_VISIBLE_DEVICES=0 fairseq-train --fp16 --ddp-backend=legacy_ddp \ + $DATA_DIR \ + --user-dir $FAIRSEQ_USER_DIR \ + --restore-file $ROBERTA_PATH \ + --reset-optimizer --reset-dataloader --reset-meters \ + --no-epoch-checkpoints --no-last-checkpoints --no-save-optimizer-state \ + --best-checkpoint-metric accuracy --maximize-best-checkpoint-metric \ + --task commonsense_qa --init-token 0 --bpe gpt2 \ + --arch roberta_large --max-positions 512 \ + --dropout 0.1 --attention-dropout 0.1 --weight-decay 0.01 \ + --criterion sentence_ranking --num-classes 5 \ + --optimizer adam --adam-betas '(0.9, 0.98)' --adam-eps 1e-06 --clip-norm 0.0 \ + --lr-scheduler polynomial_decay --lr $LR \ + --warmup-updates $WARMUP_UPDATES --total-num-update $MAX_UPDATES \ + --batch-size $MAX_SENTENCES \ + --max-update $MAX_UPDATES \ + --log-format simple --log-interval 25 \ + --seed $SEED +``` + +The above command assumes training on 1 GPU with 32GB of RAM. For GPUs with +less memory, decrease `--batch-size` and increase `--update-freq` +accordingly to compensate. + +### 3) Evaluate +```python +import json +import torch +from fairseq.models.roberta import RobertaModel +from examples.roberta import commonsense_qa # load the Commonsense QA task +roberta = RobertaModel.from_pretrained('checkpoints', 'checkpoint_best.pt', 'data/CommonsenseQA') +roberta.eval() # disable dropout +roberta.cuda() # use the GPU (optional) +nsamples, ncorrect = 0, 0 +with open('data/CommonsenseQA/valid.jsonl') as h: + for line in h: + example = json.loads(line) + scores = [] + for choice in example['question']['choices']: + input = roberta.encode( + 'Q: ' + example['question']['stem'], + 'A: ' + choice['text'], + no_separator=True + ) + score = roberta.predict('sentence_classification_head', input, return_logits=True) + scores.append(score) + pred = torch.cat(scores).argmax() + answer = ord(example['answerKey']) - ord('A') + nsamples += 1 + if pred == answer: + ncorrect += 1 + +print('Accuracy: ' + str(ncorrect / float(nsamples))) +# Accuracy: 0.7846027846027847 +``` + +The above snippet is not batched, which makes it quite slow. See [instructions +for batched prediction with RoBERTa](https://github.com/pytorch/fairseq/tree/master/examples/roberta#batched-prediction). diff --git a/examples/roberta/commonsense_qa/__init__.py b/examples/roberta/commonsense_qa/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..42d21f35eb3dd33a053dcf0edd5eadd2dff11294 --- /dev/null +++ b/examples/roberta/commonsense_qa/__init__.py @@ -0,0 +1,6 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from . import commonsense_qa_task # noqa diff --git a/examples/roberta/commonsense_qa/commonsense_qa_task.py b/examples/roberta/commonsense_qa/commonsense_qa_task.py new file mode 100644 index 0000000000000000000000000000000000000000..216093f7087a61060767babf5a3f3f4e716a4dfe --- /dev/null +++ b/examples/roberta/commonsense_qa/commonsense_qa_task.py @@ -0,0 +1,190 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import json +import os + +import numpy as np +import torch +from fairseq.data import ( + Dictionary, + IdDataset, + ListDataset, + NestedDictionaryDataset, + NumelDataset, + NumSamplesDataset, + RawLabelDataset, + RightPadDataset, + SortDataset, + data_utils, + encoders, +) +from fairseq.tasks import LegacyFairseqTask, register_task + + +@register_task("commonsense_qa") +class CommonsenseQATask(LegacyFairseqTask): + """Task to finetune RoBERTa for Commonsense QA.""" + + @staticmethod + def add_args(parser): + """Add task-specific arguments to the parser.""" + parser.add_argument( + "data", metavar="DIR", help="path to data directory; we load <split>.jsonl" + ) + parser.add_argument( + "--init-token", + type=int, + default=None, + help="add token at the beginning of each batch item", + ) + parser.add_argument("--num-classes", type=int, default=5) + + def __init__(self, args, vocab): + super().__init__(args) + self.vocab = vocab + self.mask = vocab.add_symbol("<mask>") + + self.bpe = encoders.build_bpe(args) + + @classmethod + def load_dictionary(cls, filename): + """Load the dictionary from the filename + + Args: + filename (str): the filename + """ + dictionary = Dictionary.load(filename) + dictionary.add_symbol("<mask>") + return dictionary + + @classmethod + def setup_task(cls, args, **kwargs): + assert ( + args.criterion == "sentence_ranking" + ), "Must set --criterion=sentence_ranking" + + # load data and label dictionaries + vocab = cls.load_dictionary(os.path.join(args.data, "dict.txt")) + print("| dictionary: {} types".format(len(vocab))) + + return cls(args, vocab) + + def load_dataset( + self, split, epoch=1, combine=False, data_path=None, return_only=False, **kwargs + ): + """Load a given dataset split. + + Args: + split (str): name of the split (e.g., train, valid, test) + """ + + def binarize(s, append_bos=False): + if self.bpe is not None: + s = self.bpe.encode(s) + tokens = self.vocab.encode_line( + s, + append_eos=True, + add_if_not_exist=False, + ).long() + if append_bos and self.args.init_token is not None: + tokens = torch.cat([tokens.new([self.args.init_token]), tokens]) + return tokens + + if data_path is None: + data_path = os.path.join(self.args.data, split + ".jsonl") + if not os.path.exists(data_path): + raise FileNotFoundError("Cannot find data: {}".format(data_path)) + + src_tokens = [[] for i in range(self.args.num_classes)] + src_lengths = [[] for i in range(self.args.num_classes)] + labels = [] + + with open(data_path) as h: + for line in h: + example = json.loads(line.strip()) + if "answerKey" in example: + label = ord(example["answerKey"]) - ord("A") + labels.append(label) + question = example["question"]["stem"] + assert len(example["question"]["choices"]) == self.args.num_classes + # format: `<s> Q: Where would I not want a fox? </s> A: hen house </s>` + question = "Q: " + question + question_toks = binarize(question, append_bos=True) + for i, choice in enumerate(example["question"]["choices"]): + src = "A: " + choice["text"] + src_bin = torch.cat([question_toks, binarize(src)]) + src_tokens[i].append(src_bin) + src_lengths[i].append(len(src_bin)) + assert all( + len(src_tokens[0]) == len(src_tokens[i]) + for i in range(self.args.num_classes) + ) + assert len(src_tokens[0]) == len(src_lengths[0]) + assert len(labels) == 0 or len(labels) == len(src_tokens[0]) + + for i in range(self.args.num_classes): + src_lengths[i] = np.array(src_lengths[i]) + src_tokens[i] = ListDataset(src_tokens[i], src_lengths[i]) + src_lengths[i] = ListDataset(src_lengths[i]) + + dataset = { + "id": IdDataset(), + "nsentences": NumSamplesDataset(), + "ntokens": NumelDataset(src_tokens[0], reduce=True), + } + + for i in range(self.args.num_classes): + dataset.update( + { + "net_input{}".format(i + 1): { + "src_tokens": RightPadDataset( + src_tokens[i], + pad_idx=self.source_dictionary.pad(), + ), + "src_lengths": src_lengths[i], + } + } + ) + + if len(labels) > 0: + dataset.update({"target": RawLabelDataset(labels)}) + + dataset = NestedDictionaryDataset( + dataset, + sizes=[np.maximum.reduce([src_token.sizes for src_token in src_tokens])], + ) + + with data_utils.numpy_seed(self.args.seed): + dataset = SortDataset( + dataset, + # shuffle + sort_order=[np.random.permutation(len(dataset))], + ) + + print("| Loaded {} with {} samples".format(split, len(dataset))) + + self.datasets[split] = dataset + return self.datasets[split] + + def build_model(self, args): + from fairseq import models + + model = models.build_model(args, self) + + model.register_classification_head( + "sentence_classification_head", + num_classes=1, + ) + + return model + + @property + def source_dictionary(self): + return self.vocab + + @property + def target_dictionary(self): + return self.vocab diff --git a/examples/roberta/commonsense_qa/download_cqa_data.sh b/examples/roberta/commonsense_qa/download_cqa_data.sh new file mode 100644 index 0000000000000000000000000000000000000000..5f300093fa0a0feb819d8b6aed307b59e3891d01 --- /dev/null +++ b/examples/roberta/commonsense_qa/download_cqa_data.sh @@ -0,0 +1,14 @@ +#!/bin/bash +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +OUTDIR=data/CommonsenseQA + +mkdir -p $OUTDIR + +wget -O $OUTDIR/train.jsonl https://s3.amazonaws.com/commensenseqa/train_rand_split.jsonl +wget -O $OUTDIR/valid.jsonl https://s3.amazonaws.com/commensenseqa/dev_rand_split.jsonl +wget -O $OUTDIR/test.jsonl https://s3.amazonaws.com/commensenseqa/test_rand_split_no_answers.jsonl +wget -O $OUTDIR/dict.txt https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/dict.txt diff --git a/examples/roberta/multiprocessing_bpe_encoder.py b/examples/roberta/multiprocessing_bpe_encoder.py new file mode 100644 index 0000000000000000000000000000000000000000..43fe0451bf4d5762d734314075b1402c2a8db2bb --- /dev/null +++ b/examples/roberta/multiprocessing_bpe_encoder.py @@ -0,0 +1,130 @@ +#!/usr/bin/env python +# Copyright (c) Facebook, Inc. and its affiliates. +# All rights reserved. +# +# This source code is licensed under the license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import contextlib +import sys +from collections import Counter +from multiprocessing import Pool + +from fairseq.data.encoders.gpt2_bpe import get_encoder + + +def main(): + """ + Helper script to encode raw text with the GPT-2 BPE using multiple processes. + + The encoder.json and vocab.bpe files can be obtained here: + - https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/encoder.json + - https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/vocab.bpe + """ + parser = argparse.ArgumentParser() + parser.add_argument( + "--encoder-json", + help="path to encoder.json", + ) + parser.add_argument( + "--vocab-bpe", + type=str, + help="path to vocab.bpe", + ) + parser.add_argument( + "--inputs", + nargs="+", + default=["-"], + help="input files to filter/encode", + ) + parser.add_argument( + "--outputs", + nargs="+", + default=["-"], + help="path to save encoded outputs", + ) + parser.add_argument( + "--keep-empty", + action="store_true", + help="keep empty lines", + ) + parser.add_argument("--workers", type=int, default=20) + args = parser.parse_args() + + assert len(args.inputs) == len( + args.outputs + ), "number of input and output paths should match" + + with contextlib.ExitStack() as stack: + inputs = [ + stack.enter_context(open(input, "r", encoding="utf-8")) + if input != "-" + else sys.stdin + for input in args.inputs + ] + outputs = [ + stack.enter_context(open(output, "w", encoding="utf-8")) + if output != "-" + else sys.stdout + for output in args.outputs + ] + + encoder = MultiprocessingEncoder(args) + pool = Pool(args.workers, initializer=encoder.initializer) + encoded_lines = pool.imap(encoder.encode_lines, zip(*inputs), 100) + + stats = Counter() + for i, (filt, enc_lines) in enumerate(encoded_lines, start=1): + if filt == "PASS": + for enc_line, output_h in zip(enc_lines, outputs): + print(enc_line, file=output_h) + else: + stats["num_filtered_" + filt] += 1 + if i % 10000 == 0: + print("processed {} lines".format(i), file=sys.stderr) + + for k, v in stats.most_common(): + print("[{}] filtered {} lines".format(k, v), file=sys.stderr) + + +class MultiprocessingEncoder(object): + def __init__(self, args): + self.args = args + + def initializer(self): + global bpe + bpe = get_encoder(self.args.encoder_json, self.args.vocab_bpe) + + def encode(self, line): + global bpe + ids = bpe.encode(line) + return list(map(str, ids)) + + def decode(self, tokens): + global bpe + return bpe.decode(tokens) + + def encode_lines(self, lines): + """ + Encode a set of lines. All lines will be encoded together. + """ + enc_lines = [] + for line in lines: + line = line.strip() + if len(line) == 0 and not self.args.keep_empty: + return ["EMPTY", None] + tokens = self.encode(line) + enc_lines.append(" ".join(tokens)) + return ["PASS", enc_lines] + + def decode_lines(self, lines): + dec_lines = [] + for line in lines: + tokens = map(int, line.strip().split()) + dec_lines.append(self.decode(tokens)) + return ["PASS", dec_lines] + + +if __name__ == "__main__": + main() diff --git a/examples/roberta/preprocess_GLUE_tasks.sh b/examples/roberta/preprocess_GLUE_tasks.sh new file mode 100755 index 0000000000000000000000000000000000000000..7f215a3b53e1c4a7b1f0320102915a49d84a5015 --- /dev/null +++ b/examples/roberta/preprocess_GLUE_tasks.sh @@ -0,0 +1,185 @@ +#!/bin/bash +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +# raw glue data as downloaded by glue download script (https://gist.github.com/W4ngatang/60c2bdb54d156a41194446737ce03e2e) +if [[ $# -ne 2 ]]; then + echo "Run as following:" + echo "./examples/roberta/preprocess_GLUE_tasks.sh <glud_data_folder> <task_name>" + exit 1 +fi + +GLUE_DATA_FOLDER=$1 + +# download bpe encoder.json, vocabulary and fairseq dictionary +wget -N 'https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/encoder.json' +wget -N 'https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/vocab.bpe' +wget -N 'https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/dict.txt' + +TASKS=$2 # QQP + +if [ "$TASKS" = "ALL" ] +then + TASKS="QQP MNLI QNLI MRPC RTE STS-B SST-2 CoLA" +fi + +for TASK in $TASKS +do + echo "Preprocessing $TASK" + + TASK_DATA_FOLDER="$GLUE_DATA_FOLDER/$TASK" + echo "Raw data as downloaded from glue website: $TASK_DATA_FOLDER" + + SPLITS="train dev test" + INPUT_COUNT=2 + if [ "$TASK" = "QQP" ] + then + INPUT_COLUMNS=( 4 5 ) + TEST_INPUT_COLUMNS=( 2 3 ) + LABEL_COLUMN=6 + elif [ "$TASK" = "MNLI" ] + then + SPLITS="train dev_matched dev_mismatched test_matched test_mismatched" + INPUT_COLUMNS=( 9 10 ) + TEST_INPUT_COLUMNS=( 9 10 ) + DEV_LABEL_COLUMN=16 + LABEL_COLUMN=12 + elif [ "$TASK" = "QNLI" ] + then + INPUT_COLUMNS=( 2 3 ) + TEST_INPUT_COLUMNS=( 2 3 ) + LABEL_COLUMN=4 + elif [ "$TASK" = "MRPC" ] + then + INPUT_COLUMNS=( 4 5 ) + TEST_INPUT_COLUMNS=( 4 5 ) + LABEL_COLUMN=1 + elif [ "$TASK" = "RTE" ] + then + INPUT_COLUMNS=( 2 3 ) + TEST_INPUT_COLUMNS=( 2 3 ) + LABEL_COLUMN=4 + elif [ "$TASK" = "STS-B" ] + then + INPUT_COLUMNS=( 8 9 ) + TEST_INPUT_COLUMNS=( 8 9 ) + LABEL_COLUMN=10 + # Following are single sentence tasks. + elif [ "$TASK" = "SST-2" ] + then + INPUT_COLUMNS=( 1 ) + TEST_INPUT_COLUMNS=( 2 ) + LABEL_COLUMN=2 + INPUT_COUNT=1 + elif [ "$TASK" = "CoLA" ] + then + INPUT_COLUMNS=( 4 ) + TEST_INPUT_COLUMNS=( 2 ) + LABEL_COLUMN=2 + INPUT_COUNT=1 + fi + + # Strip out header and filter lines that don't have expected number of fields. + rm -rf "$TASK_DATA_FOLDER/processed" + mkdir -p "$TASK_DATA_FOLDER/processed" + for SPLIT in $SPLITS + do + # CoLA train and dev doesn't have header. + if [[ ( "$TASK" = "CoLA") && ( "$SPLIT" != "test" ) ]] + then + cp "$TASK_DATA_FOLDER/$SPLIT.tsv" "$TASK_DATA_FOLDER/processed/$SPLIT.tsv.temp"; + else + tail -n +2 "$TASK_DATA_FOLDER/$SPLIT.tsv" > "$TASK_DATA_FOLDER/processed/$SPLIT.tsv.temp"; + fi + + # Remove unformatted lines from train and dev files for QQP dataset. + if [[ ( "$TASK" = "QQP") && ( "$SPLIT" != "test" ) ]] + then + awk -F '\t' -v NUM_FIELDS=6 'NF==NUM_FIELDS{print}{}' "$TASK_DATA_FOLDER/processed/$SPLIT.tsv.temp" > "$TASK_DATA_FOLDER/processed/$SPLIT.tsv"; + else + cp "$TASK_DATA_FOLDER/processed/$SPLIT.tsv.temp" "$TASK_DATA_FOLDER/processed/$SPLIT.tsv"; + fi + rm "$TASK_DATA_FOLDER/processed/$SPLIT.tsv.temp"; + done + + # Split into input0, input1 and label + for SPLIT in $SPLITS + do + for INPUT_TYPE in $(seq 0 $((INPUT_COUNT-1))) + do + if [[ "$SPLIT" != test* ]] + then + COLUMN_NUMBER=${INPUT_COLUMNS[$INPUT_TYPE]} + else + COLUMN_NUMBER=${TEST_INPUT_COLUMNS[$INPUT_TYPE]} + fi + cut -f"$COLUMN_NUMBER" "$TASK_DATA_FOLDER/processed/$SPLIT.tsv" > "$TASK_DATA_FOLDER/processed/$SPLIT.raw.input$INPUT_TYPE"; + done + + if [[ "$SPLIT" != test* ]] + then + if [ "$TASK" = "MNLI" ] && [ "$SPLIT" != "train" ] + then + cut -f"$DEV_LABEL_COLUMN" "$TASK_DATA_FOLDER/processed/$SPLIT.tsv" > "$TASK_DATA_FOLDER/processed/$SPLIT.label"; + else + cut -f"$LABEL_COLUMN" "$TASK_DATA_FOLDER/processed/$SPLIT.tsv" > "$TASK_DATA_FOLDER/processed/$SPLIT.label"; + fi + fi + + # BPE encode. + for INPUT_TYPE in $(seq 0 $((INPUT_COUNT-1))) + do + LANG="input$INPUT_TYPE" + echo "BPE encoding $SPLIT/$LANG" + python -m examples.roberta.multiprocessing_bpe_encoder \ + --encoder-json encoder.json \ + --vocab-bpe vocab.bpe \ + --inputs "$TASK_DATA_FOLDER/processed/$SPLIT.raw.$LANG" \ + --outputs "$TASK_DATA_FOLDER/processed/$SPLIT.$LANG" \ + --workers 60 \ + --keep-empty; + done + done + + # Remove output directory. + rm -rf "$TASK-bin" + + DEVPREF="$TASK_DATA_FOLDER/processed/dev.LANG" + TESTPREF="$TASK_DATA_FOLDER/processed/test.LANG" + if [ "$TASK" = "MNLI" ] + then + DEVPREF="$TASK_DATA_FOLDER/processed/dev_matched.LANG,$TASK_DATA_FOLDER/processed/dev_mismatched.LANG" + TESTPREF="$TASK_DATA_FOLDER/processed/test_matched.LANG,$TASK_DATA_FOLDER/processed/test_mismatched.LANG" + fi + + # Run fairseq preprocessing: + for INPUT_TYPE in $(seq 0 $((INPUT_COUNT-1))) + do + LANG="input$INPUT_TYPE" + fairseq-preprocess \ + --only-source \ + --trainpref "$TASK_DATA_FOLDER/processed/train.$LANG" \ + --validpref "${DEVPREF//LANG/$LANG}" \ + --testpref "${TESTPREF//LANG/$LANG}" \ + --destdir "$TASK-bin/$LANG" \ + --workers 60 \ + --srcdict dict.txt; + done + if [[ "$TASK" != "STS-B" ]] + then + fairseq-preprocess \ + --only-source \ + --trainpref "$TASK_DATA_FOLDER/processed/train.label" \ + --validpref "${DEVPREF//LANG/label}" \ + --destdir "$TASK-bin/label" \ + --workers 60; + else + # For STS-B output range is converted to be between: [0.0, 1.0] + mkdir -p "$TASK-bin/label" + awk '{print $1 / 5.0 }' "$TASK_DATA_FOLDER/processed/train.label" > "$TASK-bin/label/train.label" + awk '{print $1 / 5.0 }' "$TASK_DATA_FOLDER/processed/dev.label" > "$TASK-bin/label/valid.label" + fi +done diff --git a/examples/roberta/preprocess_RACE.py b/examples/roberta/preprocess_RACE.py new file mode 100644 index 0000000000000000000000000000000000000000..cdd66072718ccb6033304c97926271909a17f9d6 --- /dev/null +++ b/examples/roberta/preprocess_RACE.py @@ -0,0 +1,102 @@ +#!/usr/bin/env python +# Copyright (c) Facebook, Inc. and its affiliates. +# All rights reserved. +# +# This source code is licensed under the license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import json +import os +import re + + +class InputExample: + def __init__(self, paragraph, qa_list, label): + self.paragraph = paragraph + self.qa_list = qa_list + self.label = label + + +def get_examples(data_dir, set_type): + """ + Extract paragraph and question-answer list from each json file + """ + examples = [] + + levels = ["middle", "high"] + set_type_c = set_type.split("-") + if len(set_type_c) == 2: + levels = [set_type_c[1]] + set_type = set_type_c[0] + for level in levels: + cur_dir = os.path.join(data_dir, set_type, level) + for filename in os.listdir(cur_dir): + cur_path = os.path.join(cur_dir, filename) + with open(cur_path, "r") as f: + cur_data = json.load(f) + answers = cur_data["answers"] + options = cur_data["options"] + questions = cur_data["questions"] + context = cur_data["article"].replace("\n", " ") + context = re.sub(r"\s+", " ", context) + for i in range(len(answers)): + label = ord(answers[i]) - ord("A") + qa_list = [] + question = questions[i] + for j in range(4): + option = options[i][j] + if "_" in question: + qa_cat = question.replace("_", option) + else: + qa_cat = " ".join([question, option]) + qa_cat = re.sub(r"\s+", " ", qa_cat) + qa_list.append(qa_cat) + examples.append(InputExample(context, qa_list, label)) + + return examples + + +def main(): + """ + Helper script to extract paragraphs questions and answers from RACE datasets. + """ + parser = argparse.ArgumentParser() + parser.add_argument( + "--input-dir", + help="input directory for downloaded RACE dataset", + ) + parser.add_argument( + "--output-dir", + help="output directory for extracted data", + ) + args = parser.parse_args() + + if not os.path.exists(args.output_dir): + os.makedirs(args.output_dir, exist_ok=True) + + for set_type in ["train", "dev", "test-middle", "test-high"]: + examples = get_examples(args.input_dir, set_type) + qa_file_paths = [ + os.path.join(args.output_dir, set_type + ".input" + str(i + 1)) + for i in range(4) + ] + qa_files = [open(qa_file_path, "w") for qa_file_path in qa_file_paths] + outf_context_path = os.path.join(args.output_dir, set_type + ".input0") + outf_label_path = os.path.join(args.output_dir, set_type + ".label") + outf_context = open(outf_context_path, "w") + outf_label = open(outf_label_path, "w") + for example in examples: + outf_context.write(example.paragraph + "\n") + for i in range(4): + qa_files[i].write(example.qa_list[i] + "\n") + outf_label.write(str(example.label) + "\n") + + for f in qa_files: + f.close() + outf_label.close() + outf_context.close() + + +if __name__ == "__main__": + main() diff --git a/examples/roberta/preprocess_RACE.sh b/examples/roberta/preprocess_RACE.sh new file mode 100755 index 0000000000000000000000000000000000000000..932d2ab6e521fecc7d0297f26a8c43857541ef3b --- /dev/null +++ b/examples/roberta/preprocess_RACE.sh @@ -0,0 +1,59 @@ +#!/bin/bash +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +# data should be downloaded and processed with reprocess_RACE.py +if [[ $# -ne 2 ]]; then + echo "Run as following:" + echo "./examples/roberta/preprocess_RACE.sh <race_data_folder> <output_folder>" + exit 1 +fi + +RACE_DATA_FOLDER=$1 +OUT_DATA_FOLDER=$2 + +# download bpe encoder.json, vocabulary and fairseq dictionary +wget -N 'https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/encoder.json' +wget -N 'https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/vocab.bpe' +wget -N 'https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/dict.txt' + +SPLITS="train dev test-middle test-high" +INPUT_TYPES="input0 input1 input2 input3 input4" +for INPUT_TYPE in $INPUT_TYPES +do + for SPLIT in $SPLITS + do + echo "BPE encoding $SPLIT/$INPUT_TYPE" + python -m examples.roberta.multiprocessing_bpe_encoder \ + --encoder-json encoder.json \ + --vocab-bpe vocab.bpe \ + --inputs "$RACE_DATA_FOLDER/$SPLIT.$INPUT_TYPE" \ + --outputs "$RACE_DATA_FOLDER/$SPLIT.$INPUT_TYPE.bpe" \ + --workers 10 \ + --keep-empty; + + done +done + +for INPUT_TYPE in $INPUT_TYPES + do + LANG="input$INPUT_TYPE" + fairseq-preprocess \ + --only-source \ + --trainpref "$RACE_DATA_FOLDER/train.$INPUT_TYPE.bpe" \ + --validpref "$RACE_DATA_FOLDER/dev.$INPUT_TYPE.bpe" \ + --testpref "$RACE_DATA_FOLDER/test-middle.$INPUT_TYPE.bpe,$RACE_DATA_FOLDER/test-high.$INPUT_TYPE.bpe" \ + --destdir "$OUT_DATA_FOLDER/$INPUT_TYPE" \ + --workers 10 \ + --srcdict dict.txt; +done + +rm -rf "$OUT_DATA_FOLDER/label" +mkdir -p "$OUT_DATA_FOLDER/label" +cp "$RACE_DATA_FOLDER/train.label" "$OUT_DATA_FOLDER/label/" +cp "$RACE_DATA_FOLDER/dev.label" "$OUT_DATA_FOLDER/label/valid.label" +cp "$RACE_DATA_FOLDER/test-middle.label" "$OUT_DATA_FOLDER/label/test.label" +cp "$RACE_DATA_FOLDER/test-high.label" "$OUT_DATA_FOLDER/label/test1.label" diff --git a/examples/roberta/wsc/README.md b/examples/roberta/wsc/README.md new file mode 100644 index 0000000000000000000000000000000000000000..21a045d999739836a17574593292e42131315ae9 --- /dev/null +++ b/examples/roberta/wsc/README.md @@ -0,0 +1,125 @@ +# Finetuning RoBERTa on Winograd Schema Challenge (WSC) data + +The following instructions can be used to finetune RoBERTa on the WSC training +data provided by [SuperGLUE](https://super.gluebenchmark.com/). + +Note that there is high variance in the results. For our GLUE/SuperGLUE +submission we swept over the learning rate (1e-5, 2e-5, 3e-5), batch size (16, +32, 64) and total number of updates (500, 1000, 2000, 3000), as well as the +random seed. Out of ~100 runs we chose the best 7 models and ensembled them. + +**Approach:** The instructions below use a slightly different loss function than +what's described in the original RoBERTa arXiv paper. In particular, +[Kocijan et al. (2019)](https://arxiv.org/abs/1905.06290) introduce a margin +ranking loss between `(query, candidate)` pairs with tunable hyperparameters +alpha and beta. This is supported in our code as well with the `--wsc-alpha` and +`--wsc-beta` arguments. However, we achieved slightly better (and more robust) +results on the development set by instead using a single cross entropy loss term +over the log-probabilities for the query and all mined candidates. **The +candidates are mined using spaCy from each input sentence in isolation, so the +approach remains strictly pointwise.** This reduces the number of +hyperparameters and our best model achieved 92.3% development set accuracy, +compared to ~90% accuracy for the margin loss. Later versions of the RoBERTa +arXiv paper will describe this updated formulation. + +### 1) Download the WSC data from the SuperGLUE website: +```bash +wget https://dl.fbaipublicfiles.com/glue/superglue/data/v2/WSC.zip +unzip WSC.zip + +# we also need to copy the RoBERTa dictionary into the same directory +wget -O WSC/dict.txt https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/dict.txt +``` + +### 2) Finetune over the provided training data: +```bash +TOTAL_NUM_UPDATES=2000 # Total number of training steps. +WARMUP_UPDATES=250 # Linearly increase LR over this many steps. +LR=2e-05 # Peak LR for polynomial LR scheduler. +MAX_SENTENCES=16 # Batch size per GPU. +SEED=1 # Random seed. +ROBERTA_PATH=/path/to/roberta/model.pt + +# we use the --user-dir option to load the task and criterion +# from the examples/roberta/wsc directory: +FAIRSEQ_PATH=/path/to/fairseq +FAIRSEQ_USER_DIR=${FAIRSEQ_PATH}/examples/roberta/wsc + +CUDA_VISIBLE_DEVICES=0,1,2,3 fairseq-train WSC/ \ + --restore-file $ROBERTA_PATH \ + --reset-optimizer --reset-dataloader --reset-meters \ + --no-epoch-checkpoints --no-last-checkpoints --no-save-optimizer-state \ + --best-checkpoint-metric accuracy --maximize-best-checkpoint-metric \ + --valid-subset val \ + --fp16 --ddp-backend legacy_ddp \ + --user-dir $FAIRSEQ_USER_DIR \ + --task wsc --criterion wsc --wsc-cross-entropy \ + --arch roberta_large --bpe gpt2 --max-positions 512 \ + --dropout 0.1 --attention-dropout 0.1 --weight-decay 0.01 \ + --optimizer adam --adam-betas '(0.9, 0.98)' --adam-eps 1e-06 \ + --lr-scheduler polynomial_decay --lr $LR \ + --warmup-updates $WARMUP_UPDATES --total-num-update $TOTAL_NUM_UPDATES \ + --batch-size $MAX_SENTENCES \ + --max-update $TOTAL_NUM_UPDATES \ + --log-format simple --log-interval 100 \ + --seed $SEED +``` + +The above command assumes training on 4 GPUs, but you can achieve the same +results on a single GPU by adding `--update-freq=4`. + +### 3) Evaluate +```python +from fairseq.models.roberta import RobertaModel +from examples.roberta.wsc import wsc_utils # also loads WSC task and criterion +roberta = RobertaModel.from_pretrained('checkpoints', 'checkpoint_best.pt', 'WSC/') +roberta.cuda() +nsamples, ncorrect = 0, 0 +for sentence, label in wsc_utils.jsonl_iterator('WSC/val.jsonl', eval=True): + pred = roberta.disambiguate_pronoun(sentence) + nsamples += 1 + if pred == label: + ncorrect += 1 +print('Accuracy: ' + str(ncorrect / float(nsamples))) +# Accuracy: 0.9230769230769231 +``` + +## RoBERTa training on WinoGrande dataset +We have also provided `winogrande` task and criterion for finetuning on the +[WinoGrande](https://mosaic.allenai.org/projects/winogrande) like datasets +where there are always two candidates and one is correct. +It's more efficient implementation for such subcases. + +```bash +TOTAL_NUM_UPDATES=23750 # Total number of training steps. +WARMUP_UPDATES=2375 # Linearly increase LR over this many steps. +LR=1e-05 # Peak LR for polynomial LR scheduler. +MAX_SENTENCES=32 # Batch size per GPU. +SEED=1 # Random seed. +ROBERTA_PATH=/path/to/roberta/model.pt + +# we use the --user-dir option to load the task and criterion +# from the examples/roberta/wsc directory: +FAIRSEQ_PATH=/path/to/fairseq +FAIRSEQ_USER_DIR=${FAIRSEQ_PATH}/examples/roberta/wsc + +cd fairseq +CUDA_VISIBLE_DEVICES=0 fairseq-train winogrande_1.0/ \ + --restore-file $ROBERTA_PATH \ + --reset-optimizer --reset-dataloader --reset-meters \ + --no-epoch-checkpoints --no-last-checkpoints --no-save-optimizer-state \ + --best-checkpoint-metric accuracy --maximize-best-checkpoint-metric \ + --valid-subset val \ + --fp16 --ddp-backend legacy_ddp \ + --user-dir $FAIRSEQ_USER_DIR \ + --task winogrande --criterion winogrande \ + --wsc-margin-alpha 5.0 --wsc-margin-beta 0.4 \ + --arch roberta_large --bpe gpt2 --max-positions 512 \ + --dropout 0.1 --attention-dropout 0.1 --weight-decay 0.01 \ + --optimizer adam --adam-betas '(0.9, 0.98)' --adam-eps 1e-06 \ + --lr-scheduler polynomial_decay --lr $LR \ + --warmup-updates $WARMUP_UPDATES --total-num-update $TOTAL_NUM_UPDATES \ + --batch-size $MAX_SENTENCES \ + --max-update $TOTAL_NUM_UPDATES \ + --log-format simple --log-interval 100 +``` diff --git a/examples/roberta/wsc/__init__.py b/examples/roberta/wsc/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..78afa4728eeed96142900118f6452730023466c9 --- /dev/null +++ b/examples/roberta/wsc/__init__.py @@ -0,0 +1,7 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from . import wsc_criterion # noqa +from . import wsc_task # noqa diff --git a/examples/roberta/wsc/wsc_criterion.py b/examples/roberta/wsc/wsc_criterion.py new file mode 100644 index 0000000000000000000000000000000000000000..ed0251fdecc3573228ad271f1090aaf914b48cd1 --- /dev/null +++ b/examples/roberta/wsc/wsc_criterion.py @@ -0,0 +1,167 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +import torch +import torch.nn.functional as F +from fairseq import utils +from fairseq.criterions import LegacyFairseqCriterion, register_criterion +from fairseq.data import encoders + + +@register_criterion("wsc") +class WSCCriterion(LegacyFairseqCriterion): + def __init__(self, args, task): + super().__init__(args, task) + if self.args.save_predictions is not None: + self.prediction_h = open(self.args.save_predictions, "w") + else: + self.prediction_h = None + self.bpe = encoders.build_bpe(args.bpe) + self.tokenizer = encoders.build_tokenizer(args.tokenizer) + + def __del__(self): + if self.prediction_h is not None: + self.prediction_h.close() + + @staticmethod + def add_args(parser): + """Add criterion-specific arguments to the parser.""" + parser.add_argument("--wsc-margin-alpha", type=float, metavar="A", default=1.0) + parser.add_argument("--wsc-margin-beta", type=float, metavar="B", default=0.0) + parser.add_argument( + "--wsc-cross-entropy", + action="store_true", + help="use cross entropy formulation instead of margin loss", + ) + parser.add_argument( + "--save-predictions", metavar="FILE", help="file to save predictions to" + ) + + def get_masked_input(self, tokens, mask): + masked_tokens = tokens.clone() + masked_tokens[mask] = self.task.mask + return masked_tokens + + def get_lprobs(self, model, tokens, mask): + logits, _ = model(src_tokens=self.get_masked_input(tokens, mask)) + lprobs = F.log_softmax(logits, dim=-1, dtype=torch.float) + scores = lprobs.gather(2, tokens.unsqueeze(-1)).squeeze(-1) + mask = mask.type_as(scores) + scores = (scores * mask).sum(dim=-1) / mask.sum(dim=-1) + return scores + + def get_loss(self, query_lprobs, cand_lprobs): + if self.args.wsc_cross_entropy: + return F.cross_entropy( + torch.cat([query_lprobs, cand_lprobs]).unsqueeze(0), + query_lprobs.new([0]).long(), + ) + else: + return ( + -query_lprobs + + self.args.wsc_margin_alpha + * (cand_lprobs - query_lprobs + self.args.wsc_margin_beta).clamp(min=0) + ).sum() + + def forward(self, model, sample, reduce=True): + # compute loss and accuracy + loss, nloss = 0.0, 0 + ncorrect, nqueries = 0, 0 + + for i, label in enumerate(sample["labels"]): + query_lprobs = self.get_lprobs( + model, + sample["query_tokens"][i].unsqueeze(0), + sample["query_masks"][i].unsqueeze(0), + ) + cand_lprobs = self.get_lprobs( + model, + sample["candidate_tokens"][i], + sample["candidate_masks"][i], + ) + + pred = (query_lprobs >= cand_lprobs).all().item() + + if label is not None: + label = 1 if label else 0 + ncorrect += 1 if pred == label else 0 + nqueries += 1 + + if label: + # only compute a loss for positive instances + nloss += 1 + loss += self.get_loss(query_lprobs, cand_lprobs) + + id = sample["id"][i].item() + if self.prediction_h is not None: + print("{}\t{}\t{}".format(id, pred, label), file=self.prediction_h) + + if nloss == 0: + loss = torch.tensor(0.0, requires_grad=True) + + sample_size = nqueries if nqueries > 0 else 1 + logging_output = { + "loss": utils.item(loss.data) if reduce else loss.data, + "ntokens": sample["ntokens"], + "nsentences": sample["nsentences"], + "sample_size": sample_size, + "ncorrect": ncorrect, + "nqueries": nqueries, + } + return loss, sample_size, logging_output + + @staticmethod + def aggregate_logging_outputs(logging_outputs): + """Aggregate logging outputs from data parallel training.""" + loss_sum = sum(log.get("loss", 0) for log in logging_outputs) + ntokens = sum(log.get("ntokens", 0) for log in logging_outputs) + nsentences = sum(log.get("nsentences", 0) for log in logging_outputs) + sample_size = sum(log.get("sample_size", 0) for log in logging_outputs) + + agg_output = { + "loss": loss_sum / sample_size / math.log(2), + "ntokens": ntokens, + "nsentences": nsentences, + "sample_size": sample_size, + } + + ncorrect = sum(log.get("ncorrect", 0) for log in logging_outputs) + nqueries = sum(log.get("nqueries", 0) for log in logging_outputs) + if nqueries > 0: + agg_output["accuracy"] = ncorrect / float(nqueries) + + return agg_output + + +@register_criterion("winogrande") +class WinograndeCriterion(WSCCriterion): + def forward(self, model, sample, reduce=True): + # compute loss and accuracy + query_lprobs = self.get_lprobs( + model, + sample["query_tokens"], + sample["query_masks"], + ) + cand_lprobs = self.get_lprobs( + model, + sample["candidate_tokens"], + sample["candidate_masks"], + ) + pred = query_lprobs >= cand_lprobs + loss = self.get_loss(query_lprobs, cand_lprobs) + + sample_size = sample["query_tokens"].size(0) + ncorrect = pred.sum().item() + logging_output = { + "loss": utils.item(loss.data) if reduce else loss.data, + "ntokens": sample["ntokens"], + "nsentences": sample["nsentences"], + "sample_size": sample_size, + "ncorrect": ncorrect, + "nqueries": sample_size, + } + return loss, sample_size, logging_output diff --git a/examples/roberta/wsc/wsc_task.py b/examples/roberta/wsc/wsc_task.py new file mode 100644 index 0000000000000000000000000000000000000000..602ea737ed75a33fddf44dd859e999ecfce2730d --- /dev/null +++ b/examples/roberta/wsc/wsc_task.py @@ -0,0 +1,401 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import json +import os +import tempfile + +import numpy as np +import torch +import torch.nn.functional as F +from fairseq import utils +from fairseq.data import ( + Dictionary, + IdDataset, + ListDataset, + NestedDictionaryDataset, + NumelDataset, + NumSamplesDataset, + PadDataset, + SortDataset, + data_utils, + encoders, +) +from fairseq.tasks import LegacyFairseqTask, register_task + +from . import wsc_utils + + +@register_task("wsc") +class WSCTask(LegacyFairseqTask): + """Task to finetune RoBERTa for Winograd Schemas.""" + + @staticmethod + def add_args(parser): + """Add task-specific arguments to the parser.""" + parser.add_argument( + "data", metavar="DIR", help="path to data directory; we load <split>.jsonl" + ) + parser.add_argument( + "--init-token", + type=int, + default=None, + help="add token at the beginning of each batch item", + ) + + def __init__(self, args, vocab): + super().__init__(args) + self.vocab = vocab + self.mask = vocab.add_symbol("<mask>") + + self.bpe = encoders.build_bpe(args) + self.tokenizer = encoders.build_tokenizer(args) + + # hack to handle GPT-2 BPE, which includes leading spaces + if args.bpe == "gpt2": + self.leading_space = True + self.trailing_space = False + else: + self.leading_space = False + self.trailing_space = True + + @classmethod + def load_dictionary(cls, filename): + """Load the dictionary from the filename + + Args: + filename (str): the filename + """ + dictionary = Dictionary.load(filename) + dictionary.add_symbol("<mask>") + return dictionary + + @classmethod + def setup_task(cls, args, **kwargs): + assert args.criterion == "wsc", "Must set --criterion=wsc" + + # load data and label dictionaries + vocab = cls.load_dictionary(os.path.join(args.data, "dict.txt")) + print("| dictionary: {} types".format(len(vocab))) + + return cls(args, vocab) + + def binarize(self, s: str, append_eos: bool = False): + if self.tokenizer is not None: + s = self.tokenizer.encode(s) + if self.bpe is not None: + s = self.bpe.encode(s) + tokens = self.vocab.encode_line( + s, + append_eos=append_eos, + add_if_not_exist=False, + ).long() + if self.args.init_token is not None: + tokens = torch.cat([tokens.new([self.args.init_token]), tokens]) + return tokens + + def binarize_with_mask(self, txt, prefix, suffix, leading_space, trailing_space): + toks = self.binarize( + prefix + leading_space + txt + trailing_space + suffix, + append_eos=True, + ) + mask = torch.zeros_like(toks, dtype=torch.bool) + mask_start = len(self.binarize(prefix)) + mask_size = len(self.binarize(leading_space + txt)) + mask[mask_start : mask_start + mask_size] = 1 + return toks, mask + + def load_dataset( + self, split, epoch=1, combine=False, data_path=None, return_only=False, **kwargs + ): + """Load a given dataset split. + + Args: + split (str): name of the split (e.g., train, valid, test) + """ + if data_path is None: + data_path = os.path.join(self.args.data, split + ".jsonl") + if not os.path.exists(data_path): + raise FileNotFoundError("Cannot find data: {}".format(data_path)) + + query_tokens = [] + query_masks = [] + query_lengths = [] + candidate_tokens = [] + candidate_masks = [] + candidate_lengths = [] + labels = [] + + for sentence, pronoun_span, query, label in wsc_utils.jsonl_iterator(data_path): + prefix = sentence[: pronoun_span.start].text + suffix = sentence[pronoun_span.end :].text_with_ws + + # spaCy spans include trailing spaces, but we need to know about + # leading spaces for the GPT-2 BPE + leading_space = ( + " " if sentence[: pronoun_span.start].text_with_ws.endswith(" ") else "" + ) + trailing_space = " " if pronoun_span.text_with_ws.endswith(" ") else "" + + # get noun phrases, excluding pronouns and anything overlapping with the query + cand_spans = wsc_utils.filter_noun_chunks( + wsc_utils.extended_noun_chunks(sentence), + exclude_pronouns=True, + exclude_query=query, + exact_match=False, + ) + + if query is not None: + query_toks, query_mask = self.binarize_with_mask( + query, prefix, suffix, leading_space, trailing_space + ) + query_len = len(query_toks) + else: + query_toks, query_mask, query_len = None, None, 0 + + query_tokens.append(query_toks) + query_masks.append(query_mask) + query_lengths.append(query_len) + + cand_toks, cand_masks = [], [] + for cand_span in cand_spans: + toks, mask = self.binarize_with_mask( + cand_span.text, + prefix, + suffix, + leading_space, + trailing_space, + ) + cand_toks.append(toks) + cand_masks.append(mask) + + # collate candidates + cand_toks = data_utils.collate_tokens(cand_toks, pad_idx=self.vocab.pad()) + cand_masks = data_utils.collate_tokens(cand_masks, pad_idx=0) + assert cand_toks.size() == cand_masks.size() + + candidate_tokens.append(cand_toks) + candidate_masks.append(cand_masks) + candidate_lengths.append(cand_toks.size(1)) + + labels.append(label) + + query_lengths = np.array(query_lengths) + query_tokens = ListDataset(query_tokens, query_lengths) + query_masks = ListDataset(query_masks, query_lengths) + + candidate_lengths = np.array(candidate_lengths) + candidate_tokens = ListDataset(candidate_tokens, candidate_lengths) + candidate_masks = ListDataset(candidate_masks, candidate_lengths) + + labels = ListDataset(labels, [1] * len(labels)) + + dataset = { + "id": IdDataset(), + "query_tokens": query_tokens, + "query_masks": query_masks, + "candidate_tokens": candidate_tokens, + "candidate_masks": candidate_masks, + "labels": labels, + "nsentences": NumSamplesDataset(), + "ntokens": NumelDataset(query_tokens, reduce=True), + } + + nested_dataset = NestedDictionaryDataset( + dataset, + sizes=[query_lengths], + ) + + with data_utils.numpy_seed(self.args.seed): + shuffle = np.random.permutation(len(query_tokens)) + dataset = SortDataset( + nested_dataset, + # shuffle + sort_order=[shuffle], + ) + + if return_only: + return dataset + + self.datasets[split] = dataset + return self.datasets[split] + + def build_dataset_for_inference(self, sample_json): + with tempfile.NamedTemporaryFile(buffering=0) as h: + h.write((json.dumps(sample_json) + "\n").encode("utf-8")) + dataset = self.load_dataset( + "disambiguate_pronoun", + data_path=h.name, + return_only=True, + ) + return dataset + + def disambiguate_pronoun(self, model, sentence, use_cuda=False): + sample_json = wsc_utils.convert_sentence_to_json(sentence) + dataset = self.build_dataset_for_inference(sample_json) + sample = dataset.collater([dataset[0]]) + if use_cuda: + sample = utils.move_to_cuda(sample) + + def get_masked_input(tokens, mask): + masked_tokens = tokens.clone() + masked_tokens[mask.bool()] = self.mask + return masked_tokens + + def get_lprobs(tokens, mask): + logits, _ = model(src_tokens=get_masked_input(tokens, mask)) + lprobs = F.log_softmax(logits, dim=-1, dtype=torch.float) + scores = lprobs.gather(2, tokens.unsqueeze(-1)).squeeze(-1) + mask = mask.type_as(scores) + scores = (scores * mask).sum(dim=-1) / mask.sum(dim=-1) + return scores + + cand_lprobs = get_lprobs( + sample["candidate_tokens"][0], + sample["candidate_masks"][0], + ) + if sample["query_tokens"][0] is not None: + query_lprobs = get_lprobs( + sample["query_tokens"][0].unsqueeze(0), + sample["query_masks"][0].unsqueeze(0), + ) + return (query_lprobs >= cand_lprobs).all().item() == 1 + else: + best_idx = cand_lprobs.argmax().item() + full_cand = sample["candidate_tokens"][0][best_idx] + mask = sample["candidate_masks"][0][best_idx] + toks = full_cand[mask.bool()] + return self.bpe.decode(self.source_dictionary.string(toks)).strip() + + @property + def source_dictionary(self): + return self.vocab + + @property + def target_dictionary(self): + return self.vocab + + +@register_task("winogrande") +class WinograndeTask(WSCTask): + """ + Task for WinoGrande dataset. Efficient implementation for Winograd schema + tasks with exactly two candidates, one of which is correct. + """ + + @classmethod + def setup_task(cls, args, **kwargs): + assert args.criterion == "winogrande", "Must set --criterion=winogrande" + + # load data and label dictionaries + vocab = cls.load_dictionary(os.path.join(args.data, "dict.txt")) + print("| dictionary: {} types".format(len(vocab))) + + return cls(args, vocab) + + def load_dataset( + self, split, epoch=1, combine=False, data_path=None, return_only=False, **kwargs + ): + """Load a given dataset split. + + Args: + split (str): name of the split (e.g., train, valid, test) + """ + if data_path is None: + data_path = os.path.join(self.args.data, split + ".jsonl") + if not os.path.exists(data_path): + raise FileNotFoundError("Cannot find data: {}".format(data_path)) + + query_tokens = [] + query_masks = [] + query_lengths = [] + candidate_tokens = [] + candidate_masks = [] + candidate_lengths = [] + + itr = wsc_utils.winogrande_jsonl_iterator(data_path, eval=(split == "test")) + + for sample in itr: + sentence, pronoun_span, query, cand_text = sample + prefix = sentence[: pronoun_span[0]].rstrip() + suffix = sentence[pronoun_span[1] :] + + leading_space = " " if sentence[: pronoun_span[0]].endswith(" ") else "" + trailing_space = "" + + if query is not None: + query_toks, query_mask = self.binarize_with_mask( + query, + prefix, + suffix, + leading_space, + trailing_space, + ) + query_len = len(query_toks) + else: + query_toks, query_mask, query_len = None, None, 0 + + query_tokens.append(query_toks) + query_masks.append(query_mask) + query_lengths.append(query_len) + + cand_toks, cand_mask = self.binarize_with_mask( + cand_text, + prefix, + suffix, + leading_space, + trailing_space, + ) + + candidate_tokens.append(cand_toks) + candidate_masks.append(cand_mask) + candidate_lengths.append(cand_toks.size(0)) + + query_lengths = np.array(query_lengths) + + def get_pad_dataset_fn(tokens, length, pad_idx): + return PadDataset( + ListDataset(tokens, length), + pad_idx=pad_idx, + left_pad=False, + ) + + query_tokens = get_pad_dataset_fn(query_tokens, query_lengths, self.vocab.pad()) + query_masks = get_pad_dataset_fn(query_masks, query_lengths, 0) + + candidate_lengths = np.array(candidate_lengths) + candidate_tokens = get_pad_dataset_fn( + candidate_tokens, candidate_lengths, self.vocab.pad() + ) + candidate_masks = get_pad_dataset_fn(candidate_masks, candidate_lengths, 0) + + dataset = { + "id": IdDataset(), + "query_tokens": query_tokens, + "query_masks": query_masks, + "candidate_tokens": candidate_tokens, + "candidate_masks": candidate_masks, + "nsentences": NumSamplesDataset(), + "ntokens": NumelDataset(query_tokens, reduce=True), + } + + nested_dataset = NestedDictionaryDataset( + dataset, + sizes=[query_lengths], + ) + + with data_utils.numpy_seed(self.args.seed): + shuffle = np.random.permutation(len(query_tokens)) + dataset = SortDataset( + nested_dataset, + # shuffle + sort_order=[shuffle], + ) + + if return_only: + return dataset + + self.datasets[split] = dataset + return self.datasets[split] diff --git a/examples/roberta/wsc/wsc_utils.py b/examples/roberta/wsc/wsc_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..da6ba74383a2490e1108609f315f44ad4b3bf002 --- /dev/null +++ b/examples/roberta/wsc/wsc_utils.py @@ -0,0 +1,241 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import json +from functools import lru_cache + + +def convert_sentence_to_json(sentence): + if "_" in sentence: + prefix, rest = sentence.split("_", 1) + query, rest = rest.split("_", 1) + query_index = len(prefix.rstrip().split(" ")) + else: + query, query_index = None, None + + prefix, rest = sentence.split("[", 1) + pronoun, rest = rest.split("]", 1) + pronoun_index = len(prefix.rstrip().split(" ")) + + sentence = sentence.replace("_", "").replace("[", "").replace("]", "") + + return { + "idx": 0, + "text": sentence, + "target": { + "span1_index": query_index, + "span1_text": query, + "span2_index": pronoun_index, + "span2_text": pronoun, + }, + } + + +def extended_noun_chunks(sentence): + noun_chunks = {(np.start, np.end) for np in sentence.noun_chunks} + np_start, cur_np = 0, "NONE" + for i, token in enumerate(sentence): + np_type = token.pos_ if token.pos_ in {"NOUN", "PROPN"} else "NONE" + if np_type != cur_np: + if cur_np != "NONE": + noun_chunks.add((np_start, i)) + if np_type != "NONE": + np_start = i + cur_np = np_type + if cur_np != "NONE": + noun_chunks.add((np_start, len(sentence))) + return [sentence[s:e] for (s, e) in sorted(noun_chunks)] + + +def find_token(sentence, start_pos): + found_tok = None + for tok in sentence: + if tok.idx == start_pos: + found_tok = tok + break + return found_tok + + +def find_span(sentence, search_text, start=0): + search_text = search_text.lower() + for tok in sentence[start:]: + remainder = sentence[tok.i :].text.lower() + if remainder.startswith(search_text): + len_to_consume = len(search_text) + start_idx = tok.idx + for next_tok in sentence[tok.i :]: + end_idx = next_tok.idx + len(next_tok.text) + if end_idx - start_idx == len_to_consume: + span = sentence[tok.i : next_tok.i + 1] + return span + return None + + +@lru_cache(maxsize=1) +def get_detokenizer(): + from sacremoses import MosesDetokenizer + + detok = MosesDetokenizer(lang="en") + return detok + + +@lru_cache(maxsize=1) +def get_spacy_nlp(): + import en_core_web_lg + + nlp = en_core_web_lg.load() + return nlp + + +def jsonl_iterator(input_fname, positive_only=False, ngram_order=3, eval=False): + detok = get_detokenizer() + nlp = get_spacy_nlp() + + with open(input_fname) as fin: + for line in fin: + sample = json.loads(line.strip()) + + if positive_only and "label" in sample and not sample["label"]: + # only consider examples where the query is correct + continue + + target = sample["target"] + + # clean up the query + query = target["span1_text"] + if query is not None: + if "\n" in query: + continue + if query.endswith(".") or query.endswith(","): + query = query[:-1] + + # split tokens + tokens = sample["text"].split(" ") + + def strip_pronoun(x): + return x.rstrip('.,"') + + # find the pronoun + pronoun_idx = target["span2_index"] + pronoun = strip_pronoun(target["span2_text"]) + if strip_pronoun(tokens[pronoun_idx]) != pronoun: + # hack: sometimes the index is misaligned + if strip_pronoun(tokens[pronoun_idx + 1]) == pronoun: + pronoun_idx += 1 + else: + raise Exception("Misaligned pronoun!") + assert strip_pronoun(tokens[pronoun_idx]) == pronoun + + # split tokens before and after the pronoun + before = tokens[:pronoun_idx] + after = tokens[pronoun_idx + 1 :] + + # the GPT BPE attaches leading spaces to tokens, so we keep track + # of whether we need spaces before or after the pronoun + leading_space = " " if pronoun_idx > 0 else "" + trailing_space = " " if len(after) > 0 else "" + + # detokenize + before = detok.detokenize(before, return_str=True) + pronoun = detok.detokenize([pronoun], return_str=True) + after = detok.detokenize(after, return_str=True) + + # hack: when the pronoun ends in a period (or comma), move the + # punctuation to the "after" part + if pronoun.endswith(".") or pronoun.endswith(","): + after = pronoun[-1] + trailing_space + after + pronoun = pronoun[:-1] + + # hack: when the "after" part begins with a comma or period, remove + # the trailing space + if after.startswith(".") or after.startswith(","): + trailing_space = "" + + # parse sentence with spacy + sentence = nlp(before + leading_space + pronoun + trailing_space + after) + + # find pronoun span + start = len(before + leading_space) + first_pronoun_tok = find_token(sentence, start_pos=start) + pronoun_span = find_span(sentence, pronoun, start=first_pronoun_tok.i) + assert pronoun_span.text == pronoun + + if eval: + # convert to format where pronoun is surrounded by "[]" and + # query is surrounded by "_" + query_span = find_span(sentence, query) + query_with_ws = "_{}_{}".format( + query_span.text, + (" " if query_span.text_with_ws.endswith(" ") else ""), + ) + pronoun_with_ws = "[{}]{}".format( + pronoun_span.text, + (" " if pronoun_span.text_with_ws.endswith(" ") else ""), + ) + if query_span.start < pronoun_span.start: + first = (query_span, query_with_ws) + second = (pronoun_span, pronoun_with_ws) + else: + first = (pronoun_span, pronoun_with_ws) + second = (query_span, query_with_ws) + sentence = ( + sentence[: first[0].start].text_with_ws + + first[1] + + sentence[first[0].end : second[0].start].text_with_ws + + second[1] + + sentence[second[0].end :].text + ) + yield sentence, sample.get("label", None) + else: + yield sentence, pronoun_span, query, sample.get("label", None) + + +def winogrande_jsonl_iterator(input_fname, eval=False): + with open(input_fname) as fin: + for line in fin: + sample = json.loads(line.strip()) + sentence, option1, option2 = ( + sample["sentence"], + sample["option1"], + sample["option2"], + ) + + pronoun_span = (sentence.index("_"), sentence.index("_") + 1) + + if eval: + query, cand = option1, option2 + else: + query = option1 if sample["answer"] == "1" else option2 + cand = option2 if sample["answer"] == "1" else option1 + yield sentence, pronoun_span, query, cand + + +def filter_noun_chunks( + chunks, exclude_pronouns=False, exclude_query=None, exact_match=False +): + if exclude_pronouns: + chunks = [ + np + for np in chunks + if (np.lemma_ != "-PRON-" and not all(tok.pos_ == "PRON" for tok in np)) + ] + + if exclude_query is not None: + excl_txt = [exclude_query.lower()] + filtered_chunks = [] + for chunk in chunks: + lower_chunk = chunk.text.lower() + found = False + for excl in excl_txt: + if ( + not exact_match and (lower_chunk in excl or excl in lower_chunk) + ) or lower_chunk == excl: + found = True + break + if not found: + filtered_chunks.append(chunk) + chunks = filtered_chunks + + return chunks diff --git a/examples/rxf/README.md b/examples/rxf/README.md new file mode 100644 index 0000000000000000000000000000000000000000..22a1cc47df23c7e0ebbf0ad805031478d1b4a95e --- /dev/null +++ b/examples/rxf/README.md @@ -0,0 +1,52 @@ +[Better Fine-Tuning by Reducing Representational Collapse](https://arxiv.org/abs/2008.03156) +===================== +This repo contains the code to replicate all experiments from the _Better Fine-Tuning by Reducing Representational Collapse_ paper excluding the probing results. + +The R3F sentence prediction criterion is registered as `sentence_prediction_r3f` while the label smoothing version of it is implemented as `label_smoothed_cross_entropy_r3f`. The R4F version of the sentence prediction criterion can be achieved by applying spectral norm to the classification head via the `--spectral-norm-classification-head` parameter. + +## Hyper-parameters +Our methods introduce 3 new hyper-parameters; `--eps` which sets the standard deviation or range of the distribution we're sampling from, `--r3f-lambda` which controls the combining of logistic loss and noisy KL loss and `--noise-type` which controls which parametric distribution we use ('normal', 'uniform'). + +For example to run R3F on RTE from GLUE + +``` +TOTAL_NUM_UPDATES=3120 +WARMUP_UPDATES=187 +LR=1e-05 +NUM_CLASSES=2 +MAX_SENTENCES=8 # Batch size. +ROBERTA_PATH=/path/to/roberta/model.pt + +CUDA_VISIBLE_DEVICES=0 fairseq-train RTE-bin \ + --restore-file $ROBERTA_PATH \ + --max-positions 512 \ + --max-sentences $MAX_SENTENCES \ + --max-tokens 4400 \ + --task sentence_prediction \ + --reset-optimizer --reset-dataloader --reset-meters \ + --required-batch-size-multiple 1 \ + --init-token 0 --separator-token 2 \ + --arch roberta_large \ + --criterion sentence_prediction_r3f \ + --num-classes $NUM_CLASSES \ + --dropout 0.1 --attention-dropout 0.1 \ + --weight-decay 0.1 --optimizer adam --adam-betas "(0.9, 0.98)" --adam-eps 1e-06 \ + --clip-norm 0.0 \ + --lr-scheduler polynomial_decay --lr $LR --total-num-update $TOTAL_NUM_UPDATES --warmup-updates $WARMUP_UPDATES \ + --fp16 --fp16-init-scale 4 --threshold-loss-scale 1 --fp16-scale-window 128 \ + --max-epoch 10 \ + --find-unused-parameters \ + --best-checkpoint-metric accuracy --maximize-best-checkpoint-metric \ + --noise-type uniform --r3f-lambda 0.7 \ + --user-dir examples/rxf/rxf_src +``` + +## Citation +```bibtex +@article{aghajanyan2020better, + title={Better Fine-Tuning by Reducing Representational Collapse}, + author={Aghajanyan, Armen and Shrivastava, Akshat and Gupta, Anchit and Goyal, Naman and Zettlemoyer, Luke and Gupta, Sonal}, + journal={arXiv preprint arXiv:2008.03156}, + year={2020} +} +``` diff --git a/examples/rxf/__init__.py b/examples/rxf/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..b24cb6b797b4159c9862bab1f882ee6ae95614ab --- /dev/null +++ b/examples/rxf/__init__.py @@ -0,0 +1,6 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from . import rxf_src # noqa diff --git a/examples/rxf/rxf_src/__init__.py b/examples/rxf/rxf_src/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..306e232d6f386b26153864601114e162080dcee4 --- /dev/null +++ b/examples/rxf/rxf_src/__init__.py @@ -0,0 +1,6 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from . import label_smoothed_cross_entropy_r3f, sentence_prediction_r3f # noqa diff --git a/examples/rxf/rxf_src/label_smoothed_cross_entropy_r3f.py b/examples/rxf/rxf_src/label_smoothed_cross_entropy_r3f.py new file mode 100644 index 0000000000000000000000000000000000000000..079db13e61c5ef46d1b1d288012145148eb0be04 --- /dev/null +++ b/examples/rxf/rxf_src/label_smoothed_cross_entropy_r3f.py @@ -0,0 +1,157 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +import torch +import torch.nn.functional as F +from fairseq import metrics, utils +from fairseq.criterions import FairseqCriterion, register_criterion +from fairseq.criterions.label_smoothed_cross_entropy import label_smoothed_nll_loss + + +@register_criterion("label_smoothed_cross_entropy_r3f") +class LabelSmoothedCrossEntropyR3FCriterion(FairseqCriterion): + def __init__( + self, task, sentence_avg, label_smoothing, eps, r3f_lambda, noise_type + ): + super().__init__(task) + self.sentence_avg = sentence_avg + self.label_smoothing = label_smoothing + self.eps = eps + self.r3f_lambda = r3f_lambda + self.noise_type = noise_type + if self.noise_type in {"normal"}: + self.noise_sampler = torch.distributions.normal.Normal( + loc=0.0, scale=self.eps + ) + elif self.noise_type == "uniform": + self.noise_sampler = torch.distributions.uniform.Uniform( + low=-self.eps, high=self.eps + ) + else: + raise Exception(f"unrecognized noise type {self.noise_type}") + + @staticmethod + def add_args(parser): + """Add criterion-specific arguments to the parser.""" + # fmt: off + parser.add_argument('--label-smoothing', default=0., type=float, metavar='D', + help='epsilon for label smoothing, 0 means no label smoothing') + parser.add_argument('--eps', type=float, default=1e-5, + help='noise eps') + parser.add_argument('--r3f-lambda', type=float, default=1.0, + help='lambda for combining logistic loss and noisy KL loss') + parser.add_argument('--noise-type', type=str, default='normal', + choices=['normal', 'uniform'], + help='type of noises') + # fmt: on + + def _get_symm_kl(self, noised_logits, input_logits): + return ( + F.kl_div( + F.log_softmax(noised_logits, dim=-1, dtype=torch.float32), + F.softmax(input_logits, dim=-1, dtype=torch.float32), + None, + None, + "sum", + ) + + F.kl_div( + F.log_softmax(input_logits, dim=-1, dtype=torch.float32), + F.softmax(noised_logits, dim=-1, dtype=torch.float32), + None, + None, + "sum", + ) + ) / noised_logits.size(0) + + def forward(self, model, sample, reduce=True): + """Compute the loss for the given sample. + + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + token_embeddings = model.encoder.embed_tokens(sample["net_input"]["src_tokens"]) + input_logits, extra = model(**sample["net_input"]) + loss, nll_loss = self.compute_loss( + model, (input_logits, extra), sample, reduce=reduce + ) + sample_size = ( + sample["target"].size(0) if self.sentence_avg else sample["ntokens"] + ) + + if model.training: + noise = self.noise_sampler.sample(sample_shape=token_embeddings.shape).to( + token_embeddings + ) + noised_embeddings = token_embeddings.clone() + noise + + noised_logits, _ = model( + **sample["net_input"], token_embeddings=noised_embeddings + ) + symm_kl = self._get_symm_kl(noised_logits, input_logits) + + if model.training: + symm_kl = symm_kl * sample_size + loss = loss + self.r3f_lambda * symm_kl + + logging_output = { + "loss": loss.data, + "nll_loss": nll_loss.data, + "ntokens": sample["ntokens"], + "nsentences": sample["target"].size(0), + "sample_size": sample_size, + } + + if model.training: + logging_output.update( + symm_kl=utils.item(symm_kl.data) if reduce else symm_kl.data + ) + + return loss, sample_size, logging_output + + def compute_loss(self, model, net_output, sample, reduce=True): + lprobs = model.get_normalized_probs(net_output, log_probs=True) + lprobs = lprobs.view(-1, lprobs.size(-1)) + target = model.get_targets(sample, net_output).view(-1, 1) + loss, nll_loss = label_smoothed_nll_loss( + lprobs, + target, + self.label_smoothing, + ignore_index=self.padding_idx, + reduce=reduce, + ) + return loss, nll_loss + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + """Aggregate logging outputs from data parallel training.""" + loss_sum = sum(log.get("loss", 0) for log in logging_outputs) + nll_loss_sum = sum(log.get("nll_loss", 0) for log in logging_outputs) + ntokens = sum(log.get("ntokens", 0) for log in logging_outputs) + sample_size = sum(log.get("sample_size", 0) for log in logging_outputs) + symm_kl_sum = sum(log.get("symm_kl", 0) for log in logging_outputs) + + metrics.log_scalar("symm_kl", symm_kl_sum / sample_size, sample_size, round=3) + metrics.log_scalar( + "loss", loss_sum / sample_size / math.log(2), sample_size, round=3 + ) + metrics.log_scalar( + "nll_loss", nll_loss_sum / ntokens / math.log(2), ntokens, round=3 + ) + metrics.log_derived( + "ppl", lambda meters: utils.get_perplexity(meters["nll_loss"].avg) + ) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return True diff --git a/examples/rxf/rxf_src/sentence_prediction_r3f.py b/examples/rxf/rxf_src/sentence_prediction_r3f.py new file mode 100644 index 0000000000000000000000000000000000000000..62dd63390c24445e2610b9b0609edbd36045ce8a --- /dev/null +++ b/examples/rxf/rxf_src/sentence_prediction_r3f.py @@ -0,0 +1,170 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +import torch +import torch.nn.functional as F +from fairseq import utils +from fairseq.criterions import FairseqCriterion, register_criterion + + +@register_criterion("sentence_prediction_r3f") +class SentencePredictionR3F(FairseqCriterion): + def __init__( + self, + task, + eps, + r3f_lambda, + noise_type, + classification_head_name, + regression_target, + ): + super().__init__(task) + self.eps = eps + self.r3f_lambda = r3f_lambda + self.noise_type = noise_type + self.classification_head_name = classification_head_name + self.regression_target = regression_target + if self.noise_type in {"normal"}: + self.noise_sampler = torch.distributions.normal.Normal( + loc=0.0, scale=self.eps + ) + elif self.noise_type == "uniform": + self.noise_sampler = torch.distributions.uniform.Uniform( + low=-self.eps, high=self.eps + ) + else: + raise Exception(f"unrecognized noise type {self.noise_type}") + + @staticmethod + def add_args(parser): + # fmt: off + parser.add_argument('--eps', type=float, default=1e-5, + help='noise eps') + parser.add_argument('--r3f-lambda', type=float, default=1.0, + help='lambda for combining logistic loss and noisy KL loss') + parser.add_argument('--noise-type', type=str, default='uniform', + choices=['normal', 'uniform'], + help='type of noises for RXF methods') + parser.add_argument('--classification-head-name', + default='sentence_classification_head', + help='name of the classification head to use') + # fmt: on + + def _get_symm_kl(self, noised_logits, input_logits): + return ( + F.kl_div( + F.log_softmax(noised_logits, dim=-1, dtype=torch.float32), + F.softmax(input_logits, dim=-1, dtype=torch.float32), + None, + None, + "sum", + ) + + F.kl_div( + F.log_softmax(input_logits, dim=-1, dtype=torch.float32), + F.softmax(noised_logits, dim=-1, dtype=torch.float32), + None, + None, + "sum", + ) + ) / noised_logits.size(0) + + def forward(self, model, sample, reduce=True): + """Compute the loss for the given sample. + + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + assert ( + hasattr(model, "classification_heads") + and self.classification_head_name in model.classification_heads + ), "model must provide sentence classification head for --criterion=sentence_prediction" + + token_embeddings = model.encoder.sentence_encoder.embed_tokens( + sample["net_input"]["src_tokens"] + ) + input_logits, _ = model( + **sample["net_input"], + features_only=True, + classification_head_name=self.classification_head_name, + token_embeddings=token_embeddings, + ) + if model.training and self.noise_sampler: + noise = self.noise_sampler.sample(sample_shape=token_embeddings.shape).to( + token_embeddings + ) + noised_embeddings = token_embeddings.detach().clone() + noise + + noised_logits, _ = model( + **sample["net_input"], + features_only=True, + classification_head_name=self.classification_head_name, + token_embeddings=noised_embeddings, + ) + symm_kl = self._get_symm_kl(noised_logits, input_logits) + else: + symm_kl = 0 + + targets = model.get_targets(sample, [input_logits]).view(-1) + sample_size = targets.numel() + + if not self.regression_target: + loss = F.nll_loss( + F.log_softmax(input_logits, dim=-1, dtype=torch.float32), + targets, + reduction="sum", + ) + if model.training: + symm_kl = symm_kl * sample_size + loss = loss + self.r3f_lambda * symm_kl + else: + logits = input_logits.squeeze().float() + targets = targets.float() + loss = F.mse_loss(logits, targets, reduction="sum") + + logging_output = { + "loss": utils.item(loss.data) if reduce else loss.data, + "ntokens": sample["ntokens"], + "nsentences": sample_size, + "sample_size": sample_size, + } + + if not self.regression_target: + preds = input_logits.max(dim=1)[1] + logging_output.update(ncorrect=(preds == targets).sum().item()) + + if model.training and self.noise_sampler: + logging_output.update( + symm_kl=utils.item(symm_kl.data) if reduce else symm_kl.data + ) + return loss, sample_size, logging_output + + @staticmethod + def aggregate_logging_outputs(logging_outputs): + """Aggregate logging outputs from data parallel training.""" + loss_sum = sum(log.get("loss", 0) for log in logging_outputs) + symm_kl_sum = sum(log.get("symm_kl", 0) for log in logging_outputs) + ntokens = sum(log.get("ntokens", 0) for log in logging_outputs) + nsentences = sum(log.get("nsentences", 0) for log in logging_outputs) + sample_size = sum(log.get("sample_size", 0) for log in logging_outputs) + + agg_output = { + "loss": loss_sum / sample_size / math.log(2), + "symm_kl": symm_kl_sum / sample_size, + "ntokens": ntokens, + "nsentences": nsentences, + "sample_size": sample_size, + } + + if len(logging_outputs) > 0 and "ncorrect" in logging_outputs[0]: + ncorrect = sum(log.get("ncorrect", 0) for log in logging_outputs) + agg_output.update(accuracy=ncorrect / nsentences) + + if sample_size != ntokens: + agg_output["nll_loss"] = loss_sum / ntokens / math.log(2) + return agg_output diff --git a/examples/scaling_nmt/README.md b/examples/scaling_nmt/README.md new file mode 100644 index 0000000000000000000000000000000000000000..0cc3360c3bbd58fe35a51591db8f081fc8576877 --- /dev/null +++ b/examples/scaling_nmt/README.md @@ -0,0 +1,114 @@ +# Scaling Neural Machine Translation (Ott et al., 2018) + +This page includes instructions for reproducing results from the paper [Scaling Neural Machine Translation (Ott et al., 2018)](https://arxiv.org/abs/1806.00187). + +## Pre-trained models + +Model | Description | Dataset | Download +---|---|---|--- +`transformer.wmt14.en-fr` | Transformer <br> ([Ott et al., 2018](https://arxiv.org/abs/1806.00187)) | [WMT14 English-French](http://statmt.org/wmt14/translation-task.html#Download) | model: <br> [download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/models/wmt14.en-fr.joined-dict.transformer.tar.bz2) <br> newstest2014: <br> [download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/data/wmt14.en-fr.joined-dict.newstest2014.tar.bz2) +`transformer.wmt16.en-de` | Transformer <br> ([Ott et al., 2018](https://arxiv.org/abs/1806.00187)) | [WMT16 English-German](https://drive.google.com/uc?export=download&id=0B_bZck-ksdkpM25jRUN2X2UxMm8) | model: <br> [download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/models/wmt16.en-de.joined-dict.transformer.tar.bz2) <br> newstest2014: <br> [download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/data/wmt16.en-de.joined-dict.newstest2014.tar.bz2) + +## Training a new model on WMT'16 En-De + +First download the [preprocessed WMT'16 En-De data provided by Google](https://drive.google.com/uc?export=download&id=0B_bZck-ksdkpM25jRUN2X2UxMm8). + +Then: + +##### 1. Extract the WMT'16 En-De data +```bash +TEXT=wmt16_en_de_bpe32k +mkdir -p $TEXT +tar -xzvf wmt16_en_de.tar.gz -C $TEXT +``` + +##### 2. Preprocess the dataset with a joined dictionary +```bash +fairseq-preprocess \ + --source-lang en --target-lang de \ + --trainpref $TEXT/train.tok.clean.bpe.32000 \ + --validpref $TEXT/newstest2013.tok.bpe.32000 \ + --testpref $TEXT/newstest2014.tok.bpe.32000 \ + --destdir data-bin/wmt16_en_de_bpe32k \ + --nwordssrc 32768 --nwordstgt 32768 \ + --joined-dictionary \ + --workers 20 +``` + +##### 3. Train a model +```bash +fairseq-train \ + data-bin/wmt16_en_de_bpe32k \ + --arch transformer_vaswani_wmt_en_de_big --share-all-embeddings \ + --optimizer adam --adam-betas '(0.9, 0.98)' --clip-norm 0.0 \ + --lr 0.0005 --lr-scheduler inverse_sqrt --warmup-updates 4000 --warmup-init-lr 1e-07 \ + --dropout 0.3 --weight-decay 0.0 \ + --criterion label_smoothed_cross_entropy --label-smoothing 0.1 \ + --max-tokens 3584 \ + --fp16 +``` + +Note that the `--fp16` flag requires you have CUDA 9.1 or greater and a Volta GPU or newer. + +***IMPORTANT:*** You will get better performance by training with big batches and +increasing the learning rate. If you want to train the above model with big batches +(assuming your machine has 8 GPUs): +- add `--update-freq 16` to simulate training on 8x16=128 GPUs +- increase the learning rate; 0.001 works well for big batches + +##### 4. Evaluate + +Now we can evaluate our trained model. + +Note that the original [Attention Is All You Need](https://arxiv.org/abs/1706.03762) +paper used a couple tricks to achieve better BLEU scores. We use these same tricks in +the Scaling NMT paper, so it's important to apply them when reproducing our results. + +First, use the [average_checkpoints.py](/scripts/average_checkpoints.py) script to +average the last few checkpoints. Averaging the last 5-10 checkpoints is usually +good, but you may need to adjust this depending on how long you've trained: +```bash +python scripts/average_checkpoints \ + --inputs /path/to/checkpoints \ + --num-epoch-checkpoints 10 \ + --output checkpoint.avg10.pt +``` + +Next, generate translations using a beam width of 4 and length penalty of 0.6: +```bash +fairseq-generate \ + data-bin/wmt16_en_de_bpe32k \ + --path checkpoint.avg10.pt \ + --beam 4 --lenpen 0.6 --remove-bpe > gen.out +``` + +Finally, we apply the ["compound splitting" script](/scripts/compound_split_bleu.sh) to +add spaces around dashes. For example "Café-Liebhaber" would become three tokens: +"Café - Liebhaber". This typically results in larger BLEU scores, but it is not +appropriate to compare these inflated scores to work which does not include this trick. +This trick was used in the [original AIAYN code](https://github.com/tensorflow/tensor2tensor/blob/fc9335c0203685cbbfe2b30c92db4352d8f60779/tensor2tensor/utils/get_ende_bleu.sh), +so we used it in the Scaling NMT paper as well. That said, it's strongly advised to +report [sacrebleu](https://github.com/mjpost/sacrebleu) scores instead. + +To compute "compound split" tokenized BLEU (not recommended!): +```bash +bash scripts/compound_split_bleu.sh gen.out +# BLEU4 = 29.29, 60.3/35.0/22.8/15.3 (BP=1.000, ratio=1.004, syslen=64763, reflen=64496) +``` + +To compute detokenized BLEU with sacrebleu (preferred): +```bash +bash scripts/sacrebleu.sh wmt14/full en de gen.out +# BLEU+case.mixed+lang.en-de+numrefs.1+smooth.exp+test.wmt14/full+tok.13a+version.1.4.3 = 28.6 59.3/34.3/22.1/14.9 (BP = 1.000 ratio = 1.016 hyp_len = 63666 ref_len = 62688) +``` + +## Citation + +```bibtex +@inproceedings{ott2018scaling, + title = {Scaling Neural Machine Translation}, + author = {Ott, Myle and Edunov, Sergey and Grangier, David and Auli, Michael}, + booktitle = {Proceedings of the Third Conference on Machine Translation (WMT)}, + year = 2018, +} +``` diff --git a/examples/simultaneous_translation/README.md b/examples/simultaneous_translation/README.md new file mode 100644 index 0000000000000000000000000000000000000000..62a005e0ec6f15af9015d335e34b45df6ed89b6c --- /dev/null +++ b/examples/simultaneous_translation/README.md @@ -0,0 +1,5 @@ +# Simultaneous Translation +Examples of simultaneous translation in fairseq +- [English-to-Japanese text-to-text wait-k model](docs/enja-waitk.md) +- [English-to-Germen text-to-text monotonic multihead attention model](docs/ende-mma.md) +- [English-to-Germen speech-to-text simultaneous translation model](../speech_to_text/docs/simulst_mustc_example.md) diff --git a/examples/simultaneous_translation/__init__.py b/examples/simultaneous_translation/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..5835316ba9b23c0d99d1a8f109ee047682211546 --- /dev/null +++ b/examples/simultaneous_translation/__init__.py @@ -0,0 +1,6 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from . import models # noqa diff --git a/examples/simultaneous_translation/docs/ende-mma.md b/examples/simultaneous_translation/docs/ende-mma.md new file mode 100644 index 0000000000000000000000000000000000000000..241d604a3b31a37755da68aad6ff47d46891d3fc --- /dev/null +++ b/examples/simultaneous_translation/docs/ende-mma.md @@ -0,0 +1,74 @@ +# Simultaneous Machine Translation + +This directory contains the code for the paper [Monotonic Multihead Attention](https://openreview.net/forum?id=Hyg96gBKPS) + +## Prepare Data + +[Please follow the instructions to download and preprocess the WMT'15 En-De dataset.](https://github.com/pytorch/fairseq/tree/simulastsharedtask/examples/translation#prepare-wmt14en2desh) + +Another example of training an English to Japanese model can be found [here](docs/enja.md) + +## Training + +- MMA-IL + +```shell +fairseq-train \ + data-bin/wmt15_en_de_32k \ + --simul-type infinite_lookback \ + --user-dir $FAIRSEQ/example/simultaneous_translation \ + --mass-preservation \ + --criterion latency_augmented_label_smoothed_cross_entropy \ + --latency-weight-avg 0.1 \ + --max-update 50000 \ + --arch transformer_monotonic_iwslt_de_en save_dir_key=lambda \ + --optimizer adam --adam-betas '(0.9, 0.98)' \ + --lr-scheduler 'inverse_sqrt' \ + --warmup-init-lr 1e-7 --warmup-updates 4000 \ + --lr 5e-4 --stop-min-lr 1e-9 --clip-norm 0.0 --weight-decay 0.0001\ + --dropout 0.3 \ + --label-smoothing 0.1\ + --max-tokens 3584 +``` + +- MMA-H + +```shell +fairseq-train \ + data-bin/wmt15_en_de_32k \ + --simul-type hard_aligned \ + --user-dir $FAIRSEQ/example/simultaneous_translation \ + --mass-preservation \ + --criterion latency_augmented_label_smoothed_cross_entropy \ + --latency-weight-var 0.1 \ + --max-update 50000 \ + --arch transformer_monotonic_iwslt_de_en save_dir_key=lambda \ + --optimizer adam --adam-betas '(0.9, 0.98)' \ + --lr-scheduler 'inverse_sqrt' \ + --warmup-init-lr 1e-7 --warmup-updates 4000 \ + --lr 5e-4 --stop-min-lr 1e-9 --clip-norm 0.0 --weight-decay 0.0001\ + --dropout 0.3 \ + --label-smoothing 0.1\ + --max-tokens 3584 +``` + +- wait-k + +```shell +fairseq-train \ + data-bin/wmt15_en_de_32k \ + --simul-type wait-k \ + --waitk-lagging 3 \ + --user-dir $FAIRSEQ/example/simultaneous_translation \ + --mass-preservation \ + --criterion latency_augmented_label_smoothed_cross_entropy \ + --max-update 50000 \ + --arch transformer_monotonic_iwslt_de_en save_dir_key=lambda \ + --optimizer adam --adam-betas '(0.9, 0.98)' \ + --lr-scheduler 'inverse_sqrt' \ + --warmup-init-lr 1e-7 --warmup-updates 4000 \ + --lr 5e-4 --stop-min-lr 1e-9 --clip-norm 0.0 --weight-decay 0.0001\ + --dropout 0.3 \ + --label-smoothing 0.1\ + --max-tokens 3584 +``` diff --git a/examples/simultaneous_translation/docs/enja-waitk.md b/examples/simultaneous_translation/docs/enja-waitk.md new file mode 100644 index 0000000000000000000000000000000000000000..fb9d82576f80b4405564a99774fc98ac2fe6ad3b --- /dev/null +++ b/examples/simultaneous_translation/docs/enja-waitk.md @@ -0,0 +1,106 @@ +# An example of English to Japaneses Simultaneous Translation System + +This is an example of training and evaluating a transformer *wait-k* English to Japanese simultaneous text-to-text translation model. + +## Data Preparation +This section introduces the data preparation for training and evaluation. +If you only want to evaluate the model, please jump to [Inference & Evaluation](#inference-&-evaluation) + +For illustration, we only use the following subsets of the available data from [WMT20 news translation task](http://www.statmt.org/wmt20/translation-task.html), which results in 7,815,391 sentence pairs. +- News Commentary v16 +- Wiki Titles v3 +- WikiMatrix V1 +- Japanese-English Subtitle Corpus +- The Kyoto Free Translation Task Corpus + +We use WMT20 development data as development set. Training `transformer_vaswani_wmt_en_de_big` model on such amount of data will result in 17.3 BLEU with greedy search and 19.7 with beam (10) search. Notice that a better performance can be achieved with the full WMT training data. + +We use [sentencepiece](https://github.com/google/sentencepiece) toolkit to tokenize the data with a vocabulary size of 32000. +Additionally, we filtered out the sentences longer than 200 words after tokenization. +Assuming the tokenized text data is saved at `${DATA_DIR}`, +we prepare the data binary with the following command. + +```bash +fairseq-preprocess \ + --source-lang en --target-lang ja \ + --trainpref ${DATA_DIR}/train \ + --validpref ${DATA_DIR}/dev \ + --testpref ${DATA_DIR}/test \ + --destdir ${WMT20_ENJA_DATA_BIN} \ + --nwordstgt 32000 --nwordssrc 32000 \ + --workers 20 +``` + +## Simultaneous Translation Model Training +To train a wait-k `(k=10)` model. +```bash +fairseq-train ${WMT20_ENJA_DATA_BIN} \ + --save-dir ${SAVEDIR} + --simul-type waitk \ + --waitk-lagging 10 \ + --max-epoch 70 \ + --arch transformer_monotonic_vaswani_wmt_en_de_big \ + --optimizer adam \ + --adam-betas '(0.9, 0.98)' \ + --lr-scheduler inverse_sqrt \ + --warmup-init-lr 1e-07 \ + --warmup-updates 4000 \ + --lr 0.0005 \ + --stop-min-lr 1e-09 \ + --clip-norm 10.0 \ + --dropout 0.3 \ + --weight-decay 0.0 \ + --criterion label_smoothed_cross_entropy \ + --label-smoothing 0.1 \ + --max-tokens 3584 +``` +This command is for training on 8 GPUs. Equivalently, the model can be trained on one GPU with `--update-freq 8`. + +## Inference & Evaluation +First of all, install [SimulEval](https://github.com/facebookresearch/SimulEval) for evaluation. + +```bash +git clone https://github.com/facebookresearch/SimulEval.git +cd SimulEval +pip install -e . +``` + +The following command is for the evaluation. +Assuming the source and reference files are `${SRC_FILE}` and `${REF_FILE}`, the sentencepiece model file for English is saved at `${SRC_SPM_PATH}` + + +```bash +simuleval \ + --source ${SRC_FILE} \ + --target ${TGT_FILE} \ + --data-bin ${WMT20_ENJA_DATA_BIN} \ + --sacrebleu-tokenizer ja-mecab \ + --eval-latency-unit char \ + --no-space \ + --src-splitter-type sentencepiecemodel \ + --src-splitter-path ${SRC_SPM_PATH} \ + --agent ${FAIRSEQ}/examples/simultaneous_translation/agents/simul_trans_text_agent_enja.py \ + --model-path ${SAVE_DIR}/${CHECKPOINT_FILENAME} \ + --output ${OUTPUT} \ + --scores +``` + +The `--data-bin` should be the same in previous sections if you prepare the data from the scratch. +If only for evaluation, a prepared data directory can be found [here](https://dl.fbaipublicfiles.com/simultaneous_translation/wmt20_enja_medium_databin.tgz) and a pretrained checkpoint (wait-k=10 model) can be downloaded from [here](https://dl.fbaipublicfiles.com/simultaneous_translation/wmt20_enja_medium_wait10_ckpt.pt). + +The output should look like this: +```bash +{ + "Quality": { + "BLEU": 11.442253287568398 + }, + "Latency": { + "AL": 8.6587861866951, + "AP": 0.7863304776251316, + "DAL": 9.477850951194764 + } +} +``` +The latency is evaluated by characters (`--eval-latency-unit`) on the target side. The latency is evaluated with `sacrebleu` with `MeCab` tokenizer `--sacrebleu-tokenizer ja-mecab`. `--no-space` indicates that do not add space when merging the predicted words. + +If `--output ${OUTPUT}` option is used, the detailed log and scores will be stored under the `${OUTPUT}` directory. diff --git a/examples/simultaneous_translation/eval/agents/simul_t2t_enja.py b/examples/simultaneous_translation/eval/agents/simul_t2t_enja.py new file mode 100644 index 0000000000000000000000000000000000000000..8f3c8703ca37398b9d389ce5181bdfac2333cdf2 --- /dev/null +++ b/examples/simultaneous_translation/eval/agents/simul_t2t_enja.py @@ -0,0 +1,226 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import os + +from fairseq import checkpoint_utils, tasks +import sentencepiece as spm +import torch + +try: + from simuleval import READ_ACTION, WRITE_ACTION, DEFAULT_EOS + from simuleval.agents import TextAgent +except ImportError: + print("Please install simuleval 'pip install simuleval'") + + +BOS_PREFIX = "\u2581" + + +class SimulTransTextAgentJA(TextAgent): + """ + Simultaneous Translation + Text agent for Japanese + """ + def __init__(self, args): + + # Whether use gpu + self.gpu = getattr(args, "gpu", False) + + # Max len + self.max_len = args.max_len + + # Load Model + self.load_model_vocab(args) + + # build word splitter + self.build_word_splitter(args) + + self.eos = DEFAULT_EOS + + def initialize_states(self, states): + states.incremental_states = dict() + states.incremental_states["online"] = dict() + + def to_device(self, tensor): + if self.gpu: + return tensor.cuda() + else: + return tensor.cpu() + + def load_model_vocab(self, args): + + filename = args.model_path + if not os.path.exists(filename): + raise IOError("Model file not found: {}".format(filename)) + + state = checkpoint_utils.load_checkpoint_to_cpu(filename) + + task_args = state["cfg"]["task"] + task_args.data = args.data_bin + + task = tasks.setup_task(task_args) + + # build model for ensemble + state["cfg"]["model"].load_pretrained_encoder_from = None + state["cfg"]["model"].load_pretrained_decoder_from = None + + self.model = task.build_model(state["cfg"]["model"]) + self.model.load_state_dict(state["model"], strict=True) + self.model.eval() + self.model.share_memory() + + if self.gpu: + self.model.cuda() + + # Set dictionary + self.dict = {} + self.dict["tgt"] = task.target_dictionary + self.dict["src"] = task.source_dictionary + + @staticmethod + def add_args(parser): + # fmt: off + parser.add_argument('--model-path', type=str, required=True, + help='path to your pretrained model.') + parser.add_argument("--data-bin", type=str, required=True, + help="Path of data binary") + parser.add_argument("--max-len", type=int, default=100, + help="Max length of translation") + parser.add_argument("--tgt-splitter-type", type=str, default="SentencePiece", + help="Subword splitter type for target text.") + parser.add_argument("--tgt-splitter-path", type=str, default=None, + help="Subword splitter model path for target text.") + parser.add_argument("--src-splitter-type", type=str, default="SentencePiece", + help="Subword splitter type for source text.") + parser.add_argument("--src-splitter-path", type=str, default=None, + help="Subword splitter model path for source text.") + # fmt: on + return parser + + def build_word_splitter(self, args): + self.spm = {} + for lang in ['src', 'tgt']: + if getattr(args, f'{lang}_splitter_type', None): + path = getattr(args, f'{lang}_splitter_path', None) + if path: + self.spm[lang] = spm.SentencePieceProcessor() + self.spm[lang].Load(path) + + def segment_to_units(self, segment, states): + # Split a full word (segment) into subwords (units) + return self.spm['src'].EncodeAsPieces(segment) + + def update_model_encoder(self, states): + if len(states.units.source) == 0: + return + + src_indices = [ + self.dict['src'].index(x) + for x in states.units.source.value + ] + + if states.finish_read(): + # Append the eos index when the prediction is over + src_indices += [self.dict["tgt"].eos_index] + + src_indices = self.to_device( + torch.LongTensor(src_indices).unsqueeze(0) + ) + src_lengths = self.to_device( + torch.LongTensor([src_indices.size(1)]) + ) + + states.encoder_states = self.model.encoder(src_indices, src_lengths) + + torch.cuda.empty_cache() + + def update_states_read(self, states): + # Happens after a read action. + self.update_model_encoder(states) + + def units_to_segment(self, units, states): + # Merge sub words (units) to full word (segment). + # For Japanese, we can directly send + # the untokenized token to server except the BOS token + # with following option + # --sacrebleu-tokenizer MeCab + # --eval-latency-unit char + # --no-space + token = units.value.pop() + + if ( + token == self.dict["tgt"].eos_word + or len(states.segments.target) > self.max_len + ): + return DEFAULT_EOS + + if BOS_PREFIX == token: + return None + if token[0] == BOS_PREFIX: + return token[1:] + else: + return token + + def policy(self, states): + + if not getattr(states, "encoder_states", None): + # No encoder states, read a token first + return READ_ACTION + + # encode previous predicted target tokens + tgt_indices = self.to_device( + torch.LongTensor( + [self.model.decoder.dictionary.eos()] + + [ + self.dict['tgt'].index(x) + for x in states.units.target.value + if x is not None + ] + ).unsqueeze(0) + ) + + # Current steps + states.incremental_states["steps"] = { + "src": states.encoder_states["encoder_out"][0].size(0), + "tgt": 1 + len(states.units.target), + } + + # Online only means the reading is not finished + states.incremental_states["online"]["only"] = ( + torch.BoolTensor([not states.finish_read()]) + ) + + x, outputs = self.model.decoder.forward( + prev_output_tokens=tgt_indices, + encoder_out=states.encoder_states, + incremental_state=states.incremental_states, + ) + + states.decoder_out = x + + torch.cuda.empty_cache() + + if outputs.action == 0: + return READ_ACTION + else: + return WRITE_ACTION + + def predict(self, states): + # Predict target token from decoder states + decoder_states = states.decoder_out + + lprobs = self.model.get_normalized_probs( + [decoder_states[:, -1:]], log_probs=True + ) + + index = lprobs.argmax(dim=-1)[0, 0].item() + + if index != self.dict['tgt'].eos_index: + token = self.dict['tgt'].string([index]) + else: + token = self.dict['tgt'].eos_word + + return token diff --git a/examples/simultaneous_translation/models/__init__.py b/examples/simultaneous_translation/models/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..257a96593ff7af93c206c066d8db4ad795b2ae36 --- /dev/null +++ b/examples/simultaneous_translation/models/__init__.py @@ -0,0 +1,15 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import importlib +import os + + +for file in sorted(os.listdir(os.path.dirname(__file__))): + if file.endswith(".py") and not file.startswith("_"): + model_name = file[: file.find(".py")] + importlib.import_module( + "examples.simultaneous_translation.models." + model_name + ) diff --git a/examples/simultaneous_translation/models/convtransformer_simul_trans.py b/examples/simultaneous_translation/models/convtransformer_simul_trans.py new file mode 100644 index 0000000000000000000000000000000000000000..4a26422f650cf13ee7d4e8d2228b50ec49876fb8 --- /dev/null +++ b/examples/simultaneous_translation/models/convtransformer_simul_trans.py @@ -0,0 +1,204 @@ +# Copyright (c) 2017-present, Facebook, Inc. +# All rights reserved. +# +# This source code is licensed under the license found in the LICENSE file in +# the root directory of this source tree. An additional grant of patent rights +# can be found in the PATENTS file in the same directory. + +from fairseq import checkpoint_utils +from fairseq.models import ( + register_model, + register_model_architecture, +) +from fairseq.models.speech_to_text import ( + ConvTransformerModel, + convtransformer_espnet, + ConvTransformerEncoder, +) +from fairseq.models.speech_to_text.modules.augmented_memory_attention import ( + augmented_memory, + SequenceEncoder, + AugmentedMemoryConvTransformerEncoder, +) + +from torch import nn, Tensor +from typing import Dict, List +from fairseq.models.speech_to_text.modules.emformer import NoSegAugmentedMemoryTransformerEncoderLayer + +@register_model("convtransformer_simul_trans") +class SimulConvTransformerModel(ConvTransformerModel): + """ + Implementation of the paper: + + SimulMT to SimulST: Adapting Simultaneous Text Translation to + End-to-End Simultaneous Speech Translation + + https://www.aclweb.org/anthology/2020.aacl-main.58.pdf + """ + + @staticmethod + def add_args(parser): + super(SimulConvTransformerModel, SimulConvTransformerModel).add_args(parser) + parser.add_argument( + "--train-monotonic-only", + action="store_true", + default=False, + help="Only train monotonic attention", + ) + + @classmethod + def build_decoder(cls, args, task, embed_tokens): + tgt_dict = task.tgt_dict + + from examples.simultaneous_translation.models.transformer_monotonic_attention import ( + TransformerMonotonicDecoder, + ) + + decoder = TransformerMonotonicDecoder(args, tgt_dict, embed_tokens) + + if getattr(args, "load_pretrained_decoder_from", None): + decoder = checkpoint_utils.load_pretrained_component_from_model( + component=decoder, checkpoint=args.load_pretrained_decoder_from + ) + return decoder + + +@register_model_architecture( + "convtransformer_simul_trans", "convtransformer_simul_trans_espnet" +) +def convtransformer_simul_trans_espnet(args): + convtransformer_espnet(args) + + +@register_model("convtransformer_augmented_memory") +@augmented_memory +class AugmentedMemoryConvTransformerModel(SimulConvTransformerModel): + @classmethod + def build_encoder(cls, args): + encoder = SequenceEncoder(args, AugmentedMemoryConvTransformerEncoder(args)) + + if getattr(args, "load_pretrained_encoder_from", None) is not None: + encoder = checkpoint_utils.load_pretrained_component_from_model( + component=encoder, checkpoint=args.load_pretrained_encoder_from + ) + + return encoder + + +@register_model_architecture( + "convtransformer_augmented_memory", "convtransformer_augmented_memory" +) +def augmented_memory_convtransformer_espnet(args): + convtransformer_espnet(args) + + +# ============================================================================ # +# Convtransformer +# with monotonic attention decoder +# with emformer encoder +# ============================================================================ # + + +class ConvTransformerEmformerEncoder(ConvTransformerEncoder): + def __init__(self, args): + super().__init__(args) + stride = self.conv_layer_stride(args) + trf_left_context = args.segment_left_context // stride + trf_right_context = args.segment_right_context // stride + context_config = [trf_left_context, trf_right_context] + self.transformer_layers = nn.ModuleList( + [ + NoSegAugmentedMemoryTransformerEncoderLayer( + input_dim=args.encoder_embed_dim, + num_heads=args.encoder_attention_heads, + ffn_dim=args.encoder_ffn_embed_dim, + num_layers=args.encoder_layers, + dropout_in_attn=args.dropout, + dropout_on_attn=args.dropout, + dropout_on_fc1=args.dropout, + dropout_on_fc2=args.dropout, + activation_fn=args.activation_fn, + context_config=context_config, + segment_size=args.segment_length, + max_memory_size=args.max_memory_size, + scaled_init=True, # TODO: use constant for now. + tanh_on_mem=args.amtrf_tanh_on_mem, + ) + ] + ) + self.conv_transformer_encoder = ConvTransformerEncoder(args) + + def forward(self, src_tokens, src_lengths): + encoder_out: Dict[str, List[Tensor]] = self.conv_transformer_encoder(src_tokens, src_lengths.to(src_tokens.device)) + output = encoder_out["encoder_out"][0] + encoder_padding_masks = encoder_out["encoder_padding_mask"] + + return { + "encoder_out": [output], + # This is because that in the original implementation + # the output didn't consider the last segment as right context. + "encoder_padding_mask": [encoder_padding_masks[0][:, : output.size(0)]] if len(encoder_padding_masks) > 0 + else [], + "encoder_embedding": [], + "encoder_states": [], + "src_tokens": [], + "src_lengths": [], + } + + @staticmethod + def conv_layer_stride(args): + # TODO: make it configurable from the args + return 4 + + +@register_model("convtransformer_emformer") +class ConvtransformerEmformer(SimulConvTransformerModel): + @staticmethod + def add_args(parser): + super(ConvtransformerEmformer, ConvtransformerEmformer).add_args(parser) + + parser.add_argument( + "--segment-length", + type=int, + metavar="N", + help="length of each segment (not including left context / right context)", + ) + parser.add_argument( + "--segment-left-context", + type=int, + help="length of left context in a segment", + ) + parser.add_argument( + "--segment-right-context", + type=int, + help="length of right context in a segment", + ) + parser.add_argument( + "--max-memory-size", + type=int, + default=-1, + help="Right context for the segment.", + ) + parser.add_argument( + "--amtrf-tanh-on-mem", + default=False, + action="store_true", + help="whether to use tanh on memory vector", + ) + + @classmethod + def build_encoder(cls, args): + encoder = ConvTransformerEmformerEncoder(args) + if getattr(args, "load_pretrained_encoder_from", None): + encoder = checkpoint_utils.load_pretrained_component_from_model( + component=encoder, checkpoint=args.load_pretrained_encoder_from + ) + return encoder + + +@register_model_architecture( + "convtransformer_emformer", + "convtransformer_emformer", +) +def convtransformer_emformer_base(args): + convtransformer_espnet(args) diff --git a/examples/simultaneous_translation/models/transformer_monotonic_attention.py b/examples/simultaneous_translation/models/transformer_monotonic_attention.py new file mode 100644 index 0000000000000000000000000000000000000000..1062e9b955be475b2eaca4255bac3e7f219fd2a1 --- /dev/null +++ b/examples/simultaneous_translation/models/transformer_monotonic_attention.py @@ -0,0 +1,315 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from typing import Dict, List, NamedTuple, Optional + +import torch +import torch.nn as nn +import torch.nn.functional as F +from examples.simultaneous_translation.modules.monotonic_transformer_layer import ( + TransformerMonotonicDecoderLayer, + TransformerMonotonicEncoderLayer, +) +from fairseq.models import ( + register_model, + register_model_architecture, +) +from fairseq.models.transformer import ( + TransformerModel, + TransformerEncoder, + TransformerDecoder, + base_architecture, + transformer_iwslt_de_en, + transformer_vaswani_wmt_en_de_big, + transformer_vaswani_wmt_en_fr_big, +) +from torch import Tensor + +DEFAULT_MAX_SOURCE_POSITIONS = 1024 +DEFAULT_MAX_TARGET_POSITIONS = 1024 + +TransformerMonotonicDecoderOut = NamedTuple( + "TransformerMonotonicDecoderOut", + [ + ("action", int), + ("attn_list", Optional[List[Optional[Dict[str, Tensor]]]]), + ("step_list", Optional[List[Optional[Tensor]]]), + ("encoder_out", Optional[Dict[str, List[Tensor]]]), + ("encoder_padding_mask", Optional[Tensor]), + ], +) + + +@register_model("transformer_unidirectional") +class TransformerUnidirectionalModel(TransformerModel): + @classmethod + def build_encoder(cls, args, src_dict, embed_tokens): + return TransformerMonotonicEncoder(args, src_dict, embed_tokens) + + +@register_model("transformer_monotonic") +class TransformerModelSimulTrans(TransformerModel): + @classmethod + def build_encoder(cls, args, src_dict, embed_tokens): + return TransformerMonotonicEncoder(args, src_dict, embed_tokens) + + @classmethod + def build_decoder(cls, args, tgt_dict, embed_tokens): + return TransformerMonotonicDecoder(args, tgt_dict, embed_tokens) + + def _indices_from_states(self, states): + if type(states["indices"]["src"]) == list: + if next(self.parameters()).is_cuda: + tensor = torch.cuda.LongTensor + else: + tensor = torch.LongTensor + + src_indices = tensor( + [states["indices"]["src"][: 1 + states["steps"]["src"]]] + ) + + tgt_indices = tensor( + [[self.decoder.dictionary.eos()] + states["indices"]["tgt"]] + ) + else: + src_indices = states["indices"]["src"][: 1 + states["steps"]["src"]] + tgt_indices = states["indices"]["tgt"] + + return src_indices, None, tgt_indices + + +class TransformerMonotonicEncoder(TransformerEncoder): + def __init__(self, args, dictionary, embed_tokens): + super().__init__(args, dictionary, embed_tokens) + + self.dictionary = dictionary + self.layers = nn.ModuleList([]) + self.layers.extend( + [TransformerMonotonicEncoderLayer(args) for i in range(args.encoder_layers)] + ) + + +class TransformerMonotonicDecoder(TransformerDecoder): + """ + Transformer decoder consisting of *args.decoder_layers* layers. Each layer + is a :class:`TransformerDecoderLayer`. + + Args: + args (argparse.Namespace): parsed command-line arguments + dictionary (~fairseq.data.Dictionary): decoding dictionary + embed_tokens (torch.nn.Embedding): output embedding + no_encoder_attn (bool, optional): whether to attend to encoder outputs + (default: False). + """ + + def __init__(self, args, dictionary, embed_tokens, no_encoder_attn=False): + super().__init__(args, dictionary, embed_tokens, no_encoder_attn=False) + + self.dictionary = dictionary + self.layers = nn.ModuleList([]) + self.layers.extend( + [ + TransformerMonotonicDecoderLayer(args, no_encoder_attn) + for _ in range(args.decoder_layers) + ] + ) + + def pre_attention( + self, + prev_output_tokens, + encoder_out_dict: Dict[str, List[Tensor]], + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + ): + positions = ( + self.embed_positions( + prev_output_tokens, + incremental_state=incremental_state, + ) + if self.embed_positions is not None + else None + ) + + if incremental_state is not None: + prev_output_tokens = prev_output_tokens[:, -1:] + if positions is not None: + positions = positions[:, -1:] + # embed tokens and positions + x = self.embed_scale * self.embed_tokens(prev_output_tokens) + + if self.project_in_dim is not None: + x = self.project_in_dim(x) + + if positions is not None: + x += positions + + x = self.dropout_module(x) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + + encoder_out = encoder_out_dict["encoder_out"][0] + encoder_padding_mask = ( + encoder_out_dict["encoder_padding_mask"][0] + if encoder_out_dict["encoder_padding_mask"] + and len(encoder_out_dict["encoder_padding_mask"]) > 0 + else None + ) + + return x, encoder_out, encoder_padding_mask + + def post_attention(self, x): + if self.layer_norm is not None: + x = self.layer_norm(x) + + # T x B x C -> B x T x C + x = x.transpose(0, 1) + + if self.project_out_dim is not None: + x = self.project_out_dim(x) + + return x + + def clear_cache( + self, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]], + end_id: Optional[int] = None, + ): + """ + Clear cache in the monotonic layers. + The cache is generated because of a forward pass of decode but no prediction. + end_id is the last idx of the layers + """ + if end_id is None: + end_id = len(self.layers) + + for index, layer in enumerate(self.layers): + if index < end_id: + layer.prune_incremental_state(incremental_state) + + def extract_features( + self, + prev_output_tokens, + encoder_out: Optional[Dict[str, List[Tensor]]], + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + full_context_alignment: bool = False, # unused + alignment_layer: Optional[int] = None, # unused + alignment_heads: Optional[int] = None, # unsed + ): + """ + Similar to *forward* but only return features. + + Returns: + tuple: + - the decoder's features of shape `(batch, tgt_len, embed_dim)` + - a dictionary with any model-specific outputs + """ + # incremental_state = None + assert encoder_out is not None + (x, encoder_outs, encoder_padding_mask) = self.pre_attention( + prev_output_tokens, encoder_out, incremental_state + ) + attn = None + inner_states = [x] + attn_list: List[Optional[Dict[str, Tensor]]] = [] + step_list: List[Optional[Tensor]] = [] + + for i, layer in enumerate(self.layers): + + x, attn, _ = layer( + x=x, + encoder_out=encoder_outs, + encoder_padding_mask=encoder_padding_mask, + incremental_state=incremental_state, + self_attn_mask=self.buffered_future_mask(x) + if incremental_state is None + else None, + ) + + inner_states.append(x) + attn_list.append(attn) + + if incremental_state is not None: + curr_steps = layer.get_head_steps(incremental_state) + step_list.append(curr_steps) + if_online = incremental_state["online"]["only"] + assert if_online is not None + if if_online.to(torch.bool): + # Online indicates that the encoder states are still changing + assert attn is not None + assert curr_steps is not None + p_choose = ( + attn["p_choose"].squeeze(0).squeeze(1).gather(1, curr_steps.t()) + ) + + new_steps = curr_steps + (p_choose < 0.5).t().type_as(curr_steps) + src = incremental_state["steps"]["src"] + assert src is not None + + if (new_steps >= src).any(): + # We need to prune the last self_attn saved_state + # if model decide not to read + # otherwise there will be duplicated saved_state + self.clear_cache(incremental_state, i + 1) + + return x, TransformerMonotonicDecoderOut( + action=0, + attn_list=None, + step_list=None, + encoder_out=None, + encoder_padding_mask=None, + ) + + x = self.post_attention(x) + + return x, TransformerMonotonicDecoderOut( + action=1, + attn_list=attn_list, + step_list=step_list, + encoder_out=encoder_out, + encoder_padding_mask=encoder_padding_mask, + ) + + def reorder_incremental_state(self, incremental_state, new_order): + super().reorder_incremental_state(incremental_state, new_order) + if "fastest_step" in incremental_state: + incremental_state["fastest_step"] = incremental_state[ + "fastest_step" + ].index_select(0, new_order) + + +@register_model_architecture("transformer_monotonic", "transformer_monotonic") +def base_monotonic_architecture(args): + base_architecture(args) + args.encoder_unidirectional = getattr(args, "encoder_unidirectional", False) + + +@register_model_architecture( + "transformer_monotonic", "transformer_monotonic_iwslt_de_en" +) +def transformer_monotonic_iwslt_de_en(args): + transformer_iwslt_de_en(args) + base_monotonic_architecture(args) + + +# parameters used in the "Attention Is All You Need" paper (Vaswani et al., 2017) +@register_model_architecture( + "transformer_monotonic", "transformer_monotonic_vaswani_wmt_en_de_big" +) +def transformer_monotonic_vaswani_wmt_en_de_big(args): + transformer_vaswani_wmt_en_de_big(args) + + +@register_model_architecture( + "transformer_monotonic", "transformer_monotonic_vaswani_wmt_en_fr_big" +) +def transformer_monotonic_vaswani_wmt_en_fr_big(args): + transformer_monotonic_vaswani_wmt_en_fr_big(args) + + +@register_model_architecture( + "transformer_unidirectional", "transformer_unidirectional_iwslt_de_en" +) +def transformer_unidirectional_iwslt_de_en(args): + transformer_iwslt_de_en(args) diff --git a/examples/simultaneous_translation/modules/__init__.py b/examples/simultaneous_translation/modules/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..c695850c04952c1095edc676cf062f7ee43eb788 --- /dev/null +++ b/examples/simultaneous_translation/modules/__init__.py @@ -0,0 +1,24 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import importlib +import os + +from fairseq import registry + + +( + build_monotonic_attention, + register_monotonic_attention, + MONOTONIC_ATTENTION_REGISTRY, + _, +) = registry.setup_registry("--simul-type") + +for file in sorted(os.listdir(os.path.dirname(__file__))): + if file.endswith(".py") and not file.startswith("_"): + model_name = file[: file.find(".py")] + importlib.import_module( + "examples.simultaneous_translation.modules." + model_name + ) diff --git a/examples/simultaneous_translation/modules/fixed_pre_decision.py b/examples/simultaneous_translation/modules/fixed_pre_decision.py new file mode 100644 index 0000000000000000000000000000000000000000..dd29c031b3b23401dcf61bbe48991934099429a8 --- /dev/null +++ b/examples/simultaneous_translation/modules/fixed_pre_decision.py @@ -0,0 +1,254 @@ +from functools import partial + +import torch +from torch import Tensor +import math +import torch.nn.functional as F + +from . import register_monotonic_attention +from .monotonic_multihead_attention import ( + MonotonicMultiheadAttentionWaitK, + MonotonicMultiheadAttentionHardAligned, + MonotonicMultiheadAttentionInfiniteLookback, +) +from typing import Dict, Optional +from examples.simultaneous_translation.utils import p_choose_strategy + +def fixed_pooling_monotonic_attention(monotonic_attention): + def create_model(monotonic_attention, klass): + class FixedStrideMonotonicAttention(monotonic_attention): + def __init__(self, args): + self.waitk_lagging = 0 + self.num_heads = 0 + self.noise_mean = 0.0 + self.noise_var = 0.0 + super().__init__(args) + self.pre_decision_type = args.fixed_pre_decision_type + self.pre_decision_ratio = args.fixed_pre_decision_ratio + self.pre_decision_pad_threshold = args.fixed_pre_decision_pad_threshold + if self.pre_decision_ratio == 1: + return + + self.strategy = args.simul_type + + if args.fixed_pre_decision_type == "average": + self.pooling_layer = torch.nn.AvgPool1d( + kernel_size=self.pre_decision_ratio, + stride=self.pre_decision_ratio, + ceil_mode=True, + ) + elif args.fixed_pre_decision_type == "last": + + def last(key): + if key.size(2) < self.pre_decision_ratio: + return key + else: + k = key[ + :, + :, + self.pre_decision_ratio - 1 :: self.pre_decision_ratio, + ].contiguous() + if key.size(-1) % self.pre_decision_ratio != 0: + k = torch.cat([k, key[:, :, -1:]], dim=-1).contiguous() + return k + + self.pooling_layer = last + else: + raise NotImplementedError + + @staticmethod + def add_args(parser): + super( + FixedStrideMonotonicAttention, FixedStrideMonotonicAttention + ).add_args(parser) + parser.add_argument( + "--fixed-pre-decision-ratio", + type=int, + required=True, + help=( + "Ratio for the fixed pre-decision," + "indicating how many encoder steps will start" + "simultaneous decision making process." + ), + ) + parser.add_argument( + "--fixed-pre-decision-type", + default="average", + choices=["average", "last"], + help="Pooling type", + ) + parser.add_argument( + "--fixed-pre-decision-pad-threshold", + type=float, + default=0.3, + help="If a part of the sequence has pad" + ",the threshold the pooled part is a pad.", + ) + + def insert_zeros(self, x): + bsz_num_heads, tgt_len, src_len = x.size() + stride = self.pre_decision_ratio + weight = F.pad(torch.ones(1, 1, 1).to(x), (stride - 1, 0)) + x_upsample = F.conv_transpose1d( + x.view(-1, src_len).unsqueeze(1), + weight, + stride=stride, + padding=0, + ) + return x_upsample.squeeze(1).view(bsz_num_heads, tgt_len, -1) + + def p_choose_waitk( + self, query, key, key_padding_mask: Optional[Tensor] = None, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None + ): + """ + query: bsz, tgt_len + key: bsz, src_len + key_padding_mask: bsz, src_len + """ + if incremental_state is not None: + # Retrieve target length from incremental states + # For inference the length of query is always 1 + tgt = incremental_state["steps"]["tgt"] + assert tgt is not None + tgt_len = int(tgt) + else: + tgt_len, bsz, _ = query.size() + + src_len, bsz, _ = key.size() + + p_choose = torch.ones(bsz, tgt_len, src_len).to(query) + p_choose = torch.tril(p_choose, diagonal=self.waitk_lagging - 1) + p_choose = torch.triu(p_choose, diagonal=self.waitk_lagging - 1) + + if incremental_state is not None: + p_choose = p_choose[:, -1:] + tgt_len = 1 + + # Extend to each head + p_choose = ( + p_choose.contiguous() + .unsqueeze(1) + .expand(-1, self.num_heads, -1, -1) + .contiguous() + .view(-1, tgt_len, src_len) + ) + + return p_choose + + def p_choose( + self, + query: Optional[Tensor], + key: Optional[Tensor], + key_padding_mask: Optional[Tensor] = None, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + ): + assert key is not None + assert query is not None + src_len = key.size(0) + tgt_len = query.size(0) + batch_size = query.size(1) + + if self.pre_decision_ratio == 1: + if self.strategy == "waitk": + return p_choose_strategy.waitk( + query, + key, + self.waitk_lagging, + self.num_heads, + key_padding_mask, + incremental_state=incremental_state, + ) + else: # hard_aligned or infinite_lookback + q_proj, k_proj, _ = self.input_projections(query, key, None, "monotonic") + attn_energy = self.attn_energy(q_proj, k_proj, key_padding_mask) + return p_choose_strategy.hard_aligned( + q_proj, + k_proj, + attn_energy, + self.noise_mean, + self.noise_var, + self.training + ) + + key_pool = self.pooling_layer(key.transpose(0, 2)).transpose(0, 2) + + if key_padding_mask is not None: + key_padding_mask_pool = ( + self.pooling_layer(key_padding_mask.unsqueeze(0).float()) + .squeeze(0) + .gt(self.pre_decision_pad_threshold) + ) + # Make sure at least one element is not pad + key_padding_mask_pool[:, 0] = 0 + else: + key_padding_mask_pool = None + + if incremental_state is not None: + # The floor instead of ceil is used for inference + # But make sure the length key_pool at least 1 + if ( + max(1, math.floor(key.size(0) / self.pre_decision_ratio)) + ) < key_pool.size(0): + key_pool = key_pool[:-1] + if key_padding_mask_pool is not None: + key_padding_mask_pool = key_padding_mask_pool[:-1] + + p_choose_pooled = self.p_choose_waitk( + query, + key_pool, + key_padding_mask_pool, + incremental_state=incremental_state, + ) + + # Upsample, interpolate zeros + p_choose = self.insert_zeros(p_choose_pooled) + + if p_choose.size(-1) < src_len: + # Append zeros if the upsampled p_choose is shorter than src_len + p_choose = torch.cat( + [ + p_choose, + torch.zeros( + p_choose.size(0), + tgt_len, + src_len - p_choose.size(-1) + ).to(p_choose) + ], + dim=2 + ) + else: + # can be larger than src_len because we used ceil before + p_choose = p_choose[:, :, :src_len] + p_choose[:, :, -1] = p_choose_pooled[:, :, -1] + + assert list(p_choose.size()) == [ + batch_size * self.num_heads, + tgt_len, + src_len, + ] + + return p_choose + + FixedStrideMonotonicAttention.__name__ = klass.__name__ + return FixedStrideMonotonicAttention + + return partial(create_model, monotonic_attention) + + +@register_monotonic_attention("waitk_fixed_pre_decision") +@fixed_pooling_monotonic_attention(MonotonicMultiheadAttentionWaitK) +class MonotonicMultiheadAttentionWaitkFixedStride: + pass + + +@register_monotonic_attention("hard_aligned_fixed_pre_decision") +@fixed_pooling_monotonic_attention(MonotonicMultiheadAttentionHardAligned) +class MonotonicMultiheadAttentionHardFixedStride: + pass + + +@register_monotonic_attention("infinite_lookback_fixed_pre_decision") +@fixed_pooling_monotonic_attention(MonotonicMultiheadAttentionInfiniteLookback) +class MonotonicMultiheadAttentionInfiniteLookbackFixedStride: + pass diff --git a/examples/simultaneous_translation/modules/monotonic_multihead_attention.py b/examples/simultaneous_translation/modules/monotonic_multihead_attention.py new file mode 100644 index 0000000000000000000000000000000000000000..f49b1daa2fbe920c290055b44a09bfe404fc4f89 --- /dev/null +++ b/examples/simultaneous_translation/modules/monotonic_multihead_attention.py @@ -0,0 +1,910 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +import torch +from torch import Tensor +import torch.nn as nn + +from examples.simultaneous_translation.utils.functions import ( + exclusive_cumprod, + lengths_to_mask, +) +from fairseq.incremental_decoding_utils import with_incremental_state +from fairseq.modules import MultiheadAttention + +from . import register_monotonic_attention +from typing import Dict, Optional + +from examples.simultaneous_translation.utils import p_choose_strategy + +@with_incremental_state +class MonotonicAttention(nn.Module): + """ + Abstract class of monotonic attentions + """ + + def __init__(self, args): + self.eps = args.attention_eps + self.mass_preservation = args.mass_preservation + + self.noise_type = args.noise_type + self.noise_mean = args.noise_mean + self.noise_var = args.noise_var + + self.energy_bias_init = args.energy_bias_init + self.energy_bias = ( + nn.Parameter(self.energy_bias_init * torch.ones([1])) + if args.energy_bias is True + else 0 + ) + + @staticmethod + def add_args(parser): + # fmt: off + parser.add_argument('--no-mass-preservation', action="store_false", + dest="mass_preservation", + help='Do not stay on the last token when decoding') + parser.add_argument('--mass-preservation', action="store_true", + dest="mass_preservation", + help='Stay on the last token when decoding') + parser.set_defaults(mass_preservation=True) + parser.add_argument('--noise-var', type=float, default=1.0, + help='Variance of discretness noise') + parser.add_argument('--noise-mean', type=float, default=0.0, + help='Mean of discretness noise') + parser.add_argument('--noise-type', type=str, default="flat", + help='Type of discretness noise') + parser.add_argument('--energy-bias', action="store_true", + default=False, + help='Bias for energy') + parser.add_argument('--energy-bias-init', type=float, default=-2.0, + help='Initial value of the bias for energy') + parser.add_argument('--attention-eps', type=float, default=1e-6, + help='Epsilon when calculating expected attention') + + def p_choose(self, *args): + raise NotImplementedError + + def input_projections(self, *args): + raise NotImplementedError + + def attn_energy( + self, q_proj, k_proj, key_padding_mask=None, attn_mask=None + ): + """ + Calculating monotonic energies + + ============================================================ + Expected input size + q_proj: bsz * num_heads, tgt_len, self.head_dim + k_proj: bsz * num_heads, src_len, self.head_dim + key_padding_mask: bsz, src_len + attn_mask: tgt_len, src_len + """ + bsz, tgt_len, embed_dim = q_proj.size() + bsz = bsz // self.num_heads + src_len = k_proj.size(1) + + attn_energy = ( + torch.bmm(q_proj, k_proj.transpose(1, 2)) + self.energy_bias + ) + + if attn_mask is not None: + attn_mask = attn_mask.unsqueeze(0) + attn_energy += attn_mask + + attn_energy = attn_energy.view(bsz, self.num_heads, tgt_len, src_len) + + if key_padding_mask is not None: + attn_energy = attn_energy.masked_fill( + key_padding_mask.unsqueeze(1).unsqueeze(2).to(torch.bool), + float("-inf"), + ) + + return attn_energy + + def expected_alignment_train(self, p_choose, key_padding_mask: Optional[Tensor]): + """ + Calculating expected alignment for MMA + Mask is not need because p_choose will be 0 if masked + + q_ij = (1 − p_{ij−1})q_{ij−1} + a+{i−1j} + a_ij = p_ij q_ij + + Parallel solution: + ai = p_i * cumprod(1 − pi) * cumsum(a_i / cumprod(1 − pi)) + + ============================================================ + Expected input size + p_choose: bsz * num_heads, tgt_len, src_len + """ + + # p_choose: bsz * num_heads, tgt_len, src_len + bsz_num_heads, tgt_len, src_len = p_choose.size() + + # cumprod_1mp : bsz * num_heads, tgt_len, src_len + cumprod_1mp = exclusive_cumprod(1 - p_choose, dim=2, eps=self.eps) + cumprod_1mp_clamp = torch.clamp(cumprod_1mp, self.eps, 1.0) + + init_attention = p_choose.new_zeros([bsz_num_heads, 1, src_len]) + init_attention[:, :, 0] = 1.0 + + previous_attn = [init_attention] + + for i in range(tgt_len): + # p_choose: bsz * num_heads, tgt_len, src_len + # cumprod_1mp_clamp : bsz * num_heads, tgt_len, src_len + # previous_attn[i]: bsz * num_heads, 1, src_len + # alpha_i: bsz * num_heads, src_len + alpha_i = ( + p_choose[:, i] + * cumprod_1mp[:, i] + * torch.cumsum(previous_attn[i][:, 0] / cumprod_1mp_clamp[:, i], dim=1) + ).clamp(0, 1.0) + previous_attn.append(alpha_i.unsqueeze(1)) + + # alpha: bsz * num_heads, tgt_len, src_len + alpha = torch.cat(previous_attn[1:], dim=1) + + if self.mass_preservation: + # Last token has the residual probabilities + if key_padding_mask is not None and key_padding_mask[:, -1].any(): + # right padding + batch_size = key_padding_mask.size(0) + residuals = 1 - alpha.sum(dim=-1, keepdim=True).clamp(0.0, 1.0) + src_lens = src_len - key_padding_mask.sum(dim=1, keepdim=True) + src_lens = src_lens.expand( + batch_size, self.num_heads + ).contiguous().view(-1, 1) + src_lens = src_lens.expand(-1, tgt_len).contiguous() + # add back the last value + residuals += alpha.gather(2, src_lens.unsqueeze(-1) - 1) + alpha = alpha.scatter(2, src_lens.unsqueeze(-1) - 1, residuals) + else: + residuals = 1 - alpha[:, :, :-1].sum(dim=-1).clamp(0.0, 1.0) + alpha[:, :, -1] = residuals + + if torch.isnan(alpha).any(): + # Something is wrong + raise RuntimeError("NaN in alpha.") + + return alpha + + def expected_alignment_infer( + self, p_choose, encoder_padding_mask: Optional[Tensor], incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] + ): + # TODO modify this function + """ + Calculating mo alignment for MMA during inference time + + ============================================================ + Expected input size + p_choose: bsz * num_heads, tgt_len, src_len + incremental_state: dict + encodencoder_padding_mask: bsz * src_len + """ + # p_choose: bsz * self.num_heads, src_len + bsz_num_heads, tgt_len, src_len = p_choose.size() + # One token at a time + assert tgt_len == 1 + p_choose = p_choose[:, 0, :] + + monotonic_cache = self._get_monotonic_buffer(incremental_state) + + # prev_monotonic_step: bsz, num_heads + bsz = bsz_num_heads // self.num_heads + prev_monotonic_step = monotonic_cache.get( + "head_step", + p_choose.new_zeros([bsz, self.num_heads]).long() + ) + assert prev_monotonic_step is not None + bsz, num_heads = prev_monotonic_step.size() + assert num_heads == self.num_heads + assert bsz * num_heads == bsz_num_heads + + # p_choose: bsz, num_heads, src_len + p_choose = p_choose.view(bsz, num_heads, src_len) + + if encoder_padding_mask is not None: + src_lengths = src_len - \ + encoder_padding_mask.sum(dim=1, keepdim=True).long() + else: + src_lengths = prev_monotonic_step.new_ones(bsz, 1) * src_len + + # src_lengths: bsz, num_heads + src_lengths = src_lengths.expand_as(prev_monotonic_step) + # new_monotonic_step: bsz, num_heads + new_monotonic_step = prev_monotonic_step + + step_offset = 0 + if encoder_padding_mask is not None: + if encoder_padding_mask[:, 0].any(): + # left_pad_source = True: + step_offset = encoder_padding_mask.sum(dim=-1, keepdim=True) + + max_steps = src_lengths - 1 if self.mass_preservation else src_lengths + + # finish_read: bsz, num_heads + finish_read = new_monotonic_step.eq(max_steps) + p_choose_i = 1 + while finish_read.sum().item() < bsz * self.num_heads: + # p_choose: bsz * self.num_heads, src_len + # only choose the p at monotonic steps + # p_choose_i: bsz , self.num_heads + p_choose_i = ( + p_choose.gather( + 2, + (step_offset + new_monotonic_step) + .unsqueeze(2) + .clamp(0, src_len - 1), + ) + ).squeeze(2) + + action = ( + (p_choose_i < 0.5) + .type_as(prev_monotonic_step) + .masked_fill(finish_read, 0) + ) + # 1 x bsz + # sample actions on unfinished seq + # 1 means stay, finish reading + # 0 means leave, continue reading + # dist = torch.distributions.bernoulli.Bernoulli(p_choose) + # action = dist.sample().type_as(finish_read) * (1 - finish_read) + + new_monotonic_step += action + + finish_read = new_monotonic_step.eq(max_steps) | (action == 0) + + monotonic_cache["head_step"] = new_monotonic_step + # Whether a head is looking for new input + monotonic_cache["head_read"] = ( + new_monotonic_step.eq(max_steps) & (p_choose_i < 0.5) + ) + + # alpha: bsz * num_heads, 1, src_len + # new_monotonic_step: bsz, num_heads + alpha = ( + p_choose + .new_zeros([bsz * self.num_heads, src_len]) + .scatter( + 1, + (step_offset + new_monotonic_step) + .view(bsz * self.num_heads, 1).clamp(0, src_len - 1), + 1 + ) + ) + + if not self.mass_preservation: + alpha = alpha.masked_fill( + (new_monotonic_step == max_steps) + .view(bsz * self.num_heads, 1), + 0 + ) + + alpha = alpha.unsqueeze(1) + + self._set_monotonic_buffer(incremental_state, monotonic_cache) + + return alpha + + def _get_monotonic_buffer(self, incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]]): + return self.get_incremental_state( + incremental_state, + 'monotonic', + ) or {} + + def _set_monotonic_buffer(self, incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]], buffer: Dict[str, Optional[Tensor]]): + self.set_incremental_state( + incremental_state, + 'monotonic', + buffer, + ) + + def v_proj_output(self, value): + raise NotImplementedError + + def forward( + self, query, key, value, + key_padding_mask=None, attn_mask=None, incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + need_weights=True, static_kv=False + ): + + tgt_len, bsz, embed_dim = query.size() + src_len = value.size(0) + + # stepwise prob + # p_choose: bsz * self.num_heads, tgt_len, src_len + p_choose = self.p_choose( + query, key, key_padding_mask, incremental_state, + ) + + # expected alignment alpha + # bsz * self.num_heads, tgt_len, src_len + if incremental_state is not None: + alpha = self.expected_alignment_infer( + p_choose, key_padding_mask, incremental_state) + else: + alpha = self.expected_alignment_train( + p_choose, key_padding_mask) + + # expected attention beta + # bsz * self.num_heads, tgt_len, src_len + beta = self.expected_attention( + alpha, query, key, value, + key_padding_mask, attn_mask, + incremental_state + ) + + attn_weights = beta + + v_proj = self.v_proj_output(value) + + attn = torch.bmm(attn_weights.type_as(v_proj), v_proj) + + attn = attn.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim) + + attn = self.out_proj(attn) + + beta = beta.view(bsz, self.num_heads, tgt_len, src_len) + alpha = alpha.view(bsz, self.num_heads, tgt_len, src_len) + p_choose = p_choose.view(bsz, self.num_heads, tgt_len, src_len) + + return attn, { + "alpha": alpha, + "beta": beta, + "p_choose": p_choose, + } + + +@register_monotonic_attention("hard_aligned") +class MonotonicMultiheadAttentionHardAligned( + MonotonicAttention, MultiheadAttention +): + def __init__(self, args): + MultiheadAttention.__init__( + self, + embed_dim=args.decoder_embed_dim, + num_heads=args.decoder_attention_heads, + kdim=getattr(args, "encoder_embed_dim", None), + vdim=getattr(args, "encoder_embed_dim", None), + dropout=args.attention_dropout, + encoder_decoder_attention=True, + ) + + MonotonicAttention.__init__(self, args) + + self.k_in_proj = {"monotonic": self.k_proj} + self.q_in_proj = {"monotonic": self.q_proj} + self.v_in_proj = {"output": self.v_proj} + + @staticmethod + def add_args(parser): + # fmt: off + parser.add_argument('--no-mass-preservation', action="store_false", + dest="mass_preservation", + help='Do not stay on the last token when decoding') + parser.add_argument('--mass-preservation', action="store_true", + dest="mass_preservation", + help='Stay on the last token when decoding') + parser.set_defaults(mass_preservation=True) + parser.add_argument('--noise-var', type=float, default=1.0, + help='Variance of discretness noise') + parser.add_argument('--noise-mean', type=float, default=0.0, + help='Mean of discretness noise') + parser.add_argument('--noise-type', type=str, default="flat", + help='Type of discretness noise') + parser.add_argument('--energy-bias', action="store_true", + default=False, + help='Bias for energy') + parser.add_argument('--energy-bias-init', type=float, default=-2.0, + help='Initial value of the bias for energy') + parser.add_argument('--attention-eps', type=float, default=1e-6, + help='Epsilon when calculating expected attention') + + def attn_energy( + self, q_proj: Optional[Tensor], k_proj: Optional[Tensor], key_padding_mask: Optional[Tensor] = None, attn_mask: Optional[Tensor] = None + ): + """ + Calculating monotonic energies + + ============================================================ + Expected input size + q_proj: bsz * num_heads, tgt_len, self.head_dim + k_proj: bsz * num_heads, src_len, self.head_dim + key_padding_mask: bsz, src_len + attn_mask: tgt_len, src_len + """ + assert q_proj is not None # Optional[Tensor] annotations in the signature above are to make the JIT compiler happy + assert k_proj is not None + bsz, tgt_len, embed_dim = q_proj.size() + bsz = bsz // self.num_heads + src_len = k_proj.size(1) + + attn_energy = ( + torch.bmm(q_proj, k_proj.transpose(1, 2)) + self.energy_bias + ) + + if attn_mask is not None: + attn_mask = attn_mask.unsqueeze(0) + attn_energy += attn_mask + + attn_energy = attn_energy.view(bsz, self.num_heads, tgt_len, src_len) + + if key_padding_mask is not None: + attn_energy = attn_energy.masked_fill( + key_padding_mask.unsqueeze(1).unsqueeze(2).to(torch.bool), + float("-inf"), + ) + + return attn_energy + + def expected_alignment_train(self, p_choose, key_padding_mask: Optional[Tensor]): + """ + Calculating expected alignment for MMA + Mask is not need because p_choose will be 0 if masked + + q_ij = (1 − p_{ij−1})q_{ij−1} + a+{i−1j} + a_ij = p_ij q_ij + + Parallel solution: + ai = p_i * cumprod(1 − pi) * cumsum(a_i / cumprod(1 − pi)) + + ============================================================ + Expected input size + p_choose: bsz * num_heads, tgt_len, src_len + """ + + # p_choose: bsz * num_heads, tgt_len, src_len + bsz_num_heads, tgt_len, src_len = p_choose.size() + + # cumprod_1mp : bsz * num_heads, tgt_len, src_len + cumprod_1mp = exclusive_cumprod(1 - p_choose, dim=2, eps=self.eps) + cumprod_1mp_clamp = torch.clamp(cumprod_1mp, self.eps, 1.0) + + init_attention = p_choose.new_zeros([bsz_num_heads, 1, src_len]) + init_attention[:, :, 0] = 1.0 + + previous_attn = [init_attention] + + for i in range(tgt_len): + # p_choose: bsz * num_heads, tgt_len, src_len + # cumprod_1mp_clamp : bsz * num_heads, tgt_len, src_len + # previous_attn[i]: bsz * num_heads, 1, src_len + # alpha_i: bsz * num_heads, src_len + alpha_i = ( + p_choose[:, i] + * cumprod_1mp[:, i] + * torch.cumsum(previous_attn[i][:, 0] / cumprod_1mp_clamp[:, i], dim=1) + ).clamp(0, 1.0) + previous_attn.append(alpha_i.unsqueeze(1)) + + # alpha: bsz * num_heads, tgt_len, src_len + alpha = torch.cat(previous_attn[1:], dim=1) + + if self.mass_preservation: + # Last token has the residual probabilities + if key_padding_mask is not None and key_padding_mask[:, -1].any(): + # right padding + batch_size = key_padding_mask.size(0) + residuals = 1 - alpha.sum(dim=-1, keepdim=True).clamp(0.0, 1.0) + src_lens = src_len - key_padding_mask.sum(dim=1, keepdim=True) + src_lens = src_lens.expand( + batch_size, self.num_heads + ).contiguous().view(-1, 1) + src_lens = src_lens.expand(-1, tgt_len).contiguous() + # add back the last value + residuals += alpha.gather(2, src_lens.unsqueeze(-1) - 1) + alpha = alpha.scatter(2, src_lens.unsqueeze(-1) - 1, residuals) + else: + residuals = 1 - alpha[:, :, :-1].sum(dim=-1).clamp(0.0, 1.0) + alpha[:, :, -1] = residuals + + if torch.isnan(alpha).any(): + # Something is wrong + raise RuntimeError("NaN in alpha.") + + return alpha + + def expected_alignment_infer( + self, p_choose, encoder_padding_mask: Optional[Tensor], incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] + ): + # TODO modify this function + """ + Calculating mo alignment for MMA during inference time + + ============================================================ + Expected input size + p_choose: bsz * num_heads, tgt_len, src_len + incremental_state: dict + encodencoder_padding_mask: bsz * src_len + """ + # p_choose: bsz * self.num_heads, src_len + bsz_num_heads, tgt_len, src_len = p_choose.size() + # One token at a time + assert tgt_len == 1 + p_choose = p_choose[:, 0, :] + + monotonic_cache = self._get_monotonic_buffer(incremental_state) + + # prev_monotonic_step: bsz, num_heads + bsz = bsz_num_heads // self.num_heads + prev_monotonic_step = monotonic_cache.get( + "head_step", + p_choose.new_zeros([bsz, self.num_heads]).long() + ) + assert prev_monotonic_step is not None + bsz, num_heads = prev_monotonic_step.size() + assert num_heads == self.num_heads + assert bsz * num_heads == bsz_num_heads + + # p_choose: bsz, num_heads, src_len + p_choose = p_choose.view(bsz, num_heads, src_len) + + if encoder_padding_mask is not None: + src_lengths = src_len - \ + encoder_padding_mask.sum(dim=1, keepdim=True).long() + else: + src_lengths = torch.ones(bsz, 1).to(prev_monotonic_step) * src_len + + # src_lengths: bsz, num_heads + src_lengths = src_lengths.expand_as(prev_monotonic_step) + # new_monotonic_step: bsz, num_heads + new_monotonic_step = prev_monotonic_step + + step_offset = torch.tensor(0) + if encoder_padding_mask is not None: + if encoder_padding_mask[:, 0].any(): + # left_pad_source = True: + step_offset = encoder_padding_mask.sum(dim=-1, keepdim=True) + + max_steps = src_lengths - 1 if self.mass_preservation else src_lengths + + # finish_read: bsz, num_heads + finish_read = new_monotonic_step.eq(max_steps) + p_choose_i = torch.tensor(1) + while finish_read.sum().item() < bsz * self.num_heads: + # p_choose: bsz * self.num_heads, src_len + # only choose the p at monotonic steps + # p_choose_i: bsz , self.num_heads + p_choose_i = ( + p_choose.gather( + 2, + (step_offset + new_monotonic_step) + .unsqueeze(2) + .clamp(0, src_len - 1), + ) + ).squeeze(2) + + action = ( + (p_choose_i < 0.5) + .type_as(prev_monotonic_step) + .masked_fill(finish_read, 0) + ) + # 1 x bsz + # sample actions on unfinished seq + # 1 means stay, finish reading + # 0 means leave, continue reading + # dist = torch.distributions.bernoulli.Bernoulli(p_choose) + # action = dist.sample().type_as(finish_read) * (1 - finish_read) + + new_monotonic_step += action + + finish_read = new_monotonic_step.eq(max_steps) | (action == 0) + + monotonic_cache["head_step"] = new_monotonic_step + # Whether a head is looking for new input + monotonic_cache["head_read"] = ( + new_monotonic_step.eq(max_steps) & (p_choose_i < 0.5) + ) + + # alpha: bsz * num_heads, 1, src_len + # new_monotonic_step: bsz, num_heads + alpha = ( + p_choose + .new_zeros([bsz * self.num_heads, src_len]) + .scatter( + 1, + (step_offset + new_monotonic_step) + .view(bsz * self.num_heads, 1).clamp(0, src_len - 1), + 1 + ) + ) + + if not self.mass_preservation: + alpha = alpha.masked_fill( + (new_monotonic_step == max_steps) + .view(bsz * self.num_heads, 1), + 0 + ) + + alpha = alpha.unsqueeze(1) + + self._set_monotonic_buffer(incremental_state, monotonic_cache) + + return alpha + + def _get_monotonic_buffer(self, incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]]): + maybe_incremental_state = self.get_incremental_state( + incremental_state, + 'monotonic', + ) + if maybe_incremental_state is None: + typed_empty_dict: Dict[str, Optional[Tensor]] = {} + return typed_empty_dict + else: + return maybe_incremental_state + + def _set_monotonic_buffer(self, incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]], buffer: Dict[str, Optional[Tensor]]): + self.set_incremental_state( + incremental_state, + 'monotonic', + buffer, + ) + + def forward( + self, query: Optional[Tensor], key: Optional[Tensor], value: Optional[Tensor], + key_padding_mask: Optional[Tensor] = None, attn_mask: Optional[Tensor] = None, incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + need_weights: bool = True, static_kv: bool = False, need_head_weights: bool = False, + ): + assert query is not None + assert value is not None + tgt_len, bsz, embed_dim = query.size() + src_len = value.size(0) + + # stepwise prob + # p_choose: bsz * self.num_heads, tgt_len, src_len + p_choose = self.p_choose( + query, key, key_padding_mask, incremental_state, + ) + + # expected alignment alpha + # bsz * self.num_heads, tgt_len, src_len + if incremental_state is not None: + alpha = self.expected_alignment_infer( + p_choose, key_padding_mask, incremental_state) + else: + alpha = self.expected_alignment_train( + p_choose, key_padding_mask) + + # expected attention beta + # bsz * self.num_heads, tgt_len, src_len + beta = self.expected_attention( + alpha, query, key, value, + key_padding_mask, attn_mask, + incremental_state + ) + + attn_weights = beta + + v_proj = self.v_proj_output(value) + assert v_proj is not None + + attn = torch.bmm(attn_weights.type_as(v_proj), v_proj) + + attn = attn.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim) + + attn = self.out_proj(attn) + + beta = beta.view(bsz, self.num_heads, tgt_len, src_len) + alpha = alpha.view(bsz, self.num_heads, tgt_len, src_len) + p_choose = p_choose.view(bsz, self.num_heads, tgt_len, src_len) + + return attn, { + "alpha": alpha, + "beta": beta, + "p_choose": p_choose, + } + + def input_projections(self, query: Optional[Tensor], key: Optional[Tensor], value: Optional[Tensor], name: str): + """ + Prepare inputs for multihead attention + + ============================================================ + Expected input size + query: tgt_len, bsz, embed_dim + key: src_len, bsz, embed_dim + value: src_len, bsz, embed_dim + name: monotonic or soft + """ + + if query is not None: + bsz = query.size(1) + q = self.q_proj(query) + q *= self.scaling + q = q.contiguous().view( + -1, bsz * self.num_heads, self.head_dim + ).transpose(0, 1) + else: + q = None + + if key is not None: + bsz = key.size(1) + k = self.k_proj(key) + k = k.contiguous().view( + -1, bsz * self.num_heads, self.head_dim + ).transpose(0, 1) + else: + k = None + + if value is not None: + bsz = value.size(1) + v = self.v_proj(value) + v = v.contiguous().view( + -1, bsz * self.num_heads, self.head_dim + ).transpose(0, 1) + else: + v = None + + return q, k, v + + def p_choose( + self, query: Optional[Tensor], key: Optional[Tensor], key_padding_mask: Optional[Tensor] = None, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None + ): + """ + Calculating step wise prob for reading and writing + 1 to read, 0 to write + + ============================================================ + Expected input size + query: bsz, tgt_len, embed_dim + key: bsz, src_len, embed_dim + value: bsz, src_len, embed_dim + key_padding_mask: bsz, src_len + attn_mask: bsz, src_len + query: bsz, tgt_len, embed_dim + """ + + # prepare inputs + q_proj, k_proj, _ = self.input_projections( + query, key, None, "monotonic" + ) + + # attention energy + attn_energy = self.attn_energy(q_proj, k_proj, key_padding_mask) + + return p_choose_strategy.hard_aligned(q_proj, k_proj, attn_energy, self.noise_mean, self.noise_var, self.training) + + def expected_attention(self, alpha, *args): + """ + For MMA-H, beta = alpha + """ + return alpha + + def v_proj_output(self, value): + _, _, v_proj = self.input_projections(None, None, value, "output") + return v_proj + + +@register_monotonic_attention("infinite_lookback") +class MonotonicMultiheadAttentionInfiniteLookback( + MonotonicMultiheadAttentionHardAligned +): + def __init__(self, args): + super().__init__(args) + self.init_soft_attention() + + def init_soft_attention(self): + self.k_proj_soft = nn.Linear(self.kdim, self.embed_dim, bias=True) + self.q_proj_soft = nn.Linear(self.embed_dim, self.embed_dim, bias=True) + self.k_in_proj["soft"] = self.k_proj_soft + self.q_in_proj["soft"] = self.q_proj_soft + + if self.qkv_same_dim: + # Empirically observed the convergence to be much better with + # the scaled initialization + nn.init.xavier_uniform_( + self.k_in_proj["soft"].weight, gain=1 / math.sqrt(2) + ) + nn.init.xavier_uniform_( + self.q_in_proj["soft"].weight, gain=1 / math.sqrt(2) + ) + else: + nn.init.xavier_uniform_(self.k_in_proj["soft"].weight) + nn.init.xavier_uniform_(self.q_in_proj["soft"].weight) + + def expected_attention( + self, alpha, query: Optional[Tensor], key: Optional[Tensor], value: Optional[Tensor], + key_padding_mask: Optional[Tensor], attn_mask: Optional[Tensor], incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] + ): + # monotonic attention, we will calculate milk here + bsz_x_num_heads, tgt_len, src_len = alpha.size() + bsz = int(bsz_x_num_heads / self.num_heads) + + q, k, _ = self.input_projections(query, key, None, "soft") + soft_energy = self.attn_energy(q, k, key_padding_mask, attn_mask) + + assert list(soft_energy.size()) == \ + [bsz, self.num_heads, tgt_len, src_len] + + soft_energy = soft_energy.view(bsz * self.num_heads, tgt_len, src_len) + + if incremental_state is not None: + monotonic_cache = self._get_monotonic_buffer(incremental_state) + head_step = monotonic_cache["head_step"] + assert head_step is not None + monotonic_length = head_step + 1 + step_offset = 0 + if key_padding_mask is not None: + if key_padding_mask[:, 0].any(): + # left_pad_source = True: + step_offset = key_padding_mask.sum(dim=-1, keepdim=True) + monotonic_length += step_offset + mask = lengths_to_mask( + monotonic_length.view(-1), + soft_energy.size(2), 1 + ).unsqueeze(1) + + soft_energy = soft_energy.masked_fill(~mask.to(torch.bool), float("-inf")) + soft_energy = soft_energy - soft_energy.max(dim=2, keepdim=True)[0] + exp_soft_energy = torch.exp(soft_energy) + exp_soft_energy_sum = exp_soft_energy.sum(dim=2) + beta = exp_soft_energy / exp_soft_energy_sum.unsqueeze(2) + + else: + soft_energy = soft_energy - soft_energy.max(dim=2, keepdim=True)[0] + exp_soft_energy = torch.exp(soft_energy) + self.eps + inner_items = alpha / (torch.cumsum(exp_soft_energy, dim=2)) + + beta = ( + exp_soft_energy + * torch.cumsum(inner_items.flip(dims=[2]), dim=2) + .flip(dims=[2]) + ) + + beta = beta.view(bsz, self.num_heads, tgt_len, src_len) + + if key_padding_mask is not None: + beta = beta.masked_fill( + key_padding_mask.unsqueeze(1).unsqueeze(2).to(torch.bool), 0) + + beta = beta / beta.sum(dim=3, keepdim=True) + beta = beta.view(bsz * self.num_heads, tgt_len, src_len) + beta = self.dropout_module(beta) + + if torch.isnan(beta).any(): + # Something is wrong + raise RuntimeError("NaN in beta.") + + return beta + + +@register_monotonic_attention("waitk") +class MonotonicMultiheadAttentionWaitK( + MonotonicMultiheadAttentionInfiniteLookback +): + def __init__(self, args): + super().__init__(args) + self.q_in_proj["soft"] = self.q_in_proj["monotonic"] + self.k_in_proj["soft"] = self.k_in_proj["monotonic"] + self.waitk_lagging = args.waitk_lagging + assert self.waitk_lagging > 0, ( + f"Lagging has to been larger than 0, get {self.waitk_lagging}." + ) + + @staticmethod + def add_args(parser): + super( + MonotonicMultiheadAttentionWaitK, + MonotonicMultiheadAttentionWaitK, + ).add_args(parser) + + parser.add_argument( + "--waitk-lagging", type=int, required=True, help="Wait K lagging" + ) + + def p_choose( + self, query: Optional[Tensor], key: Optional[Tensor], key_padding_mask: Optional[Tensor] = None, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + ): + """ + query: bsz, tgt_len + key: bsz, src_len + key_padding_mask: bsz, src_len + """ + return p_choose_strategy.waitk(query, key, self.waitk_lagging, self.num_heads, key_padding_mask, incremental_state) diff --git a/examples/simultaneous_translation/modules/monotonic_transformer_layer.py b/examples/simultaneous_translation/modules/monotonic_transformer_layer.py new file mode 100644 index 0000000000000000000000000000000000000000..bcd45aa8a6bbe86d2e3826c9589cc0ae648730a2 --- /dev/null +++ b/examples/simultaneous_translation/modules/monotonic_transformer_layer.py @@ -0,0 +1,198 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq.modules import LayerNorm, TransformerDecoderLayer, TransformerEncoderLayer + +from . import build_monotonic_attention + +from typing import Dict, List, Optional + +import torch +from torch import Tensor + + +class TransformerMonotonicEncoderLayer(TransformerEncoderLayer): + def forward(self, x, encoder_padding_mask): + seq_len, _, _ = x.size() + attn_mask = x.new_ones([seq_len, seq_len]).triu(1) + attn_mask = attn_mask.masked_fill(attn_mask.bool(), float("-inf")) + return super().forward(x, encoder_padding_mask, attn_mask) + + +class TransformerMonotonicDecoderLayer(TransformerDecoderLayer): + def __init__( + self, args, no_encoder_attn=False, add_bias_kv=False, add_zero_attn=False + ): + super().__init__( + args, + no_encoder_attn=True, + add_bias_kv=add_bias_kv, + add_zero_attn=add_zero_attn, + ) + + assert args.simul_type is not None, "A --simul-type is needed." + + self.encoder_attn = build_monotonic_attention(args) + self.encoder_attn_layer_norm = LayerNorm( + self.embed_dim, export=getattr(args, "char_inputs", False) + ) + + def get_head_steps(self, incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]]): + return self.encoder_attn._get_monotonic_buffer(incremental_state).get( + "head_step" + ) + + def prune_incremental_state(self, incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]]): + input_buffer = self.self_attn._get_input_buffer(incremental_state) + for key in ["prev_key", "prev_value"]: + input_buffer_key = input_buffer[key] + assert input_buffer_key is not None + if input_buffer_key.size(2) > 1: + input_buffer[key] = input_buffer_key[:, :, :-1, :] + else: + typed_empty_dict: Dict[str, Optional[Tensor]] = {} + input_buffer = typed_empty_dict + break + assert incremental_state is not None + self.self_attn._set_input_buffer(incremental_state, input_buffer) + + def get_steps(self, incremental_state): + return self.encoder_attn._get_monotonic_buffer(incremental_state).get("step", 0) + + def forward( + self, + x, + encoder_out: Optional[torch.Tensor] = None, + encoder_padding_mask: Optional[torch.Tensor] = None, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + prev_self_attn_state: Optional[List[torch.Tensor]] = None, + prev_attn_state: Optional[List[torch.Tensor]] = None, + self_attn_mask: Optional[torch.Tensor] = None, + self_attn_padding_mask: Optional[torch.Tensor] = None, + need_attn: bool = False, + need_head_weights: bool = False, + ): + """ + Args: + x (Tensor): input to the layer of shape `(seq_len, batch, embed_dim)` + encoder_padding_mask (ByteTensor, optional): binary + ByteTensor of shape `(batch, src_len)` where padding + elements are indicated by ``1``. + need_attn (bool, optional): return attention weights + need_head_weights (bool, optional): return attention weights + for each head (default: return average over heads). + + Returns: + encoded output of shape `(seq_len, batch, embed_dim)` + """ + if need_head_weights: + need_attn = True + + residual = x + if self.normalize_before: + x = self.self_attn_layer_norm(x) + if prev_self_attn_state is not None: + prev_key, prev_value = prev_self_attn_state[:2] + saved_state: Dict[str, Optional[Tensor]] = { + "prev_key": prev_key, + "prev_value": prev_value, + } + if len(prev_self_attn_state) >= 3: + saved_state["prev_key_padding_mask"] = prev_self_attn_state[2] + assert incremental_state is not None + self.self_attn._set_input_buffer(incremental_state, saved_state) + _self_attn_input_buffer = self.self_attn._get_input_buffer(incremental_state) + if self.cross_self_attention and not ( + incremental_state is not None + and _self_attn_input_buffer is not None + and "prev_key" in _self_attn_input_buffer + ): + if self_attn_mask is not None: + assert encoder_out is not None + self_attn_mask = torch.cat( + (x.new_zeros(x.size(0), encoder_out.size(0)), self_attn_mask), dim=1 + ) + if self_attn_padding_mask is not None: + if encoder_padding_mask is None: + assert encoder_out is not None + encoder_padding_mask = self_attn_padding_mask.new_zeros( + encoder_out.size(1), encoder_out.size(0) + ) + self_attn_padding_mask = torch.cat( + (encoder_padding_mask, self_attn_padding_mask), dim=1 + ) + assert encoder_out is not None + y = torch.cat((encoder_out, x), dim=0) + else: + y = x + + x, attn = self.self_attn( + query=x, + key=y, + value=y, + key_padding_mask=self_attn_padding_mask, + incremental_state=incremental_state, + need_weights=False, + attn_mask=self_attn_mask, + ) + x = self.dropout_module(x) + x = self.residual_connection(x, residual) + if not self.normalize_before: + x = self.self_attn_layer_norm(x) + + assert self.encoder_attn is not None + residual = x + if self.normalize_before: + x = self.encoder_attn_layer_norm(x) + if prev_attn_state is not None: + prev_key, prev_value = prev_attn_state[:2] + saved_state: Dict[str, Optional[Tensor]] = { + "prev_key": prev_key, + "prev_value": prev_value, + } + if len(prev_attn_state) >= 3: + saved_state["prev_key_padding_mask"] = prev_attn_state[2] + assert incremental_state is not None + self.encoder_attn._set_input_buffer(incremental_state, saved_state) + + x, attn = self.encoder_attn( + query=x, + key=encoder_out, + value=encoder_out, + key_padding_mask=encoder_padding_mask, + incremental_state=incremental_state, + static_kv=True, + need_weights=need_attn or (not self.training and self.need_attn), + need_head_weights=need_head_weights, + ) + x = self.dropout_module(x) + x = self.residual_connection(x, residual) + if not self.normalize_before: + x = self.encoder_attn_layer_norm(x) + + residual = x + if self.normalize_before: + x = self.final_layer_norm(x) + + x = self.activation_fn(self.fc1(x)) + x = self.activation_dropout_module(x) + x = self.fc2(x) + x = self.dropout_module(x) + x = self.residual_connection(x, residual) + if not self.normalize_before: + x = self.final_layer_norm(x) + if self.onnx_trace and incremental_state is not None: + saved_state = self.self_attn._get_input_buffer(incremental_state) + assert saved_state is not None + if self_attn_padding_mask is not None: + self_attn_state = [ + saved_state["prev_key"], + saved_state["prev_value"], + saved_state["prev_key_padding_mask"], + ] + else: + self_attn_state = [saved_state["prev_key"], saved_state["prev_value"]] + return x, attn, self_attn_state + return x, attn, None diff --git a/examples/simultaneous_translation/utils/__init__.py b/examples/simultaneous_translation/utils/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..1e9ce844f59a4211061392084cc81075e6bab19f --- /dev/null +++ b/examples/simultaneous_translation/utils/__init__.py @@ -0,0 +1,14 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import importlib +import os + + +# automatically import any Python files in the criterions/ directory +for file in sorted(os.listdir(os.path.dirname(__file__))): + if file.endswith(".py") and not file.startswith("_"): + module = file[: file.find(".py")] + importlib.import_module("examples.simultaneous_translation.utils." + module) diff --git a/examples/simultaneous_translation/utils/data_utils.py b/examples/simultaneous_translation/utils/data_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..a763ea6686c66024dca4c84a4a2ca238fad0d856 --- /dev/null +++ b/examples/simultaneous_translation/utils/data_utils.py @@ -0,0 +1,100 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch + + +def calc_mean_invstddev(feature): + if len(feature.size()) != 2: + raise ValueError("We expect the input feature to be 2-D tensor") + mean = feature.mean(0) + var = feature.var(0) + # avoid division by ~zero + eps = 1e-8 + if (var < eps).any(): + return mean, 1.0 / (torch.sqrt(var) + eps) + return mean, 1.0 / torch.sqrt(var) + + +def apply_mv_norm(features): + # If there is less than 2 spectrograms, the variance cannot be computed (is NaN) + # and normalization is not possible, so return the item as it is + if features.size(0) < 2: + return features + mean, invstddev = calc_mean_invstddev(features) + res = (features - mean) * invstddev + return res + + +def lengths_to_encoder_padding_mask(lengths, batch_first: bool = False): + """ + convert lengths (a 1-D Long/Int tensor) to 2-D binary tensor + + Args: + lengths: a (B, )-shaped tensor + + Return: + max_length: maximum length of B sequences + encoder_padding_mask: a (max_length, B) binary mask, where + [t, b] = 0 for t < lengths[b] and 1 otherwise + + TODO: + kernelize this function if benchmarking shows this function is slow + """ + max_lengths = torch.max(lengths).item() + bsz = lengths.size(0) + encoder_padding_mask = torch.arange( + max_lengths + ).to( # a (T, ) tensor with [0, ..., T-1] + lengths.device + ).view( # move to the right device + 1, max_lengths + ).expand( # reshape to (1, T)-shaped tensor + bsz, -1 + ) >= lengths.view( # expand to (B, T)-shaped tensor + bsz, 1 + ).expand( + -1, max_lengths + ) + if not batch_first: + return encoder_padding_mask.t(), max_lengths + else: + return encoder_padding_mask, max_lengths + + +def encoder_padding_mask_to_lengths( + encoder_padding_mask, max_lengths, batch_size, device +): + """ + convert encoder_padding_mask (2-D binary tensor) to a 1-D tensor + + Conventionally, encoder output contains a encoder_padding_mask, which is + a 2-D mask in a shape (T, B), whose (t, b) element indicate whether + encoder_out[t, b] is a valid output (=0) or not (=1). Occasionally, we + need to convert this mask tensor to a 1-D tensor in shape (B, ), where + [b] denotes the valid length of b-th sequence + + Args: + encoder_padding_mask: a (T, B)-shaped binary tensor or None; if None, + indicating all are valid + Return: + seq_lengths: a (B,)-shaped tensor, where its (b, )-th element is the + number of valid elements of b-th sequence + + max_lengths: maximum length of all sequence, if encoder_padding_mask is + not None, max_lengths must equal to encoder_padding_mask.size(0) + + batch_size: batch size; if encoder_padding_mask is + not None, max_lengths must equal to encoder_padding_mask.size(1) + + device: which device to put the result on + """ + if encoder_padding_mask is None: + return torch.Tensor([max_lengths] * batch_size).to(torch.int32).to(device) + + assert encoder_padding_mask.size(0) == max_lengths, "max_lengths does not match" + assert encoder_padding_mask.size(1) == batch_size, "batch_size does not match" + + return max_lengths - torch.sum(encoder_padding_mask, dim=0) diff --git a/examples/simultaneous_translation/utils/functions.py b/examples/simultaneous_translation/utils/functions.py new file mode 100644 index 0000000000000000000000000000000000000000..f795b5f31cee6d9f8387d6402994b9cbb4c98190 --- /dev/null +++ b/examples/simultaneous_translation/utils/functions.py @@ -0,0 +1,149 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch + + +def exclusive_cumprod(tensor, dim: int, eps: float = 1e-10): + """ + Implementing exclusive cumprod. + There is cumprod in pytorch, however there is no exclusive mode. + cumprod(x) = [x1, x1x2, x2x3x4, ..., prod_{i=1}^n x_i] + exclusive means cumprod(x) = [1, x1, x1x2, x1x2x3, ..., prod_{i=1}^{n-1} x_i] + """ + tensor_size = list(tensor.size()) + tensor_size[dim] = 1 + return_tensor = safe_cumprod( + torch.cat([torch.ones(tensor_size).type_as(tensor), tensor], dim=dim), + dim=dim, + eps=eps, + ) + + if dim == 0: + return return_tensor[:-1] + elif dim == 1: + return return_tensor[:, :-1] + elif dim == 2: + return return_tensor[:, :, :-1] + else: + raise RuntimeError("Cumprod on dimension 3 and more is not implemented") + + +def safe_cumprod(tensor, dim: int, eps: float = 1e-10): + """ + An implementation of cumprod to prevent precision issue. + cumprod(x) + = [x1, x1x2, x1x2x3, ....] + = [exp(log(x1)), exp(log(x1) + log(x2)), exp(log(x1) + log(x2) + log(x3)), ...] + = exp(cumsum(log(x))) + """ + + if (tensor + eps < 0).any().item(): + raise RuntimeError( + "Safe cumprod can only take non-negative tensors as input." + "Consider use torch.cumprod if you want to calculate negative values." + ) + + log_tensor = torch.log(tensor + eps) + cumsum_log_tensor = torch.cumsum(log_tensor, dim) + exp_cumsum_log_tensor = torch.exp(cumsum_log_tensor) + return exp_cumsum_log_tensor + + +def lengths_to_mask(lengths, max_len: int, dim: int = 0, negative_mask: bool = False): + """ + Convert a tensor of lengths to mask + For example, lengths = [[2, 3, 4]], max_len = 5 + mask = + [[1, 1, 1], + [1, 1, 1], + [0, 1, 1], + [0, 0, 1], + [0, 0, 0]] + """ + assert len(lengths.size()) <= 2 + if len(lengths) == 2: + if dim == 1: + lengths = lengths.t() + lengths = lengths + else: + lengths = lengths.unsqueeze(1) + + # lengths : batch_size, 1 + lengths = lengths.view(-1, 1) + + batch_size = lengths.size(0) + # batch_size, max_len + mask = torch.arange(max_len).expand(batch_size, max_len).type_as(lengths) < lengths + + if negative_mask: + mask = ~mask + + if dim == 0: + # max_len, batch_size + mask = mask.t() + + return mask + + +def moving_sum(x, start_idx: int, end_idx: int): + """ + From MONOTONIC CHUNKWISE ATTENTION + https://arxiv.org/pdf/1712.05382.pdf + Equation (18) + + x = [x_1, x_2, ..., x_N] + MovingSum(x, start_idx, end_idx)_n = Sigma_{m=n−(start_idx−1)}^{n+end_idx-1} x_m + for n in {1, 2, 3, ..., N} + + x : src_len, batch_size + start_idx : start idx + end_idx : end idx + + Example + src_len = 5 + batch_size = 3 + x = + [[ 0, 5, 10], + [ 1, 6, 11], + [ 2, 7, 12], + [ 3, 8, 13], + [ 4, 9, 14]] + + MovingSum(x, 3, 1) = + [[ 0, 5, 10], + [ 1, 11, 21], + [ 3, 18, 33], + [ 6, 21, 36], + [ 9, 24, 39]] + + MovingSum(x, 1, 3) = + [[ 3, 18, 33], + [ 6, 21, 36], + [ 9, 24, 39], + [ 7, 17, 27], + [ 4, 9, 14]] + """ + assert start_idx > 0 and end_idx > 0 + assert len(x.size()) == 2 + src_len, batch_size = x.size() + # batch_size, 1, src_len + x = x.t().unsqueeze(1) + # batch_size, 1, src_len + moving_sum_weight = x.new_ones([1, 1, end_idx + start_idx - 1]) + + moving_sum = ( + torch.nn.functional.conv1d( + x, moving_sum_weight, padding=start_idx + end_idx - 1 + ) + .squeeze(1) + .t() + ) + moving_sum = moving_sum[end_idx:-start_idx] + + assert src_len == moving_sum.size(0) + assert batch_size == moving_sum.size(1) + + return moving_sum diff --git a/examples/simultaneous_translation/utils/latency.py b/examples/simultaneous_translation/utils/latency.py new file mode 100644 index 0000000000000000000000000000000000000000..5d800a5d9e992be49cedc72b7a9604a32e35fbcc --- /dev/null +++ b/examples/simultaneous_translation/utils/latency.py @@ -0,0 +1,451 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch + + +class LatencyMetric(object): + @staticmethod + def length_from_padding_mask(padding_mask, batch_first: bool = False): + dim = 1 if batch_first else 0 + return padding_mask.size(dim) - padding_mask.sum(dim=dim, keepdim=True) + + def prepare_latency_metric( + self, + delays, + src_lens, + target_padding_mask=None, + batch_first: bool = False, + start_from_zero: bool = True, + ): + assert len(delays.size()) == 2 + assert len(src_lens.size()) == 2 + + if start_from_zero: + delays = delays + 1 + + if batch_first: + # convert to batch_last + delays = delays.t() + src_lens = src_lens.t() + tgt_len, bsz = delays.size() + _, bsz_1 = src_lens.size() + + if target_padding_mask is not None: + target_padding_mask = target_padding_mask.t() + tgt_len_1, bsz_2 = target_padding_mask.size() + assert tgt_len == tgt_len_1 + assert bsz == bsz_2 + + assert bsz == bsz_1 + + if target_padding_mask is None: + tgt_lens = tgt_len * delays.new_ones([1, bsz]).float() + else: + # 1, batch_size + tgt_lens = self.length_from_padding_mask(target_padding_mask, False).float() + delays = delays.masked_fill(target_padding_mask, 0) + + return delays, src_lens, tgt_lens, target_padding_mask + + def __call__( + self, + delays, + src_lens, + target_padding_mask=None, + batch_first: bool = False, + start_from_zero: bool = True, + ): + delays, src_lens, tgt_lens, target_padding_mask = self.prepare_latency_metric( + delays, src_lens, target_padding_mask, batch_first, start_from_zero + ) + return self.cal_metric(delays, src_lens, tgt_lens, target_padding_mask) + + @staticmethod + def cal_metric(delays, src_lens, tgt_lens, target_padding_mask): + """ + Expected sizes: + delays: tgt_len, batch_size + src_lens: 1, batch_size + target_padding_mask: tgt_len, batch_size + """ + raise NotImplementedError + + +class AverageProportion(LatencyMetric): + """ + Function to calculate Average Proportion from + Can neural machine translation do simultaneous translation? + (https://arxiv.org/abs/1606.02012) + + Delays are monotonic steps, range from 1 to src_len. + Give src x tgt y, AP is calculated as: + + AP = 1 / (|x||y]) sum_i^|Y| deleys_i + """ + + @staticmethod + def cal_metric(delays, src_lens, tgt_lens, target_padding_mask): + if target_padding_mask is not None: + AP = torch.sum( + delays.masked_fill(target_padding_mask, 0), dim=0, keepdim=True + ) + else: + AP = torch.sum(delays, dim=0, keepdim=True) + + AP = AP / (src_lens * tgt_lens) + return AP + + +class AverageLagging(LatencyMetric): + """ + Function to calculate Average Lagging from + STACL: Simultaneous Translation with Implicit Anticipation + and Controllable Latency using Prefix-to-Prefix Framework + (https://arxiv.org/abs/1810.08398) + + Delays are monotonic steps, range from 1 to src_len. + Give src x tgt y, AP is calculated as: + + AL = 1 / tau sum_i^tau delays_i - (i - 1) / gamma + + Where + gamma = |y| / |x| + tau = argmin_i(delays_i = |x|) + """ + + @staticmethod + def cal_metric(delays, src_lens, tgt_lens, target_padding_mask): + # tau = argmin_i(delays_i = |x|) + tgt_len, bsz = delays.size() + lagging_padding_mask = delays >= src_lens + lagging_padding_mask = torch.nn.functional.pad( + lagging_padding_mask.t(), (1, 0) + ).t()[:-1, :] + gamma = tgt_lens / src_lens + lagging = ( + delays + - torch.arange(delays.size(0)) + .unsqueeze(1) + .type_as(delays) + .expand_as(delays) + / gamma + ) + lagging.masked_fill_(lagging_padding_mask, 0) + tau = (1 - lagging_padding_mask.type_as(lagging)).sum(dim=0, keepdim=True) + AL = lagging.sum(dim=0, keepdim=True) / tau + + return AL + + +class DifferentiableAverageLagging(LatencyMetric): + """ + Function to calculate Differentiable Average Lagging from + Monotonic Infinite Lookback Attention for Simultaneous Machine Translation + (https://arxiv.org/abs/1906.05218) + + Delays are monotonic steps, range from 0 to src_len-1. + (In the original paper thery are from 1 to src_len) + Give src x tgt y, AP is calculated as: + + DAL = 1 / |Y| sum_i^|Y| delays'_i - (i - 1) / gamma + + Where + delays'_i = + 1. delays_i if i == 1 + 2. max(delays_i, delays'_{i-1} + 1 / gamma) + + """ + + @staticmethod + def cal_metric(delays, src_lens, tgt_lens, target_padding_mask): + tgt_len, bsz = delays.size() + + gamma = tgt_lens / src_lens + new_delays = torch.zeros_like(delays) + + for i in range(delays.size(0)): + if i == 0: + new_delays[i] = delays[i] + else: + new_delays[i] = torch.cat( + [ + new_delays[i - 1].unsqueeze(0) + 1 / gamma, + delays[i].unsqueeze(0), + ], + dim=0, + ).max(dim=0)[0] + + DAL = ( + new_delays + - torch.arange(delays.size(0)) + .unsqueeze(1) + .type_as(delays) + .expand_as(delays) + / gamma + ) + if target_padding_mask is not None: + DAL = DAL.masked_fill(target_padding_mask, 0) + + DAL = DAL.sum(dim=0, keepdim=True) / tgt_lens + + return DAL + + +class LatencyMetricVariance(LatencyMetric): + def prepare_latency_metric( + self, + delays, + src_lens, + target_padding_mask=None, + batch_first: bool = True, + start_from_zero: bool = True, + ): + assert batch_first + assert len(delays.size()) == 3 + assert len(src_lens.size()) == 2 + + if start_from_zero: + delays = delays + 1 + + # convert to batch_last + bsz, num_heads_x_layers, tgt_len = delays.size() + bsz_1, _ = src_lens.size() + assert bsz == bsz_1 + + if target_padding_mask is not None: + bsz_2, tgt_len_1 = target_padding_mask.size() + assert tgt_len == tgt_len_1 + assert bsz == bsz_2 + + if target_padding_mask is None: + tgt_lens = tgt_len * delays.new_ones([bsz, tgt_len]).float() + else: + # batch_size, 1 + tgt_lens = self.length_from_padding_mask(target_padding_mask, True).float() + delays = delays.masked_fill(target_padding_mask.unsqueeze(1), 0) + + return delays, src_lens, tgt_lens, target_padding_mask + + +class VarianceDelay(LatencyMetricVariance): + @staticmethod + def cal_metric(delays, src_lens, tgt_lens, target_padding_mask): + """ + delays : bsz, num_heads_x_layers, tgt_len + src_lens : bsz, 1 + target_lens : bsz, 1 + target_padding_mask: bsz, tgt_len or None + """ + if delays.size(1) == 1: + return delays.new_zeros([1]) + + variance_delays = delays.var(dim=1) + + if target_padding_mask is not None: + variance_delays.masked_fill_(target_padding_mask, 0) + + return variance_delays.sum(dim=1, keepdim=True) / tgt_lens + + +class LatencyInference(object): + def __init__(self, start_from_zero=True): + self.metric_calculator = { + "differentiable_average_lagging": DifferentiableAverageLagging(), + "average_lagging": AverageLagging(), + "average_proportion": AverageProportion(), + } + + self.start_from_zero = start_from_zero + + def __call__(self, monotonic_step, src_lens): + """ + monotonic_step range from 0 to src_len. src_len means eos + delays: bsz, tgt_len + src_lens: bsz, 1 + """ + if not self.start_from_zero: + monotonic_step -= 1 + + src_lens = src_lens + + delays = monotonic_step.view( + monotonic_step.size(0), -1, monotonic_step.size(-1) + ).max(dim=1)[0] + + delays = delays.masked_fill(delays >= src_lens, 0) + (src_lens - 1).expand_as( + delays + ).masked_fill(delays < src_lens, 0) + return_dict = {} + for key, func in self.metric_calculator.items(): + return_dict[key] = func( + delays.float(), + src_lens.float(), + target_padding_mask=None, + batch_first=True, + start_from_zero=True, + ).t() + + return return_dict + + +class LatencyTraining(object): + def __init__( + self, + avg_weight, + var_weight, + avg_type, + var_type, + stay_on_last_token, + average_method, + ): + self.avg_weight = avg_weight + self.var_weight = var_weight + self.avg_type = avg_type + self.var_type = var_type + self.stay_on_last_token = stay_on_last_token + self.average_method = average_method + + self.metric_calculator = { + "differentiable_average_lagging": DifferentiableAverageLagging(), + "average_lagging": AverageLagging(), + "average_proportion": AverageProportion(), + } + + self.variance_calculator = { + "variance_delay": VarianceDelay(), + } + + def expected_delays_from_attention( + self, attention, source_padding_mask=None, target_padding_mask=None + ): + if type(attention) == list: + # bsz, num_heads, tgt_len, src_len + bsz, num_heads, tgt_len, src_len = attention[0].size() + attention = torch.cat(attention, dim=1) + bsz, num_heads_x_layers, tgt_len, src_len = attention.size() + # bsz * num_heads * num_layers, tgt_len, src_len + attention = attention.view(-1, tgt_len, src_len) + else: + # bsz * num_heads * num_layers, tgt_len, src_len + bsz, tgt_len, src_len = attention.size() + num_heads_x_layers = 1 + attention = attention.view(-1, tgt_len, src_len) + + if not self.stay_on_last_token: + residual_attention = 1 - attention[:, :, :-1].sum(dim=2, keepdim=True) + attention = torch.cat([attention[:, :, :-1], residual_attention], dim=2) + + # bsz * num_heads_x_num_layers, tgt_len, src_len for MMA + steps = ( + torch.arange(1, 1 + src_len) + .unsqueeze(0) + .unsqueeze(1) + .expand_as(attention) + .type_as(attention) + ) + + if source_padding_mask is not None: + src_offset = ( + source_padding_mask.type_as(attention) + .sum(dim=1, keepdim=True) + .expand(bsz, num_heads_x_layers) + .contiguous() + .view(-1, 1) + ) + src_lens = src_len - src_offset + if source_padding_mask[:, 0].any(): + # Pad left + src_offset = src_offset.view(-1, 1, 1) + steps = steps - src_offset + steps = steps.masked_fill(steps <= 0, 0) + else: + src_lens = attention.new_ones([bsz, num_heads_x_layers]) * src_len + src_lens = src_lens.view(-1, 1) + + # bsz * num_heads_num_layers, tgt_len, src_len + expected_delays = ( + (steps * attention).sum(dim=2).view(bsz, num_heads_x_layers, tgt_len) + ) + + if target_padding_mask is not None: + expected_delays.masked_fill_(target_padding_mask.unsqueeze(1), 0) + + return expected_delays, src_lens + + def avg_loss(self, expected_delays, src_lens, target_padding_mask): + + bsz, num_heads_x_layers, tgt_len = expected_delays.size() + target_padding_mask = ( + target_padding_mask.unsqueeze(1) + .expand_as(expected_delays) + .contiguous() + .view(-1, tgt_len) + ) + + if self.average_method == "average": + # bsz * tgt_len + expected_delays = expected_delays.mean(dim=1) + elif self.average_method == "weighted_average": + weights = torch.nn.functional.softmax(expected_delays, dim=1) + expected_delays = torch.sum(expected_delays * weights, dim=1) + elif self.average_method == "max": + # bsz * num_heads_x_num_layers, tgt_len + expected_delays = expected_delays.max(dim=1)[0] + else: + raise RuntimeError(f"{self.average_method} is not supported") + + src_lens = src_lens.view(bsz, -1)[:, :1] + target_padding_mask = target_padding_mask.view(bsz, -1, tgt_len)[:, 0] + + if self.avg_weight > 0.0: + if self.avg_type in self.metric_calculator: + average_delays = self.metric_calculator[self.avg_type]( + expected_delays, + src_lens, + target_padding_mask, + batch_first=True, + start_from_zero=False, + ) + else: + raise RuntimeError(f"{self.avg_type} is not supported.") + + # bsz * num_heads_x_num_layers, 1 + return self.avg_weight * average_delays.sum() + else: + return 0.0 + + def var_loss(self, expected_delays, src_lens, target_padding_mask): + src_lens = src_lens.view(expected_delays.size(0), expected_delays.size(1))[ + :, :1 + ] + if self.var_weight > 0.0: + if self.var_type in self.variance_calculator: + variance_delays = self.variance_calculator[self.var_type]( + expected_delays, + src_lens, + target_padding_mask, + batch_first=True, + start_from_zero=False, + ) + else: + raise RuntimeError(f"{self.var_type} is not supported.") + + return self.var_weight * variance_delays.sum() + else: + return 0.0 + + def loss(self, attention, source_padding_mask=None, target_padding_mask=None): + expected_delays, src_lens = self.expected_delays_from_attention( + attention, source_padding_mask, target_padding_mask + ) + + latency_loss = 0 + + latency_loss += self.avg_loss(expected_delays, src_lens, target_padding_mask) + + latency_loss += self.var_loss(expected_delays, src_lens, target_padding_mask) + + return latency_loss diff --git a/examples/simultaneous_translation/utils/p_choose_strategy.py b/examples/simultaneous_translation/utils/p_choose_strategy.py new file mode 100644 index 0000000000000000000000000000000000000000..308227ed96d8ee94b66bc0df343c96abbe2c55cc --- /dev/null +++ b/examples/simultaneous_translation/utils/p_choose_strategy.py @@ -0,0 +1,124 @@ +from typing import Optional, Dict +from torch import Tensor +import torch + + +def waitk( + query, key, waitk_lagging: int, num_heads: int, key_padding_mask: Optional[Tensor] = None, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None +): + if incremental_state is not None: + # Retrieve target length from incremental states + # For inference the length of query is always 1 + tgt_len = incremental_state["steps"]["tgt"] + assert tgt_len is not None + tgt_len = int(tgt_len) + else: + tgt_len, bsz, _ = query.size() + + max_src_len, bsz, _ = key.size() + + if max_src_len < waitk_lagging: + if incremental_state is not None: + tgt_len = 1 + return query.new_zeros( + bsz * num_heads, tgt_len, max_src_len + ) + + # Assuming the p_choose looks like this for wait k=3 + # src_len = 6, tgt_len = 5 + # [0, 0, 1, 0, 0, 0, 0] + # [0, 0, 0, 1, 0, 0, 0] + # [0, 0, 0, 0, 1, 0, 0] + # [0, 0, 0, 0, 0, 1, 0] + # [0, 0, 0, 0, 0, 0, 1] + # linearize the p_choose matrix: + # [0, 0, 1, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0...] + # The indices of linearized matrix that equals 1 is + # 2 + 6 * 0 + # 3 + 6 * 1 + # ... + # n + src_len * n + k - 1 = n * (src_len + 1) + k - 1 + # n from 0 to tgt_len - 1 + # + # First, generate the indices (activate_indices_offset: bsz, tgt_len) + # Second, scatter a zeros tensor (bsz, tgt_len * src_len) + # with activate_indices_offset + # Third, resize the tensor to (bsz, tgt_len, src_len) + + activate_indices_offset = ( + ( + torch.arange(tgt_len) * (max_src_len + 1) + + waitk_lagging - 1 + ) + .unsqueeze(0) + .expand(bsz, tgt_len) + .to(query) + .long() + ) + + if key_padding_mask is not None: + if key_padding_mask[:, 0].any(): + # Left padding + activate_indices_offset += ( + key_padding_mask.sum(dim=1, keepdim=True) + ) + + # Need to clamp the indices that are too large + activate_indices_offset = ( + activate_indices_offset + .clamp( + 0, + min( + [ + tgt_len, + max_src_len - waitk_lagging + 1 + ] + ) * max_src_len - 1 + ) + ) + + p_choose = torch.zeros(bsz, tgt_len * max_src_len).to(query) + + p_choose = p_choose.scatter( + 1, + activate_indices_offset, + 1.0 + ).view(bsz, tgt_len, max_src_len) + + if incremental_state is not None: + p_choose = p_choose[:, -1:] + tgt_len = 1 + + # Extend to each head + p_choose = ( + p_choose.contiguous() + .unsqueeze(1) + .expand(-1, num_heads, -1, -1) + .contiguous() + .view(-1, tgt_len, max_src_len) + ) + + return p_choose + + +def hard_aligned(q_proj: Optional[Tensor], k_proj: Optional[Tensor], attn_energy, noise_mean: float = 0.0, noise_var: float = 0.0, training: bool = True): + """ + Calculating step wise prob for reading and writing + 1 to read, 0 to write + """ + + noise = 0 + if training: + # add noise here to encourage discretness + noise = ( + torch.normal(noise_mean, noise_var, attn_energy.size()) + .type_as(attn_energy) + .to(attn_energy.device) + ) + + p_choose = torch.sigmoid(attn_energy + noise) + _, _, tgt_len, src_len = p_choose.size() + + # p_choose: bsz * self.num_heads, tgt_len, src_len + return p_choose.view(-1, tgt_len, src_len) diff --git a/examples/speech_recognition/README.md b/examples/speech_recognition/README.md new file mode 100644 index 0000000000000000000000000000000000000000..17030bf0fd50bb843a508e13e97ed436eae33287 --- /dev/null +++ b/examples/speech_recognition/README.md @@ -0,0 +1,83 @@ +### 2021 Update: We are merging this example into the [S2T framework](../speech_to_text), which supports more generic speech-to-text tasks (e.g. speech translation) and more flexible data processing pipelines. Please stay tuned. + +# Speech Recognition +`examples/speech_recognition` is implementing ASR task in Fairseq, along with needed features, datasets, models and loss functions to train and infer model described in [Transformers with convolutional context for ASR (Abdelrahman Mohamed et al., 2019)](https://arxiv.org/abs/1904.11660). + + +## Additional dependencies +On top of main fairseq dependencies there are couple more additional requirements. + +1) Please follow the instructions to install [torchaudio](https://github.com/pytorch/audio). This is required to compute audio fbank features. +2) [Sclite](http://www1.icsi.berkeley.edu/Speech/docs/sctk-1.2/sclite.htm#sclite_name_0) is used to measure WER. Sclite can be downloaded and installed from source from sctk package [here](http://www.openslr.org/4/). Training and inference doesn't require Sclite dependency. +3) [sentencepiece](https://github.com/google/sentencepiece) is required in order to create dataset with word-piece targets. + +## Preparing librispeech data +``` +./examples/speech_recognition/datasets/prepare-librispeech.sh $DIR_TO_SAVE_RAW_DATA $DIR_FOR_PREPROCESSED_DATA +``` + +## Training librispeech data +``` +python train.py $DIR_FOR_PREPROCESSED_DATA --save-dir $MODEL_PATH --max-epoch 80 --task speech_recognition --arch vggtransformer_2 --optimizer adadelta --lr 1.0 --adadelta-eps 1e-8 --adadelta-rho 0.95 --clip-norm 10.0 --max-tokens 5000 --log-format json --log-interval 1 --criterion cross_entropy_acc --user-dir examples/speech_recognition/ +``` + +## Inference for librispeech +`$SET` can be `test_clean` or `test_other` +Any checkpoint in `$MODEL_PATH` can be selected. In this example we are working with `checkpoint_last.pt` +``` +python examples/speech_recognition/infer.py $DIR_FOR_PREPROCESSED_DATA --task speech_recognition --max-tokens 25000 --nbest 1 --path $MODEL_PATH/checkpoint_last.pt --beam 20 --results-path $RES_DIR --batch-size 40 --gen-subset $SET --user-dir examples/speech_recognition/ +``` + +## Inference for librispeech +``` +sclite -r ${RES_DIR}/ref.word-checkpoint_last.pt-${SET}.txt -h ${RES_DIR}/hypo.word-checkpoint_last.pt-${SET}.txt -i rm -o all stdout > $RES_REPORT +``` +`Sum/Avg` row from first table of the report has WER + +## Using flashlight (previously called [wav2letter](https://github.com/facebookresearch/wav2letter)) components +[flashlight](https://github.com/facebookresearch/flashlight) now has integration with fairseq. Currently this includes: + +* AutoSegmentationCriterion (ASG) +* flashlight-style Conv/GLU model +* flashlight's beam search decoder + +To use these, follow the instructions on [this page](https://github.com/facebookresearch/flashlight/tree/master/bindings/python) to install python bindings. + +## Training librispeech data (flashlight style, Conv/GLU + ASG loss) +Training command: +``` +python train.py $DIR_FOR_PREPROCESSED_DATA --save-dir $MODEL_PATH --max-epoch 100 --task speech_recognition --arch w2l_conv_glu_enc --batch-size 4 --optimizer sgd --lr 0.3,0.8 --momentum 0.8 --clip-norm 0.2 --max-tokens 50000 --log-format json --log-interval 100 --num-workers 0 --sentence-avg --criterion asg_loss --asg-transitions-init 5 --max-replabel 2 --linseg-updates 8789 --user-dir examples/speech_recognition +``` + +Note that ASG loss currently doesn't do well with word-pieces. You should prepare a dataset with character targets by setting `nbpe=31` in `prepare-librispeech.sh`. + +## Inference for librispeech (flashlight decoder, n-gram LM) +Inference command: +``` +python examples/speech_recognition/infer.py $DIR_FOR_PREPROCESSED_DATA --task speech_recognition --seed 1 --nbest 1 --path $MODEL_PATH/checkpoint_last.pt --gen-subset $SET --results-path $RES_DIR --w2l-decoder kenlm --kenlm-model $KENLM_MODEL_PATH --lexicon $LEXICON_PATH --beam 200 --beam-threshold 15 --lm-weight 1.5 --word-score 1.5 --sil-weight -0.3 --criterion asg_loss --max-replabel 2 --user-dir examples/speech_recognition +``` + +`$KENLM_MODEL_PATH` should be a standard n-gram language model file. `$LEXICON_PATH` should be a flashlight-style lexicon (list of known words and their spellings). For ASG inference, a lexicon line should look like this (note the repetition labels): +``` +doorbell D O 1 R B E L 1 ▁ +``` +For CTC inference with word-pieces, repetition labels are not used and the lexicon should have most common spellings for each word (one can use sentencepiece's `NBestEncodeAsPieces` for this): +``` +doorbell ▁DOOR BE LL +doorbell ▁DOOR B E LL +doorbell ▁DO OR BE LL +doorbell ▁DOOR B EL L +doorbell ▁DOOR BE L L +doorbell ▁DO OR B E LL +doorbell ▁DOOR B E L L +doorbell ▁DO OR B EL L +doorbell ▁DO O R BE LL +doorbell ▁DO OR BE L L +``` +Lowercase vs. uppercase matters: the *word* should match the case of the n-gram language model (i.e. `$KENLM_MODEL_PATH`), while the *spelling* should match the case of the token dictionary (i.e. `$DIR_FOR_PREPROCESSED_DATA/dict.txt`). + +## Inference for librispeech (flashlight decoder, viterbi only) +Inference command: +``` +python examples/speech_recognition/infer.py $DIR_FOR_PREPROCESSED_DATA --task speech_recognition --seed 1 --nbest 1 --path $MODEL_PATH/checkpoint_last.pt --gen-subset $SET --results-path $RES_DIR --w2l-decoder viterbi --criterion asg_loss --max-replabel 2 --user-dir examples/speech_recognition +``` diff --git a/examples/speech_recognition/__init__.py b/examples/speech_recognition/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..0278f6a27340c7ff7e207d09348483d1b0d3a100 --- /dev/null +++ b/examples/speech_recognition/__init__.py @@ -0,0 +1 @@ +from . import criterions, models, tasks # noqa diff --git a/examples/speech_recognition/criterions/ASG_loss.py b/examples/speech_recognition/criterions/ASG_loss.py new file mode 100644 index 0000000000000000000000000000000000000000..41f50bbd70388ce723f2d316d4e9776bcd6be3c9 --- /dev/null +++ b/examples/speech_recognition/criterions/ASG_loss.py @@ -0,0 +1,170 @@ +#!/usr/bin/env python3 + +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +from examples.speech_recognition.data.replabels import pack_replabels +from fairseq import utils +from fairseq.criterions import FairseqCriterion, register_criterion + + +@register_criterion("asg_loss") +class ASGCriterion(FairseqCriterion): + @staticmethod + def add_args(parser): + group = parser.add_argument_group("ASG Loss") + group.add_argument( + "--asg-transitions-init", + help="initial diagonal value of transition matrix", + type=float, + default=0.0, + ) + group.add_argument( + "--max-replabel", help="maximum # of replabels", type=int, default=2 + ) + group.add_argument( + "--linseg-updates", + help="# of training updates to use LinSeg initialization", + type=int, + default=0, + ) + group.add_argument( + "--hide-linseg-messages", + help="hide messages about LinSeg initialization", + action="store_true", + ) + + def __init__( + self, + task, + silence_token, + asg_transitions_init, + max_replabel, + linseg_updates, + hide_linseg_messages, + ): + from flashlight.lib.sequence.criterion import ASGLoss, CriterionScaleMode + + super().__init__(task) + self.tgt_dict = task.target_dictionary + self.eos = self.tgt_dict.eos() + self.silence = ( + self.tgt_dict.index(silence_token) + if silence_token in self.tgt_dict + else None + ) + self.max_replabel = max_replabel + + num_labels = len(self.tgt_dict) + self.asg = ASGLoss(num_labels, scale_mode=CriterionScaleMode.TARGET_SZ_SQRT) + self.asg.trans = torch.nn.Parameter( + asg_transitions_init * torch.eye(num_labels), requires_grad=True + ) + + self.linseg_progress = torch.nn.Parameter( + torch.tensor([0], dtype=torch.int), requires_grad=False + ) + self.linseg_maximum = linseg_updates + self.linseg_message_state = "none" if hide_linseg_messages else "start" + + @classmethod + def build_criterion(cls, args, task): + return cls( + task, + args.silence_token, + args.asg_transitions_init, + args.max_replabel, + args.linseg_updates, + args.hide_linseg_messages, + ) + + def linseg_step(self): + if not self.training: + return False + if self.linseg_progress.item() < self.linseg_maximum: + if self.linseg_message_state == "start": + print("| using LinSeg to initialize ASG") + self.linseg_message_state = "finish" + self.linseg_progress.add_(1) + return True + elif self.linseg_message_state == "finish": + print("| finished LinSeg initialization") + self.linseg_message_state = "none" + return False + + def replace_eos_with_silence(self, tgt): + if tgt[-1] != self.eos: + return tgt + elif self.silence is None or (len(tgt) > 1 and tgt[-2] == self.silence): + return tgt[:-1] + else: + return tgt[:-1] + [self.silence] + + def forward(self, model, sample, reduce=True): + """Compute the loss for the given sample. + + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + + net_output = model(**sample["net_input"]) + emissions = net_output["encoder_out"].transpose(0, 1).contiguous() + B = emissions.size(0) + T = emissions.size(1) + device = emissions.device + + target = torch.IntTensor(B, T) + target_size = torch.IntTensor(B) + using_linseg = self.linseg_step() + + for b in range(B): + initial_target_size = sample["target_lengths"][b].item() + if initial_target_size == 0: + raise ValueError("target size cannot be zero") + + tgt = sample["target"][b, :initial_target_size].tolist() + tgt = self.replace_eos_with_silence(tgt) + tgt = pack_replabels(tgt, self.tgt_dict, self.max_replabel) + tgt = tgt[:T] + + if using_linseg: + tgt = [tgt[t * len(tgt) // T] for t in range(T)] + + target[b][: len(tgt)] = torch.IntTensor(tgt) + target_size[b] = len(tgt) + + loss = self.asg.forward(emissions, target.to(device), target_size.to(device)) + + if reduce: + loss = torch.sum(loss) + + sample_size = ( + sample["target"].size(0) if self.args.sentence_avg else sample["ntokens"] + ) + logging_output = { + "loss": utils.item(loss.data) if reduce else loss.data, + "ntokens": sample["ntokens"], + "nsentences": sample["target"].size(0), + "sample_size": sample_size, + } + return loss, sample_size, logging_output + + @staticmethod + def aggregate_logging_outputs(logging_outputs): + """Aggregate logging outputs from data parallel training.""" + loss_sum = sum(log.get("loss", 0) for log in logging_outputs) + ntokens = sum(log.get("ntokens", 0) for log in logging_outputs) + nsentences = sum(log.get("nsentences", 0) for log in logging_outputs) + sample_size = sum(log.get("sample_size", 0) for log in logging_outputs) + agg_output = { + "loss": loss_sum / nsentences, + "ntokens": ntokens, + "nsentences": nsentences, + "sample_size": sample_size, + } + return agg_output diff --git a/examples/speech_recognition/criterions/__init__.py b/examples/speech_recognition/criterions/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..579abd2ace1b14b80f5e53e5c96583e4d5b14c52 --- /dev/null +++ b/examples/speech_recognition/criterions/__init__.py @@ -0,0 +1,17 @@ +import importlib +import os + + +# ASG loss requires flashlight bindings +files_to_skip = set() +try: + import flashlight.lib.sequence.criterion +except ImportError: + files_to_skip.add("ASG_loss.py") + +for file in sorted(os.listdir(os.path.dirname(__file__))): + if file.endswith(".py") and not file.startswith("_") and file not in files_to_skip: + criterion_name = file[: file.find(".py")] + importlib.import_module( + "examples.speech_recognition.criterions." + criterion_name + ) diff --git a/examples/speech_recognition/criterions/cross_entropy_acc.py b/examples/speech_recognition/criterions/cross_entropy_acc.py new file mode 100644 index 0000000000000000000000000000000000000000..7c4d8ba3802a2da9467c42b0aa18653c7bbb2ec9 --- /dev/null +++ b/examples/speech_recognition/criterions/cross_entropy_acc.py @@ -0,0 +1,130 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from __future__ import absolute_import, division, print_function, unicode_literals + +import logging +import math + +import torch +import torch.nn.functional as F +from fairseq import utils +from fairseq.criterions import FairseqCriterion, register_criterion + + +@register_criterion("cross_entropy_acc") +class CrossEntropyWithAccCriterion(FairseqCriterion): + def __init__(self, task, sentence_avg): + super().__init__(task) + self.sentence_avg = sentence_avg + + def compute_loss(self, model, net_output, target, reduction, log_probs): + # N, T -> N * T + target = target.view(-1) + lprobs = model.get_normalized_probs(net_output, log_probs=log_probs) + if not hasattr(lprobs, "batch_first"): + logging.warning( + "ERROR: we need to know whether " + "batch first for the net output; " + "you need to set batch_first attribute for the return value of " + "model.get_normalized_probs. Now, we assume this is true, but " + "in the future, we will raise exception instead. " + ) + batch_first = getattr(lprobs, "batch_first", True) + if not batch_first: + lprobs = lprobs.transpose(0, 1) + + # N, T, D -> N * T, D + lprobs = lprobs.view(-1, lprobs.size(-1)) + loss = F.nll_loss( + lprobs, target, ignore_index=self.padding_idx, reduction=reduction + ) + return lprobs, loss + + def get_logging_output(self, sample, target, lprobs, loss): + target = target.view(-1) + mask = target != self.padding_idx + correct = torch.sum( + lprobs.argmax(1).masked_select(mask) == target.masked_select(mask) + ) + total = torch.sum(mask) + sample_size = ( + sample["target"].size(0) if self.sentence_avg else sample["ntokens"] + ) + + logging_output = { + "loss": utils.item(loss.data), # * sample['ntokens'], + "ntokens": sample["ntokens"], + "nsentences": sample["target"].size(0), + "sample_size": sample_size, + "correct": utils.item(correct.data), + "total": utils.item(total.data), + "nframes": torch.sum(sample["net_input"]["src_lengths"]).item(), + } + + return sample_size, logging_output + + def forward(self, model, sample, reduction="sum", log_probs=True): + """Computes the cross entropy with accuracy metric for the given sample. + + This is similar to CrossEntropyCriterion in fairseq, but also + computes accuracy metrics as part of logging + + Args: + logprobs (Torch.tensor) of shape N, T, D i.e. + batchsize, timesteps, dimensions + targets (Torch.tensor) of shape N, T i.e batchsize, timesteps + + Returns: + tuple: With three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + + TODO: + * Currently this Criterion will only work with LSTMEncoderModels or + FairseqModels which have decoder, or Models which return TorchTensor + as net_output. + We need to make a change to support all FairseqEncoder models. + """ + net_output = model(**sample["net_input"]) + target = model.get_targets(sample, net_output) + lprobs, loss = self.compute_loss( + model, net_output, target, reduction, log_probs + ) + sample_size, logging_output = self.get_logging_output( + sample, target, lprobs, loss + ) + return loss, sample_size, logging_output + + @staticmethod + def aggregate_logging_outputs(logging_outputs): + """Aggregate logging outputs from data parallel training.""" + correct_sum = sum(log.get("correct", 0) for log in logging_outputs) + total_sum = sum(log.get("total", 0) for log in logging_outputs) + loss_sum = sum(log.get("loss", 0) for log in logging_outputs) + ntokens = sum(log.get("ntokens", 0) for log in logging_outputs) + nsentences = sum(log.get("nsentences", 0) for log in logging_outputs) + sample_size = sum(log.get("sample_size", 0) for log in logging_outputs) + nframes = sum(log.get("nframes", 0) for log in logging_outputs) + agg_output = { + "loss": loss_sum / sample_size / math.log(2) if sample_size > 0 else 0.0, + # if args.sentence_avg, then sample_size is nsentences, then loss + # is per-sentence loss; else sample_size is ntokens, the loss + # becomes per-output token loss + "ntokens": ntokens, + "nsentences": nsentences, + "nframes": nframes, + "sample_size": sample_size, + "acc": correct_sum * 100.0 / total_sum if total_sum > 0 else 0.0, + "correct": correct_sum, + "total": total_sum, + # total is the number of validate tokens + } + if sample_size != ntokens: + agg_output["nll_loss"] = loss_sum / ntokens / math.log(2) + # loss: per output token loss + # nll_loss: per sentence loss + return agg_output diff --git a/examples/speech_recognition/data/__init__.py b/examples/speech_recognition/data/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..47bb6e24ddf25aa4fd5bf0fe9672f89099efb9b4 --- /dev/null +++ b/examples/speech_recognition/data/__init__.py @@ -0,0 +1,11 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .asr_dataset import AsrDataset + + +__all__ = [ + "AsrDataset", +] diff --git a/examples/speech_recognition/data/asr_dataset.py b/examples/speech_recognition/data/asr_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..63a6fcac85d73b1fce8e4d044b4209b1b67fa8ce --- /dev/null +++ b/examples/speech_recognition/data/asr_dataset.py @@ -0,0 +1,122 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import os + +import numpy as np +from fairseq.data import FairseqDataset + +from . import data_utils +from .collaters import Seq2SeqCollater + + +class AsrDataset(FairseqDataset): + """ + A dataset representing speech and corresponding transcription. + + Args: + aud_paths: (List[str]): A list of str with paths to audio files. + aud_durations_ms (List[int]): A list of int containing the durations of + audio files. + tgt (List[torch.LongTensor]): A list of LongTensors containing the indices + of target transcriptions. + tgt_dict (~fairseq.data.Dictionary): target vocabulary. + ids (List[str]): A list of utterance IDs. + speakers (List[str]): A list of speakers corresponding to utterances. + num_mel_bins (int): Number of triangular mel-frequency bins (default: 80) + frame_length (float): Frame length in milliseconds (default: 25.0) + frame_shift (float): Frame shift in milliseconds (default: 10.0) + """ + + def __init__( + self, + aud_paths, + aud_durations_ms, + tgt, + tgt_dict, + ids, + speakers, + num_mel_bins=80, + frame_length=25.0, + frame_shift=10.0, + ): + assert frame_length > 0 + assert frame_shift > 0 + assert all(x > frame_length for x in aud_durations_ms) + self.frame_sizes = [ + int(1 + (d - frame_length) / frame_shift) for d in aud_durations_ms + ] + + assert len(aud_paths) > 0 + assert len(aud_paths) == len(aud_durations_ms) + assert len(aud_paths) == len(tgt) + assert len(aud_paths) == len(ids) + assert len(aud_paths) == len(speakers) + self.aud_paths = aud_paths + self.tgt_dict = tgt_dict + self.tgt = tgt + self.ids = ids + self.speakers = speakers + self.num_mel_bins = num_mel_bins + self.frame_length = frame_length + self.frame_shift = frame_shift + + self.s2s_collater = Seq2SeqCollater( + 0, + 1, + pad_index=self.tgt_dict.pad(), + eos_index=self.tgt_dict.eos(), + move_eos_to_beginning=True, + ) + + def __getitem__(self, index): + import torchaudio + import torchaudio.compliance.kaldi as kaldi + + tgt_item = self.tgt[index] if self.tgt is not None else None + + path = self.aud_paths[index] + if not os.path.exists(path): + raise FileNotFoundError("Audio file not found: {}".format(path)) + sound, sample_rate = torchaudio.load_wav(path) + output = kaldi.fbank( + sound, + num_mel_bins=self.num_mel_bins, + frame_length=self.frame_length, + frame_shift=self.frame_shift, + ) + output_cmvn = data_utils.apply_mv_norm(output) + + return {"id": index, "data": [output_cmvn.detach(), tgt_item]} + + def __len__(self): + return len(self.aud_paths) + + def collater(self, samples): + """Merge a list of samples to form a mini-batch. + + Args: + samples (List[int]): sample indices to collate + + Returns: + dict: a mini-batch suitable for forwarding with a Model + """ + return self.s2s_collater.collate(samples) + + def num_tokens(self, index): + return self.frame_sizes[index] + + def size(self, index): + """Return an example's size as a float or tuple. This value is used when + filtering a dataset with ``--max-positions``.""" + return ( + self.frame_sizes[index], + len(self.tgt[index]) if self.tgt is not None else 0, + ) + + def ordered_indices(self): + """Return an ordered list of indices. Batches will be constructed based + on this order.""" + return np.arange(len(self)) diff --git a/examples/speech_recognition/data/collaters.py b/examples/speech_recognition/data/collaters.py new file mode 100644 index 0000000000000000000000000000000000000000..6acfec876b87e5a00bc92083b1181301a2a18e3f --- /dev/null +++ b/examples/speech_recognition/data/collaters.py @@ -0,0 +1,131 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" + This module contains collection of classes which implement + collate functionalities for various tasks. + + Collaters should know what data to expect for each sample + and they should pack / collate them into batches +""" + + +from __future__ import absolute_import, division, print_function, unicode_literals + +import numpy as np +import torch +from fairseq.data import data_utils as fairseq_data_utils + + +class Seq2SeqCollater(object): + """ + Implements collate function mainly for seq2seq tasks + This expects each sample to contain feature (src_tokens) and + targets. + This collator is also used for aligned training task. + """ + + def __init__( + self, + feature_index=0, + label_index=1, + pad_index=1, + eos_index=2, + move_eos_to_beginning=True, + ): + self.feature_index = feature_index + self.label_index = label_index + self.pad_index = pad_index + self.eos_index = eos_index + self.move_eos_to_beginning = move_eos_to_beginning + + def _collate_frames(self, frames): + """Convert a list of 2d frames into a padded 3d tensor + Args: + frames (list): list of 2d frames of size L[i]*f_dim. Where L[i] is + length of i-th frame and f_dim is static dimension of features + Returns: + 3d tensor of size len(frames)*len_max*f_dim where len_max is max of L[i] + """ + len_max = max(frame.size(0) for frame in frames) + f_dim = frames[0].size(1) + res = frames[0].new(len(frames), len_max, f_dim).fill_(0.0) + + for i, v in enumerate(frames): + res[i, : v.size(0)] = v + + return res + + def collate(self, samples): + """ + utility function to collate samples into batch for speech recognition. + """ + if len(samples) == 0: + return {} + + # parse samples into torch tensors + parsed_samples = [] + for s in samples: + # skip invalid samples + if s["data"][self.feature_index] is None: + continue + source = s["data"][self.feature_index] + if isinstance(source, (np.ndarray, np.generic)): + source = torch.from_numpy(source) + target = s["data"][self.label_index] + if isinstance(target, (np.ndarray, np.generic)): + target = torch.from_numpy(target).long() + elif isinstance(target, list): + target = torch.LongTensor(target) + + parsed_sample = {"id": s["id"], "source": source, "target": target} + parsed_samples.append(parsed_sample) + samples = parsed_samples + + id = torch.LongTensor([s["id"] for s in samples]) + frames = self._collate_frames([s["source"] for s in samples]) + # sort samples by descending number of frames + frames_lengths = torch.LongTensor([s["source"].size(0) for s in samples]) + frames_lengths, sort_order = frames_lengths.sort(descending=True) + id = id.index_select(0, sort_order) + frames = frames.index_select(0, sort_order) + + target = None + target_lengths = None + prev_output_tokens = None + if samples[0].get("target", None) is not None: + ntokens = sum(len(s["target"]) for s in samples) + target = fairseq_data_utils.collate_tokens( + [s["target"] for s in samples], + self.pad_index, + self.eos_index, + left_pad=False, + move_eos_to_beginning=False, + ) + target = target.index_select(0, sort_order) + target_lengths = torch.LongTensor( + [s["target"].size(0) for s in samples] + ).index_select(0, sort_order) + prev_output_tokens = fairseq_data_utils.collate_tokens( + [s["target"] for s in samples], + self.pad_index, + self.eos_index, + left_pad=False, + move_eos_to_beginning=self.move_eos_to_beginning, + ) + prev_output_tokens = prev_output_tokens.index_select(0, sort_order) + else: + ntokens = sum(len(s["source"]) for s in samples) + + batch = { + "id": id, + "ntokens": ntokens, + "net_input": {"src_tokens": frames, "src_lengths": frames_lengths}, + "target": target, + "target_lengths": target_lengths, + "nsentences": len(samples), + } + if prev_output_tokens is not None: + batch["net_input"]["prev_output_tokens"] = prev_output_tokens + return batch diff --git a/examples/speech_recognition/data/data_utils.py b/examples/speech_recognition/data/data_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..cc4729e63c8ef551b29617d1169a44c24f509ad0 --- /dev/null +++ b/examples/speech_recognition/data/data_utils.py @@ -0,0 +1,100 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch + + +def calc_mean_invstddev(feature): + if len(feature.size()) != 2: + raise ValueError("We expect the input feature to be 2-D tensor") + mean = feature.mean(0) + var = feature.var(0) + # avoid division by ~zero + eps = 1e-8 + if (var < eps).any(): + return mean, 1.0 / (torch.sqrt(var) + eps) + return mean, 1.0 / torch.sqrt(var) + + +def apply_mv_norm(features): + # If there is less than 2 spectrograms, the variance cannot be computed (is NaN) + # and normalization is not possible, so return the item as it is + if features.size(0) < 2: + return features + mean, invstddev = calc_mean_invstddev(features) + res = (features - mean) * invstddev + return res + + +def lengths_to_encoder_padding_mask(lengths, batch_first=False): + """ + convert lengths (a 1-D Long/Int tensor) to 2-D binary tensor + + Args: + lengths: a (B, )-shaped tensor + + Return: + max_length: maximum length of B sequences + encoder_padding_mask: a (max_length, B) binary mask, where + [t, b] = 0 for t < lengths[b] and 1 otherwise + + TODO: + kernelize this function if benchmarking shows this function is slow + """ + max_lengths = torch.max(lengths).item() + bsz = lengths.size(0) + encoder_padding_mask = torch.arange( + max_lengths + ).to( # a (T, ) tensor with [0, ..., T-1] + lengths.device + ).view( # move to the right device + 1, max_lengths + ).expand( # reshape to (1, T)-shaped tensor + bsz, -1 + ) >= lengths.view( # expand to (B, T)-shaped tensor + bsz, 1 + ).expand( + -1, max_lengths + ) + if not batch_first: + return encoder_padding_mask.t(), max_lengths + else: + return encoder_padding_mask, max_lengths + + +def encoder_padding_mask_to_lengths( + encoder_padding_mask, max_lengths, batch_size, device +): + """ + convert encoder_padding_mask (2-D binary tensor) to a 1-D tensor + + Conventionally, encoder output contains a encoder_padding_mask, which is + a 2-D mask in a shape (T, B), whose (t, b) element indicate whether + encoder_out[t, b] is a valid output (=0) or not (=1). Occasionally, we + need to convert this mask tensor to a 1-D tensor in shape (B, ), where + [b] denotes the valid length of b-th sequence + + Args: + encoder_padding_mask: a (T, B)-shaped binary tensor or None; if None, + indicating all are valid + Return: + seq_lengths: a (B,)-shaped tensor, where its (b, )-th element is the + number of valid elements of b-th sequence + + max_lengths: maximum length of all sequence, if encoder_padding_mask is + not None, max_lengths must equal to encoder_padding_mask.size(0) + + batch_size: batch size; if encoder_padding_mask is + not None, max_lengths must equal to encoder_padding_mask.size(1) + + device: which device to put the result on + """ + if encoder_padding_mask is None: + return torch.Tensor([max_lengths] * batch_size).to(torch.int32).to(device) + + assert encoder_padding_mask.size(0) == max_lengths, "max_lengths does not match" + assert encoder_padding_mask.size(1) == batch_size, "batch_size does not match" + + return max_lengths - torch.sum(encoder_padding_mask, dim=0) diff --git a/examples/speech_recognition/data/replabels.py b/examples/speech_recognition/data/replabels.py new file mode 100644 index 0000000000000000000000000000000000000000..441f1bd432b95865fc981c6c695cee299b07ed62 --- /dev/null +++ b/examples/speech_recognition/data/replabels.py @@ -0,0 +1,70 @@ +#!/usr/bin/env python3 + +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +""" +Replabel transforms for use with flashlight's ASG criterion. +""" + + +def replabel_symbol(i): + """ + Replabel symbols used in flashlight, currently just "1", "2", ... + This prevents training with numeral tokens, so this might change in the future + """ + return str(i) + + +def pack_replabels(tokens, dictionary, max_reps): + """ + Pack a token sequence so that repeated symbols are replaced by replabels + """ + if len(tokens) == 0 or max_reps <= 0: + return tokens + + replabel_value_to_idx = [0] * (max_reps + 1) + for i in range(1, max_reps + 1): + replabel_value_to_idx[i] = dictionary.index(replabel_symbol(i)) + + result = [] + prev_token = -1 + num_reps = 0 + for token in tokens: + if token == prev_token and num_reps < max_reps: + num_reps += 1 + else: + if num_reps > 0: + result.append(replabel_value_to_idx[num_reps]) + num_reps = 0 + result.append(token) + prev_token = token + if num_reps > 0: + result.append(replabel_value_to_idx[num_reps]) + return result + + +def unpack_replabels(tokens, dictionary, max_reps): + """ + Unpack a token sequence so that replabels are replaced by repeated symbols + """ + if len(tokens) == 0 or max_reps <= 0: + return tokens + + replabel_idx_to_value = {} + for i in range(1, max_reps + 1): + replabel_idx_to_value[dictionary.index(replabel_symbol(i))] = i + + result = [] + prev_token = -1 + for token in tokens: + try: + for _ in range(replabel_idx_to_value[token]): + result.append(prev_token) + prev_token = -1 + except KeyError: + result.append(token) + prev_token = token + return result diff --git a/examples/speech_recognition/datasets/asr_prep_json.py b/examples/speech_recognition/datasets/asr_prep_json.py new file mode 100644 index 0000000000000000000000000000000000000000..b8db8ff16691158fae034a8ab3faad622b351caf --- /dev/null +++ b/examples/speech_recognition/datasets/asr_prep_json.py @@ -0,0 +1,125 @@ +#!/usr/bin/env python3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from __future__ import absolute_import, division, print_function, unicode_literals + +import argparse +import concurrent.futures +import json +import multiprocessing +import os +from collections import namedtuple +from itertools import chain + +import sentencepiece as spm +from fairseq.data import Dictionary + + +MILLISECONDS_TO_SECONDS = 0.001 + + +def process_sample(aud_path, lable, utt_id, sp, tgt_dict): + import torchaudio + + input = {} + output = {} + si, ei = torchaudio.info(aud_path) + input["length_ms"] = int( + si.length / si.channels / si.rate / MILLISECONDS_TO_SECONDS + ) + input["path"] = aud_path + + token = " ".join(sp.EncodeAsPieces(lable)) + ids = tgt_dict.encode_line(token, append_eos=False) + output["text"] = lable + output["token"] = token + output["tokenid"] = ", ".join(map(str, [t.tolist() for t in ids])) + return {utt_id: {"input": input, "output": output}} + + +def main(): + parser = argparse.ArgumentParser() + parser.add_argument( + "--audio-dirs", + nargs="+", + default=["-"], + required=True, + help="input directories with audio files", + ) + parser.add_argument( + "--labels", + required=True, + help="aggregated input labels with format <ID LABEL> per line", + type=argparse.FileType("r", encoding="UTF-8"), + ) + parser.add_argument( + "--spm-model", + required=True, + help="sentencepiece model to use for encoding", + type=argparse.FileType("r", encoding="UTF-8"), + ) + parser.add_argument( + "--dictionary", + required=True, + help="file to load fairseq dictionary from", + type=argparse.FileType("r", encoding="UTF-8"), + ) + parser.add_argument("--audio-format", choices=["flac", "wav"], default="wav") + parser.add_argument( + "--output", + required=True, + type=argparse.FileType("w"), + help="path to save json output", + ) + args = parser.parse_args() + + sp = spm.SentencePieceProcessor() + sp.Load(args.spm_model.name) + + tgt_dict = Dictionary.load(args.dictionary) + + labels = {} + for line in args.labels: + (utt_id, label) = line.split(" ", 1) + labels[utt_id] = label + if len(labels) == 0: + raise Exception("No labels found in ", args.labels_path) + + Sample = namedtuple("Sample", "aud_path utt_id") + samples = [] + for path, _, files in chain.from_iterable( + os.walk(path) for path in args.audio_dirs + ): + for f in files: + if f.endswith(args.audio_format): + if len(os.path.splitext(f)) != 2: + raise Exception("Expect <utt_id.extension> file name. Got: ", f) + utt_id = os.path.splitext(f)[0] + if utt_id not in labels: + continue + samples.append(Sample(os.path.join(path, f), utt_id)) + + utts = {} + num_cpu = multiprocessing.cpu_count() + with concurrent.futures.ThreadPoolExecutor(max_workers=num_cpu) as executor: + future_to_sample = { + executor.submit( + process_sample, s.aud_path, labels[s.utt_id], s.utt_id, sp, tgt_dict + ): s + for s in samples + } + for future in concurrent.futures.as_completed(future_to_sample): + try: + data = future.result() + except Exception as exc: + print("generated an exception: ", exc) + else: + utts.update(data) + json.dump({"utts": utts}, args.output, indent=4) + + +if __name__ == "__main__": + main() diff --git a/examples/speech_recognition/datasets/prepare-librispeech.sh b/examples/speech_recognition/datasets/prepare-librispeech.sh new file mode 100755 index 0000000000000000000000000000000000000000..9e9297f08947027685ff508bfa91ff26b0d8ea0c --- /dev/null +++ b/examples/speech_recognition/datasets/prepare-librispeech.sh @@ -0,0 +1,88 @@ +#!/usr/bin/env bash +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +# Prepare librispeech dataset + +base_url=www.openslr.org/resources/12 +train_dir=train_960 + +if [ "$#" -ne 2 ]; then + echo "Usage: $0 <download_dir> <out_dir>" + echo "e.g.: $0 /tmp/librispeech_raw/ ~/data/librispeech_final" + exit 1 +fi + +download_dir=${1%/} +out_dir=${2%/} + +fairseq_root=~/fairseq-py/ +mkdir -p ${out_dir} +cd ${out_dir} || exit + +nbpe=5000 +bpemode=unigram + +if [ ! -d "$fairseq_root" ]; then + echo "$0: Please set correct fairseq_root" + exit 1 +fi + +echo "Data Download" +for part in dev-clean test-clean dev-other test-other train-clean-100 train-clean-360 train-other-500; do + url=$base_url/$part.tar.gz + if ! wget -P $download_dir $url; then + echo "$0: wget failed for $url" + exit 1 + fi + if ! tar -C $download_dir -xvzf $download_dir/$part.tar.gz; then + echo "$0: error un-tarring archive $download_dir/$part.tar.gz" + exit 1 + fi +done + +echo "Merge all train packs into one" +mkdir -p ${download_dir}/LibriSpeech/${train_dir}/ +for part in train-clean-100 train-clean-360 train-other-500; do + mv ${download_dir}/LibriSpeech/${part}/* $download_dir/LibriSpeech/${train_dir}/ +done +echo "Merge train text" +find ${download_dir}/LibriSpeech/${train_dir}/ -name '*.txt' -exec cat {} \; >> ${download_dir}/LibriSpeech/${train_dir}/text + +# Use combined dev-clean and dev-other as validation set +find ${download_dir}/LibriSpeech/dev-clean/ ${download_dir}/LibriSpeech/dev-other/ -name '*.txt' -exec cat {} \; >> ${download_dir}/LibriSpeech/valid_text +find ${download_dir}/LibriSpeech/test-clean/ -name '*.txt' -exec cat {} \; >> ${download_dir}/LibriSpeech/test-clean/text +find ${download_dir}/LibriSpeech/test-other/ -name '*.txt' -exec cat {} \; >> ${download_dir}/LibriSpeech/test-other/text + + +dict=data/lang_char/${train_dir}_${bpemode}${nbpe}_units.txt +encoded=data/lang_char/${train_dir}_${bpemode}${nbpe}_encoded.txt +fairseq_dict=data/lang_char/${train_dir}_${bpemode}${nbpe}_fairseq_dict.txt +bpemodel=data/lang_char/${train_dir}_${bpemode}${nbpe} +echo "dictionary: ${dict}" +echo "Dictionary preparation" +mkdir -p data/lang_char/ +echo "<unk> 3" > ${dict} +echo "</s> 2" >> ${dict} +echo "<pad> 1" >> ${dict} +cut -f 2- -d" " ${download_dir}/LibriSpeech/${train_dir}/text > data/lang_char/input.txt +spm_train --input=data/lang_char/input.txt --vocab_size=${nbpe} --model_type=${bpemode} --model_prefix=${bpemodel} --input_sentence_size=100000000 --unk_id=3 --eos_id=2 --pad_id=1 --bos_id=-1 --character_coverage=1 +spm_encode --model=${bpemodel}.model --output_format=piece < data/lang_char/input.txt > ${encoded} +cat ${encoded} | tr ' ' '\n' | sort | uniq | awk '{print $0 " " NR+3}' >> ${dict} +cat ${encoded} | tr ' ' '\n' | sort | uniq -c | awk '{print $2 " " $1}' > ${fairseq_dict} +wc -l ${dict} + +echo "Prepare train and test jsons" +for part in train_960 test-other test-clean; do + python ${fairseq_root}/examples/speech_recognition/datasets/asr_prep_json.py --audio-dirs ${download_dir}/LibriSpeech/${part} --labels ${download_dir}/LibriSpeech/${part}/text --spm-model ${bpemodel}.model --audio-format flac --dictionary ${fairseq_dict} --output ${part}.json +done +# fairseq expects to find train.json and valid.json during training +mv train_960.json train.json + +echo "Prepare valid json" +python ${fairseq_root}/examples/speech_recognition/datasets/asr_prep_json.py --audio-dirs ${download_dir}/LibriSpeech/dev-clean ${download_dir}/LibriSpeech/dev-other --labels ${download_dir}/LibriSpeech/valid_text --spm-model ${bpemodel}.model --audio-format flac --dictionary ${fairseq_dict} --output valid.json + +cp ${fairseq_dict} ./dict.txt +cp ${bpemodel}.model ./spm.model diff --git a/examples/speech_recognition/infer.py b/examples/speech_recognition/infer.py new file mode 100644 index 0000000000000000000000000000000000000000..6e9a878af46242ced57cfcd0e876a3d2ef3820ae --- /dev/null +++ b/examples/speech_recognition/infer.py @@ -0,0 +1,427 @@ +#!/usr/bin/env python3 -u +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +""" +Run inference for pre-processed data with a trained model. +""" + +import ast +import logging +import math +import os +import sys + +import editdistance +import numpy as np +import torch +from fairseq import checkpoint_utils, options, progress_bar, tasks, utils +from fairseq.data.data_utils import post_process +from fairseq.logging.meters import StopwatchMeter, TimeMeter + + +logging.basicConfig() +logging.root.setLevel(logging.INFO) +logging.basicConfig(level=logging.INFO) +logger = logging.getLogger(__name__) + + +def add_asr_eval_argument(parser): + parser.add_argument("--kspmodel", default=None, help="sentence piece model") + parser.add_argument( + "--wfstlm", default=None, help="wfstlm on dictonary output units" + ) + parser.add_argument( + "--rnnt_decoding_type", + default="greedy", + help="wfstlm on dictonary\ +output units", + ) + try: + parser.add_argument( + "--lm-weight", + "--lm_weight", + type=float, + default=0.2, + help="weight for lm while interpolating with neural score", + ) + except: + pass + parser.add_argument( + "--rnnt_len_penalty", default=-0.5, help="rnnt length penalty on word level" + ) + parser.add_argument( + "--w2l-decoder", + choices=["viterbi", "kenlm", "fairseqlm"], + help="use a w2l decoder", + ) + parser.add_argument("--lexicon", help="lexicon for w2l decoder") + parser.add_argument("--unit-lm", action="store_true", help="if using a unit lm") + parser.add_argument("--kenlm-model", "--lm-model", help="lm model for w2l decoder") + parser.add_argument("--beam-threshold", type=float, default=25.0) + parser.add_argument("--beam-size-token", type=float, default=100) + parser.add_argument("--word-score", type=float, default=1.0) + parser.add_argument("--unk-weight", type=float, default=-math.inf) + parser.add_argument("--sil-weight", type=float, default=0.0) + parser.add_argument( + "--dump-emissions", + type=str, + default=None, + help="if present, dumps emissions into this file and exits", + ) + parser.add_argument( + "--dump-features", + type=str, + default=None, + help="if present, dumps features into this file and exits", + ) + parser.add_argument( + "--load-emissions", + type=str, + default=None, + help="if present, loads emissions from this file", + ) + return parser + + +def check_args(args): + # assert args.path is not None, "--path required for generation!" + # assert args.results_path is not None, "--results_path required for generation!" + assert ( + not args.sampling or args.nbest == args.beam + ), "--sampling requires --nbest to be equal to --beam" + assert ( + args.replace_unk is None or args.raw_text + ), "--replace-unk requires a raw text dataset (--raw-text)" + + +def get_dataset_itr(args, task, models): + return task.get_batch_iterator( + dataset=task.dataset(args.gen_subset), + max_tokens=args.max_tokens, + max_sentences=args.batch_size, + max_positions=(sys.maxsize, sys.maxsize), + ignore_invalid_inputs=args.skip_invalid_size_inputs_valid_test, + required_batch_size_multiple=args.required_batch_size_multiple, + num_shards=args.num_shards, + shard_id=args.shard_id, + num_workers=args.num_workers, + data_buffer_size=args.data_buffer_size, + ).next_epoch_itr(shuffle=False) + + +def process_predictions( + args, hypos, sp, tgt_dict, target_tokens, res_files, speaker, id +): + for hypo in hypos[: min(len(hypos), args.nbest)]: + hyp_pieces = tgt_dict.string(hypo["tokens"].int().cpu()) + + if "words" in hypo: + hyp_words = " ".join(hypo["words"]) + else: + hyp_words = post_process(hyp_pieces, args.post_process) + + if res_files is not None: + print( + "{} ({}-{})".format(hyp_pieces, speaker, id), + file=res_files["hypo.units"], + ) + print( + "{} ({}-{})".format(hyp_words, speaker, id), + file=res_files["hypo.words"], + ) + + tgt_pieces = tgt_dict.string(target_tokens) + tgt_words = post_process(tgt_pieces, args.post_process) + + if res_files is not None: + print( + "{} ({}-{})".format(tgt_pieces, speaker, id), + file=res_files["ref.units"], + ) + print( + "{} ({}-{})".format(tgt_words, speaker, id), file=res_files["ref.words"] + ) + + if not args.quiet: + logger.info("HYPO:" + hyp_words) + logger.info("TARGET:" + tgt_words) + logger.info("___________________") + + hyp_words = hyp_words.split() + tgt_words = tgt_words.split() + return editdistance.eval(hyp_words, tgt_words), len(tgt_words) + + +def prepare_result_files(args): + def get_res_file(file_prefix): + if args.num_shards > 1: + file_prefix = f"{args.shard_id}_{file_prefix}" + path = os.path.join( + args.results_path, + "{}-{}-{}.txt".format( + file_prefix, os.path.basename(args.path), args.gen_subset + ), + ) + return open(path, "w", buffering=1) + + if not args.results_path: + return None + + return { + "hypo.words": get_res_file("hypo.word"), + "hypo.units": get_res_file("hypo.units"), + "ref.words": get_res_file("ref.word"), + "ref.units": get_res_file("ref.units"), + } + + +def optimize_models(args, use_cuda, models): + """Optimize ensemble for generation""" + for model in models: + model.make_generation_fast_( + beamable_mm_beam_size=None if args.no_beamable_mm else args.beam, + need_attn=args.print_alignment, + ) + if args.fp16: + model.half() + if use_cuda: + model.cuda() + + +class ExistingEmissionsDecoder(object): + def __init__(self, decoder, emissions): + self.decoder = decoder + self.emissions = emissions + + def generate(self, models, sample, **unused): + ids = sample["id"].cpu().numpy() + try: + emissions = np.stack(self.emissions[ids]) + except: + print([x.shape for x in self.emissions[ids]]) + raise Exception("invalid sizes") + emissions = torch.from_numpy(emissions) + return self.decoder.decode(emissions) + + +def main(args, task=None, model_state=None): + check_args(args) + + if args.max_tokens is None and args.batch_size is None: + args.max_tokens = 4000000 + logger.info(args) + + use_cuda = torch.cuda.is_available() and not args.cpu + + logger.info("| decoding with criterion {}".format(args.criterion)) + + task = tasks.setup_task(args) + + # Load ensemble + if args.load_emissions: + models, criterions = [], [] + task.load_dataset(args.gen_subset) + else: + logger.info("| loading model(s) from {}".format(args.path)) + models, saved_cfg, task = checkpoint_utils.load_model_ensemble_and_task( + utils.split_paths(args.path, separator="\\"), + arg_overrides=ast.literal_eval(args.model_overrides), + task=task, + suffix=args.checkpoint_suffix, + strict=(args.checkpoint_shard_count == 1), + num_shards=args.checkpoint_shard_count, + state=model_state, + ) + optimize_models(args, use_cuda, models) + task.load_dataset(args.gen_subset, task_cfg=saved_cfg.task) + + + # Set dictionary + tgt_dict = task.target_dictionary + + logger.info( + "| {} {} {} examples".format( + args.data, args.gen_subset, len(task.dataset(args.gen_subset)) + ) + ) + + # hack to pass transitions to W2lDecoder + if args.criterion == "asg_loss": + raise NotImplementedError("asg_loss is currently not supported") + # trans = criterions[0].asg.trans.data + # args.asg_transitions = torch.flatten(trans).tolist() + + # Load dataset (possibly sharded) + itr = get_dataset_itr(args, task, models) + + # Initialize generator + gen_timer = StopwatchMeter() + + def build_generator(args): + w2l_decoder = getattr(args, "w2l_decoder", None) + if w2l_decoder == "viterbi": + from examples.speech_recognition.w2l_decoder import W2lViterbiDecoder + + return W2lViterbiDecoder(args, task.target_dictionary) + elif w2l_decoder == "kenlm": + from examples.speech_recognition.w2l_decoder import W2lKenLMDecoder + + return W2lKenLMDecoder(args, task.target_dictionary) + elif w2l_decoder == "fairseqlm": + from examples.speech_recognition.w2l_decoder import W2lFairseqLMDecoder + + return W2lFairseqLMDecoder(args, task.target_dictionary) + else: + print( + "only flashlight decoders with (viterbi, kenlm, fairseqlm) options are supported at the moment" + ) + + # please do not touch this unless you test both generate.py and infer.py with audio_pretraining task + generator = build_generator(args) + + if args.load_emissions: + generator = ExistingEmissionsDecoder( + generator, np.load(args.load_emissions, allow_pickle=True) + ) + logger.info("loaded emissions from " + args.load_emissions) + + num_sentences = 0 + + if args.results_path is not None and not os.path.exists(args.results_path): + os.makedirs(args.results_path) + + max_source_pos = ( + utils.resolve_max_positions( + task.max_positions(), *[model.max_positions() for model in models] + ), + ) + + if max_source_pos is not None: + max_source_pos = max_source_pos[0] + if max_source_pos is not None: + max_source_pos = max_source_pos[0] - 1 + + if args.dump_emissions: + emissions = {} + if args.dump_features: + features = {} + models[0].bert.proj = None + else: + res_files = prepare_result_files(args) + errs_t = 0 + lengths_t = 0 + with progress_bar.build_progress_bar(args, itr) as t: + wps_meter = TimeMeter() + for sample in t: + sample = utils.move_to_cuda(sample) if use_cuda else sample + if "net_input" not in sample: + continue + + prefix_tokens = None + if args.prefix_size > 0: + prefix_tokens = sample["target"][:, : args.prefix_size] + + gen_timer.start() + if args.dump_emissions: + with torch.no_grad(): + encoder_out = models[0](**sample["net_input"]) + emm = models[0].get_normalized_probs(encoder_out, log_probs=True) + emm = emm.transpose(0, 1).cpu().numpy() + for i, id in enumerate(sample["id"]): + emissions[id.item()] = emm[i] + continue + elif args.dump_features: + with torch.no_grad(): + encoder_out = models[0](**sample["net_input"]) + feat = encoder_out["encoder_out"].transpose(0, 1).cpu().numpy() + for i, id in enumerate(sample["id"]): + padding = ( + encoder_out["encoder_padding_mask"][i].cpu().numpy() + if encoder_out["encoder_padding_mask"] is not None + else None + ) + features[id.item()] = (feat[i], padding) + continue + hypos = task.inference_step(generator, models, sample, prefix_tokens) + num_generated_tokens = sum(len(h[0]["tokens"]) for h in hypos) + gen_timer.stop(num_generated_tokens) + + for i, sample_id in enumerate(sample["id"].tolist()): + speaker = None + # id = task.dataset(args.gen_subset).ids[int(sample_id)] + id = sample_id + toks = ( + sample["target"][i, :] + if "target_label" not in sample + else sample["target_label"][i, :] + ) + target_tokens = utils.strip_pad(toks, tgt_dict.pad()).int().cpu() + # Process top predictions + errs, length = process_predictions( + args, + hypos[i], + None, + tgt_dict, + target_tokens, + res_files, + speaker, + id, + ) + errs_t += errs + lengths_t += length + + wps_meter.update(num_generated_tokens) + t.log({"wps": round(wps_meter.avg)}) + num_sentences += ( + sample["nsentences"] if "nsentences" in sample else sample["id"].numel() + ) + + wer = None + if args.dump_emissions: + emm_arr = [] + for i in range(len(emissions)): + emm_arr.append(emissions[i]) + np.save(args.dump_emissions, emm_arr) + logger.info(f"saved {len(emissions)} emissions to {args.dump_emissions}") + elif args.dump_features: + feat_arr = [] + for i in range(len(features)): + feat_arr.append(features[i]) + np.save(args.dump_features, feat_arr) + logger.info(f"saved {len(features)} emissions to {args.dump_features}") + else: + if lengths_t > 0: + wer = errs_t * 100.0 / lengths_t + logger.info(f"WER: {wer}") + + logger.info( + "| Processed {} sentences ({} tokens) in {:.1f}s ({:.2f}" + "sentences/s, {:.2f} tokens/s)".format( + num_sentences, + gen_timer.n, + gen_timer.sum, + num_sentences / gen_timer.sum, + 1.0 / gen_timer.avg, + ) + ) + logger.info("| Generate {} with beam={}".format(args.gen_subset, args.beam)) + return task, wer + + +def make_parser(): + parser = options.get_generation_parser() + parser = add_asr_eval_argument(parser) + return parser + + +def cli_main(): + parser = make_parser() + args = options.parse_args_and_arch(parser) + main(args) + + +if __name__ == "__main__": + cli_main() diff --git a/examples/speech_recognition/kaldi/__init__.py b/examples/speech_recognition/kaldi/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/examples/speech_recognition/kaldi/add-self-loop-simple.cc b/examples/speech_recognition/kaldi/add-self-loop-simple.cc new file mode 100644 index 0000000000000000000000000000000000000000..89754b925ea2b770e569b24d8ee07c408102733c --- /dev/null +++ b/examples/speech_recognition/kaldi/add-self-loop-simple.cc @@ -0,0 +1,94 @@ +/* +* Copyright (c) Facebook, Inc. and its affiliates. +* +* This source code is licensed under the MIT license found in the +* LICENSE file in the root directory of this source tree. +*/ + +#include <iostream> +#include "fstext/fstext-lib.h" // @manual +#include "util/common-utils.h" // @manual + +/* + * This program is to modify a FST without self-loop by: + * for each incoming arc with non-eps input symbol, add a self-loop arc + * with that non-eps symbol as input and eps as output. + * + * This is to make sure the resultant FST can do deduplication for repeated + * symbols, which is very common in acoustic model + * + */ +namespace { +int32 AddSelfLoopsSimple(fst::StdVectorFst* fst) { + typedef fst::MutableArcIterator<fst::StdVectorFst> IterType; + + int32 num_states_before = fst->NumStates(); + fst::MakePrecedingInputSymbolsSame(false, fst); + int32 num_states_after = fst->NumStates(); + KALDI_LOG << "There are " << num_states_before + << " states in the original FST; " + << " after MakePrecedingInputSymbolsSame, there are " + << num_states_after << " states " << std::endl; + + auto weight_one = fst::StdArc::Weight::One(); + + int32 num_arc_added = 0; + + fst::StdArc self_loop_arc; + self_loop_arc.weight = weight_one; + + int32 num_states = fst->NumStates(); + std::vector<std::set<int32>> incoming_non_eps_label_per_state(num_states); + + for (int32 state = 0; state < num_states; state++) { + for (IterType aiter(fst, state); !aiter.Done(); aiter.Next()) { + fst::StdArc arc(aiter.Value()); + if (arc.ilabel != 0) { + incoming_non_eps_label_per_state[arc.nextstate].insert(arc.ilabel); + } + } + } + + for (int32 state = 0; state < num_states; state++) { + if (!incoming_non_eps_label_per_state[state].empty()) { + auto& ilabel_set = incoming_non_eps_label_per_state[state]; + for (auto it = ilabel_set.begin(); it != ilabel_set.end(); it++) { + self_loop_arc.ilabel = *it; + self_loop_arc.olabel = 0; + self_loop_arc.nextstate = state; + fst->AddArc(state, self_loop_arc); + num_arc_added++; + } + } + } + return num_arc_added; +} + +void print_usage() { + std::cout << "add-self-loop-simple usage:\n" + "\tadd-self-loop-simple <in-fst> <out-fst> \n"; +} +} // namespace + +int main(int argc, char** argv) { + if (argc != 3) { + print_usage(); + exit(1); + } + + auto input = argv[1]; + auto output = argv[2]; + + auto fst = fst::ReadFstKaldi(input); + auto num_states = fst->NumStates(); + KALDI_LOG << "Loading FST from " << input << " with " << num_states + << " states." << std::endl; + + int32 num_arc_added = AddSelfLoopsSimple(fst); + KALDI_LOG << "Adding " << num_arc_added << " self-loop arcs " << std::endl; + + fst::WriteFstKaldi(*fst, std::string(output)); + KALDI_LOG << "Writing FST to " << output << std::endl; + + delete fst; +} \ No newline at end of file diff --git a/examples/speech_recognition/kaldi/config/kaldi_initializer.yaml b/examples/speech_recognition/kaldi/config/kaldi_initializer.yaml new file mode 100644 index 0000000000000000000000000000000000000000..be9ba98f55463d41d5d5ea35e306abc0886dbead --- /dev/null +++ b/examples/speech_recognition/kaldi/config/kaldi_initializer.yaml @@ -0,0 +1,8 @@ +# @package _group_ + +data_dir: ??? +fst_dir: ??? +in_labels: ??? +kaldi_root: ??? +lm_arpa: ??? +blank_symbol: <s> diff --git a/examples/speech_recognition/kaldi/kaldi_decoder.py b/examples/speech_recognition/kaldi/kaldi_decoder.py new file mode 100644 index 0000000000000000000000000000000000000000..5f62cc58ae8c0c5a3ba7d17713fedf0abc302942 --- /dev/null +++ b/examples/speech_recognition/kaldi/kaldi_decoder.py @@ -0,0 +1,244 @@ +#!/usr/bin/env python3 + +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from concurrent.futures import ThreadPoolExecutor +import logging +from omegaconf import MISSING +import os +import torch +from typing import Optional +import warnings + + +from dataclasses import dataclass +from fairseq.dataclass import FairseqDataclass +from .kaldi_initializer import KaldiInitializerConfig, initalize_kaldi + + +logger = logging.getLogger(__name__) + + +@dataclass +class KaldiDecoderConfig(FairseqDataclass): + hlg_graph_path: Optional[str] = None + output_dict: str = MISSING + + kaldi_initializer_config: Optional[KaldiInitializerConfig] = None + + acoustic_scale: float = 0.5 + max_active: int = 10000 + beam_delta: float = 0.5 + hash_ratio: float = 2.0 + + is_lattice: bool = False + lattice_beam: float = 10.0 + prune_interval: int = 25 + determinize_lattice: bool = True + prune_scale: float = 0.1 + max_mem: int = 0 + phone_determinize: bool = True + word_determinize: bool = True + minimize: bool = True + + num_threads: int = 1 + + +class KaldiDecoder(object): + def __init__( + self, + cfg: KaldiDecoderConfig, + beam: int, + nbest: int = 1, + ): + try: + from kaldi.asr import FasterRecognizer, LatticeFasterRecognizer + from kaldi.base import set_verbose_level + from kaldi.decoder import ( + FasterDecoder, + FasterDecoderOptions, + LatticeFasterDecoder, + LatticeFasterDecoderOptions, + ) + from kaldi.lat.functions import DeterminizeLatticePhonePrunedOptions + from kaldi.fstext import read_fst_kaldi, SymbolTable + except: + warnings.warn( + "pykaldi is required for this functionality. Please install from https://github.com/pykaldi/pykaldi" + ) + + # set_verbose_level(2) + + self.acoustic_scale = cfg.acoustic_scale + self.nbest = nbest + + if cfg.hlg_graph_path is None: + assert ( + cfg.kaldi_initializer_config is not None + ), "Must provide hlg graph path or kaldi initializer config" + cfg.hlg_graph_path = initalize_kaldi(cfg.kaldi_initializer_config) + + assert os.path.exists(cfg.hlg_graph_path), cfg.hlg_graph_path + + if cfg.is_lattice: + self.dec_cls = LatticeFasterDecoder + opt_cls = LatticeFasterDecoderOptions + self.rec_cls = LatticeFasterRecognizer + else: + assert self.nbest == 1, "nbest > 1 requires lattice decoder" + self.dec_cls = FasterDecoder + opt_cls = FasterDecoderOptions + self.rec_cls = FasterRecognizer + + self.decoder_options = opt_cls() + self.decoder_options.beam = beam + self.decoder_options.max_active = cfg.max_active + self.decoder_options.beam_delta = cfg.beam_delta + self.decoder_options.hash_ratio = cfg.hash_ratio + + if cfg.is_lattice: + self.decoder_options.lattice_beam = cfg.lattice_beam + self.decoder_options.prune_interval = cfg.prune_interval + self.decoder_options.determinize_lattice = cfg.determinize_lattice + self.decoder_options.prune_scale = cfg.prune_scale + det_opts = DeterminizeLatticePhonePrunedOptions() + det_opts.max_mem = cfg.max_mem + det_opts.phone_determinize = cfg.phone_determinize + det_opts.word_determinize = cfg.word_determinize + det_opts.minimize = cfg.minimize + self.decoder_options.det_opts = det_opts + + self.output_symbols = {} + with open(cfg.output_dict, "r") as f: + for line in f: + items = line.rstrip().split() + assert len(items) == 2 + self.output_symbols[int(items[1])] = items[0] + + logger.info(f"Loading FST from {cfg.hlg_graph_path}") + self.fst = read_fst_kaldi(cfg.hlg_graph_path) + self.symbol_table = SymbolTable.read_text(cfg.output_dict) + + self.executor = ThreadPoolExecutor(max_workers=cfg.num_threads) + + def generate(self, models, sample, **unused): + """Generate a batch of inferences.""" + # model.forward normally channels prev_output_tokens into the decoder + # separately, but SequenceGenerator directly calls model.encoder + encoder_input = { + k: v for k, v in sample["net_input"].items() if k != "prev_output_tokens" + } + emissions, padding = self.get_emissions(models, encoder_input) + return self.decode(emissions, padding) + + def get_emissions(self, models, encoder_input): + """Run encoder and normalize emissions""" + model = models[0] + + all_encoder_out = [m(**encoder_input) for m in models] + + if len(all_encoder_out) > 1: + + if "encoder_out" in all_encoder_out[0]: + encoder_out = { + "encoder_out": sum(e["encoder_out"] for e in all_encoder_out) + / len(all_encoder_out), + "encoder_padding_mask": all_encoder_out[0]["encoder_padding_mask"], + } + padding = encoder_out["encoder_padding_mask"] + else: + encoder_out = { + "logits": sum(e["logits"] for e in all_encoder_out) + / len(all_encoder_out), + "padding_mask": all_encoder_out[0]["padding_mask"], + } + padding = encoder_out["padding_mask"] + else: + encoder_out = all_encoder_out[0] + padding = ( + encoder_out["padding_mask"] + if "padding_mask" in encoder_out + else encoder_out["encoder_padding_mask"] + ) + + if hasattr(model, "get_logits"): + emissions = model.get_logits(encoder_out, normalize=True) + else: + emissions = model.get_normalized_probs(encoder_out, log_probs=True) + + return ( + emissions.cpu().float().transpose(0, 1), + padding.cpu() if padding is not None and padding.any() else None, + ) + + def decode_one(self, logits, padding): + from kaldi.matrix import Matrix + + decoder = self.dec_cls(self.fst, self.decoder_options) + asr = self.rec_cls( + decoder, self.symbol_table, acoustic_scale=self.acoustic_scale + ) + + if padding is not None: + logits = logits[~padding] + + mat = Matrix(logits.numpy()) + + out = asr.decode(mat) + + if self.nbest > 1: + from kaldi.fstext import shortestpath + from kaldi.fstext.utils import ( + convert_compact_lattice_to_lattice, + convert_lattice_to_std, + convert_nbest_to_list, + get_linear_symbol_sequence, + ) + + lat = out["lattice"] + + sp = shortestpath(lat, nshortest=self.nbest) + + sp = convert_compact_lattice_to_lattice(sp) + sp = convert_lattice_to_std(sp) + seq = convert_nbest_to_list(sp) + + results = [] + for s in seq: + _, o, w = get_linear_symbol_sequence(s) + words = list(self.output_symbols[z] for z in o) + results.append( + { + "tokens": words, + "words": words, + "score": w.value, + "emissions": logits, + } + ) + return results + else: + words = out["text"].split() + return [ + { + "tokens": words, + "words": words, + "score": out["likelihood"], + "emissions": logits, + } + ] + + def decode(self, emissions, padding): + if padding is None: + padding = [None] * len(emissions) + + ret = list( + map( + lambda e, p: self.executor.submit(self.decode_one, e, p), + emissions, + padding, + ) + ) + return ret diff --git a/examples/speech_recognition/kaldi/kaldi_initializer.py b/examples/speech_recognition/kaldi/kaldi_initializer.py new file mode 100644 index 0000000000000000000000000000000000000000..6d2a2a4b6b809ba1106f9a57cb6f241dc083e670 --- /dev/null +++ b/examples/speech_recognition/kaldi/kaldi_initializer.py @@ -0,0 +1,698 @@ +#!/usr/bin/env python3 + +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from dataclasses import dataclass +import hydra +from hydra.core.config_store import ConfigStore +import logging +from omegaconf import MISSING, OmegaConf +import os +import os.path as osp +from pathlib import Path +import subprocess +from typing import Optional + +from fairseq.data.dictionary import Dictionary +from fairseq.dataclass import FairseqDataclass + +script_dir = Path(__file__).resolve().parent +config_path = script_dir / "config" + + +logger = logging.getLogger(__name__) + + +@dataclass +class KaldiInitializerConfig(FairseqDataclass): + data_dir: str = MISSING + fst_dir: Optional[str] = None + in_labels: str = MISSING + out_labels: Optional[str] = None + wav2letter_lexicon: Optional[str] = None + lm_arpa: str = MISSING + kaldi_root: str = MISSING + blank_symbol: str = "<s>" + silence_symbol: Optional[str] = None + + +def create_units(fst_dir: Path, in_labels: str, vocab: Dictionary) -> Path: + in_units_file = fst_dir / f"kaldi_dict.{in_labels}.txt" + if not in_units_file.exists(): + + logger.info(f"Creating {in_units_file}") + + with open(in_units_file, "w") as f: + print("<eps> 0", file=f) + i = 1 + for symb in vocab.symbols[vocab.nspecial :]: + if not symb.startswith("madeupword"): + print(f"{symb} {i}", file=f) + i += 1 + return in_units_file + + +def create_lexicon( + cfg: KaldiInitializerConfig, + fst_dir: Path, + unique_label: str, + in_units_file: Path, + out_words_file: Path, +) -> (Path, Path): + + disambig_in_units_file = fst_dir / f"kaldi_dict.{cfg.in_labels}_disambig.txt" + lexicon_file = fst_dir / f"kaldi_lexicon.{unique_label}.txt" + disambig_lexicon_file = fst_dir / f"kaldi_lexicon.{unique_label}_disambig.txt" + if ( + not lexicon_file.exists() + or not disambig_lexicon_file.exists() + or not disambig_in_units_file.exists() + ): + logger.info(f"Creating {lexicon_file} (in units file: {in_units_file})") + + assert cfg.wav2letter_lexicon is not None or cfg.in_labels == cfg.out_labels + + if cfg.wav2letter_lexicon is not None: + lm_words = set() + with open(out_words_file, "r") as lm_dict_f: + for line in lm_dict_f: + lm_words.add(line.split()[0]) + + num_skipped = 0 + total = 0 + with open(cfg.wav2letter_lexicon, "r") as w2l_lex_f, open( + lexicon_file, "w" + ) as out_f: + for line in w2l_lex_f: + items = line.rstrip().split("\t") + assert len(items) == 2, items + if items[0] in lm_words: + print(items[0], items[1], file=out_f) + else: + num_skipped += 1 + logger.debug( + f"Skipping word {items[0]} as it was not found in LM" + ) + total += 1 + if num_skipped > 0: + logger.warning( + f"Skipped {num_skipped} out of {total} words as they were not found in LM" + ) + else: + with open(in_units_file, "r") as in_f, open(lexicon_file, "w") as out_f: + for line in in_f: + symb = line.split()[0] + if symb != "<eps>" and symb != "<ctc_blank>" and symb != "<SIL>": + print(symb, symb, file=out_f) + + lex_disambig_path = ( + Path(cfg.kaldi_root) / "egs/wsj/s5/utils/add_lex_disambig.pl" + ) + res = subprocess.run( + [lex_disambig_path, lexicon_file, disambig_lexicon_file], + check=True, + capture_output=True, + ) + ndisambig = int(res.stdout) + disamib_path = Path(cfg.kaldi_root) / "egs/wsj/s5/utils/add_disambig.pl" + res = subprocess.run( + [disamib_path, "--include-zero", in_units_file, str(ndisambig)], + check=True, + capture_output=True, + ) + with open(disambig_in_units_file, "wb") as f: + f.write(res.stdout) + + return disambig_lexicon_file, disambig_in_units_file + + +def create_G( + kaldi_root: Path, fst_dir: Path, lm_arpa: Path, arpa_base: str +) -> (Path, Path): + + out_words_file = fst_dir / f"kaldi_dict.{arpa_base}.txt" + grammar_graph = fst_dir / f"G_{arpa_base}.fst" + if not grammar_graph.exists() or not out_words_file.exists(): + logger.info(f"Creating {grammar_graph}") + arpa2fst = kaldi_root / "src/lmbin/arpa2fst" + subprocess.run( + [ + arpa2fst, + "--disambig-symbol=#0", + f"--write-symbol-table={out_words_file}", + lm_arpa, + grammar_graph, + ], + check=True, + ) + return grammar_graph, out_words_file + + +def create_L( + kaldi_root: Path, + fst_dir: Path, + unique_label: str, + lexicon_file: Path, + in_units_file: Path, + out_words_file: Path, +) -> Path: + lexicon_graph = fst_dir / f"L.{unique_label}.fst" + + if not lexicon_graph.exists(): + logger.info(f"Creating {lexicon_graph} (in units: {in_units_file})") + make_lex = kaldi_root / "egs/wsj/s5/utils/make_lexicon_fst.pl" + fstcompile = kaldi_root / "tools/openfst-1.6.7/bin/fstcompile" + fstaddselfloops = kaldi_root / "src/fstbin/fstaddselfloops" + fstarcsort = kaldi_root / "tools/openfst-1.6.7/bin/fstarcsort" + + def write_disambig_symbol(file): + with open(file, "r") as f: + for line in f: + items = line.rstrip().split() + if items[0] == "#0": + out_path = str(file) + "_disamig" + with open(out_path, "w") as out_f: + print(items[1], file=out_f) + return out_path + + return None + + in_disambig_sym = write_disambig_symbol(in_units_file) + assert in_disambig_sym is not None + out_disambig_sym = write_disambig_symbol(out_words_file) + assert out_disambig_sym is not None + + try: + with open(lexicon_graph, "wb") as out_f: + res = subprocess.run( + [make_lex, lexicon_file], capture_output=True, check=True + ) + assert len(res.stderr) == 0, res.stderr.decode("utf-8") + res = subprocess.run( + [ + fstcompile, + f"--isymbols={in_units_file}", + f"--osymbols={out_words_file}", + "--keep_isymbols=false", + "--keep_osymbols=false", + ], + input=res.stdout, + capture_output=True, + ) + assert len(res.stderr) == 0, res.stderr.decode("utf-8") + res = subprocess.run( + [fstaddselfloops, in_disambig_sym, out_disambig_sym], + input=res.stdout, + capture_output=True, + check=True, + ) + res = subprocess.run( + [fstarcsort, "--sort_type=olabel"], + input=res.stdout, + capture_output=True, + check=True, + ) + out_f.write(res.stdout) + except subprocess.CalledProcessError as e: + logger.error(f"cmd: {e.cmd}, err: {e.stderr.decode('utf-8')}") + os.remove(lexicon_graph) + raise + except AssertionError: + os.remove(lexicon_graph) + raise + + return lexicon_graph + + +def create_LG( + kaldi_root: Path, + fst_dir: Path, + unique_label: str, + lexicon_graph: Path, + grammar_graph: Path, +) -> Path: + lg_graph = fst_dir / f"LG.{unique_label}.fst" + + if not lg_graph.exists(): + logger.info(f"Creating {lg_graph}") + + fsttablecompose = kaldi_root / "src/fstbin/fsttablecompose" + fstdeterminizestar = kaldi_root / "src/fstbin/fstdeterminizestar" + fstminimizeencoded = kaldi_root / "src/fstbin/fstminimizeencoded" + fstpushspecial = kaldi_root / "src/fstbin/fstpushspecial" + fstarcsort = kaldi_root / "tools/openfst-1.6.7/bin/fstarcsort" + + try: + with open(lg_graph, "wb") as out_f: + res = subprocess.run( + [fsttablecompose, lexicon_graph, grammar_graph], + capture_output=True, + check=True, + ) + res = subprocess.run( + [ + fstdeterminizestar, + "--use-log=true", + ], + input=res.stdout, + capture_output=True, + ) + res = subprocess.run( + [fstminimizeencoded], + input=res.stdout, + capture_output=True, + check=True, + ) + res = subprocess.run( + [fstpushspecial], + input=res.stdout, + capture_output=True, + check=True, + ) + res = subprocess.run( + [fstarcsort, "--sort_type=ilabel"], + input=res.stdout, + capture_output=True, + check=True, + ) + out_f.write(res.stdout) + except subprocess.CalledProcessError as e: + logger.error(f"cmd: {e.cmd}, err: {e.stderr.decode('utf-8')}") + os.remove(lg_graph) + raise + + return lg_graph + + +def create_H( + kaldi_root: Path, + fst_dir: Path, + disambig_out_units_file: Path, + in_labels: str, + vocab: Dictionary, + blk_sym: str, + silence_symbol: Optional[str], +) -> (Path, Path, Path): + h_graph = ( + fst_dir / f"H.{in_labels}{'_' + silence_symbol if silence_symbol else ''}.fst" + ) + h_out_units_file = fst_dir / f"kaldi_dict.h_out.{in_labels}.txt" + disambig_in_units_file_int = Path(str(h_graph) + "isym_disambig.int") + disambig_out_units_file_int = Path(str(disambig_out_units_file) + ".int") + if ( + not h_graph.exists() + or not h_out_units_file.exists() + or not disambig_in_units_file_int.exists() + ): + logger.info(f"Creating {h_graph}") + eps_sym = "<eps>" + + num_disambig = 0 + osymbols = [] + + with open(disambig_out_units_file, "r") as f, open( + disambig_out_units_file_int, "w" + ) as out_f: + for line in f: + symb, id = line.rstrip().split() + if line.startswith("#"): + num_disambig += 1 + print(id, file=out_f) + else: + if len(osymbols) == 0: + assert symb == eps_sym, symb + osymbols.append((symb, id)) + + i_idx = 0 + isymbols = [(eps_sym, 0)] + + imap = {} + + for i, s in enumerate(vocab.symbols): + i_idx += 1 + isymbols.append((s, i_idx)) + imap[s] = i_idx + + fst_str = [] + + node_idx = 0 + root_node = node_idx + + special_symbols = [blk_sym] + if silence_symbol is not None: + special_symbols.append(silence_symbol) + + for ss in special_symbols: + fst_str.append("{} {} {} {}".format(root_node, root_node, ss, eps_sym)) + + for symbol, _ in osymbols: + if symbol == eps_sym or symbol.startswith("#"): + continue + + node_idx += 1 + # 1. from root to emitting state + fst_str.append("{} {} {} {}".format(root_node, node_idx, symbol, symbol)) + # 2. from emitting state back to root + fst_str.append("{} {} {} {}".format(node_idx, root_node, eps_sym, eps_sym)) + # 3. from emitting state to optional blank state + pre_node = node_idx + node_idx += 1 + for ss in special_symbols: + fst_str.append("{} {} {} {}".format(pre_node, node_idx, ss, eps_sym)) + # 4. from blank state back to root + fst_str.append("{} {} {} {}".format(node_idx, root_node, eps_sym, eps_sym)) + + fst_str.append("{}".format(root_node)) + + fst_str = "\n".join(fst_str) + h_str = str(h_graph) + isym_file = h_str + ".isym" + + with open(isym_file, "w") as f: + for sym, id in isymbols: + f.write("{} {}\n".format(sym, id)) + + with open(h_out_units_file, "w") as f: + for sym, id in osymbols: + f.write("{} {}\n".format(sym, id)) + + with open(disambig_in_units_file_int, "w") as f: + disam_sym_id = len(isymbols) + for _ in range(num_disambig): + f.write("{}\n".format(disam_sym_id)) + disam_sym_id += 1 + + fstcompile = kaldi_root / "tools/openfst-1.6.7/bin/fstcompile" + fstaddselfloops = kaldi_root / "src/fstbin/fstaddselfloops" + fstarcsort = kaldi_root / "tools/openfst-1.6.7/bin/fstarcsort" + + try: + with open(h_graph, "wb") as out_f: + res = subprocess.run( + [ + fstcompile, + f"--isymbols={isym_file}", + f"--osymbols={h_out_units_file}", + "--keep_isymbols=false", + "--keep_osymbols=false", + ], + input=str.encode(fst_str), + capture_output=True, + check=True, + ) + res = subprocess.run( + [ + fstaddselfloops, + disambig_in_units_file_int, + disambig_out_units_file_int, + ], + input=res.stdout, + capture_output=True, + check=True, + ) + res = subprocess.run( + [fstarcsort, "--sort_type=olabel"], + input=res.stdout, + capture_output=True, + check=True, + ) + out_f.write(res.stdout) + except subprocess.CalledProcessError as e: + logger.error(f"cmd: {e.cmd}, err: {e.stderr.decode('utf-8')}") + os.remove(h_graph) + raise + return h_graph, h_out_units_file, disambig_in_units_file_int + + +def create_HLGa( + kaldi_root: Path, + fst_dir: Path, + unique_label: str, + h_graph: Path, + lg_graph: Path, + disambig_in_words_file_int: Path, +) -> Path: + hlga_graph = fst_dir / f"HLGa.{unique_label}.fst" + + if not hlga_graph.exists(): + logger.info(f"Creating {hlga_graph}") + + fsttablecompose = kaldi_root / "src/fstbin/fsttablecompose" + fstdeterminizestar = kaldi_root / "src/fstbin/fstdeterminizestar" + fstrmsymbols = kaldi_root / "src/fstbin/fstrmsymbols" + fstrmepslocal = kaldi_root / "src/fstbin/fstrmepslocal" + fstminimizeencoded = kaldi_root / "src/fstbin/fstminimizeencoded" + + try: + with open(hlga_graph, "wb") as out_f: + res = subprocess.run( + [ + fsttablecompose, + h_graph, + lg_graph, + ], + capture_output=True, + check=True, + ) + res = subprocess.run( + [fstdeterminizestar, "--use-log=true"], + input=res.stdout, + capture_output=True, + check=True, + ) + res = subprocess.run( + [fstrmsymbols, disambig_in_words_file_int], + input=res.stdout, + capture_output=True, + check=True, + ) + res = subprocess.run( + [fstrmepslocal], + input=res.stdout, + capture_output=True, + check=True, + ) + res = subprocess.run( + [fstminimizeencoded], + input=res.stdout, + capture_output=True, + check=True, + ) + out_f.write(res.stdout) + except subprocess.CalledProcessError as e: + logger.error(f"cmd: {e.cmd}, err: {e.stderr.decode('utf-8')}") + os.remove(hlga_graph) + raise + + return hlga_graph + + +def create_HLa( + kaldi_root: Path, + fst_dir: Path, + unique_label: str, + h_graph: Path, + l_graph: Path, + disambig_in_words_file_int: Path, +) -> Path: + hla_graph = fst_dir / f"HLa.{unique_label}.fst" + + if not hla_graph.exists(): + logger.info(f"Creating {hla_graph}") + + fsttablecompose = kaldi_root / "src/fstbin/fsttablecompose" + fstdeterminizestar = kaldi_root / "src/fstbin/fstdeterminizestar" + fstrmsymbols = kaldi_root / "src/fstbin/fstrmsymbols" + fstrmepslocal = kaldi_root / "src/fstbin/fstrmepslocal" + fstminimizeencoded = kaldi_root / "src/fstbin/fstminimizeencoded" + + try: + with open(hla_graph, "wb") as out_f: + res = subprocess.run( + [ + fsttablecompose, + h_graph, + l_graph, + ], + capture_output=True, + check=True, + ) + res = subprocess.run( + [fstdeterminizestar, "--use-log=true"], + input=res.stdout, + capture_output=True, + check=True, + ) + res = subprocess.run( + [fstrmsymbols, disambig_in_words_file_int], + input=res.stdout, + capture_output=True, + check=True, + ) + res = subprocess.run( + [fstrmepslocal], + input=res.stdout, + capture_output=True, + check=True, + ) + res = subprocess.run( + [fstminimizeencoded], + input=res.stdout, + capture_output=True, + check=True, + ) + out_f.write(res.stdout) + except subprocess.CalledProcessError as e: + logger.error(f"cmd: {e.cmd}, err: {e.stderr.decode('utf-8')}") + os.remove(hla_graph) + raise + + return hla_graph + + +def create_HLG( + kaldi_root: Path, + fst_dir: Path, + unique_label: str, + hlga_graph: Path, + prefix: str = "HLG", +) -> Path: + hlg_graph = fst_dir / f"{prefix}.{unique_label}.fst" + + if not hlg_graph.exists(): + logger.info(f"Creating {hlg_graph}") + + add_self_loop = script_dir / "add-self-loop-simple" + kaldi_src = kaldi_root / "src" + kaldi_lib = kaldi_src / "lib" + + try: + if not add_self_loop.exists(): + fst_include = kaldi_root / "tools/openfst-1.6.7/include" + add_self_loop_src = script_dir / "add-self-loop-simple.cc" + + subprocess.run( + [ + "c++", + f"-I{kaldi_src}", + f"-I{fst_include}", + f"-L{kaldi_lib}", + add_self_loop_src, + "-lkaldi-base", + "-lkaldi-fstext", + "-o", + add_self_loop, + ], + check=True, + ) + + my_env = os.environ.copy() + my_env["LD_LIBRARY_PATH"] = f"{kaldi_lib}:{my_env['LD_LIBRARY_PATH']}" + + subprocess.run( + [ + add_self_loop, + hlga_graph, + hlg_graph, + ], + check=True, + capture_output=True, + env=my_env, + ) + except subprocess.CalledProcessError as e: + logger.error(f"cmd: {e.cmd}, err: {e.stderr.decode('utf-8')}") + raise + + return hlg_graph + + +def initalize_kaldi(cfg: KaldiInitializerConfig) -> Path: + if cfg.fst_dir is None: + cfg.fst_dir = osp.join(cfg.data_dir, "kaldi") + if cfg.out_labels is None: + cfg.out_labels = cfg.in_labels + + kaldi_root = Path(cfg.kaldi_root) + data_dir = Path(cfg.data_dir) + fst_dir = Path(cfg.fst_dir) + fst_dir.mkdir(parents=True, exist_ok=True) + + arpa_base = osp.splitext(osp.basename(cfg.lm_arpa))[0] + unique_label = f"{cfg.in_labels}.{arpa_base}" + + with open(data_dir / f"dict.{cfg.in_labels}.txt", "r") as f: + vocab = Dictionary.load(f) + + in_units_file = create_units(fst_dir, cfg.in_labels, vocab) + + grammar_graph, out_words_file = create_G( + kaldi_root, fst_dir, Path(cfg.lm_arpa), arpa_base + ) + + disambig_lexicon_file, disambig_L_in_units_file = create_lexicon( + cfg, fst_dir, unique_label, in_units_file, out_words_file + ) + + h_graph, h_out_units_file, disambig_in_units_file_int = create_H( + kaldi_root, + fst_dir, + disambig_L_in_units_file, + cfg.in_labels, + vocab, + cfg.blank_symbol, + cfg.silence_symbol, + ) + lexicon_graph = create_L( + kaldi_root, + fst_dir, + unique_label, + disambig_lexicon_file, + disambig_L_in_units_file, + out_words_file, + ) + lg_graph = create_LG( + kaldi_root, fst_dir, unique_label, lexicon_graph, grammar_graph + ) + hlga_graph = create_HLGa( + kaldi_root, fst_dir, unique_label, h_graph, lg_graph, disambig_in_units_file_int + ) + hlg_graph = create_HLG(kaldi_root, fst_dir, unique_label, hlga_graph) + + # for debugging + # hla_graph = create_HLa(kaldi_root, fst_dir, unique_label, h_graph, lexicon_graph, disambig_in_units_file_int) + # hl_graph = create_HLG(kaldi_root, fst_dir, unique_label, hla_graph, prefix="HL_looped") + # create_HLG(kaldi_root, fst_dir, "phnc", h_graph, prefix="H_looped") + + return hlg_graph + + +@hydra.main(config_path=config_path, config_name="kaldi_initializer") +def cli_main(cfg: KaldiInitializerConfig) -> None: + container = OmegaConf.to_container(cfg, resolve=True, enum_to_str=True) + cfg = OmegaConf.create(container) + OmegaConf.set_struct(cfg, True) + initalize_kaldi(cfg) + + +if __name__ == "__main__": + + logging.root.setLevel(logging.INFO) + logging.basicConfig(level=logging.INFO) + + try: + from hydra._internal.utils import ( + get_args, + ) # pylint: disable=import-outside-toplevel + + cfg_name = get_args().config_name or "kaldi_initializer" + except ImportError: + logger.warning("Failed to get config name from hydra args") + cfg_name = "kaldi_initializer" + + cs = ConfigStore.instance() + cs.store(name=cfg_name, node=KaldiInitializerConfig) + + cli_main() diff --git a/examples/speech_recognition/models/__init__.py b/examples/speech_recognition/models/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..54b5a1c31243e55d384f80ef9514461cd35b15c6 --- /dev/null +++ b/examples/speech_recognition/models/__init__.py @@ -0,0 +1,8 @@ +import importlib +import os + + +for file in sorted(os.listdir(os.path.dirname(__file__))): + if file.endswith(".py") and not file.startswith("_"): + model_name = file[: file.find(".py")] + importlib.import_module("examples.speech_recognition.models." + model_name) diff --git a/examples/speech_recognition/models/vggtransformer.py b/examples/speech_recognition/models/vggtransformer.py new file mode 100644 index 0000000000000000000000000000000000000000..97974360a454b581eb63bdfd2af2e2afa05596c7 --- /dev/null +++ b/examples/speech_recognition/models/vggtransformer.py @@ -0,0 +1,1019 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import math +from collections.abc import Iterable + +import torch +import torch.nn as nn +from examples.speech_recognition.data.data_utils import lengths_to_encoder_padding_mask +from fairseq import utils +from fairseq.models import ( + FairseqEncoder, + FairseqEncoderDecoderModel, + FairseqEncoderModel, + FairseqIncrementalDecoder, + register_model, + register_model_architecture, +) +from fairseq.modules import ( + LinearizedConvolution, + TransformerDecoderLayer, + TransformerEncoderLayer, + VGGBlock, +) + + +@register_model("asr_vggtransformer") +class VGGTransformerModel(FairseqEncoderDecoderModel): + """ + Transformers with convolutional context for ASR + https://arxiv.org/abs/1904.11660 + """ + + def __init__(self, encoder, decoder): + super().__init__(encoder, decoder) + + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + parser.add_argument( + "--input-feat-per-channel", + type=int, + metavar="N", + help="encoder input dimension per input channel", + ) + parser.add_argument( + "--vggblock-enc-config", + type=str, + metavar="EXPR", + help=""" + an array of tuples each containing the configuration of one vggblock: + [(out_channels, + conv_kernel_size, + pooling_kernel_size, + num_conv_layers, + use_layer_norm), ...]) + """, + ) + parser.add_argument( + "--transformer-enc-config", + type=str, + metavar="EXPR", + help="""" + a tuple containing the configuration of the encoder transformer layers + configurations: + [(input_dim, + num_heads, + ffn_dim, + normalize_before, + dropout, + attention_dropout, + relu_dropout), ...]') + """, + ) + parser.add_argument( + "--enc-output-dim", + type=int, + metavar="N", + help=""" + encoder output dimension, can be None. If specified, projecting the + transformer output to the specified dimension""", + ) + parser.add_argument( + "--in-channels", + type=int, + metavar="N", + help="number of encoder input channels", + ) + parser.add_argument( + "--tgt-embed-dim", + type=int, + metavar="N", + help="embedding dimension of the decoder target tokens", + ) + parser.add_argument( + "--transformer-dec-config", + type=str, + metavar="EXPR", + help=""" + a tuple containing the configuration of the decoder transformer layers + configurations: + [(input_dim, + num_heads, + ffn_dim, + normalize_before, + dropout, + attention_dropout, + relu_dropout), ...] + """, + ) + parser.add_argument( + "--conv-dec-config", + type=str, + metavar="EXPR", + help=""" + an array of tuples for the decoder 1-D convolution config + [(out_channels, conv_kernel_size, use_layer_norm), ...]""", + ) + + @classmethod + def build_encoder(cls, args, task): + return VGGTransformerEncoder( + input_feat_per_channel=args.input_feat_per_channel, + vggblock_config=eval(args.vggblock_enc_config), + transformer_config=eval(args.transformer_enc_config), + encoder_output_dim=args.enc_output_dim, + in_channels=args.in_channels, + ) + + @classmethod + def build_decoder(cls, args, task): + return TransformerDecoder( + dictionary=task.target_dictionary, + embed_dim=args.tgt_embed_dim, + transformer_config=eval(args.transformer_dec_config), + conv_config=eval(args.conv_dec_config), + encoder_output_dim=args.enc_output_dim, + ) + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + # make sure that all args are properly defaulted + # (in case there are any new ones) + base_architecture(args) + + encoder = cls.build_encoder(args, task) + decoder = cls.build_decoder(args, task) + return cls(encoder, decoder) + + def get_normalized_probs(self, net_output, log_probs, sample=None): + # net_output['encoder_out'] is a (B, T, D) tensor + lprobs = super().get_normalized_probs(net_output, log_probs, sample) + lprobs.batch_first = True + return lprobs + + +DEFAULT_ENC_VGGBLOCK_CONFIG = ((32, 3, 2, 2, False),) * 2 +DEFAULT_ENC_TRANSFORMER_CONFIG = ((256, 4, 1024, True, 0.2, 0.2, 0.2),) * 2 +# 256: embedding dimension +# 4: number of heads +# 1024: FFN +# True: apply layerNorm before (dropout + resiaul) instead of after +# 0.2 (dropout): dropout after MultiheadAttention and second FC +# 0.2 (attention_dropout): dropout in MultiheadAttention +# 0.2 (relu_dropout): dropout after ReLu +DEFAULT_DEC_TRANSFORMER_CONFIG = ((256, 2, 1024, True, 0.2, 0.2, 0.2),) * 2 +DEFAULT_DEC_CONV_CONFIG = ((256, 3, True),) * 2 + + +# TODO: repace transformer encoder config from one liner +# to explicit args to get rid of this transformation +def prepare_transformer_encoder_params( + input_dim, + num_heads, + ffn_dim, + normalize_before, + dropout, + attention_dropout, + relu_dropout, +): + args = argparse.Namespace() + args.encoder_embed_dim = input_dim + args.encoder_attention_heads = num_heads + args.attention_dropout = attention_dropout + args.dropout = dropout + args.activation_dropout = relu_dropout + args.encoder_normalize_before = normalize_before + args.encoder_ffn_embed_dim = ffn_dim + return args + + +def prepare_transformer_decoder_params( + input_dim, + num_heads, + ffn_dim, + normalize_before, + dropout, + attention_dropout, + relu_dropout, +): + args = argparse.Namespace() + args.decoder_embed_dim = input_dim + args.decoder_attention_heads = num_heads + args.attention_dropout = attention_dropout + args.dropout = dropout + args.activation_dropout = relu_dropout + args.decoder_normalize_before = normalize_before + args.decoder_ffn_embed_dim = ffn_dim + return args + + +class VGGTransformerEncoder(FairseqEncoder): + """VGG + Transformer encoder""" + + def __init__( + self, + input_feat_per_channel, + vggblock_config=DEFAULT_ENC_VGGBLOCK_CONFIG, + transformer_config=DEFAULT_ENC_TRANSFORMER_CONFIG, + encoder_output_dim=512, + in_channels=1, + transformer_context=None, + transformer_sampling=None, + ): + """constructor for VGGTransformerEncoder + + Args: + - input_feat_per_channel: feature dim (not including stacked, + just base feature) + - in_channel: # input channels (e.g., if stack 8 feature vector + together, this is 8) + - vggblock_config: configuration of vggblock, see comments on + DEFAULT_ENC_VGGBLOCK_CONFIG + - transformer_config: configuration of transformer layer, see comments + on DEFAULT_ENC_TRANSFORMER_CONFIG + - encoder_output_dim: final transformer output embedding dimension + - transformer_context: (left, right) if set, self-attention will be focused + on (t-left, t+right) + - transformer_sampling: an iterable of int, must match with + len(transformer_config), transformer_sampling[i] indicates sampling + factor for i-th transformer layer, after multihead att and feedfoward + part + """ + super().__init__(None) + + self.num_vggblocks = 0 + if vggblock_config is not None: + if not isinstance(vggblock_config, Iterable): + raise ValueError("vggblock_config is not iterable") + self.num_vggblocks = len(vggblock_config) + + self.conv_layers = nn.ModuleList() + self.in_channels = in_channels + self.input_dim = input_feat_per_channel + self.pooling_kernel_sizes = [] + + if vggblock_config is not None: + for _, config in enumerate(vggblock_config): + ( + out_channels, + conv_kernel_size, + pooling_kernel_size, + num_conv_layers, + layer_norm, + ) = config + self.conv_layers.append( + VGGBlock( + in_channels, + out_channels, + conv_kernel_size, + pooling_kernel_size, + num_conv_layers, + input_dim=input_feat_per_channel, + layer_norm=layer_norm, + ) + ) + self.pooling_kernel_sizes.append(pooling_kernel_size) + in_channels = out_channels + input_feat_per_channel = self.conv_layers[-1].output_dim + + transformer_input_dim = self.infer_conv_output_dim( + self.in_channels, self.input_dim + ) + # transformer_input_dim is the output dimension of VGG part + + self.validate_transformer_config(transformer_config) + self.transformer_context = self.parse_transformer_context(transformer_context) + self.transformer_sampling = self.parse_transformer_sampling( + transformer_sampling, len(transformer_config) + ) + + self.transformer_layers = nn.ModuleList() + + if transformer_input_dim != transformer_config[0][0]: + self.transformer_layers.append( + Linear(transformer_input_dim, transformer_config[0][0]) + ) + self.transformer_layers.append( + TransformerEncoderLayer( + prepare_transformer_encoder_params(*transformer_config[0]) + ) + ) + + for i in range(1, len(transformer_config)): + if transformer_config[i - 1][0] != transformer_config[i][0]: + self.transformer_layers.append( + Linear(transformer_config[i - 1][0], transformer_config[i][0]) + ) + self.transformer_layers.append( + TransformerEncoderLayer( + prepare_transformer_encoder_params(*transformer_config[i]) + ) + ) + + self.encoder_output_dim = encoder_output_dim + self.transformer_layers.extend( + [ + Linear(transformer_config[-1][0], encoder_output_dim), + LayerNorm(encoder_output_dim), + ] + ) + + def forward(self, src_tokens, src_lengths, **kwargs): + """ + src_tokens: padded tensor (B, T, C * feat) + src_lengths: tensor of original lengths of input utterances (B,) + """ + bsz, max_seq_len, _ = src_tokens.size() + x = src_tokens.view(bsz, max_seq_len, self.in_channels, self.input_dim) + x = x.transpose(1, 2).contiguous() + # (B, C, T, feat) + + for layer_idx in range(len(self.conv_layers)): + x = self.conv_layers[layer_idx](x) + + bsz, _, output_seq_len, _ = x.size() + + # (B, C, T, feat) -> (B, T, C, feat) -> (T, B, C, feat) -> (T, B, C * feat) + x = x.transpose(1, 2).transpose(0, 1) + x = x.contiguous().view(output_seq_len, bsz, -1) + + input_lengths = src_lengths.clone() + for s in self.pooling_kernel_sizes: + input_lengths = (input_lengths.float() / s).ceil().long() + + encoder_padding_mask, _ = lengths_to_encoder_padding_mask( + input_lengths, batch_first=True + ) + if not encoder_padding_mask.any(): + encoder_padding_mask = None + + subsampling_factor = int(max_seq_len * 1.0 / output_seq_len + 0.5) + attn_mask = self.lengths_to_attn_mask(input_lengths, subsampling_factor) + + transformer_layer_idx = 0 + + for layer_idx in range(len(self.transformer_layers)): + + if isinstance(self.transformer_layers[layer_idx], TransformerEncoderLayer): + x = self.transformer_layers[layer_idx]( + x, encoder_padding_mask, attn_mask + ) + + if self.transformer_sampling[transformer_layer_idx] != 1: + sampling_factor = self.transformer_sampling[transformer_layer_idx] + x, encoder_padding_mask, attn_mask = self.slice( + x, encoder_padding_mask, attn_mask, sampling_factor + ) + + transformer_layer_idx += 1 + + else: + x = self.transformer_layers[layer_idx](x) + + # encoder_padding_maks is a (T x B) tensor, its [t, b] elements indicate + # whether encoder_output[t, b] is valid or not (valid=0, invalid=1) + + return { + "encoder_out": x, # (T, B, C) + "encoder_padding_mask": encoder_padding_mask.t() + if encoder_padding_mask is not None + else None, + # (B, T) --> (T, B) + } + + def infer_conv_output_dim(self, in_channels, input_dim): + sample_seq_len = 200 + sample_bsz = 10 + x = torch.randn(sample_bsz, in_channels, sample_seq_len, input_dim) + for i, _ in enumerate(self.conv_layers): + x = self.conv_layers[i](x) + x = x.transpose(1, 2) + mb, seq = x.size()[:2] + return x.contiguous().view(mb, seq, -1).size(-1) + + def validate_transformer_config(self, transformer_config): + for config in transformer_config: + input_dim, num_heads = config[:2] + if input_dim % num_heads != 0: + msg = ( + "ERROR in transformer config {}: ".format(config) + + "input dimension {} ".format(input_dim) + + "not dividable by number of heads {}".format(num_heads) + ) + raise ValueError(msg) + + def parse_transformer_context(self, transformer_context): + """ + transformer_context can be the following: + - None; indicates no context is used, i.e., + transformer can access full context + - a tuple/list of two int; indicates left and right context, + any number <0 indicates infinite context + * e.g., (5, 6) indicates that for query at x_t, transformer can + access [t-5, t+6] (inclusive) + * e.g., (-1, 6) indicates that for query at x_t, transformer can + access [0, t+6] (inclusive) + """ + if transformer_context is None: + return None + + if not isinstance(transformer_context, Iterable): + raise ValueError("transformer context must be Iterable if it is not None") + + if len(transformer_context) != 2: + raise ValueError("transformer context must have length 2") + + left_context = transformer_context[0] + if left_context < 0: + left_context = None + + right_context = transformer_context[1] + if right_context < 0: + right_context = None + + if left_context is None and right_context is None: + return None + + return (left_context, right_context) + + def parse_transformer_sampling(self, transformer_sampling, num_layers): + """ + parsing transformer sampling configuration + + Args: + - transformer_sampling, accepted input: + * None, indicating no sampling + * an Iterable with int (>0) as element + - num_layers, expected number of transformer layers, must match with + the length of transformer_sampling if it is not None + + Returns: + - A tuple with length num_layers + """ + if transformer_sampling is None: + return (1,) * num_layers + + if not isinstance(transformer_sampling, Iterable): + raise ValueError( + "transformer_sampling must be an iterable if it is not None" + ) + + if len(transformer_sampling) != num_layers: + raise ValueError( + "transformer_sampling {} does not match with the number " + "of layers {}".format(transformer_sampling, num_layers) + ) + + for layer, value in enumerate(transformer_sampling): + if not isinstance(value, int): + raise ValueError("Invalid value in transformer_sampling: ") + if value < 1: + raise ValueError( + "{} layer's subsampling is {}.".format(layer, value) + + " This is not allowed! " + ) + return transformer_sampling + + def slice(self, embedding, padding_mask, attn_mask, sampling_factor): + """ + embedding is a (T, B, D) tensor + padding_mask is a (B, T) tensor or None + attn_mask is a (T, T) tensor or None + """ + embedding = embedding[::sampling_factor, :, :] + if padding_mask is not None: + padding_mask = padding_mask[:, ::sampling_factor] + if attn_mask is not None: + attn_mask = attn_mask[::sampling_factor, ::sampling_factor] + + return embedding, padding_mask, attn_mask + + def lengths_to_attn_mask(self, input_lengths, subsampling_factor=1): + """ + create attention mask according to sequence lengths and transformer + context + + Args: + - input_lengths: (B, )-shape Int/Long tensor; input_lengths[b] is + the length of b-th sequence + - subsampling_factor: int + * Note that the left_context and right_context is specified in + the input frame-level while input to transformer may already + go through subsampling (e.g., the use of striding in vggblock) + we use subsampling_factor to scale the left/right context + + Return: + - a (T, T) binary tensor or None, where T is max(input_lengths) + * if self.transformer_context is None, None + * if left_context is None, + * attn_mask[t, t + right_context + 1:] = 1 + * others = 0 + * if right_context is None, + * attn_mask[t, 0:t - left_context] = 1 + * others = 0 + * elsif + * attn_mask[t, t - left_context: t + right_context + 1] = 0 + * others = 1 + """ + if self.transformer_context is None: + return None + + maxT = torch.max(input_lengths).item() + attn_mask = torch.zeros(maxT, maxT) + + left_context = self.transformer_context[0] + right_context = self.transformer_context[1] + if left_context is not None: + left_context = math.ceil(self.transformer_context[0] / subsampling_factor) + if right_context is not None: + right_context = math.ceil(self.transformer_context[1] / subsampling_factor) + + for t in range(maxT): + if left_context is not None: + st = 0 + en = max(st, t - left_context) + attn_mask[t, st:en] = 1 + if right_context is not None: + st = t + right_context + 1 + st = min(st, maxT - 1) + attn_mask[t, st:] = 1 + + return attn_mask.to(input_lengths.device) + + def reorder_encoder_out(self, encoder_out, new_order): + encoder_out["encoder_out"] = encoder_out["encoder_out"].index_select( + 1, new_order + ) + if encoder_out["encoder_padding_mask"] is not None: + encoder_out["encoder_padding_mask"] = encoder_out[ + "encoder_padding_mask" + ].index_select(1, new_order) + return encoder_out + + +class TransformerDecoder(FairseqIncrementalDecoder): + """ + Transformer decoder consisting of *args.decoder_layers* layers. Each layer + is a :class:`TransformerDecoderLayer`. + Args: + args (argparse.Namespace): parsed command-line arguments + dictionary (~fairseq.data.Dictionary): decoding dictionary + embed_tokens (torch.nn.Embedding): output embedding + no_encoder_attn (bool, optional): whether to attend to encoder outputs. + Default: ``False`` + left_pad (bool, optional): whether the input is left-padded. Default: + ``False`` + """ + + def __init__( + self, + dictionary, + embed_dim=512, + transformer_config=DEFAULT_ENC_TRANSFORMER_CONFIG, + conv_config=DEFAULT_DEC_CONV_CONFIG, + encoder_output_dim=512, + ): + + super().__init__(dictionary) + vocab_size = len(dictionary) + self.padding_idx = dictionary.pad() + self.embed_tokens = Embedding(vocab_size, embed_dim, self.padding_idx) + + self.conv_layers = nn.ModuleList() + for i in range(len(conv_config)): + out_channels, kernel_size, layer_norm = conv_config[i] + if i == 0: + conv_layer = LinearizedConv1d( + embed_dim, out_channels, kernel_size, padding=kernel_size - 1 + ) + else: + conv_layer = LinearizedConv1d( + conv_config[i - 1][0], + out_channels, + kernel_size, + padding=kernel_size - 1, + ) + self.conv_layers.append(conv_layer) + if layer_norm: + self.conv_layers.append(nn.LayerNorm(out_channels)) + self.conv_layers.append(nn.ReLU()) + + self.layers = nn.ModuleList() + if conv_config[-1][0] != transformer_config[0][0]: + self.layers.append(Linear(conv_config[-1][0], transformer_config[0][0])) + self.layers.append( + TransformerDecoderLayer( + prepare_transformer_decoder_params(*transformer_config[0]) + ) + ) + + for i in range(1, len(transformer_config)): + if transformer_config[i - 1][0] != transformer_config[i][0]: + self.layers.append( + Linear(transformer_config[i - 1][0], transformer_config[i][0]) + ) + self.layers.append( + TransformerDecoderLayer( + prepare_transformer_decoder_params(*transformer_config[i]) + ) + ) + self.fc_out = Linear(transformer_config[-1][0], vocab_size) + + def forward(self, prev_output_tokens, encoder_out=None, incremental_state=None): + """ + Args: + prev_output_tokens (LongTensor): previous decoder outputs of shape + `(batch, tgt_len)`, for input feeding/teacher forcing + encoder_out (Tensor, optional): output from the encoder, used for + encoder-side attention + incremental_state (dict): dictionary used for storing state during + :ref:`Incremental decoding` + Returns: + tuple: + - the last decoder layer's output of shape `(batch, tgt_len, + vocab)` + - the last decoder layer's attention weights of shape `(batch, + tgt_len, src_len)` + """ + target_padding_mask = ( + (prev_output_tokens == self.padding_idx).to(prev_output_tokens.device) + if incremental_state is None + else None + ) + + if incremental_state is not None: + prev_output_tokens = prev_output_tokens[:, -1:] + + # embed tokens + x = self.embed_tokens(prev_output_tokens) + + # B x T x C -> T x B x C + x = self._transpose_if_training(x, incremental_state) + + for layer in self.conv_layers: + if isinstance(layer, LinearizedConvolution): + x = layer(x, incremental_state) + else: + x = layer(x) + + # B x T x C -> T x B x C + x = self._transpose_if_inference(x, incremental_state) + + # decoder layers + for layer in self.layers: + if isinstance(layer, TransformerDecoderLayer): + x, *_ = layer( + x, + (encoder_out["encoder_out"] if encoder_out is not None else None), + ( + encoder_out["encoder_padding_mask"].t() + if encoder_out["encoder_padding_mask"] is not None + else None + ), + incremental_state, + self_attn_mask=( + self.buffered_future_mask(x) + if incremental_state is None + else None + ), + self_attn_padding_mask=( + target_padding_mask if incremental_state is None else None + ), + ) + else: + x = layer(x) + + # T x B x C -> B x T x C + x = x.transpose(0, 1) + + x = self.fc_out(x) + + return x, None + + def buffered_future_mask(self, tensor): + dim = tensor.size(0) + if ( + not hasattr(self, "_future_mask") + or self._future_mask is None + or self._future_mask.device != tensor.device + ): + self._future_mask = torch.triu( + utils.fill_with_neg_inf(tensor.new(dim, dim)), 1 + ) + if self._future_mask.size(0) < dim: + self._future_mask = torch.triu( + utils.fill_with_neg_inf(self._future_mask.resize_(dim, dim)), 1 + ) + return self._future_mask[:dim, :dim] + + def _transpose_if_training(self, x, incremental_state): + if incremental_state is None: + x = x.transpose(0, 1) + return x + + def _transpose_if_inference(self, x, incremental_state): + if incremental_state: + x = x.transpose(0, 1) + return x + + +@register_model("asr_vggtransformer_encoder") +class VGGTransformerEncoderModel(FairseqEncoderModel): + def __init__(self, encoder): + super().__init__(encoder) + + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + parser.add_argument( + "--input-feat-per-channel", + type=int, + metavar="N", + help="encoder input dimension per input channel", + ) + parser.add_argument( + "--vggblock-enc-config", + type=str, + metavar="EXPR", + help=""" + an array of tuples each containing the configuration of one vggblock + [(out_channels, conv_kernel_size, pooling_kernel_size,num_conv_layers), ...] + """, + ) + parser.add_argument( + "--transformer-enc-config", + type=str, + metavar="EXPR", + help=""" + a tuple containing the configuration of the Transformer layers + configurations: + [(input_dim, + num_heads, + ffn_dim, + normalize_before, + dropout, + attention_dropout, + relu_dropout), ]""", + ) + parser.add_argument( + "--enc-output-dim", + type=int, + metavar="N", + help="encoder output dimension, projecting the LSTM output", + ) + parser.add_argument( + "--in-channels", + type=int, + metavar="N", + help="number of encoder input channels", + ) + parser.add_argument( + "--transformer-context", + type=str, + metavar="EXPR", + help=""" + either None or a tuple of two ints, indicating left/right context a + transformer can have access to""", + ) + parser.add_argument( + "--transformer-sampling", + type=str, + metavar="EXPR", + help=""" + either None or a tuple of ints, indicating sampling factor in each layer""", + ) + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + base_architecture_enconly(args) + encoder = VGGTransformerEncoderOnly( + vocab_size=len(task.target_dictionary), + input_feat_per_channel=args.input_feat_per_channel, + vggblock_config=eval(args.vggblock_enc_config), + transformer_config=eval(args.transformer_enc_config), + encoder_output_dim=args.enc_output_dim, + in_channels=args.in_channels, + transformer_context=eval(args.transformer_context), + transformer_sampling=eval(args.transformer_sampling), + ) + return cls(encoder) + + def get_normalized_probs(self, net_output, log_probs, sample=None): + # net_output['encoder_out'] is a (T, B, D) tensor + lprobs = super().get_normalized_probs(net_output, log_probs, sample) + # lprobs is a (T, B, D) tensor + # we need to transoose to get (B, T, D) tensor + lprobs = lprobs.transpose(0, 1).contiguous() + lprobs.batch_first = True + return lprobs + + +class VGGTransformerEncoderOnly(VGGTransformerEncoder): + def __init__( + self, + vocab_size, + input_feat_per_channel, + vggblock_config=DEFAULT_ENC_VGGBLOCK_CONFIG, + transformer_config=DEFAULT_ENC_TRANSFORMER_CONFIG, + encoder_output_dim=512, + in_channels=1, + transformer_context=None, + transformer_sampling=None, + ): + super().__init__( + input_feat_per_channel=input_feat_per_channel, + vggblock_config=vggblock_config, + transformer_config=transformer_config, + encoder_output_dim=encoder_output_dim, + in_channels=in_channels, + transformer_context=transformer_context, + transformer_sampling=transformer_sampling, + ) + self.fc_out = Linear(self.encoder_output_dim, vocab_size) + + def forward(self, src_tokens, src_lengths, **kwargs): + """ + src_tokens: padded tensor (B, T, C * feat) + src_lengths: tensor of original lengths of input utterances (B,) + """ + + enc_out = super().forward(src_tokens, src_lengths) + x = self.fc_out(enc_out["encoder_out"]) + # x = F.log_softmax(x, dim=-1) + # Note: no need this line, because model.get_normalized_prob will call + # log_softmax + return { + "encoder_out": x, # (T, B, C) + "encoder_padding_mask": enc_out["encoder_padding_mask"], # (T, B) + } + + def max_positions(self): + """Maximum input length supported by the encoder.""" + return (1e6, 1e6) # an arbitrary large number + + +def Embedding(num_embeddings, embedding_dim, padding_idx): + m = nn.Embedding(num_embeddings, embedding_dim, padding_idx=padding_idx) + # nn.init.uniform_(m.weight, -0.1, 0.1) + # nn.init.constant_(m.weight[padding_idx], 0) + return m + + +def Linear(in_features, out_features, bias=True, dropout=0): + """Linear layer (input: N x T x C)""" + m = nn.Linear(in_features, out_features, bias=bias) + # m.weight.data.uniform_(-0.1, 0.1) + # if bias: + # m.bias.data.uniform_(-0.1, 0.1) + return m + + +def LinearizedConv1d(in_channels, out_channels, kernel_size, dropout=0, **kwargs): + """Weight-normalized Conv1d layer optimized for decoding""" + m = LinearizedConvolution(in_channels, out_channels, kernel_size, **kwargs) + std = math.sqrt((4 * (1.0 - dropout)) / (m.kernel_size[0] * in_channels)) + nn.init.normal_(m.weight, mean=0, std=std) + nn.init.constant_(m.bias, 0) + return nn.utils.weight_norm(m, dim=2) + + +def LayerNorm(embedding_dim): + m = nn.LayerNorm(embedding_dim) + return m + + +# seq2seq models +def base_architecture(args): + args.input_feat_per_channel = getattr(args, "input_feat_per_channel", 40) + args.vggblock_enc_config = getattr( + args, "vggblock_enc_config", DEFAULT_ENC_VGGBLOCK_CONFIG + ) + args.transformer_enc_config = getattr( + args, "transformer_enc_config", DEFAULT_ENC_TRANSFORMER_CONFIG + ) + args.enc_output_dim = getattr(args, "enc_output_dim", 512) + args.in_channels = getattr(args, "in_channels", 1) + args.tgt_embed_dim = getattr(args, "tgt_embed_dim", 128) + args.transformer_dec_config = getattr( + args, "transformer_dec_config", DEFAULT_ENC_TRANSFORMER_CONFIG + ) + args.conv_dec_config = getattr(args, "conv_dec_config", DEFAULT_DEC_CONV_CONFIG) + args.transformer_context = getattr(args, "transformer_context", "None") + + +@register_model_architecture("asr_vggtransformer", "vggtransformer_1") +def vggtransformer_1(args): + args.input_feat_per_channel = getattr(args, "input_feat_per_channel", 80) + args.vggblock_enc_config = getattr( + args, "vggblock_enc_config", "[(64, 3, 2, 2, True), (128, 3, 2, 2, True)]" + ) + args.transformer_enc_config = getattr( + args, + "transformer_enc_config", + "((1024, 16, 4096, True, 0.15, 0.15, 0.15),) * 14", + ) + args.enc_output_dim = getattr(args, "enc_output_dim", 1024) + args.tgt_embed_dim = getattr(args, "tgt_embed_dim", 128) + args.conv_dec_config = getattr(args, "conv_dec_config", "((256, 3, True),) * 4") + args.transformer_dec_config = getattr( + args, + "transformer_dec_config", + "((1024, 16, 4096, True, 0.15, 0.15, 0.15),) * 4", + ) + + +@register_model_architecture("asr_vggtransformer", "vggtransformer_2") +def vggtransformer_2(args): + args.input_feat_per_channel = getattr(args, "input_feat_per_channel", 80) + args.vggblock_enc_config = getattr( + args, "vggblock_enc_config", "[(64, 3, 2, 2, True), (128, 3, 2, 2, True)]" + ) + args.transformer_enc_config = getattr( + args, + "transformer_enc_config", + "((1024, 16, 4096, True, 0.15, 0.15, 0.15),) * 16", + ) + args.enc_output_dim = getattr(args, "enc_output_dim", 1024) + args.tgt_embed_dim = getattr(args, "tgt_embed_dim", 512) + args.conv_dec_config = getattr(args, "conv_dec_config", "((256, 3, True),) * 4") + args.transformer_dec_config = getattr( + args, + "transformer_dec_config", + "((1024, 16, 4096, True, 0.15, 0.15, 0.15),) * 6", + ) + + +@register_model_architecture("asr_vggtransformer", "vggtransformer_base") +def vggtransformer_base(args): + args.input_feat_per_channel = getattr(args, "input_feat_per_channel", 80) + args.vggblock_enc_config = getattr( + args, "vggblock_enc_config", "[(64, 3, 2, 2, True), (128, 3, 2, 2, True)]" + ) + args.transformer_enc_config = getattr( + args, "transformer_enc_config", "((512, 8, 2048, True, 0.15, 0.15, 0.15),) * 12" + ) + + args.enc_output_dim = getattr(args, "enc_output_dim", 512) + args.tgt_embed_dim = getattr(args, "tgt_embed_dim", 512) + args.conv_dec_config = getattr(args, "conv_dec_config", "((256, 3, True),) * 4") + args.transformer_dec_config = getattr( + args, "transformer_dec_config", "((512, 8, 2048, True, 0.15, 0.15, 0.15),) * 6" + ) + # Size estimations: + # Encoder: + # - vggblock param: 64*1*3*3 + 64*64*3*3 + 128*64*3*3 + 128*128*3 = 258K + # Transformer: + # - input dimension adapter: 2560 x 512 -> 1.31M + # - transformer_layers (x12) --> 37.74M + # * MultiheadAttention: 512*512*3 (in_proj) + 512*512 (out_proj) = 1.048M + # * FFN weight: 512*2048*2 = 2.097M + # - output dimension adapter: 512 x 512 -> 0.26 M + # Decoder: + # - LinearizedConv1d: 512 * 256 * 3 + 256 * 256 * 3 * 3 + # - transformer_layer: (x6) --> 25.16M + # * MultiheadAttention (self-attention): 512*512*3 + 512*512 = 1.048M + # * MultiheadAttention (encoder-attention): 512*512*3 + 512*512 = 1.048M + # * FFN: 512*2048*2 = 2.097M + # Final FC: + # - FC: 512*5000 = 256K (assuming vocab size 5K) + # In total: + # ~65 M + + +# CTC models +def base_architecture_enconly(args): + args.input_feat_per_channel = getattr(args, "input_feat_per_channel", 40) + args.vggblock_enc_config = getattr( + args, "vggblock_enc_config", "[(32, 3, 2, 2, True)] * 2" + ) + args.transformer_enc_config = getattr( + args, "transformer_enc_config", "((256, 4, 1024, True, 0.2, 0.2, 0.2),) * 2" + ) + args.enc_output_dim = getattr(args, "enc_output_dim", 512) + args.in_channels = getattr(args, "in_channels", 1) + args.transformer_context = getattr(args, "transformer_context", "None") + args.transformer_sampling = getattr(args, "transformer_sampling", "None") + + +@register_model_architecture("asr_vggtransformer_encoder", "vggtransformer_enc_1") +def vggtransformer_enc_1(args): + # vggtransformer_1 is the same as vggtransformer_enc_big, except the number + # of layers is increased to 16 + # keep it here for backward compatiablity purpose + args.input_feat_per_channel = getattr(args, "input_feat_per_channel", 80) + args.vggblock_enc_config = getattr( + args, "vggblock_enc_config", "[(64, 3, 2, 2, True), (128, 3, 2, 2, True)]" + ) + args.transformer_enc_config = getattr( + args, + "transformer_enc_config", + "((1024, 16, 4096, True, 0.15, 0.15, 0.15),) * 16", + ) + args.enc_output_dim = getattr(args, "enc_output_dim", 1024) diff --git a/examples/speech_recognition/models/w2l_conv_glu_enc.py b/examples/speech_recognition/models/w2l_conv_glu_enc.py new file mode 100644 index 0000000000000000000000000000000000000000..655a9b0d19d11e35511392a016f9d6b7d7aa2925 --- /dev/null +++ b/examples/speech_recognition/models/w2l_conv_glu_enc.py @@ -0,0 +1,177 @@ +#!/usr/bin/env python3 + +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +import torch +import torch.nn as nn +import torch.nn.functional as F +from fairseq.models import ( + FairseqEncoder, + FairseqEncoderModel, + register_model, + register_model_architecture, +) +from fairseq.modules.fairseq_dropout import FairseqDropout + + +default_conv_enc_config = """[ + (400, 13, 170, 0.2), + (440, 14, 0, 0.214), + (484, 15, 0, 0.22898), + (532, 16, 0, 0.2450086), + (584, 17, 0, 0.262159202), + (642, 18, 0, 0.28051034614), + (706, 19, 0, 0.30014607037), + (776, 20, 0, 0.321156295296), + (852, 21, 0, 0.343637235966), + (936, 22, 0, 0.367691842484), + (1028, 23, 0, 0.393430271458), + (1130, 24, 0, 0.42097039046), + (1242, 25, 0, 0.450438317792), + (1366, 26, 0, 0.481969000038), + (1502, 27, 0, 0.51570683004), + (1652, 28, 0, 0.551806308143), + (1816, 29, 0, 0.590432749713), +]""" + + +@register_model("asr_w2l_conv_glu_encoder") +class W2lConvGluEncoderModel(FairseqEncoderModel): + def __init__(self, encoder): + super().__init__(encoder) + + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + parser.add_argument( + "--input-feat-per-channel", + type=int, + metavar="N", + help="encoder input dimension per input channel", + ) + parser.add_argument( + "--in-channels", + type=int, + metavar="N", + help="number of encoder input channels", + ) + parser.add_argument( + "--conv-enc-config", + type=str, + metavar="EXPR", + help=""" + an array of tuples each containing the configuration of one conv layer + [(out_channels, kernel_size, padding, dropout), ...] + """, + ) + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + conv_enc_config = getattr(args, "conv_enc_config", default_conv_enc_config) + encoder = W2lConvGluEncoder( + vocab_size=len(task.target_dictionary), + input_feat_per_channel=args.input_feat_per_channel, + in_channels=args.in_channels, + conv_enc_config=eval(conv_enc_config), + ) + return cls(encoder) + + def get_normalized_probs(self, net_output, log_probs, sample=None): + lprobs = super().get_normalized_probs(net_output, log_probs, sample) + lprobs.batch_first = False + return lprobs + + +class W2lConvGluEncoder(FairseqEncoder): + def __init__( + self, vocab_size, input_feat_per_channel, in_channels, conv_enc_config + ): + super().__init__(None) + + self.input_dim = input_feat_per_channel + if in_channels != 1: + raise ValueError("only 1 input channel is currently supported") + + self.conv_layers = nn.ModuleList() + self.linear_layers = nn.ModuleList() + self.dropouts = [] + cur_channels = input_feat_per_channel + + for out_channels, kernel_size, padding, dropout in conv_enc_config: + layer = nn.Conv1d(cur_channels, out_channels, kernel_size, padding=padding) + layer.weight.data.mul_(math.sqrt(3)) # match wav2letter init + self.conv_layers.append(nn.utils.weight_norm(layer)) + self.dropouts.append( + FairseqDropout(dropout, module_name=self.__class__.__name__) + ) + if out_channels % 2 != 0: + raise ValueError("odd # of out_channels is incompatible with GLU") + cur_channels = out_channels // 2 # halved by GLU + + for out_channels in [2 * cur_channels, vocab_size]: + layer = nn.Linear(cur_channels, out_channels) + layer.weight.data.mul_(math.sqrt(3)) + self.linear_layers.append(nn.utils.weight_norm(layer)) + cur_channels = out_channels // 2 + + def forward(self, src_tokens, src_lengths, **kwargs): + + """ + src_tokens: padded tensor (B, T, C * feat) + src_lengths: tensor of original lengths of input utterances (B,) + """ + B, T, _ = src_tokens.size() + x = src_tokens.transpose(1, 2).contiguous() # (B, feat, T) assuming C == 1 + + for layer_idx in range(len(self.conv_layers)): + x = self.conv_layers[layer_idx](x) + x = F.glu(x, dim=1) + x = self.dropouts[layer_idx](x) + + x = x.transpose(1, 2).contiguous() # (B, T, 908) + x = self.linear_layers[0](x) + x = F.glu(x, dim=2) + x = self.dropouts[-1](x) + x = self.linear_layers[1](x) + + assert x.size(0) == B + assert x.size(1) == T + + encoder_out = x.transpose(0, 1) # (T, B, vocab_size) + + # need to debug this -- find a simpler/elegant way in pytorch APIs + encoder_padding_mask = ( + torch.arange(T).view(1, T).expand(B, -1).to(x.device) + >= src_lengths.view(B, 1).expand(-1, T) + ).t() # (B x T) -> (T x B) + + return { + "encoder_out": encoder_out, # (T, B, vocab_size) + "encoder_padding_mask": encoder_padding_mask, # (T, B) + } + + def reorder_encoder_out(self, encoder_out, new_order): + encoder_out["encoder_out"] = encoder_out["encoder_out"].index_select( + 1, new_order + ) + encoder_out["encoder_padding_mask"] = encoder_out[ + "encoder_padding_mask" + ].index_select(1, new_order) + return encoder_out + + def max_positions(self): + """Maximum input length supported by the encoder.""" + return (1e6, 1e6) # an arbitrary large number + + +@register_model_architecture("asr_w2l_conv_glu_encoder", "w2l_conv_glu_enc") +def w2l_conv_glu_enc(args): + args.input_feat_per_channel = getattr(args, "input_feat_per_channel", 80) + args.in_channels = getattr(args, "in_channels", 1) + args.conv_enc_config = getattr(args, "conv_enc_config", default_conv_enc_config) diff --git a/examples/speech_recognition/new/README.md b/examples/speech_recognition/new/README.md new file mode 100644 index 0000000000000000000000000000000000000000..5fa0e97245d3ba6db69d11222261b0644960183d --- /dev/null +++ b/examples/speech_recognition/new/README.md @@ -0,0 +1,43 @@ +# Flashlight Decoder + +This script runs decoding for pre-trained speech recognition models. + +## Usage + +Assuming a few variables: + +```bash +checkpoint=<path-to-checkpoint> +data=<path-to-data-directory> +lm_model=<path-to-language-model> +lexicon=<path-to-lexicon> +``` + +Example usage for decoding a fine-tuned Wav2Vec model: + +```bash +python $FAIRSEQ_ROOT/examples/speech_recognition/new/infer.py --multirun \ + task=audio_pretraining \ + task.data=$data \ + task.labels=ltr \ + common_eval.path=$checkpoint \ + decoding.type=kenlm \ + decoding.lexicon=$lexicon \ + decoding.lmpath=$lm_model \ + dataset.gen_subset=dev_clean,dev_other,test_clean,test_other +``` + +Example usage for using Ax to sweep WER parameters (requires `pip install hydra-ax-sweeper`): + +```bash +python $FAIRSEQ_ROOT/examples/speech_recognition/new/infer.py --multirun \ + hydra/sweeper=ax \ + task=audio_pretraining \ + task.data=$data \ + task.labels=ltr \ + common_eval.path=$checkpoint \ + decoding.type=kenlm \ + decoding.lexicon=$lexicon \ + decoding.lmpath=$lm_model \ + dataset.gen_subset=dev_other +``` diff --git a/examples/speech_recognition/new/__init__.py b/examples/speech_recognition/new/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/examples/speech_recognition/new/conf/hydra/sweeper/ax.yaml b/examples/speech_recognition/new/conf/hydra/sweeper/ax.yaml new file mode 100644 index 0000000000000000000000000000000000000000..fbeff17ca6b5fb0a1b44de0abe0d1a3d3d2aeeb2 --- /dev/null +++ b/examples/speech_recognition/new/conf/hydra/sweeper/ax.yaml @@ -0,0 +1,26 @@ +# @package hydra.sweeper +_target_: hydra_plugins.hydra_ax_sweeper.ax_sweeper.AxSweeper +max_batch_size: null +ax_config: + max_trials: 128 + early_stop: + minimize: true + max_epochs_without_improvement: 32 + epsilon: 1.0e-05 + experiment: + name: ${dataset.gen_subset} + objective_name: wer + minimize: true + parameter_constraints: null + outcome_constraints: null + status_quo: null + client: + verbose_logging: false + random_seed: null + params: + decoding.lmweight: + type: range + bounds: [0.0, 5.0] + decoding.wordscore: + type: range + bounds: [-5.0, 5.0] diff --git a/examples/speech_recognition/new/conf/infer.yaml b/examples/speech_recognition/new/conf/infer.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f176228082478fae0586a6da60a437e7b377b9ae --- /dev/null +++ b/examples/speech_recognition/new/conf/infer.yaml @@ -0,0 +1,25 @@ +# @package _group_ + +defaults: + - task: null + - model: null + +hydra: + run: + dir: ${common_eval.results_path}/${dataset.gen_subset} + sweep: + dir: ${common_eval.results_path} + subdir: ${dataset.gen_subset} +common_eval: + results_path: null + path: null + post_process: letter + quiet: true +dataset: + max_tokens: 1000000 + gen_subset: test +distributed_training: + distributed_world_size: 1 +decoding: + beam: 5 + type: viterbi diff --git a/examples/speech_recognition/new/decoders/__init__.py b/examples/speech_recognition/new/decoders/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/examples/speech_recognition/new/decoders/base_decoder.py b/examples/speech_recognition/new/decoders/base_decoder.py new file mode 100644 index 0000000000000000000000000000000000000000..a097969b3c0650cf8ea2ab5f8e96bbc68ea9b97f --- /dev/null +++ b/examples/speech_recognition/new/decoders/base_decoder.py @@ -0,0 +1,62 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import itertools as it +from typing import Any, Dict, List + +import torch +from fairseq.data.dictionary import Dictionary +from fairseq.models.fairseq_model import FairseqModel + + +class BaseDecoder: + def __init__(self, tgt_dict: Dictionary) -> None: + self.tgt_dict = tgt_dict + self.vocab_size = len(tgt_dict) + + self.blank = ( + tgt_dict.index("<ctc_blank>") + if "<ctc_blank>" in tgt_dict.indices + else tgt_dict.bos() + ) + if "<sep>" in tgt_dict.indices: + self.silence = tgt_dict.index("<sep>") + elif "|" in tgt_dict.indices: + self.silence = tgt_dict.index("|") + else: + self.silence = tgt_dict.eos() + + def generate( + self, models: List[FairseqModel], sample: Dict[str, Any], **unused + ) -> List[List[Dict[str, torch.LongTensor]]]: + encoder_input = { + k: v for k, v in sample["net_input"].items() if k != "prev_output_tokens" + } + emissions = self.get_emissions(models, encoder_input) + return self.decode(emissions) + + def get_emissions( + self, + models: List[FairseqModel], + encoder_input: Dict[str, Any], + ) -> torch.FloatTensor: + model = models[0] + encoder_out = model(**encoder_input) + if hasattr(model, "get_logits"): + emissions = model.get_logits(encoder_out) + else: + emissions = model.get_normalized_probs(encoder_out, log_probs=True) + return emissions.transpose(0, 1).float().cpu().contiguous() + + def get_tokens(self, idxs: torch.IntTensor) -> torch.LongTensor: + idxs = (g[0] for g in it.groupby(idxs)) + idxs = filter(lambda x: x != self.blank, idxs) + return torch.LongTensor(list(idxs)) + + def decode( + self, + emissions: torch.FloatTensor, + ) -> List[List[Dict[str, torch.LongTensor]]]: + raise NotImplementedError diff --git a/examples/speech_recognition/new/decoders/decoder.py b/examples/speech_recognition/new/decoders/decoder.py new file mode 100644 index 0000000000000000000000000000000000000000..b5bec8cf707b53104ef7a45993a5db2893d3443b --- /dev/null +++ b/examples/speech_recognition/new/decoders/decoder.py @@ -0,0 +1,32 @@ +#!/usr/bin/env python3 + +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from typing import Union + +from fairseq.data.dictionary import Dictionary + +from .decoder_config import DecoderConfig, FlashlightDecoderConfig +from .base_decoder import BaseDecoder + + +def Decoder( + cfg: Union[DecoderConfig, FlashlightDecoderConfig], tgt_dict: Dictionary +) -> BaseDecoder: + + if cfg.type == "viterbi": + from .viterbi_decoder import ViterbiDecoder + + return ViterbiDecoder(tgt_dict) + if cfg.type == "kenlm": + from .flashlight_decoder import KenLMDecoder + + return KenLMDecoder(cfg, tgt_dict) + if cfg.type == "fairseqlm": + from .flashlight_decoder import FairseqLMDecoder + + return FairseqLMDecoder(cfg, tgt_dict) + raise NotImplementedError(f"Invalid decoder name: {cfg.name}") diff --git a/examples/speech_recognition/new/decoders/decoder_config.py b/examples/speech_recognition/new/decoders/decoder_config.py new file mode 100644 index 0000000000000000000000000000000000000000..659eb94a9b8187a7c126d7b439ac2742f9d72022 --- /dev/null +++ b/examples/speech_recognition/new/decoders/decoder_config.py @@ -0,0 +1,70 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +from dataclasses import dataclass, field +from typing import Optional + +from fairseq.dataclass.configs import FairseqDataclass +from fairseq.dataclass.constants import ChoiceEnum +from omegaconf import MISSING + + +DECODER_CHOICES = ChoiceEnum(["viterbi", "kenlm", "fairseqlm"]) + + +@dataclass +class DecoderConfig(FairseqDataclass): + type: DECODER_CHOICES = field( + default="viterbi", + metadata={"help": "The type of decoder to use"}, + ) + + +@dataclass +class FlashlightDecoderConfig(FairseqDataclass): + nbest: int = field( + default=1, + metadata={"help": "Number of decodings to return"}, + ) + unitlm: bool = field( + default=False, + metadata={"help": "If set, use unit language model"}, + ) + lmpath: str = field( + default=MISSING, + metadata={"help": "Language model for KenLM decoder"}, + ) + lexicon: Optional[str] = field( + default=None, + metadata={"help": "Lexicon for Flashlight decoder"}, + ) + beam: int = field( + default=50, + metadata={"help": "Number of beams to use for decoding"}, + ) + beamthreshold: float = field( + default=50.0, + metadata={"help": "Threshold for beam search decoding"}, + ) + beamsizetoken: Optional[int] = field( + default=None, metadata={"help": "Beam size to use"} + ) + wordscore: float = field( + default=-1, + metadata={"help": "Word score for KenLM decoder"}, + ) + unkweight: float = field( + default=-math.inf, + metadata={"help": "Unknown weight for KenLM decoder"}, + ) + silweight: float = field( + default=0, + metadata={"help": "Silence weight for KenLM decoder"}, + ) + lmweight: float = field( + default=2, + metadata={"help": "Weight for LM while interpolating score"}, + ) diff --git a/examples/speech_recognition/new/decoders/flashlight_decoder.py b/examples/speech_recognition/new/decoders/flashlight_decoder.py new file mode 100644 index 0000000000000000000000000000000000000000..8a548bdf6613abc7d9a31d04be6158602ccf0967 --- /dev/null +++ b/examples/speech_recognition/new/decoders/flashlight_decoder.py @@ -0,0 +1,409 @@ +#!/usr/bin/env python3 + +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import gc +import os.path as osp +import warnings +from collections import deque, namedtuple +from typing import Any, Dict, Tuple + +import numpy as np +import torch +from fairseq import tasks +from fairseq.data.dictionary import Dictionary +from fairseq.dataclass.utils import convert_namespace_to_omegaconf +from fairseq.models.fairseq_model import FairseqModel +from fairseq.utils import apply_to_sample +from omegaconf import open_dict, OmegaConf + +from typing import List + +from .decoder_config import FlashlightDecoderConfig +from .base_decoder import BaseDecoder + +try: + from flashlight.lib.text.decoder import ( + LM, + CriterionType, + DecodeResult, + KenLM, + LexiconDecoder, + LexiconDecoderOptions, + LexiconFreeDecoder, + LexiconFreeDecoderOptions, + LMState, + SmearingMode, + Trie, + ) + from flashlight.lib.text.dictionary import create_word_dict, load_words +except ImportError: + warnings.warn( + "flashlight python bindings are required to use this functionality. " + "Please install from " + "https://github.com/facebookresearch/flashlight/tree/master/bindings/python" + ) + LM = object + LMState = object + + +class KenLMDecoder(BaseDecoder): + def __init__(self, cfg: FlashlightDecoderConfig, tgt_dict: Dictionary) -> None: + super().__init__(tgt_dict) + + self.nbest = cfg.nbest + self.unitlm = cfg.unitlm + + if cfg.lexicon: + self.lexicon = load_words(cfg.lexicon) + self.word_dict = create_word_dict(self.lexicon) + self.unk_word = self.word_dict.get_index("<unk>") + + self.lm = KenLM(cfg.lmpath, self.word_dict) + self.trie = Trie(self.vocab_size, self.silence) + + start_state = self.lm.start(False) + for word, spellings in self.lexicon.items(): + word_idx = self.word_dict.get_index(word) + _, score = self.lm.score(start_state, word_idx) + for spelling in spellings: + spelling_idxs = [tgt_dict.index(token) for token in spelling] + assert ( + tgt_dict.unk() not in spelling_idxs + ), f"{word} {spelling} {spelling_idxs}" + self.trie.insert(spelling_idxs, word_idx, score) + self.trie.smear(SmearingMode.MAX) + + self.decoder_opts = LexiconDecoderOptions( + beam_size=cfg.beam, + beam_size_token=cfg.beamsizetoken or len(tgt_dict), + beam_threshold=cfg.beamthreshold, + lm_weight=cfg.lmweight, + word_score=cfg.wordscore, + unk_score=cfg.unkweight, + sil_score=cfg.silweight, + log_add=False, + criterion_type=CriterionType.CTC, + ) + + self.decoder = LexiconDecoder( + self.decoder_opts, + self.trie, + self.lm, + self.silence, + self.blank, + self.unk_word, + [], + self.unitlm, + ) + else: + assert self.unitlm, "Lexicon-free decoding requires unit LM" + + d = {w: [[w]] for w in tgt_dict.symbols} + self.word_dict = create_word_dict(d) + self.lm = KenLM(cfg.lmpath, self.word_dict) + self.decoder_opts = LexiconFreeDecoderOptions( + beam_size=cfg.beam, + beam_size_token=cfg.beamsizetoken or len(tgt_dict), + beam_threshold=cfg.beamthreshold, + lm_weight=cfg.lmweight, + sil_score=cfg.silweight, + log_add=False, + criterion_type=CriterionType.CTC, + ) + self.decoder = LexiconFreeDecoder( + self.decoder_opts, self.lm, self.silence, self.blank, [] + ) + + def decode( + self, + emissions: torch.FloatTensor, + ) -> List[List[Dict[str, torch.LongTensor]]]: + B, T, N = emissions.size() + hypos = [] + for b in range(B): + emissions_ptr = emissions.data_ptr() + 4 * b * emissions.stride(0) + results = self.decoder.decode(emissions_ptr, T, N) + + nbest_results = results[: self.nbest] + hypos.append( + [ + { + "tokens": self.get_tokens(result.tokens), + "score": result.score, + "words": [ + self.word_dict.get_entry(x) for x in result.words if x >= 0 + ], + } + for result in nbest_results + ] + ) + return hypos + + +FairseqLMState = namedtuple( + "FairseqLMState", + [ + "prefix", + "incremental_state", + "probs", + ], +) + + +class FairseqLM(LM): + def __init__(self, dictionary: Dictionary, model: FairseqModel) -> None: + super().__init__() + + self.dictionary = dictionary + self.model = model + self.unk = self.dictionary.unk() + + self.save_incremental = False # this currently does not work properly + self.max_cache = 20_000 + + if torch.cuda.is_available(): + model.cuda() + model.eval() + model.make_generation_fast_() + + self.states = {} + self.stateq = deque() + + def start(self, start_with_nothing: bool) -> LMState: + state = LMState() + prefix = torch.LongTensor([[self.dictionary.eos()]]) + incremental_state = {} if self.save_incremental else None + with torch.no_grad(): + res = self.model(prefix.cuda(), incremental_state=incremental_state) + probs = self.model.get_normalized_probs(res, log_probs=True, sample=None) + + if incremental_state is not None: + incremental_state = apply_to_sample(lambda x: x.cpu(), incremental_state) + self.states[state] = FairseqLMState( + prefix.numpy(), incremental_state, probs[0, -1].cpu().numpy() + ) + self.stateq.append(state) + + return state + + def score( + self, + state: LMState, + token_index: int, + no_cache: bool = False, + ) -> Tuple[LMState, int]: + """ + Evaluate language model based on the current lm state and new word + Parameters: + ----------- + state: current lm state + token_index: index of the word + (can be lexicon index then you should store inside LM the + mapping between indices of lexicon and lm, or lm index of a word) + Returns: + -------- + (LMState, float): pair of (new state, score for the current word) + """ + curr_state = self.states[state] + + def trim_cache(targ_size: int) -> None: + while len(self.stateq) > targ_size: + rem_k = self.stateq.popleft() + rem_st = self.states[rem_k] + rem_st = FairseqLMState(rem_st.prefix, None, None) + self.states[rem_k] = rem_st + + if curr_state.probs is None: + new_incremental_state = ( + curr_state.incremental_state.copy() + if curr_state.incremental_state is not None + else None + ) + with torch.no_grad(): + if new_incremental_state is not None: + new_incremental_state = apply_to_sample( + lambda x: x.cuda(), new_incremental_state + ) + elif self.save_incremental: + new_incremental_state = {} + + res = self.model( + torch.from_numpy(curr_state.prefix).cuda(), + incremental_state=new_incremental_state, + ) + probs = self.model.get_normalized_probs( + res, log_probs=True, sample=None + ) + + if new_incremental_state is not None: + new_incremental_state = apply_to_sample( + lambda x: x.cpu(), new_incremental_state + ) + + curr_state = FairseqLMState( + curr_state.prefix, new_incremental_state, probs[0, -1].cpu().numpy() + ) + + if not no_cache: + self.states[state] = curr_state + self.stateq.append(state) + + score = curr_state.probs[token_index].item() + + trim_cache(self.max_cache) + + outstate = state.child(token_index) + if outstate not in self.states and not no_cache: + prefix = np.concatenate( + [curr_state.prefix, torch.LongTensor([[token_index]])], -1 + ) + incr_state = curr_state.incremental_state + + self.states[outstate] = FairseqLMState(prefix, incr_state, None) + + if token_index == self.unk: + score = float("-inf") + + return outstate, score + + def finish(self, state: LMState) -> Tuple[LMState, int]: + """ + Evaluate eos for language model based on the current lm state + Returns: + -------- + (LMState, float): pair of (new state, score for the current word) + """ + return self.score(state, self.dictionary.eos()) + + def empty_cache(self) -> None: + self.states = {} + self.stateq = deque() + gc.collect() + + +class FairseqLMDecoder(BaseDecoder): + def __init__(self, cfg: FlashlightDecoderConfig, tgt_dict: Dictionary) -> None: + super().__init__(tgt_dict) + + self.nbest = cfg.nbest + self.unitlm = cfg.unitlm + + self.lexicon = load_words(cfg.lexicon) if cfg.lexicon else None + self.idx_to_wrd = {} + + checkpoint = torch.load(cfg.lmpath, map_location="cpu") + + if "cfg" in checkpoint and checkpoint["cfg"] is not None: + lm_args = checkpoint["cfg"] + else: + lm_args = convert_namespace_to_omegaconf(checkpoint["args"]) + + if not OmegaConf.is_dict(lm_args): + lm_args = OmegaConf.create(lm_args) + + with open_dict(lm_args.task): + lm_args.task.data = osp.dirname(cfg.lmpath) + + task = tasks.setup_task(lm_args.task) + model = task.build_model(lm_args.model) + model.load_state_dict(checkpoint["model"], strict=False) + + self.trie = Trie(self.vocab_size, self.silence) + + self.word_dict = task.dictionary + self.unk_word = self.word_dict.unk() + self.lm = FairseqLM(self.word_dict, model) + + if self.lexicon: + start_state = self.lm.start(False) + for i, (word, spellings) in enumerate(self.lexicon.items()): + if self.unitlm: + word_idx = i + self.idx_to_wrd[i] = word + score = 0 + else: + word_idx = self.word_dict.index(word) + _, score = self.lm.score(start_state, word_idx, no_cache=True) + + for spelling in spellings: + spelling_idxs = [tgt_dict.index(token) for token in spelling] + assert ( + tgt_dict.unk() not in spelling_idxs + ), f"{spelling} {spelling_idxs}" + self.trie.insert(spelling_idxs, word_idx, score) + self.trie.smear(SmearingMode.MAX) + + self.decoder_opts = LexiconDecoderOptions( + beam_size=cfg.beam, + beam_size_token=cfg.beamsizetoken or len(tgt_dict), + beam_threshold=cfg.beamthreshold, + lm_weight=cfg.lmweight, + word_score=cfg.wordscore, + unk_score=cfg.unkweight, + sil_score=cfg.silweight, + log_add=False, + criterion_type=CriterionType.CTC, + ) + + self.decoder = LexiconDecoder( + self.decoder_opts, + self.trie, + self.lm, + self.silence, + self.blank, + self.unk_word, + [], + self.unitlm, + ) + else: + assert self.unitlm, "Lexicon-free decoding requires unit LM" + + d = {w: [[w]] for w in tgt_dict.symbols} + self.word_dict = create_word_dict(d) + self.lm = KenLM(cfg.lmpath, self.word_dict) + self.decoder_opts = LexiconFreeDecoderOptions( + beam_size=cfg.beam, + beam_size_token=cfg.beamsizetoken or len(tgt_dict), + beam_threshold=cfg.beamthreshold, + lm_weight=cfg.lmweight, + sil_score=cfg.silweight, + log_add=False, + criterion_type=CriterionType.CTC, + ) + self.decoder = LexiconFreeDecoder( + self.decoder_opts, self.lm, self.silence, self.blank, [] + ) + + def decode( + self, + emissions: torch.FloatTensor, + ) -> List[List[Dict[str, torch.LongTensor]]]: + B, T, N = emissions.size() + hypos = [] + + def make_hypo(result: DecodeResult) -> Dict[str, Any]: + hypo = { + "tokens": self.get_tokens(result.tokens), + "score": result.score, + } + if self.lexicon: + hypo["words"] = [ + self.idx_to_wrd[x] if self.unitlm else self.word_dict[x] + for x in result.words + if x >= 0 + ] + return hypo + + for b in range(B): + emissions_ptr = emissions.data_ptr() + 4 * b * emissions.stride(0) + results = self.decoder.decode(emissions_ptr, T, N) + + nbest_results = results[: self.nbest] + hypos.append([make_hypo(result) for result in nbest_results]) + self.lm.empty_cache() + + return hypos diff --git a/examples/speech_recognition/new/decoders/viterbi_decoder.py b/examples/speech_recognition/new/decoders/viterbi_decoder.py new file mode 100644 index 0000000000000000000000000000000000000000..b1c47868fa3b4e21f939b0695ede8d14ba1b168d --- /dev/null +++ b/examples/speech_recognition/new/decoders/viterbi_decoder.py @@ -0,0 +1,24 @@ +#!/usr/bin/env python3 + +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch + +from typing import List, Dict + +from .base_decoder import BaseDecoder + + +class ViterbiDecoder(BaseDecoder): + def decode( + self, + emissions: torch.FloatTensor, + ) -> List[List[Dict[str, torch.LongTensor]]]: + def get_pred(e): + toks = e.argmax(dim=-1).unique_consecutive() + return toks[toks != self.blank] + + return [[{"tokens": get_pred(x), "score": 0}] for x in emissions] diff --git a/examples/speech_recognition/new/infer.py b/examples/speech_recognition/new/infer.py new file mode 100644 index 0000000000000000000000000000000000000000..79afbc426d4655b6aa3eb4d12b2293fb57c9a568 --- /dev/null +++ b/examples/speech_recognition/new/infer.py @@ -0,0 +1,471 @@ +#!/usr/bin/env python -u +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import ast +import hashlib +import logging +import os +import shutil +import sys +from dataclasses import dataclass, field, is_dataclass +from pathlib import Path +from typing import Any, Dict, List, Optional, Tuple, Union + +import editdistance +import torch +import torch.distributed as dist +from examples.speech_recognition.new.decoders.decoder_config import ( + DecoderConfig, + FlashlightDecoderConfig, +) +from examples.speech_recognition.new.decoders.decoder import Decoder +from fairseq import checkpoint_utils, distributed_utils, progress_bar, tasks, utils +from fairseq.data.data_utils import post_process +from fairseq.dataclass.configs import ( + CheckpointConfig, + CommonConfig, + CommonEvalConfig, + DatasetConfig, + DistributedTrainingConfig, + FairseqDataclass, +) +from fairseq.logging.meters import StopwatchMeter, TimeMeter +from fairseq.logging.progress_bar import BaseProgressBar +from fairseq.models.fairseq_model import FairseqModel +from omegaconf import OmegaConf + +import hydra +from hydra.core.config_store import ConfigStore + +logging.root.setLevel(logging.INFO) +logging.basicConfig(level=logging.INFO) +logger = logging.getLogger(__name__) + +config_path = Path(__file__).resolve().parent / "conf" + + +@dataclass +class DecodingConfig(DecoderConfig, FlashlightDecoderConfig): + unique_wer_file: bool = field( + default=False, + metadata={"help": "If set, use a unique file for storing WER"}, + ) + results_path: Optional[str] = field( + default=None, + metadata={ + "help": "If set, write hypothesis and reference sentences into this directory" + }, + ) + + +@dataclass +class InferConfig(FairseqDataclass): + task: Any = None + decoding: DecodingConfig = DecodingConfig() + common: CommonConfig = CommonConfig() + common_eval: CommonEvalConfig = CommonEvalConfig() + checkpoint: CheckpointConfig = CheckpointConfig() + distributed_training: DistributedTrainingConfig = DistributedTrainingConfig() + dataset: DatasetConfig = DatasetConfig() + is_ax: bool = field( + default=False, + metadata={ + "help": "if true, assumes we are using ax for tuning and returns a tuple for ax to consume" + }, + ) + + +def reset_logging(): + root = logging.getLogger() + for handler in root.handlers: + root.removeHandler(handler) + root.setLevel(os.environ.get("LOGLEVEL", "INFO").upper()) + handler = logging.StreamHandler(sys.stdout) + handler.setFormatter( + logging.Formatter( + fmt="%(asctime)s | %(levelname)s | %(name)s | %(message)s", + datefmt="%Y-%m-%d %H:%M:%S", + ) + ) + root.addHandler(handler) + + +class InferenceProcessor: + cfg: InferConfig + + def __init__(self, cfg: InferConfig) -> None: + self.cfg = cfg + self.task = tasks.setup_task(cfg.task) + self.tgt_dict = self.task.target_dictionary + + models, saved_cfg = self.load_model_ensemble() + self.models = models + self.saved_cfg = saved_cfg + + self.task.load_dataset( + self.cfg.dataset.gen_subset, + task_cfg=saved_cfg.task, + ) + self.generator = Decoder(cfg.decoding, self.tgt_dict) + self.gen_timer = StopwatchMeter() + self.wps_meter = TimeMeter() + self.num_sentences = 0 + self.total_errors = 0 + self.total_length = 0 + + self.hypo_words_file = None + self.hypo_units_file = None + self.ref_words_file = None + self.ref_units_file = None + + self.progress_bar = self.build_progress_bar() + + def __enter__(self) -> "InferenceProcessor": + if self.cfg.decoding.results_path is not None: + self.hypo_words_file = self.get_res_file("hypo.word") + self.hypo_units_file = self.get_res_file("hypo.units") + self.ref_words_file = self.get_res_file("ref.word") + self.ref_units_file = self.get_res_file("ref.units") + return self + + def __exit__(self, *exc) -> bool: + if self.cfg.decoding.results_path is not None: + self.hypo_words_file.close() + self.hypo_units_file.close() + self.ref_words_file.close() + self.ref_units_file.close() + return False + + def __iter__(self) -> Any: + for sample in self.progress_bar: + if not self.cfg.common.cpu: + sample = utils.move_to_cuda(sample) + + # Happens on the last batch. + if "net_input" not in sample: + continue + yield sample + + def log(self, *args, **kwargs): + self.progress_bar.log(*args, **kwargs) + + def print(self, *args, **kwargs): + self.progress_bar.print(*args, **kwargs) + + def get_res_file(self, fname: str) -> None: + fname = os.path.join(self.cfg.decoding.results_path, fname) + if self.data_parallel_world_size > 1: + fname = f"{fname}.{self.data_parallel_rank}" + return open(fname, "w", buffering=1) + + def merge_shards(self) -> None: + """Merges all shard files into shard 0, then removes shard suffix.""" + + shard_id = self.data_parallel_rank + num_shards = self.data_parallel_world_size + + if self.data_parallel_world_size > 1: + + def merge_shards_with_root(fname: str) -> None: + fname = os.path.join(self.cfg.decoding.results_path, fname) + logger.info("Merging %s on shard %d", fname, shard_id) + base_fpath = Path(f"{fname}.0") + with open(base_fpath, "a") as out_file: + for s in range(1, num_shards): + shard_fpath = Path(f"{fname}.{s}") + with open(shard_fpath, "r") as in_file: + for line in in_file: + out_file.write(line) + shard_fpath.unlink() + shutil.move(f"{fname}.0", fname) + + dist.barrier() # ensure all shards finished writing + if shard_id == (0 % num_shards): + merge_shards_with_root("hypo.word") + if shard_id == (1 % num_shards): + merge_shards_with_root("hypo.units") + if shard_id == (2 % num_shards): + merge_shards_with_root("ref.word") + if shard_id == (3 % num_shards): + merge_shards_with_root("ref.units") + dist.barrier() + + def optimize_model(self, model: FairseqModel) -> None: + model.make_generation_fast_() + if self.cfg.common.fp16: + model.half() + if not self.cfg.common.cpu: + model.cuda() + + def load_model_ensemble(self) -> Tuple[List[FairseqModel], FairseqDataclass]: + arg_overrides = ast.literal_eval(self.cfg.common_eval.model_overrides) + models, saved_cfg = checkpoint_utils.load_model_ensemble( + utils.split_paths(self.cfg.common_eval.path, separator="\\"), + arg_overrides=arg_overrides, + task=self.task, + suffix=self.cfg.checkpoint.checkpoint_suffix, + strict=(self.cfg.checkpoint.checkpoint_shard_count == 1), + num_shards=self.cfg.checkpoint.checkpoint_shard_count, + ) + for model in models: + self.optimize_model(model) + return models, saved_cfg + + def get_dataset_itr(self, disable_iterator_cache: bool = False) -> None: + return self.task.get_batch_iterator( + dataset=self.task.dataset(self.cfg.dataset.gen_subset), + max_tokens=self.cfg.dataset.max_tokens, + max_sentences=self.cfg.dataset.batch_size, + max_positions=(sys.maxsize, sys.maxsize), + ignore_invalid_inputs=self.cfg.dataset.skip_invalid_size_inputs_valid_test, + required_batch_size_multiple=self.cfg.dataset.required_batch_size_multiple, + seed=self.cfg.common.seed, + num_shards=self.data_parallel_world_size, + shard_id=self.data_parallel_rank, + num_workers=self.cfg.dataset.num_workers, + data_buffer_size=self.cfg.dataset.data_buffer_size, + disable_iterator_cache=disable_iterator_cache, + ).next_epoch_itr(shuffle=False) + + def build_progress_bar( + self, + epoch: Optional[int] = None, + prefix: Optional[str] = None, + default_log_format: str = "tqdm", + ) -> BaseProgressBar: + return progress_bar.progress_bar( + iterator=self.get_dataset_itr(), + log_format=self.cfg.common.log_format, + log_interval=self.cfg.common.log_interval, + epoch=epoch, + prefix=prefix, + tensorboard_logdir=self.cfg.common.tensorboard_logdir, + default_log_format=default_log_format, + ) + + @property + def data_parallel_world_size(self): + if self.cfg.distributed_training.distributed_world_size == 1: + return 1 + return distributed_utils.get_data_parallel_world_size() + + @property + def data_parallel_rank(self): + if self.cfg.distributed_training.distributed_world_size == 1: + return 0 + return distributed_utils.get_data_parallel_rank() + + def process_sentence( + self, + sample: Dict[str, Any], + hypo: Dict[str, Any], + sid: int, + batch_id: int, + ) -> Tuple[int, int]: + speaker = None # Speaker can't be parsed from dataset. + + if "target_label" in sample: + toks = sample["target_label"] + else: + toks = sample["target"] + toks = toks[batch_id, :] + + # Processes hypothesis. + hyp_pieces = self.tgt_dict.string(hypo["tokens"].int().cpu()) + if "words" in hypo: + hyp_words = " ".join(hypo["words"]) + else: + hyp_words = post_process(hyp_pieces, self.cfg.common_eval.post_process) + + # Processes target. + target_tokens = utils.strip_pad(toks, self.tgt_dict.pad()) + tgt_pieces = self.tgt_dict.string(target_tokens.int().cpu()) + tgt_words = post_process(tgt_pieces, self.cfg.common_eval.post_process) + + if self.cfg.decoding.results_path is not None: + print(f"{hyp_pieces} ({speaker}-{sid})", file=self.hypo_units_file) + print(f"{hyp_words} ({speaker}-{sid})", file=self.hypo_words_file) + print(f"{tgt_pieces} ({speaker}-{sid})", file=self.ref_units_file) + print(f"{tgt_words} ({speaker}-{sid})", file=self.ref_words_file) + + if not self.cfg.common_eval.quiet: + logger.info(f"HYPO: {hyp_words}") + logger.info(f"REF: {tgt_words}") + logger.info("---------------------") + + hyp_words, tgt_words = hyp_words.split(), tgt_words.split() + + return editdistance.eval(hyp_words, tgt_words), len(tgt_words) + + def process_sample(self, sample: Dict[str, Any]) -> None: + self.gen_timer.start() + hypos = self.task.inference_step( + generator=self.generator, + models=self.models, + sample=sample, + ) + num_generated_tokens = sum(len(h[0]["tokens"]) for h in hypos) + self.gen_timer.stop(num_generated_tokens) + self.wps_meter.update(num_generated_tokens) + + for batch_id, sample_id in enumerate(sample["id"].tolist()): + errs, length = self.process_sentence( + sample=sample, + sid=sample_id, + batch_id=batch_id, + hypo=hypos[batch_id][0], + ) + self.total_errors += errs + self.total_length += length + + self.log({"wps": round(self.wps_meter.avg)}) + if "nsentences" in sample: + self.num_sentences += sample["nsentences"] + else: + self.num_sentences += sample["id"].numel() + + def log_generation_time(self) -> None: + logger.info( + "Processed %d sentences (%d tokens) in %.1fs %.2f " + "sentences per second, %.2f tokens per second)", + self.num_sentences, + self.gen_timer.n, + self.gen_timer.sum, + self.num_sentences / self.gen_timer.sum, + 1.0 / self.gen_timer.avg, + ) + + +def parse_wer(wer_file: Path) -> float: + with open(wer_file, "r") as f: + return float(f.readline().strip().split(" ")[1]) + + +def get_wer_file(cfg: InferConfig) -> Path: + """Hashes the decoding parameters to a unique file ID.""" + base_path = "wer" + if cfg.decoding.results_path is not None: + base_path = os.path.join(cfg.decoding.results_path, base_path) + + if cfg.decoding.unique_wer_file: + yaml_str = OmegaConf.to_yaml(cfg.decoding) + fid = int(hashlib.md5(yaml_str.encode("utf-8")).hexdigest(), 16) + return Path(f"{base_path}.{fid % 1000000}") + else: + return Path(base_path) + + +def main(cfg: InferConfig) -> float: + """Entry point for main processing logic. + + Args: + cfg: The inferance configuration to use. + wer: Optional shared memory pointer for returning the WER. If not None, + the final WER value will be written here instead of being returned. + + Returns: + The final WER if `wer` is None, otherwise None. + """ + + yaml_str, wer_file = OmegaConf.to_yaml(cfg.decoding), get_wer_file(cfg) + + # Validates the provided configuration. + if cfg.dataset.max_tokens is None and cfg.dataset.batch_size is None: + cfg.dataset.max_tokens = 4000000 + if not cfg.common.cpu and not torch.cuda.is_available(): + raise ValueError("CUDA not found; set `cpu=True` to run without CUDA") + + with InferenceProcessor(cfg) as processor: + for sample in processor: + processor.process_sample(sample) + + processor.log_generation_time() + + if cfg.decoding.results_path is not None: + processor.merge_shards() + + errs_t, leng_t = processor.total_errors, processor.total_length + + if cfg.common.cpu: + logger.warning("Merging WER requires CUDA.") + elif processor.data_parallel_world_size > 1: + stats = torch.LongTensor([errs_t, leng_t]).cuda() + dist.all_reduce(stats, op=dist.ReduceOp.SUM) + errs_t, leng_t = stats[0].item(), stats[1].item() + + wer = errs_t * 100.0 / leng_t + + if distributed_utils.is_master(cfg.distributed_training): + with open(wer_file, "w") as f: + f.write( + ( + f"WER: {wer}\n" + f"err / num_ref_words = {errs_t} / {leng_t}\n\n" + f"{yaml_str}" + ) + ) + + return wer + + +@hydra.main(config_path=config_path, config_name="infer") +def hydra_main(cfg: InferConfig) -> Union[float, Tuple[float, Optional[float]]]: + container = OmegaConf.to_container(cfg, resolve=True, enum_to_str=True) + cfg = OmegaConf.create(container) + OmegaConf.set_struct(cfg, True) + + if cfg.common.reset_logging: + reset_logging() + + # logger.info("Config:\n%s", OmegaConf.to_yaml(cfg)) + wer = float("inf") + + try: + if cfg.common.profile: + with torch.cuda.profiler.profile(): + with torch.autograd.profiler.emit_nvtx(): + distributed_utils.call_main(cfg, main) + else: + distributed_utils.call_main(cfg, main) + + wer = parse_wer(get_wer_file(cfg)) + except BaseException as e: # pylint: disable=broad-except + if not cfg.common.suppress_crashes: + raise + else: + logger.error("Crashed! %s", str(e)) + + logger.info("Word error rate: %.4f", wer) + if cfg.is_ax: + return wer, None + + return wer + + +def cli_main() -> None: + try: + from hydra._internal.utils import ( + get_args, + ) # pylint: disable=import-outside-toplevel + + cfg_name = get_args().config_name or "infer" + except ImportError: + logger.warning("Failed to get config name from hydra args") + cfg_name = "infer" + + cs = ConfigStore.instance() + cs.store(name=cfg_name, node=InferConfig) + + for k in InferConfig.__dataclass_fields__: + if is_dataclass(InferConfig.__dataclass_fields__[k].type): + v = InferConfig.__dataclass_fields__[k].default + cs.store(name=k, node=v) + + hydra_main() # pylint: disable=no-value-for-parameter + + +if __name__ == "__main__": + cli_main() diff --git a/examples/speech_recognition/tasks/__init__.py b/examples/speech_recognition/tasks/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..7ac3b8dc69639c92cc129294356e9012745e3fb2 --- /dev/null +++ b/examples/speech_recognition/tasks/__init__.py @@ -0,0 +1,8 @@ +import importlib +import os + + +for file in sorted(os.listdir(os.path.dirname(__file__))): + if file.endswith(".py") and not file.startswith("_"): + task_name = file[: file.find(".py")] + importlib.import_module("examples.speech_recognition.tasks." + task_name) diff --git a/examples/speech_recognition/tasks/speech_recognition.py b/examples/speech_recognition/tasks/speech_recognition.py new file mode 100644 index 0000000000000000000000000000000000000000..d9f011d55ff4fdfeb4c04ca790c314d685708c3a --- /dev/null +++ b/examples/speech_recognition/tasks/speech_recognition.py @@ -0,0 +1,157 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import json +import os +import re +import sys + +import torch +from examples.speech_recognition.data import AsrDataset +from examples.speech_recognition.data.replabels import replabel_symbol +from fairseq.data import Dictionary +from fairseq.tasks import LegacyFairseqTask, register_task + + +def get_asr_dataset_from_json(data_json_path, tgt_dict): + """ + Parse data json and create dataset. + See scripts/asr_prep_json.py which pack json from raw files + + Json example: + { + "utts": { + "4771-29403-0025": { + "input": { + "length_ms": 170, + "path": "/tmp/file1.flac" + }, + "output": { + "text": "HELLO \n", + "token": "HE LLO", + "tokenid": "4815, 861" + } + }, + "1564-142299-0096": { + ... + } + } + """ + if not os.path.isfile(data_json_path): + raise FileNotFoundError("Dataset not found: {}".format(data_json_path)) + with open(data_json_path, "rb") as f: + data_samples = json.load(f)["utts"] + assert len(data_samples) != 0 + sorted_samples = sorted( + data_samples.items(), + key=lambda sample: int(sample[1]["input"]["length_ms"]), + reverse=True, + ) + aud_paths = [s[1]["input"]["path"] for s in sorted_samples] + ids = [s[0] for s in sorted_samples] + speakers = [] + for s in sorted_samples: + m = re.search("(.+?)-(.+?)-(.+?)", s[0]) + speakers.append(m.group(1) + "_" + m.group(2)) + frame_sizes = [s[1]["input"]["length_ms"] for s in sorted_samples] + tgt = [ + [int(i) for i in s[1]["output"]["tokenid"].split(", ")] + for s in sorted_samples + ] + # append eos + tgt = [[*t, tgt_dict.eos()] for t in tgt] + return AsrDataset(aud_paths, frame_sizes, tgt, tgt_dict, ids, speakers) + + +@register_task("speech_recognition") +class SpeechRecognitionTask(LegacyFairseqTask): + """ + Task for training speech recognition model. + """ + + @staticmethod + def add_args(parser): + """Add task-specific arguments to the parser.""" + parser.add_argument("data", help="path to data directory") + parser.add_argument( + "--silence-token", default="\u2581", help="token for silence (used by w2l)" + ) + parser.add_argument( + "--max-source-positions", + default=sys.maxsize, + type=int, + metavar="N", + help="max number of frames in the source sequence", + ) + parser.add_argument( + "--max-target-positions", + default=1024, + type=int, + metavar="N", + help="max number of tokens in the target sequence", + ) + + def __init__(self, args, tgt_dict): + super().__init__(args) + self.tgt_dict = tgt_dict + + @classmethod + def setup_task(cls, args, **kwargs): + """Setup the task (e.g., load dictionaries).""" + dict_path = os.path.join(args.data, "dict.txt") + if not os.path.isfile(dict_path): + raise FileNotFoundError("Dict not found: {}".format(dict_path)) + tgt_dict = Dictionary.load(dict_path) + + if args.criterion == "ctc_loss": + tgt_dict.add_symbol("<ctc_blank>") + elif args.criterion == "asg_loss": + for i in range(1, args.max_replabel + 1): + tgt_dict.add_symbol(replabel_symbol(i)) + + print("| dictionary: {} types".format(len(tgt_dict))) + return cls(args, tgt_dict) + + def load_dataset(self, split, combine=False, **kwargs): + """Load a given dataset split. + + Args: + split (str): name of the split (e.g., train, valid, test) + """ + data_json_path = os.path.join(self.args.data, "{}.json".format(split)) + self.datasets[split] = get_asr_dataset_from_json(data_json_path, self.tgt_dict) + + def build_generator(self, models, args, **unused): + w2l_decoder = getattr(args, "w2l_decoder", None) + if w2l_decoder == "viterbi": + from examples.speech_recognition.w2l_decoder import W2lViterbiDecoder + + return W2lViterbiDecoder(args, self.target_dictionary) + elif w2l_decoder == "kenlm": + from examples.speech_recognition.w2l_decoder import W2lKenLMDecoder + + return W2lKenLMDecoder(args, self.target_dictionary) + elif w2l_decoder == "fairseqlm": + from examples.speech_recognition.w2l_decoder import W2lFairseqLMDecoder + + return W2lFairseqLMDecoder(args, self.target_dictionary) + else: + return super().build_generator(models, args) + + @property + def target_dictionary(self): + """Return the :class:`~fairseq.data.Dictionary` for the language + model.""" + return self.tgt_dict + + @property + def source_dictionary(self): + """Return the source :class:`~fairseq.data.Dictionary` (if applicable + for this task).""" + return None + + def max_positions(self): + """Return the max speech and sentence length allowed by the task.""" + return (self.args.max_source_positions, self.args.max_target_positions) diff --git a/examples/speech_recognition/utils/wer_utils.py b/examples/speech_recognition/utils/wer_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..cf6f3d09ba41a46ad4d7968fb3c286dd53d15c38 --- /dev/null +++ b/examples/speech_recognition/utils/wer_utils.py @@ -0,0 +1,381 @@ +#!/usr/bin/env python3 + +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from __future__ import absolute_import, division, print_function, unicode_literals + +import re +from collections import deque +from enum import Enum + +import numpy as np + + +""" + Utility modules for computation of Word Error Rate, + Alignments, as well as more granular metrics like + deletion, insersion and substitutions. +""" + + +class Code(Enum): + match = 1 + substitution = 2 + insertion = 3 + deletion = 4 + + +class Token(object): + def __init__(self, lbl="", st=np.nan, en=np.nan): + if np.isnan(st): + self.label, self.start, self.end = "", 0.0, 0.0 + else: + self.label, self.start, self.end = lbl, st, en + + +class AlignmentResult(object): + def __init__(self, refs, hyps, codes, score): + self.refs = refs # std::deque<int> + self.hyps = hyps # std::deque<int> + self.codes = codes # std::deque<Code> + self.score = score # float + + +def coordinate_to_offset(row, col, ncols): + return int(row * ncols + col) + + +def offset_to_row(offset, ncols): + return int(offset / ncols) + + +def offset_to_col(offset, ncols): + return int(offset % ncols) + + +def trimWhitespace(str): + return re.sub(" +", " ", re.sub(" *$", "", re.sub("^ *", "", str))) + + +def str2toks(str): + pieces = trimWhitespace(str).split(" ") + toks = [] + for p in pieces: + toks.append(Token(p, 0.0, 0.0)) + return toks + + +class EditDistance(object): + def __init__(self, time_mediated): + self.time_mediated_ = time_mediated + self.scores_ = np.nan # Eigen::Matrix<float, Eigen::Dynamic, Eigen::Dynamic> + self.backtraces_ = ( + np.nan + ) # Eigen::Matrix<size_t, Eigen::Dynamic, Eigen::Dynamic> backtraces_; + self.confusion_pairs_ = {} + + def cost(self, ref, hyp, code): + if self.time_mediated_: + if code == Code.match: + return abs(ref.start - hyp.start) + abs(ref.end - hyp.end) + elif code == Code.insertion: + return hyp.end - hyp.start + elif code == Code.deletion: + return ref.end - ref.start + else: # substitution + return abs(ref.start - hyp.start) + abs(ref.end - hyp.end) + 0.1 + else: + if code == Code.match: + return 0 + elif code == Code.insertion or code == Code.deletion: + return 3 + else: # substitution + return 4 + + def get_result(self, refs, hyps): + res = AlignmentResult(refs=deque(), hyps=deque(), codes=deque(), score=np.nan) + + num_rows, num_cols = self.scores_.shape + res.score = self.scores_[num_rows - 1, num_cols - 1] + + curr_offset = coordinate_to_offset(num_rows - 1, num_cols - 1, num_cols) + + while curr_offset != 0: + curr_row = offset_to_row(curr_offset, num_cols) + curr_col = offset_to_col(curr_offset, num_cols) + + prev_offset = self.backtraces_[curr_row, curr_col] + + prev_row = offset_to_row(prev_offset, num_cols) + prev_col = offset_to_col(prev_offset, num_cols) + + res.refs.appendleft(curr_row - 1) # Note: this was .push_front() in C++ + res.hyps.appendleft(curr_col - 1) + if curr_row - 1 == prev_row and curr_col == prev_col: + res.codes.appendleft(Code.deletion) + elif curr_row == prev_row and curr_col - 1 == prev_col: + res.codes.appendleft(Code.insertion) + else: + # assert(curr_row - 1 == prev_row and curr_col - 1 == prev_col) + ref_str = refs[res.refs[0]].label + hyp_str = hyps[res.hyps[0]].label + + if ref_str == hyp_str: + res.codes.appendleft(Code.match) + else: + res.codes.appendleft(Code.substitution) + + confusion_pair = "%s -> %s" % (ref_str, hyp_str) + if confusion_pair not in self.confusion_pairs_: + self.confusion_pairs_[confusion_pair] = 1 + else: + self.confusion_pairs_[confusion_pair] += 1 + + curr_offset = prev_offset + + return res + + def align(self, refs, hyps): + if len(refs) == 0 and len(hyps) == 0: + return np.nan + + # NOTE: we're not resetting the values in these matrices because every value + # will be overridden in the loop below. If this assumption doesn't hold, + # be sure to set all entries in self.scores_ and self.backtraces_ to 0. + self.scores_ = np.zeros((len(refs) + 1, len(hyps) + 1)) + self.backtraces_ = np.zeros((len(refs) + 1, len(hyps) + 1)) + + num_rows, num_cols = self.scores_.shape + + for i in range(num_rows): + for j in range(num_cols): + if i == 0 and j == 0: + self.scores_[i, j] = 0.0 + self.backtraces_[i, j] = 0 + continue + + if i == 0: + self.scores_[i, j] = self.scores_[i, j - 1] + self.cost( + None, hyps[j - 1], Code.insertion + ) + self.backtraces_[i, j] = coordinate_to_offset(i, j - 1, num_cols) + continue + + if j == 0: + self.scores_[i, j] = self.scores_[i - 1, j] + self.cost( + refs[i - 1], None, Code.deletion + ) + self.backtraces_[i, j] = coordinate_to_offset(i - 1, j, num_cols) + continue + + # Below here both i and j are greater than 0 + ref = refs[i - 1] + hyp = hyps[j - 1] + best_score = self.scores_[i - 1, j - 1] + ( + self.cost(ref, hyp, Code.match) + if (ref.label == hyp.label) + else self.cost(ref, hyp, Code.substitution) + ) + + prev_row = i - 1 + prev_col = j - 1 + ins = self.scores_[i, j - 1] + self.cost(None, hyp, Code.insertion) + if ins < best_score: + best_score = ins + prev_row = i + prev_col = j - 1 + + delt = self.scores_[i - 1, j] + self.cost(ref, None, Code.deletion) + if delt < best_score: + best_score = delt + prev_row = i - 1 + prev_col = j + + self.scores_[i, j] = best_score + self.backtraces_[i, j] = coordinate_to_offset( + prev_row, prev_col, num_cols + ) + + return self.get_result(refs, hyps) + + +class WERTransformer(object): + def __init__(self, hyp_str, ref_str, verbose=True): + self.ed_ = EditDistance(False) + self.id2oracle_errs_ = {} + self.utts_ = 0 + self.words_ = 0 + self.insertions_ = 0 + self.deletions_ = 0 + self.substitutions_ = 0 + + self.process(["dummy_str", hyp_str, ref_str]) + + if verbose: + print("'%s' vs '%s'" % (hyp_str, ref_str)) + self.report_result() + + def process(self, input): # std::vector<std::string>&& input + if len(input) < 3: + print( + "Input must be of the form <id> ... <hypo> <ref> , got ", + len(input), + " inputs:", + ) + return None + + # Align + # std::vector<Token> hyps; + # std::vector<Token> refs; + + hyps = str2toks(input[-2]) + refs = str2toks(input[-1]) + + alignment = self.ed_.align(refs, hyps) + if alignment is None: + print("Alignment is null") + return np.nan + + # Tally errors + ins = 0 + dels = 0 + subs = 0 + for code in alignment.codes: + if code == Code.substitution: + subs += 1 + elif code == Code.insertion: + ins += 1 + elif code == Code.deletion: + dels += 1 + + # Output + row = input + row.append(str(len(refs))) + row.append(str(ins)) + row.append(str(dels)) + row.append(str(subs)) + # print(row) + + # Accumulate + kIdIndex = 0 + kNBestSep = "/" + + pieces = input[kIdIndex].split(kNBestSep) + + if len(pieces) == 0: + print( + "Error splitting ", + input[kIdIndex], + " on '", + kNBestSep, + "', got empty list", + ) + return np.nan + + id = pieces[0] + if id not in self.id2oracle_errs_: + self.utts_ += 1 + self.words_ += len(refs) + self.insertions_ += ins + self.deletions_ += dels + self.substitutions_ += subs + self.id2oracle_errs_[id] = [ins, dels, subs] + else: + curr_err = ins + dels + subs + prev_err = np.sum(self.id2oracle_errs_[id]) + if curr_err < prev_err: + self.id2oracle_errs_[id] = [ins, dels, subs] + + return 0 + + def report_result(self): + # print("---------- Summary ---------------") + if self.words_ == 0: + print("No words counted") + return + + # 1-best + best_wer = ( + 100.0 + * (self.insertions_ + self.deletions_ + self.substitutions_) + / self.words_ + ) + + print( + "\tWER = %0.2f%% (%i utts, %i words, %0.2f%% ins, " + "%0.2f%% dels, %0.2f%% subs)" + % ( + best_wer, + self.utts_, + self.words_, + 100.0 * self.insertions_ / self.words_, + 100.0 * self.deletions_ / self.words_, + 100.0 * self.substitutions_ / self.words_, + ) + ) + + def wer(self): + if self.words_ == 0: + wer = np.nan + else: + wer = ( + 100.0 + * (self.insertions_ + self.deletions_ + self.substitutions_) + / self.words_ + ) + return wer + + def stats(self): + if self.words_ == 0: + stats = {} + else: + wer = ( + 100.0 + * (self.insertions_ + self.deletions_ + self.substitutions_) + / self.words_ + ) + stats = dict( + { + "wer": wer, + "utts": self.utts_, + "numwords": self.words_, + "ins": self.insertions_, + "dels": self.deletions_, + "subs": self.substitutions_, + "confusion_pairs": self.ed_.confusion_pairs_, + } + ) + return stats + + +def calc_wer(hyp_str, ref_str): + t = WERTransformer(hyp_str, ref_str, verbose=0) + return t.wer() + + +def calc_wer_stats(hyp_str, ref_str): + t = WERTransformer(hyp_str, ref_str, verbose=0) + return t.stats() + + +def get_wer_alignment_codes(hyp_str, ref_str): + """ + INPUT: hypothesis string, reference string + OUTPUT: List of alignment codes (intermediate results from WER computation) + """ + t = WERTransformer(hyp_str, ref_str, verbose=0) + return t.ed_.align(str2toks(ref_str), str2toks(hyp_str)).codes + + +def merge_counts(x, y): + # Merge two hashes which have 'counts' as their values + # This can be used for example to merge confusion pair counts + # conf_pairs = merge_counts(conf_pairs, stats['confusion_pairs']) + for k, v in y.items(): + if k not in x: + x[k] = 0 + x[k] += v + return x diff --git a/examples/speech_recognition/w2l_decoder.py b/examples/speech_recognition/w2l_decoder.py new file mode 100644 index 0000000000000000000000000000000000000000..aef44815932fb2d33bae2cf2d8777c762b9f8182 --- /dev/null +++ b/examples/speech_recognition/w2l_decoder.py @@ -0,0 +1,464 @@ +#!/usr/bin/env python3 + +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +""" +Flashlight decoders. +""" + +import gc +import itertools as it +import os.path as osp +import warnings +from collections import deque, namedtuple + +import numpy as np +import torch +from examples.speech_recognition.data.replabels import unpack_replabels +from fairseq import tasks +from fairseq.utils import apply_to_sample +from omegaconf import open_dict +from fairseq.dataclass.utils import convert_namespace_to_omegaconf + + +try: + from flashlight.lib.text.dictionary import create_word_dict, load_words + from flashlight.lib.sequence.criterion import CpuViterbiPath, get_data_ptr_as_bytes + from flashlight.lib.text.decoder import ( + CriterionType, + LexiconDecoderOptions, + KenLM, + LM, + LMState, + SmearingMode, + Trie, + LexiconDecoder, + ) +except: + warnings.warn( + "flashlight python bindings are required to use this functionality. Please install from https://github.com/facebookresearch/flashlight/tree/master/bindings/python" + ) + LM = object + LMState = object + + +class W2lDecoder(object): + def __init__(self, args, tgt_dict): + self.tgt_dict = tgt_dict + self.vocab_size = len(tgt_dict) + self.nbest = args.nbest + + # criterion-specific init + self.criterion_type = CriterionType.CTC + self.blank = ( + tgt_dict.index("<ctc_blank>") + if "<ctc_blank>" in tgt_dict.indices + else tgt_dict.bos() + ) + if "<sep>" in tgt_dict.indices: + self.silence = tgt_dict.index("<sep>") + elif "|" in tgt_dict.indices: + self.silence = tgt_dict.index("|") + else: + self.silence = tgt_dict.eos() + self.asg_transitions = None + + def generate(self, models, sample, **unused): + """Generate a batch of inferences.""" + # model.forward normally channels prev_output_tokens into the decoder + # separately, but SequenceGenerator directly calls model.encoder + encoder_input = { + k: v for k, v in sample["net_input"].items() if k != "prev_output_tokens" + } + emissions = self.get_emissions(models, encoder_input) + return self.decode(emissions) + + def get_emissions(self, models, encoder_input): + """Run encoder and normalize emissions""" + model = models[0] + encoder_out = model(**encoder_input) + if hasattr(model, "get_logits"): + emissions = model.get_logits(encoder_out) # no need to normalize emissions + else: + emissions = model.get_normalized_probs(encoder_out, log_probs=True) + return emissions.transpose(0, 1).float().cpu().contiguous() + + def get_tokens(self, idxs): + """Normalize tokens by handling CTC blank, ASG replabels, etc.""" + idxs = (g[0] for g in it.groupby(idxs)) + idxs = filter(lambda x: x != self.blank, idxs) + return torch.LongTensor(list(idxs)) + + +class W2lViterbiDecoder(W2lDecoder): + def __init__(self, args, tgt_dict): + super().__init__(args, tgt_dict) + + def decode(self, emissions): + B, T, N = emissions.size() + hypos = [] + if self.asg_transitions is None: + transitions = torch.FloatTensor(N, N).zero_() + else: + transitions = torch.FloatTensor(self.asg_transitions).view(N, N) + viterbi_path = torch.IntTensor(B, T) + workspace = torch.ByteTensor(CpuViterbiPath.get_workspace_size(B, T, N)) + CpuViterbiPath.compute( + B, + T, + N, + get_data_ptr_as_bytes(emissions), + get_data_ptr_as_bytes(transitions), + get_data_ptr_as_bytes(viterbi_path), + get_data_ptr_as_bytes(workspace), + ) + return [ + [{"tokens": self.get_tokens(viterbi_path[b].tolist()), "score": 0}] + for b in range(B) + ] + + +class W2lKenLMDecoder(W2lDecoder): + def __init__(self, args, tgt_dict): + super().__init__(args, tgt_dict) + + self.unit_lm = getattr(args, "unit_lm", False) + + if args.lexicon: + self.lexicon = load_words(args.lexicon) + self.word_dict = create_word_dict(self.lexicon) + self.unk_word = self.word_dict.get_index("<unk>") + + self.lm = KenLM(args.kenlm_model, self.word_dict) + self.trie = Trie(self.vocab_size, self.silence) + + start_state = self.lm.start(False) + for i, (word, spellings) in enumerate(self.lexicon.items()): + word_idx = self.word_dict.get_index(word) + _, score = self.lm.score(start_state, word_idx) + for spelling in spellings: + spelling_idxs = [tgt_dict.index(token) for token in spelling] + assert ( + tgt_dict.unk() not in spelling_idxs + ), f"{spelling} {spelling_idxs}" + self.trie.insert(spelling_idxs, word_idx, score) + self.trie.smear(SmearingMode.MAX) + + self.decoder_opts = LexiconDecoderOptions( + beam_size=args.beam, + beam_size_token=int(getattr(args, "beam_size_token", len(tgt_dict))), + beam_threshold=args.beam_threshold, + lm_weight=args.lm_weight, + word_score=args.word_score, + unk_score=args.unk_weight, + sil_score=args.sil_weight, + log_add=False, + criterion_type=self.criterion_type, + ) + + if self.asg_transitions is None: + N = 768 + # self.asg_transitions = torch.FloatTensor(N, N).zero_() + self.asg_transitions = [] + + self.decoder = LexiconDecoder( + self.decoder_opts, + self.trie, + self.lm, + self.silence, + self.blank, + self.unk_word, + self.asg_transitions, + self.unit_lm, + ) + else: + assert args.unit_lm, "lexicon free decoding can only be done with a unit language model" + from flashlight.lib.text.decoder import LexiconFreeDecoder, LexiconFreeDecoderOptions + + d = {w: [[w]] for w in tgt_dict.symbols} + self.word_dict = create_word_dict(d) + self.lm = KenLM(args.kenlm_model, self.word_dict) + self.decoder_opts = LexiconFreeDecoderOptions( + beam_size=args.beam, + beam_size_token=int(getattr(args, "beam_size_token", len(tgt_dict))), + beam_threshold=args.beam_threshold, + lm_weight=args.lm_weight, + sil_score=args.sil_weight, + log_add=False, + criterion_type=self.criterion_type, + ) + self.decoder = LexiconFreeDecoder( + self.decoder_opts, self.lm, self.silence, self.blank, [] + ) + + + def decode(self, emissions): + B, T, N = emissions.size() + hypos = [] + for b in range(B): + emissions_ptr = emissions.data_ptr() + 4 * b * emissions.stride(0) + results = self.decoder.decode(emissions_ptr, T, N) + + nbest_results = results[: self.nbest] + hypos.append( + [ + { + "tokens": self.get_tokens(result.tokens), + "score": result.score, + "words": [ + self.word_dict.get_entry(x) for x in result.words if x >= 0 + ], + } + for result in nbest_results + ] + ) + return hypos + + +FairseqLMState = namedtuple("FairseqLMState", ["prefix", "incremental_state", "probs"]) + + +class FairseqLM(LM): + def __init__(self, dictionary, model): + LM.__init__(self) + self.dictionary = dictionary + self.model = model + self.unk = self.dictionary.unk() + + self.save_incremental = False # this currently does not work properly + self.max_cache = 20_000 + + model.cuda() + model.eval() + model.make_generation_fast_() + + self.states = {} + self.stateq = deque() + + def start(self, start_with_nothing): + state = LMState() + prefix = torch.LongTensor([[self.dictionary.eos()]]) + incremental_state = {} if self.save_incremental else None + with torch.no_grad(): + res = self.model(prefix.cuda(), incremental_state=incremental_state) + probs = self.model.get_normalized_probs(res, log_probs=True, sample=None) + + if incremental_state is not None: + incremental_state = apply_to_sample(lambda x: x.cpu(), incremental_state) + self.states[state] = FairseqLMState( + prefix.numpy(), incremental_state, probs[0, -1].cpu().numpy() + ) + self.stateq.append(state) + + return state + + def score(self, state: LMState, token_index: int, no_cache: bool = False): + """ + Evaluate language model based on the current lm state and new word + Parameters: + ----------- + state: current lm state + token_index: index of the word + (can be lexicon index then you should store inside LM the + mapping between indices of lexicon and lm, or lm index of a word) + + Returns: + -------- + (LMState, float): pair of (new state, score for the current word) + """ + curr_state = self.states[state] + + def trim_cache(targ_size): + while len(self.stateq) > targ_size: + rem_k = self.stateq.popleft() + rem_st = self.states[rem_k] + rem_st = FairseqLMState(rem_st.prefix, None, None) + self.states[rem_k] = rem_st + + if curr_state.probs is None: + new_incremental_state = ( + curr_state.incremental_state.copy() + if curr_state.incremental_state is not None + else None + ) + with torch.no_grad(): + if new_incremental_state is not None: + new_incremental_state = apply_to_sample( + lambda x: x.cuda(), new_incremental_state + ) + elif self.save_incremental: + new_incremental_state = {} + + res = self.model( + torch.from_numpy(curr_state.prefix).cuda(), + incremental_state=new_incremental_state, + ) + probs = self.model.get_normalized_probs( + res, log_probs=True, sample=None + ) + + if new_incremental_state is not None: + new_incremental_state = apply_to_sample( + lambda x: x.cpu(), new_incremental_state + ) + + curr_state = FairseqLMState( + curr_state.prefix, new_incremental_state, probs[0, -1].cpu().numpy() + ) + + if not no_cache: + self.states[state] = curr_state + self.stateq.append(state) + + score = curr_state.probs[token_index].item() + + trim_cache(self.max_cache) + + outstate = state.child(token_index) + if outstate not in self.states and not no_cache: + prefix = np.concatenate( + [curr_state.prefix, torch.LongTensor([[token_index]])], -1 + ) + incr_state = curr_state.incremental_state + + self.states[outstate] = FairseqLMState(prefix, incr_state, None) + + if token_index == self.unk: + score = float("-inf") + + return outstate, score + + def finish(self, state: LMState): + """ + Evaluate eos for language model based on the current lm state + + Returns: + -------- + (LMState, float): pair of (new state, score for the current word) + """ + return self.score(state, self.dictionary.eos()) + + def empty_cache(self): + self.states = {} + self.stateq = deque() + gc.collect() + + +class W2lFairseqLMDecoder(W2lDecoder): + def __init__(self, args, tgt_dict): + super().__init__(args, tgt_dict) + + self.unit_lm = getattr(args, "unit_lm", False) + + self.lexicon = load_words(args.lexicon) if args.lexicon else None + self.idx_to_wrd = {} + + checkpoint = torch.load(args.kenlm_model, map_location="cpu") + + if "cfg" in checkpoint and checkpoint["cfg"] is not None: + lm_args = checkpoint["cfg"] + else: + lm_args = convert_namespace_to_omegaconf(checkpoint["args"]) + + with open_dict(lm_args.task): + lm_args.task.data = osp.dirname(args.kenlm_model) + + task = tasks.setup_task(lm_args.task) + model = task.build_model(lm_args.model) + model.load_state_dict(checkpoint["model"], strict=False) + + self.trie = Trie(self.vocab_size, self.silence) + + self.word_dict = task.dictionary + self.unk_word = self.word_dict.unk() + self.lm = FairseqLM(self.word_dict, model) + + if self.lexicon: + start_state = self.lm.start(False) + for i, (word, spellings) in enumerate(self.lexicon.items()): + if self.unit_lm: + word_idx = i + self.idx_to_wrd[i] = word + score = 0 + else: + word_idx = self.word_dict.index(word) + _, score = self.lm.score(start_state, word_idx, no_cache=True) + + for spelling in spellings: + spelling_idxs = [tgt_dict.index(token) for token in spelling] + assert ( + tgt_dict.unk() not in spelling_idxs + ), f"{spelling} {spelling_idxs}" + self.trie.insert(spelling_idxs, word_idx, score) + self.trie.smear(SmearingMode.MAX) + + self.decoder_opts = LexiconDecoderOptions( + beam_size=args.beam, + beam_size_token=int(getattr(args, "beam_size_token", len(tgt_dict))), + beam_threshold=args.beam_threshold, + lm_weight=args.lm_weight, + word_score=args.word_score, + unk_score=args.unk_weight, + sil_score=args.sil_weight, + log_add=False, + criterion_type=self.criterion_type, + ) + + self.decoder = LexiconDecoder( + self.decoder_opts, + self.trie, + self.lm, + self.silence, + self.blank, + self.unk_word, + [], + self.unit_lm, + ) + else: + assert args.unit_lm, "lexicon free decoding can only be done with a unit language model" + from flashlight.lib.text.decoder import LexiconFreeDecoder, LexiconFreeDecoderOptions + + d = {w: [[w]] for w in tgt_dict.symbols} + self.word_dict = create_word_dict(d) + self.lm = KenLM(args.kenlm_model, self.word_dict) + self.decoder_opts = LexiconFreeDecoderOptions( + beam_size=args.beam, + beam_size_token=int(getattr(args, "beam_size_token", len(tgt_dict))), + beam_threshold=args.beam_threshold, + lm_weight=args.lm_weight, + sil_score=args.sil_weight, + log_add=False, + criterion_type=self.criterion_type, + ) + self.decoder = LexiconFreeDecoder( + self.decoder_opts, self.lm, self.silence, self.blank, [] + ) + + def decode(self, emissions): + B, T, N = emissions.size() + hypos = [] + + def idx_to_word(idx): + if self.unit_lm: + return self.idx_to_wrd[idx] + else: + return self.word_dict[idx] + + def make_hypo(result): + hypo = {"tokens": self.get_tokens(result.tokens), "score": result.score} + if self.lexicon: + hypo["words"] = [idx_to_word(x) for x in result.words if x >= 0] + return hypo + + for b in range(B): + emissions_ptr = emissions.data_ptr() + 4 * b * emissions.stride(0) + results = self.decoder.decode(emissions_ptr, T, N) + + nbest_results = results[: self.nbest] + hypos.append([make_hypo(result) for result in nbest_results]) + self.lm.empty_cache() + + return hypos diff --git a/examples/speech_to_text/README.md b/examples/speech_to_text/README.md new file mode 100644 index 0000000000000000000000000000000000000000..f639d300d342f8de1392c98bfc44ec8690188539 --- /dev/null +++ b/examples/speech_to_text/README.md @@ -0,0 +1,77 @@ +# Speech-to-Text (S2T) Modeling + +[https://www.aclweb.org/anthology/2020.aacl-demo.6](https://www.aclweb.org/anthology/2020.aacl-demo.6.pdf) + +Speech recognition (ASR) and speech-to-text translation (ST) with fairseq. + +## Data Preparation +S2T modeling data consists of source speech features, target text and other optional information +(source text, speaker id, etc.). Fairseq S2T uses per-dataset-split TSV manifest files +to store these information. Each data field is represented by a column in the TSV file. + +Unlike text token embeddings, speech features (e.g. log mel-scale filter banks) are usually fixed +during model training and can be pre-computed. The manifest file contains the path to +either the feature file in NumPy format or the WAV/FLAC audio file. For the latter, +features will be extracted on-the-fly by fairseq S2T. Optionally, feature/audio files can be packed +into uncompressed ZIP files (then accessed via byte offset and length) to improve I/O performance. + +Fairseq S2T also employs a YAML file for data related configurations: tokenizer type and dictionary path +for the target text, feature transforms such as CMVN (cepstral mean and variance normalization) and SpecAugment, +temperature-based resampling, etc. + +## Model Training +Fairseq S2T uses the unified `fairseq-train` interface for model training. It requires arguments `--task speech_to_text`, + `--arch <model architecture in fairseq.models.speech_to_text.*>` and `--config-yaml <config YAML filename>`. + +## Inference & Evaluation +Fairseq S2T uses the unified `fairseq-generate`/`fairseq-interactive` interface for inference and evaluation. It +requires arguments `--task speech_to_text` and `--config-yaml <config YAML filename>`. The interactive console takes +audio paths (one per line) as inputs. + + +## Examples +- [Speech Recognition (ASR) on LibriSpeech](docs/librispeech_example.md) + +- [Speech-to-Text Translation (ST) on MuST-C](docs/mustc_example.md) + +- [Speech-to-Text Translation (ST) on CoVoST 2](docs/covost_example.md) + +- [Speech-to-Text Translation (ST) on Multilingual TEDx](docs/mtedx_example.md) +- [Simultaneous Speech-to-Text Translation (SimulST) on MuST-C](docs/simulst_mustc_example.md) + +## Updates +- 02/04/2021: Added interactive decoding (`fairseq-interactive`) support. Examples: + [ASR (LibriSpeech)](docs/librispeech_example.md#interactive-decoding) + and [ST (CoVoST 2)](docs/covost_example.md#interactive-decoding). +- 01/08/2021: Several fixes for S2T Transformer model, inference-time de-tokenization, scorer configuration and data + preparation scripts. We also add pre-trained models to the examples and revise the instructions. + Breaking changes: the data preparation scripts now extract filterbank features without CMVN. CMVN is instead applied + on-the-fly (defined in the config YAML). + +## What's Next +- We are migrating the old fairseq [ASR example](../speech_recognition) into this S2T framework and + merging the features from both sides. +- The following papers also base their experiments on fairseq S2T. We are adding more examples for replication. + - [Improving Cross-Lingual Transfer Learning for End-to-End Speech Recognition with Speech Translation (Wang et al., 2020)](https://arxiv.org/abs/2006.05474) + - [Self-Supervised Representations Improve End-to-End Speech Translation (Wu et al., 2020)](https://arxiv.org/abs/2006.12124) + - [Self-Training for End-to-End Speech Translation (Pino et al., 2020)](https://arxiv.org/abs/2006.02490) + - [CoVoST: A Diverse Multilingual Speech-To-Text Translation Corpus (Wang et al., 2020)](https://arxiv.org/abs/2002.01320) + - [Harnessing Indirect Training Data for End-to-End Automatic Speech Translation: Tricks of the Trade (Pino et al., 2019)](https://arxiv.org/abs/1909.06515) + +## Citation +Please cite as: +``` +@inproceedings{wang2020fairseqs2t, + title = {fairseq S2T: Fast Speech-to-Text Modeling with fairseq}, + author = {Changhan Wang and Yun Tang and Xutai Ma and Anne Wu and Dmytro Okhonko and Juan Pino}, + booktitle = {Proceedings of the 2020 Conference of the Asian Chapter of the Association for Computational Linguistics (AACL): System Demonstrations}, + year = {2020}, +} + +@inproceedings{ott2019fairseq, + title = {fairseq: A Fast, Extensible Toolkit for Sequence Modeling}, + author = {Myle Ott and Sergey Edunov and Alexei Baevski and Angela Fan and Sam Gross and Nathan Ng and David Grangier and Michael Auli}, + booktitle = {Proceedings of NAACL-HLT 2019: Demonstrations}, + year = {2019}, +} +``` diff --git a/examples/speech_to_text/data_utils.py b/examples/speech_to_text/data_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..2bcff046f7e4daf7f6029f9e89936d2d0b708dae --- /dev/null +++ b/examples/speech_to_text/data_utils.py @@ -0,0 +1,339 @@ +#!/usr/bin/env python3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import csv +from pathlib import Path +import zipfile +from functools import reduce +from multiprocessing import cpu_count +from typing import Any, Dict, List, Optional, Union + +import numpy as np +import pandas as pd +import sentencepiece as sp +from fairseq.data.audio.audio_utils import ( + _convert_to_mono, _get_kaldi_fbank, _get_torchaudio_fbank +) +import torch +from tqdm import tqdm + + +UNK_TOKEN, UNK_TOKEN_ID = "<unk>", 3 +BOS_TOKEN, BOS_TOKEN_ID = "<s>", 0 +EOS_TOKEN, EOS_TOKEN_ID = "</s>", 2 +PAD_TOKEN, PAD_TOKEN_ID = "<pad>", 1 + + +def gen_vocab( + input_path: Path, output_path_prefix: Path, model_type="bpe", + vocab_size=1000, special_symbols: Optional[List[str]] = None +): + # Train SentencePiece Model + arguments = [ + f"--input={input_path.as_posix()}", + f"--model_prefix={output_path_prefix.as_posix()}", + f"--model_type={model_type}", + f"--vocab_size={vocab_size}", + "--character_coverage=1.0", + f"--num_threads={cpu_count()}", + f"--unk_id={UNK_TOKEN_ID}", + f"--bos_id={BOS_TOKEN_ID}", + f"--eos_id={EOS_TOKEN_ID}", + f"--pad_id={PAD_TOKEN_ID}", + ] + if special_symbols is not None: + _special_symbols = ",".join(special_symbols) + arguments.append(f"--user_defined_symbols={_special_symbols}") + sp.SentencePieceTrainer.Train(" ".join(arguments)) + # Export fairseq dictionary + spm = sp.SentencePieceProcessor() + spm.Load(output_path_prefix.as_posix() + ".model") + vocab = {i: spm.IdToPiece(i) for i in range(spm.GetPieceSize())} + assert ( + vocab.get(UNK_TOKEN_ID) == UNK_TOKEN + and vocab.get(PAD_TOKEN_ID) == PAD_TOKEN + and vocab.get(BOS_TOKEN_ID) == BOS_TOKEN + and vocab.get(EOS_TOKEN_ID) == EOS_TOKEN + ) + vocab = { + i: s + for i, s in vocab.items() + if s not in {UNK_TOKEN, BOS_TOKEN, EOS_TOKEN, PAD_TOKEN} + } + with open(output_path_prefix.as_posix() + ".txt", "w") as f_out: + for _, s in sorted(vocab.items(), key=lambda x: x[0]): + f_out.write(f"{s} 1\n") + + +def extract_fbank_features( + waveform: torch.FloatTensor, + sample_rate: int, + output_path: Optional[Path] = None, + n_mel_bins: int = 80, + overwrite: bool = False, +): + if output_path is not None and output_path.is_file() and not overwrite: + return + + _waveform = _convert_to_mono(waveform, sample_rate) + _waveform = _waveform * (2 ** 15) # Kaldi compliance: 16-bit signed integers + _waveform = _waveform.numpy() + + features = _get_kaldi_fbank(_waveform, sample_rate, n_mel_bins) + if features is None: + features = _get_torchaudio_fbank(_waveform, sample_rate, n_mel_bins) + if features is None: + raise ImportError( + "Please install pyKaldi or torchaudio to enable fbank feature extraction" + ) + + if output_path is not None: + np.save(output_path.as_posix(), features) + else: + return features + + +def create_zip(data_root: Path, zip_path: Path): + paths = list(data_root.glob("*.npy")) + with zipfile.ZipFile(zip_path, "w", zipfile.ZIP_STORED) as f: + for path in tqdm(paths): + f.write(path, arcname=path.name) + + +def is_npy_data(data: bytes) -> bool: + return data[0] == 147 and data[1] == 78 + + +def get_zip_manifest(zip_path: Path, zip_root: Optional[Path] = None): + _zip_path = zip_path if zip_root is None else Path.joinpath(zip_root, zip_path) + with zipfile.ZipFile(_zip_path, mode="r") as f: + info = f.infolist() + manifest = {} + for i in tqdm(info): + utt_id = Path(i.filename).stem + offset, file_size = i.header_offset + 30 + len(i.filename), i.file_size + manifest[utt_id] = f"{zip_path.as_posix()}:{offset}:{file_size}" + with open(_zip_path, "rb") as f: + f.seek(offset) + data = f.read(file_size) + assert len(data) > 1 and is_npy_data(data) + return manifest + + +def gen_config_yaml( + manifest_root: Path, + spm_filename: str, + yaml_filename: str = "config.yaml", + specaugment_policy: str = "lb", + prepend_tgt_lang_tag: bool = False, + sampling_alpha: float = 1.0, + audio_root: str = "", + cmvn_type: str = "utterance", + gcmvn_path: Optional[Path] = None, +): + manifest_root = manifest_root.absolute() + writer = S2TDataConfigWriter(manifest_root / yaml_filename) + writer.set_vocab_filename(spm_filename.replace(".model", ".txt")) + writer.set_input_channels(1) + writer.set_input_feat_per_channel(80) + specaugment_setters = { + "lb": writer.set_specaugment_lb_policy, + "ld": writer.set_specaugment_ld_policy, + "sm": writer.set_specaugment_sm_policy, + "ss": writer.set_specaugment_ss_policy, + } + specaugment_setter = specaugment_setters.get(specaugment_policy, None) + if specaugment_setter is not None: + specaugment_setter() + writer.set_bpe_tokenizer( + { + "bpe": "sentencepiece", + "sentencepiece_model": (manifest_root / spm_filename).as_posix(), + } + ) + if prepend_tgt_lang_tag: + writer.set_prepend_tgt_lang_tag(True) + writer.set_sampling_alpha(sampling_alpha) + + if cmvn_type not in ["global", "utterance"]: + raise NotImplementedError + + writer.set_feature_transforms("_train", [f"{cmvn_type}_cmvn", "specaugment"]) + writer.set_feature_transforms("*", [f"{cmvn_type}_cmvn"]) + + if cmvn_type == "global": + assert gcmvn_path is not None, ( + 'Please provide path of global cmvn file.' + ) + writer.set_global_cmvn(str(gcmvn_path)) + + if len(audio_root) > 0: + writer.set_audio_root(audio_root) + writer.flush() + + +def load_df_from_tsv(path: Union[str, Path]): + _path = path if isinstance(path, str) else path.as_posix() + return pd.read_csv( + _path, + sep="\t", + header=0, + encoding="utf-8", + escapechar="\\", + quoting=csv.QUOTE_NONE, + na_filter=False, + ) + + +def save_df_to_tsv(dataframe, path: Union[str, Path]): + _path = path if isinstance(path, str) else path.as_posix() + dataframe.to_csv( + _path, + sep="\t", + header=True, + index=False, + encoding="utf-8", + escapechar="\\", + quoting=csv.QUOTE_NONE, + ) + + +def filter_manifest_df( + df, is_train_split=False, extra_filters=None, min_n_frames=5, max_n_frames=3000 +): + filters = { + "no speech": df["audio"] == "", + f"short speech (<{min_n_frames} frames)": df["n_frames"] < min_n_frames, + "empty sentence": df["tgt_text"] == "", + } + if is_train_split: + filters[f"long speech (>{max_n_frames} frames)"] = df["n_frames"] > max_n_frames + if extra_filters is not None: + filters.update(extra_filters) + invalid = reduce(lambda x, y: x | y, filters.values()) + valid = ~invalid + print( + "| " + + ", ".join(f"{n}: {f.sum()}" for n, f in filters.items()) + + f", total {invalid.sum()} filtered, {valid.sum()} remained." + ) + return df[valid] + + +def cal_gcmvn_stats(features_list): + features = np.concatenate(features_list) + square_sums = (features ** 2).sum(axis=0) + mean = features.mean(axis=0) + features = np.subtract(features, mean) + var = square_sums / features.shape[0] - mean ** 2 + std = np.sqrt(np.maximum(var, 1e-8)) + return {"mean": mean.astype("float32"), "std": std.astype("float32")} + + +class S2TDataConfigWriter(object): + DEFAULT_VOCAB_FILENAME = "dict.txt" + DEFAULT_INPUT_FEAT_PER_CHANNEL = 80 + DEFAULT_INPUT_CHANNELS = 1 + + def __init__(self, yaml_path: Path): + try: + import yaml + except ImportError: + print("Please install PyYAML for S2T data config YAML files") + self.yaml = yaml + self.yaml_path = yaml_path + self.config = {} + + def flush(self): + with open(self.yaml_path, "w") as f: + self.yaml.dump(self.config, f) + + def set_audio_root(self, audio_root=""): + self.config["audio_root"] = audio_root + + def set_vocab_filename(self, vocab_filename: str = "dict.txt"): + self.config["vocab_filename"] = vocab_filename + + def set_specaugment( + self, + time_wrap_w: int, + freq_mask_n: int, + freq_mask_f: int, + time_mask_n: int, + time_mask_t: int, + time_mask_p: float, + ): + self.config["specaugment"] = { + "time_wrap_W": time_wrap_w, + "freq_mask_N": freq_mask_n, + "freq_mask_F": freq_mask_f, + "time_mask_N": time_mask_n, + "time_mask_T": time_mask_t, + "time_mask_p": time_mask_p, + } + + def set_specaugment_lb_policy(self): + self.set_specaugment( + time_wrap_w=0, + freq_mask_n=1, + freq_mask_f=27, + time_mask_n=1, + time_mask_t=100, + time_mask_p=1.0, + ) + + def set_specaugment_ld_policy(self): + self.set_specaugment( + time_wrap_w=0, + freq_mask_n=2, + freq_mask_f=27, + time_mask_n=2, + time_mask_t=100, + time_mask_p=1.0, + ) + + def set_specaugment_sm_policy(self): + self.set_specaugment( + time_wrap_w=0, + freq_mask_n=2, + freq_mask_f=15, + time_mask_n=2, + time_mask_t=70, + time_mask_p=0.2, + ) + + def set_specaugment_ss_policy(self): + self.set_specaugment( + time_wrap_w=0, + freq_mask_n=2, + freq_mask_f=27, + time_mask_n=2, + time_mask_t=70, + time_mask_p=0.2, + ) + + def set_input_channels(self, input_channels: int = 1): + self.config["input_channels"] = input_channels + + def set_input_feat_per_channel(self, input_feat_per_channel: int = 80): + self.config["input_feat_per_channel"] = input_feat_per_channel + + def set_bpe_tokenizer(self, bpe_tokenizer: Dict[str, Any]): + self.config["bpe_tokenizer"] = bpe_tokenizer + + def set_global_cmvn(self, stats_npz_path: str): + self.config["global_cmvn"] = {"stats_npz_path": stats_npz_path} + + def set_feature_transforms(self, split: str, transforms: List[str]): + if "transforms" not in self.config: + self.config["transforms"] = {} + self.config["transforms"][split] = transforms + + def set_prepend_tgt_lang_tag(self, flag: bool = True): + self.config["prepend_tgt_lang_tag"] = flag + + def set_sampling_alpha(self, sampling_alpha: float = 1.0): + self.config["sampling_alpha"] = sampling_alpha diff --git a/examples/speech_to_text/docs/covost_example.md b/examples/speech_to_text/docs/covost_example.md new file mode 100644 index 0000000000000000000000000000000000000000..16447f041e4751f79d9f7848b33ef2ff943d63c2 --- /dev/null +++ b/examples/speech_to_text/docs/covost_example.md @@ -0,0 +1,102 @@ +[[Back]](..) + +# S2T Example: ST on CoVoST +We replicate the experiments in +[CoVoST 2 and Massively Multilingual Speech-to-Text Translation (Wang et al., 2020)](https://arxiv.org/abs/2007.10310). + +## Data Preparation +[Download](https://commonvoice.mozilla.org/en/datasets) and unpack Common Voice v4 to a path +`${COVOST_ROOT}/${SOURCE_LANG_ID}`, then preprocess it with +```bash +# additional Python packages for S2T data processing/model training +pip install pandas torchaudio sentencepiece + +# En ASR +python examples/speech_to_text/prep_covost_data.py \ + --data-root ${COVOST_ROOT} --vocab-type char --src-lang en +# ST +python examples/speech_to_text/prep_covost_data.py \ + --data-root ${COVOST_ROOT} --vocab-type char \ + --src-lang fr --tgt-lang en +``` +The generated files (manifest, features, vocabulary and data configuration) will be added to +`${COVOST_ROOT}/${SOURCE_LANG_ID}`. + +Download our vocabulary files if you want to use our pre-trained models: +- ASR: [En](https://dl.fbaipublicfiles.com/fairseq/s2t/covost2_en_asr_vocab_char.zip) +- ST: [Fr-En](https://dl.fbaipublicfiles.com/fairseq/s2t/covost2_fr_en_st_vocab_char.zip), [De-En](https://dl.fbaipublicfiles.com/fairseq/s2t/covost2_de_en_st_vocab_char.zip), [Es-En](https://dl.fbaipublicfiles.com/fairseq/s2t/covost2_es_en_st_vocab_char.zip), [Ca-En](https://dl.fbaipublicfiles.com/fairseq/s2t/covost2_ca_en_st_vocab_char.zip), [En-De](https://dl.fbaipublicfiles.com/fairseq/s2t/covost2_en_de_st_vocab_char.zip), [En-Ca](https://dl.fbaipublicfiles.com/fairseq/s2t/covost2_en_ca_st_vocab_char.zip), [En-Fa](https://dl.fbaipublicfiles.com/fairseq/s2t/covost2_en_fa_st_vocab_char.zip), [En-Et](https://dl.fbaipublicfiles.com/fairseq/s2t/covost2_en_et_st_vocab_char.zip) + +## ASR +#### Training +We train an En ASR model for encoder pre-training of all ST models: +```bash +fairseq-train ${COVOST_ROOT}/en \ + --config-yaml config_asr_en.yaml --train-subset train_asr_en --valid-subset dev_asr_en \ + --save-dir ${ASR_SAVE_DIR} --num-workers 4 --max-tokens 50000 --max-update 60000 \ + --task speech_to_text --criterion label_smoothed_cross_entropy --label-smoothing 0.1 \ + --report-accuracy --arch s2t_transformer_s --dropout 0.15 --optimizer adam --lr 2e-3 \ + --lr-scheduler inverse_sqrt --warmup-updates 10000 --clip-norm 10.0 --seed 1 --update-freq 8 +``` +where `ASR_SAVE_DIR` is the checkpoint root path. We set `--update-freq 8` to simulate 8 GPUs with 1 GPU. +You may want to update it accordingly when using more than 1 GPU. + +#### Inference & Evaluation +```bash +CHECKPOINT_FILENAME=avg_last_10_checkpoint.pt +python scripts/average_checkpoints.py \ + --inputs ${ASR_SAVE_DIR} --num-epoch-checkpoints 10 \ + --output "${ASR_SAVE_DIR}/${CHECKPOINT_FILENAME}" +fairseq-generate ${COVOST_ROOT}/en \ + --config-yaml config_asr_en.yaml --gen-subset test_asr_en --task speech_to_text \ + --path ${ASR_SAVE_DIR}/${CHECKPOINT_FILENAME} --max-tokens 50000 --beam 5 \ + --scoring wer --wer-tokenizer 13a --wer-lowercase --wer-remove-punct +``` +#### Results +| --arch | Params | En | Model | +|---|---|---|---| +| s2t_transformer_s | 31M | 25.6 | [Download](https://dl.fbaipublicfiles.com/fairseq/s2t/covost2_en_asr_transformer_s.pt) | + +## ST +#### Training +Fr-En as example: +```bash +fairseq-train ${COVOST_ROOT}/fr \ + --config-yaml config_st_fr_en.yaml --train-subset train_st_fr_en --valid-subset dev_st_fr_en \ + --save-dir ${ST_SAVE_DIR} --num-workers 4 --max-update 30000 --max-tokens 40000 \ # --max-tokens 50000 for en-* + --task speech_to_text --criterion label_smoothed_cross_entropy --label-smoothing 0.1 --report-accuracy \ + --arch s2t_transformer_s --encoder-freezing-updates 1000 --optimizer adam --lr 2e-3 \ + --lr-scheduler inverse_sqrt --warmup-updates 10000 --clip-norm 10.0 --seed 1 --update-freq 8 \ + --load-pretrained-encoder-from ${ASR_SAVE_DIR}/${CHECKPOINT_FILENAME} +``` +where `ST_SAVE_DIR` is the checkpoint root path. The ST encoder is pre-trained by En ASR for faster training and better +performance: `--load-pretrained-encoder-from <ASR checkpoint path>`. We set `--update-freq 8` to simulate 8 GPUs with 1 GPU. +You may want to update it accordingly when using more than 1 GPU. + +#### Inference & Evaluation +Average the last 10 checkpoints and evaluate on test split: +```bash +CHECKPOINT_FILENAME=avg_last_10_checkpoint.pt +python scripts/average_checkpoints.py \ + --inputs ${ST_SAVE_DIR} --num-epoch-checkpoints 10 \ + --output "${ST_SAVE_DIR}/${CHECKPOINT_FILENAME}" +fairseq-generate ${COVOST_ROOT}/fr \ + --config-yaml config_st_fr_en.yaml --gen-subset test_st_fr_en --task speech_to_text \ + --path ${ST_SAVE_DIR}/${CHECKPOINT_FILENAME} \ + --max-tokens 50000 --beam 5 --scoring sacrebleu +``` + +## Interactive Decoding +Launch the interactive console via +```bash +fairseq-interactive ${COVOST_ROOT}/fr --config-yaml config_st_fr_en.yaml \ + --task speech_to_text --path ${SAVE_DIR}/${CHECKPOINT_FILENAME} \ + --max-tokens 50000 --beam 5 +``` +Type in WAV/FLAC/OGG audio paths (one per line) after the prompt. + +#### Results +| --arch | Params | Fr-En | De-En | Es-En | Ca-En | En-De | En-Ca | En-Fa | En-Et | Model | +|---|---|---|---|---|---|---|---|---|---|---| +| s2t_transformer_s | 31M | [27.2](https://dl.fbaipublicfiles.com/fairseq/s2t/covost2_fr_en_st_transformer_s.pt) | [17.7](https://dl.fbaipublicfiles.com/fairseq/s2t/covost2_de_en_st_transformer_s.pt) | [23.1](https://dl.fbaipublicfiles.com/fairseq/s2t/covost2_es_en_st_transformer_s.pt) | [19.3](https://dl.fbaipublicfiles.com/fairseq/s2t/covost2_ca_en_st_transformer_s.pt) | [16.1](https://dl.fbaipublicfiles.com/fairseq/s2t/covost2_en_de_st_transformer_s.pt) | [21.6](https://dl.fbaipublicfiles.com/fairseq/s2t/covost2_en_ca_st_transformer_s.pt) | [12.9](https://dl.fbaipublicfiles.com/fairseq/s2t/covost2_en_fa_st_transformer_s.pt) | [12.8](https://dl.fbaipublicfiles.com/fairseq/s2t/covost2_en_et_st_transformer_s.pt) | (<-Download) | + +[[Back]](..) diff --git a/examples/speech_to_text/docs/librispeech_example.md b/examples/speech_to_text/docs/librispeech_example.md new file mode 100644 index 0000000000000000000000000000000000000000..4040fda9426027537036ba987d087a43e734bfd9 --- /dev/null +++ b/examples/speech_to_text/docs/librispeech_example.md @@ -0,0 +1,69 @@ +[[Back]](..) + +# S2T Example: Speech Recognition (ASR) on LibriSpeech +[LibriSpeech](https://www.danielpovey.com/files/2015_icassp_librispeech.pdf) is a de-facto standard English ASR +benchmark. We provide competitive +vanilla [Transformer](https://papers.nips.cc/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf) baselines. + +## Data preparation +Download and preprocess LibriSpeech data with +```bash +# additional Python packages for S2T data processing/model training +pip install pandas torchaudio sentencepiece + +python examples/speech_to_text/prep_librispeech_data.py \ + --output-root ${LS_ROOT} --vocab-type unigram --vocab-size 10000 +``` +where `LS_ROOT` is the root path for downloaded data as well as generated files (manifest, features, vocabulary and +data configuration). + +[Download](https://dl.fbaipublicfiles.com/fairseq/s2t/librispeech_vocab_unigram10000.zip) our vocabulary files +if you want to use our pre-trained models. + +## Training +```bash +fairseq-train ${LS_ROOT} --save-dir ${SAVE_DIR} \ + --config-yaml config.yaml --train-subset train-clean-100,train-clean-360,train-other-500 --valid-subset dev-clean,dev-other \ + --num-workers 4 --max-tokens 40000 --max-update 300000 \ + --task speech_to_text --criterion label_smoothed_cross_entropy --label-smoothing 0.1 --report-accuracy \ + --arch s2t_transformer_s --share-decoder-input-output-embed \ + --optimizer adam --lr 2e-3 --lr-scheduler inverse_sqrt --warmup-updates 10000 \ + --clip-norm 10.0 --seed 1 --update-freq 8 +``` +where `SAVE_DIR` is the checkpoint root path. Here we use `--arch s2t_transformer_s` (31M parameters) as example. +For better performance, you may switch to `s2t_transformer_m` (71M, with `--lr 1e-3`) or `s2t_transformer_l` +(268M, with `--lr 5e-4`). We set `--update-freq 8` to simulate 8 GPUs with 1 GPU. You may want to update it accordingly +when using more than 1 GPU. + +## Inference & Evaluation +Average the last 10 checkpoints and evaluate on the 4 splits +(`dev-clean`, `dev-other`, `test-clean` and `test-other`): +```bash +CHECKPOINT_FILENAME=avg_last_10_checkpoint.pt +python scripts/average_checkpoints.py --inputs ${SAVE_DIR} \ + --num-epoch-checkpoints 10 \ + --output "${SAVE_DIR}/${CHECKPOINT_FILENAME}" +for SUBSET in dev-clean dev-other test-clean test-other; do + fairseq-generate ${LS_ROOT} --config-yaml config.yaml --gen-subset ${SUBSET} \ + --task speech_to_text --path ${SAVE_DIR}/${CHECKPOINT_FILENAME} \ + --max-tokens 50000 --beam 5 --scoring wer +done +``` + +## Interactive Decoding +Launch the interactive console via +```bash +fairseq-interactive ${LS_ROOT} --config-yaml config.yaml --task speech_to_text \ + --path ${SAVE_DIR}/${CHECKPOINT_FILENAME} --max-tokens 50000 --beam 5 +``` +Type in WAV/FLAC/OGG audio paths (one per line) after the prompt. + +## Results + +| --arch | Params | dev-clean | dev-other | test-clean | test-other | Model | +|---|---|---|---|---|---|---| +| s2t_transformer_s | 30M | 3.8 | 8.9 | 4.4 | 9.0 | [Download](https://dl.fbaipublicfiles.com/fairseq/s2t/librispeech_transformer_s.pt) | +| s2t_transformer_m | 71M | 3.2 | 8.0 | 3.4 | 7.9 | [Download](https://dl.fbaipublicfiles.com/fairseq/s2t/librispeech_transformer_m.pt) | +| s2t_transformer_l | 268M | 3.0 | 7.5 | 3.2 | 7.5 | [Download](https://dl.fbaipublicfiles.com/fairseq/s2t/librispeech_transformer_l.pt) | + +[[Back]](..) diff --git a/examples/speech_to_text/docs/mtedx_example.md b/examples/speech_to_text/docs/mtedx_example.md new file mode 100644 index 0000000000000000000000000000000000000000..25b4556affbf5bc141b103095d15fffef6225c0e --- /dev/null +++ b/examples/speech_to_text/docs/mtedx_example.md @@ -0,0 +1,200 @@ +[[Back]](..) + +# S2T Example: Speech Translation (ST) on Multilingual TEDx + +[Multilingual TEDx](https://arxiv.org/abs/2102.01757) is multilingual corpus for speech recognition and +speech translation. The data is derived from TEDx talks in 8 source languages +with translations to a subset of 5 target languages. + +## Data Preparation +[Download](http://openslr.org/100/) and unpack Multilingual TEDx data to a path +`${MTEDX_ROOT}/${LANG_PAIR}`, then preprocess it with +```bash +# additional Python packages for S2T data processing/model training +pip install pandas torchaudio soundfile sentencepiece + +# Generate TSV manifests, features, vocabulary +# and configuration for each language +python examples/speech_to_text/prep_mtedx_data.py \ + --data-root ${MTEDX_ROOT} --task asr \ + --vocab-type unigram --vocab-size 1000 +python examples/speech_to_text/prep_mtedx_data.py \ + --data-root ${MTEDX_ROOT} --task st \ + --vocab-type unigram --vocab-size 1000 + +# Add vocabulary and configuration for joint data +# (based on the manifests and features generated above) +python examples/speech_to_text/prep_mtedx_data.py \ + --data-root ${MTEDX_ROOT} --task asr --joint \ + --vocab-type unigram --vocab-size 8000 +python examples/speech_to_text/prep_mtedx_data.py \ + --data-root ${MTEDX_ROOT} --task st --joint \ + --vocab-type unigram --vocab-size 8000 +``` +The generated files (manifest, features, vocabulary and data configuration) will be added to +`${MTEDX_ROOT}/${LANG_PAIR}` (per-language data) and `MTEDX_ROOT` (joint data). + + +## ASR +#### Training +Spanish as example: +```bash +fairseq-train ${MTEDX_ROOT}/es-es \ + --config-yaml config_asr.yaml --train-subset train_asr --valid-subset valid_asr \ + --save-dir ${ASR_SAVE_DIR} --num-workers 4 --max-tokens 40000 --max-epoch 200 \ + --task speech_to_text --criterion label_smoothed_cross_entropy --report-accuracy \ + --arch s2t_transformer_xs --optimizer adam --lr 2e-3 --lr-scheduler inverse_sqrt \ + --warmup-updates 10000 --clip-norm 10.0 --seed 1 --dropout 0.3 --label-smoothing 0.1 \ + --load-pretrained-encoder-from ${PRETRAINED_ENCODER} \ + --skip-invalid-size-inputs-valid-test \ + --keep-last-epochs 10 --update-freq 8 --patience 10 +``` +For joint model (using ASR data from all 8 languages): +```bash +fairseq-train ${MTEDX_ROOT} \ + --config-yaml config_asr.yaml \ + --train-subset train_es-es_asr,train_fr-fr_asr,train_pt-pt_asr,train_it-it_asr,train_ru-ru_asr,train_el-el_asr,train_ar-ar_asr,train_de-de_asr \ + --valid-subset valid_es-es_asr,valid_fr-fr_asr,valid_pt-pt_asr,valid_it-it_asr,valid_ru-ru_asr,valid_el-el_asr,valid_ar-ar_asr,valid_de-de_asr \ + --save-dir ${MULTILINGUAL_ASR_SAVE_DIR} --num-workers 4 --max-tokens 40000 --max-epoch 200 \ + --task speech_to_text --criterion label_smoothed_cross_entropy --report-accuracy \ + --arch s2t_transformer_s --optimizer adam --lr 2e-3 --lr-scheduler inverse_sqrt \ + --warmup-updates 10000 --clip-norm 10.0 --seed 1 --dropout 0.3 --label-smoothing 0.1 \ + --skip-invalid-size-inputs-valid-test \ + --keep-last-epochs 10 --update-freq 8 --patience 10 \ + --ignore-prefix-size 1 +``` +where `MULTILINGUAL_ASR_SAVE_DIR` is the checkpoint root path. We set `--update-freq 8` to simulate 8 GPUs +with 1 GPU. You may want to update it accordingly when using more than 1 GPU. +For multilingual models, we prepend target language ID token as target BOS, which should be excluded from +the training loss via `--ignore-prefix-size 1`. + +#### Inference & Evaluation +```bash +CHECKPOINT_FILENAME=avg_last_10_checkpoint.pt +python scripts/average_checkpoints.py \ + --inputs ${ASR_SAVE_DIR} --num-epoch-checkpoints 10 \ + --output "${ASR_SAVE_DIR}/${CHECKPOINT_FILENAME}" + +fairseq-generate ${MTEDX_ROOT}/es-es \ + --config-yaml config_asr.yaml --gen-subset test --task speech_to_text \ + --path ${ASR_SAVE_DIR}/${CHECKPOINT_FILENAME} --max-tokens 50000 --beam 5 \ + --skip-invalid-size-inputs-valid-test \ + --scoring wer --wer-tokenizer 13a --wer-lowercase --wer-remove-punct --remove-bpe + +# For models trained on joint data +CHECKPOINT_FILENAME=avg_last_10_checkpoint.pt +python scripts/average_checkpoints.py \ + --inputs ${MULTILINGUAL_ASR_SAVE_DIR} --num-epoch-checkpoints 10 \ + --output "${MULTILINGUAL_ASR_SAVE_DIR}/${CHECKPOINT_FILENAME}" + +for LANG in es fr pt it ru el ar de; do + fairseq-generate ${MTEDX_ROOT} \ + --config-yaml config_asr.yaml --gen-subset test_${LANG}-${LANG}_asr --task speech_to_text \ + --prefix-size 1 --path ${MULTILINGUAL_ASR_SAVE_DIR}/${CHECKPOINT_FILENAME} \ + --max-tokens 40000 --beam 5 \ + --skip-invalid-size-inputs-valid-test \ + --scoring wer --wer-tokenizer 13a --wer-lowercase --wer-remove-punct --remove-bpe +done +``` +#### Results +| Data | --arch | Params | Es | Fr | Pt | It | Ru | El | Ar | De | +|--------------|--------------------|--------|------|------|------|------|------|-------|-------|-------| +| Monolingual | s2t_transformer_xs | 10M | 46.4 | 45.6 | 54.8 | 48.0 | 74.7 | 109.5 | 104.4 | 111.1 | + + +## ST +#### Training +Es-En as example: +```bash +fairseq-train ${MTEDX_ROOT}/es-en \ + --config-yaml config_st.yaml --train-subset train_st --valid-subset valid_st \ + --save-dir ${ST_SAVE_DIR} --num-workers 4 --max-tokens 40000 --max-epoch 200 \ + --task speech_to_text --criterion label_smoothed_cross_entropy --report-accuracy \ + --arch s2t_transformer_xs --optimizer adam --lr 2e-3 --lr-scheduler inverse_sqrt \ + --warmup-updates 10000 --clip-norm 10.0 --seed 1 --dropout 0.3 --label-smoothing 0.1 \ + --load-pretrained-encoder-from ${PRETRAINED_ENCODER} \ + --skip-invalid-size-inputs-valid-test \ + --keep-last-epochs 10 --update-freq 8 --patience 10 +``` +For multilingual model (all 12 directions): +```bash +fairseq-train ${MTEDX_ROOT} \ + --config-yaml config_st.yaml \ + --train-subset train_el-en_st,train_es-en_st,train_es-fr_st,train_es-it_st,train_es-pt_st,train_fr-en_st,train_fr-es_st,train_fr-pt_st,train_it-en_st,train_it-es_st,train_pt-en_st,train_pt-es_st,train_ru-en_st \ + --valid-subset valid_el-en_st,valid_es-en_st,valid_es-fr_st,valid_es-it_st,valid_es-pt_st,valid_fr-en_st,valid_fr-es_st,valid_fr-pt_st,valid_it-en_st,valid_it-es_st,valid_pt-en_st,valid_pt-es_st,valid_ru-en_st \ + --save-dir ${MULTILINGUAL_ST_SAVE_DIR} --num-workers 4 --max-tokens 40000 --max-epoch 200 \ + --task speech_to_text --criterion label_smoothed_cross_entropy --report-accuracy \ + --arch s2t_transformer_s --optimizer adam --lr 2e-3 --lr-scheduler inverse_sqrt \ + --warmup-updates 10000 --clip-norm 10.0 --seed 1 --dropout 0.3 --label-smoothing 0.1 \ + --skip-invalid-size-inputs-valid-test \ + --keep-last-epochs 10 --update-freq 8 --patience 10 \ + --ignore-prefix-size 1 \ + --load-pretrained-encoder-from ${PRETRAINED_ENCODER} +``` +where `ST_SAVE_DIR` (`MULTILINGUAL_ST_SAVE_DIR`) is the checkpoint root path. The ST encoder is pre-trained by ASR +for faster training and better performance: `--load-pretrained-encoder-from <(JOINT_)ASR checkpoint path>`. We set +`--update-freq 8` to simulate 8 GPUs with 1 GPU. You may want to update it accordingly when using more than 1 GPU. +For multilingual models, we prepend target language ID token as target BOS, which should be excluded from +the training loss via `--ignore-prefix-size 1`. + +#### Inference & Evaluation +Average the last 10 checkpoints and evaluate on the `test` split: +```bash +CHECKPOINT_FILENAME=avg_last_10_checkpoint.pt +python scripts/average_checkpoints.py \ + --inputs ${ST_SAVE_DIR} --num-epoch-checkpoints 10 \ + --output "${ST_SAVE_DIR}/${CHECKPOINT_FILENAME}" + +fairseq-generate ${MTEDX_ROOT}/es-en \ + --config-yaml config_st.yaml --gen-subset test --task speech_to_text \ + --path ${ST_SAVE_DIR}/${CHECKPOINT_FILENAME} \ + --max-tokens 50000 --beam 5 --scoring sacrebleu --remove-bpe + +# For multilingual models +python scripts/average_checkpoints.py \ + --inputs ${MULTILINGUAL_ST_SAVE_DIR} --num-epoch-checkpoints 10 \ + --output "${MULTILINGUAL_ST_SAVE_DIR}/${CHECKPOINT_FILENAME}" + +for LANGPAIR in es-en es-fr es-pt fr-en fr-es fr-pt pt-en pt-es it-en it-es ru-en el-en; do + fairseq-generate ${MTEDX_ROOT} \ + --config-yaml config_st.yaml --gen-subset test_${LANGPAIR}_st --task speech_to_text \ + --prefix-size 1 --path ${MULTILINGUAL_ST_SAVE_DIR}/${CHECKPOINT_FILENAME} \ + --max-tokens 40000 --beam 5 \ + --skip-invalid-size-inputs-valid-test \ + --scoring sacrebleu --remove-bpe +done +``` +For multilingual models, we force decoding from the target language ID token (as BOS) via `--prefix-size 1`. + +#### Results +| Data | --arch | Params | Es-En | Es-Pt | Es-Fr | Fr-En | Fr-Es | Fr-Pt | Pt-En | Pt-Es | It-En | It-Es | Ru-En | El-En | +|--------------|--------------------|-----|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------| +| Bilingual | s2t_transformer_xs | 10M | 7.0 | 12.2 | 1.7 | 8.9 | 10.6 | 7.9 | 8.1 | 8.7 | 6.4 | 1.0 | 0.7 | 0.6 | +| Multilingual | s2t_transformer_s | 31M | 12.3 | 17.4 | 6.1 | 12.0 | 13.6 | 13.2 | 12.0 | 13.7 | 10.7 | 13.1 | 0.6 | 0.8 | + + +## Citation +Please cite as: +``` +@misc{salesky2021mtedx, + title={Multilingual TEDx Corpus for Speech Recognition and Translation}, + author={Elizabeth Salesky and Matthew Wiesner and Jacob Bremerman and Roldano Cattoni and Matteo Negri and Marco Turchi and Douglas W. Oard and Matt Post}, + year={2021}, +} + +@inproceedings{wang2020fairseqs2t, + title = {fairseq S2T: Fast Speech-to-Text Modeling with fairseq}, + author = {Changhan Wang and Yun Tang and Xutai Ma and Anne Wu and Dmytro Okhonko and Juan Pino}, + booktitle = {Proceedings of the 2020 Conference of the Asian Chapter of the Association for Computational Linguistics (AACL): System Demonstrations}, + year = {2020}, +} + +@inproceedings{ott2019fairseq, + title = {fairseq: A Fast, Extensible Toolkit for Sequence Modeling}, + author = {Myle Ott and Sergey Edunov and Alexei Baevski and Angela Fan and Sam Gross and Nathan Ng and David Grangier and Michael Auli}, + booktitle = {Proceedings of NAACL-HLT 2019: Demonstrations}, + year = {2019}, +} +``` + +[[Back]](..) diff --git a/examples/speech_to_text/docs/mustc_example.md b/examples/speech_to_text/docs/mustc_example.md new file mode 100644 index 0000000000000000000000000000000000000000..c95ef3e15660107c3384f87c1680f005044e7f3b --- /dev/null +++ b/examples/speech_to_text/docs/mustc_example.md @@ -0,0 +1,155 @@ +[[Back]](..) + +# S2T Example: Speech Translation (ST) on MuST-C + +[MuST-C](https://www.aclweb.org/anthology/N19-1202) is multilingual speech-to-text translation corpus with +8-language translations on English TED talks. We match the state-of-the-art performance in +[ESPNet-ST](https://arxiv.org/pdf/2004.10234.pdf) with a simpler model training pipeline. + +## Data Preparation +[Download](https://ict.fbk.eu/must-c) and unpack MuST-C data to a path +`${MUSTC_ROOT}/en-${TARGET_LANG_ID}`, then preprocess it with +```bash +# additional Python packages for S2T data processing/model training +pip install pandas torchaudio soundfile sentencepiece + +# Generate TSV manifests, features, vocabulary +# and configuration for each language +python examples/speech_to_text/prep_mustc_data.py \ + --data-root ${MUSTC_ROOT} --task asr \ + --vocab-type unigram --vocab-size 5000 +python examples/speech_to_text/prep_mustc_data.py \ + --data-root ${MUSTC_ROOT} --task st \ + --vocab-type unigram --vocab-size 8000 + +# Add vocabulary and configuration for joint data +# (based on the manifests and features generated above) +python examples/speech_to_text/prep_mustc_data.py \ + --data-root ${MUSTC_ROOT} --task asr --joint \ + --vocab-type unigram --vocab-size 10000 +python examples/speech_to_text/prep_mustc_data.py \ + --data-root ${MUSTC_ROOT} --task st --joint \ + --vocab-type unigram --vocab-size 10000 +``` +The generated files (manifest, features, vocabulary and data configuration) will be added to +`${MUSTC_ROOT}/en-${TARGET_LANG_ID}` (per-language data) and `MUSTC_ROOT` (joint data). + +Download our vocabulary files if you want to use our pre-trained models: +- ASR: [En-De](https://dl.fbaipublicfiles.com/fairseq/s2t/mustc_de_asr_vocab_unigram5000.zip), [En-Nl](https://dl.fbaipublicfiles.com/fairseq/s2t/mustc_nl_asr_vocab_unigram5000.zip), [En-Es](https://dl.fbaipublicfiles.com/fairseq/s2t/mustc_es_asr_vocab_unigram5000.zip), [En-Fr](https://dl.fbaipublicfiles.com/fairseq/s2t/mustc_fr_asr_vocab_unigram5000.zip), [En-It](https://dl.fbaipublicfiles.com/fairseq/s2t/mustc_it_asr_vocab_unigram5000.zip), [En-Pt](https://dl.fbaipublicfiles.com/fairseq/s2t/mustc_pt_asr_vocab_unigram5000.zip), [En-Ro](https://dl.fbaipublicfiles.com/fairseq/s2t/mustc_ro_asr_vocab_unigram5000.zip), [En-Ru](https://dl.fbaipublicfiles.com/fairseq/s2t/mustc_ru_asr_vocab_unigram5000.zip), [Joint](https://dl.fbaipublicfiles.com/fairseq/s2t/mustc_joint_asr_vocab_unigram10000.zip) +- ST: [En-De](https://dl.fbaipublicfiles.com/fairseq/s2t/mustc_de_st_vocab_unigram8000.zip), [En-Nl](https://dl.fbaipublicfiles.com/fairseq/s2t/mustc_nl_st_vocab_unigram8000.zip), [En-Es](https://dl.fbaipublicfiles.com/fairseq/s2t/mustc_es_st_vocab_unigram8000.zip), [En-Fr](https://dl.fbaipublicfiles.com/fairseq/s2t/mustc_fr_st_vocab_unigram8000.zip), [En-It](https://dl.fbaipublicfiles.com/fairseq/s2t/mustc_it_st_vocab_unigram8000.zip), [En-Pt](https://dl.fbaipublicfiles.com/fairseq/s2t/mustc_pt_st_vocab_unigram8000.zip), [En-Ro](https://dl.fbaipublicfiles.com/fairseq/s2t/mustc_ro_st_vocab_unigram8000.zip), [En-Ru](https://dl.fbaipublicfiles.com/fairseq/s2t/mustc_ru_st_vocab_unigram8000.zip), [Multilingual](https://dl.fbaipublicfiles.com/fairseq/s2t/mustc_multilingual_st_vocab_unigram10000.zip) + +## ASR +#### Training +En-De as example: +```bash +fairseq-train ${MUSTC_ROOT}/en-de \ + --config-yaml config_asr.yaml --train-subset train_asr --valid-subset dev_asr \ + --save-dir ${ASR_SAVE_DIR} --num-workers 4 --max-tokens 40000 --max-update 100000 \ + --task speech_to_text --criterion label_smoothed_cross_entropy --label-smoothing 0.1 --report-accuracy \ + --arch s2t_transformer_s --optimizer adam --lr 1e-3 --lr-scheduler inverse_sqrt \ + --warmup-updates 10000 --clip-norm 10.0 --seed 1 --update-freq 8 +``` +For joint model (using ASR data from all 8 directions): +```bash +fairseq-train ${MUSTC_ROOT} \ + --config-yaml config_asr.yaml \ + --train-subset train_de_asr,train_nl_asr,train_es_asr,train_fr_asr,train_it_asr,train_pt_asr,train_ro_asr,train_ru_asr \ + --valid-subset dev_de_asr,dev_nl_asr,dev_es_asr,dev_fr_asr,dev_it_asr,dev_pt_asr,dev_ro_asr,dev_ru_asr \ + --save-dir ${JOINT_ASR_SAVE_DIR} --num-workers 4 --max-tokens 40000 --max-update 100000 \ + --task speech_to_text --criterion label_smoothed_cross_entropy --label-smoothing 0.1 --report-accuracy \ + --arch s2t_transformer_s --optimizer adam --lr 1e-3 --lr-scheduler inverse_sqrt \ + --warmup-updates 10000 --clip-norm 10.0 --seed 1 --update-freq 8 +``` +where `ASR_SAVE_DIR` (`JOINT_ASR_SAVE_DIR`) is the checkpoint root path. We set `--update-freq 8` to simulate 8 GPUs +with 1 GPU. You may want to update it accordingly when using more than 1 GPU. + +#### Inference & Evaluation +```bash +CHECKPOINT_FILENAME=avg_last_10_checkpoint.pt +python scripts/average_checkpoints.py \ + --inputs ${ASR_SAVE_DIR} --num-epoch-checkpoints 10 \ + --output "${ASR_SAVE_DIR}/${CHECKPOINT_FILENAME}" +fairseq-generate ${MUSTC_ROOT}/en-de \ + --config-yaml config_asr.yaml --gen-subset tst-COMMON_asr --task speech_to_text \ + --path ${ASR_SAVE_DIR}/${CHECKPOINT_FILENAME} --max-tokens 50000 --beam 5 \ + --scoring wer --wer-tokenizer 13a --wer-lowercase --wer-remove-punct + +# For models trained on joint data +python scripts/average_checkpoints.py \ + --inputs ${JOINT_ASR_SAVE_DIR} --num-epoch-checkpoints 10 \ + --output "${JOINT_ASR_SAVE_DIR}/${CHECKPOINT_FILENAME}" +for LANG in de nl es fr it pt ro ru; do + fairseq-generate ${MUSTC_ROOT} \ + --config-yaml config_asr.yaml --gen-subset tst-COMMON_${LANG}_asr --task speech_to_text \ + --path ${JOINT_ASR_SAVE_DIR}/${CHECKPOINT_FILENAME} --max-tokens 50000 --beam 5 \ + --scoring wer --wer-tokenizer 13a --wer-lowercase --wer-remove-punct +done +``` +#### Results +| Data | --arch | Params | En-De | En-Nl | En-Es | En-Fr | En-It | En-Pt | En-Ro | En-Ru | Model | +|---|---|---|---|---|---|---|---|---|---|---|---| +| Single | s2t_transformer_s | 31M | [18.2](https://dl.fbaipublicfiles.com/fairseq/s2t/mustc_de_asr_transformer_s.pt) | [17.6](https://dl.fbaipublicfiles.com/fairseq/s2t/mustc_nl_asr_transformer_s.pt) | [17.7](https://dl.fbaipublicfiles.com/fairseq/s2t/mustc_es_asr_transformer_s.pt) | [17.2](https://dl.fbaipublicfiles.com/fairseq/s2t/mustc_fr_asr_transformer_s.pt) | [17.9](https://dl.fbaipublicfiles.com/fairseq/s2t/mustc_it_asr_transformer_s.pt) | [19.1](https://dl.fbaipublicfiles.com/fairseq/s2t/mustc_pt_asr_transformer_s.pt) | [18.1](https://dl.fbaipublicfiles.com/fairseq/s2t/mustc_ro_asr_transformer_s.pt) | [17.7](https://dl.fbaipublicfiles.com/fairseq/s2t/mustc_ru_asr_transformer_s.pt) | (<-Download) | +| Joint | s2t_transformer_m | 76M | 16.8 | 16.7 | 16.9 | 16.9 | 17.0 | 17.4 | 17.0 | 16.9 | [Download](https://dl.fbaipublicfiles.com/fairseq/s2t/mustc_joint_asr_transformer_m.pt) | + +## ST +#### Training +En-De as example: +```bash +fairseq-train ${MUSTC_ROOT}/en-de \ + --config-yaml config_st.yaml --train-subset train_st --valid-subset dev_st \ + --save-dir ${ST_SAVE_DIR} --num-workers 4 --max-tokens 40000 --max-update 100000 \ + --task speech_to_text --criterion label_smoothed_cross_entropy --label-smoothing 0.1 --report-accuracy \ + --arch s2t_transformer_s --optimizer adam --lr 2e-3 --lr-scheduler inverse_sqrt \ + --warmup-updates 10000 --clip-norm 10.0 --seed 1 --update-freq 8 \ + --load-pretrained-encoder-from ${ASR_SAVE_DIR}/${CHECKPOINT_FILENAME} +``` +For multilingual model (all 8 directions): +```bash +fairseq-train ${MUSTC_ROOT} \ + --config-yaml config_st.yaml \ + --train-subset train_de_st,train_nl_st,train_es_st,train_fr_st,train_it_st,train_pt_st,train_ro_st,train_ru_st \ + --valid-subset dev_de_st,dev_nl_st,dev_es_st,dev_fr_st,dev_it_st,dev_pt_st,dev_ro_st,dev_ru_st \ + --save-dir ${MULTILINGUAL_ST_SAVE_DIR} --num-workers 4 --max-tokens 40000 --max-update 100000 \ + --task speech_to_text --criterion label_smoothed_cross_entropy --label-smoothing 0.1 --report-accuracy \ + --arch s2t_transformer_s --ignore-prefix-size 1 --optimizer adam --lr 2e-3 --lr-scheduler inverse_sqrt \ + --warmup-updates 10000 --clip-norm 10.0 --seed 1 --update-freq 8 \ + --load-pretrained-encoder-from ${JOINT_ASR_SAVE_DIR}/${CHECKPOINT_FILENAME} +``` +where `ST_SAVE_DIR` (`MULTILINGUAL_ST_SAVE_DIR`) is the checkpoint root path. The ST encoder is pre-trained by ASR +for faster training and better performance: `--load-pretrained-encoder-from <(JOINT_)ASR checkpoint path>`. We set +`--update-freq 8` to simulate 8 GPUs with 1 GPU. You may want to update it accordingly when using more than 1 GPU. +For multilingual models, we prepend target language ID token as target BOS, which should be excluded from +the training loss via `--ignore-prefix-size 1`. + +#### Inference & Evaluation +Average the last 10 checkpoints and evaluate on the `tst-COMMON` split: +```bash +CHECKPOINT_FILENAME=avg_last_10_checkpoint.pt +python scripts/average_checkpoints.py \ + --inputs ${ST_SAVE_DIR} --num-epoch-checkpoints 10 \ + --output "${ST_SAVE_DIR}/${CHECKPOINT_FILENAME}" +fairseq-generate ${MUSTC_ROOT}/en-de \ + --config-yaml config_st.yaml --gen-subset tst-COMMON_st --task speech_to_text \ + --path ${ST_SAVE_DIR}/${CHECKPOINT_FILENAME} \ + --max-tokens 50000 --beam 5 --scoring sacrebleu + +# For multilingual models +python scripts/average_checkpoints.py \ + --inputs ${MULTILINGUAL_ST_SAVE_DIR} --num-epoch-checkpoints 10 \ + --output "${MULTILINGUAL_ST_SAVE_DIR}/${CHECKPOINT_FILENAME}" +for LANG in de nl es fr it pt ro ru; do + fairseq-generate ${MUSTC_ROOT} \ + --config-yaml config_st.yaml --gen-subset tst-COMMON_${LANG}_st --task speech_to_text \ + --prefix-size 1 --path ${MULTILINGUAL_ST_SAVE_DIR}/${CHECKPOINT_FILENAME} \ + --max-tokens 50000 --beam 5 --scoring sacrebleu +done +``` +For multilingual models, we force decoding from the target language ID token (as BOS) via `--prefix-size 1`. + +#### Results +| Data | --arch | Params | En-De | En-Nl | En-Es | En-Fr | En-It | En-Pt | En-Ro | En-Ru | Model | +|---|---|---|---|---|---|---|---|---|---|---|---| +| Bilingual | s2t_transformer_s | 31M | [22.7](https://dl.fbaipublicfiles.com/fairseq/s2t/mustc_de_st_transformer_s.pt) | [27.3](https://dl.fbaipublicfiles.com/fairseq/s2t/mustc_nl_st_transformer_s.pt) | [27.2](https://dl.fbaipublicfiles.com/fairseq/s2t/mustc_es_st_transformer_s.pt) | [32.9](https://dl.fbaipublicfiles.com/fairseq/s2t/mustc_fr_st_transformer_s.pt) | [22.7](https://dl.fbaipublicfiles.com/fairseq/s2t/mustc_it_st_transformer_s.pt) | [28.1](https://dl.fbaipublicfiles.com/fairseq/s2t/mustc_pt_st_transformer_s.pt) | [21.9](https://dl.fbaipublicfiles.com/fairseq/s2t/mustc_ro_st_transformer_s.pt) | [15.3](https://dl.fbaipublicfiles.com/fairseq/s2t/mustc_ru_st_transformer_s.pt) | (<-Download) | +| Multilingual | s2t_transformer_m | 76M | 24.5 | 28.6 | 28.2 | 34.9 | 24.6 | 31.1 | 23.8 | 16.0 | [Download](https://dl.fbaipublicfiles.com/fairseq/s2t/mustc_multilingual_st_transformer_m.pt) | + +[[Back]](..) diff --git a/examples/speech_to_text/docs/simulst_mustc_example.md b/examples/speech_to_text/docs/simulst_mustc_example.md new file mode 100644 index 0000000000000000000000000000000000000000..52ca9ac0625b0da6b5202111b247cd91cc531be7 --- /dev/null +++ b/examples/speech_to_text/docs/simulst_mustc_example.md @@ -0,0 +1,190 @@ +# Simultaneous Speech Translation (SimulST) on MuST-C + +This is a tutorial of training and evaluating a transformer *wait-k* simultaneous model on MUST-C English-Germen Dataset, from [SimulMT to SimulST: Adapting Simultaneous Text Translation to End-to-End Simultaneous Speech Translation](https://www.aclweb.org/anthology/2020.aacl-main.58.pdf). + +[MuST-C](https://www.aclweb.org/anthology/N19-1202) is multilingual speech-to-text translation corpus with 8-language translations on English TED talks. + +## Data Preparation +This section introduces the data preparation for training and evaluation. +If you only want to evaluate the model, please jump to [Inference & Evaluation](#inference-&-evaluation) + +[Download](https://ict.fbk.eu/must-c) and unpack MuST-C data to a path +`${MUSTC_ROOT}/en-${TARGET_LANG_ID}`, then preprocess it with +```bash +# Additional Python packages for S2T data processing/model training +pip install pandas torchaudio sentencepiece + +# Generate TSV manifests, features, vocabulary, +# global cepstral and mean estimation, +# and configuration for each language +cd fairseq + +python examples/speech_to_text/prep_mustc_data.py \ + --data-root ${MUSTC_ROOT} --task asr \ + --vocab-type unigram --vocab-size 10000 \ + --cmvn-type global + +python examples/speech_to_text/prep_mustc_data.py \ + --data-root ${MUSTC_ROOT} --task st \ + --vocab-type unigram --vocab-size 10000 \ + --cmvn-type global +``` + +## ASR Pretraining +We need a pretrained offline ASR model. Assuming the save directory of the ASR model is `${ASR_SAVE_DIR}`. +The following command (and the subsequent training commands in this tutorial) assume training on 1 GPU (you can also train on 8 GPUs and remove the `--update-freq 8` option). +``` +fairseq-train ${MUSTC_ROOT}/en-de \ + --config-yaml config_asr.yaml --train-subset train_asr --valid-subset dev_asr \ + --save-dir ${ASR_SAVE_DIR} --num-workers 4 --max-tokens 40000 --max-update 100000 \ + --task speech_to_text --criterion label_smoothed_cross_entropy --report-accuracy \ + --arch convtransformer_espnet --optimizer adam --lr 0.0005 --lr-scheduler inverse_sqrt \ + --warmup-updates 10000 --clip-norm 10.0 --seed 1 --update-freq 8 +``` +A pretrained ASR checkpoint can be downloaded [here](https://dl.fbaipublicfiles.com/simultaneous_translation/must_c_v1_en_de_pretrained_asr) + +## Simultaneous Speech Translation Training + +### Wait-K with fixed pre-decision module +Fixed pre-decision indicates that the model operate simultaneous policy on the boundaries of fixed chunks. +Here is a example of fixed pre-decision ratio 7 (the simultaneous decision is made every 7 encoder states) and +a wait-3 policy model. Assuming the save directory is `${ST_SAVE_DIR}` +```bash + fairseq-train ${MUSTC_ROOT}/en-de \ + --config-yaml config_st.yaml --train-subset train_st --valid-subset dev_st \ + --save-dir ${ST_SAVE_DIR} --num-workers 8 \ + --optimizer adam --lr 0.0001 --lr-scheduler inverse_sqrt --clip-norm 10.0 \ + --criterion label_smoothed_cross_entropy \ + --warmup-updates 4000 --max-update 100000 --max-tokens 40000 --seed 2 \ + --load-pretrained-encoder-from ${ASR_SAVE_DIR}/checkpoint_best.pt \ + --task speech_to_text \ + --arch convtransformer_simul_trans_espnet \ + --simul-type waitk_fixed_pre_decision \ + --waitk-lagging 3 \ + --fixed-pre-decision-ratio 7 \ + --update-freq 8 + +``` +### Monotonic multihead attention with fixed pre-decision module +``` + fairseq-train ${MUSTC_ROOT}/en-de \ + --config-yaml config_st.yaml --train-subset train_st --valid-subset dev_st \ + --save-dir ${ST_SAVE_DIR} --num-workers 8 \ + --optimizer adam --lr 0.0001 --lr-scheduler inverse_sqrt --clip-norm 10.0 \ + --warmup-updates 4000 --max-update 100000 --max-tokens 40000 --seed 2 \ + --load-pretrained-encoder-from ${ASR_SAVE_DIR}/${CHECKPOINT_FILENAME} \ + --task speech_to_text \ + --criterion latency_augmented_label_smoothed_cross_entropy \ + --latency-weight-avg 0.1 \ + --arch convtransformer_simul_trans_espnet \ + --simul-type infinite_lookback_fixed_pre_decision \ + --fixed-pre-decision-ratio 7 \ + --update-freq 8 +``` +## Inference & Evaluation +[SimulEval](https://github.com/facebookresearch/SimulEval) is used for evaluation. +The following command is for evaluation. + +``` +git clone https://github.com/facebookresearch/SimulEval.git +cd SimulEval +pip install -e . + +simuleval \ + --agent ${FAIRSEQ}/examples/speech_to_text/simultaneous_translation/agents/fairseq_simul_st_agent.py + --source ${SRC_LIST_OF_AUDIO} + --target ${TGT_FILE} + --data-bin ${MUSTC_ROOT}/en-de \ + --config config_st.yaml \ + --model-path ${ST_SAVE_DIR}/${CHECKPOINT_FILENAME} \ + --output ${OUTPUT} \ + --scores +``` + +The source file `${SRC_LIST_OF_AUDIO}` is a list of paths of audio files. Assuming your audio files stored at `/home/user/data`, +it should look like this + +```bash +/home/user/data/audio-1.wav +/home/user/data/audio-2.wav +``` + +Each line of target file `${TGT_FILE}` is the translation for each audio file input. +```bash +Translation_1 +Translation_2 +``` +The evaluation runs on the original MUSTC segmentation. +The following command will generate the wav list and text file for a evaluation set `${SPLIT}` (chose from `dev`, `tst-COMMON` and `tst-HE`) in MUSTC to `${EVAL_DATA}`. +```bash +python ${FAIRSEQ}/examples/speech_to_text/seg_mustc_data.py \ + --data-root ${MUSTC_ROOT} --lang de \ + --split ${SPLIT} --task st \ + --output ${EVAL_DATA} +``` + +The `--data-bin` and `--config` should be the same in previous section if you prepare the data from the scratch. +If only for evaluation, a prepared data directory can be found [here](https://dl.fbaipublicfiles.com/simultaneous_translation/must_c_v1.0_en_de_databin.tgz). It contains +- `spm_unigram10000_st.model`: a sentencepiece model binary. +- `spm_unigram10000_st.txt`: the dictionary file generated by the sentencepiece model. +- `gcmvn.npz`: the binary for global cepstral mean and variance. +- `config_st.yaml`: the config yaml file. It looks like this. +You will need to set the absolute paths for `sentencepiece_model` and `stats_npz_path` if the data directory is downloaded. +```yaml +bpe_tokenizer: + bpe: sentencepiece + sentencepiece_model: ABS_PATH_TO_SENTENCEPIECE_MODEL +global_cmvn: + stats_npz_path: ABS_PATH_TO_GCMVN_FILE +input_channels: 1 +input_feat_per_channel: 80 +sampling_alpha: 1.0 +specaugment: + freq_mask_F: 27 + freq_mask_N: 1 + time_mask_N: 1 + time_mask_T: 100 + time_mask_p: 1.0 + time_wrap_W: 0 +transforms: + '*': + - global_cmvn + _train: + - global_cmvn + - specaugment +vocab_filename: spm_unigram10000_st.txt +``` + +Notice that once a `--data-bin` is set, the `--config` is the base name of the config yaml, not the full path. + +Set `--model-path` to the model checkpoint. +A pretrained checkpoint can be downloaded from [here](https://dl.fbaipublicfiles.com/simultaneous_translation/convtransformer_wait5_pre7), which is a wait-5 model with a pre-decision of 280 ms. + +The result of this model on `tst-COMMON` is: +```bash +{ + "Quality": { + "BLEU": 13.94974229366959 + }, + "Latency": { + "AL": 1751.8031870037803, + "AL_CA": 2338.5911762796536, + "AP": 0.7931395378788959, + "AP_CA": 0.9405103863210942, + "DAL": 1987.7811616943081, + "DAL_CA": 2425.2751560926167 + } +} +``` + +If `--output ${OUTPUT}` option is used, the detailed log and scores will be stored under the `${OUTPUT}` directory. + + +The quality is measured by detokenized BLEU. So make sure that the predicted words sent to the server are detokenized. + +The latency metrics are +* Average Proportion +* Average Lagging +* Differentiable Average Lagging + +Again they will also be evaluated on detokenized text. diff --git a/examples/speech_to_text/prep_covost_data.py b/examples/speech_to_text/prep_covost_data.py new file mode 100644 index 0000000000000000000000000000000000000000..af1d3fc6b854c961dc460b4db23e86ff4fcbcbf3 --- /dev/null +++ b/examples/speech_to_text/prep_covost_data.py @@ -0,0 +1,280 @@ +#!/usr/bin/env python3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import logging +from pathlib import Path +import shutil +from tempfile import NamedTemporaryFile +from typing import Optional, Tuple + +import pandas as pd +import torchaudio +from examples.speech_to_text.data_utils import ( + create_zip, + extract_fbank_features, + filter_manifest_df, + gen_config_yaml, + gen_vocab, + get_zip_manifest, + load_df_from_tsv, + save_df_to_tsv, +) +from torch import Tensor +from torch.utils.data import Dataset +from torchaudio.datasets.utils import download_url, extract_archive +from tqdm import tqdm + + +log = logging.getLogger(__name__) + + +MANIFEST_COLUMNS = ["id", "audio", "n_frames", "tgt_text", "speaker"] + + +class CoVoST(Dataset): + """Create a Dataset for CoVoST (https://github.com/facebookresearch/covost). + + Args: + root (str): root path to the dataset and generated manifests/features + source_language (str): source (audio) language + target_language (str, optional): target (text) language, + None for no translation (default: None) + version (int, optional): CoVoST version. (default: 2) + download (bool, optional): Whether to download the dataset if it is not + found at root path. (default: ``False``). + """ + + COVOST_URL_TEMPLATE = ( + "https://dl.fbaipublicfiles.com/covost/" + "covost_v2.{src_lang}_{tgt_lang}.tsv.tar.gz" + ) + + VERSIONS = {2} + SPLITS = ["train", "dev", "test"] + + XX_EN_LANGUAGES = { + 1: ["fr", "de", "nl", "ru", "es", "it", "tr", "fa", "sv-SE", "mn", "zh-CN"], + 2: [ + "fr", + "de", + "es", + "ca", + "it", + "ru", + "zh-CN", + "pt", + "fa", + "et", + "mn", + "nl", + "tr", + "ar", + "sv-SE", + "lv", + "sl", + "ta", + "ja", + "id", + "cy", + ], + } + EN_XX_LANGUAGES = { + 1: [], + 2: [ + "de", + "tr", + "fa", + "sv-SE", + "mn", + "zh-CN", + "cy", + "ca", + "sl", + "et", + "id", + "ar", + "ta", + "lv", + "ja", + ], + } + + def __init__( + self, + root: str, + split: str, + source_language: str, + target_language: Optional[str] = None, + version: int = 2, + ) -> None: + assert version in self.VERSIONS and split in self.SPLITS + assert source_language is not None + self.no_translation = target_language is None + if not self.no_translation: + assert "en" in {source_language, target_language} + if source_language == "en": + assert target_language in self.EN_XX_LANGUAGES[version] + else: + assert source_language in self.XX_EN_LANGUAGES[version] + else: + # Hack here so that we can get "split" column from CoVoST TSV. + # Note that we use CoVoST train split for ASR which is an extension + # to Common Voice train split. + target_language = "de" if source_language == "en" else "en" + + self.root: Path = Path(root) + + cv_tsv_path = self.root / "validated.tsv" + assert cv_tsv_path.is_file() + + covost_url = self.COVOST_URL_TEMPLATE.format( + src_lang=source_language, tgt_lang=target_language + ) + covost_archive = self.root / Path(covost_url).name + if not covost_archive.is_file(): + download_url(covost_url, self.root.as_posix(), hash_value=None) + extract_archive(covost_archive.as_posix()) + + cv_tsv = load_df_from_tsv(cv_tsv_path) + covost_tsv = load_df_from_tsv( + self.root / Path(covost_url).name.replace(".tar.gz", "") + ) + df = pd.merge( + left=cv_tsv[["path", "sentence", "client_id"]], + right=covost_tsv[["path", "translation", "split"]], + how="inner", + on="path", + ) + if split == "train": + df = df[(df["split"] == split) | (df["split"] == f"{split}_covost")] + else: + df = df[df["split"] == split] + data = df.to_dict(orient="index").items() + data = [v for k, v in sorted(data, key=lambda x: x[0])] + self.data = [] + for e in data: + try: + path = self.root / "clips" / e["path"] + _ = torchaudio.info(path.as_posix()) + self.data.append(e) + except RuntimeError: + pass + + def __getitem__( + self, n: int + ) -> Tuple[Tensor, int, str, str, Optional[str], str, str]: + """Load the n-th sample from the dataset. + + Args: + n (int): The index of the sample to be loaded + + Returns: + tuple: ``(waveform, sample_rate, sentence, translation, speaker_id, + sample_id)`` + """ + data = self.data[n] + path = self.root / "clips" / data["path"] + waveform, sample_rate = torchaudio.load(path) + sentence = data["sentence"] + translation = None if self.no_translation else data["translation"] + speaker_id = data["client_id"] + _id = data["path"].replace(".mp3", "") + return waveform, sample_rate, sentence, translation, speaker_id, _id + + def __len__(self) -> int: + return len(self.data) + + +def process(args): + root = Path(args.data_root).absolute() / args.src_lang + if not root.is_dir(): + raise NotADirectoryError(f"{root} does not exist") + # Extract features + feature_root = root / "fbank80" + feature_root.mkdir(exist_ok=True) + for split in CoVoST.SPLITS: + print(f"Fetching split {split}...") + dataset = CoVoST(root, split, args.src_lang, args.tgt_lang) + print("Extracting log mel filter bank features...") + for waveform, sample_rate, _, _, _, utt_id in tqdm(dataset): + extract_fbank_features( + waveform, sample_rate, feature_root / f"{utt_id}.npy" + ) + # Pack features into ZIP + zip_path = root / "fbank80.zip" + print("ZIPing features...") + create_zip(feature_root, zip_path) + print("Fetching ZIP manifest...") + zip_manifest = get_zip_manifest(zip_path) + # Generate TSV manifest + print("Generating manifest...") + train_text = [] + task = f"asr_{args.src_lang}" + if args.tgt_lang is not None: + task = f"st_{args.src_lang}_{args.tgt_lang}" + for split in CoVoST.SPLITS: + manifest = {c: [] for c in MANIFEST_COLUMNS} + dataset = CoVoST(root, split, args.src_lang, args.tgt_lang) + for wav, sr, src_utt, tgt_utt, speaker_id, utt_id in tqdm(dataset): + manifest["id"].append(utt_id) + manifest["audio"].append(zip_manifest[utt_id]) + duration_ms = int(wav.size(1) / sr * 1000) + manifest["n_frames"].append(int(1 + (duration_ms - 25) / 10)) + manifest["tgt_text"].append(src_utt if args.tgt_lang is None else tgt_utt) + manifest["speaker"].append(speaker_id) + is_train_split = split.startswith("train") + if is_train_split: + train_text.extend(manifest["tgt_text"]) + df = pd.DataFrame.from_dict(manifest) + df = filter_manifest_df(df, is_train_split=is_train_split) + save_df_to_tsv(df, root / f"{split}_{task}.tsv") + # Generate vocab + vocab_size_str = "" if args.vocab_type == "char" else str(args.vocab_size) + spm_filename_prefix = f"spm_{args.vocab_type}{vocab_size_str}_{task}" + with NamedTemporaryFile(mode="w") as f: + for t in train_text: + f.write(t + "\n") + gen_vocab( + Path(f.name), + root / spm_filename_prefix, + args.vocab_type, + args.vocab_size + ) + # Generate config YAML + gen_config_yaml( + root, + spm_filename_prefix + ".model", + yaml_filename=f"config_{task}.yaml", + specaugment_policy="lb", + ) + # Clean up + shutil.rmtree(feature_root) + + +def main(): + parser = argparse.ArgumentParser() + parser.add_argument( + "--data-root", "-d", required=True, type=str, + help="data root with sub-folders for each language <root>/<src_lang>" + ) + parser.add_argument( + "--vocab-type", + default="unigram", + required=True, + type=str, + choices=["bpe", "unigram", "char"], + ), + parser.add_argument("--vocab-size", default=1000, type=int) + parser.add_argument("--src-lang", "-s", required=True, type=str) + parser.add_argument("--tgt-lang", "-t", type=str) + args = parser.parse_args() + + process(args) + + +if __name__ == "__main__": + main() diff --git a/examples/speech_to_text/prep_librispeech_data.py b/examples/speech_to_text/prep_librispeech_data.py new file mode 100644 index 0000000000000000000000000000000000000000..7b08447190b8e7af4d81c49abdf42461fdd6760b --- /dev/null +++ b/examples/speech_to_text/prep_librispeech_data.py @@ -0,0 +1,118 @@ +#!/usr/bin/env python3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import logging +from pathlib import Path +import shutil +from tempfile import NamedTemporaryFile + +import pandas as pd +from examples.speech_to_text.data_utils import ( + create_zip, + extract_fbank_features, + gen_config_yaml, + gen_vocab, + get_zip_manifest, + save_df_to_tsv, +) +from torchaudio.datasets import LIBRISPEECH +from tqdm import tqdm + + +log = logging.getLogger(__name__) + +SPLITS = [ + "train-clean-100", + "train-clean-360", + "train-other-500", + "dev-clean", + "dev-other", + "test-clean", + "test-other", +] + +MANIFEST_COLUMNS = ["id", "audio", "n_frames", "tgt_text", "speaker"] + + +def process(args): + out_root = Path(args.output_root).absolute() + out_root.mkdir(exist_ok=True) + # Extract features + feature_root = out_root / "fbank80" + feature_root.mkdir(exist_ok=True) + for split in SPLITS: + print(f"Fetching split {split}...") + dataset = LIBRISPEECH(out_root.as_posix(), url=split, download=True) + print("Extracting log mel filter bank features...") + for wav, sample_rate, _, spk_id, chapter_no, utt_no in tqdm(dataset): + sample_id = f"{spk_id}-{chapter_no}-{utt_no}" + extract_fbank_features( + wav, sample_rate, feature_root / f"{sample_id}.npy" + ) + # Pack features into ZIP + zip_path = out_root / "fbank80.zip" + print("ZIPing features...") + create_zip(feature_root, zip_path) + print("Fetching ZIP manifest...") + zip_manifest = get_zip_manifest(zip_path) + # Generate TSV manifest + print("Generating manifest...") + train_text = [] + for split in SPLITS: + manifest = {c: [] for c in MANIFEST_COLUMNS} + dataset = LIBRISPEECH(out_root.as_posix(), url=split) + for wav, sample_rate, utt, spk_id, chapter_no, utt_no in tqdm(dataset): + sample_id = f"{spk_id}-{chapter_no}-{utt_no}" + manifest["id"].append(sample_id) + manifest["audio"].append(zip_manifest[sample_id]) + duration_ms = int(wav.size(1) / sample_rate * 1000) + manifest["n_frames"].append(int(1 + (duration_ms - 25) / 10)) + manifest["tgt_text"].append(utt.lower()) + manifest["speaker"].append(spk_id) + save_df_to_tsv( + pd.DataFrame.from_dict(manifest), out_root / f"{split}.tsv" + ) + if split.startswith("train"): + train_text.extend(manifest["tgt_text"]) + # Generate vocab + vocab_size = "" if args.vocab_type == "char" else str(args.vocab_size) + spm_filename_prefix = f"spm_{args.vocab_type}{vocab_size}" + with NamedTemporaryFile(mode="w") as f: + for t in train_text: + f.write(t + "\n") + gen_vocab( + Path(f.name), + out_root / spm_filename_prefix, + args.vocab_type, + args.vocab_size, + ) + # Generate config YAML + gen_config_yaml( + out_root, spm_filename_prefix + ".model", specaugment_policy="ld" + ) + # Clean up + shutil.rmtree(feature_root) + + +def main(): + parser = argparse.ArgumentParser() + parser.add_argument("--output-root", "-o", required=True, type=str) + parser.add_argument( + "--vocab-type", + default="unigram", + required=True, + type=str, + choices=["bpe", "unigram", "char"], + ), + parser.add_argument("--vocab-size", default=10000, type=int) + args = parser.parse_args() + + process(args) + + +if __name__ == "__main__": + main() diff --git a/examples/speech_to_text/prep_mtedx_data.py b/examples/speech_to_text/prep_mtedx_data.py new file mode 100644 index 0000000000000000000000000000000000000000..34b1c398c8df5168cfc731cf68592b2ae8d5897b --- /dev/null +++ b/examples/speech_to_text/prep_mtedx_data.py @@ -0,0 +1,238 @@ +#!/usr/bin/env python3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import logging +import os +from pathlib import Path +import shutil +from itertools import groupby +from tempfile import NamedTemporaryFile +from typing import Tuple + +import pandas as pd +import soundfile as sf +from examples.speech_to_text.data_utils import ( + create_zip, + extract_fbank_features, + filter_manifest_df, + gen_config_yaml, + gen_vocab, + get_zip_manifest, + load_df_from_tsv, + save_df_to_tsv, +) +import torch +from torch.utils.data import Dataset +from tqdm import tqdm + +from fairseq.data.audio.audio_utils import get_waveform + + +log = logging.getLogger(__name__) + + +MANIFEST_COLUMNS = ["id", "audio", "n_frames", "tgt_text", "speaker", "tgt_lang"] + + +class mTEDx(Dataset): + """ + Create a Dataset for Multilingual TEDx. + Each item is a tuple of the form: waveform, sample_rate, source utterance, + target utterance, speaker_id, utterance_id + """ + + SPLITS = ["train", "valid", "test"] + LANGPAIRS = ["es-es", "fr-fr", "pt-pt", "it-it", "ru-ru", "el-el", "ar-ar", "de-de", + "es-en", "es-fr", "es-pt", "es-it", "fr-en", "fr-es", "fr-pt", + "pt-en", "pt-es", "it-en", "it-es", "ru-en", "el-en"] + + def __init__(self, root: str, lang: str, split: str) -> None: + assert split in self.SPLITS and lang in self.LANGPAIRS + _root = Path(root) / f"{lang}" / "data" / split + wav_root, txt_root = _root / "wav", _root / "txt" + assert _root.is_dir() and wav_root.is_dir() and txt_root.is_dir() + # Load audio segments + try: + import yaml + except ImportError: + print("Please install PyYAML to load the Multilingual TEDx YAML files") + with open(txt_root / f"{split}.yaml") as f: + segments = yaml.load(f, Loader=yaml.BaseLoader) + # Load source and target utterances + src, tgt = lang.split("-") + for _lang in [src, tgt]: + with open(txt_root / f"{split}.{_lang}") as f: + utterances = [r.strip() for r in f] + assert len(segments) == len(utterances) + for i, u in enumerate(utterances): + segments[i][_lang] = u + # Gather info + self.data = [] + for wav_filename, _seg_group in groupby(segments, lambda x: x["wav"]): + wav_filename = wav_filename.replace(".wav", ".flac") + wav_path = wav_root / wav_filename + sample_rate = sf.info(wav_path.as_posix()).samplerate + seg_group = sorted(_seg_group, key=lambda x: float(x["offset"])) + for i, segment in enumerate(seg_group): + offset = int(float(segment["offset"]) * sample_rate) + n_frames = int(float(segment["duration"]) * sample_rate) + _id = f"{wav_path.stem}_{i}" + self.data.append( + ( + wav_path.as_posix(), + offset, + n_frames, + sample_rate, + segment[src], + segment[tgt], + segment["speaker_id"], + tgt, + _id, + ) + ) + + def __getitem__(self, n: int) -> Tuple[torch.Tensor, int, str, str, str, str, str]: + wav_path, offset, n_frames, sr, src_utt, tgt_utt, spk_id, tgt_lang, utt_id = self.data[n] + waveform, _ = get_waveform(wav_path, frames=n_frames, start=offset) + waveform = torch.from_numpy(waveform) + return waveform, sr, src_utt, tgt_utt, spk_id, tgt_lang, utt_id + + def __len__(self) -> int: + return len(self.data) + + +def process(args): + root = Path(args.data_root).absolute() + for lang in mTEDx.LANGPAIRS: + cur_root = root / f"{lang}" + if not cur_root.is_dir(): + print(f"{cur_root.as_posix()} does not exist. Skipped.") + continue + # Extract features + feature_root = cur_root / "fbank80" + feature_root.mkdir(exist_ok=True) + for split in mTEDx.SPLITS: + print(f"Fetching split {split}...") + dataset = mTEDx(root.as_posix(), lang, split) + print("Extracting log mel filter bank features...") + for waveform, sample_rate, _, _, _, _, utt_id in tqdm(dataset): + extract_fbank_features( + waveform, sample_rate, feature_root / f"{utt_id}.npy" + ) + # Pack features into ZIP + zip_path = cur_root / "fbank80.zip" + print("ZIPing features...") + create_zip(feature_root, zip_path) + print("Fetching ZIP manifest...") + zip_manifest = get_zip_manifest(zip_path) + # Generate TSV manifest + print("Generating manifest...") + train_text = [] + for split in mTEDx.SPLITS: + is_train_split = split.startswith("train") + manifest = {c: [] for c in MANIFEST_COLUMNS} + dataset = mTEDx(args.data_root, lang, split) + for wav, sr, src_utt, tgt_utt, speaker_id, tgt_lang, utt_id in tqdm(dataset): + manifest["id"].append(utt_id) + manifest["audio"].append(zip_manifest[utt_id]) + duration_ms = int(wav.size(1) / sr * 1000) + manifest["n_frames"].append(int(1 + (duration_ms - 25) / 10)) + manifest["tgt_text"].append(src_utt if args.task == "asr" else tgt_utt) + manifest["speaker"].append(speaker_id) + manifest["tgt_lang"].append(tgt_lang) + if is_train_split: + train_text.extend(manifest["tgt_text"]) + df = pd.DataFrame.from_dict(manifest) + df = filter_manifest_df(df, is_train_split=is_train_split) + save_df_to_tsv(df, cur_root / f"{split}_{args.task}.tsv") + # Generate vocab + v_size_str = "" if args.vocab_type == "char" else str(args.vocab_size) + spm_filename_prefix = f"spm_{args.vocab_type}{v_size_str}_{args.task}" + with NamedTemporaryFile(mode="w") as f: + for t in train_text: + f.write(t + "\n") + gen_vocab( + Path(f.name), + cur_root / spm_filename_prefix, + args.vocab_type, + args.vocab_size, + ) + # Generate config YAML + gen_config_yaml( + cur_root, + spm_filename_prefix + ".model", + yaml_filename=f"config_{args.task}.yaml", + specaugment_policy="lb", + ) + # Clean up + shutil.rmtree(feature_root) + + +def process_joint(args): + cur_root = Path(args.data_root) + assert all((cur_root / f"{lang}").is_dir() for lang in mTEDx.LANGPAIRS), \ + "do not have downloaded data available for all languages" + # Generate vocab + vocab_size_str = "" if args.vocab_type == "char" else str(args.vocab_size) + spm_filename_prefix = f"spm_{args.vocab_type}{vocab_size_str}_{args.task}" + with NamedTemporaryFile(mode="w") as f: + for lang in mTEDx.LANGPAIRS: + tsv_path = cur_root / f"{lang}" / f"train_{args.task}.tsv" + df = load_df_from_tsv(tsv_path) + for t in df["tgt_text"]: + f.write(t + "\n") + special_symbols = None + if args.joint: + # Add tgt_lang tags to dict + special_symbols = list({f'<lang:{lang.split("-")[1]}>' for lang in mTEDx.LANGPAIRS}) + gen_vocab( + Path(f.name), + cur_root / spm_filename_prefix, + args.vocab_type, + args.vocab_size, + special_symbols=special_symbols + ) + # Generate config YAML + gen_config_yaml( + cur_root, + spm_filename_prefix + ".model", + yaml_filename=f"config_{args.task}.yaml", + specaugment_policy="ld", + prepend_tgt_lang_tag=(args.joint), + ) + # Make symbolic links to manifests + for lang in mTEDx.LANGPAIRS: + for split in mTEDx.SPLITS: + src_path = cur_root / f"{lang}" / f"{split}_{args.task}.tsv" + desc_path = cur_root / f"{split}_{lang}_{args.task}.tsv" + if not desc_path.is_symlink(): + os.symlink(src_path, desc_path) + + +def main(): + parser = argparse.ArgumentParser() + parser.add_argument("--data-root", "-d", required=True, type=str) + parser.add_argument( + "--vocab-type", + default="unigram", + required=True, + type=str, + choices=["bpe", "unigram", "char"], + ), + parser.add_argument("--vocab-size", default=8000, type=int) + parser.add_argument("--task", type=str, choices=["asr", "st"]) + parser.add_argument("--joint", action="store_true", help="") + args = parser.parse_args() + + if args.joint: + process_joint(args) + else: + process(args) + + +if __name__ == "__main__": + main() diff --git a/examples/speech_to_text/prep_mustc_data.py b/examples/speech_to_text/prep_mustc_data.py new file mode 100644 index 0000000000000000000000000000000000000000..0ee204e651a4e5a305cd27274101740a2a35d7bc --- /dev/null +++ b/examples/speech_to_text/prep_mustc_data.py @@ -0,0 +1,264 @@ +#!/usr/bin/env python3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import logging +import os +from pathlib import Path +import shutil +from itertools import groupby +from tempfile import NamedTemporaryFile +from typing import Tuple + +import numpy as np +import pandas as pd +import soundfile as sf +from examples.speech_to_text.data_utils import ( + create_zip, + extract_fbank_features, + filter_manifest_df, + gen_config_yaml, + gen_vocab, + get_zip_manifest, + load_df_from_tsv, + save_df_to_tsv, + cal_gcmvn_stats, +) +import torch +from torch.utils.data import Dataset +from tqdm import tqdm + +from fairseq.data.audio.audio_utils import get_waveform + + +log = logging.getLogger(__name__) + + +MANIFEST_COLUMNS = ["id", "audio", "n_frames", "tgt_text", "speaker"] + + +class MUSTC(Dataset): + """ + Create a Dataset for MuST-C. Each item is a tuple of the form: + waveform, sample_rate, source utterance, target utterance, speaker_id, + utterance_id + """ + + SPLITS = ["train", "dev", "tst-COMMON", "tst-HE"] + LANGUAGES = ["de", "es", "fr", "it", "nl", "pt", "ro", "ru"] + + def __init__(self, root: str, lang: str, split: str) -> None: + assert split in self.SPLITS and lang in self.LANGUAGES + _root = Path(root) / f"en-{lang}" / "data" / split + wav_root, txt_root = _root / "wav", _root / "txt" + assert _root.is_dir() and wav_root.is_dir() and txt_root.is_dir() + # Load audio segments + try: + import yaml + except ImportError: + print("Please install PyYAML to load the MuST-C YAML files") + with open(txt_root / f"{split}.yaml") as f: + segments = yaml.load(f, Loader=yaml.BaseLoader) + # Load source and target utterances + for _lang in ["en", lang]: + with open(txt_root / f"{split}.{_lang}") as f: + utterances = [r.strip() for r in f] + assert len(segments) == len(utterances) + for i, u in enumerate(utterances): + segments[i][_lang] = u + # Gather info + self.data = [] + for wav_filename, _seg_group in groupby(segments, lambda x: x["wav"]): + wav_path = wav_root / wav_filename + sample_rate = sf.info(wav_path.as_posix()).samplerate + seg_group = sorted(_seg_group, key=lambda x: x["offset"]) + for i, segment in enumerate(seg_group): + offset = int(float(segment["offset"]) * sample_rate) + n_frames = int(float(segment["duration"]) * sample_rate) + _id = f"{wav_path.stem}_{i}" + self.data.append( + ( + wav_path.as_posix(), + offset, + n_frames, + sample_rate, + segment["en"], + segment[lang], + segment["speaker_id"], + _id, + ) + ) + + def __getitem__(self, n: int) -> Tuple[torch.Tensor, int, str, str, str, str]: + wav_path, offset, n_frames, sr, src_utt, tgt_utt, spk_id, utt_id = self.data[n] + waveform, _ = get_waveform(wav_path, frames=n_frames, start=offset) + waveform = torch.from_numpy(waveform) + return waveform, sr, src_utt, tgt_utt, spk_id, utt_id + + def __len__(self) -> int: + return len(self.data) + + +def process(args): + root = Path(args.data_root).absolute() + for lang in MUSTC.LANGUAGES: + cur_root = root / f"en-{lang}" + if not cur_root.is_dir(): + print(f"{cur_root.as_posix()} does not exist. Skipped.") + continue + # Extract features + feature_root = cur_root / "fbank80" + feature_root.mkdir(exist_ok=True) + for split in MUSTC.SPLITS: + print(f"Fetching split {split}...") + dataset = MUSTC(root.as_posix(), lang, split) + print("Extracting log mel filter bank features...") + if split == 'train' and args.cmvn_type == "global": + print("And estimating cepstral mean and variance stats...") + gcmvn_feature_list = [] + + for waveform, sample_rate, _, _, _, utt_id in tqdm(dataset): + features = extract_fbank_features(waveform, sample_rate) + + np.save( + (feature_root / f"{utt_id}.npy").as_posix(), + features + ) + + if split == 'train' and args.cmvn_type == "global": + if len(gcmvn_feature_list) < args.gcmvn_max_num: + gcmvn_feature_list.append(features) + + if split == 'train' and args.cmvn_type == "global": + # Estimate and save cmv + stats = cal_gcmvn_stats(gcmvn_feature_list) + with open(cur_root / "gcmvn.npz", "wb") as f: + np.savez(f, mean=stats["mean"], std=stats["std"]) + + # Pack features into ZIP + zip_path = cur_root / "fbank80.zip" + print("ZIPing features...") + create_zip(feature_root, zip_path) + print("Fetching ZIP manifest...") + zip_manifest = get_zip_manifest(zip_path) + # Generate TSV manifest + print("Generating manifest...") + train_text = [] + for split in MUSTC.SPLITS: + is_train_split = split.startswith("train") + manifest = {c: [] for c in MANIFEST_COLUMNS} + dataset = MUSTC(args.data_root, lang, split) + for wav, sr, src_utt, tgt_utt, speaker_id, utt_id in tqdm(dataset): + manifest["id"].append(utt_id) + manifest["audio"].append(zip_manifest[utt_id]) + duration_ms = int(wav.size(1) / sr * 1000) + manifest["n_frames"].append(int(1 + (duration_ms - 25) / 10)) + manifest["tgt_text"].append(src_utt if args.task == "asr" else tgt_utt) + manifest["speaker"].append(speaker_id) + if is_train_split: + train_text.extend(manifest["tgt_text"]) + df = pd.DataFrame.from_dict(manifest) + df = filter_manifest_df(df, is_train_split=is_train_split) + save_df_to_tsv(df, cur_root / f"{split}_{args.task}.tsv") + # Generate vocab + v_size_str = "" if args.vocab_type == "char" else str(args.vocab_size) + spm_filename_prefix = f"spm_{args.vocab_type}{v_size_str}_{args.task}" + with NamedTemporaryFile(mode="w") as f: + for t in train_text: + f.write(t + "\n") + gen_vocab( + Path(f.name), + cur_root / spm_filename_prefix, + args.vocab_type, + args.vocab_size, + ) + # Generate config YAML + gen_config_yaml( + cur_root, + spm_filename_prefix + ".model", + yaml_filename=f"config_{args.task}.yaml", + specaugment_policy="lb", + cmvn_type=args.cmvn_type, + gcmvn_path=( + cur_root / "gcmvn.npz" if args.cmvn_type == "global" + else None + ), + ) + # Clean up + shutil.rmtree(feature_root) + + +def process_joint(args): + cur_root = Path(args.data_root) + assert all((cur_root / f"en-{lang}").is_dir() for lang in MUSTC.LANGUAGES), \ + "do not have downloaded data available for all 8 languages" + # Generate vocab + vocab_size_str = "" if args.vocab_type == "char" else str(args.vocab_size) + spm_filename_prefix = f"spm_{args.vocab_type}{vocab_size_str}_{args.task}" + with NamedTemporaryFile(mode="w") as f: + for lang in MUSTC.LANGUAGES: + tsv_path = cur_root / f"en-{lang}" / f"train_{args.task}.tsv" + df = load_df_from_tsv(tsv_path) + for t in df["tgt_text"]: + f.write(t + "\n") + special_symbols = None + if args.task == 'st': + special_symbols = [f'<lang:{lang}>' for lang in MUSTC.LANGUAGES] + gen_vocab( + Path(f.name), + cur_root / spm_filename_prefix, + args.vocab_type, + args.vocab_size, + special_symbols=special_symbols + ) + # Generate config YAML + gen_config_yaml( + cur_root, + spm_filename_prefix + ".model", + yaml_filename=f"config_{args.task}.yaml", + specaugment_policy="ld", + prepend_tgt_lang_tag=(args.task == "st"), + ) + # Make symbolic links to manifests + for lang in MUSTC.LANGUAGES: + for split in MUSTC.SPLITS: + src_path = cur_root / f"en-{lang}" / f"{split}_{args.task}.tsv" + desc_path = cur_root / f"{split}_{lang}_{args.task}.tsv" + if not desc_path.is_symlink(): + os.symlink(src_path, desc_path) + + +def main(): + parser = argparse.ArgumentParser() + parser.add_argument("--data-root", "-d", required=True, type=str) + parser.add_argument( + "--vocab-type", + default="unigram", + required=True, + type=str, + choices=["bpe", "unigram", "char"], + ), + parser.add_argument("--vocab-size", default=8000, type=int) + parser.add_argument("--task", type=str, choices=["asr", "st"]) + parser.add_argument("--joint", action="store_true", help="") + parser.add_argument("--cmvn-type", default="utterance", + choices=["global", "utterance"], + help="The type of cepstral mean and variance normalization") + parser.add_argument("--gcmvn-max-num", default=150000, type=int, + help=( + "Maximum number of sentences to use to estimate" + "global mean and variance" + )) + args = parser.parse_args() + + if args.joint: + process_joint(args) + else: + process(args) + + +if __name__ == "__main__": + main() diff --git a/examples/speech_to_text/seg_mustc_data.py b/examples/speech_to_text/seg_mustc_data.py new file mode 100644 index 0000000000000000000000000000000000000000..1ee665d6399729afe17d790d872eff34de124900 --- /dev/null +++ b/examples/speech_to_text/seg_mustc_data.py @@ -0,0 +1,54 @@ +#!/usr/bin/env python3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import logging +from pathlib import Path +import soundfile as sf +from examples.speech_to_text.prep_mustc_data import ( + MUSTC +) + +from tqdm import tqdm + +log = logging.getLogger(__name__) + + +def main(args): + root = Path(args.data_root).absolute() + lang = args.lang + split = args.split + + cur_root = root / f"en-{lang}" + assert cur_root.is_dir(), ( + f"{cur_root.as_posix()} does not exist. Skipped." + ) + + dataset = MUSTC(root.as_posix(), lang, split) + output = Path(args.output).absolute() + output.mkdir(exist_ok=True) + f_text = open(output / f"{split}.{lang}", "w") + f_wav_list = open(output / f"{split}.wav_list", "w") + for waveform, sample_rate, _, text, _, utt_id in tqdm(dataset): + sf.write( + output / f"{utt_id}.wav", + waveform.squeeze(0).numpy(), + samplerate=int(sample_rate) + ) + f_text.write(text + "\n") + f_wav_list.write(str(output / f"{utt_id}.wav") + "\n") + + +if __name__ == "__main__": + parser = argparse.ArgumentParser() + parser.add_argument("--data-root", "-d", required=True, type=str) + parser.add_argument("--task", required=True, type=str, choices=["asr", "st"]) + parser.add_argument("--lang", required=True, type=str) + parser.add_argument("--output", required=True, type=str) + parser.add_argument("--split", required=True, choices=MUSTC.SPLITS) + args = parser.parse_args() + + main(args) diff --git a/examples/speech_to_text/simultaneous_translation/agents/fairseq_simul_st_agent.py b/examples/speech_to_text/simultaneous_translation/agents/fairseq_simul_st_agent.py new file mode 100644 index 0000000000000000000000000000000000000000..61617a1739ce196abba1e9a6f9ad9e9f4b37b9c1 --- /dev/null +++ b/examples/speech_to_text/simultaneous_translation/agents/fairseq_simul_st_agent.py @@ -0,0 +1,363 @@ +import math +import os +import json +import numpy as np +import torch +import torchaudio.compliance.kaldi as kaldi +import yaml +from fairseq import checkpoint_utils, tasks +from fairseq.file_io import PathManager + +try: + from simuleval import READ_ACTION, WRITE_ACTION, DEFAULT_EOS + from simuleval.agents import SpeechAgent + from simuleval.states import ListEntry, SpeechStates +except ImportError: + print("Please install simuleval 'pip install simuleval'") + +SHIFT_SIZE = 10 +WINDOW_SIZE = 25 +SAMPLE_RATE = 16000 +FEATURE_DIM = 80 +BOW_PREFIX = "\u2581" + + +class OnlineFeatureExtractor: + """ + Extract speech feature on the fly. + """ + + def __init__(self, args): + self.shift_size = args.shift_size + self.window_size = args.window_size + assert self.window_size >= self.shift_size + + self.sample_rate = args.sample_rate + self.feature_dim = args.feature_dim + self.num_samples_per_shift = int(self.shift_size * self.sample_rate / 1000) + self.num_samples_per_window = int(self.window_size * self.sample_rate / 1000) + self.len_ms_to_samples = lambda x: x * self.sample_rate / 1000 + self.previous_residual_samples = [] + self.global_cmvn = args.global_cmvn + + def clear_cache(self): + self.previous_residual_samples = [] + + def __call__(self, new_samples): + samples = self.previous_residual_samples + new_samples + if len(samples) < self.num_samples_per_window: + self.previous_residual_samples = samples + return + + # num_frames is the number of frames from the new segment + num_frames = math.floor( + (len(samples) - self.len_ms_to_samples(self.window_size - self.shift_size)) + / self.num_samples_per_shift + ) + + # the number of frames used for feature extraction + # including some part of thte previous segment + effective_num_samples = int( + num_frames * self.len_ms_to_samples(self.shift_size) + + self.len_ms_to_samples(self.window_size - self.shift_size) + ) + + input_samples = samples[:effective_num_samples] + self.previous_residual_samples = samples[ + num_frames * self.num_samples_per_shift: + ] + + torch.manual_seed(1) + output = kaldi.fbank( + torch.FloatTensor(input_samples).unsqueeze(0), + num_mel_bins=self.feature_dim, + frame_length=self.window_size, + frame_shift=self.shift_size, + ).numpy() + + output = self.transform(output) + + return torch.from_numpy(output) + + def transform(self, input): + if self.global_cmvn is None: + return input + + mean = self.global_cmvn["mean"] + std = self.global_cmvn["std"] + + x = np.subtract(input, mean) + x = np.divide(x, std) + return x + + +class TensorListEntry(ListEntry): + """ + Data structure to store a list of tensor. + """ + + def append(self, value): + + if len(self.value) == 0: + self.value = value + return + + self.value = torch.cat([self.value] + [value], dim=0) + + def info(self): + return { + "type": str(self.new_value_type), + "length": self.__len__(), + "value": "" if type(self.value) is list else self.value.size(), + } + + +class FairseqSimulSTAgent(SpeechAgent): + + speech_segment_size = 40 # in ms, 4 pooling ratio * 10 ms step size + + def __init__(self, args): + super().__init__(args) + + self.eos = DEFAULT_EOS + + self.gpu = getattr(args, "gpu", False) + + self.args = args + + self.load_model_vocab(args) + + if getattr( + self.model.decoder.layers[0].encoder_attn, + 'pre_decision_ratio', + None + ) is not None: + self.speech_segment_size *= ( + self.model.decoder.layers[0].encoder_attn.pre_decision_ratio + ) + + args.global_cmvn = None + if args.config: + with open(os.path.join(args.data_bin, args.config), "r") as f: + config = yaml.load(f, Loader=yaml.BaseLoader) + + if "global_cmvn" in config: + args.global_cmvn = np.load(config["global_cmvn"]["stats_npz_path"]) + + if args.global_stats: + with PathManager.open(args.global_stats, "r") as f: + global_cmvn = json.loads(f.read()) + self.global_cmvn = {"mean": global_cmvn["mean"], "std": global_cmvn["stddev"]} + + self.feature_extractor = OnlineFeatureExtractor(args) + + self.max_len = args.max_len + + self.force_finish = args.force_finish + + torch.set_grad_enabled(False) + + def build_states(self, args, client, sentence_id): + # Initialize states here, for example add customized entry to states + # This function will be called at beginning of every new sentence + states = SpeechStates(args, client, sentence_id, self) + self.initialize_states(states) + return states + + def to_device(self, tensor): + if self.gpu: + return tensor.cuda() + else: + return tensor.cpu() + + @staticmethod + def add_args(parser): + # fmt: off + parser.add_argument('--model-path', type=str, required=True, + help='path to your pretrained model.') + parser.add_argument("--data-bin", type=str, required=True, + help="Path of data binary") + parser.add_argument("--config", type=str, default=None, + help="Path to config yaml file") + parser.add_argument("--global-stats", type=str, default=None, + help="Path to json file containing cmvn stats") + parser.add_argument("--tgt-splitter-type", type=str, default="SentencePiece", + help="Subword splitter type for target text") + parser.add_argument("--tgt-splitter-path", type=str, default=None, + help="Subword splitter model path for target text") + parser.add_argument("--user-dir", type=str, default="examples/simultaneous_translation", + help="User directory for simultaneous translation") + parser.add_argument("--max-len", type=int, default=200, + help="Max length of translation") + parser.add_argument("--force-finish", default=False, action="store_true", + help="Force the model to finish the hypothsis if the source is not finished") + parser.add_argument("--shift-size", type=int, default=SHIFT_SIZE, + help="Shift size of feature extraction window.") + parser.add_argument("--window-size", type=int, default=WINDOW_SIZE, + help="Window size of feature extraction window.") + parser.add_argument("--sample-rate", type=int, default=SAMPLE_RATE, + help="Sample rate") + parser.add_argument("--feature-dim", type=int, default=FEATURE_DIM, + help="Acoustic feature dimension.") + + # fmt: on + return parser + + def load_model_vocab(self, args): + + filename = args.model_path + if not os.path.exists(filename): + raise IOError("Model file not found: {}".format(filename)) + + state = checkpoint_utils.load_checkpoint_to_cpu(filename) + + task_args = state["cfg"]["task"] + task_args.data = args.data_bin + + if args.config is not None: + task_args.config_yaml = args.config + + task = tasks.setup_task(task_args) + + # build model for ensemble + state["cfg"]["model"].load_pretrained_encoder_from = None + state["cfg"]["model"].load_pretrained_decoder_from = None + self.model = task.build_model(state["cfg"]["model"]) + self.model.load_state_dict(state["model"], strict=True) + self.model.eval() + self.model.share_memory() + + if self.gpu: + self.model.cuda() + + # Set dictionary + self.dict = {} + self.dict["tgt"] = task.target_dictionary + + def initialize_states(self, states): + self.feature_extractor.clear_cache() + states.units.source = TensorListEntry() + states.units.target = ListEntry() + states.incremental_states = dict() + + def segment_to_units(self, segment, states): + # Convert speech samples to features + features = self.feature_extractor(segment) + if features is not None: + return [features] + else: + return [] + + def units_to_segment(self, units, states): + # Merge sub word to full word. + if self.model.decoder.dictionary.eos() == units[0]: + return DEFAULT_EOS + + segment = [] + if None in units.value: + units.value.remove(None) + + for index in units: + if index is None: + units.pop() + token = self.model.decoder.dictionary.string([index]) + if token.startswith(BOW_PREFIX): + if len(segment) == 0: + segment += [token.replace(BOW_PREFIX, "")] + else: + for j in range(len(segment)): + units.pop() + + string_to_return = ["".join(segment)] + + if self.model.decoder.dictionary.eos() == units[0]: + string_to_return += [DEFAULT_EOS] + + return string_to_return + else: + segment += [token.replace(BOW_PREFIX, "")] + + if ( + len(units) > 0 + and self.model.decoder.dictionary.eos() == units[-1] + or len(states.units.target) > self.max_len + ): + tokens = [self.model.decoder.dictionary.string([unit]) for unit in units] + return ["".join(tokens).replace(BOW_PREFIX, "")] + [DEFAULT_EOS] + + return None + + def update_model_encoder(self, states): + if len(states.units.source) == 0: + return + src_indices = self.to_device( + states.units.source.value.unsqueeze(0) + ) + src_lengths = self.to_device( + torch.LongTensor([states.units.source.value.size(0)]) + ) + + states.encoder_states = self.model.encoder(src_indices, src_lengths) + torch.cuda.empty_cache() + + def update_states_read(self, states): + # Happens after a read action. + self.update_model_encoder(states) + + def policy(self, states): + if not getattr(states, "encoder_states", None): + return READ_ACTION + + tgt_indices = self.to_device( + torch.LongTensor( + [self.model.decoder.dictionary.eos()] + + [x for x in states.units.target.value if x is not None] + ).unsqueeze(0) + ) + + states.incremental_states["steps"] = { + "src": states.encoder_states["encoder_out"][0].size(0), + "tgt": 1 + len(states.units.target), + } + + states.incremental_states["online"] = {"only": torch.tensor(not states.finish_read())} + + x, outputs = self.model.decoder.forward( + prev_output_tokens=tgt_indices, + encoder_out=states.encoder_states, + incremental_state=states.incremental_states, + ) + + states.decoder_out = x + + states.decoder_out_extra = outputs + + torch.cuda.empty_cache() + + if outputs.action == 0: + return READ_ACTION + else: + return WRITE_ACTION + + def predict(self, states): + decoder_states = states.decoder_out + + lprobs = self.model.get_normalized_probs( + [decoder_states[:, -1:]], log_probs=True + ) + + index = lprobs.argmax(dim=-1) + + index = index[0, 0].item() + + if ( + self.force_finish + and index == self.model.decoder.dictionary.eos() + and not states.finish_read() + ): + # If we want to force finish the translation + # (don't stop before finish reading), return a None + # self.model.decoder.clear_cache(states.incremental_states) + index = None + + return index diff --git a/examples/stories/README.md b/examples/stories/README.md new file mode 100644 index 0000000000000000000000000000000000000000..588941eddc5f0280f5254affd40ef49de874c885 --- /dev/null +++ b/examples/stories/README.md @@ -0,0 +1,66 @@ +# Hierarchical Neural Story Generation (Fan et al., 2018) + +The following commands provide an example of pre-processing data, training a model, and generating text for story generation with the WritingPrompts dataset. + +## Pre-trained models + +Description | Dataset | Model | Test set(s) +---|---|---|--- +Stories with Convolutional Model <br> ([Fan et al., 2018](https://arxiv.org/abs/1805.04833)) | [WritingPrompts](https://dl.fbaipublicfiles.com/fairseq/data/writingPrompts.tar.gz) | [download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/models/stories_checkpoint.tar.bz2) | [download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/data/stories_test.tar.bz2) + +We provide sample stories generated by the [convolutional seq2seq model](https://dl.fbaipublicfiles.com/fairseq/data/seq2seq_stories.txt) and [fusion model](https://dl.fbaipublicfiles.com/fairseq/data/fusion_stories.txt) from [Fan et al., 2018](https://arxiv.org/abs/1805.04833). The corresponding prompts for the fusion model can be found [here](https://dl.fbaipublicfiles.com/fairseq/data/fusion_prompts.txt). Note that there are unk in the file, as we modeled a small full vocabulary (no BPE or pre-training). We did not use these unk prompts for human evaluation. + +## Dataset + +The dataset can be downloaded like this: + +```bash +cd examples/stories +curl https://dl.fbaipublicfiles.com/fairseq/data/writingPrompts.tar.gz | tar xvzf - +``` + +and contains a train, test, and valid split. The dataset is described here: https://arxiv.org/abs/1805.04833. We model only the first 1000 words of each story, including one newLine token. + +## Example usage + +First we will preprocess the dataset. Note that the dataset release is the full data, but the paper models the first 1000 words of each story. Here is example code that trims the dataset to the first 1000 words of each story: +```python +data = ["train", "test", "valid"] +for name in data: + with open(name + ".wp_target") as f: + stories = f.readlines() + stories = [" ".join(i.split()[0:1000]) for i in stories] + with open(name + ".wp_target", "w") as o: + for line in stories: + o.write(line.strip() + "\n") +``` + +Once we've trimmed the data we can binarize it and train our model: +```bash +# Binarize the dataset: +export TEXT=examples/stories/writingPrompts +fairseq-preprocess --source-lang wp_source --target-lang wp_target \ + --trainpref $TEXT/train --validpref $TEXT/valid --testpref $TEXT/test \ + --destdir data-bin/writingPrompts --padding-factor 1 --thresholdtgt 10 --thresholdsrc 10 + +# Train the model: +fairseq-train data-bin/writingPrompts -a fconv_self_att_wp --lr 0.25 --optimizer nag --clip-norm 0.1 --max-tokens 1500 --lr-scheduler reduce_lr_on_plateau --decoder-attention True --encoder-attention False --criterion label_smoothed_cross_entropy --weight-decay .0000001 --label-smoothing 0 --source-lang wp_source --target-lang wp_target --gated-attention True --self-attention True --project-input True --pretrained False + +# Train a fusion model: +# add the arguments: --pretrained True --pretrained-checkpoint path/to/checkpoint + +# Generate: +# Note: to load the pretrained model at generation time, you need to pass in a model-override argument to communicate to the fusion model at generation time where you have placed the pretrained checkpoint. By default, it will load the exact path of the fusion model's pretrained model from training time. You should use model-override if you have moved the pretrained model (or are using our provided models). If you are generating from a non-fusion model, the model-override argument is not necessary. + +fairseq-generate data-bin/writingPrompts --path /path/to/trained/model/checkpoint_best.pt --batch-size 32 --beam 1 --sampling --sampling-topk 10 --temperature 0.8 --nbest 1 --model-overrides "{'pretrained_checkpoint':'/path/to/pretrained/model/checkpoint'}" +``` + +## Citation +```bibtex +@inproceedings{fan2018hierarchical, + title = {Hierarchical Neural Story Generation}, + author = {Fan, Angela and Lewis, Mike and Dauphin, Yann}, + booktitle = {Conference of the Association for Computational Linguistics (ACL)}, + year = 2018, +} +``` diff --git a/examples/translation/README.md b/examples/translation/README.md new file mode 100644 index 0000000000000000000000000000000000000000..2941f5eb8482dab61dca5eca27a71abd7ee5bf5c --- /dev/null +++ b/examples/translation/README.md @@ -0,0 +1,301 @@ +# Neural Machine Translation + +This README contains instructions for [using pretrained translation models](#example-usage-torchhub) +as well as [training new models](#training-a-new-model). + +## Pre-trained models + +Model | Description | Dataset | Download +---|---|---|--- +`conv.wmt14.en-fr` | Convolutional <br> ([Gehring et al., 2017](https://arxiv.org/abs/1705.03122)) | [WMT14 English-French](http://statmt.org/wmt14/translation-task.html#Download) | model: <br> [download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/models/wmt14.v2.en-fr.fconv-py.tar.bz2) <br> newstest2014: <br> [download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/data/wmt14.v2.en-fr.newstest2014.tar.bz2) <br> newstest2012/2013: <br> [download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/data/wmt14.v2.en-fr.ntst1213.tar.bz2) +`conv.wmt14.en-de` | Convolutional <br> ([Gehring et al., 2017](https://arxiv.org/abs/1705.03122)) | [WMT14 English-German](http://statmt.org/wmt14/translation-task.html#Download) | model: <br> [download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/models/wmt14.en-de.fconv-py.tar.bz2) <br> newstest2014: <br> [download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/data/wmt14.en-de.newstest2014.tar.bz2) +`conv.wmt17.en-de` | Convolutional <br> ([Gehring et al., 2017](https://arxiv.org/abs/1705.03122)) | [WMT17 English-German](http://statmt.org/wmt17/translation-task.html#Download) | model: <br> [download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/models/wmt17.v2.en-de.fconv-py.tar.bz2) <br> newstest2014: <br> [download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/data/wmt17.v2.en-de.newstest2014.tar.bz2) +`transformer.wmt14.en-fr` | Transformer <br> ([Ott et al., 2018](https://arxiv.org/abs/1806.00187)) | [WMT14 English-French](http://statmt.org/wmt14/translation-task.html#Download) | model: <br> [download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/models/wmt14.en-fr.joined-dict.transformer.tar.bz2) <br> newstest2014: <br> [download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/data/wmt14.en-fr.joined-dict.newstest2014.tar.bz2) +`transformer.wmt16.en-de` | Transformer <br> ([Ott et al., 2018](https://arxiv.org/abs/1806.00187)) | [WMT16 English-German](https://drive.google.com/uc?export=download&id=0B_bZck-ksdkpM25jRUN2X2UxMm8) | model: <br> [download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/models/wmt16.en-de.joined-dict.transformer.tar.bz2) <br> newstest2014: <br> [download (.tar.bz2)](https://dl.fbaipublicfiles.com/fairseq/data/wmt16.en-de.joined-dict.newstest2014.tar.bz2) +`transformer.wmt18.en-de` | Transformer <br> ([Edunov et al., 2018](https://arxiv.org/abs/1808.09381)) <br> WMT'18 winner | [WMT'18 English-German](http://www.statmt.org/wmt18/translation-task.html) | model: <br> [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/wmt18.en-de.ensemble.tar.gz) <br> See NOTE in the archive +`transformer.wmt19.en-de` | Transformer <br> ([Ng et al., 2019](https://arxiv.org/abs/1907.06616)) <br> WMT'19 winner | [WMT'19 English-German](http://www.statmt.org/wmt19/translation-task.html) | model: <br> [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/wmt19.en-de.joined-dict.ensemble.tar.gz) +`transformer.wmt19.de-en` | Transformer <br> ([Ng et al., 2019](https://arxiv.org/abs/1907.06616)) <br> WMT'19 winner | [WMT'19 German-English](http://www.statmt.org/wmt19/translation-task.html) | model: <br> [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/wmt19.de-en.joined-dict.ensemble.tar.gz) +`transformer.wmt19.en-ru` | Transformer <br> ([Ng et al., 2019](https://arxiv.org/abs/1907.06616)) <br> WMT'19 winner | [WMT'19 English-Russian](http://www.statmt.org/wmt19/translation-task.html) | model: <br> [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/wmt19.en-ru.ensemble.tar.gz) +`transformer.wmt19.ru-en` | Transformer <br> ([Ng et al., 2019](https://arxiv.org/abs/1907.06616)) <br> WMT'19 winner | [WMT'19 Russian-English](http://www.statmt.org/wmt19/translation-task.html) | model: <br> [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/wmt19.ru-en.ensemble.tar.gz) + +## Example usage (torch.hub) + +We require a few additional Python dependencies for preprocessing: +```bash +pip install fastBPE sacremoses subword_nmt +``` + +Interactive translation via PyTorch Hub: +```python +import torch + +# List available models +torch.hub.list('pytorch/fairseq') # [..., 'transformer.wmt16.en-de', ... ] + +# Load a transformer trained on WMT'16 En-De +# Note: WMT'19 models use fastBPE instead of subword_nmt, see instructions below +en2de = torch.hub.load('pytorch/fairseq', 'transformer.wmt16.en-de', + tokenizer='moses', bpe='subword_nmt') +en2de.eval() # disable dropout + +# The underlying model is available under the *models* attribute +assert isinstance(en2de.models[0], fairseq.models.transformer.TransformerModel) + +# Move model to GPU for faster translation +en2de.cuda() + +# Translate a sentence +en2de.translate('Hello world!') +# 'Hallo Welt!' + +# Batched translation +en2de.translate(['Hello world!', 'The cat sat on the mat.']) +# ['Hallo Welt!', 'Die Katze saß auf der Matte.'] +``` + +Loading custom models: +```python +from fairseq.models.transformer import TransformerModel +zh2en = TransformerModel.from_pretrained( + '/path/to/checkpoints', + checkpoint_file='checkpoint_best.pt', + data_name_or_path='data-bin/wmt17_zh_en_full', + bpe='subword_nmt', + bpe_codes='data-bin/wmt17_zh_en_full/zh.code' +) +zh2en.translate('你好 世界') +# 'Hello World' +``` + +If you are using a `transformer.wmt19` models, you will need to set the `bpe` +argument to `'fastbpe'` and (optionally) load the 4-model ensemble: +```python +en2de = torch.hub.load('pytorch/fairseq', 'transformer.wmt19.en-de', + checkpoint_file='model1.pt:model2.pt:model3.pt:model4.pt', + tokenizer='moses', bpe='fastbpe') +en2de.eval() # disable dropout +``` + +## Example usage (CLI tools) + +Generation with the binarized test sets can be run in batch mode as follows, e.g. for WMT 2014 English-French on a GTX-1080ti: +```bash +mkdir -p data-bin +curl https://dl.fbaipublicfiles.com/fairseq/models/wmt14.v2.en-fr.fconv-py.tar.bz2 | tar xvjf - -C data-bin +curl https://dl.fbaipublicfiles.com/fairseq/data/wmt14.v2.en-fr.newstest2014.tar.bz2 | tar xvjf - -C data-bin +fairseq-generate data-bin/wmt14.en-fr.newstest2014 \ + --path data-bin/wmt14.en-fr.fconv-py/model.pt \ + --beam 5 --batch-size 128 --remove-bpe | tee /tmp/gen.out +# ... +# | Translated 3003 sentences (96311 tokens) in 166.0s (580.04 tokens/s) +# | Generate test with beam=5: BLEU4 = 40.83, 67.5/46.9/34.4/25.5 (BP=1.000, ratio=1.006, syslen=83262, reflen=82787) + +# Compute BLEU score +grep ^H /tmp/gen.out | cut -f3- > /tmp/gen.out.sys +grep ^T /tmp/gen.out | cut -f2- > /tmp/gen.out.ref +fairseq-score --sys /tmp/gen.out.sys --ref /tmp/gen.out.ref +# BLEU4 = 40.83, 67.5/46.9/34.4/25.5 (BP=1.000, ratio=1.006, syslen=83262, reflen=82787) +``` + +## Training a new model + +### IWSLT'14 German to English (Transformer) + +The following instructions can be used to train a Transformer model on the [IWSLT'14 German to English dataset](http://workshop2014.iwslt.org/downloads/proceeding.pdf). + +First download and preprocess the data: +```bash +# Download and prepare the data +cd examples/translation/ +bash prepare-iwslt14.sh +cd ../.. + +# Preprocess/binarize the data +TEXT=examples/translation/iwslt14.tokenized.de-en +fairseq-preprocess --source-lang de --target-lang en \ + --trainpref $TEXT/train --validpref $TEXT/valid --testpref $TEXT/test \ + --destdir data-bin/iwslt14.tokenized.de-en \ + --workers 20 +``` + +Next we'll train a Transformer translation model over this data: +```bash +CUDA_VISIBLE_DEVICES=0 fairseq-train \ + data-bin/iwslt14.tokenized.de-en \ + --arch transformer_iwslt_de_en --share-decoder-input-output-embed \ + --optimizer adam --adam-betas '(0.9, 0.98)' --clip-norm 0.0 \ + --lr 5e-4 --lr-scheduler inverse_sqrt --warmup-updates 4000 \ + --dropout 0.3 --weight-decay 0.0001 \ + --criterion label_smoothed_cross_entropy --label-smoothing 0.1 \ + --max-tokens 4096 \ + --eval-bleu \ + --eval-bleu-args '{"beam": 5, "max_len_a": 1.2, "max_len_b": 10}' \ + --eval-bleu-detok moses \ + --eval-bleu-remove-bpe \ + --eval-bleu-print-samples \ + --best-checkpoint-metric bleu --maximize-best-checkpoint-metric +``` + +Finally we can evaluate our trained model: +```bash +fairseq-generate data-bin/iwslt14.tokenized.de-en \ + --path checkpoints/checkpoint_best.pt \ + --batch-size 128 --beam 5 --remove-bpe +``` + +### WMT'14 English to German (Convolutional) + +The following instructions can be used to train a Convolutional translation model on the WMT English to German dataset. +See the [Scaling NMT README](../scaling_nmt/README.md) for instructions to train a Transformer translation model on this data. + +The WMT English to German dataset can be preprocessed using the `prepare-wmt14en2de.sh` script. +By default it will produce a dataset that was modeled after [Attention Is All You Need (Vaswani et al., 2017)](https://arxiv.org/abs/1706.03762), but with additional news-commentary-v12 data from WMT'17. + +To use only data available in WMT'14 or to replicate results obtained in the original [Convolutional Sequence to Sequence Learning (Gehring et al., 2017)](https://arxiv.org/abs/1705.03122) paper, please use the `--icml17` option. + +```bash +# Download and prepare the data +cd examples/translation/ +# WMT'17 data: +bash prepare-wmt14en2de.sh +# or to use WMT'14 data: +# bash prepare-wmt14en2de.sh --icml17 +cd ../.. + +# Binarize the dataset +TEXT=examples/translation/wmt17_en_de +fairseq-preprocess \ + --source-lang en --target-lang de \ + --trainpref $TEXT/train --validpref $TEXT/valid --testpref $TEXT/test \ + --destdir data-bin/wmt17_en_de --thresholdtgt 0 --thresholdsrc 0 \ + --workers 20 + +# Train the model +mkdir -p checkpoints/fconv_wmt_en_de +fairseq-train \ + data-bin/wmt17_en_de \ + --arch fconv_wmt_en_de \ + --dropout 0.2 \ + --criterion label_smoothed_cross_entropy --label-smoothing 0.1 \ + --optimizer nag --clip-norm 0.1 \ + --lr 0.5 --lr-scheduler fixed --force-anneal 50 \ + --max-tokens 4000 \ + --save-dir checkpoints/fconv_wmt_en_de + +# Evaluate +fairseq-generate data-bin/wmt17_en_de \ + --path checkpoints/fconv_wmt_en_de/checkpoint_best.pt \ + --beam 5 --remove-bpe +``` + +### WMT'14 English to French +```bash +# Download and prepare the data +cd examples/translation/ +bash prepare-wmt14en2fr.sh +cd ../.. + +# Binarize the dataset +TEXT=examples/translation/wmt14_en_fr +fairseq-preprocess \ + --source-lang en --target-lang fr \ + --trainpref $TEXT/train --validpref $TEXT/valid --testpref $TEXT/test \ + --destdir data-bin/wmt14_en_fr --thresholdtgt 0 --thresholdsrc 0 \ + --workers 60 + +# Train the model +mkdir -p checkpoints/fconv_wmt_en_fr +fairseq-train \ + data-bin/wmt14_en_fr \ + --arch fconv_wmt_en_fr \ + --dropout 0.1 \ + --criterion label_smoothed_cross_entropy --label-smoothing 0.1 \ + --optimizer nag --clip-norm 0.1 \ + --lr 0.5 --lr-scheduler fixed --force-anneal 50 \ + --max-tokens 3000 \ + --save-dir checkpoints/fconv_wmt_en_fr + +# Evaluate +fairseq-generate \ + data-bin/fconv_wmt_en_fr \ + --path checkpoints/fconv_wmt_en_fr/checkpoint_best.pt \ + --beam 5 --remove-bpe +``` + +## Multilingual Translation + +We also support training multilingual translation models. In this example we'll +train a multilingual `{de,fr}-en` translation model using the IWSLT'17 datasets. + +Note that we use slightly different preprocessing here than for the IWSLT'14 +En-De data above. In particular we learn a joint BPE code for all three +languages and use fairseq-interactive and sacrebleu for scoring the test set. + +```bash +# First install sacrebleu and sentencepiece +pip install sacrebleu sentencepiece + +# Then download and preprocess the data +cd examples/translation/ +bash prepare-iwslt17-multilingual.sh +cd ../.. + +# Binarize the de-en dataset +TEXT=examples/translation/iwslt17.de_fr.en.bpe16k +fairseq-preprocess --source-lang de --target-lang en \ + --trainpref $TEXT/train.bpe.de-en \ + --validpref $TEXT/valid0.bpe.de-en,$TEXT/valid1.bpe.de-en,$TEXT/valid2.bpe.de-en,$TEXT/valid3.bpe.de-en,$TEXT/valid4.bpe.de-en,$TEXT/valid5.bpe.de-en \ + --destdir data-bin/iwslt17.de_fr.en.bpe16k \ + --workers 10 + +# Binarize the fr-en dataset +# NOTE: it's important to reuse the en dictionary from the previous step +fairseq-preprocess --source-lang fr --target-lang en \ + --trainpref $TEXT/train.bpe.fr-en \ + --validpref $TEXT/valid0.bpe.fr-en,$TEXT/valid1.bpe.fr-en,$TEXT/valid2.bpe.fr-en,$TEXT/valid3.bpe.fr-en,$TEXT/valid4.bpe.fr-en,$TEXT/valid5.bpe.fr-en \ + --tgtdict data-bin/iwslt17.de_fr.en.bpe16k/dict.en.txt \ + --destdir data-bin/iwslt17.de_fr.en.bpe16k \ + --workers 10 + +# Train a multilingual transformer model +# NOTE: the command below assumes 1 GPU, but accumulates gradients from +# 8 fwd/bwd passes to simulate training on 8 GPUs +mkdir -p checkpoints/multilingual_transformer +CUDA_VISIBLE_DEVICES=0 fairseq-train data-bin/iwslt17.de_fr.en.bpe16k/ \ + --max-epoch 50 \ + --ddp-backend=legacy_ddp \ + --task multilingual_translation --lang-pairs de-en,fr-en \ + --arch multilingual_transformer_iwslt_de_en \ + --share-decoders --share-decoder-input-output-embed \ + --optimizer adam --adam-betas '(0.9, 0.98)' \ + --lr 0.0005 --lr-scheduler inverse_sqrt \ + --warmup-updates 4000 --warmup-init-lr '1e-07' \ + --label-smoothing 0.1 --criterion label_smoothed_cross_entropy \ + --dropout 0.3 --weight-decay 0.0001 \ + --save-dir checkpoints/multilingual_transformer \ + --max-tokens 4000 \ + --update-freq 8 + +# Generate and score the test set with sacrebleu +SRC=de +sacrebleu --test-set iwslt17 --language-pair ${SRC}-en --echo src \ + | python scripts/spm_encode.py --model examples/translation/iwslt17.de_fr.en.bpe16k/sentencepiece.bpe.model \ + > iwslt17.test.${SRC}-en.${SRC}.bpe +cat iwslt17.test.${SRC}-en.${SRC}.bpe \ + | fairseq-interactive data-bin/iwslt17.de_fr.en.bpe16k/ \ + --task multilingual_translation --lang-pairs de-en,fr-en \ + --source-lang ${SRC} --target-lang en \ + --path checkpoints/multilingual_transformer/checkpoint_best.pt \ + --buffer-size 2000 --batch-size 128 \ + --beam 5 --remove-bpe=sentencepiece \ + > iwslt17.test.${SRC}-en.en.sys +grep ^H iwslt17.test.${SRC}-en.en.sys | cut -f3 \ + | sacrebleu --test-set iwslt17 --language-pair ${SRC}-en +``` + +##### Argument format during inference + +During inference it is required to specify a single `--source-lang` and +`--target-lang`, which indicates the inference langauge direction. +`--lang-pairs`, `--encoder-langtok`, `--decoder-langtok` have to be set to +the same value as training. diff --git a/examples/translation/prepare-iwslt14.sh b/examples/translation/prepare-iwslt14.sh new file mode 100644 index 0000000000000000000000000000000000000000..2fb6643fbccb58701dcbb77d91430e68a821ba38 --- /dev/null +++ b/examples/translation/prepare-iwslt14.sh @@ -0,0 +1,115 @@ +#!/usr/bin/env bash +# +# Adapted from https://github.com/facebookresearch/MIXER/blob/master/prepareData.sh + +echo 'Cloning Moses github repository (for tokenization scripts)...' +git clone https://github.com/moses-smt/mosesdecoder.git + +echo 'Cloning Subword NMT repository (for BPE pre-processing)...' +git clone https://github.com/rsennrich/subword-nmt.git + +SCRIPTS=mosesdecoder/scripts +TOKENIZER=$SCRIPTS/tokenizer/tokenizer.perl +LC=$SCRIPTS/tokenizer/lowercase.perl +CLEAN=$SCRIPTS/training/clean-corpus-n.perl +BPEROOT=subword-nmt/subword_nmt +BPE_TOKENS=10000 + +URL="http://dl.fbaipublicfiles.com/fairseq/data/iwslt14/de-en.tgz" +GZ=de-en.tgz + +if [ ! -d "$SCRIPTS" ]; then + echo "Please set SCRIPTS variable correctly to point to Moses scripts." + exit +fi + +src=de +tgt=en +lang=de-en +prep=iwslt14.tokenized.de-en +tmp=$prep/tmp +orig=orig + +mkdir -p $orig $tmp $prep + +echo "Downloading data from ${URL}..." +cd $orig +wget "$URL" + +if [ -f $GZ ]; then + echo "Data successfully downloaded." +else + echo "Data not successfully downloaded." + exit +fi + +tar zxvf $GZ +cd .. + +echo "pre-processing train data..." +for l in $src $tgt; do + f=train.tags.$lang.$l + tok=train.tags.$lang.tok.$l + + cat $orig/$lang/$f | \ + grep -v '<url>' | \ + grep -v '<talkid>' | \ + grep -v '<keywords>' | \ + sed -e 's/<title>//g' | \ + sed -e 's/<\/title>//g' | \ + sed -e 's/<description>//g' | \ + sed -e 's/<\/description>//g' | \ + perl $TOKENIZER -threads 8 -l $l > $tmp/$tok + echo "" +done +perl $CLEAN -ratio 1.5 $tmp/train.tags.$lang.tok $src $tgt $tmp/train.tags.$lang.clean 1 175 +for l in $src $tgt; do + perl $LC < $tmp/train.tags.$lang.clean.$l > $tmp/train.tags.$lang.$l +done + +echo "pre-processing valid/test data..." +for l in $src $tgt; do + for o in `ls $orig/$lang/IWSLT14.TED*.$l.xml`; do + fname=${o##*/} + f=$tmp/${fname%.*} + echo $o $f + grep '<seg id' $o | \ + sed -e 's/<seg id="[0-9]*">\s*//g' | \ + sed -e 's/\s*<\/seg>\s*//g' | \ + sed -e "s/\’/\'/g" | \ + perl $TOKENIZER -threads 8 -l $l | \ + perl $LC > $f + echo "" + done +done + + +echo "creating train, valid, test..." +for l in $src $tgt; do + awk '{if (NR%23 == 0) print $0; }' $tmp/train.tags.de-en.$l > $tmp/valid.$l + awk '{if (NR%23 != 0) print $0; }' $tmp/train.tags.de-en.$l > $tmp/train.$l + + cat $tmp/IWSLT14.TED.dev2010.de-en.$l \ + $tmp/IWSLT14.TEDX.dev2012.de-en.$l \ + $tmp/IWSLT14.TED.tst2010.de-en.$l \ + $tmp/IWSLT14.TED.tst2011.de-en.$l \ + $tmp/IWSLT14.TED.tst2012.de-en.$l \ + > $tmp/test.$l +done + +TRAIN=$tmp/train.en-de +BPE_CODE=$prep/code +rm -f $TRAIN +for l in $src $tgt; do + cat $tmp/train.$l >> $TRAIN +done + +echo "learn_bpe.py on ${TRAIN}..." +python $BPEROOT/learn_bpe.py -s $BPE_TOKENS < $TRAIN > $BPE_CODE + +for L in $src $tgt; do + for f in train.$L valid.$L test.$L; do + echo "apply_bpe.py to ${f}..." + python $BPEROOT/apply_bpe.py -c $BPE_CODE < $tmp/$f > $prep/$f + done +done diff --git a/examples/translation/prepare-iwslt17-multilingual.sh b/examples/translation/prepare-iwslt17-multilingual.sh new file mode 100644 index 0000000000000000000000000000000000000000..23be87555322bc03b13e9d95951d88b1a442f97a --- /dev/null +++ b/examples/translation/prepare-iwslt17-multilingual.sh @@ -0,0 +1,133 @@ +#!/bin/bash +# Copyright (c) Facebook, Inc. and its affiliates. +# All rights reserved. +# +# This source code is licensed under the license found in the +# LICENSE file in the root directory of this source tree. + +SRCS=( + "de" + "fr" +) +TGT=en + +ROOT=$(dirname "$0") +SCRIPTS=$ROOT/../../scripts +SPM_TRAIN=$SCRIPTS/spm_train.py +SPM_ENCODE=$SCRIPTS/spm_encode.py + +BPESIZE=16384 +ORIG=$ROOT/iwslt17_orig +DATA=$ROOT/iwslt17.de_fr.en.bpe16k +mkdir -p "$ORIG" "$DATA" + +TRAIN_MINLEN=1 # remove sentences with <1 BPE token +TRAIN_MAXLEN=250 # remove sentences with >250 BPE tokens + +URLS=( + "https://wit3.fbk.eu/archive/2017-01-trnted/texts/de/en/de-en.tgz" + "https://wit3.fbk.eu/archive/2017-01-trnted/texts/fr/en/fr-en.tgz" +) +ARCHIVES=( + "de-en.tgz" + "fr-en.tgz" +) +VALID_SETS=( + "IWSLT17.TED.dev2010.de-en IWSLT17.TED.tst2010.de-en IWSLT17.TED.tst2011.de-en IWSLT17.TED.tst2012.de-en IWSLT17.TED.tst2013.de-en IWSLT17.TED.tst2014.de-en IWSLT17.TED.tst2015.de-en" + "IWSLT17.TED.dev2010.fr-en IWSLT17.TED.tst2010.fr-en IWSLT17.TED.tst2011.fr-en IWSLT17.TED.tst2012.fr-en IWSLT17.TED.tst2013.fr-en IWSLT17.TED.tst2014.fr-en IWSLT17.TED.tst2015.fr-en" +) + +# download and extract data +for ((i=0;i<${#URLS[@]};++i)); do + ARCHIVE=$ORIG/${ARCHIVES[i]} + if [ -f "$ARCHIVE" ]; then + echo "$ARCHIVE already exists, skipping download" + else + URL=${URLS[i]} + wget -P "$ORIG" "$URL" + if [ -f "$ARCHIVE" ]; then + echo "$URL successfully downloaded." + else + echo "$URL not successfully downloaded." + exit 1 + fi + fi + FILE=${ARCHIVE: -4} + if [ -e "$FILE" ]; then + echo "$FILE already exists, skipping extraction" + else + tar -C "$ORIG" -xzvf "$ARCHIVE" + fi +done + +echo "pre-processing train data..." +for SRC in "${SRCS[@]}"; do + for LANG in "${SRC}" "${TGT}"; do + cat "$ORIG/${SRC}-${TGT}/train.tags.${SRC}-${TGT}.${LANG}" \ + | grep -v '<url>' \ + | grep -v '<talkid>' \ + | grep -v '<keywords>' \ + | grep -v '<speaker>' \ + | grep -v '<reviewer' \ + | grep -v '<translator' \ + | grep -v '<doc' \ + | grep -v '</doc>' \ + | sed -e 's/<title>//g' \ + | sed -e 's/<\/title>//g' \ + | sed -e 's/<description>//g' \ + | sed -e 's/<\/description>//g' \ + | sed 's/^\s*//g' \ + | sed 's/\s*$//g' \ + > "$DATA/train.${SRC}-${TGT}.${LANG}" + done +done + +echo "pre-processing valid data..." +for ((i=0;i<${#SRCS[@]};++i)); do + SRC=${SRCS[i]} + VALID_SET=(${VALID_SETS[i]}) + for ((j=0;j<${#VALID_SET[@]};++j)); do + FILE=${VALID_SET[j]} + for LANG in "$SRC" "$TGT"; do + grep '<seg id' "$ORIG/${SRC}-${TGT}/${FILE}.${LANG}.xml" \ + | sed -e 's/<seg id="[0-9]*">\s*//g' \ + | sed -e 's/\s*<\/seg>\s*//g' \ + | sed -e "s/\’/\'/g" \ + > "$DATA/valid${j}.${SRC}-${TGT}.${LANG}" + done + done +done + +# learn BPE with sentencepiece +TRAIN_FILES=$(for SRC in "${SRCS[@]}"; do echo $DATA/train.${SRC}-${TGT}.${SRC}; echo $DATA/train.${SRC}-${TGT}.${TGT}; done | tr "\n" ",") +echo "learning joint BPE over ${TRAIN_FILES}..." +python "$SPM_TRAIN" \ + --input=$TRAIN_FILES \ + --model_prefix=$DATA/sentencepiece.bpe \ + --vocab_size=$BPESIZE \ + --character_coverage=1.0 \ + --model_type=bpe + +# encode train/valid +echo "encoding train with learned BPE..." +for SRC in "${SRCS[@]}"; do + python "$SPM_ENCODE" \ + --model "$DATA/sentencepiece.bpe.model" \ + --output_format=piece \ + --inputs $DATA/train.${SRC}-${TGT}.${SRC} $DATA/train.${SRC}-${TGT}.${TGT} \ + --outputs $DATA/train.bpe.${SRC}-${TGT}.${SRC} $DATA/train.bpe.${SRC}-${TGT}.${TGT} \ + --min-len $TRAIN_MINLEN --max-len $TRAIN_MAXLEN +done + +echo "encoding valid with learned BPE..." +for ((i=0;i<${#SRCS[@]};++i)); do + SRC=${SRCS[i]} + VALID_SET=(${VALID_SETS[i]}) + for ((j=0;j<${#VALID_SET[@]};++j)); do + python "$SPM_ENCODE" \ + --model "$DATA/sentencepiece.bpe.model" \ + --output_format=piece \ + --inputs $DATA/valid${j}.${SRC}-${TGT}.${SRC} $DATA/valid${j}.${SRC}-${TGT}.${TGT} \ + --outputs $DATA/valid${j}.bpe.${SRC}-${TGT}.${SRC} $DATA/valid${j}.bpe.${SRC}-${TGT}.${TGT} + done +done diff --git a/examples/translation/prepare-wmt14en2de.sh b/examples/translation/prepare-wmt14en2de.sh new file mode 100644 index 0000000000000000000000000000000000000000..6702c88b568c9e680b525593ff0c9fb0a474825d --- /dev/null +++ b/examples/translation/prepare-wmt14en2de.sh @@ -0,0 +1,142 @@ +#!/bin/bash +# Adapted from https://github.com/facebookresearch/MIXER/blob/master/prepareData.sh + +echo 'Cloning Moses github repository (for tokenization scripts)...' +git clone https://github.com/moses-smt/mosesdecoder.git + +echo 'Cloning Subword NMT repository (for BPE pre-processing)...' +git clone https://github.com/rsennrich/subword-nmt.git + +SCRIPTS=mosesdecoder/scripts +TOKENIZER=$SCRIPTS/tokenizer/tokenizer.perl +CLEAN=$SCRIPTS/training/clean-corpus-n.perl +NORM_PUNC=$SCRIPTS/tokenizer/normalize-punctuation.perl +REM_NON_PRINT_CHAR=$SCRIPTS/tokenizer/remove-non-printing-char.perl +BPEROOT=subword-nmt/subword_nmt +BPE_TOKENS=40000 + +URLS=( + "http://statmt.org/wmt13/training-parallel-europarl-v7.tgz" + "http://statmt.org/wmt13/training-parallel-commoncrawl.tgz" + "http://data.statmt.org/wmt17/translation-task/training-parallel-nc-v12.tgz" + "http://data.statmt.org/wmt17/translation-task/dev.tgz" + "http://statmt.org/wmt14/test-full.tgz" +) +FILES=( + "training-parallel-europarl-v7.tgz" + "training-parallel-commoncrawl.tgz" + "training-parallel-nc-v12.tgz" + "dev.tgz" + "test-full.tgz" +) +CORPORA=( + "training/europarl-v7.de-en" + "commoncrawl.de-en" + "training/news-commentary-v12.de-en" +) + +# This will make the dataset compatible to the one used in "Convolutional Sequence to Sequence Learning" +# https://arxiv.org/abs/1705.03122 +if [ "$1" == "--icml17" ]; then + URLS[2]="http://statmt.org/wmt14/training-parallel-nc-v9.tgz" + FILES[2]="training-parallel-nc-v9.tgz" + CORPORA[2]="training/news-commentary-v9.de-en" + OUTDIR=wmt14_en_de +else + OUTDIR=wmt17_en_de +fi + +if [ ! -d "$SCRIPTS" ]; then + echo "Please set SCRIPTS variable correctly to point to Moses scripts." + exit +fi + +src=en +tgt=de +lang=en-de +prep=$OUTDIR +tmp=$prep/tmp +orig=orig +dev=dev/newstest2013 + +mkdir -p $orig $tmp $prep + +cd $orig + +for ((i=0;i<${#URLS[@]};++i)); do + file=${FILES[i]} + if [ -f $file ]; then + echo "$file already exists, skipping download" + else + url=${URLS[i]} + wget "$url" + if [ -f $file ]; then + echo "$url successfully downloaded." + else + echo "$url not successfully downloaded." + exit -1 + fi + if [ ${file: -4} == ".tgz" ]; then + tar zxvf $file + elif [ ${file: -4} == ".tar" ]; then + tar xvf $file + fi + fi +done +cd .. + +echo "pre-processing train data..." +for l in $src $tgt; do + rm $tmp/train.tags.$lang.tok.$l + for f in "${CORPORA[@]}"; do + cat $orig/$f.$l | \ + perl $NORM_PUNC $l | \ + perl $REM_NON_PRINT_CHAR | \ + perl $TOKENIZER -threads 8 -a -l $l >> $tmp/train.tags.$lang.tok.$l + done +done + +echo "pre-processing test data..." +for l in $src $tgt; do + if [ "$l" == "$src" ]; then + t="src" + else + t="ref" + fi + grep '<seg id' $orig/test-full/newstest2014-deen-$t.$l.sgm | \ + sed -e 's/<seg id="[0-9]*">\s*//g' | \ + sed -e 's/\s*<\/seg>\s*//g' | \ + sed -e "s/\’/\'/g" | \ + perl $TOKENIZER -threads 8 -a -l $l > $tmp/test.$l + echo "" +done + +echo "splitting train and valid..." +for l in $src $tgt; do + awk '{if (NR%100 == 0) print $0; }' $tmp/train.tags.$lang.tok.$l > $tmp/valid.$l + awk '{if (NR%100 != 0) print $0; }' $tmp/train.tags.$lang.tok.$l > $tmp/train.$l +done + +TRAIN=$tmp/train.de-en +BPE_CODE=$prep/code +rm -f $TRAIN +for l in $src $tgt; do + cat $tmp/train.$l >> $TRAIN +done + +echo "learn_bpe.py on ${TRAIN}..." +python $BPEROOT/learn_bpe.py -s $BPE_TOKENS < $TRAIN > $BPE_CODE + +for L in $src $tgt; do + for f in train.$L valid.$L test.$L; do + echo "apply_bpe.py to ${f}..." + python $BPEROOT/apply_bpe.py -c $BPE_CODE < $tmp/$f > $tmp/bpe.$f + done +done + +perl $CLEAN -ratio 1.5 $tmp/bpe.train $src $tgt $prep/train 1 250 +perl $CLEAN -ratio 1.5 $tmp/bpe.valid $src $tgt $prep/valid 1 250 + +for L in $src $tgt; do + cp $tmp/bpe.test.$L $prep/test.$L +done diff --git a/examples/translation/prepare-wmt14en2fr.sh b/examples/translation/prepare-wmt14en2fr.sh new file mode 100644 index 0000000000000000000000000000000000000000..2ac97a5b76fab255449493488ed8bd67350a7bac --- /dev/null +++ b/examples/translation/prepare-wmt14en2fr.sh @@ -0,0 +1,136 @@ +#!/bin/bash +# Adapted from https://github.com/facebookresearch/MIXER/blob/master/prepareData.sh + +echo 'Cloning Moses github repository (for tokenization scripts)...' +git clone https://github.com/moses-smt/mosesdecoder.git + +echo 'Cloning Subword NMT repository (for BPE pre-processing)...' +git clone https://github.com/rsennrich/subword-nmt.git + +SCRIPTS=mosesdecoder/scripts +TOKENIZER=$SCRIPTS/tokenizer/tokenizer.perl +CLEAN=$SCRIPTS/training/clean-corpus-n.perl +NORM_PUNC=$SCRIPTS/tokenizer/normalize-punctuation.perl +REM_NON_PRINT_CHAR=$SCRIPTS/tokenizer/remove-non-printing-char.perl +BPEROOT=subword-nmt/subword_nmt +BPE_TOKENS=40000 + +URLS=( + "http://statmt.org/wmt13/training-parallel-europarl-v7.tgz" + "http://statmt.org/wmt13/training-parallel-commoncrawl.tgz" + "http://statmt.org/wmt13/training-parallel-un.tgz" + "http://statmt.org/wmt14/training-parallel-nc-v9.tgz" + "http://statmt.org/wmt10/training-giga-fren.tar" + "http://statmt.org/wmt14/test-full.tgz" +) +FILES=( + "training-parallel-europarl-v7.tgz" + "training-parallel-commoncrawl.tgz" + "training-parallel-un.tgz" + "training-parallel-nc-v9.tgz" + "training-giga-fren.tar" + "test-full.tgz" +) +CORPORA=( + "training/europarl-v7.fr-en" + "commoncrawl.fr-en" + "un/undoc.2000.fr-en" + "training/news-commentary-v9.fr-en" + "giga-fren.release2.fixed" +) + +if [ ! -d "$SCRIPTS" ]; then + echo "Please set SCRIPTS variable correctly to point to Moses scripts." + exit +fi + +src=en +tgt=fr +lang=en-fr +prep=wmt14_en_fr +tmp=$prep/tmp +orig=orig + +mkdir -p $orig $tmp $prep + +cd $orig + +for ((i=0;i<${#URLS[@]};++i)); do + file=${FILES[i]} + if [ -f $file ]; then + echo "$file already exists, skipping download" + else + url=${URLS[i]} + wget "$url" + if [ -f $file ]; then + echo "$url successfully downloaded." + else + echo "$url not successfully downloaded." + exit -1 + fi + if [ ${file: -4} == ".tgz" ]; then + tar zxvf $file + elif [ ${file: -4} == ".tar" ]; then + tar xvf $file + fi + fi +done + +gunzip giga-fren.release2.fixed.*.gz +cd .. + +echo "pre-processing train data..." +for l in $src $tgt; do + rm $tmp/train.tags.$lang.tok.$l + for f in "${CORPORA[@]}"; do + cat $orig/$f.$l | \ + perl $NORM_PUNC $l | \ + perl $REM_NON_PRINT_CHAR | \ + perl $TOKENIZER -threads 8 -a -l $l >> $tmp/train.tags.$lang.tok.$l + done +done + +echo "pre-processing test data..." +for l in $src $tgt; do + if [ "$l" == "$src" ]; then + t="src" + else + t="ref" + fi + grep '<seg id' $orig/test-full/newstest2014-fren-$t.$l.sgm | \ + sed -e 's/<seg id="[0-9]*">\s*//g' | \ + sed -e 's/\s*<\/seg>\s*//g' | \ + sed -e "s/\’/\'/g" | \ + perl $TOKENIZER -threads 8 -a -l $l > $tmp/test.$l + echo "" +done + +echo "splitting train and valid..." +for l in $src $tgt; do + awk '{if (NR%1333 == 0) print $0; }' $tmp/train.tags.$lang.tok.$l > $tmp/valid.$l + awk '{if (NR%1333 != 0) print $0; }' $tmp/train.tags.$lang.tok.$l > $tmp/train.$l +done + +TRAIN=$tmp/train.fr-en +BPE_CODE=$prep/code +rm -f $TRAIN +for l in $src $tgt; do + cat $tmp/train.$l >> $TRAIN +done + +echo "learn_bpe.py on ${TRAIN}..." +python $BPEROOT/learn_bpe.py -s $BPE_TOKENS < $TRAIN > $BPE_CODE + +for L in $src $tgt; do + for f in train.$L valid.$L test.$L; do + echo "apply_bpe.py to ${f}..." + python $BPEROOT/apply_bpe.py -c $BPE_CODE < $tmp/$f > $tmp/bpe.$f + done +done + +perl $CLEAN -ratio 1.5 $tmp/bpe.train $src $tgt $prep/train 1 250 +perl $CLEAN -ratio 1.5 $tmp/bpe.valid $src $tgt $prep/valid 1 250 + +for L in $src $tgt; do + cp $tmp/bpe.test.$L $prep/test.$L +done diff --git a/examples/translation_moe/README.md b/examples/translation_moe/README.md new file mode 100644 index 0000000000000000000000000000000000000000..2e5c8af617f410f64ca38d29447bd05b6af8c5a8 --- /dev/null +++ b/examples/translation_moe/README.md @@ -0,0 +1,89 @@ +# Mixture Models for Diverse Machine Translation: Tricks of the Trade (Shen et al., 2019) + +This page includes instructions for reproducing results from the paper [Mixture Models for Diverse Machine Translation: Tricks of the Trade (Shen et al., 2019)](https://arxiv.org/abs/1902.07816). + +## Download data + +First, follow the [instructions to download and preprocess the WMT'17 En-De dataset](../translation#prepare-wmt14en2desh). +Make sure to learn a joint vocabulary by passing the `--joined-dictionary` option to `fairseq-preprocess`. + +## Train a model + +Then we can train a mixture of experts model using the `translation_moe` task. +Use the `--method` flag to choose the MoE variant; we support hard mixtures with a learned or uniform prior (`--method hMoElp` and `hMoEup`, respectively) and soft mixures (`--method sMoElp` and `sMoEup`). +The model is trained with online responsibility assignment and shared parameterization. + +The following command will train a `hMoElp` model with `3` experts: +```bash +fairseq-train --ddp-backend='legacy_ddp' \ + data-bin/wmt17_en_de \ + --max-update 100000 \ + --task translation_moe --user-dir examples/translation_moe/translation_moe_src \ + --method hMoElp --mean-pool-gating-network \ + --num-experts 3 \ + --arch transformer_wmt_en_de --share-all-embeddings \ + --optimizer adam --adam-betas '(0.9, 0.98)' --clip-norm 0.0 \ + --lr-scheduler inverse_sqrt --warmup-init-lr 1e-07 --warmup-updates 4000 \ + --lr 0.0007 \ + --dropout 0.1 --weight-decay 0.0 --criterion cross_entropy \ + --max-tokens 3584 +``` + +## Translate + +Once a model is trained, we can generate translations from different experts using the `--gen-expert` option. +For example, to generate from expert 0: +```bash +fairseq-generate data-bin/wmt17_en_de \ + --path checkpoints/checkpoint_best.pt \ + --beam 1 --remove-bpe \ + --task translation_moe --user-dir examples/translation_moe/translation_moe_src \ + --method hMoElp --mean-pool-gating-network \ + --num-experts 3 \ + --gen-expert 0 +``` + +## Evaluate + +First download a tokenized version of the WMT'14 En-De test set with multiple references: +```bash +wget dl.fbaipublicfiles.com/fairseq/data/wmt14-en-de.extra_refs.tok +``` + +Next apply BPE on the fly and run generation for each expert: +```bash +BPE_CODE=examples/translation/wmt17_en_de/code +for EXPERT in $(seq 0 2); do \ + cat wmt14-en-de.extra_refs.tok \ + | grep ^S | cut -f 2 \ + | fairseq-interactive data-bin/wmt17_en_de \ + --path checkpoints/checkpoint_best.pt \ + --beam 1 \ + --bpe subword_nmt --bpe-codes $BPE_CODE \ + --buffer-size 500 --max-tokens 6000 \ + --task translation_moe --user-dir examples/translation_moe/translation_moe_src \ + --method hMoElp --mean-pool-gating-network \ + --num-experts 3 \ + --gen-expert $EXPERT ; \ +done > wmt14-en-de.extra_refs.tok.gen.3experts +``` + +Finally use `score_moe.py` to compute pairwise BLUE and average oracle BLEU: +```bash +python examples/translation_moe/score.py --sys wmt14-en-de.extra_refs.tok.gen.3experts --ref wmt14-en-de.extra_refs.tok +# pairwise BLEU: 48.26 +# #refs covered: 2.11 +# multi-reference BLEU (leave-one-out): 59.46 +``` +This matches row 3 from Table 7 in the paper. + +## Citation + +```bibtex +@article{shen2019mixture, + title = {Mixture Models for Diverse Machine Translation: Tricks of the Trade}, + author = {Tianxiao Shen and Myle Ott and Michael Auli and Marc'Aurelio Ranzato}, + journal = {International Conference on Machine Learning}, + year = 2019, +} +``` diff --git a/examples/translation_moe/score.py b/examples/translation_moe/score.py new file mode 100644 index 0000000000000000000000000000000000000000..9a529a985019710ea202cb6bf28ae071c0ce4135 --- /dev/null +++ b/examples/translation_moe/score.py @@ -0,0 +1,197 @@ +#!/usr/bin/env python3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +Scoring script for computing pairwise BLEU and multi-ref BLEU over a set of +candidate hypotheses. + +See `"Mixture Models for Diverse Machine Translation: Tricks of the Trade" +(Shen et al., 2019) <https://arxiv.org/abs/1902.07816>`_. +""" + +import argparse +import random +import sys +from itertools import chain + +import numpy as np +from sacrebleu import compute_bleu, corpus_bleu as _corpus_bleu + + +def main(): + parser = argparse.ArgumentParser(sys.argv[0]) + parser.add_argument( + "--sys", nargs="*", default="", metavar="FILE", help="path to system output" + ) + parser.add_argument("--ref", default="", metavar="FILE", help="path to references") + parser.add_argument( + "--output", + default="", + metavar="FILE", + help="print outputs into a pretty format", + ) + args = parser.parse_args() + + if args.sys: + src, tgt, hypos, log_probs = load_sys(args.sys) + print("pairwise BLEU: %.2f" % pairwise(hypos)) + if args.output: + merge(src, tgt, hypos, log_probs, args.output) + + if args.ref: + _, _, refs = load_ref(args.ref) + if args.sys: + multi_ref(refs, hypos) + else: + intra_ref(refs) + + +def dictolist(d): + a = sorted(d.items(), key=lambda i: i[0]) + return [i[1] for i in a] + + +def load_sys(paths): + src, tgt, hypos, log_probs = {}, {}, {}, {} + for path in paths: + with open(path) as f: + for line in f: + line = line.rstrip() + # S: source + # T: target + # D: detokenized system output + if line.startswith(("S-", "T-", "D-")): + i = int(line[line.find("-") + 1 : line.find("\t")]) + if line.startswith("S-"): + src[i] = line.split("\t")[1] + if line.startswith("T-"): + tgt[i] = line.split("\t")[1] + if line.startswith("D-"): + if i not in hypos: + hypos[i] = [] + log_probs[i] = [] + hypos[i].append(line.split("\t")[2]) + log_probs[i].append(float(line.split("\t")[1])) + return dictolist(src), dictolist(tgt), dictolist(hypos), dictolist(log_probs) + + +def load_ref(path): + with open(path) as f: + lines = f.readlines() + src, tgt, refs = [], [], [] + i = 0 + while i < len(lines): + if lines[i].startswith("S-"): + src.append(lines[i].split("\t")[1].rstrip()) + i += 1 + elif lines[i].startswith("T-"): + tgt.append(lines[i].split("\t")[1].rstrip()) + i += 1 + else: + a = [] + while i < len(lines) and lines[i].startswith("R"): + a.append(lines[i].split("\t")[1].rstrip()) + i += 1 + refs.append(a) + return src, tgt, refs + + +def merge(src, tgt, hypos, log_probs, path): + with open(path, "w") as f: + for s, t, hs, lps in zip(src, tgt, hypos, log_probs): + f.write(s + "\n") + f.write(t + "\n") + f.write("\n") + for h, lp in zip(hs, lps): + f.write("\t%f\t%s\n" % (lp, h.strip())) + f.write("------------------------------------------------------\n") + + +def corpus_bleu(sys_stream, ref_streams): + bleu = _corpus_bleu(sys_stream, ref_streams, tokenize="none") + return bleu.score + + +def sentence_bleu(hypothesis, reference): + bleu = _corpus_bleu(hypothesis, reference) + for i in range(1, 4): + bleu.counts[i] += 1 + bleu.totals[i] += 1 + bleu = compute_bleu( + bleu.counts, + bleu.totals, + bleu.sys_len, + bleu.ref_len, + smooth_method="exp", + ) + return bleu.score + + +def pairwise(sents): + _ref, _hypo = [], [] + for s in sents: + for i in range(len(s)): + for j in range(len(s)): + if i != j: + _ref.append(s[i]) + _hypo.append(s[j]) + return corpus_bleu(_hypo, [_ref]) + + +def multi_ref(refs, hypos): + _ref, _hypo = [], [] + ref_cnt = 0 + assert len(refs) == len(hypos) + + # count number of refs covered + for rs, hs in zip(refs, hypos): + a = set() + for h in hs: + s = [sentence_bleu(h, r) for r in rs] + j = np.argmax(s) + _ref.append(rs[j]) + _hypo.append(h) + best = [k for k in range(len(rs)) if s[k] == s[j]] + a.add(random.choice(best)) + ref_cnt += len(a) + print("#refs covered: %.2f" % (ref_cnt / len(refs))) + + # transpose refs and hypos + refs = list(zip(*refs)) + hypos = list(zip(*hypos)) + + # compute multi-ref corpus BLEU (leave-one-out to be comparable to intra_ref) + k = len(hypos) + m = len(refs) + flat_hypos = [hypos[j][i] for i in range(len(hypos[0])) for j in range(k)] + duplicated_refs = [[ref for ref in refs_i for _ in range(k)] for refs_i in refs] + loo_bleus = [] + for held_out_ref in range(m): + remaining_refs = ( + duplicated_refs[:held_out_ref] + duplicated_refs[held_out_ref + 1 :] + ) + assert len(remaining_refs) == m - 1 + loo_bleus.append(corpus_bleu(flat_hypos, remaining_refs)) + print("average multi-reference BLEU (leave-one-out): %.2f" % np.mean(loo_bleus)) + + +def intra_ref(refs): + print("ref pairwise BLEU: %.2f" % pairwise(refs)) + refs = list(zip(*refs)) + m = len(refs) + concat_h = [] + concat_rest = [[] for j in range(m - 1)] + for i, h in enumerate(refs): + rest = refs[:i] + refs[i + 1 :] + concat_h.append(h) + for j in range(m - 1): + concat_rest[j].extend(rest[j]) + concat_h = list(chain.from_iterable(concat_h)) + bleu = corpus_bleu(concat_h, concat_rest) + print("multi-reference BLEU (leave-one-out): %.2f" % bleu) + + +if __name__ == "__main__": + main() diff --git a/examples/translation_moe/translation_moe_src/__init__.py b/examples/translation_moe/translation_moe_src/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..c0abe53e973b4bb31cfb062708965d002c79b6e7 --- /dev/null +++ b/examples/translation_moe/translation_moe_src/__init__.py @@ -0,0 +1,6 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from . import translation_moe # noqa diff --git a/examples/translation_moe/translation_moe_src/logsumexp_moe.py b/examples/translation_moe/translation_moe_src/logsumexp_moe.py new file mode 100644 index 0000000000000000000000000000000000000000..fb299daecbc2b15fb66555bbfb8d1d983e481518 --- /dev/null +++ b/examples/translation_moe/translation_moe_src/logsumexp_moe.py @@ -0,0 +1,26 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch + + +class LogSumExpMoE(torch.autograd.Function): + """Standard LogSumExp forward pass, but use *posterior* for the backward. + + See `"Mixture Models for Diverse Machine Translation: Tricks of the Trade" + (Shen et al., 2019) <https://arxiv.org/abs/1902.07816>`_. + """ + + @staticmethod + def forward(ctx, logp, posterior, dim=-1): + ctx.save_for_backward(posterior) + ctx.dim = dim + return torch.logsumexp(logp, dim=dim) + + @staticmethod + def backward(ctx, grad_output): + (posterior,) = ctx.saved_tensors + grad_logp = grad_output.unsqueeze(ctx.dim) * posterior + return grad_logp, None, None diff --git a/examples/translation_moe/translation_moe_src/mean_pool_gating_network.py b/examples/translation_moe/translation_moe_src/mean_pool_gating_network.py new file mode 100644 index 0000000000000000000000000000000000000000..efc7ae40bf8fed6c2384cbc6f94477c4caa4c10c --- /dev/null +++ b/examples/translation_moe/translation_moe_src/mean_pool_gating_network.py @@ -0,0 +1,50 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn.functional as F + + +class MeanPoolGatingNetwork(torch.nn.Module): + """A simple mean-pooling gating network for selecting experts. + + This module applies mean pooling over an encoder's output and returns + reponsibilities for each expert. The encoder format is expected to match + :class:`fairseq.models.transformer.TransformerEncoder`. + """ + + def __init__(self, embed_dim, num_experts, dropout=None): + super().__init__() + self.embed_dim = embed_dim + self.num_experts = num_experts + + self.fc1 = torch.nn.Linear(embed_dim, embed_dim) + self.dropout = torch.nn.Dropout(dropout) if dropout is not None else None + self.fc2 = torch.nn.Linear(embed_dim, num_experts) + + def forward(self, encoder_out): + if not ( + "encoder_out" in encoder_out + and "encoder_padding_mask" in encoder_out + and encoder_out["encoder_out"][0].size(2) == self.embed_dim + ): + raise ValueError("Unexpected format for encoder_out") + + # mean pooling over time + encoder_padding_mask = encoder_out["encoder_padding_mask"][0] # B x T + encoder_out = encoder_out["encoder_out"][0].transpose(0, 1) # B x T x C + if encoder_padding_mask is not None: + encoder_out = encoder_out.clone() # required because of transpose above + encoder_out[encoder_padding_mask] = 0 + ntokens = torch.sum(~encoder_padding_mask, dim=1, keepdim=True) + x = torch.sum(encoder_out, dim=1) / ntokens.type_as(encoder_out) + else: + x = torch.mean(encoder_out, dim=1) + + x = torch.tanh(self.fc1(x)) + if self.dropout is not None: + x = self.dropout(x) + x = self.fc2(x) + return F.log_softmax(x, dim=-1, dtype=torch.float32).type_as(x) diff --git a/examples/translation_moe/translation_moe_src/translation_moe.py b/examples/translation_moe/translation_moe_src/translation_moe.py new file mode 100644 index 0000000000000000000000000000000000000000..7f28c32dd6152f53d6922cdfccfa903e0bdc5829 --- /dev/null +++ b/examples/translation_moe/translation_moe_src/translation_moe.py @@ -0,0 +1,258 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from dataclasses import dataclass, field +import torch +from omegaconf import II + +from fairseq import metrics, utils +from fairseq.dataclass import ChoiceEnum +from fairseq.tasks import register_task +from fairseq.tasks.translation import TranslationConfig, TranslationTask + +from .logsumexp_moe import LogSumExpMoE +from .mean_pool_gating_network import MeanPoolGatingNetwork + + +METHOD_CHOICES = ChoiceEnum(["sMoElp", "sMoEup", "hMoElp", "hMoEup"]) + + +@dataclass +class TranslationMoEConfig(TranslationConfig): + method: METHOD_CHOICES = field( + default="hMoEup", + metadata={"help": "MoE method"}, + ) + num_experts: int = field( + default=3, + metadata={"help": "number of experts"}, + ) + mean_pool_gating_network: bool = field( + default=False, + metadata={"help": "use a simple mean-pooling gating network"}, + ) + mean_pool_gating_network_dropout: float = field( + default=0, + metadata={"help": "dropout for mean-pooling gating network"}, + ) + mean_pool_gating_network_encoder_dim: int = field( + default=0, + metadata={"help": "encoder output dim for mean-pooling gating network"}, + ) + gen_expert: int = field( + default=0, + metadata={"help": "which expert to use for generation"}, + ) + sentence_avg: bool = II("optimization.sentence_avg") + + +@register_task("translation_moe", dataclass=TranslationMoEConfig) +class TranslationMoETask(TranslationTask): + """ + Translation task for Mixture of Experts (MoE) models. + + See `"Mixture Models for Diverse Machine Translation: Tricks of the Trade" + (Shen et al., 2019) <https://arxiv.org/abs/1902.07816>`_. + + Args: + src_dict (~fairseq.data.Dictionary): dictionary for the source language + tgt_dict (~fairseq.data.Dictionary): dictionary for the target language + + .. note:: + + The translation task is compatible with :mod:`fairseq-train`, + :mod:`fairseq-generate` and :mod:`fairseq-interactive`. + + The translation task provides the following additional command-line + arguments: + + .. argparse:: + :ref: fairseq.tasks.translation_parser + :prog: + """ + + cfg: TranslationMoEConfig + + def __init__(self, cfg: TranslationMoEConfig, src_dict, tgt_dict): + if cfg.method == "sMoElp": + # soft MoE with learned prior + self.uniform_prior = False + self.hard_selection = False + elif cfg.method == "sMoEup": + # soft MoE with uniform prior + self.uniform_prior = True + self.hard_selection = False + elif cfg.method == "hMoElp": + # hard MoE with learned prior + self.uniform_prior = False + self.hard_selection = True + elif cfg.method == "hMoEup": + # hard MoE with uniform prior + self.uniform_prior = True + self.hard_selection = True + + # add indicator tokens for each expert + for i in range(cfg.num_experts): + # add to both dictionaries in case we're sharing embeddings + src_dict.add_symbol("<expert_{}>".format(i)) + tgt_dict.add_symbol("<expert_{}>".format(i)) + + super().__init__(cfg, src_dict, tgt_dict) + + def build_model(self, cfg): + from fairseq import models + + model = models.build_model(cfg, self) + if not self.uniform_prior and not hasattr(model, "gating_network"): + if self.cfg.mean_pool_gating_network: + if self.cfg.mean_pool_gating_network_encoder_dim > 0: + encoder_dim = self.cfg.mean_pool_gating_network_encoder_dim + elif getattr(cfg, "encoder_embed_dim", None): + # assume that encoder_embed_dim is the encoder's output dimension + encoder_dim = cfg.encoder_embed_dim + else: + raise ValueError( + "Must specify --mean-pool-gating-network-encoder-dim" + ) + + if self.cfg.mean_pool_gating_network_dropout > 0: + dropout = self.cfg.mean_pool_gating_network_dropout + elif getattr(cfg, "dropout", None): + dropout = cfg.dropout + else: + raise ValueError("Must specify task.mean_pool_gating_network_dropout") + + model.gating_network = MeanPoolGatingNetwork( + encoder_dim, + self.cfg.num_experts, + dropout, + ) + else: + raise ValueError( + "translation_moe task with learned prior requires the model to " + "have a gating network; try using --mean-pool-gating-network" + ) + return model + + def expert_index(self, i): + return i + self.tgt_dict.index("<expert_0>") + + def _get_loss(self, sample, model, criterion): + assert hasattr( + criterion, "compute_loss" + ), "translation_moe task requires the criterion to implement the compute_loss() method" + + k = self.cfg.num_experts + bsz = sample["target"].size(0) + + def get_lprob_y(encoder_out, prev_output_tokens_k): + net_output = model.decoder( + prev_output_tokens=prev_output_tokens_k, + encoder_out=encoder_out, + ) + loss, _ = criterion.compute_loss(model, net_output, sample, reduce=False) + loss = loss.view(bsz, -1) + return -loss.sum(dim=1, keepdim=True) # -> B x 1 + + def get_lprob_yz(winners=None): + encoder_out = model.encoder( + src_tokens=sample["net_input"]["src_tokens"], + src_lengths=sample["net_input"]["src_lengths"], + ) + + if winners is None: + lprob_y = [] + for i in range(k): + prev_output_tokens_k = sample["net_input"][ + "prev_output_tokens" + ].clone() + assert not prev_output_tokens_k.requires_grad + prev_output_tokens_k[:, 0] = self.expert_index(i) + lprob_y.append(get_lprob_y(encoder_out, prev_output_tokens_k)) + lprob_y = torch.cat(lprob_y, dim=1) # -> B x K + else: + prev_output_tokens_k = sample["net_input"]["prev_output_tokens"].clone() + prev_output_tokens_k[:, 0] = self.expert_index(winners) + lprob_y = get_lprob_y(encoder_out, prev_output_tokens_k) # -> B + + if self.uniform_prior: + lprob_yz = lprob_y + else: + lprob_z = model.gating_network(encoder_out) # B x K + if winners is not None: + lprob_z = lprob_z.gather(dim=1, index=winners.unsqueeze(-1)) + lprob_yz = lprob_y + lprob_z.type_as(lprob_y) # B x K + + return lprob_yz + + # compute responsibilities without dropout + with utils.model_eval(model): # disable dropout + with torch.no_grad(): # disable autograd + lprob_yz = get_lprob_yz() # B x K + prob_z_xy = torch.nn.functional.softmax(lprob_yz, dim=1) + assert not prob_z_xy.requires_grad + + # compute loss with dropout + if self.hard_selection: + winners = prob_z_xy.max(dim=1)[1] + loss = -get_lprob_yz(winners) + else: + lprob_yz = get_lprob_yz() # B x K + loss = -LogSumExpMoE.apply(lprob_yz, prob_z_xy, 1) + + loss = loss.sum() + sample_size = ( + sample["target"].size(0) if self.cfg.sentence_avg else sample["ntokens"] + ) + logging_output = { + "loss": utils.item(loss.data), + "ntokens": sample["ntokens"], + "nsentences": bsz, + "sample_size": sample_size, + "posterior": prob_z_xy.float().sum(dim=0).cpu(), + } + return loss, sample_size, logging_output + + def train_step( + self, sample, model, criterion, optimizer, update_num, ignore_grad=False + ): + model.train() + loss, sample_size, logging_output = self._get_loss(sample, model, criterion) + if ignore_grad: + loss *= 0 + optimizer.backward(loss) + return loss, sample_size, logging_output + + def valid_step(self, sample, model, criterion): + model.eval() + with torch.no_grad(): + loss, sample_size, logging_output = self._get_loss(sample, model, criterion) + return loss, sample_size, logging_output + + def inference_step( + self, + generator, + models, + sample, + prefix_tokens=None, + expert=None, + constraints=None, + ): + expert = expert or self.cfg.gen_expert + with torch.no_grad(): + return generator.generate( + models, + sample, + prefix_tokens=prefix_tokens, + constraints=constraints, + bos_token=self.expert_index(expert), + ) + + def reduce_metrics(self, logging_outputs, criterion): + super().reduce_metrics(logging_outputs, criterion) + metrics.log_scalar( + "posterior", + sum(log["posterior"] for log in logging_outputs if "posterior" in log), + ) diff --git a/examples/truncated_bptt/README.md b/examples/truncated_bptt/README.md new file mode 100644 index 0000000000000000000000000000000000000000..86518c9d5ef09fbd4fed1512a52e9431b74f08fa --- /dev/null +++ b/examples/truncated_bptt/README.md @@ -0,0 +1,70 @@ +# Truncated Backpropagation Through Time (BPTT) + +Truncated BPTT is a useful technique for training language models on very long +sequences. Typically a long sequences is split into chunks and a language model +is trained over the chunks sequentially. The LM may condition on previous +chunks, but gradients only flow through the current chunk. This technique was +the basis for the paper: [Transformer-XL: Attentive Language Models Beyond a +Fixed-Length Context](https://arxiv.org/abs/1901.02860), which achieved +state-of-the-art language modeling results at the time of publication. + +It is slightly tricky to implement Truncated BPTT efficiently in fairseq, since +we need to iterate over the data sequentially and disable any batch shuffling +logic. The code provided in this example illustrates how to implement Truncated +BPTT in fairseq by overriding ``FairseqTask::get_batch_iterator`` to iterate +over the data sequentially. Crucially, this example supports batching and +multi-GPU (data parallel) training. + +##### 0. Setup + +First, see the general [language modeling README](README.md) for instructions on +preprocessing the WikiText-103 data. + +##### 1. Train a Transformer-XL model on WikiText-103 + +We will train a 16-layer Transformer-XL model following the [hyperparameters +used in the original +paper](https://github.com/kimiyoung/transformer-xl/blob/master/pytorch/run_wt103_base.sh). + +The following command assumes 4 GPUs, so that the total batch size is 60 +sequences (15 x 4). Training should take ~24 hours on 4 V100 GPUs: +```bash +CUDA_VISIBLE_DEVICES=0,1,2,3 fairseq-train \ + --user-dir examples/truncated_bptt \ + data-bin/wikitext-103/ \ + --task truncated_bptt_lm --tokens-per-sample 150 \ + --batch-size 15 --max-update 200000 \ + --arch transformer_xl --n-layer 16 --d-model 410 --n-head 10 \ + --d-head 41 --d-inner 2100 --dropout 0.1 --dropatt 0.0 --mem-len 150 \ + --optimizer adam --clip-norm 0.25 \ + --lr-scheduler cosine --warmup-updates 0 --min-lr 0.0 --lr 0.00025 \ + --log-format json --log-interval 25 \ + --fp16 +``` + +If training on a single GPU, set `--update-freq=4` to accumulate 4x gradients +and simulate training on 4 GPUs. + +##### 2. Evaluate + +```bash +fairseq-eval-lm data-bin/wikitext-103/ \ + --path checkpoints/checkpoint_best.pt \ + --user-dir examples/truncated_bptt/ \ + --task truncated_bptt_lm \ + --batch-size 1 --required-batch-size-multiple 1 \ + --model-overrides '{"mem_len":640,"clamp_len":400,"same_length":True}' \ + --tokens-per-sample 64 +# ... | INFO | fairseq_cli.eval_lm | num. model params: 151123537 +# ... | INFO | fairseq_cli.eval_lm | Evaluated 245569 tokens in 83.1s (2956.82 tokens/s) +# ... | INFO | fairseq_cli.eval_lm | Loss (base 2): 4.5668, Perplexity: 23.70 +# Compare to 24.0 test perplexity from the paper +``` + +*Note:* During training the model saw 150 tokens of context +(``--tokens-per-sample=150``) and 150 extra memory tokens (``--mem-len=150``). +During evaluation we measure perplexity on sequences of 64 tokens +(``--tokens-per-sample=64``) and increase the memory length +(``--model-overrides='{"mem_len":640}'``). These settings match the evaluation +settings from [the original +paper](https://github.com/kimiyoung/transformer-xl/blob/master/pytorch/run_wt103_base.sh). diff --git a/examples/truncated_bptt/__init__.py b/examples/truncated_bptt/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..eee484d427a68828462469d133144a8d7c052c40 --- /dev/null +++ b/examples/truncated_bptt/__init__.py @@ -0,0 +1,6 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from . import transformer_xl_model, truncated_bptt_lm_task # noqa diff --git a/examples/truncated_bptt/transformer_xl_model.py b/examples/truncated_bptt/transformer_xl_model.py new file mode 100644 index 0000000000000000000000000000000000000000..a6c8b25a07276c2ee30c0aa5f0e4b0a2837ed5ca --- /dev/null +++ b/examples/truncated_bptt/transformer_xl_model.py @@ -0,0 +1,155 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +from dataclasses import dataclass, field +from typing import Dict, List, Optional + +import torch +from fairseq.dataclass import FairseqDataclass +from fairseq.models import ( + FairseqIncrementalDecoder, + FairseqLanguageModel, + register_model, +) +from fairseq.modules.checkpoint_activations import checkpoint_wrapper +from omegaconf import II + + +logger = logging.getLogger(__name__) + + +@dataclass +class TransformerXLConfig(FairseqDataclass): + # defaults come from the original Transformer-XL code + cutoffs: List[int] = field(default_factory=lambda: [20000, 40000, 200000]) + d_model: int = 500 + n_head: int = 10 + d_head: int = 50 + d_inner: int = 1000 + div_val: int = 1 + n_layer: int = 12 + mem_len: int = 0 + clamp_len: int = -1 + same_length: bool = False + dropout: float = 0.0 + dropatt: float = 0.0 + checkpoint_activations: bool = False + offload_activations: bool = False + max_target_positions: int = II("task.max_target_positions") + + +@register_model("transformer_xl", dataclass=TransformerXLConfig) +class TransformerXLLanguageModel(FairseqLanguageModel): + @classmethod + def build_model(cls, cfg: TransformerXLConfig, task): + return cls(TransformerXLDecoder(cfg, task)) + + +class TransformerXLDecoder(FairseqIncrementalDecoder): + def __init__(self, cfg, task): + try: + from transformers.models.transfo_xl import ( + TransfoXLConfig, + TransfoXLLMHeadModel, + ) + except ImportError: + from transformers.configuration_transfo_xl import TransfoXLConfig + from transformers.modeling_transfo_xl import TransfoXLLMHeadModel + + super().__init__(task.target_dictionary) + self.cfg = cfg + + # remove any cutoffs larger than the vocab size + cutoffs = [ + cutoff for cutoff in cfg.cutoffs if cutoff < len(task.target_dictionary) + ] + + config = TransfoXLConfig( + vocab_size=len(task.target_dictionary), + cutoffs=cutoffs, + d_model=cfg.d_model, + d_embed=cfg.d_model, + n_head=cfg.n_head, + d_head=cfg.d_head, + d_inner=cfg.d_inner, + div_val=cfg.div_val, + n_layer=cfg.n_layer, + mem_len=cfg.mem_len, + clamp_len=cfg.clamp_len, + same_length=cfg.same_length, + dropout=cfg.dropout, + dropatt=cfg.dropatt, + ) + logger.info(config) + self.model = TransfoXLLMHeadModel(config) + + # Workaround a bug in huggingface's ``ProjectedAdaptiveLogSoftmax`` + # which adds ``None`` values to an ``nn.ParameterList``, which is not + # supported in PyTorch. Instead we can replace this with an + # ``nn.ModuleList``, which does support ``None`` values. + try: + if all(p is None for p in self.model.crit.out_projs._parameters.values()): + self.model.crit.out_projs = torch.nn.ModuleList( + [None] * len(self.model.crit.out_projs._parameters) + ) + except Exception: + pass + + if cfg.checkpoint_activations or cfg.offload_activations: + for i in range(len(self.model.transformer.layers)): + self.model.transformer.layers[i] = checkpoint_wrapper( + self.model.transformer.layers[i], + offload_to_cpu=cfg.offload_activations, + ) + # TODO: may save mem to wrap(layer.pos_ff.CoreNet[3]) + + self._mems = None + + def forward( + self, + src_tokens, + src_lengths=None, # unused + incremental_state: Optional[Dict[str, List[torch.Tensor]]] = None, + encoder_out=None, + ): + if incremental_state is not None: # used during inference + mems = self.get_incremental_state(incremental_state, "mems") + src_tokens = src_tokens[:, -1:] # only keep the most recent token + else: + mems = self._mems + + output = self.model( + input_ids=src_tokens, + mems=mems, + return_dict=False, + ) + + if len(output) >= 2: + if incremental_state is not None: + self.set_incremental_state(incremental_state, "mems", output[1]) + else: + self._mems = output[1] + + return (output[0],) + + def max_positions(self): + return self.cfg.max_target_positions + + def reorder_incremental_state( + self, + incremental_state: Dict[str, Dict[str, Optional[torch.Tensor]]], + new_order: torch.Tensor, + ): + """Reorder incremental state. + + This will be called when the order of the input has changed from the + previous time step. A typical use case is beam search, where the input + order changes between time steps based on the selection of beams. + """ + mems = self.get_incremental_state(incremental_state, "mems") + if mems is not None: + new_mems = [mems_i.index_select(1, new_order) for mems_i in mems] + self.set_incremental_state(incremental_state, "mems", new_mems) diff --git a/examples/truncated_bptt/truncated_bptt_lm_task.py b/examples/truncated_bptt/truncated_bptt_lm_task.py new file mode 100644 index 0000000000000000000000000000000000000000..02be0e7fb4213b98798c85b79e9046e9990b97fc --- /dev/null +++ b/examples/truncated_bptt/truncated_bptt_lm_task.py @@ -0,0 +1,281 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import os +from dataclasses import dataclass, field +from typing import List, Optional, Tuple + +import torch +from fairseq import utils +from fairseq.data import ( + Dictionary, + TokenBlockDataset, + data_utils, + iterators, +) +from fairseq.dataclass import FairseqDataclass +from fairseq.distributed import utils as dist_utils +from fairseq.tasks import FairseqTask, register_task +from omegaconf import II + + +logger = logging.getLogger(__name__) + + +@dataclass +class TruncatedBPTTLMConfig(FairseqDataclass): + data: str = field(default="???", metadata={"help": "path to data directory"}) + tokens_per_sample: int = field( + default=1024, + metadata={"help": "max number of tokens per sequence"}, + ) + batch_size: int = II("dataset.batch_size") + # Some models use *max_target_positions* to know how many positional + # embeddings to learn. We use II(...) to make it default to + # *tokens_per_sample*, but in principle there could be more positional + # embeddings than tokens in a single batch. This may also be irrelevant for + # custom model implementations. + max_target_positions: int = II("task.tokens_per_sample") + # these will be populated automatically if not provided + data_parallel_rank: Optional[int] = None + data_parallel_size: Optional[int] = None + + +@register_task("truncated_bptt_lm", dataclass=TruncatedBPTTLMConfig) +class TruncatedBPTTLMTask(FairseqTask): + def __init__(self, cfg: TruncatedBPTTLMConfig): + super().__init__(cfg) + + if cfg.data_parallel_rank is None or cfg.data_parallel_size is None: + if torch.distributed.is_initialized(): + cfg.data_parallel_rank = dist_utils.get_data_parallel_rank() + cfg.data_parallel_size = dist_utils.get_data_parallel_world_size() + else: + cfg.data_parallel_rank = 0 + cfg.data_parallel_size = 1 + + # load the dictionary + paths = utils.split_paths(cfg.data) + assert len(paths) > 0 + self.dictionary = Dictionary.load(os.path.join(paths[0], "dict.txt")) + logger.info("dictionary: {} types".format(len(self.dictionary))) + + def load_dataset(self, split, epoch=1, combine=False, **kwargs): + """Load a given dataset split (e.g., train, valid, test)""" + + # support sharded datasets + paths = utils.split_paths(self.cfg.data) + assert len(paths) > 0 + data_path = paths[(epoch - 1) % len(paths)] + split_path = os.path.join(data_path, split) + + # each element of *data* will be a tensorized line from the original + # text dataset, similar to ``open(split_path).readlines()`` + data = data_utils.load_indexed_dataset( + split_path, self.dictionary, combine=combine + ) + if data is None: + raise FileNotFoundError( + "Dataset not found: {} ({})".format(split, split_path) + ) + + # this is similar to ``data.view(-1).split(tokens_per_sample)`` + data = TokenBlockDataset( + data, + data.sizes, + block_size=self.cfg.tokens_per_sample, + pad=None, # unused + eos=None, # unused + break_mode="none", + ) + + self.datasets[split] = TruncatedBPTTDataset( + data=data, + bsz_per_shard=self.cfg.batch_size, + shard_id=self.cfg.data_parallel_rank, + num_shards=self.cfg.data_parallel_size, + ) + + def dataset(self, split): + return self.datasets[split] + + def get_batch_iterator( + self, dataset, num_workers=0, epoch=1, data_buffer_size=0, **kwargs + ): + return iterators.EpochBatchIterator( + dataset=dataset, + collate_fn=self._collate_fn, + num_workers=num_workers, + epoch=epoch, + buffer_size=data_buffer_size, + # we don't use the batching functionality from EpochBatchIterator; + # instead every item in *dataset* is a whole batch + batch_sampler=[[i] for i in range(len(dataset))], + disable_shuffling=True, + ) + + def _collate_fn(self, items: List[List[torch.Tensor]]): + # we don't use fairseq's batching functionality, so we expect a single + # Tensor of type List[torch.Tensor] + assert len(items) == 1 + + # item will have shape B x T (the last batch may have length < T) + id, item = items[0] + item = data_utils.collate_tokens(item, pad_idx=self.source_dictionary.pad()) + B, T = item.size() + + # shift item one position over and append a padding token for the target + target = torch.nn.functional.pad( + item[:, 1:], (0, 1, 0, 0), value=self.target_dictionary.pad() + ) + + # fairseq expects batches to have the following structure + return { + "id": torch.tensor([id]*item.size(0)), + "net_input": { + "src_tokens": item, + }, + "target": target, + "nsentences": item.size(0), + "ntokens": item.numel(), + } + + def build_dataset_for_inference( + self, src_tokens: List[torch.Tensor], src_lengths: List[int], **kwargs + ) -> torch.utils.data.Dataset: + eos = self.source_dictionary.eos() + dataset = TokenBlockDataset( + src_tokens, + src_lengths, + block_size=None, # ignored for "eos" break mode + pad=self.source_dictionary.pad(), + eos=eos, + break_mode="eos", + ) + + class Dataset(torch.utils.data.Dataset): + def __getitem__(self, i): + item = dataset[i] + if item[-1] == eos: + # remove eos to support generating with a prefix + item = item[:-1] + return (i, [item]) + + def __len__(self): + return len(dataset) + + return Dataset() + + def inference_step( + self, generator, models, sample, prefix_tokens=None, constraints=None + ): + with torch.no_grad(): + if constraints is not None: + raise NotImplementedError + + # SequenceGenerator doesn't use *src_tokens* directly, we need to + # pass the *prefix_tokens* argument instead. + if prefix_tokens is None and sample["net_input"]["src_tokens"].nelement(): + prefix_tokens = sample["net_input"]["src_tokens"] + + # begin generation with the end-of-sentence token + bos_token = self.source_dictionary.eos() + + return generator.generate( + models, sample, prefix_tokens=prefix_tokens, bos_token=bos_token + ) + + def eval_lm_dataloader( + self, + dataset, + max_tokens: Optional[int] = 36000, + batch_size: Optional[int] = None, + max_positions: Optional[int] = None, + num_shards: int = 1, + shard_id: int = 0, + num_workers: int = 1, + data_buffer_size: int = 10, + context_window: int = 0, + ): + if context_window > 0: + raise NotImplementedError( + "Transformer-XL doesn't need --context-window, try " + "--model-overrides '{\"mem_len\":42}' instead " + ) + return self.get_batch_iterator( + dataset=dataset, + max_tokens=max_tokens, + max_sentences=batch_size, + max_positions=max_positions, + ignore_invalid_inputs=True, + num_shards=num_shards, + shard_id=shard_id, + num_workers=num_workers, + data_buffer_size=data_buffer_size, + ).next_epoch_itr(shuffle=False) + + @property + def source_dictionary(self): + return self.dictionary + + @property + def target_dictionary(self): + return self.dictionary + + +class TruncatedBPTTDataset(torch.utils.data.Dataset): + def __init__( + self, + data: List[torch.Tensor], # ordered list of items + bsz_per_shard, # number of items processed per GPUs per forward + shard_id, # current GPU ID + num_shards, # number of GPUs + ): + super().__init__() + self.data = data + + def batchify(data, bsz): + # Work out how cleanly we can divide the dataset into bsz parts. + nbatch = data.size(0) // bsz + # Trim off any extra elements that wouldn't cleanly fit (remainders). + data = data.narrow(0, 0, nbatch * bsz) + # Evenly divide the data across the bsz batches. + data = data.view(bsz, -1).contiguous() + return data + + # total number of sequences processed by all GPUs in each forward pass + global_batch_size = bsz_per_shard * num_shards + + """ + With a 16 item dataset, bsz_per_shard=2 and num_shards=3, + *indices* might look like: + + indices = [[0, 1], + [2, 3], + [4, 5], + [6, 7], + [8, 9], + [10, 11]] + + The size of the TruncatedBPTTDataset instance will be 2, + and shard 1 will see items: + + [(0, [data[4], data[6]]), + (1, [data[5], data[7]])] + """ + indices = batchify(torch.arange(len(data)), global_batch_size) + assert indices.size(0) == global_batch_size + + self.my_indices = indices[ + shard_id * bsz_per_shard : (shard_id + 1) * bsz_per_shard + ] + assert self.my_indices.size(0) == bsz_per_shard + + def __len__(self): + return self.my_indices.size(1) + + def __getitem__(self, i) -> Tuple[int, List[torch.Tensor]]: + return (i, [self.data[idx] for idx in self.my_indices[:, i]]) diff --git a/examples/unsupervised_quality_estimation/README.md b/examples/unsupervised_quality_estimation/README.md new file mode 100644 index 0000000000000000000000000000000000000000..e86a0d13b883af0c37fdc2c1fee9b0b9dff4d18c --- /dev/null +++ b/examples/unsupervised_quality_estimation/README.md @@ -0,0 +1,126 @@ +# Unsupervised Quality Estimation for Neural Machine Translation (Fomicheva et al., 2020) + +This page includes instructions for reproducing results from the paper [Unsupervised Quality Estimation for Neural +Machine Translation (Fomicheva et al., 2020)](https://arxiv.org/abs/2005.10608) + +## Requirements: + +* mosesdecoder: https://github.com/moses-smt/mosesdecoder +* subword-nmt: https://github.com/rsennrich/subword-nmt +* flores: https://github.com/facebookresearch/flores + +## Download Models and Test Data + +Download translation models and test data from [MLQE dataset repository](https://github.com/facebookresearch/mlqe). + +## Set up: + +Given a testset consisting of source sentences and reference translations: + +* `SRC_LANG`: source language +* `TGT_LANG`: target language +* `INPUT`: input prefix, such that the file `$INPUT.$SRC_LANG` contains source sentences and `$INPUT.$TGT_LANG` +contains the reference sentences +* `OUTPUT_DIR`: output path to store results +* `MOSES_DECODER`: path to mosesdecoder installation +* `BPE_ROOT`: path to subword-nmt installation +* `BPE`: path to BPE model +* `MODEL_DIR`: directory containing the NMT model `.pt` file as well as the source and target vocabularies. +* `TMP`: directory for intermediate temporary files +* `GPU`: if translating with GPU, id of the GPU to use for inference +* `DROPOUT_N`: number of stochastic forward passes + +`$DROPOUT_N` is set to 30 in the experiments reported in the paper. However, we observed that increasing it beyond 10 +does not bring substantial improvements. + +## Translate the data using standard decoding + +Preprocess the input data: +``` +for LANG in $SRC_LANG $TGT_LANG; do + perl $MOSES_DECODER/scripts/tokenizer/tokenizer.perl -threads 80 -a -l $LANG < $INPUT.$LANG > $TMP/preprocessed.tok.$LANG + python $BPE_ROOT/apply_bpe.py -c ${BPE} < $TMP/preprocessed.tok.$LANG > $TMP/preprocessed.tok.bpe.$LANG +done +``` + +Binarize the data for faster translation: + +``` +fairseq-preprocess --srcdict $MODEL_DIR/dict.$SRC_LANG.txt --tgtdict $MODEL_DIR/dict.$TGT_LANG.txt +--source-lang ${SRC_LANG} --target-lang ${TGT_LANG} --testpref $TMP/preprocessed.tok.bpe --destdir $TMP/bin --workers 4 +``` + +Translate + +``` +CUDA_VISIBLE_DEVICES=$GPU fairseq-generate $TMP/bin --path ${MODEL_DIR}/${SRC_LANG}-${TGT_LANG}.pt --beam 5 +--source-lang $SRC_LANG --target-lang $TGT_LANG --no-progress-bar --unkpen 5 > $TMP/fairseq.out +grep ^H $TMP/fairseq.out | cut -d- -f2- | sort -n | cut -f3- > $TMP/mt.out +``` + +Post-process + +``` +sed -r 's/(@@ )| (@@ ?$)//g' < $TMP/mt.out | perl $MOSES_DECODER/scripts/tokenizer/detokenizer.perl +-l $TGT_LANG > $OUTPUT_DIR/mt.out +``` + +## Produce uncertainty estimates + +### Scoring + +Make temporary files to store the translations repeated N times. + +``` +python ${SCRIPTS}/scripts/uncertainty/repeat_lines.py -i $TMP/preprocessed.tok.bpe.$SRC_LANG -n $DROPOUT_N +-o $TMP/repeated.$SRC_LANG +python ${SCRIPTS}/scripts/uncertainty/repeat_lines.py -i $TMP/mt.out -n $DROPOUT_N -o $TMP/repeated.$TGT_LANG + +fairseq-preprocess --srcdict ${MODEL_DIR}/dict.${SRC_LANG}.txt $TGT_DIC --source-lang ${SRC_LANG} +--target-lang ${TGT_LANG} --testpref ${TMP}/repeated --destdir ${TMP}/bin-repeated +``` + +Produce model scores for the generated translations using `--retain-dropout` option to apply dropout at inference time: + +``` +CUDA_VISIBLE_DEVICES=${GPU} fairseq-generate ${TMP}/bin-repeated --path ${MODEL_DIR}/${LP}.pt --beam 5 + --source-lang $SRC_LANG --target-lang $TGT_LANG --no-progress-bar --unkpen 5 --score-reference --retain-dropout + --retain-dropout-modules '["TransformerModel","TransformerEncoder","TransformerDecoder","TransformerEncoderLayer"]' + TransformerDecoderLayer --seed 46 > $TMP/dropout.scoring.out + +grep ^H $TMP/dropout.scoring.out | cut -d- -f2- | sort -n | cut -f2 > $TMP/dropout.scores + +``` + +Use `--retain-dropout-modules` to specify the modules. By default, dropout is applied in the same places +as for training. + +Compute the mean of the resulting output distribution: + +``` +python $SCRIPTS/scripts/uncertainty/aggregate_scores.py -i $TMP/dropout.scores -o $OUTPUT_DIR/dropout.scores.mean +-n $DROPOUT_N +``` + +### Generation + +Produce multiple translation hypotheses for the same source using `--retain-dropout` option: + +``` +CUDA_VISIBLE_DEVICES=${GPU} fairseq-generate ${TMP}/bin-repeated --path ${MODEL_DIR}/${LP}.pt + --beam 5 --source-lang $SRC_LANG --target-lang $TGT_LANG --no-progress-bar --retain-dropout + --unkpen 5 --retain-dropout-modules TransformerModel TransformerEncoder TransformerDecoder +TransformerEncoderLayer TransformerDecoderLayer --seed 46 > $TMP/dropout.generation.out + +grep ^H $TMP/dropout.generation.out | cut -d- -f2- | sort -n | cut -f3- > $TMP/dropout.hypotheses_ + +sed -r 's/(@@ )| (@@ ?$)//g' < $TMP/dropout.hypotheses_ | perl $MOSES_DECODER/scripts/tokenizer/detokenizer.perl +-l $TGT_LANG > $TMP/dropout.hypotheses +``` + +Compute similarity between multiple hypotheses corresponding to the same source sentence using Meteor +evaluation metric: +``` +python meteor.py -i $TMP/dropout.hypotheses -m <path_to_meteor_installation> -n $DROPOUT_N -o +$OUTPUT_DIR/dropout.gen.sim.meteor +``` diff --git a/examples/unsupervised_quality_estimation/aggregate_scores.py b/examples/unsupervised_quality_estimation/aggregate_scores.py new file mode 100644 index 0000000000000000000000000000000000000000..66d50d07ff2067b802b90a2aadd88df23153830a --- /dev/null +++ b/examples/unsupervised_quality_estimation/aggregate_scores.py @@ -0,0 +1,41 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import sys + +import numpy as np + + +aggregate_funcs = { + "std": np.std, + "var": np.var, + "median": np.median, + "mean": np.mean, + "min": np.min, + "max": np.max, +} + + +def main(): + parser = argparse.ArgumentParser() + parser.add_argument("-i", "--input_file", required=True, type=str) + parser.add_argument("-n", "--repeat_times", required=True, type=int) + parser.add_argument("-o", "--output_file", required=False) + parser.add_argument("-f", "--func", required=False, default="mean") + args = parser.parse_args() + + stream = open(args.output_file, "w") if args.output_file else sys.stdout + + segment_scores = [] + for line in open(args.input_file): + segment_scores.append(float(line.strip())) + if len(segment_scores) == args.repeat_times: + stream.write("{}\n".format(aggregate_funcs[args.func](segment_scores))) + segment_scores = [] + + +if __name__ == "__main__": + main() diff --git a/examples/unsupervised_quality_estimation/meteor.py b/examples/unsupervised_quality_estimation/meteor.py new file mode 100644 index 0000000000000000000000000000000000000000..2ee0448cf1f167f6f3ecee56ad807922cffb0956 --- /dev/null +++ b/examples/unsupervised_quality_estimation/meteor.py @@ -0,0 +1,109 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import math +import os +import subprocess +import sys +import tempfile +from collections import defaultdict +from itertools import combinations + + +def read_translations(path, n_repeats): + segment_counter = 0 + segment_translations = [] + translations = defaultdict(list) + for line in open(path): + segment_translations.append(" ".join(line.split())) + if len(segment_translations) == n_repeats: + translations[segment_counter] = segment_translations + segment_translations = [] + segment_counter += 1 + return translations + + +def generate_input(translations, n_repeats): + _, ref_path = tempfile.mkstemp() + _, mt_path = tempfile.mkstemp() + ref_fh = open(ref_path, "w") + mt_fh = open(mt_path, "w") + for segid in sorted(translations.keys()): + assert len(translations[segid]) == n_repeats + indexes = combinations(range(n_repeats), 2) + for idx1, idx2 in indexes: + mt_fh.write(translations[segid][idx1].strip() + "\n") + ref_fh.write(translations[segid][idx2].strip() + "\n") + sys.stderr.write("\nSaved translations to %s and %s" % (ref_path, mt_path)) + return ref_path, mt_path + + +def run_meteor(ref_path, mt_path, metric_path, lang="en"): + _, out_path = tempfile.mkstemp() + subprocess.call( + [ + "java", + "-Xmx2G", + "-jar", + metric_path, + mt_path, + ref_path, + "-p", + "0.5 0.2 0.6 0.75", # default parameters, only changed alpha to give equal weight to P and R + "-norm", + "-l", + lang, + ], + stdout=open(out_path, "w"), + ) + os.remove(ref_path) + os.remove(mt_path) + sys.stderr.write("\nSaved Meteor output to %s" % out_path) + return out_path + + +def read_output(meteor_output_path, n_repeats): + n_combinations = math.factorial(n_repeats) / ( + math.factorial(2) * math.factorial(n_repeats - 2) + ) + raw_scores = [] + average_scores = [] + for line in open(meteor_output_path): + if not line.startswith("Segment "): + continue + score = float(line.strip().split("\t")[1]) + raw_scores.append(score) + if len(raw_scores) == n_combinations: + average_scores.append(sum(raw_scores) / n_combinations) + raw_scores = [] + os.remove(meteor_output_path) + return average_scores + + +def main(): + parser = argparse.ArgumentParser() + parser.add_argument("-i", "--infile") + parser.add_argument("-n", "--repeat_times", type=int) + parser.add_argument("-m", "--meteor") + parser.add_argument("-o", "--output") + args = parser.parse_args() + + translations = read_translations(args.infile, args.repeat_times) + sys.stderr.write("\nGenerating input for Meteor...") + ref_path, mt_path = generate_input(translations, args.repeat_times) + sys.stderr.write("\nRunning Meteor...") + out_path = run_meteor(ref_path, mt_path, args.meteor) + sys.stderr.write("\nReading output...") + scores = read_output(out_path, args.repeat_times) + sys.stderr.write("\nWriting results...") + with open(args.output, "w") as o: + for scr in scores: + o.write("{}\n".format(scr)) + o.close() + + +if __name__ == "__main__": + main() diff --git a/examples/unsupervised_quality_estimation/repeat_lines.py b/examples/unsupervised_quality_estimation/repeat_lines.py new file mode 100644 index 0000000000000000000000000000000000000000..5a04851a74624e9c8ebc259805b7aed6c638b0de --- /dev/null +++ b/examples/unsupervised_quality_estimation/repeat_lines.py @@ -0,0 +1,28 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import sys + + +def _normalize_spaces(line): + return " ".join(line.split()) + + +def main(): + parser = argparse.ArgumentParser() + parser.add_argument("-i", "--input_file", required=True, type=str) + parser.add_argument("-n", "--repeat_times", required=True, type=int) + parser.add_argument("-o", "--output_file", required=False, type=str) + args = parser.parse_args() + stream = open(args.output_file, "w") if args.output_file else sys.stdout + + for line in open(args.input_file): + for _ in range(args.repeat_times): + stream.write(_normalize_spaces(line) + "\n") + + +if __name__ == "__main__": + main() diff --git a/examples/wav2vec/README.md b/examples/wav2vec/README.md new file mode 100644 index 0000000000000000000000000000000000000000..238639a9ba2474481cfb93bb94d42ac62613897d --- /dev/null +++ b/examples/wav2vec/README.md @@ -0,0 +1,369 @@ +# wav2vec 2.0 + +wav2vec 2.0 learns speech representations on unlabeled data as described in [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations (Baevski et al., 2020)](https://arxiv.org/abs/2006.11477). + +We learned speech representations in multiple languages as well in [Unsupervised Cross-lingual Representation Learning for Speech Recognition (Conneau et al., 2020)](https://arxiv.org/abs/2006.13979). + +We also combined wav2vec 2.0 with self-training in [Self-training and Pre-training are Complementary for Speech Recognition (Xu et al., 2020)](https://arxiv.org/abs/2010.11430). + +## Pre-trained models + +Model | Finetuning split | Dataset | Model +|---|---|---|--- +Wav2Vec 2.0 Base | No finetuning | [Librispeech](http://www.openslr.org/12) | [download](https://dl.fbaipublicfiles.com/fairseq/wav2vec/wav2vec_small.pt) +Wav2Vec 2.0 Base | 10 minutes | [Librispeech](http://www.openslr.org/12) | [download](https://dl.fbaipublicfiles.com/fairseq/wav2vec/wav2vec_small_10m.pt) +Wav2Vec 2.0 Base | 100 hours | [Librispeech](http://www.openslr.org/12) | [download](https://dl.fbaipublicfiles.com/fairseq/wav2vec/wav2vec_small_100h.pt) +Wav2Vec 2.0 Base | 960 hours | [Librispeech](http://www.openslr.org/12) | [download](https://dl.fbaipublicfiles.com/fairseq/wav2vec/wav2vec_small_960h.pt) +Wav2Vec 2.0 Large | No finetuning | [Librispeech](http://www.openslr.org/12) | [download](https://dl.fbaipublicfiles.com/fairseq/wav2vec/libri960_big.pt) +Wav2Vec 2.0 Large | 10 minutes | [Librispeech](http://www.openslr.org/12) | [download](https://dl.fbaipublicfiles.com/fairseq/wav2vec/wav2vec_big_10m.pt) +Wav2Vec 2.0 Large | 100 hours | [Librispeech](http://www.openslr.org/12) | [download](https://dl.fbaipublicfiles.com/fairseq/wav2vec/wav2vec_big_100h.pt) +Wav2Vec 2.0 Large | 960 hours | [Librispeech](http://www.openslr.org/12) | [download](https://dl.fbaipublicfiles.com/fairseq/wav2vec/wav2vec_big_960h.pt) +Wav2Vec 2.0 Large (LV-60)* | No finetuning | [Libri-Light](https://github.com/facebookresearch/libri-light) | [download](https://dl.fbaipublicfiles.com/fairseq/wav2vec/wav2vec_vox_new.pt) +Wav2Vec 2.0 Large (LV-60)* | 10 minutes | [Libri-Light](https://github.com/facebookresearch/libri-light) + [Librispeech](http://www.openslr.org/12) | [download](https://dl.fbaipublicfiles.com/fairseq/wav2vec/wav2vec_vox_10m_new.pt) +Wav2Vec 2.0 Large (LV-60)* | 100 hours | [Libri-Light](https://github.com/facebookresearch/libri-light) + [Librispeech](http://www.openslr.org/12) | [download](https://dl.fbaipublicfiles.com/fairseq/wav2vec/wav2vec_vox_100h_new.pt) +Wav2Vec 2.0 Large (LV-60)* | 960 hours | [Libri-Light](https://github.com/facebookresearch/libri-light) + [Librispeech](http://www.openslr.org/12) | [download](https://dl.fbaipublicfiles.com/fairseq/wav2vec/wav2vec2_vox_960h_new.pt) +Wav2Vec 2.0 Large (LV-60) + Self Training * | 10 minutes | [Libri-Light](https://github.com/facebookresearch/libri-light) + [Librispeech](http://www.openslr.org/12) | [download](https://dl.fbaipublicfiles.com/fairseq/wav2vec/wav2vec_vox_10m_pl.pt) +Wav2Vec 2.0 Large (LV-60) + Self Training * | 100 hours | [Libri-Light](https://github.com/facebookresearch/libri-light) + [Librispeech](http://www.openslr.org/12) | [download](https://dl.fbaipublicfiles.com/fairseq/wav2vec/wav2vec_vox_100h_pl.pt) +Wav2Vec 2.0 Large (LV-60) + Self Training * | 960 hours | [Libri-Light](https://github.com/facebookresearch/libri-light) + [Librispeech](http://www.openslr.org/12) | [download](https://dl.fbaipublicfiles.com/fairseq/wav2vec/wav2vec_vox_960h_pl.pt) + +\* updated (Oct. 24, 2020) + +We also release multilingual pre-trained wav2vec 2.0 (XLSR) models: + +Model | Architecture | Hours | Languages | Datasets | Model +|---|---|---|---|---|--- +XLSR-53 | Large | 56k | 53 | MLS, CommonVoice, BABEL | [download](https://dl.fbaipublicfiles.com/fairseq/wav2vec/xlsr_53_56k.pt) + +The XLSR model uses the following datasets for multilingual pretraining: + +* **[MLS: Multilingual LibriSpeech](https://indico2.conference4me.psnc.pl/event/35/contributions/3585/attachments/1060/1101/Wed-2-6-10.pdf)** (8 languages, 50.7k hours): *Dutch, English, French, German, Italian, Polish, Portuguese, Spanish* + +* **[CommonVoice](https://commonvoice.mozilla.org/en/languages)** (36 languages, 3.6k hours): *Arabic, Basque, Breton, Chinese (CN), Chinese (HK), Chinese (TW), Chuvash, Dhivehi, Dutch, English, Esperanto, Estonian, French, German, Hakh-Chin, Indonesian, Interlingua, Irish, Italian, Japanese, Kabyle, Kinyarwanda, Kyrgyz, Latvian, Mongolian, Persian, Portuguese, Russian, Sakha, Slovenian, Spanish, Swedish, Tamil, Tatar, Turkish, Welsh* (see also [finetuning splits]([https://dl.fbaipublicfiles.com/cpc_audio/common_voices_splits.tar.gz]) from [this paper](https://arxiv.org/abs/2002.02848)). + +* **[Babel](https://catalog.ldc.upenn.edu/byyear)** (17 languages, 1.7k hours): *Assamese, Bengali, Cantonese, Cebuano, Georgian, Haitian, Kazakh, Kurmanji, Lao, Pashto, Swahili, Tagalog, Tamil, Tok, Turkish, Vietnamese, Zulu* + + +## Training a new model with the CLI tools + +Given a directory containing wav files to be used for pretraining (we recommend splitting each file into separate file 10 to 30 seconds in length) + +### Prepare training data manifest: + +First, install the `soundfile` library: +```shell script +pip install soundfile +``` + +Next, run: + +```shell script +$ python examples/wav2vec/wav2vec_manifest.py /path/to/waves --dest /manifest/path --ext $ext --valid-percent $valid +``` + +$ext should be set to flac, wav, or whatever format your dataset happens to use that soundfile can read. + +$valid should be set to some reasonable percentage (like 0.01) of training data to use for validation. +To use a pre-defined validation set (like dev-other from librispeech), set to it 0 and then overwrite valid.tsv with a +separately pre-processed manifest file. + +### Train a wav2vec 2.0 base model: + +This configuration was used for the base model trained on the Librispeech dataset in the wav2vec 2.0 paper + +Note that the input is expected to be single channel, sampled at 16 kHz + +```shell script +$ fairseq-hydra-train \ + task.data=/path/to/data \ + --config-dir /path/to/fairseq-py/examples/wav2vec/config/pretraining \ + --config-name wav2vec2_base_librispeech +``` + +Note: you can simulate 64 GPUs by using k GPUs and adding command line parameters (before `--config-dir`) +`distributed_training.distributed_world_size=k` `+optimization.update_freq='[x]'` where x = 64/k + +### Train a wav2vec 2.0 large model: + +This configuration was used for the large model trained on the Libri-light dataset in the wav2vec 2.0 paper + +```shell script +$ fairseq-hydra-train \ + task.data=/path/to/data \ + --config-dir /path/to/fairseq-py/examples/wav2vec/config/pretraining \ + --config-name wav2vec2_large_librivox +``` + +Note: you can simulate 128 GPUs by using k GPUs and adding command line parameters (before `--config-dir`) +`distributed_training.distributed_world_size=k` `+optimization.update_freq='[x]'` where x = 128/k + +### Fine-tune a pre-trained model with CTC: + +Fine-tuning a model requires parallel audio and labels file, as well as a vocabulary file in fairseq format. +A letter vocabulary can be downloaded [here](https://dl.fbaipublicfiles.com/fairseq/wav2vec/dict.ltr.txt). +An example [script](libri_labels.py) that generates labels for the Librispeech dataset from the tsv file produced by wav2vec_manifest.py can be used as follows: + +```shell script +split=train +$ python libri_labels.py /path/to/tsv --output-dir /output/dir --output-name $split +``` + +Fine-tuning on 100h of Librispeech with letter targets: +```shell script +$ fairseq-hydra-train \ + distributed_training.distributed_port=$PORT \ + task.data=/path/to/data \ + model.w2v_path=/path/to/model.pt \ + --config-dir /path/to/fairseq-py/examples/wav2vec/config/finetuning \ + --config-name base_100h +``` + +There are other config files in the config/finetuning directory that can be used to fine-tune on other splits. +You can specify the right config via the `--config-name` parameter. + +Note: you can simulate 24 GPUs by using k GPUs and adding command line parameters (before `--config-dir`) +`distributed_training.distributed_world_size=k` `+optimization.update_freq='[x]'` where x = 24/k + +Decoding with a language model during training requires flashlight [python bindings](https://github.com/facebookresearch/flashlight/tree/master/bindings/python) (previously called [wav2letter](https://github.com/facebookresearch/wav2letter). +If you want to use a language model, add `+criterion.wer_args='[/path/to/kenlm, /path/to/lexicon, 2, -1]'` to the command line. + +### Evaluating a CTC model: + +Evaluating a CTC model with a language model requires [flashlight python bindings](https://github.com/facebookresearch/flashlight/tree/master/bindings/python) (previously called [wav2letter](https://github.com/facebookresearch/wav2letter) to be installed. + +Fairseq transformer language model used in the wav2vec 2.0 paper can be obtained from the [wav2letter model repository](https://github.com/facebookresearch/wav2letter/tree/master/recipes/sota/2019). +Be sure to upper-case the language model vocab after downloading it. + +Letter dictionary for pre-trained models can be found [here](https://dl.fbaipublicfiles.com/fairseq/wav2vec/dict.ltr.txt). + +Next, run the evaluation command: + +```shell script +$subset=dev_other +python examples/speech_recognition/infer.py /checkpoint/abaevski/data/speech/libri/10h/wav2vec/raw --task audio_pretraining \ +--nbest 1 --path /path/to/model --gen-subset $subset --results-path /path/to/save/results/for/sclite --w2l-decoder kenlm \ +--lm-model /path/to/kenlm.bin --lm-weight 2 --word-score -1 --sil-weight 0 --criterion ctc --labels ltr --max-tokens 4000000 \ +--post-process letter +``` + +To get raw numbers, use --w2l-decoder viterbi and omit the lexicon. To use the transformer language model, use --w2l-decoder fairseqlm. + +## Use wav2vec 2.0 with 🤗Transformers: + +Wav2Vec2 is also available in the [🤗Transformers library](https://github.com/huggingface/transformers) since version 4.4. + +Pretrained Models can be found on the [hub](https://huggingface.co/models?filter=wav2vec2) +and documentation can be found [here](https://huggingface.co/transformers/master/model_doc/wav2vec2.html). + +Usage example: + +```python +# !pip install transformers +# !pip install datasets +import soundfile as sf +import torch +from datasets import load_dataset +from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor + +# load pretrained model +processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-960h") +model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-960h") + + +librispeech_samples_ds = load_dataset("patrickvonplaten/librispeech_asr_dummy", "clean", split="validation") + +# load audio +audio_input, sample_rate = sf.read(librispeech_samples_ds[0]["file"]) + +# pad input values and return pt tensor +input_values = processor(audio_input, sampling_rate=sample_rate, return_tensors="pt").input_values + +# INFERENCE + +# retrieve logits & take argmax +logits = model(input_values).logits +predicted_ids = torch.argmax(logits, dim=-1) + +# transcribe +transcription = processor.decode(predicted_ids[0]) + +# FINE-TUNE + +target_transcription = "A MAN SAID TO THE UNIVERSE I EXIST" + +# encode labels +with processor.as_target_processor(): + labels = processor(target_transcription, return_tensors="pt").input_ids + +# compute loss by passing labels +loss = model(input_values, labels=labels).loss +loss.backward() +``` + +# wav2vec + +Example to train a wav2vec model as described in [wav2vec: Unsupervised Pre-training for Speech Recognition (Schneider et al., 2019)](https://arxiv.org/abs/1904.05862). + +## Pre-trained models + +Description | Dataset | Model +---|---|--- +Wav2Vec large | [Librispeech](http://www.openslr.org/12) | [download](https://dl.fbaipublicfiles.com/fairseq/wav2vec/wav2vec_large.pt) + +#### Example usage: +```python +import torch +import fairseq + +cp_path = '/path/to/wav2vec.pt' +model, cfg, task = fairseq.checkpoint_utils.load_model_ensemble_and_task([cp_path]) +model = model[0] +model.eval() + +wav_input_16khz = torch.randn(1,10000) +z = model.feature_extractor(wav_input_16khz) +c = model.feature_aggregator(z) +``` + +## Training a new model with the CLI tools + +Given a directory containing wav files to be used for pretraining (we recommend splitting each file into separate files 10 to 30 seconds in length) + +### Prepare training data manifest: + +``` +$ python examples/wav2vec/wav2vec_manifest.py /path/to/waves --dest /manifest/path --ext wav +``` + +### Train a wav2vec model: + +``` +$ python train.py /manifest/path --save-dir /model/path --num-workers 6 --fp16 --max-update 400000 --save-interval 1 --no-epoch-checkpoints \ +--arch wav2vec --task audio_pretraining --min-lr 1e-06 --stop-min-lr 1e-09 --optimizer adam --lr 0.005 --lr-scheduler cosine \ +--conv-feature-layers [(512, 10, 5), (512, 8, 4), (512, 4, 2), (512, 4, 2), (512, 4, 2), (512, 1, 1), (512, 1, 1)] \ +--conv-aggregator-layers [(512, 2, 1), (512, 3, 1), (512, 4, 1), (512, 5, 1), (512, 6, 1), (512, 7, 1), (512, 8, 1), (512, 9, 1), (512, 10, 1), (512, 11, 1), (512, 12, 1), (512, 13, 1)] \ +--skip-connections-agg --residual-scale 0.5 --log-compression --warmup-updates 500 --warmup-init-lr 1e-07 --criterion wav2vec --num-negatives 10 \ +--max-sample-size 150000 --max-tokens 1500000 --skip-invalid-size-inputs-valid-test +``` + +### Run wav2vec2 pre-training on Google Cloud TPUs: + +Wav2Vec2 is now supported on TPUs! It's currently pre-training only. + +#### Using hydra on a v3-8: + +``` +$ OMP_NUM_THREADS=1 fairseq-hydra-train \ + task.data=/manifest/path \ + --config-dir /PATH/TO/FAIRSEQ/examples/wav2vec/config/pretraining \ + --config-name wav2vec2_large_librivox_tpu.yaml +``` + +#### Using command line arguments on a v3-8: + +``` +$ OMP_NUM_THREADS=1 python train.py /manifest/path --save-dir /model/path --num-workers 6 --fp16 --max-update 400000 --save-interval 1 --no-epoch-checkpoints \ +--arch wav2vec2 --task audio_pretraining --min-lr 1e-06 --stop-min-lr 1e-09 --optimizer adam --lr 0.005 --lr-scheduler cosine \ +--conv-feature-layers [(512, 10, 5), (512, 8, 4), (512, 4, 2), (512, 4, 2), (512, 4, 2), (512, 1, 1), (512, 1, 1)] \ +--conv-aggregator-layers [(512, 2, 1), (512, 3, 1), (512, 4, 1), (512, 5, 1), (512, 6, 1), (512, 7, 1), (512, 8, 1), (512, 9, 1), (512, 10, 1), (512, 11, 1), (512, 12, 1), (512, 13, 1)] \ +--skip-connections-agg --residual-scale 0.5 --log-compression --warmup-updates 500 --warmup-init-lr 1e-07 --criterion wav2vec --num-negatives 10 \ +--max-sample-size 150000 --max-tokens 1500000 --skip-invalid-size-inputs-valid-test \ +--tpu --distributed-world-size 8 --num-batch-buckets 3 --enable-padding \ +--encoder-layerdrop 0 --mask-channel-prob 0.1 +``` + +#### Using hydra on a pod slice (v3-N with N > 8): + +``` +$ OMP_NUM_THREADS=1 fairseq-hydra-train \ + task.data=/manifest/path \ + --config-dir /PATH/TO/FAIRSEQ/examples/wav2vec/config/pretraining \ + --config-name wav2vec2_large_librivox_tpu-pod.yaml # edit distributed-world-size accordingly +``` + +#### Using command line arguments on a pod slice (v3-N with N > 8): + + +``` +$ python -m torch_xla.distributed.xla_dist \ + --tpu ${TPUNAME} --conda-env=torch-xla-${TORCH_XLA_VERSION} --env OMP_NUM_THREADS=1 \ + -- \ +python train.py /manifest/path --save-dir /model/path --num-workers 6 --fp16 --max-update 400000 --save-interval 1 --no-epoch-checkpoints \ +--arch wav2vec2 --task audio_pretraining --min-lr 1e-06 --stop-min-lr 1e-09 --optimizer adam --lr 0.005 --lr-scheduler cosine \ +--conv-feature-layers [(512, 10, 5), (512, 8, 4), (512, 4, 2), (512, 4, 2), (512, 4, 2), (512, 1, 1), (512, 1, 1)] \ +--conv-aggregator-layers [(512, 2, 1), (512, 3, 1), (512, 4, 1), (512, 5, 1), (512, 6, 1), (512, 7, 1), (512, 8, 1), (512, 9, 1), (512, 10, 1), (512, 11, 1), (512, 12, 1), (512, 13, 1)] \ +--skip-connections-agg --residual-scale 0.5 --log-compression --warmup-updates 500 --warmup-init-lr 1e-07 --criterion wav2vec --num-negatives 10 \ +--max-sample-size 150000 --max-tokens 1500000 --skip-invalid-size-inputs-valid-test \ +--tpu --distributed-world-size ${WORLD_SIZE} --num-batch-buckets 3 --enable-padding \ +--encoder-layerdrop 0 --mask-channel-prob 0.1 +``` + +### Extract embeddings from the downstream task data: + +``` +$ PYTHONPATH=/path/to/fairseq python examples/wav2vec/wav2vec_featurize.py --input /path/to/task/waves --output /path/to/output \ +--model /model/path/checkpoint_best.pt --split train valid test +``` + +# vq-wav2vec + +Example to train a vq-wav2vec model as described in [vq-wav2vec: Self-Supervised Learning of Discrete Speech Representations (Baevski et al., 2019)](https://arxiv.org/abs/1910.05453). + +These models are also used in [Effectiveness of self-supervised pre-training for speech recognition (Baevski et al., 2019)](https://arxiv.org/abs/1911.03912). + +## Pre-trained models + +Description | Dataset | Model +---|---|--- +vq-wav2vec Gumbel | [Librispeech](http://www.openslr.org/12) | [download](https://dl.fbaipublicfiles.com/fairseq/wav2vec/vq-wav2vec.pt) +vq-wav2vec K-means | [Librispeech](http://www.openslr.org/12) | [download](https://dl.fbaipublicfiles.com/fairseq/wav2vec/vq-wav2vec_kmeans.pt) +Roberta on K-means codes | [Librispeech](http://www.openslr.org/12) | [download](https://dl.fbaipublicfiles.com/fairseq/wav2vec/bert_kmeans.tar) + +#### Example usage: +```python +import torch +import fairseq + +cp = torch.load('/path/to/vq-wav2vec.pt') +model, cfg, task = fairseq.checkpoint_utils.load_model_ensemble_and_task([cp]) +model = model[0] +model.eval() + +wav_input_16khz = torch.randn(1,10000) +z = model.feature_extractor(wav_input_16khz) +_, idxs = model.vector_quantizer.forward_idx(z) +print(idxs.shape) # output: torch.Size([1, 60, 2]), 60 timesteps with 2 indexes corresponding to 2 groups in the model +``` + +## Training a new model with the CLI tools + +Given a directory containing wav files to be used for pretraining (we recommend splitting each file into separate file 10 to 30 seconds in length) + +### Prepare training data manifest: + +``` +$ python examples/wav2vec/wav2vec_manifest.py /path/to/waves --dest /manifest/path --ext wav +``` + +### Train a gumbel vq-wav2vec model: + +``` +$ python train.py /manifest/path --save-dir /model/path --num-workers 6 --fp16 --max-update 400000 \ +--save-interval 1 --no-epoch-checkpoints --arch wav2vec --task audio_pretraining --min-lr 1e-06 --stop-min-lr 1e-09 \ +--optimizer adam --lr 1e-05 --lr-scheduler cosine \ +--conv-feature-layers [(512, 10, 5), (512, 8, 4), (512, 4, 2), (512, 4, 2), (512, 4, 2), (512, 1, 1), (512, 1, 1), (512, 1, 1)] \ +--conv-aggregator-layers [(512, 2, 1), (512, 3, 1), (512, 4, 1), (512, 5, 1), (512, 6, 1), (512, 7, 1), (512, 8, 1), (512, 9, 1), (512, 10, 1), (512, 11, 1), (512, 12, 1), (512, 13, 1)] \ +--activation gelu --offset auto --skip-connections-agg --residual-scale 0.5 \ +--log-keys ["prob_perplexity","code_perplexity","temp"] --vq-type gumbel --vq-groups 2 --vq-depth 2 \ +--combine-groups --vq-vars 320 --vq-temp (2,0.5,0.999995) --prediction-steps 12 --warmup-updates 1000 \ +--warmup-init-lr 1e-07 --criterion wav2vec --num-negatives 10 --max-sample-size 150000 \ +--max-tokens 300000 --cross-sample-negatives 0 --update-freq 1 --seed 2 --skip-invalid-size-inputs-valid-test +``` + +for k-means training, set vq-type with "kmeans" and add --loss-weights [1] argument. Pre-trained models were trained on 16 GPUs. + +### Tokenize audio data (e.g. for BERT training): + +``` +$ PYTHONPATH=/path/to/fairseq python examples/wav2vec/vq-wav2vec_featurize.py --data-dir /manifest/path --output-dir /path/to/output \ +--checkpoint /model/path/checkpoint_best.pt --split train valid test --extension tsv +``` diff --git a/examples/wav2vec/__init__.py b/examples/wav2vec/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/examples/wav2vec/config/finetuning/base_100h.yaml b/examples/wav2vec/config/finetuning/base_100h.yaml new file mode 100644 index 0000000000000000000000000000000000000000..539dabb047d02089c3e633c01960dba787134e53 --- /dev/null +++ b/examples/wav2vec/config/finetuning/base_100h.yaml @@ -0,0 +1,59 @@ +# @package _group_ + +common: + fp16: true + log_format: json + log_interval: 200 + +checkpoint: + no_epoch_checkpoints: true + best_checkpoint_metric: wer + +task: + _name: audio_pretraining + data: ??? + normalize: false + labels: ltr + +dataset: + num_workers: 6 + max_tokens: 3200000 + skip_invalid_size_inputs_valid_test: true + valid_subset: dev_other + +distributed_training: + ddp_backend: legacy_ddp + distributed_world_size: 2 + +criterion: + _name: ctc + zero_infinity: true + +optimization: + max_update: 80000 + lr: [0.00003] + sentence_avg: true + update_freq: [4] + +optimizer: + _name: adam + adam_betas: (0.9,0.98) + adam_eps: 1e-08 + +lr_scheduler: + _name: tri_stage + phase_ratio: [0.1, 0.4, 0.5] + final_lr_scale: 0.05 + +model: + _name: wav2vec_ctc + w2v_path: ??? + apply_mask: true + mask_prob: 0.65 + mask_channel_prob: 0.5 + mask_channel_length: 64 + layerdrop: 0.1 + activation_dropout: 0.1 + feature_grad_mult: 0.0 + freeze_finetune_updates: 0 + diff --git a/examples/wav2vec/config/finetuning/base_10h.yaml b/examples/wav2vec/config/finetuning/base_10h.yaml new file mode 100644 index 0000000000000000000000000000000000000000..16a3c4d96cf7f676b4314b3cd4632cec7ec2cebf --- /dev/null +++ b/examples/wav2vec/config/finetuning/base_10h.yaml @@ -0,0 +1,64 @@ +# @package _group_ + +common: + fp16: true + log_format: json + log_interval: 200 + +checkpoint: + save_interval: 50 + save_interval_updates: 10000 + keep_interval_updates: 1 + no_epoch_checkpoints: true + best_checkpoint_metric: wer + +task: + _name: audio_pretraining + data: ??? + normalize: false + labels: ltr + +dataset: + num_workers: 6 + max_tokens: 3200000 + skip_invalid_size_inputs_valid_test: true + validate_after_updates: 10000 + validate_interval: 50 + valid_subset: dev_other + +distributed_training: + ddp_backend: legacy_ddp + distributed_world_size: 2 + +criterion: + _name: ctc + zero_infinity: true + +optimization: + max_update: 20000 + lr: [0.00005] + sentence_avg: true + update_freq: [4] + +optimizer: + _name: adam + adam_betas: (0.9,0.98) + adam_eps: 1e-08 + +lr_scheduler: + _name: tri_stage + phase_ratio: [0.1, 0.4, 0.5] + final_lr_scale: 0.05 + +model: + _name: wav2vec_ctc + w2v_path: ??? + apply_mask: true + mask_prob: 0.65 + mask_channel_prob: 0.5 + mask_channel_length: 64 + layerdrop: 0.05 + activation_dropout: 0.1 + feature_grad_mult: 0.0 + freeze_finetune_updates: 10000 + diff --git a/examples/wav2vec/config/finetuning/base_10m.yaml b/examples/wav2vec/config/finetuning/base_10m.yaml new file mode 100644 index 0000000000000000000000000000000000000000..3ceb77a252de06e51e960bbda5952e6db3ea13e2 --- /dev/null +++ b/examples/wav2vec/config/finetuning/base_10m.yaml @@ -0,0 +1,64 @@ +# @package _group_ + +common: + fp16: true + log_format: json + log_interval: 200 + +checkpoint: + save_interval: 1000 + save_interval_updates: 50 + keep_interval_updates: 1 + no_epoch_checkpoints: true + best_checkpoint_metric: wer + +task: + _name: audio_pretraining + data: ??? + normalize: false + labels: ltr + +dataset: + num_workers: 6 + max_tokens: 3200000 + skip_invalid_size_inputs_valid_test: true + validate_after_updates: 10000 + validate_interval: 1000 + valid_subset: dev_other + +distributed_training: + ddp_backend: legacy_ddp + distributed_world_size: 2 + +criterion: + _name: ctc + zero_infinity: true + +optimization: + max_update: 13000 + lr: [0.00005] + sentence_avg: true + update_freq: [4] + +optimizer: + _name: adam + adam_betas: (0.9,0.98) + adam_eps: 1e-08 + +lr_scheduler: + _name: tri_stage + phase_ratio: [0.1, 0.4, 0.5] + final_lr_scale: 0.05 + +model: + _name: wav2vec_ctc + w2v_path: ??? + apply_mask: true + mask_prob: 0.65 + mask_channel_prob: 0.25 + mask_channel_length: 64 + layerdrop: 0.1 + activation_dropout: 0.1 + feature_grad_mult: 0.0 + freeze_finetune_updates: 10000 + diff --git a/examples/wav2vec/config/finetuning/base_1h.yaml b/examples/wav2vec/config/finetuning/base_1h.yaml new file mode 100644 index 0000000000000000000000000000000000000000..3ceb77a252de06e51e960bbda5952e6db3ea13e2 --- /dev/null +++ b/examples/wav2vec/config/finetuning/base_1h.yaml @@ -0,0 +1,64 @@ +# @package _group_ + +common: + fp16: true + log_format: json + log_interval: 200 + +checkpoint: + save_interval: 1000 + save_interval_updates: 50 + keep_interval_updates: 1 + no_epoch_checkpoints: true + best_checkpoint_metric: wer + +task: + _name: audio_pretraining + data: ??? + normalize: false + labels: ltr + +dataset: + num_workers: 6 + max_tokens: 3200000 + skip_invalid_size_inputs_valid_test: true + validate_after_updates: 10000 + validate_interval: 1000 + valid_subset: dev_other + +distributed_training: + ddp_backend: legacy_ddp + distributed_world_size: 2 + +criterion: + _name: ctc + zero_infinity: true + +optimization: + max_update: 13000 + lr: [0.00005] + sentence_avg: true + update_freq: [4] + +optimizer: + _name: adam + adam_betas: (0.9,0.98) + adam_eps: 1e-08 + +lr_scheduler: + _name: tri_stage + phase_ratio: [0.1, 0.4, 0.5] + final_lr_scale: 0.05 + +model: + _name: wav2vec_ctc + w2v_path: ??? + apply_mask: true + mask_prob: 0.65 + mask_channel_prob: 0.25 + mask_channel_length: 64 + layerdrop: 0.1 + activation_dropout: 0.1 + feature_grad_mult: 0.0 + freeze_finetune_updates: 10000 + diff --git a/examples/wav2vec/config/finetuning/base_960h.yaml b/examples/wav2vec/config/finetuning/base_960h.yaml new file mode 100644 index 0000000000000000000000000000000000000000..2d38211e919ddcec7cc9a24557fc11dc0f3f99cf --- /dev/null +++ b/examples/wav2vec/config/finetuning/base_960h.yaml @@ -0,0 +1,58 @@ +# @package _group_ + +common: + fp16: true + log_format: json + log_interval: 200 + +checkpoint: + no_epoch_checkpoints: true + best_checkpoint_metric: wer + +task: + _name: audio_pretraining + data: ??? + normalize: false + labels: ltr + +dataset: + num_workers: 6 + max_tokens: 3200000 + skip_invalid_size_inputs_valid_test: true + valid_subset: dev_other + +distributed_training: + ddp_backend: legacy_ddp + distributed_world_size: 8 + +criterion: + _name: ctc + zero_infinity: true + +optimization: + max_update: 320000 + lr: [0.0001] + sentence_avg: true + +optimizer: + _name: adam + adam_betas: (0.9,0.98) + adam_eps: 1e-08 + +lr_scheduler: + _name: tri_stage + phase_ratio: [0.1, 0.4, 0.5] + final_lr_scale: 0.05 + +model: + _name: wav2vec_ctc + w2v_path: ??? + apply_mask: true + mask_prob: 0.5 + mask_channel_prob: 0.1 + mask_channel_length: 64 + layerdrop: 0.1 + activation_dropout: 0.1 + feature_grad_mult: 0.0 + freeze_finetune_updates: 0 + diff --git a/examples/wav2vec/config/finetuning/vox_100h.yaml b/examples/wav2vec/config/finetuning/vox_100h.yaml new file mode 100644 index 0000000000000000000000000000000000000000..2fdb0c568c197186fc370cfc5b95c04d0f49b453 --- /dev/null +++ b/examples/wav2vec/config/finetuning/vox_100h.yaml @@ -0,0 +1,59 @@ +# @package _group_ + +common: + fp16: true + log_format: json + log_interval: 200 + +checkpoint: + no_epoch_checkpoints: true + best_checkpoint_metric: wer + +task: + _name: audio_pretraining + data: ??? + normalize: true + labels: ltr + +dataset: + num_workers: 6 + max_tokens: 1280000 + skip_invalid_size_inputs_valid_test: true + valid_subset: dev_other + +distributed_training: + ddp_backend: legacy_ddp + distributed_world_size: 4 + +criterion: + _name: ctc + zero_infinity: true + +optimization: + max_update: 80000 + lr: [0.00003] + sentence_avg: true + update_freq: [5] + +optimizer: + _name: adam + adam_betas: (0.9,0.98) + adam_eps: 1e-08 + +lr_scheduler: + _name: tri_stage + phase_ratio: [0.1, 0.4, 0.5] + final_lr_scale: 0.05 + +model: + _name: wav2vec_ctc + w2v_path: ??? + apply_mask: true + mask_prob: 0.5 + mask_channel_prob: 0.5 + mask_channel_length: 64 + layerdrop: 0.1 + activation_dropout: 0.1 + feature_grad_mult: 0.0 + freeze_finetune_updates: 10000 + diff --git a/examples/wav2vec/config/finetuning/vox_10h.yaml b/examples/wav2vec/config/finetuning/vox_10h.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f1a979e05dad279f77463b7cb3cb62a7d0178d5c --- /dev/null +++ b/examples/wav2vec/config/finetuning/vox_10h.yaml @@ -0,0 +1,64 @@ +# @package _group_ + +common: + fp16: true + log_format: json + log_interval: 200 + +checkpoint: + save_interval: 50 + save_interval_updates: 10000 + keep_interval_updates: 1 + no_epoch_checkpoints: true + best_checkpoint_metric: wer + +task: + _name: audio_pretraining + data: ??? + normalize: true + labels: ltr + +dataset: + num_workers: 6 + max_tokens: 1280000 + skip_invalid_size_inputs_valid_test: true + validate_after_updates: 10000 + validate_interval: 50 + valid_subset: dev_other + +distributed_training: + ddp_backend: legacy_ddp + distributed_world_size: 4 + +criterion: + _name: ctc + zero_infinity: true + +optimization: + max_update: 20000 + lr: [0.0001] + sentence_avg: true + update_freq: [5] + +optimizer: + _name: adam + adam_betas: (0.9,0.98) + adam_eps: 1e-08 + +lr_scheduler: + _name: tri_stage + phase_ratio: [0.1, 0.4, 0.5] + final_lr_scale: 0.05 + +model: + _name: wav2vec_ctc + w2v_path: ??? + apply_mask: true + mask_prob: 0.75 + mask_channel_prob: 0.25 + mask_channel_length: 64 + layerdrop: 0.1 + activation_dropout: 0.1 + feature_grad_mult: 0.0 + freeze_finetune_updates: 10000 + diff --git a/examples/wav2vec/config/finetuning/vox_10m.yaml b/examples/wav2vec/config/finetuning/vox_10m.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d12439bb28cd4a5f0ecc255b4c21a77c64ae8b38 --- /dev/null +++ b/examples/wav2vec/config/finetuning/vox_10m.yaml @@ -0,0 +1,64 @@ +# @package _group_ + +common: + fp16: true + log_format: json + log_interval: 200 + +checkpoint: + save_interval: 1000 + save_interval_updates: 50 + keep_interval_updates: 1 + no_epoch_checkpoints: true + best_checkpoint_metric: wer + +task: + _name: audio_pretraining + data: ??? + normalize: true + labels: ltr + +dataset: + num_workers: 6 + max_tokens: 1280000 + skip_invalid_size_inputs_valid_test: true + validate_after_updates: 10000 + validate_interval: 1000 + valid_subset: dev_other + +distributed_training: + ddp_backend: legacy_ddp + distributed_world_size: 4 + +criterion: + _name: ctc + zero_infinity: true + +optimization: + max_update: 13000 + lr: [0.0001] + sentence_avg: true + update_freq: [5] + +optimizer: + _name: adam + adam_betas: (0.9,0.98) + adam_eps: 1e-08 + +lr_scheduler: + _name: tri_stage + phase_ratio: [0.1, 0.4, 0.5] + final_lr_scale: 0.05 + +model: + _name: wav2vec_ctc + w2v_path: ??? + apply_mask: true + mask_prob: 0.65 + mask_channel_prob: 0.25 + mask_channel_length: 64 + layerdrop: 0.1 + activation_dropout: 0.1 + feature_grad_mult: 0.0 + freeze_finetune_updates: 10000 + diff --git a/examples/wav2vec/config/finetuning/vox_1h.yaml b/examples/wav2vec/config/finetuning/vox_1h.yaml new file mode 100644 index 0000000000000000000000000000000000000000..7f3b04c0349b8900d00992da4a6bce8acac22449 --- /dev/null +++ b/examples/wav2vec/config/finetuning/vox_1h.yaml @@ -0,0 +1,64 @@ +# @package _group_ + +common: + fp16: true + log_format: json + log_interval: 200 + +checkpoint: + save_interval: 1000 + save_interval_updates: 50 + keep_interval_updates: 1 + no_epoch_checkpoints: true + best_checkpoint_metric: wer + +task: + _name: audio_pretraining + data: ??? + normalize: true + labels: ltr + +dataset: + num_workers: 6 + max_tokens: 1280000 + skip_invalid_size_inputs_valid_test: true + validate_after_updates: 10000 + validate_interval: 1000 + valid_subset: dev_other + +distributed_training: + ddp_backend: legacy_ddp + distributed_world_size: 4 + +criterion: + _name: ctc + zero_infinity: true + +optimization: + max_update: 13000 + lr: [0.0003] + sentence_avg: true + update_freq: [5] + +optimizer: + _name: adam + adam_betas: (0.9,0.98) + adam_eps: 1e-08 + +lr_scheduler: + _name: tri_stage + phase_ratio: [0.1, 0.4, 0.5] + final_lr_scale: 0.05 + +model: + _name: wav2vec_ctc + w2v_path: ??? + apply_mask: true + mask_prob: 0.75 + mask_channel_prob: 0.25 + mask_channel_length: 64 + layerdrop: 0.1 + activation_dropout: 0.1 + feature_grad_mult: 0.0 + freeze_finetune_updates: 10000 + diff --git a/examples/wav2vec/config/finetuning/vox_960h.yaml b/examples/wav2vec/config/finetuning/vox_960h.yaml new file mode 100644 index 0000000000000000000000000000000000000000..0633915bb29a102a2275cbc49e45ac6e11bd5ad2 --- /dev/null +++ b/examples/wav2vec/config/finetuning/vox_960h.yaml @@ -0,0 +1,58 @@ +# @package _group_ + +common: + fp16: true + log_format: json + log_interval: 200 + +checkpoint: + no_epoch_checkpoints: true + best_checkpoint_metric: wer + +task: + _name: audio_pretraining + data: ??? + normalize: true + labels: ltr + +dataset: + num_workers: 6 + max_tokens: 1280000 + skip_invalid_size_inputs_valid_test: true + valid_subset: dev_other + +distributed_training: + ddp_backend: legacy_ddp + distributed_world_size: 24 + +criterion: + _name: ctc + zero_infinity: true + +optimization: + max_update: 320000 + lr: [0.00003] + sentence_avg: true + +optimizer: + _name: adam + adam_betas: (0.9,0.98) + adam_eps: 1e-08 + +lr_scheduler: + _name: tri_stage + phase_ratio: [0.1, 0.4, 0.5] + final_lr_scale: 0.05 + +model: + _name: wav2vec_ctc + w2v_path: ??? + apply_mask: true + mask_prob: 0.5 + mask_channel_prob: 0.25 + mask_channel_length: 64 + layerdrop: 0.1 + activation_dropout: 0.1 + feature_grad_mult: 0.0 + freeze_finetune_updates: 10000 + diff --git a/examples/wav2vec/config/pretraining/wav2vec2_base_librispeech.yaml b/examples/wav2vec/config/pretraining/wav2vec2_base_librispeech.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b686e21ab1d367158fe7afa4197303a4ee74df66 --- /dev/null +++ b/examples/wav2vec/config/pretraining/wav2vec2_base_librispeech.yaml @@ -0,0 +1,57 @@ +# @package _group_ + +common: + fp16: true + log_format: json + log_interval: 200 + +checkpoint: + save_interval_updates: 25000 + keep_interval_updates: 1 + no_epoch_checkpoints: true + +task: + _name: audio_pretraining + data: ??? + max_sample_size: 250000 + min_sample_size: 32000 + normalize: false + +dataset: + num_workers: 6 + max_tokens: 1400000 + skip_invalid_size_inputs_valid_test: true + +distributed_training: + distributed_world_size: 64 + ddp_backend: legacy_ddp + +criterion: + _name: wav2vec + infonce: true + log_keys: ["prob_perplexity","code_perplexity","temp"] + loss_weights: [0.1, 10] + +optimization: + max_update: 400000 + lr: [0.0005] + +optimizer: + _name: adam + adam_betas: (0.9,0.98) + adam_eps: 1e-06 + weight_decay: 0.01 + +lr_scheduler: + _name: polynomial_decay + warmup_updates: 32000 + +model: + _name: wav2vec2 + quantize_targets: true + final_dim: 256 + encoder_layerdrop: 0.05 + dropout_input: 0.1 + dropout_features: 0.1 + feature_grad_mult: 0.1 + encoder_embed_dim: 768 diff --git a/examples/wav2vec/config/pretraining/wav2vec2_large_librivox.yaml b/examples/wav2vec/config/pretraining/wav2vec2_large_librivox.yaml new file mode 100644 index 0000000000000000000000000000000000000000..bee41157a9984ea89f46dc89e6986ba6c73c3037 --- /dev/null +++ b/examples/wav2vec/config/pretraining/wav2vec2_large_librivox.yaml @@ -0,0 +1,69 @@ +# @package _group_ + +common: + fp16: true + log_format: json + log_interval: 200 + +checkpoint: + save_interval_updates: 25000 + keep_interval_updates: 1 + no_epoch_checkpoints: true + +task: + _name: audio_pretraining + data: ??? + max_sample_size: 320000 + min_sample_size: 32000 + normalize: true + +dataset: + num_workers: 6 + max_tokens: 1200000 + skip_invalid_size_inputs_valid_test: true + +distributed_training: + distributed_world_size: 128 + ddp_backend: legacy_ddp + +criterion: + _name: wav2vec + infonce: true + log_keys: ["prob_perplexity","code_perplexity","temp"] + loss_weights: [0.1, 0] + +optimization: + max_update: 1000000 + lr: [0.005] + +optimizer: + _name: adam + adam_betas: (0.9,0.98) + adam_eps: 1e-06 + weight_decay: 0.01 + +lr_scheduler: + _name: polynomial_decay + warmup_updates: 32000 + +model: + _name: wav2vec2 + quantize_targets: true + extractor_mode: layer_norm + layer_norm_first: true + final_dim: 768 + latent_temp: [2.0,0.1,0.999995] + encoder_layerdrop: 0.00 + dropout_input: 0.0 + dropout_features: 0.0 + dropout: 0.0 + attention_dropout: 0.0 + conv_bias: true + + encoder_layers: 24 + encoder_embed_dim: 1024 + encoder_ffn_embed_dim: 4096 + encoder_attention_heads: 16 + + feature_grad_mult: 1.0 + diff --git a/examples/wav2vec/config/pretraining/wav2vec2_large_librivox_tpu-pod.yaml b/examples/wav2vec/config/pretraining/wav2vec2_large_librivox_tpu-pod.yaml new file mode 100644 index 0000000000000000000000000000000000000000..ff35a95b6596b74215ef1bbdd2ec8d462d1d8542 --- /dev/null +++ b/examples/wav2vec/config/pretraining/wav2vec2_large_librivox_tpu-pod.yaml @@ -0,0 +1,72 @@ +# @package _group_ + +common: + tpu: true + fp16: false + log_format: json + log_interval: 10 + +checkpoint: + save_interval_updates: 25000 + keep_interval_updates: 1 + no_epoch_checkpoints: true + +task: + _name: audio_pretraining + data: ??? + max_sample_size: 250000 + min_sample_size: 32000 + normalize: true + num_batch_buckets: 3 + precompute_mask_indices: true + enable_padding: true + +dataset: + num_workers: 6 + max_tokens: 1200000 + skip_invalid_size_inputs_valid_test: true + +distributed_training: + distributed_world_size: 128 + ddp_backend: legacy_ddp + +criterion: + _name: wav2vec + infonce: true + log_keys: ["prob_perplexity","code_perplexity","temp"] + loss_weights: [0.1, 0] + +optimization: + max_update: 1000000 + lr: [0.005] + +optimizer: + _name: adam + adam_betas: (0.9,0.98) + adam_eps: 1e-06 + weight_decay: 0.01 + +lr_scheduler: + _name: polynomial_decay + warmup_updates: 32000 + +model: + _name: wav2vec2 + quantize_targets: true + extractor_mode: layer_norm + layer_norm_first: true + final_dim: 768 + latent_temp: [2.0,0.1,0.999995] + encoder_layerdrop: 0.00 + dropout_input: 0.0 + dropout_features: 0.0 + dropout: 0.0 + attention_dropout: 0.0 + conv_bias: true + + encoder_layers: 24 + encoder_embed_dim: 1024 + encoder_ffn_embed_dim: 4096 + encoder_attention_heads: 16 + + feature_grad_mult: 1.0 diff --git a/examples/wav2vec/config/pretraining/wav2vec2_large_librivox_tpu.yaml b/examples/wav2vec/config/pretraining/wav2vec2_large_librivox_tpu.yaml new file mode 100644 index 0000000000000000000000000000000000000000..2036e23c6bd6ba896cdd2b055915c8f66944b3e4 --- /dev/null +++ b/examples/wav2vec/config/pretraining/wav2vec2_large_librivox_tpu.yaml @@ -0,0 +1,72 @@ +# @package _group_ + +common: + tpu: true + fp16: false + log_format: json + log_interval: 10 + +checkpoint: + save_interval_updates: 25000 + keep_interval_updates: 1 + no_epoch_checkpoints: true + +task: + _name: audio_pretraining + data: ??? + max_sample_size: 250000 + min_sample_size: 32000 + normalize: true + num_batch_buckets: 3 + precompute_mask_indices: true + enable_padding: true + +dataset: + num_workers: 6 + max_tokens: 1200000 + skip_invalid_size_inputs_valid_test: true + +distributed_training: + distributed_world_size: 8 + ddp_backend: legacy_ddp + +criterion: + _name: wav2vec + infonce: true + log_keys: ["prob_perplexity","code_perplexity","temp"] + loss_weights: [0.1, 0] + +optimization: + max_update: 1000000 + lr: [0.005] + +optimizer: + _name: adam + adam_betas: (0.9,0.98) + adam_eps: 1e-06 + weight_decay: 0.01 + +lr_scheduler: + _name: polynomial_decay + warmup_updates: 32000 + +model: + _name: wav2vec2 + quantize_targets: true + extractor_mode: layer_norm + layer_norm_first: true + final_dim: 768 + latent_temp: [2.0,0.1,0.999995] + encoder_layerdrop: 0.00 + dropout_input: 0.0 + dropout_features: 0.0 + dropout: 0.0 + attention_dropout: 0.0 + conv_bias: true + + encoder_layers: 24 + encoder_embed_dim: 1024 + encoder_ffn_embed_dim: 4096 + encoder_attention_heads: 16 + + feature_grad_mult: 1.0 diff --git a/examples/wav2vec/libri_labels.py b/examples/wav2vec/libri_labels.py new file mode 100644 index 0000000000000000000000000000000000000000..694a202604c7a4a480550550679ce6c16bd10e42 --- /dev/null +++ b/examples/wav2vec/libri_labels.py @@ -0,0 +1,56 @@ +#!/usr/bin/env python3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +""" +Helper script to pre-compute embeddings for a flashlight (previously called wav2letter++) dataset +""" + +import argparse +import os + + +def main(): + parser = argparse.ArgumentParser() + parser.add_argument("tsv") + parser.add_argument("--output-dir", required=True) + parser.add_argument("--output-name", required=True) + args = parser.parse_args() + + os.makedirs(args.output_dir, exist_ok=True) + + transcriptions = {} + + with open(args.tsv, "r") as tsv, open( + os.path.join(args.output_dir, args.output_name + ".ltr"), "w" + ) as ltr_out, open( + os.path.join(args.output_dir, args.output_name + ".wrd"), "w" + ) as wrd_out: + root = next(tsv).strip() + for line in tsv: + line = line.strip() + dir = os.path.dirname(line) + if dir not in transcriptions: + parts = dir.split(os.path.sep) + trans_path = f"{parts[-2]}-{parts[-1]}.trans.txt" + path = os.path.join(root, dir, trans_path) + assert os.path.exists(path) + texts = {} + with open(path, "r") as trans_f: + for tline in trans_f: + items = tline.strip().split() + texts[items[0]] = " ".join(items[1:]) + transcriptions[dir] = texts + part = os.path.basename(line).split(".")[0] + assert part in transcriptions[dir] + print(transcriptions[dir][part], file=wrd_out) + print( + " ".join(list(transcriptions[dir][part].replace(" ", "|"))) + " |", + file=ltr_out, + ) + + +if __name__ == "__main__": + main() diff --git a/examples/wav2vec/scripts/binarize_manifest.sh b/examples/wav2vec/scripts/binarize_manifest.sh new file mode 100644 index 0000000000000000000000000000000000000000..6f201bdb524fad51a69d8c45889eaa1578efc62d --- /dev/null +++ b/examples/wav2vec/scripts/binarize_manifest.sh @@ -0,0 +1,33 @@ +#!/usr/bin/env bash + +# usage: bash binarize_manifest <dest_dir> <train_split> <valid_split> + +DEST_DIR=$1 +TRAIN_SPLIT=$2 +VALID_SPLIT=$3 +FAIRSEQ_ROOT=$4 + +mkdir -p $DEST_DIR + +# split file path and lengths into separate files +cut -f1 $TRAIN_SPLIT.tsv > $DEST_DIR/train_fnames.txt +cut -f1 $VALID_SPLIT.tsv > $DEST_DIR/valid_fnames.txt +cut -f2 $TRAIN_SPLIT.tsv > $DEST_DIR/train.lengths +cut -f2 $VALID_SPLIT.tsv > $DEST_DIR/valid.lengths + +# copy root directory +head -1 $TRAIN_SPLIT.tsv > $DEST_DIR/train.root +head -1 $VALID_SPLIT.tsv > $DEST_DIR/valid.root + +# remove root directory +sed -i '1d' $DEST_DIR/train_fnames.txt +sed -i '1d' $DEST_DIR/valid_fnames.txt +sed -i '1d' $DEST_DIR/train.lengths +sed -i '1d' $DEST_DIR/valid.lengths + +# insert spaces between characters +sed -i -e 's/\(.\)/\1 /g' $DEST_DIR/train_fnames.txt +sed -i -e 's/\(.\)/\1 /g' $DEST_DIR/valid_fnames.txt + +# run preprocessor +PYTHONPATH=$FAIRSEQ_ROOT python $FAIRSEQ_ROOT/fairseq_cli/preprocess.py --dataset-impl mmap --trainpref $DEST_DIR/train_fnames.txt --validpref $DEST_DIR/valid_fnames.txt --workers 60 --only-source --destdir $DEST_DIR diff --git a/examples/wav2vec/unsupervised/README.md b/examples/wav2vec/unsupervised/README.md new file mode 100644 index 0000000000000000000000000000000000000000..c2a935d414c5e4fe9d72c1e6e45c1059c32c8586 --- /dev/null +++ b/examples/wav2vec/unsupervised/README.md @@ -0,0 +1,93 @@ +# wav2vec Unsupervised (wav2vec-U) + +Wav2vec Unsupervised (wav2vec-U) is a framework for building speech recognition systems without any labeled training data as described in [Unsupervised Speech Recognition (Baevski et al., 2021)](https://ai.facebook.com/research/publications/unsupervised-speech-recognition). The model takes as input wav2vec 2.0 or XLSR representations (see [pretrained models](https://github.com/pytorch/fairseq/blob/master/examples/wav2vec)) as well as unlabeled speech and text data. + + The wav2vec-U training procedure consists of three consecutive main steps: +* Preparation of speech representations and text data +* Generative adversarial training (GAN) +* Iterative self-training + Kaldi LM-decoding + +## Preparation of speech and text data +Similar to [wav2vec 2.0](https://github.com/pytorch/fairseq/blob/master/examples/wav2vec/README.md), data folders contain {train,valid,test}.{tsv,wrd,phn} files, where audio paths are stored in tsv files, and word, letter or phoneme transcriptions are stored in .{wrd,ltr,phn}. + +In **/path/to/data/with_silence** you need a *train.tsv* file as well as (optionally) *{valid,test}.{tsv,wrd,phn}*. It is nice to have *10h.{tsv,phn}* files there too for reproducing the ablation study on layer selection. In **/path/to/data/without_silence** you have the same files, except *.tsv* files contain audios with silences removed using rVAD. + +Pre-requisites: +* set FAIRSEQ_ROOT environmental variable to your fairseq installation +* set RVAD_ROOT environmental variable to a checkout of [rVADfast](https://github.com/zhenghuatan/rVADfast) +* set KENLM_ROOT environmental variable to the location of [KenLM](https://github.com/kpu/kenlm) binaries +* install [PyKaldi](https://github.com/pykaldi/pykaldi) and set KALDI_ROOT environmental variable to the location of your kaldi installation. To use the version bundled with PyKaldi, you can use /path/to/pykaldi/tools/kaldi + +Create new audio files without silences: +```shell +# create a manifest file for the set original of audio files +python $FAIRSEQ_ROOT/examples/wav2vec/wav2vec_manifest.py /dir/to/save/audio/files --ext wav --dest /path/to/new/train.tsv --valid-percent 0 + +python scripts/vads.py -r $RVAD_ROOT < /path/to/train.tsv > train.vads + +python scripts/remove_silence.py --tsv /path/to/train.tsv --vads train.vads --out /dir/to/save/audio/files + +python $FAIRSEQ_ROOT/examples/wav2vec/wav2vec_manifest.py /dir/to/save/audio/files --ext wav --dest /path/to/new/train.tsv --valid-percent 0.01 +``` + +Next, we need to preprocess the audio data to better match phonemized text data: + +```shell +zsh scripts/prepare_audio.sh /dir/with/{train,test,valid}.tsv /output/dir /path/to/wav2vec2/model.pt 512 14 +``` +Note that if you have splits different than train/valid/test, you will need to modify this script. The last two arguments are the PCA dimensionality and the 0-based index of the layer from which to extract representations. + +Now we need to prepare text data: +```shell +zsh scripts/prepare_text.sh language /path/to/text/file /output/dir 1000 espeak /path/to/fasttext/lid/model +``` + +The fourth argument is minimum number observations of phones to keep. If your text corpus is small, you might want to reduce this number. + +The fifth argument is which phonemizer to use. Supported values are [espeak](http://espeak.sourceforge.net/), [espeak-ng](https://github.com/espeak-ng/espeak-ng), and [G2P](https://github.com/Kyubyong/g2p) (english only). + +Pre-trained fasttext LID models can be downloaded [here](https://fasttext.cc/docs/en/language-identification.html). + +## Generative adversarial training (GAN) + +We then use a GAN model to build a first unsupervised ASR model. The data preparation above of both speech features and text data is a necessary procedure that enables the generator to match speech to text in an unsupervised way. + +Launching GAN training on top of preprocessed features, with default hyperparameters can be done with: + +``` +PREFIX=w2v_unsup_gan_xp +TASK_DATA=/path/to/features/precompute_unfiltered_pca512_cls128_mean_pooled +TEXT_DATA=/path/to/data/phones # path to fairseq-preprocessed GAN data (phones dir) +KENLM_PATH=/path/to/data/phones/kenlm.phn.o4.bin # KenLM 4-gram phoneme language model (LM data = GAN data here) + +PYTHONPATH=$FAIRSEQ_ROOT PREFIX=$PREFIX fairseq-hydra-train \ + -m --config-dir config/gan \ + --config-name w2vu \ + task.data=${TASK_DATA} \ + task.text_data=${TEXT_DATA} \ + task.kenlm_path=${KENLM_PATH} \ + common.user_dir=${FAIRSEQ_ROOT}/examples/wav2vec/unsupervised \ + model.code_penalty=2,4 model.gradient_penalty=1.5,2.0 \ + model.smoothness_weight=0.5,0.75,1.0 'common.seed=range(0,5)' +``` + + +Once we find the best checkpoint (chosen using unsupervised metric that combined language model perplexity and vocabulary usage), we can use it to generate phone labels (or word labels with an appropriate kaldi WFST): + +```shell +python w2vu_generate.py --config-dir config/generate --config-name viterbi \ +fairseq.common.user_dir=${FAIRSEQ_ROOT}/examples/wav2vec/unsupervised \ +fairseq.task.data=/path/to/dir/with/features \ +fairseq.common_eval.path=/path/to/gan/checkpoint \ +fairseq.dataset.gen_subset=valid results_path=/where/to/save/transcriptions +``` + +The decoding without LM works best on the same adjacent-mean-pooled features that the gan was trained on, while decoding with LM works better on features before the adjacent timestep mean-pooling step (without the "_pooled" suffix). + +## Iterative self-training + Kaldi LM-decoding +After the GAN training provides a first unsupervised model, we can then progressively refine the quality of transcriptions using several iterations of semi-supervised learning. We perform two iterations: first, pseudo-label the training data with the unsupervised GAN model and train an HMM on the pseudo-labels. Second, we relabel the training data with the HMM and then fine-tune the original wav2vec 2.0 model using the HMM pseudo-labels with a CTC loss. Note that HMM models use phonemes as output, while wav2vec 2.0 use letter. Both are decoded using WFST decoders into words. + + +Please see [this README](kaldi_self_train/README.md) for more instructions on how to do iterative self-training + Kaldi LM-decoding. + +*** Note: these instructions are a work in progress and will be updated over the next few days diff --git a/examples/wav2vec/unsupervised/__init__.py b/examples/wav2vec/unsupervised/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/examples/wav2vec/unsupervised/config/finetuning/w2v_finetune.yaml b/examples/wav2vec/unsupervised/config/finetuning/w2v_finetune.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e94da2ba4e46de564cb4619b5e5f955bddc103cc --- /dev/null +++ b/examples/wav2vec/unsupervised/config/finetuning/w2v_finetune.yaml @@ -0,0 +1,62 @@ +# @package _group_ + +common: + fp16: true + log_format: json + log_interval: 200 + tensorboard_logdir: tb + +checkpoint: + no_epoch_checkpoints: true + save_interval_updates: 20000 + +task: + _name: audio_pretraining + data: ??? + normalize: true + labels: ltr + +dataset: + num_workers: 6 + max_tokens: 800000 + skip_invalid_size_inputs_valid_test: true + train_subset: train + valid_subset: valid + +distributed_training: + ddp_backend: legacy_ddp + distributed_world_size: 8 + find_unused_parameters: True + +criterion: + _name: ctc + zero_infinity: true + post_process: letter + +optimization: + max_update: 80000 + lr: [0.00003] + sentence_avg: true + update_freq: [1] + +optimizer: + _name: adam + adam_betas: (0.9,0.98) + adam_eps: 1e-08 + +lr_scheduler: + _name: tri_stage + phase_ratio: [0.1, 0.4, 0.5] + final_lr_scale: 0.05 + +model: + _name: wav2vec_ctc + w2v_path: ??? + apply_mask: true + mask_prob: 0.25 + mask_channel_prob: 0.1 + mask_channel_length: 64 + layerdrop: 0.1 + activation_dropout: 0.1 + feature_grad_mult: 0.0 + freeze_finetune_updates: 0 diff --git a/examples/wav2vec/unsupervised/config/gan/w2vu.yaml b/examples/wav2vec/unsupervised/config/gan/w2vu.yaml new file mode 100644 index 0000000000000000000000000000000000000000..74f1829d1497560f6e1e006073f19716d36bc947 --- /dev/null +++ b/examples/wav2vec/unsupervised/config/gan/w2vu.yaml @@ -0,0 +1,115 @@ +# @package _group_ + +common: + fp16: false + fp16_no_flatten_grads: true + log_format: json + log_interval: 100 + tensorboard_logdir: tb + reset_logging: false + suppress_crashes: false + +checkpoint: + save_interval: 1000 + save_interval_updates: 1000 + no_epoch_checkpoints: true + best_checkpoint_metric: weighted_lm_ppl + save_dir: . + +distributed_training: + distributed_world_size: 1 + +task: + _name: unpaired_audio_text + data: ??? + text_data: ??? + labels: phn + sort_by_length: false + unfiltered: false + max_length: null + append_eos: false + kenlm_path: ??? + +dataset: + num_workers: 6 + batch_size: 160 + skip_invalid_size_inputs_valid_test: true + valid_subset: valid + validate_interval: 1000 + validate_interval_updates: 1000 + +criterion: + _name: model + log_keys: + - accuracy_dense + - accuracy_token + - temp + - code_ppl + +optimization: + max_update: 150000 + clip_norm: 5.0 + lr: [0] + +optimizer: + _name: composite + groups: + generator: + lr: [0.0004] + lr_float: null + optimizer: + _name: adam + adam_betas: [0.5,0.98] + adam_eps: 1e-06 + weight_decay: 0 + amsgrad: false + lr_scheduler: + _name: fixed + warmup_updates: 0 + discriminator: + lr: [ 0.0005 ] + lr_float: null + optimizer: + _name: adam + adam_betas: [0.5,0.98] + adam_eps: 1e-06 + weight_decay: 0.0001 + amsgrad: false + lr_scheduler: + _name: fixed + warmup_updates: 0 + +lr_scheduler: pass_through + +model: + _name: wav2vec_u + + discriminator_dim: 384 + discriminator_depth: 2 + discriminator_kernel: 6 + discriminator_linear_emb: false + discriminator_causal: true + discriminator_max_pool: false + discriminator_act_after_linear: false + discriminator_dropout: 0.0 + discriminator_weight_norm: false + + generator_stride: 1 + generator_kernel: 4 + generator_bias: false + generator_dropout: 0.1 + + smoothness_weight: 0.5 + smoothing: 0 + smoothing_one_sided: false + gumbel: false + hard_gumbel: false + gradient_penalty: 1.5 + code_penalty: 4.0 + temp: [ 2,0.1,0.99995 ] + input_dim: 512 + + segmentation: + type: JOIN + mean_pool_join: false + remove_zeros: false diff --git a/examples/wav2vec/unsupervised/config/generate/viterbi.yaml b/examples/wav2vec/unsupervised/config/generate/viterbi.yaml new file mode 100644 index 0000000000000000000000000000000000000000..9c88beebcb15f9047195c8c7e79c21eac59db418 --- /dev/null +++ b/examples/wav2vec/unsupervised/config/generate/viterbi.yaml @@ -0,0 +1,21 @@ +# @package _group_ + +fairseq: + task: + _name: unpaired_audio_text + labels: phn + data: ??? + sort_by_length: false + shuffle: false + text_data: '' + + common_eval: + path: ??? + quiet: true + + dataset: + gen_subset: valid + batch_size: 1 + +w2l_decoder: VITERBI +post_process: silence diff --git a/examples/wav2vec/unsupervised/data/__init__.py b/examples/wav2vec/unsupervised/data/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..d0545627efc9a6f9bb180e351ead519a2cb6dea7 --- /dev/null +++ b/examples/wav2vec/unsupervised/data/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .extracted_features_dataset import ExtractedFeaturesDataset +from .random_input_dataset import RandomInputDataset + + +__all__ = [ + "ExtractedFeaturesDataset", + "RandomInputDataset", +] diff --git a/examples/wav2vec/unsupervised/data/extracted_features_dataset.py b/examples/wav2vec/unsupervised/data/extracted_features_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d6ee9c4a3602be9db8ddfe67d41ce8a96a98ad1e --- /dev/null +++ b/examples/wav2vec/unsupervised/data/extracted_features_dataset.py @@ -0,0 +1,144 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +import logging +import os +import contextlib + +import numpy as np +import torch + +from fairseq.data import FairseqDataset, data_utils + + +logger = logging.getLogger(__name__) + + +class ExtractedFeaturesDataset(FairseqDataset): + def __init__( + self, + path, + split, + min_length=3, + max_length=None, + labels=None, + label_dict=None, + shuffle=True, + sort_by_length=True, + ): + super().__init__() + + self.min_length = min_length + self.max_length = max_length + self.shuffle = shuffle + self.sort_by_length = sort_by_length + self.label_dict = label_dict + + if labels is not None: + assert label_dict is not None + + self.sizes = [] + self.offsets = [] + self.labels = [] + + path = os.path.join(path, split) + data_path = path + self.data = np.load(data_path + ".npy", mmap_mode="r") + + offset = 0 + skipped = 0 + + if not os.path.exists(path + f".{labels}"): + labels = None + + with open(data_path + ".lengths", "r") as len_f, open( + path + f".{labels}", "r" + ) if labels is not None else contextlib.ExitStack() as lbl_f: + for line in len_f: + length = int(line.rstrip()) + lbl = None if labels is None else next(lbl_f).rstrip().split() + if length >= min_length and ( + max_length is None or length <= max_length + ): + self.sizes.append(length) + self.offsets.append(offset) + if lbl is not None: + self.labels.append(lbl) + offset += length + + self.sizes = np.asarray(self.sizes) + self.offsets = np.asarray(self.offsets) + + logger.info(f"loaded {len(self.offsets)}, skipped {skipped} samples") + + def __getitem__(self, index): + offset = self.offsets[index] + end = self.sizes[index] + offset + feats = torch.from_numpy(self.data[offset:end].copy()).float() + + res = {"id": index, "features": feats} + if len(self.labels) > 0: + res["target"] = self.label_dict.encode_line( + self.labels[index], + line_tokenizer=lambda x: x, + append_eos=False, + ) + + return res + + def __len__(self): + return len(self.sizes) + + def collater(self, samples): + if len(samples) == 0: + return {} + + features = [s["features"] for s in samples] + sizes = [len(s) for s in features] + + target_size = max(sizes) + + collated_features = features[0].new_zeros( + len(features), target_size, features[0].size(-1) + ) + padding_mask = torch.BoolTensor(collated_features.shape[:-1]).fill_(False) + for i, (f, size) in enumerate(zip(features, sizes)): + collated_features[i, :size] = f + padding_mask[i, size:] = True + + res = { + "id": torch.LongTensor([s["id"] for s in samples]), + "net_input": {"features": collated_features, "padding_mask": padding_mask}, + } + + if len(self.labels) > 0: + target = data_utils.collate_tokens( + [s["target"] for s in samples], + pad_idx=self.label_dict.pad(), + left_pad=False, + ) + res["target"] = target + return res + + def num_tokens(self, index): + return self.size(index) + + def size(self, index): + return self.sizes[index] + + def ordered_indices(self): + """Return an ordered list of indices. Batches will be constructed based + on this order.""" + if self.shuffle: + order = [np.random.permutation(len(self))] + else: + order = [np.arange(len(self))] + + if self.sort_by_length: + order.append(self.sizes) + return np.lexsort(order)[::-1] + else: + return order[0] diff --git a/examples/wav2vec/unsupervised/data/random_input_dataset.py b/examples/wav2vec/unsupervised/data/random_input_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..886505616cc7f7a515ecebf34fae5c2bc541de03 --- /dev/null +++ b/examples/wav2vec/unsupervised/data/random_input_dataset.py @@ -0,0 +1,62 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import random +from typing import List + +from fairseq.data import BaseWrapperDataset, data_utils + + +class RandomInputDataset(BaseWrapperDataset): + def __init__( + self, + dataset, + random_input_dataset, + input_key_path: List[str], + add_to_input, + pad_idx, + ): + super().__init__(dataset) + self.random_input_dataset = random_input_dataset + if isinstance(input_key_path, str): + input_key_path = [input_key_path] + assert len(input_key_path) > 0 + self.input_key_path = input_key_path + self.add_to_input = add_to_input + self.pad_idx = pad_idx + + def get_target(self, item): + target_loc = item + for p in self.input_key_path[:-1]: + target_loc = target_loc[p] + return self.input_key_path[-1], target_loc + + def get_target_value(self, item): + k, target_loc = self.get_target(item) + return target_loc[k] + + def __getitem__(self, index): + item = self.dataset[index] + k, target_loc = self.get_target(item) + target_loc[k] = random.choice(self.random_input_dataset) + return item + + def collater(self, samples): + collated = self.dataset.collater(samples) + if len(collated) == 0: + return collated + indices = set(collated["id"].tolist()) + + random_inputs = data_utils.collate_tokens( + [self.get_target_value(s) for s in samples if s["id"] in indices], + pad_idx=self.pad_idx, + left_pad=False, + ) + k, target_loc = self.get_target( + collated if not self.add_to_input else collated["net_input"] + ) + target_loc[k] = random_inputs + + return collated diff --git a/examples/wav2vec/unsupervised/kaldi_self_train/README.md b/examples/wav2vec/unsupervised/kaldi_self_train/README.md new file mode 100644 index 0000000000000000000000000000000000000000..314984fcbb6825169193b21bd6bb3fca5fd2503b --- /dev/null +++ b/examples/wav2vec/unsupervised/kaldi_self_train/README.md @@ -0,0 +1,56 @@ +# Self-Training with Kaldi HMM Models +This folder contains recipes for self-training on pseudo phone transcripts and +decoding into phones or words with [kaldi](https://github.com/kaldi-asr/kaldi). + +To start, download and install kaldi follow its instruction, and place this +folder in `path/to/kaldi/egs`. + +## Training +Assuming the following has been prepared: +- `w2v_dir`: contains features `{train,valid}.{npy,lengths}`, real transcripts `{train,valid}.${label}`, and dict `dict.${label}.txt` +- `lab_dir`: contains pseudo labels `{train,valid}.txt` +- `arpa_lm`: Arpa-format n-gram phone LM for decoding +- `arpa_lm_bin`: Arpa-format n-gram phone LM for unsupervised model selection to be used with KenLM + +Set these variables in `train.sh`, as well as `out_dir`, the output directory, +and then run it. + +The output will be: +``` +==== WER w.r.t. real transcript (select based on unsupervised metric) +INFO:root:./out/exp/mono/decode_valid/scoring/14.0.0.tra.txt: score 0.9178 wer 28.71% lm_ppl 24.4500 gt_wer 25.57% +INFO:root:./out/exp/tri1/decode_valid/scoring/17.1.0.tra.txt: score 0.9257 wer 26.99% lm_ppl 30.8494 gt_wer 21.90% +INFO:root:./out/exp/tri2b/decode_valid/scoring/8.0.0.tra.txt: score 0.7506 wer 23.15% lm_ppl 25.5944 gt_wer 15.78% +``` +where `wer` is the word eror rate with respect to the pseudo label, `gt_wer` to +the ground truth label, `lm_ppl` the language model perplexity of HMM prediced +transcripts, and `score` is the unsupervised metric for model selection. We +choose the model and the LM parameter of the one with the lowest score. In the +example above, it is `tri2b`, `8.0.0`. + + +## Decoding into Phones +In `decode_phone.sh`, set `out_dir` the same as used in `train.sh`, set +`dec_exp` and `dec_lmparam` to the selected model and LM parameter (e.g. +`tri2b` and `8.0.0` in the above example). `dec_script` needs to be set +according to `dec_exp`: for mono/tri1/tri2b, use `decode.sh`; for tri3b, use +`decode_fmllr.sh`. + +The output will be saved at `out_dir/dec_data` + + +## Decoding into Words +`decode_word_step1.sh` prepares WFSTs for word decoding. Besides the variables +mentioned above, set +- `wrd_arpa_lm`: Arpa-format n-gram word LM for decoding +- `wrd_arpa_lm_bin`: Arpa-format n-gram word LM for unsupervised model selection + +`decode_word_step1.sh` decodes the `train` and `valid` split into word and runs +unsupervised model selection using the `valid` split. The output is like: +``` +INFO:root:./out/exp/tri2b/decodeword_valid/scoring/17.0.0.tra.txt: score 1.8693 wer 24.97% lm_ppl 1785.5333 gt_wer 31.45% +``` + +After determining the LM parameter (`17.0.0` in the example above), set it in +`decode_word_step2.sh` and run it. The output will be saved at +`out_dir/dec_data_word`. diff --git a/examples/wav2vec/unsupervised/kaldi_self_train/st/cmd.sh b/examples/wav2vec/unsupervised/kaldi_self_train/st/cmd.sh new file mode 100644 index 0000000000000000000000000000000000000000..e74953194d41f0d93855d41b2acef08556d92477 --- /dev/null +++ b/examples/wav2vec/unsupervised/kaldi_self_train/st/cmd.sh @@ -0,0 +1,15 @@ +# you can change cmd.sh depending on what type of queue you are using. +# If you have no queueing system and want to run on a local machine, you +# can change all instances 'queue.pl' to run.pl (but be careful and run +# commands one by one: most recipes will exhaust the memory on your +# machine). queue.pl works with GridEngine (qsub). slurm.pl works +# with slurm. Different queues are configured differently, with different +# queue names and different ways of specifying things like memory; +# to account for these differences you can create and edit the file +# conf/queue.conf to match your queue's configuration. Search for +# conf/queue.conf in http://kaldi-asr.org/doc/queue.html for more information, +# or search for the string 'default_config' in utils/queue.pl or utils/slurm.pl. + +export train_cmd="run.pl --mem 2G" +export decode_cmd="run.pl --mem 4G" +export mkgraph_cmd="run.pl --mem 8G" diff --git a/examples/wav2vec/unsupervised/kaldi_self_train/st/decode_phone.sh b/examples/wav2vec/unsupervised/kaldi_self_train/st/decode_phone.sh new file mode 100644 index 0000000000000000000000000000000000000000..947342a0b7d8f50bcf4164b284ef3303a1247b64 --- /dev/null +++ b/examples/wav2vec/unsupervised/kaldi_self_train/st/decode_phone.sh @@ -0,0 +1,33 @@ +#!/bin/bash + +# decode into phones (and prepare a new data directory for HMM outputs) + +. ./path.sh + +set -eu + +out_dir= # same as in train.sh +dec_lmparam= # LM hyperparameters (e.g., 7.0.0) +dec_exp= +dec_script= +dec_splits="train valid" +dec_data_dir=$out_dir/dec_data # where to write HMM output + +data_dir=${out_dir}/data + +local/decode.sh --nj 40 --graph_name graph \ + --val_sets "$dec_splits" --decode_script $dec_script \ + $out_dir/exp/$dec_exp $data_dir $data_dir/lang_test + +if [ ! -z $dec_lmparam ]; then + for x in $dec_splits; do + mkdir -p $dec_data_dir/$x + cp $data_dir/$x/{feats.scp,cmvn.scp,utt2spk,spk2utt} $dec_data_dir/$x/ + + tra=$out_dir/exp/$dec_exp/decode_${x}/scoring/${dec_lmparam}.tra + cat $tra | utils/int2sym.pl -f 2- $data_dir/lang/words.txt | \ + sed 's:<UNK>::g' | sed 's:<SIL>::g' > $dec_data_dir/${x}/text + utils/fix_data_dir.sh $dec_data_dir/${x} + echo "WER on ${x} is" $(compute-wer ark:$data_dir/${x}_gt/text ark:$dec_data_dir/$x/text | cut -d" " -f2-) + done +fi diff --git a/examples/wav2vec/unsupervised/kaldi_self_train/st/decode_word_step1.sh b/examples/wav2vec/unsupervised/kaldi_self_train/st/decode_word_step1.sh new file mode 100644 index 0000000000000000000000000000000000000000..c1276bbe4d0e02deb984c7c10d6c0486dce09a5f --- /dev/null +++ b/examples/wav2vec/unsupervised/kaldi_self_train/st/decode_word_step1.sh @@ -0,0 +1,46 @@ +#!/bin/bash + +# prepare word WFSTs, reference data, and decode + +set -eu + +w2v_dir= # same as in train.sh +out_dir= # same as in train.sh +lexicon= # word to phone mapping +wrd_arpa_lm= # word LM +wrd_arpa_lm_bin= # word LM for KenLM, used in unsupervised selection + +dec_exp= # what HMM stage to decode (e.g., tri3b) +dec_script= # what decoding script to use (e.g., steps/decode_fmllr.sh) +phn_label=phnc +wrd_label=wrd +dec_suffix=word +dec_splits="train valid" +valid_split="valid" + +data_dir=$out_dir/data +wrd_data_dir=$out_dir/data_word + +lexicon_clean=$(mktemp) +cat $lexicon | sort | uniq > $lexicon_clean +local/prepare_lang_word.sh $w2v_dir/dict.${phn_label}.txt $data_dir $lexicon_clean && rm $lexicon_clean +local/prepare_lm.sh --langdir $data_dir/lang_word --lmdir $data_dir/lang_test_word $wrd_arpa_lm $data_dir + +for x in $dec_splits; do + x_gt=${x}_gt + mkdir -p $wrd_data_dir/$x_gt + cp $data_dir/$x_gt/{feats.scp,cmvn.scp,utt2spk,spk2utt} $wrd_data_dir/$x_gt/ + python local/copy_aligned_text.py < $w2v_dir/$x.$wrd_label > $wrd_data_dir/$x_gt/text +done + +local/decode.sh --nj 40 --graph_name graph${dec_suffix} --decode_suffix $dec_suffix \ + --val_sets "$dec_splits" --decode_script $dec_script \ + $out_dir/exp/$dec_exp $data_dir $data_dir/lang_test_word + +local/unsup_select_decode_word.sh \ + --split $valid_split --kenlm_path $wrd_arpa_lm_bin \ + --ref_txt $wrd_data_dir/${valid_split}_gt/text \ + --psd_txt $data_dir/${valid_split}/text \ + --dec_name decode${dec_suffix} --graph_name graph${dec_suffix} \ + --phonemize_lexicon $data_dir/local/dict_word/lexicon.txt \ + $out_dir/exp diff --git a/examples/wav2vec/unsupervised/kaldi_self_train/st/decode_word_step2.sh b/examples/wav2vec/unsupervised/kaldi_self_train/st/decode_word_step2.sh new file mode 100644 index 0000000000000000000000000000000000000000..59a6cbb12539cf62658f8344f7be7cecf2e3380f --- /dev/null +++ b/examples/wav2vec/unsupervised/kaldi_self_train/st/decode_word_step2.sh @@ -0,0 +1,30 @@ +#!/bin/bash + +# prepare a new data directory of HMM word output + +. ./path.sh + +set -eu + +out_dir= # same as in train.sh +dec_lmparam= # LM hyperparameters (e.g., 7.0.0) + +dec_exp=tri3b # what HMM stage to decode (e.g., tri3b) +dec_suffix=word +dec_splits="train valid" +dec_data_dir=$out_dir/dec_data_word # where to write HMM output + +data_dir=$out_dir/data +wrd_data_dir=$out_dir/data_word + +for x in $dec_splits; do + mkdir -p $dec_data_dir/$x + cp $data_dir/$x/{feats.scp,cmvn.scp,utt2spk,spk2utt} $dec_data_dir/$x/ + + tra=$out_dir/exp/$dec_exp/decode${dec_suffix}_${x}/scoring/${dec_lmparam}.tra + cat $tra | utils/int2sym.pl -f 2- $data_dir/lang_word/words.txt | \ + sed 's:<UNK>::g' | sed 's:<SIL>::g' > $dec_data_dir/$x/text + utils/fix_data_dir.sh $dec_data_dir/$x + echo "WER on $x is" $(compute-wer ark:$wrd_data_dir/${x}_gt/text ark:$dec_data_dir/$x/text | cut -d" " -f2-) +done + diff --git a/examples/wav2vec/unsupervised/kaldi_self_train/st/local/copy_aligned_text.py b/examples/wav2vec/unsupervised/kaldi_self_train/st/local/copy_aligned_text.py new file mode 100644 index 0000000000000000000000000000000000000000..5f4faa99218b0b30c980cad167c52b2297cd92c3 --- /dev/null +++ b/examples/wav2vec/unsupervised/kaldi_self_train/st/local/copy_aligned_text.py @@ -0,0 +1,4 @@ +import sys + +for idx, line in enumerate(sys.stdin): + print(f"utt{idx:010d} {line}", end='') \ No newline at end of file diff --git a/examples/wav2vec/unsupervised/kaldi_self_train/st/local/decode.sh b/examples/wav2vec/unsupervised/kaldi_self_train/st/local/decode.sh new file mode 100755 index 0000000000000000000000000000000000000000..811cb63c88bb7cdd03b0a250ef2db32b5eaa50df --- /dev/null +++ b/examples/wav2vec/unsupervised/kaldi_self_train/st/local/decode.sh @@ -0,0 +1,38 @@ +#!/bin/bash + +set -u + +val_sets="dev_other" +graph_name=graph +decode_suffix="" +decode_script="steps/decode_fmllr.sh" +decode_args="" +nj=60 + +. ./cmd.sh +. ./path.sh +. parse_options.sh + +set -x +exp_dir=$1 +data_root=$2 +lang_test=$3 + +graph=$exp_dir/$graph_name + +if [ ! -d $graph ]; then + utils/mkgraph.sh $lang_test $exp_dir $graph +fi + +for part in $val_sets; do + dec_dir=$exp_dir/decode${decode_suffix}_${part} + if [ ! -d $dec_dir ]; then + echo "decoding $part for $exp_dir" + $decode_script --nj $nj --cmd "$decode_cmd" $decode_args \ + $graph $data_root/$part $dec_dir & + else + echo "$dec_dir exists. skip" + fi +done + +wait diff --git a/examples/wav2vec/unsupervised/kaldi_self_train/st/local/prepare_data_from_w2v.py b/examples/wav2vec/unsupervised/kaldi_self_train/st/local/prepare_data_from_w2v.py new file mode 100644 index 0000000000000000000000000000000000000000..66954ea5c9f3f3330e3230860229c7c4046a5d6a --- /dev/null +++ b/examples/wav2vec/unsupervised/kaldi_self_train/st/local/prepare_data_from_w2v.py @@ -0,0 +1,56 @@ +import kaldi_io +import numpy as np +import os + + +def get_parser(): + import argparse + parser = argparse.ArgumentParser() + parser.add_argument("w2v_dir", help="wav2vec feature and text directory") + parser.add_argument("tar_root", help="output data directory in kaldi's format") + parser.add_argument("split", help="name of the subset") + parser.add_argument("--label", default="", help="if specified, copy labels too") + return parser + +def main(): + parser = get_parser() + args = parser.parse_args() + + tar_dir = os.path.join(args.tar_root, args.split) + os.makedirs(tar_dir, exist_ok=True) + + lengths_path = os.path.join(args.w2v_dir, f"{args.split}.lengths") + with open(lengths_path) as f: + lengths = [int(line.rstrip()) for line in f] + offsets = [0] + np.cumsum(lengths[:-1]).tolist() + feats = np.load( + os.path.join(args.w2v_dir, f"{args.split}.npy"), + mmap_mode="r" + ) + assert feats.shape[0] == sum(lengths), \ + f"lengths mismatch {feats.shape[0]} != {sum(lengths)}" + + ark_path = os.path.join(tar_dir, "feats.ark") + scp_path = os.path.join(tar_dir, "feats.scp") + wspec = f"ark:| copy-feats --compress=true ark:- ark,scp:{ark_path},{scp_path}" + with kaldi_io.open_or_fd(wspec, "wb") as f: + for idx, (offset, length) in enumerate(zip(offsets, lengths)): + feat = feats[offset:offset+length] + kaldi_io.write_mat(f, feat, key=f"utt{idx:010d}") + + u2s_path = os.path.join(tar_dir, "utt2spk") + s2u_path = os.path.join(tar_dir, "spk2utt") + with open(u2s_path, "w") as f_u2s, open(s2u_path, "w") as f_s2u: + for idx in range(len(lengths)): + f_u2s.write(f"utt{idx:010d} utt{idx:010d}\n") + f_s2u.write(f"utt{idx:010d} utt{idx:010d}\n") + + if bool(args.label): + lab_path = os.path.join(args.w2v_dir, f"{args.split}.{args.label}") + txt_path = os.path.join(tar_dir, "text") + with open(lab_path) as f_lab, open(txt_path, "w") as f_txt: + for idx, line in enumerate(f_lab): + f_txt.write(f"utt{idx:010d} {line}") + +if __name__ == "__main__": + main() diff --git a/examples/wav2vec/unsupervised/kaldi_self_train/st/local/prepare_lang.sh b/examples/wav2vec/unsupervised/kaldi_self_train/st/local/prepare_lang.sh new file mode 100755 index 0000000000000000000000000000000000000000..e9a80001eb47d5af863d6aab11a59362a59cef61 --- /dev/null +++ b/examples/wav2vec/unsupervised/kaldi_self_train/st/local/prepare_lang.sh @@ -0,0 +1,37 @@ +#!/bin/bash + +sil_prob=0.5 +num_sil_states=3 +num_nonsil_states=1 + +. ./cmd.sh +. ./path.sh +. parse_options.sh + +set -eux + +dict=$1 +data_dir=$2 + +dict_dir=$data_dir/local/dict +tmplm_dir=$data_dir/local/lang_tmp +lm_dir=$data_dir/lang + +mkdir -p $dict_dir $tmplm_dir $lm_dir + +# prepare dict +echo "SIL" > $dict_dir/silence_phones.txt +echo "SIL" > $dict_dir/optional_silence.txt +awk '{print $1}' $dict > $dict_dir/nonsilence_phones.txt + +echo "SIL SIL" > $dict_dir/lexicon.txt +echo "<UNK> SIL" >> $dict_dir/lexicon.txt +awk '{print $1" "$1}' $dict >> $dict_dir/lexicon.txt + +echo "SIL" > $dict_dir/extra_questions.txt +awk '{printf $1" "} END {printf "\n"}' $dict >> $dict_dir/extra_questions.txt + +# prepare lang +utils/prepare_lang.sh --sil-prob $sil_prob --position-dependent-phones false \ + --num_sil_states $num_sil_states --num_nonsil_states $num_nonsil_states \ + $dict_dir "<UNK>" $tmplm_dir $lm_dir diff --git a/examples/wav2vec/unsupervised/kaldi_self_train/st/local/prepare_lang_word.sh b/examples/wav2vec/unsupervised/kaldi_self_train/st/local/prepare_lang_word.sh new file mode 100755 index 0000000000000000000000000000000000000000..a7ea3877beefe1d4d53f9f7e32b004d8ce01e22a --- /dev/null +++ b/examples/wav2vec/unsupervised/kaldi_self_train/st/local/prepare_lang_word.sh @@ -0,0 +1,35 @@ +#!/bin/bash + +num_sil_states=3 +num_nonsil_states=1 + +. ./cmd.sh +. ./path.sh +. parse_options.sh + +set -eux + +dict=$1 +data_dir=$2 +lexicon=$3 + +dict_dir=$data_dir/local/dict_word +tmplm_dir=$data_dir/local/lang_tmp_word +lm_dir=$data_dir/lang_word + +mkdir -p $dict_dir $tmplm_dir $lm_dir + +# prepare dict +echo "SIL" > $dict_dir/silence_phones.txt +echo "SIL" > $dict_dir/optional_silence.txt +awk '{print $1}' $dict > $dict_dir/nonsilence_phones.txt + +(echo "!SIL SIL"; echo "<UNK> SIL";) | cat - $lexicon > $dict_dir/lexicon.txt + +echo "SIL" > $dict_dir/extra_questions.txt +awk '{printf $1" "} END {printf "\n"}' $dict >> $dict_dir/extra_questions.txt + +# prepare lang +utils/prepare_lang.sh --position-dependent-phones false \ + --num_sil_states $num_sil_states --num_nonsil_states $num_nonsil_states \ + $dict_dir "<UNK>" $tmplm_dir $lm_dir diff --git a/examples/wav2vec/unsupervised/kaldi_self_train/st/local/prepare_lm.sh b/examples/wav2vec/unsupervised/kaldi_self_train/st/local/prepare_lm.sh new file mode 100755 index 0000000000000000000000000000000000000000..c2edcefede2da3b6a991b9c8fbc78c96d46d27cb --- /dev/null +++ b/examples/wav2vec/unsupervised/kaldi_self_train/st/local/prepare_lm.sh @@ -0,0 +1,35 @@ +#!/usr/bin/env bash + +langdir="" +lmdir="" + +. ./cmd.sh +. ./path.sh +. parse_options.sh + +arpa_lm=$1 +data=$2 + +if [ -z $langdir ]; then + langdir=$data/lang +fi +if [ -z $lmdir ]; then + lmdir=$data/lang_test +fi + +if [ ! -d $langdir ]; then + echo "$langdir not found. run local/prepare_lang.sh first" && exit 1 +fi + +mkdir -p $lmdir +cp -r $langdir/* $lmdir + +if [[ "$arpa_lm" == *.gz ]]; then + gunzip -c $arpa_lm | arpa2fst --disambig-symbol=#0 --read-symbol-table=$lmdir/words.txt - $lmdir/G.fst +else + arpa2fst --disambig-symbol=#0 --read-symbol-table=$lmdir/words.txt $arpa_lm $lmdir/G.fst +fi +fstisstochastic $lmdir/G.fst +utils/validate_lang.pl $lmdir || exit 1 + +echo "done preparing lm ($lmdir)" diff --git a/examples/wav2vec/unsupervised/kaldi_self_train/st/local/score.sh b/examples/wav2vec/unsupervised/kaldi_self_train/st/local/score.sh new file mode 100755 index 0000000000000000000000000000000000000000..cb5bbb7277bfb9f2d5440da0514bf7b16da8140d --- /dev/null +++ b/examples/wav2vec/unsupervised/kaldi_self_train/st/local/score.sh @@ -0,0 +1,63 @@ +#!/usr/bin/env bash +# Copyright 2012 Johns Hopkins University (Author: Daniel Povey) +# 2014 Guoguo Chen +# Apache 2.0 + +[ -f ./path.sh ] && . ./path.sh + +# begin configuration section. +cmd=run.pl +stage=0 +decode_mbr=true +word_ins_penalty=0.0,0.5,1.0 +min_lmwt=7 +max_lmwt=17 +iter=final +#end configuration section. + +[ -f ./path.sh ] && . ./path.sh +. parse_options.sh || exit 1; + +if [ $# -ne 3 ]; then + echo "Usage: local/score.sh [--cmd (run.pl|queue.pl...)] <data-dir> <lang-dir|graph-dir> <decode-dir>" + echo " Options:" + echo " --cmd (run.pl|queue.pl...) # specify how to run the sub-processes." + echo " --stage (0|1|2) # start scoring script from part-way through." + echo " --decode_mbr (true/false) # maximum bayes risk decoding (confusion network)." + echo " --min_lmwt <int> # minumum LM-weight for lattice rescoring " + echo " --max_lmwt <int> # maximum LM-weight for lattice rescoring " + exit 1; +fi + +data=$1 +lang_or_graph=$2 +dir=$3 + +symtab=$lang_or_graph/words.txt + +for f in $symtab $dir/lat.1.gz $data/text; do + [ ! -f $f ] && echo "score.sh: no such file $f" && exit 1; +done + +mkdir -p $dir/scoring/log + +cat $data/text | sed 's:<NOISE>::g' | sed 's:<SPOKEN_NOISE>::g' > $dir/scoring/test_filt.txt + +for wip in $(echo $word_ins_penalty | sed 's/,/ /g'); do + $cmd LMWT=$min_lmwt:$max_lmwt $dir/scoring/log/best_path.LMWT.$wip.log \ + lattice-scale --inv-acoustic-scale=LMWT "ark:gunzip -c $dir/lat.*.gz|" ark:- \| \ + lattice-add-penalty --word-ins-penalty=$wip ark:- ark:- \| \ + lattice-best-path --word-symbol-table=$symtab \ + ark:- ark,t:$dir/scoring/LMWT.$wip.tra || exit 1; +done + +# Note: the double level of quoting for the sed command +for wip in $(echo $word_ins_penalty | sed 's/,/ /g'); do + $cmd LMWT=$min_lmwt:$max_lmwt $dir/scoring/log/score.LMWT.$wip.log \ + cat $dir/scoring/LMWT.$wip.tra \| \ + utils/int2sym.pl -f 2- $symtab \| sed 's:\<UNK\>::g' \| \ + compute-wer --text --mode=present \ + ark:$dir/scoring/test_filt.txt ark,p:- ">&" $dir/wer_LMWT_$wip || exit 1; +done + +exit 0; diff --git a/examples/wav2vec/unsupervised/kaldi_self_train/st/local/show_wer.sh b/examples/wav2vec/unsupervised/kaldi_self_train/st/local/show_wer.sh new file mode 100755 index 0000000000000000000000000000000000000000..9ecf1690c67f8a019009ef32d973fbd45b56c7ca --- /dev/null +++ b/examples/wav2vec/unsupervised/kaldi_self_train/st/local/show_wer.sh @@ -0,0 +1,52 @@ +#!/bin/bash + +split="dev_other" +ref_data="" +get_best_wer=true +dec_name="decode" +graph_name="graph" + +. ./cmd.sh +. ./path.sh +. parse_options.sh + +exp_root=$1 + +set -eu + +echo "==== WER w.r.t. pseudo transcript" +for x in $exp_root/*/${dec_name}_${split}*; do grep WER $x/wer_* 2>/dev/null | utils/best_wer.sh; done + + +if [ ! -z $ref_data ]; then + echo "==== WER w.r.t. real transcript (select based on pseudo WER)" + ref_txt=$ref_data/$split/text + for x in $exp_root/*/${dec_name}_${split}*; do + lang=$(dirname $x)/$graph_name + + lmwt=$( + grep WER $x/wer_* 2>/dev/null | utils/best_wer.sh | + sed 's/.*wer_\(.*\)$/\1/g' | sed 's/_/./g' + ) + tra=$x/scoring/$lmwt.tra + cat $tra | utils/int2sym.pl -f 2- $lang/words.txt | sed 's:<UNK>::g' | sed 's:<SIL>::g' | \ + compute-wer --text --mode=present \ + ark:$ref_txt ark,p:- 2> /dev/null | grep WER | xargs -I{} echo {} $tra + done +fi + +if [ ! -z $ref_data ] && $get_best_wer; then + echo "==== WER w.r.t. real transcript (select based on true WER)" + ref_txt=$ref_data/$split/text + for x in $exp_root/*/${dec_name}_${split}*; do + lang=$(dirname $x)/$graph_name + + for tra in $x/scoring/*.tra; do + cat $tra | utils/int2sym.pl -f 2- $lang/words.txt | sed 's:<UNK>::g' | sed 's:<SIL>::g' | \ + compute-wer --text --mode=present \ + ark:$ref_txt ark,p:- 2> /dev/null | grep WER | xargs -I{} echo {} $tra + done | sort -k2n | head -n1 + done +fi + +exit 0; diff --git a/examples/wav2vec/unsupervised/kaldi_self_train/st/local/train_subset_lgbeam.sh b/examples/wav2vec/unsupervised/kaldi_self_train/st/local/train_subset_lgbeam.sh new file mode 100755 index 0000000000000000000000000000000000000000..913c1d8e4357c146026b86e78f0b16f921776441 --- /dev/null +++ b/examples/wav2vec/unsupervised/kaldi_self_train/st/local/train_subset_lgbeam.sh @@ -0,0 +1,129 @@ +#!/usr/bin/env bash + +out_root=/tmp +out_name=train_${RANDOM} +num_nonsil_states=1 + +valid="dev_other" +train="train" +mono_size="-1" # 2000 +tri1_size="-1" # 5000 +tri2b_size="-1" # 10000 +tri3b_size="-1" # 10000 + +# Acoustic model parameters +numLeavesTri1=2000 +numGaussTri1=10000 +numLeavesMLLT=2500 +numGaussMLLT=15000 +numLeavesSAT=2500 +numGaussSAT=15000 + +stage=1 +max_stage=1 + +. ./cmd.sh +. ./path.sh +. parse_options.sh + +data=$1 +lang=$2 +lang_test=$3 + +exp_root=$out_root/$out_name + +# you might not want to do this for interactive shells. +set -e + + +if [ $stage -le 1 ] && [ $max_stage -ge 1 ]; then + # train a monophone system + if [ ! $mono_size -eq -1 ]; then + utils/subset_data_dir.sh $data/$train $mono_size $data/${train}_${mono_size} + mono_train=${train}_${mono_size} + else + mono_train=${train} + fi + + steps/train_mono.sh --boost-silence 1.25 --nj 20 --cmd "$train_cmd" \ + --initial-beam 40 --regular-beam 60 --retry-beam 120 \ + $data/$mono_train $lang $exp_root/mono + + utils/mkgraph.sh $lang_test $exp_root/mono $exp_root/mono/graph + steps/decode.sh --nj 20 --cmd "$decode_cmd" \ + $exp_root/mono/graph $data/$valid $exp_root/mono/decode_$valid & +fi + + +if [ $stage -le 2 ] && [ $max_stage -ge 2 ]; then + # train a first delta + delta-delta triphone system on a subset of 5000 utterances + if [ ! $tri1_size -eq -1 ]; then + utils/subset_data_dir.sh $data/$train $tri1_size $data/${train}_${tri1_size} + tri1_train=${train}_${tri1_size} + else + tri1_train=${train} + fi + + steps/align_si.sh --boost-silence 1.25 --nj 10 --cmd "$train_cmd" \ + $data/$tri1_train $lang \ + $exp_root/mono $exp_root/mono_ali_${tri1_train} + + steps_gan/train_deltas.sh --boost-silence 1.25 --cmd "$train_cmd" \ + --num_nonsil_states $num_nonsil_states $numLeavesTri1 $numGaussTri1 \ + $data/$tri1_train $lang \ + $exp_root/mono_ali_${tri1_train} $exp_root/tri1 + + utils/mkgraph.sh $lang_test $exp_root/tri1 $exp_root/tri1/graph + steps/decode.sh --nj 20 --cmd "$decode_cmd" \ + $exp_root/tri1/graph $data/$valid $exp_root/tri1/decode_$valid & +fi + +if [ $stage -le 3 ] && [ $max_stage -ge 3 ]; then + # train an LDA+MLLT system. + if [ ! $tri2b_size -eq -1 ]; then + utils/subset_data_dir.sh $data/$train $tri2b_size $data/${train}_${tri2b_size} + tri2b_train=${train}_${tri2b_size} + else + tri2b_train=${train} + fi + + steps/align_si.sh --nj 10 --cmd "$train_cmd" \ + $data/$tri2b_train $lang \ + $exp_root/tri1 $exp_root/tri1_ali_${tri2b_train} + + steps_gan/train_lda_mllt.sh --cmd "$train_cmd" \ + --num_nonsil_states $num_nonsil_states \ + --splice-opts "--left-context=3 --right-context=3" $numLeavesMLLT $numGaussMLLT \ + $data/$tri2b_train $lang \ + $exp_root/tri1_ali_${tri2b_train} $exp_root/tri2b + + utils/mkgraph.sh $lang_test $exp_root/tri2b $exp_root/tri2b/graph + steps/decode.sh --nj 20 --cmd "$decode_cmd" \ + $exp_root/tri2b/graph $data/$valid $exp_root/tri2b/decode_$valid & +fi + + +if [ $stage -le 4 ] && [ $max_stage -ge 4 ]; then + # Train tri3b, which is LDA+MLLT+SAT on 10k utts + if [ ! $tri3b_size -eq -1 ]; then + utils/subset_data_dir.sh $data/$train $tri3b_size $data/${train}_${tri3b_size} + tri3b_train=${train}_${tri3b_size} + else + tri3b_train=${train} + fi + + steps/align_si.sh --nj 10 --cmd "$train_cmd" --use-graphs true \ + $data/$tri3b_train $lang \ + $exp_root/tri2b $exp_root/tri2b_ali_${tri2b_train} + + steps_gan/train_sat.sh --cmd "$train_cmd" \ + --num_nonsil_states $num_nonsil_states $numLeavesSAT $numGaussSAT \ + $data/$tri3b_train $lang \ + $exp_root/tri2b_ali_${tri2b_train} $exp_root/tri3b + + utils/mkgraph.sh $lang_test $exp_root/tri3b $exp_root/tri3b/graph + steps/decode_fmllr.sh --nj 20 --cmd "$decode_cmd" \ + $exp_root/tri3b/graph $data/$valid $exp_root/tri3b/decode_$valid & +fi + +wait diff --git a/examples/wav2vec/unsupervised/kaldi_self_train/st/local/unsup_select.py b/examples/wav2vec/unsupervised/kaldi_self_train/st/local/unsup_select.py new file mode 100644 index 0000000000000000000000000000000000000000..1122c88c1964d8beead63bc8dfe21d41602b83bc --- /dev/null +++ b/examples/wav2vec/unsupervised/kaldi_self_train/st/local/unsup_select.py @@ -0,0 +1,135 @@ +""" +Implement unsupervised metric for decoding hyperparameter selection: + $$ alpha * LM_PPL + ViterbitUER(%) * 100 $$ +""" +import argparse +import logging +import math +import sys + +import kenlm +import editdistance +from g2p_en import G2p + +logging.root.setLevel(logging.INFO) +logging.basicConfig(stream=sys.stdout, level=logging.INFO) +logger = logging.getLogger(__name__) + + +def get_parser(): + parser = argparse.ArgumentParser() + parser.add_argument("ref_tra", help="reference pseudo labels") + parser.add_argument("hyp_tra", help="decoded pseudo labels to be assess") + parser.add_argument("--kenlm_path", default="/checkpoint/abaevski/data/speech/libri/librispeech_lm_novox.phnc_o5.bin", help="") + parser.add_argument("--uppercase", action="store_true", help="") + parser.add_argument("--skipwords", default="", help="") + parser.add_argument("--gt_tra", default="", help="ground truth pseudo labels for computing oracle WER") + parser.add_argument("--min_vt_uer", default=0.0, type=float) + parser.add_argument("--phonemize", action="store_true", help="phonemize word hypotheses, used when reference is phone transcript") + parser.add_argument("--phonemize_lexicon", default="", type=str, help="use a lexicon for phonemizing") + return parser + +def load_tra(tra_path): + with open(tra_path, "r") as f: + uid_to_tra = {} + for line in f: + toks = line.rstrip().split() + uid, tra = toks[0], " ".join(toks[1:]) + uid_to_tra[uid] = tra + logger.debug(f"loaded {len(uid_to_tra)} utterances from {tra_path}") + return uid_to_tra + +def load_lex(lex_path): + with open(lex_path, "r") as f: + w2p = {} + for line in f: + w, p = line.rstrip().split(None, 1) + w2p[w] = p.split() + return w2p + +def compute_wer(ref_uid_to_tra, hyp_uid_to_tra, g2p, g2p_dict): + d_cnt = 0 + w_cnt = 0 + w_cnt_h = 0 + for uid in hyp_uid_to_tra: + ref = ref_uid_to_tra[uid].split() + if g2p_dict is not None: + hyp = [] + for word in hyp_uid_to_tra[uid].split(): + if word in g2p_dict: + hyp = hyp + g2p_dict[word] + else: + logger.warning(f"{word} not in g2p_dict") + elif g2p is not None: + hyp = g2p(hyp_uid_to_tra[uid]) + hyp = [p for p in hyp if p != "'" and p != " "] + hyp = [p[:-1] if p[-1].isnumeric() else p for p in hyp] + else: + hyp = hyp_uid_to_tra[uid].split() + logger.debug(( + f"======================\n" + f"HYP: {' '.join(hyp)}\n" + f"REF: {' '.join(ref)}" + )) + d_cnt += editdistance.eval(ref, hyp) + w_cnt += len(ref) + w_cnt_h += len(hyp) + wer = float(d_cnt) / w_cnt + logger.debug(( + f"wer = {wer*100:.2f}%; num. of ref words = {w_cnt}; " + f"num. of hyp words = {w_cnt_h}; num. of sentences = {len(ref_uid_to_tra)}" + )) + return wer + +def compute_lm_ppl(hyp_uid_to_tra, score_fn): + lm_score = 0. + w_cnt = 0 + for hyp in hyp_uid_to_tra.values(): + cur_score = score_fn(hyp) + cur_cnt = len(hyp.split()) + 1 # plus one for </s> + lm_score += cur_score + w_cnt += cur_cnt + logger.debug(( + f"======================\n" + f"score sum/avg = {cur_score:.2f}/{cur_score/cur_cnt:.2f}\n" + f"hyp = {hyp}" + )) + lm_ppl = math.pow(10, -lm_score / w_cnt) + logger.debug(f"lm ppl = {lm_ppl:.2f}; num. of words = {w_cnt}") + return lm_ppl + +def main(): + args = get_parser().parse_args() + logger.debug(f"Args: {args}") + + ref_uid_to_tra = load_tra(args.ref_tra) + hyp_uid_to_tra = load_tra(args.hyp_tra) + assert not bool(set(hyp_uid_to_tra.keys()) - set(ref_uid_to_tra.keys())) + + lm = kenlm.Model(args.kenlm_path) + skipwords = set(args.skipwords.split(",")) + def compute_lm_score(s): + s = " ".join(w for w in s.split() if w not in skipwords) + s = s.upper() if args.uppercase else s + return lm.score(s) + + g2p, g2p_dict = None, None + if args.phonemize: + if args.phonemize_lexicon: + g2p_dict = load_lex(args.phonemize_lexicon) + else: + g2p = G2p() + + wer = compute_wer(ref_uid_to_tra, hyp_uid_to_tra, g2p, g2p_dict) + lm_ppl = compute_lm_ppl(hyp_uid_to_tra, compute_lm_score) + + gt_wer = -math.inf + if args.gt_tra: + gt_uid_to_tra = load_tra(args.gt_tra) + gt_wer = compute_wer(gt_uid_to_tra, hyp_uid_to_tra, None, None) + + score = math.log(lm_ppl) * max(wer, args.min_vt_uer) + logging.info(f"{args.hyp_tra}: score={score:.4f}; wer={wer*100:.2f}%; lm_ppl={lm_ppl:.4f}; gt_wer={gt_wer*100:.2f}%") + +if __name__ == "__main__": + main() diff --git a/examples/wav2vec/unsupervised/kaldi_self_train/st/local/unsup_select_decode.sh b/examples/wav2vec/unsupervised/kaldi_self_train/st/local/unsup_select_decode.sh new file mode 100755 index 0000000000000000000000000000000000000000..b34c5b6e0688914a53515162f817a93617b609e5 --- /dev/null +++ b/examples/wav2vec/unsupervised/kaldi_self_train/st/local/unsup_select_decode.sh @@ -0,0 +1,37 @@ +#!/bin/bash + +split="dev_other" +ref_txt="" # ground truth transcript path +psd_txt="" # pseudo transcript path +get_best_wer=true +dec_name="decode" +graph_name="graph" +kenlm_path=/checkpoint/abaevski/data/speech/libri/librispeech_lm_novox.phnc_o6.bin + +. ./cmd.sh +. ./path.sh +. parse_options.sh + +exp_root=$1 +unsup_args="" +if [ $# -ge 2 ]; then + unsup_args=$2 +fi + +set -eu + +if [ ! -z $ref_txt ] && $get_best_wer; then + echo "==== WER w.r.t. real transcript (select based on unsupervised metric)" + for x in $exp_root/*/${dec_name}_${split}*; do + lang=$(dirname $x)/$graph_name + + ( + for tra in $x/scoring/*.tra; do + cat $tra | utils/int2sym.pl -f 2- $lang/words.txt | sed 's:<UNK>::g' | sed 's:<SIL>::g' > $tra.txt + python local/unsup_select.py $psd_txt $tra.txt --kenlm_path $kenlm_path --gt_tra $ref_txt $unsup_args + done 2>/dev/null | grep "score=" | sed 's/=/ /g' | sed 's/;//g' | sort -k3n | head -n1 + ) & + done +fi +wait + diff --git a/examples/wav2vec/unsupervised/kaldi_self_train/st/local/unsup_select_decode_word.sh b/examples/wav2vec/unsupervised/kaldi_self_train/st/local/unsup_select_decode_word.sh new file mode 100755 index 0000000000000000000000000000000000000000..c10a6b8809b77bca2b2c02df8b8702725bdd51c7 --- /dev/null +++ b/examples/wav2vec/unsupervised/kaldi_self_train/st/local/unsup_select_decode_word.sh @@ -0,0 +1,35 @@ +#!/bin/bash + +split="dev_other" +ref_txt="" # ground truth transcript path +psd_txt="" # pseudo transcript path +get_best_wer=true +dec_name="decode" +graph_name="graph" +kenlm_path=/checkpoint/abaevski/data/speech/libri/librispeech_lm_novox.phnc_o6.bin +phonemize_lexicon="" + +. ./cmd.sh +. ./path.sh +. parse_options.sh +. /private/home/wnhsu/unsup_asr/fairseq-py-unsup/env.sh + +exp_root=$1 + +set -eu + +if [ ! -z $ref_txt ] && $get_best_wer; then + echo "==== WER w.r.t. real transcript (select based on unsupervised metric)" + for x in $exp_root/*/${dec_name}_${split}*; do + lang=$(dirname $x)/$graph_name + + for tra in $x/scoring/*.tra; do + cat $tra | utils/int2sym.pl -f 2- $lang/words.txt | sed 's:\<UNK\>::g' > $tra.txt + python local/unsup_select.py $psd_txt $tra.txt \ + --kenlm_path $kenlm_path --gt_tra $ref_txt --phonemize \ + --phonemize_lexicon "$phonemize_lexicon" + done | grep "score=" | sed 's/=/ /g' | sed 's/;//g' | sort -k3n | head -n1 + done +fi + + diff --git a/examples/wav2vec/unsupervised/kaldi_self_train/st/path.sh b/examples/wav2vec/unsupervised/kaldi_self_train/st/path.sh new file mode 100755 index 0000000000000000000000000000000000000000..1a6fb5f891b55d9fd978cfe54565f112f7eedce7 --- /dev/null +++ b/examples/wav2vec/unsupervised/kaldi_self_train/st/path.sh @@ -0,0 +1,5 @@ +export KALDI_ROOT=`pwd`/../../.. +export PATH=$PWD/utils/:$KALDI_ROOT/tools/openfst/bin:$PWD:$PATH +[ ! -f $KALDI_ROOT/tools/config/common_path.sh ] && echo >&2 "The standard file $KALDI_ROOT/tools/config/common_path.sh is not present -> Exit!" && exit 1 +. $KALDI_ROOT/tools/config/common_path.sh +export LC_ALL=C diff --git a/examples/wav2vec/unsupervised/kaldi_self_train/st/steps b/examples/wav2vec/unsupervised/kaldi_self_train/st/steps new file mode 120000 index 0000000000000000000000000000000000000000..6e99bf5b5adab1a857cb113ced3567cc4dee8ebe --- /dev/null +++ b/examples/wav2vec/unsupervised/kaldi_self_train/st/steps @@ -0,0 +1 @@ +../../wsj/s5/steps \ No newline at end of file diff --git a/examples/wav2vec/unsupervised/kaldi_self_train/st/steps_gan/train_deltas.sh b/examples/wav2vec/unsupervised/kaldi_self_train/st/steps_gan/train_deltas.sh new file mode 100755 index 0000000000000000000000000000000000000000..af68715ab0d87ae40666596d9d877d593684f8e2 --- /dev/null +++ b/examples/wav2vec/unsupervised/kaldi_self_train/st/steps_gan/train_deltas.sh @@ -0,0 +1,175 @@ +#!/usr/bin/env bash + +# Copyright 2012 Johns Hopkins University (Author: Daniel Povey) +# Apache 2.0 + +# Begin configuration. +stage=-4 # This allows restarting after partway, when something when wrong. +config= +cmd=run.pl +scale_opts="--transition-scale=1.0 --acoustic-scale=0.1 --self-loop-scale=0.1" +realign_iters="10 20 30"; +num_iters=35 # Number of iterations of training +max_iter_inc=25 # Last iter to increase #Gauss on. +beam=10 +careful=false +retry_beam=40 +boost_silence=1.0 # Factor by which to boost silence likelihoods in alignment +power=0.25 # Exponent for number of gaussians according to occurrence counts +cluster_thresh=-1 # for build-tree control final bottom-up clustering of leaves +norm_vars=false # deprecated. Prefer --cmvn-opts "--norm-vars=true" + # use the option --cmvn-opts "--norm-means=false" +cmvn_opts= +delta_opts= +context_opts= # use"--context-width=5 --central-position=2" for quinphone +num_nonsil_states=3 +# End configuration. + +echo "$0 $@" # Print the command line for logging + +[ -f path.sh ] && . ./path.sh; +. parse_options.sh || exit 1; + +if [ $# != 6 ]; then + echo "Usage: steps/train_deltas.sh <num-leaves> <tot-gauss> <data-dir> <lang-dir> <alignment-dir> <exp-dir>" + echo "e.g.: steps/train_deltas.sh 2000 10000 data/train_si84_half data/lang exp/mono_ali exp/tri1" + echo "main options (for others, see top of script file)" + echo " --cmd (utils/run.pl|utils/queue.pl <queue opts>) # how to run jobs." + echo " --config <config-file> # config containing options" + echo " --stage <stage> # stage to do partial re-run from." + exit 1; +fi + +numleaves=$1 +totgauss=$2 +data=$3 +lang=$4 +alidir=$5 +dir=$6 + +for f in $alidir/final.mdl $alidir/ali.1.gz $data/feats.scp $lang/phones.txt; do + [ ! -f $f ] && echo "train_deltas.sh: no such file $f" && exit 1; +done + +numgauss=$numleaves +incgauss=$[($totgauss-$numgauss)/$max_iter_inc] # per-iter increment for #Gauss +oov=`cat $lang/oov.int` || exit 1; +ciphonelist=`cat $lang/phones/context_indep.csl` || exit 1; +nj=`cat $alidir/num_jobs` || exit 1; +mkdir -p $dir/log +echo $nj > $dir/num_jobs + +utils/lang/check_phones_compatible.sh $lang/phones.txt $alidir/phones.txt || exit 1; +cp $lang/phones.txt $dir || exit 1; + +sdata=$data/split$nj; +split_data.sh $data $nj || exit 1; + + +[ $(cat $alidir/cmvn_opts 2>/dev/null | wc -c) -gt 1 ] && [ -z "$cmvn_opts" ] && \ + echo "$0: warning: ignoring CMVN options from source directory $alidir" +$norm_vars && cmvn_opts="--norm-vars=true $cmvn_opts" +echo $cmvn_opts > $dir/cmvn_opts # keep track of options to CMVN. +[ ! -z $delta_opts ] && echo $delta_opts > $dir/delta_opts + +feats="ark,s,cs:apply-cmvn $cmvn_opts --utt2spk=ark:$sdata/JOB/utt2spk scp:$sdata/JOB/cmvn.scp scp:$sdata/JOB/feats.scp ark:- | add-deltas $delta_opts ark:- ark:- |" + +rm $dir/.error 2>/dev/null + +if [ $stage -le -3 ]; then + echo "$0: accumulating tree stats" + $cmd JOB=1:$nj $dir/log/acc_tree.JOB.log \ + acc-tree-stats $context_opts \ + --ci-phones=$ciphonelist $alidir/final.mdl "$feats" \ + "ark:gunzip -c $alidir/ali.JOB.gz|" $dir/JOB.treeacc || exit 1; + sum-tree-stats $dir/treeacc $dir/*.treeacc 2>$dir/log/sum_tree_acc.log || exit 1; + rm $dir/*.treeacc +fi + +if [ $stage -le -2 ]; then + echo "$0: getting questions for tree-building, via clustering" + # preparing questions, roots file... + cluster-phones --pdf-class-list=$(($num_nonsil_states / 2)) $context_opts \ + $dir/treeacc $lang/phones/sets.int \ + $dir/questions.int 2> $dir/log/questions.log || exit 1; + cat $lang/phones/extra_questions.int >> $dir/questions.int + compile-questions $context_opts $lang/topo $dir/questions.int \ + $dir/questions.qst 2>$dir/log/compile_questions.log || exit 1; + + echo "$0: building the tree" + $cmd $dir/log/build_tree.log \ + build-tree $context_opts --verbose=1 --max-leaves=$numleaves \ + --cluster-thresh=$cluster_thresh $dir/treeacc $lang/phones/roots.int \ + $dir/questions.qst $lang/topo $dir/tree || exit 1; + + $cmd $dir/log/init_model.log \ + gmm-init-model --write-occs=$dir/1.occs \ + $dir/tree $dir/treeacc $lang/topo $dir/1.mdl || exit 1; + if grep 'no stats' $dir/log/init_model.log; then + echo "** The warnings above about 'no stats' generally mean you have phones **" + echo "** (or groups of phones) in your phone set that had no corresponding data. **" + echo "** You should probably figure out whether something went wrong, **" + echo "** or whether your data just doesn't happen to have examples of those **" + echo "** phones. **" + fi + + gmm-mixup --mix-up=$numgauss $dir/1.mdl $dir/1.occs $dir/1.mdl 2>$dir/log/mixup.log || exit 1; + rm $dir/treeacc +fi + +if [ $stage -le -1 ]; then + # Convert the alignments. + echo "$0: converting alignments from $alidir to use current tree" + $cmd JOB=1:$nj $dir/log/convert.JOB.log \ + convert-ali $alidir/final.mdl $dir/1.mdl $dir/tree \ + "ark:gunzip -c $alidir/ali.JOB.gz|" "ark:|gzip -c >$dir/ali.JOB.gz" || exit 1; +fi + +if [ $stage -le 0 ]; then + echo "$0: compiling graphs of transcripts" + $cmd JOB=1:$nj $dir/log/compile_graphs.JOB.log \ + compile-train-graphs --read-disambig-syms=$lang/phones/disambig.int $dir/tree $dir/1.mdl $lang/L.fst \ + "ark:utils/sym2int.pl --map-oov $oov -f 2- $lang/words.txt < $sdata/JOB/text |" \ + "ark:|gzip -c >$dir/fsts.JOB.gz" || exit 1; +fi + +x=1 +while [ $x -lt $num_iters ]; do + echo "$0: training pass $x" + if [ $stage -le $x ]; then + if echo $realign_iters | grep -w $x >/dev/null; then + echo "$0: aligning data" + mdl="gmm-boost-silence --boost=$boost_silence `cat $lang/phones/optional_silence.csl` $dir/$x.mdl - |" + $cmd JOB=1:$nj $dir/log/align.$x.JOB.log \ + gmm-align-compiled $scale_opts --beam=$beam --retry-beam=$retry_beam --careful=$careful "$mdl" \ + "ark:gunzip -c $dir/fsts.JOB.gz|" "$feats" \ + "ark:|gzip -c >$dir/ali.JOB.gz" || exit 1; + fi + $cmd JOB=1:$nj $dir/log/acc.$x.JOB.log \ + gmm-acc-stats-ali $dir/$x.mdl "$feats" \ + "ark,s,cs:gunzip -c $dir/ali.JOB.gz|" $dir/$x.JOB.acc || exit 1; + $cmd $dir/log/update.$x.log \ + gmm-est --mix-up=$numgauss --power=$power \ + --write-occs=$dir/$[$x+1].occs $dir/$x.mdl \ + "gmm-sum-accs - $dir/$x.*.acc |" $dir/$[$x+1].mdl || exit 1; + rm $dir/$x.mdl $dir/$x.*.acc + rm $dir/$x.occs + fi + [ $x -le $max_iter_inc ] && numgauss=$[$numgauss+$incgauss]; + x=$[$x+1]; +done + +rm $dir/final.mdl $dir/final.occs 2>/dev/null +ln -s $x.mdl $dir/final.mdl +ln -s $x.occs $dir/final.occs + +steps/diagnostic/analyze_alignments.sh --cmd "$cmd" $lang $dir + +# Summarize warning messages... +utils/summarize_warnings.pl $dir/log + +steps/info/gmm_dir_info.pl $dir + +echo "$0: Done training system with delta+delta-delta features in $dir" + +exit 0 diff --git a/examples/wav2vec/unsupervised/kaldi_self_train/st/steps_gan/train_lda_mllt.sh b/examples/wav2vec/unsupervised/kaldi_self_train/st/steps_gan/train_lda_mllt.sh new file mode 100755 index 0000000000000000000000000000000000000000..9d8c319ce848e431ec47a3548156347ae3b50ced --- /dev/null +++ b/examples/wav2vec/unsupervised/kaldi_self_train/st/steps_gan/train_lda_mllt.sh @@ -0,0 +1,239 @@ +#!/usr/bin/env bash + +# Copyright 2012 Johns Hopkins University (Author: Daniel Povey) +# +# LDA+MLLT refers to the way we transform the features after computing +# the MFCCs: we splice across several frames, reduce the dimension (to 40 +# by default) using Linear Discriminant Analysis), and then later estimate, +# over multiple iterations, a diagonalizing transform known as MLLT or STC. +# See http://kaldi-asr.org/doc/transform.html for more explanation. +# +# Apache 2.0. + +# Begin configuration. +cmd=run.pl +config= +stage=-5 +scale_opts="--transition-scale=1.0 --acoustic-scale=0.1 --self-loop-scale=0.1" +realign_iters="10 20 30"; +mllt_iters="2 4 6 12"; +num_iters=35 # Number of iterations of training +max_iter_inc=25 # Last iter to increase #Gauss on. +dim=40 +beam=10 +retry_beam=40 +careful=false +boost_silence=1.0 # Factor by which to boost silence likelihoods in alignment +power=0.25 # Exponent for number of gaussians according to occurrence counts +randprune=4.0 # This is approximately the ratio by which we will speed up the + # LDA and MLLT calculations via randomized pruning. +splice_opts= +cluster_thresh=-1 # for build-tree control final bottom-up clustering of leaves +norm_vars=false # deprecated. Prefer --cmvn-opts "--norm-vars=false" +cmvn_opts= +context_opts= # use "--context-width=5 --central-position=2" for quinphone. +# End configuration. +train_tree=true # if false, don't actually train the tree. +use_lda_mat= # If supplied, use this LDA[+MLLT] matrix. +num_nonsil_states=3 + +echo "$0 $@" # Print the command line for logging + +[ -f path.sh ] && . ./path.sh +. parse_options.sh || exit 1; + +if [ $# != 6 ]; then + echo "Usage: steps/train_lda_mllt.sh [options] <#leaves> <#gauss> <data> <lang> <alignments> <dir>" + echo " e.g.: steps/train_lda_mllt.sh 2500 15000 data/train_si84 data/lang exp/tri1_ali_si84 exp/tri2b" + echo "Main options (for others, see top of script file)" + echo " --cmd (utils/run.pl|utils/queue.pl <queue opts>) # how to run jobs." + echo " --config <config-file> # config containing options" + echo " --stage <stage> # stage to do partial re-run from." + exit 1; +fi + +numleaves=$1 +totgauss=$2 +data=$3 +lang=$4 +alidir=$5 +dir=$6 + +for f in $alidir/final.mdl $alidir/ali.1.gz $data/feats.scp $lang/phones.txt; do + [ ! -f $f ] && echo "train_lda_mllt.sh: no such file $f" && exit 1; +done + +numgauss=$numleaves +incgauss=$[($totgauss-$numgauss)/$max_iter_inc] # per-iter #gauss increment +oov=`cat $lang/oov.int` || exit 1; +nj=`cat $alidir/num_jobs` || exit 1; +silphonelist=`cat $lang/phones/silence.csl` || exit 1; +ciphonelist=`cat $lang/phones/context_indep.csl` || exit 1; + +mkdir -p $dir/log + +utils/lang/check_phones_compatible.sh $lang/phones.txt $alidir/phones.txt || exit 1; +cp $lang/phones.txt $dir || exit 1; + +echo $nj >$dir/num_jobs +echo "$splice_opts" >$dir/splice_opts # keep track of frame-splicing options + # so that later stages of system building can know what they were. + + +[ $(cat $alidir/cmvn_opts 2>/dev/null | wc -c) -gt 1 ] && [ -z "$cmvn_opts" ] && \ + echo "$0: warning: ignoring CMVN options from source directory $alidir" +$norm_vars && cmvn_opts="--norm-vars=true $cmvn_opts" +echo $cmvn_opts > $dir/cmvn_opts # keep track of options to CMVN. + +sdata=$data/split$nj; +split_data.sh $data $nj || exit 1; + +splicedfeats="ark,s,cs:apply-cmvn $cmvn_opts --utt2spk=ark:$sdata/JOB/utt2spk scp:$sdata/JOB/cmvn.scp scp:$sdata/JOB/feats.scp ark:- | splice-feats $splice_opts ark:- ark:- |" +# Note: $feats gets overwritten later in the script. +feats="$splicedfeats transform-feats $dir/0.mat ark:- ark:- |" + + + +if [ $stage -le -5 ]; then + if [ -z "$use_lda_mat" ]; then + echo "$0: Accumulating LDA statistics." + rm $dir/lda.*.acc 2>/dev/null + $cmd JOB=1:$nj $dir/log/lda_acc.JOB.log \ + ali-to-post "ark:gunzip -c $alidir/ali.JOB.gz|" ark:- \| \ + weight-silence-post 0.0 $silphonelist $alidir/final.mdl ark:- ark:- \| \ + acc-lda --rand-prune=$randprune $alidir/final.mdl "$splicedfeats" ark,s,cs:- \ + $dir/lda.JOB.acc || exit 1; + est-lda --write-full-matrix=$dir/full.mat --dim=$dim $dir/0.mat $dir/lda.*.acc \ + 2>$dir/log/lda_est.log || exit 1; + rm $dir/lda.*.acc + else + echo "$0: Using supplied LDA matrix $use_lda_mat" + cp $use_lda_mat $dir/0.mat || exit 1; + [ ! -z "$mllt_iters" ] && \ + echo "$0: Warning: using supplied LDA matrix $use_lda_mat but we will do MLLT," && \ + echo " which you might not want; to disable MLLT, specify --mllt-iters ''" && \ + sleep 5 + fi +fi + +cur_lda_iter=0 + +if [ $stage -le -4 ] && $train_tree; then + echo "$0: Accumulating tree stats" + $cmd JOB=1:$nj $dir/log/acc_tree.JOB.log \ + acc-tree-stats $context_opts \ + --ci-phones=$ciphonelist $alidir/final.mdl "$feats" \ + "ark:gunzip -c $alidir/ali.JOB.gz|" $dir/JOB.treeacc || exit 1; + [ `ls $dir/*.treeacc | wc -w` -ne "$nj" ] && echo "$0: Wrong #tree-accs" && exit 1; + $cmd $dir/log/sum_tree_acc.log \ + sum-tree-stats $dir/treeacc $dir/*.treeacc || exit 1; + rm $dir/*.treeacc +fi + + +if [ $stage -le -3 ] && $train_tree; then + echo "$0: Getting questions for tree clustering." + # preparing questions, roots file... + cluster-phones --pdf-class-list=$(($num_nonsil_states / 2)) $context_opts $dir/treeacc $lang/phones/sets.int \ + $dir/questions.int 2> $dir/log/questions.log || exit 1; + cat $lang/phones/extra_questions.int >> $dir/questions.int + compile-questions $context_opts $lang/topo $dir/questions.int \ + $dir/questions.qst 2>$dir/log/compile_questions.log || exit 1; + + echo "$0: Building the tree" + $cmd $dir/log/build_tree.log \ + build-tree $context_opts --verbose=1 --max-leaves=$numleaves \ + --cluster-thresh=$cluster_thresh $dir/treeacc $lang/phones/roots.int \ + $dir/questions.qst $lang/topo $dir/tree || exit 1; +fi + +if [ $stage -le -2 ]; then + echo "$0: Initializing the model" + if $train_tree; then + gmm-init-model --write-occs=$dir/1.occs \ + $dir/tree $dir/treeacc $lang/topo $dir/1.mdl 2> $dir/log/init_model.log || exit 1; + grep 'no stats' $dir/log/init_model.log && echo "This is a bad warning."; + rm $dir/treeacc + else + cp $alidir/tree $dir/ || exit 1; + $cmd JOB=1 $dir/log/init_model.log \ + gmm-init-model-flat $dir/tree $lang/topo $dir/1.mdl \ + "$feats subset-feats ark:- ark:-|" || exit 1; + fi +fi + + +if [ $stage -le -1 ]; then + # Convert the alignments. + echo "$0: Converting alignments from $alidir to use current tree" + $cmd JOB=1:$nj $dir/log/convert.JOB.log \ + convert-ali $alidir/final.mdl $dir/1.mdl $dir/tree \ + "ark:gunzip -c $alidir/ali.JOB.gz|" "ark:|gzip -c >$dir/ali.JOB.gz" || exit 1; +fi + +if [ $stage -le 0 ] && [ "$realign_iters" != "" ]; then + echo "$0: Compiling graphs of transcripts" + $cmd JOB=1:$nj $dir/log/compile_graphs.JOB.log \ + compile-train-graphs --read-disambig-syms=$lang/phones/disambig.int $dir/tree $dir/1.mdl $lang/L.fst \ + "ark:utils/sym2int.pl --map-oov $oov -f 2- $lang/words.txt < $data/split$nj/JOB/text |" \ + "ark:|gzip -c >$dir/fsts.JOB.gz" || exit 1; +fi + + +x=1 +while [ $x -lt $num_iters ]; do + echo Training pass $x + if echo $realign_iters | grep -w $x >/dev/null && [ $stage -le $x ]; then + echo Aligning data + mdl="gmm-boost-silence --boost=$boost_silence `cat $lang/phones/optional_silence.csl` $dir/$x.mdl - |" + $cmd JOB=1:$nj $dir/log/align.$x.JOB.log \ + gmm-align-compiled $scale_opts --beam=$beam --retry-beam=$retry_beam --careful=$careful "$mdl" \ + "ark:gunzip -c $dir/fsts.JOB.gz|" "$feats" \ + "ark:|gzip -c >$dir/ali.JOB.gz" || exit 1; + fi + if echo $mllt_iters | grep -w $x >/dev/null; then + if [ $stage -le $x ]; then + echo "$0: Estimating MLLT" + $cmd JOB=1:$nj $dir/log/macc.$x.JOB.log \ + ali-to-post "ark:gunzip -c $dir/ali.JOB.gz|" ark:- \| \ + weight-silence-post 0.0 $silphonelist $dir/$x.mdl ark:- ark:- \| \ + gmm-acc-mllt --rand-prune=$randprune $dir/$x.mdl "$feats" ark:- $dir/$x.JOB.macc \ + || exit 1; + est-mllt $dir/$x.mat.new $dir/$x.*.macc 2> $dir/log/mupdate.$x.log || exit 1; + gmm-transform-means $dir/$x.mat.new $dir/$x.mdl $dir/$x.mdl \ + 2> $dir/log/transform_means.$x.log || exit 1; + compose-transforms --print-args=false $dir/$x.mat.new $dir/$cur_lda_iter.mat $dir/$x.mat || exit 1; + rm $dir/$x.*.macc + fi + feats="$splicedfeats transform-feats $dir/$x.mat ark:- ark:- |" + cur_lda_iter=$x + fi + + if [ $stage -le $x ]; then + $cmd JOB=1:$nj $dir/log/acc.$x.JOB.log \ + gmm-acc-stats-ali $dir/$x.mdl "$feats" \ + "ark,s,cs:gunzip -c $dir/ali.JOB.gz|" $dir/$x.JOB.acc || exit 1; + $cmd $dir/log/update.$x.log \ + gmm-est --write-occs=$dir/$[$x+1].occs --mix-up=$numgauss --power=$power \ + $dir/$x.mdl "gmm-sum-accs - $dir/$x.*.acc |" $dir/$[$x+1].mdl || exit 1; + rm $dir/$x.mdl $dir/$x.*.acc $dir/$x.occs + fi + [ $x -le $max_iter_inc ] && numgauss=$[$numgauss+$incgauss]; + x=$[$x+1]; +done + +rm $dir/final.{mdl,mat,occs} 2>/dev/null +ln -s $x.mdl $dir/final.mdl +ln -s $x.occs $dir/final.occs +ln -s $cur_lda_iter.mat $dir/final.mat + +steps/diagnostic/analyze_alignments.sh --cmd "$cmd" $lang $dir + +# Summarize warning messages... +utils/summarize_warnings.pl $dir/log + +steps/info/gmm_dir_info.pl $dir + +echo "$0: Done training system with LDA+MLLT features in $dir" + +exit 0 diff --git a/examples/wav2vec/unsupervised/kaldi_self_train/st/steps_gan/train_sat.sh b/examples/wav2vec/unsupervised/kaldi_self_train/st/steps_gan/train_sat.sh new file mode 100755 index 0000000000000000000000000000000000000000..f75afafb1c4ad04ee71ab8541064ab0477430616 --- /dev/null +++ b/examples/wav2vec/unsupervised/kaldi_self_train/st/steps_gan/train_sat.sh @@ -0,0 +1,281 @@ +#!/usr/bin/env bash +# Copyright 2012 Johns Hopkins University (Author: Daniel Povey). Apache 2.0. + + +# This does Speaker Adapted Training (SAT), i.e. train on +# fMLLR-adapted features. It can be done on top of either LDA+MLLT, or +# delta and delta-delta features. If there are no transforms supplied +# in the alignment directory, it will estimate transforms itself before +# building the tree (and in any case, it estimates transforms a number +# of times during training). + + +# Begin configuration section. +stage=-5 +exit_stage=-100 # you can use this to require it to exit at the + # beginning of a specific stage. Not all values are + # supported. +fmllr_update_type=full +cmd=run.pl +scale_opts="--transition-scale=1.0 --acoustic-scale=0.1 --self-loop-scale=0.1" +beam=10 +retry_beam=40 +careful=false +boost_silence=1.0 # Factor by which to boost silence likelihoods in alignment +context_opts= # e.g. set this to "--context-width 5 --central-position 2" for quinphone. +realign_iters="10 20 30"; +fmllr_iters="2 4 6 12"; +silence_weight=0.0 # Weight on silence in fMLLR estimation. +num_iters=35 # Number of iterations of training +max_iter_inc=25 # Last iter to increase #Gauss on. +power=0.2 # Exponent for number of gaussians according to occurrence counts +cluster_thresh=-1 # for build-tree control final bottom-up clustering of leaves +phone_map= +train_tree=true +tree_stats_opts= +cluster_phones_opts= +compile_questions_opts= +# End configuration section. +num_nonsil_states=3 + +echo "$0 $@" # Print the command line for logging + +[ -f path.sh ] && . ./path.sh +. parse_options.sh || exit 1; + +if [ $# != 6 ]; then + echo "Usage: steps/train_sat.sh <#leaves> <#gauss> <data> <lang> <ali-dir> <exp-dir>" + echo " e.g.: steps/train_sat.sh 2500 15000 data/train_si84 data/lang exp/tri2b_ali_si84 exp/tri3b" + echo "Main options (for others, see top of script file)" + echo " --cmd (utils/run.pl|utils/queue.pl <queue opts>) # how to run jobs." + echo " --config <config-file> # config containing options" + echo " --stage <stage> # stage to do partial re-run from." + exit 1; +fi + +numleaves=$1 +totgauss=$2 +data=$3 +lang=$4 +alidir=$5 +dir=$6 + +for f in $data/feats.scp $lang/phones.txt $alidir/final.mdl $alidir/ali.1.gz; do + [ ! -f $f ] && echo "train_sat.sh: no such file $f" && exit 1; +done + +numgauss=$numleaves +incgauss=$[($totgauss-$numgauss)/$max_iter_inc] # per-iter #gauss increment +oov=`cat $lang/oov.int` +nj=`cat $alidir/num_jobs` || exit 1; +silphonelist=`cat $lang/phones/silence.csl` +ciphonelist=`cat $lang/phones/context_indep.csl` || exit 1; +sdata=$data/split$nj; +splice_opts=`cat $alidir/splice_opts 2>/dev/null` # frame-splicing options. +cmvn_opts=`cat $alidir/cmvn_opts 2>/dev/null` +delta_opts=`cat $alidir/delta_opts 2>/dev/null` +phone_map_opt= +[ ! -z "$phone_map" ] && phone_map_opt="--phone-map='$phone_map'" + +mkdir -p $dir/log +cp $alidir/splice_opts $dir 2>/dev/null # frame-splicing options. +cp $alidir/cmvn_opts $dir 2>/dev/null # cmn/cmvn option. +cp $alidir/delta_opts $dir 2>/dev/null # delta option. + +utils/lang/check_phones_compatible.sh $lang/phones.txt $alidir/phones.txt || exit 1; +cp $lang/phones.txt $dir || exit 1; + +echo $nj >$dir/num_jobs +[[ -d $sdata && $data/feats.scp -ot $sdata ]] || split_data.sh $data $nj || exit 1; + +# Set up features. + +if [ -f $alidir/final.mat ]; then feat_type=lda; else feat_type=delta; fi +echo "$0: feature type is $feat_type" + +## Set up speaker-independent features. +case $feat_type in + delta) sifeats="ark,s,cs:apply-cmvn $cmvn_opts --utt2spk=ark:$sdata/JOB/utt2spk scp:$sdata/JOB/cmvn.scp scp:$sdata/JOB/feats.scp ark:- | add-deltas $delta_opts ark:- ark:- |";; + lda) sifeats="ark,s,cs:apply-cmvn $cmvn_opts --utt2spk=ark:$sdata/JOB/utt2spk scp:$sdata/JOB/cmvn.scp scp:$sdata/JOB/feats.scp ark:- | splice-feats $splice_opts ark:- ark:- | transform-feats $alidir/final.mat ark:- ark:- |" + cp $alidir/final.mat $dir + cp $alidir/full.mat $dir 2>/dev/null + ;; + *) echo "$0: invalid feature type $feat_type" && exit 1; +esac + +## Get initial fMLLR transforms (possibly from alignment dir) +if [ -f $alidir/trans.1 ]; then + echo "$0: Using transforms from $alidir" + feats="$sifeats transform-feats --utt2spk=ark:$sdata/JOB/utt2spk ark,s,cs:$alidir/trans.JOB ark:- ark:- |" + cur_trans_dir=$alidir +else + if [ $stage -le -5 ]; then + echo "$0: obtaining initial fMLLR transforms since not present in $alidir" + # The next line is necessary because of $silphonelist otherwise being incorrect; would require + # old $lang dir which would require another option. Not needed anyway. + [ ! -z "$phone_map" ] && \ + echo "$0: error: you must provide transforms if you use the --phone-map option." && exit 1; + $cmd JOB=1:$nj $dir/log/fmllr.0.JOB.log \ + ali-to-post "ark:gunzip -c $alidir/ali.JOB.gz|" ark:- \| \ + weight-silence-post $silence_weight $silphonelist $alidir/final.mdl ark:- ark:- \| \ + gmm-est-fmllr --fmllr-update-type=$fmllr_update_type \ + --spk2utt=ark:$sdata/JOB/spk2utt $alidir/final.mdl "$sifeats" \ + ark:- ark:$dir/trans.JOB || exit 1; + fi + feats="$sifeats transform-feats --utt2spk=ark:$sdata/JOB/utt2spk ark,s,cs:$dir/trans.JOB ark:- ark:- |" + cur_trans_dir=$dir +fi + +if [ $stage -le -4 ] && $train_tree; then + # Get tree stats. + echo "$0: Accumulating tree stats" + $cmd JOB=1:$nj $dir/log/acc_tree.JOB.log \ + acc-tree-stats $context_opts $tree_stats_opts $phone_map_opt --ci-phones=$ciphonelist $alidir/final.mdl "$feats" \ + "ark:gunzip -c $alidir/ali.JOB.gz|" $dir/JOB.treeacc || exit 1; + [ "`ls $dir/*.treeacc | wc -w`" -ne "$nj" ] && echo "$0: Wrong #tree-accs" && exit 1; + $cmd $dir/log/sum_tree_acc.log \ + sum-tree-stats $dir/treeacc $dir/*.treeacc || exit 1; + rm $dir/*.treeacc +fi + +if [ $stage -le -3 ] && $train_tree; then + echo "$0: Getting questions for tree clustering." + # preparing questions, roots file... + cluster-phones --pdf-class-list=$(($num_nonsil_states / 2)) \ + $cluster_phones_opts $context_opts \ + $dir/treeacc $lang/phones/sets.int $dir/questions.int 2>$dir/log/questions.log || exit 1; + cat $lang/phones/extra_questions.int >> $dir/questions.int + compile-questions $context_opts $compile_questions_opts $lang/topo $dir/questions.int $dir/questions.qst 2>$dir/log/compile_questions.log || exit 1; + + echo "$0: Building the tree" + $cmd $dir/log/build_tree.log \ + build-tree $context_opts --verbose=1 --max-leaves=$numleaves \ + --cluster-thresh=$cluster_thresh $dir/treeacc $lang/phones/roots.int \ + $dir/questions.qst $lang/topo $dir/tree || exit 1; +fi + +if [ $stage -le -2 ]; then + echo "$0: Initializing the model" + if $train_tree; then + gmm-init-model --write-occs=$dir/1.occs \ + $dir/tree $dir/treeacc $lang/topo $dir/1.mdl 2> $dir/log/init_model.log || exit 1; + grep 'no stats' $dir/log/init_model.log && echo "This is a bad warning."; + rm $dir/treeacc + else + cp $alidir/tree $dir/ || exit 1; + $cmd JOB=1 $dir/log/init_model.log \ + gmm-init-model-flat $dir/tree $lang/topo $dir/1.mdl \ + "$feats subset-feats ark:- ark:-|" || exit 1; + fi +fi + +if [ $stage -le -1 ]; then + # Convert the alignments. + echo "$0: Converting alignments from $alidir to use current tree" + $cmd JOB=1:$nj $dir/log/convert.JOB.log \ + convert-ali $phone_map_opt $alidir/final.mdl $dir/1.mdl $dir/tree \ + "ark:gunzip -c $alidir/ali.JOB.gz|" "ark:|gzip -c >$dir/ali.JOB.gz" || exit 1; +fi + +[ "$exit_stage" -eq 0 ] && echo "$0: Exiting early: --exit-stage $exit_stage" && exit 0; + +if [ $stage -le 0 ] && [ "$realign_iters" != "" ]; then + echo "$0: Compiling graphs of transcripts" + $cmd JOB=1:$nj $dir/log/compile_graphs.JOB.log \ + compile-train-graphs --read-disambig-syms=$lang/phones/disambig.int $dir/tree $dir/1.mdl $lang/L.fst \ + "ark:utils/sym2int.pl --map-oov $oov -f 2- $lang/words.txt < $sdata/JOB/text |" \ + "ark:|gzip -c >$dir/fsts.JOB.gz" || exit 1; +fi + +x=1 +while [ $x -lt $num_iters ]; do + echo Pass $x + if echo $realign_iters | grep -w $x >/dev/null && [ $stage -le $x ]; then + echo Aligning data + mdl="gmm-boost-silence --boost=$boost_silence `cat $lang/phones/optional_silence.csl` $dir/$x.mdl - |" + $cmd JOB=1:$nj $dir/log/align.$x.JOB.log \ + gmm-align-compiled $scale_opts --beam=$beam --retry-beam=$retry_beam --careful=$careful "$mdl" \ + "ark:gunzip -c $dir/fsts.JOB.gz|" "$feats" \ + "ark:|gzip -c >$dir/ali.JOB.gz" || exit 1; + fi + + if echo $fmllr_iters | grep -w $x >/dev/null; then + if [ $stage -le $x ]; then + echo Estimating fMLLR transforms + # We estimate a transform that's additional to the previous transform; + # we'll compose them. + $cmd JOB=1:$nj $dir/log/fmllr.$x.JOB.log \ + ali-to-post "ark:gunzip -c $dir/ali.JOB.gz|" ark:- \| \ + weight-silence-post $silence_weight $silphonelist $dir/$x.mdl ark:- ark:- \| \ + gmm-est-fmllr --fmllr-update-type=$fmllr_update_type \ + --spk2utt=ark:$sdata/JOB/spk2utt $dir/$x.mdl \ + "$feats" ark:- ark:$dir/tmp_trans.JOB || exit 1; + for n in `seq $nj`; do + ! ( compose-transforms --b-is-affine=true \ + ark:$dir/tmp_trans.$n ark:$cur_trans_dir/trans.$n ark:$dir/composed_trans.$n \ + && mv $dir/composed_trans.$n $dir/trans.$n && \ + rm $dir/tmp_trans.$n ) 2>$dir/log/compose_transforms.$x.log \ + && echo "$0: Error composing transforms" && exit 1; + done + fi + feats="$sifeats transform-feats --utt2spk=ark:$sdata/JOB/utt2spk ark:$dir/trans.JOB ark:- ark:- |" + cur_trans_dir=$dir + fi + + if [ $stage -le $x ]; then + $cmd JOB=1:$nj $dir/log/acc.$x.JOB.log \ + gmm-acc-stats-ali $dir/$x.mdl "$feats" \ + "ark,s,cs:gunzip -c $dir/ali.JOB.gz|" $dir/$x.JOB.acc || exit 1; + [ `ls $dir/$x.*.acc | wc -w` -ne "$nj" ] && echo "$0: Wrong #accs" && exit 1; + $cmd $dir/log/update.$x.log \ + gmm-est --power=$power --write-occs=$dir/$[$x+1].occs --mix-up=$numgauss $dir/$x.mdl \ + "gmm-sum-accs - $dir/$x.*.acc |" $dir/$[$x+1].mdl || exit 1; + rm $dir/$x.mdl $dir/$x.*.acc + rm $dir/$x.occs + fi + [ $x -le $max_iter_inc ] && numgauss=$[$numgauss+$incgauss]; + x=$[$x+1]; +done + + +if [ $stage -le $x ]; then + # Accumulate stats for "alignment model"-- this model is + # computed with the speaker-independent features, but matches Gaussian-for-Gaussian + # with the final speaker-adapted model. + $cmd JOB=1:$nj $dir/log/acc_alimdl.JOB.log \ + ali-to-post "ark:gunzip -c $dir/ali.JOB.gz|" ark:- \| \ + gmm-acc-stats-twofeats $dir/$x.mdl "$feats" "$sifeats" \ + ark,s,cs:- $dir/$x.JOB.acc || exit 1; + [ `ls $dir/$x.*.acc | wc -w` -ne "$nj" ] && echo "$0: Wrong #accs" && exit 1; + # Update model. + $cmd $dir/log/est_alimdl.log \ + gmm-est --power=$power --remove-low-count-gaussians=false $dir/$x.mdl \ + "gmm-sum-accs - $dir/$x.*.acc|" $dir/$x.alimdl || exit 1; + rm $dir/$x.*.acc +fi + +rm $dir/final.{mdl,alimdl,occs} 2>/dev/null +ln -s $x.mdl $dir/final.mdl +ln -s $x.occs $dir/final.occs +ln -s $x.alimdl $dir/final.alimdl + + +steps/diagnostic/analyze_alignments.sh --cmd "$cmd" $lang $dir + +utils/summarize_warnings.pl $dir/log +( + echo "$0: Likelihood evolution:" + for x in `seq $[$num_iters-1]`; do + tail -n 30 $dir/log/acc.$x.*.log | awk '/Overall avg like/{l += $(NF-3)*$(NF-1); t += $(NF-1); } + /Overall average logdet/{d += $(NF-3)*$(NF-1); t2 += $(NF-1);} + END{ d /= t2; l /= t; printf("%s ", d+l); } ' + done + echo +) | tee $dir/log/summary.log + + +steps/info/gmm_dir_info.pl $dir + +echo "$0: done training SAT system in $dir" + +exit 0 diff --git a/examples/wav2vec/unsupervised/kaldi_self_train/st/train.sh b/examples/wav2vec/unsupervised/kaldi_self_train/st/train.sh new file mode 100644 index 0000000000000000000000000000000000000000..f3a3d3fc7cc98a38d8e9d523a0b43c0c8ea51bf9 --- /dev/null +++ b/examples/wav2vec/unsupervised/kaldi_self_train/st/train.sh @@ -0,0 +1,43 @@ +#!/bin/bash + +set -eu + +w2v_dir= # contains features `{train,valid}.{npy,lengths}`, real transcripts `{train,valid}.${label}`, and dict `dict.${label}.txt` +lab_dir= # contains pseudo labels `{train,valid}.txt` +out_dir= # output root +arpa_lm= # phone LM +arpa_lm_bin= # (binary) phone LM for KenLM, used in unsupervised selection + +label=phnc +train_name="train" +valid_name="valid" +data_dir=${out_dir}/data + +mkdir -p ${out_dir}/exp +local/prepare_lang.sh $w2v_dir/dict.${label}.txt $data_dir +local/prepare_lm.sh $arpa_lm $data_dir + +for x in $train_name $valid_name; do + x_gt=${x}_gt + + # prepare pseudo data + python local/prepare_data_from_w2v.py $w2v_dir $data_dir $x + steps/compute_cmvn_stats.sh $data_dir/$x $out_dir/exp/make_feat/$x $out_dir/feats/$x + python local/copy_aligned_text.py < $lab_dir/$x.txt > $data_dir/$x/text + + # prepare ground truth data + mkdir $data_dir/$x_gt + cp $data_dir/$x/{feats.scp,cmvn.scp,utt2spk,spk2utt} $data_dir/$x_gt/ + python local/copy_aligned_text.py < $w2v_dir/$x.$label > $data_dir/$x_gt/text +done + +local/train_subset_lgbeam.sh \ + --out_root ${out_dir} --out_name exp --train $train_name --valid $valid_name \ + --mono_size 2000 --tri1_size 5000 --tri2b_size -1 --tri3b_size -1 \ + --stage 1 --max_stage 3 $data_dir $data_dir/lang $data_dir/lang_test + +local/unsup_select_decode.sh \ + --split $valid_name --kenlm_path $arpa_lm_bin \ + --ref_txt $data_dir/${valid_name}_gt/text \ + --psd_txt $data_dir/${valid_name}/text \ + $out_dir/exp diff --git a/examples/wav2vec/unsupervised/kaldi_self_train/st/utils b/examples/wav2vec/unsupervised/kaldi_self_train/st/utils new file mode 120000 index 0000000000000000000000000000000000000000..b240885218f9eaa37a81f7ca797be77746aeb44c --- /dev/null +++ b/examples/wav2vec/unsupervised/kaldi_self_train/st/utils @@ -0,0 +1 @@ +../../wsj/s5/utils \ No newline at end of file diff --git a/examples/wav2vec/unsupervised/models/__init__.py b/examples/wav2vec/unsupervised/models/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..3e3039b7081a9e3228c8abefb6391a75b4864439 --- /dev/null +++ b/examples/wav2vec/unsupervised/models/__init__.py @@ -0,0 +1,11 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .wav2vec_u import Wav2vec_U + + +__all__ = [ + "Wav2vec_U", +] diff --git a/examples/wav2vec/unsupervised/models/wav2vec_u.py b/examples/wav2vec/unsupervised/models/wav2vec_u.py new file mode 100644 index 0000000000000000000000000000000000000000..27792ebda842057e33fed3dc53dd9d8a594d0483 --- /dev/null +++ b/examples/wav2vec/unsupervised/models/wav2vec_u.py @@ -0,0 +1,637 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from dataclasses import dataclass +from enum import Enum, auto +import math +import numpy as np +from typing import Tuple, List, Optional, Dict + +import torch +import torch.nn as nn +import torch.nn.functional as F +from torch import autograd + +from fairseq import checkpoint_utils, utils +from fairseq.dataclass import FairseqDataclass +from fairseq.models import BaseFairseqModel, register_model +from fairseq.modules import ( + SamePad, + TransposeLast, +) + + +class SegmentationType(Enum): + NONE = auto() + RANDOM = auto() + UNIFORM_RANDOM = auto() + UNIFORM_RANDOM_JOIN = auto() + JOIN = auto() + + +@dataclass +class SegmentationConfig(FairseqDataclass): + type: SegmentationType = SegmentationType.NONE + subsample_rate: float = 0.25 + mean_pool: bool = True + mean_pool_join: bool = False + remove_zeros: bool = False + + +@dataclass +class Wav2vec_UConfig(FairseqDataclass): + + discriminator_kernel: int = 3 + discriminator_dilation: int = 1 + discriminator_dim: int = 256 + discriminator_causal: bool = True + discriminator_linear_emb: bool = False + discriminator_depth: int = 1 + discriminator_max_pool: bool = False + discriminator_act_after_linear: bool = False + discriminator_dropout: float = 0.0 + discriminator_spectral_norm: bool = False + discriminator_weight_norm: bool = False + + generator_kernel: int = 4 + generator_dilation: int = 1 + generator_stride: int = 1 + generator_bias: bool = False + generator_dropout: float = 0.0 + + blank_weight: float = 0 + blank_mode: str = "add" + blank_is_sil: bool = False + no_softmax: bool = False + + smoothness_weight: float = 0.0 + smoothing: float = 0.0 + smoothing_one_sided: bool = False + gradient_penalty: float = 0.0 + probabilistic_grad_penalty_slicing: bool = False + code_penalty: float = 0.0 + gumbel: bool = False + hard_gumbel: bool = True + temp: Tuple[float, float, float] = (2, 0.1, 0.99995) + input_dim: int = 128 + + segmentation: SegmentationConfig = SegmentationConfig() + + +class Segmenter(nn.Module): + cfg: SegmentationConfig + + def __init__(self, cfg: SegmentationConfig): + super().__init__() + self.cfg = cfg + self.subsample_rate = cfg.subsample_rate + + def pre_segment(self, dense_x, dense_padding_mask): + return dense_x, dense_padding_mask + + def logit_segment(self, logits, padding_mask): + return logits, padding_mask + + +class RandomSegmenter(Segmenter): + def pre_segment(self, dense_x, dense_padding_mask): + target_num = math.ceil(dense_x.size(1) * self.subsample_rate) + ones = torch.ones(dense_x.shape[:-1], device=dense_x.device) + indices, _ = ones.multinomial(target_num).sort(dim=-1) + indices_ld = indices.unsqueeze(-1).expand(-1, -1, dense_x.size(-1)) + dense_x = dense_x.gather(1, indices_ld) + dense_padding_mask = dense_padding_mask.gather(1, index=indices) + return dense_x, dense_padding_mask + + +class UniformRandomSegmenter(Segmenter): + def pre_segment(self, dense_x, dense_padding_mask): + bsz, tsz, fsz = dense_x.shape + + target_num = math.ceil(tsz * self.subsample_rate) + + rem = tsz % target_num + + if rem > 0: + dense_x = F.pad(dense_x, [0, 0, 0, target_num - rem]) + dense_padding_mask = F.pad( + dense_padding_mask, [0, target_num - rem], value=True + ) + + dense_x = dense_x.view(bsz, target_num, -1, fsz) + dense_padding_mask = dense_padding_mask.view(bsz, target_num, -1) + + if self.cfg.mean_pool: + dense_x = dense_x.mean(dim=-2) + dense_padding_mask = dense_padding_mask.all(dim=-1) + else: + ones = torch.ones((bsz, dense_x.size(2)), device=dense_x.device) + indices = ones.multinomial(1) + indices = indices.unsqueeze(-1).expand(-1, target_num, -1) + indices_ld = indices.unsqueeze(-1).expand(-1, -1, -1, fsz) + dense_x = dense_x.gather(2, indices_ld).reshape(bsz, -1, fsz) + dense_padding_mask = dense_padding_mask.gather(2, index=indices).reshape( + bsz, -1 + ) + return dense_x, dense_padding_mask + + +class JoinSegmenter(Segmenter): + def logit_segment(self, logits, padding_mask): + preds = logits.argmax(dim=-1) + + if padding_mask.any(): + preds[padding_mask] = -1 # mark pad + uniques = [] + + bsz, tsz, csz = logits.shape + + for p in preds: + uniques.append( + p.cpu().unique_consecutive(return_inverse=True, return_counts=True) + ) + + new_tsz = max(u[0].numel() for u in uniques) + new_logits = logits.new_zeros(bsz, new_tsz, csz) + new_pad = padding_mask.new_zeros(bsz, new_tsz) + + for b in range(bsz): + u, idx, c = uniques[b] + keep = u != -1 + + if self.cfg.remove_zeros: + keep.logical_and_(u != 0) + + if self.training and not self.cfg.mean_pool_join: + u[0] = 0 + u[1:] = c.cumsum(0)[:-1] + m = c > 1 + r = torch.rand(m.sum()) + o = (c[m] * r).long() + u[m] += o + new_logits[b, : u.numel()] = logits[b, u] + else: + new_logits[b].index_add_( + dim=0, index=idx.to(new_logits.device), source=logits[b] + ) + new_logits[b, : c.numel()] /= c.unsqueeze(-1).to(new_logits.device) + + new_sz = keep.sum() + if not keep.all(): + kept_logits = new_logits[b, : c.numel()][keep] + new_logits[b, :new_sz] = kept_logits + + if new_sz < new_tsz: + pad = new_tsz - new_sz + new_logits[b, -pad:] = 0 + new_pad[b, -pad:] = True + + return new_logits, new_pad + + +class UniformRandomJoinSegmenter(UniformRandomSegmenter, JoinSegmenter): + pass + + +SEGMENT_FACTORY = { + SegmentationType.NONE: Segmenter, + SegmentationType.RANDOM: RandomSegmenter, + SegmentationType.UNIFORM_RANDOM: UniformRandomSegmenter, + SegmentationType.UNIFORM_RANDOM_JOIN: UniformRandomJoinSegmenter, + SegmentationType.JOIN: JoinSegmenter, +} + + +class Discriminator(nn.Module): + def __init__(self, dim, cfg: Wav2vec_UConfig): + super().__init__() + + inner_dim = cfg.discriminator_dim + kernel = cfg.discriminator_kernel + dilation = cfg.discriminator_dilation + self.max_pool = cfg.discriminator_max_pool + + if cfg.discriminator_causal: + padding = kernel - 1 + else: + padding = kernel // 2 + + def make_conv(in_d, out_d, k, p=0, has_dilation=True): + conv = nn.Conv1d( + in_d, + out_d, + kernel_size=k, + padding=p, + dilation=dilation if has_dilation else 1, + ) + if cfg.discriminator_spectral_norm: + conv = nn.utils.spectral_norm(conv) + elif cfg.discriminator_weight_norm: + conv = nn.utils.weight_norm(conv) + return conv + + inner_net = [ + nn.Sequential( + make_conv(inner_dim, inner_dim, kernel, padding), + SamePad(kernel_size=kernel, causal=cfg.discriminator_causal), + nn.Dropout(cfg.discriminator_dropout), + nn.GELU(), + ) + for _ in range(cfg.discriminator_depth - 1) + ] + [ + make_conv(inner_dim, 1, kernel, padding, has_dilation=False), + SamePad(kernel_size=kernel, causal=cfg.discriminator_causal), + ] + + if cfg.discriminator_linear_emb: + emb_net = [make_conv(dim, inner_dim, 1)] + else: + emb_net = [ + make_conv(dim, inner_dim, kernel, padding), + SamePad(kernel_size=kernel, causal=cfg.discriminator_causal), + ] + + if cfg.discriminator_act_after_linear: + emb_net.append(nn.GELU()) + + self.net = nn.Sequential( + *emb_net, + nn.Dropout(cfg.discriminator_dropout), + *inner_net, + ) + + def forward(self, x, padding_mask): + x = x.transpose(1, 2) # BTC -> BCT + x = self.net(x) + x = x.transpose(1, 2) + x_sz = x.size(1) + if padding_mask is not None and padding_mask.any() and padding_mask.dim() > 1: + padding_mask = padding_mask[:, : x.size(1)] + x[padding_mask] = float("-inf") if self.max_pool else 0 + x_sz = x_sz - padding_mask.sum(dim=-1) + x = x.squeeze(-1) + if self.max_pool: + x, _ = x.max(dim=-1) + else: + x = x.sum(dim=-1) + x = x / x_sz + return x + + +class Generator(nn.Module): + def __init__(self, input_dim, output_dim, cfg: Wav2vec_UConfig): + super().__init__() + + self.cfg = cfg + self.output_dim = output_dim + self.stride = cfg.generator_stride + self.dropout = nn.Dropout(cfg.generator_dropout) + + padding = cfg.generator_kernel // 2 + self.proj = nn.Sequential( + TransposeLast(), + nn.Conv1d( + input_dim, + output_dim, + kernel_size=cfg.generator_kernel, + stride=cfg.generator_stride, + dilation=cfg.generator_dilation, + padding=padding, + bias=cfg.generator_bias, + ), + TransposeLast(), + ) + + def forward(self, dense_x, tokens, dense_padding_mask): + dense_x = self.dropout(dense_x) + + dense_x = self.proj(dense_x) + if self.stride > 1: + dense_padding_mask = dense_padding_mask[:, :: self.stride] + + if dense_padding_mask.size(1) != dense_x.size(1): + new_padding = dense_padding_mask.new_zeros(dense_x.shape[:-1]) + diff = new_padding.size(1) - dense_padding_mask.size(1) + assert ( + diff > 0 + ), f"{new_padding.shape}, {dense_padding_mask.shape}, {dense_x.shape}, {diff}" + if diff > 0: + new_padding[:, diff:] = dense_padding_mask + else: + assert diff < 0 + new_padding = dense_padding_mask[:, :diff] + + dense_padding_mask = new_padding + + result = {} + + token_x = None + if tokens is not None: + token_x = dense_x.new_zeros(tokens.numel(), self.output_dim) + token_x.scatter_(1, tokens.view(-1, 1).long(), 1) + token_x = token_x.view(tokens.shape + (self.output_dim,)) + + result["dense_x"] = dense_x + result["token_x"] = token_x + result["dense_padding_mask"] = dense_padding_mask + + return result + + +@register_model("wav2vec_u", dataclass=Wav2vec_UConfig) +class Wav2vec_U(BaseFairseqModel): + def calc_gradient_penalty(self, real_data, fake_data): + + b_size = min(real_data.size(0), fake_data.size(0)) + t_size = min(real_data.size(1), fake_data.size(1)) + + if self.cfg.probabilistic_grad_penalty_slicing: + + def get_slice(data, dim, target_size): + + size = data.size(dim) + diff = size - target_size + if diff <= 0: + return data + + start = np.random.randint(0, diff + 1) + return data.narrow(dim=dim, start=start, length=target_size) + + real_data = get_slice(real_data, 0, b_size) + real_data = get_slice(real_data, 1, t_size) + fake_data = get_slice(fake_data, 0, b_size) + fake_data = get_slice(fake_data, 1, t_size) + + else: + real_data = real_data[:b_size, :t_size] + fake_data = fake_data[:b_size, :t_size] + + alpha = torch.rand(real_data.size(0), 1, 1) + alpha = alpha.expand(real_data.size()) + alpha = alpha.to(real_data.device) + + interpolates = alpha * real_data + ((1 - alpha) * fake_data) + + disc_interpolates = self.discriminator(interpolates, None) + + gradients = autograd.grad( + outputs=disc_interpolates, + inputs=interpolates, + grad_outputs=torch.ones(disc_interpolates.size(), device=real_data.device), + create_graph=True, + retain_graph=True, + only_inputs=True, + )[0] + + gradient_penalty = (gradients.norm(2, dim=1) - 1) ** 2 + return gradient_penalty + + def set_num_updates(self, num_updates): + super().set_num_updates(num_updates) + self.update_num = num_updates + self.curr_temp = max( + self.max_temp * self.temp_decay ** num_updates, self.min_temp + ) + + def discrim_step(self, num_updates): + return num_updates % 2 == 1 + + def get_groups_for_update(self, num_updates): + return "discriminator" if self.discrim_step(num_updates) else "generator" + + def __init__(self, cfg: Wav2vec_UConfig, target_dict): + super().__init__() + + self.cfg = cfg + self.zero_index = target_dict.index("<SIL>") if "<SIL>" in target_dict else 0 + self.smoothness_weight = cfg.smoothness_weight + + output_size = len(target_dict) + self.pad = target_dict.pad() + self.eos = target_dict.eos() + self.smoothing = cfg.smoothing + self.smoothing_one_sided = cfg.smoothing_one_sided + self.no_softmax = cfg.no_softmax + self.gumbel = cfg.gumbel + self.hard_gumbel = cfg.hard_gumbel + self.last_acc = None + + self.gradient_penalty = cfg.gradient_penalty + self.code_penalty = cfg.code_penalty + self.blank_weight = cfg.blank_weight + self.blank_mode = cfg.blank_mode + self.blank_index = target_dict.index("<SIL>") if cfg.blank_is_sil else 0 + assert self.blank_index != target_dict.unk() + + self.discriminator = Discriminator(output_size, cfg) + for p in self.discriminator.parameters(): + p.param_group = "discriminator" + + self.pca_A = self.pca_b = None + d = cfg.input_dim + + self.segmenter = SEGMENT_FACTORY[cfg.segmentation.type](cfg.segmentation) + + self.generator = Generator(d, output_size, cfg) + + for p in self.generator.parameters(): + p.param_group = "generator" + + for p in self.segmenter.parameters(): + p.param_group = "generator" + + self.max_temp, self.min_temp, self.temp_decay = cfg.temp + self.curr_temp = self.max_temp + self.update_num = 0 + + @classmethod + def build_model(cls, cfg, task): + return cls(cfg, task.target_dictionary) + + def get_logits( + self, + net_output: Optional[Dict[str, List[Optional[torch.Tensor]]]], + normalize: bool = False, + ): + logits = net_output["logits"] + + if self.blank_weight != 0: + if self.blank_mode == "add": + logits[..., self.blank_index] += self.blank_weight + elif self.blank_mode == "set": + logits[..., self.blank_index] = self.blank_weight + else: + raise Exception(f"invalid blank mode {self.blank_mode}") + + padding = net_output["padding_mask"] + if padding.any(): + logits[padding] = float("-inf") + logits[padding][..., self.blank_index] = float("inf") + + if normalize: + logits = utils.log_softmax(logits.float(), dim=-1) + + return logits.transpose(0, 1) + + def get_normalized_probs( + self, + net_output: Tuple[ + torch.Tensor, Optional[Dict[str, List[Optional[torch.Tensor]]]] + ], + log_probs: bool, + sample: Optional[Dict[str, torch.Tensor]] = None, + ): + logits = self.get_logits(net_output) + + probs = super().get_normalized_probs(logits, log_probs, sample) + # BTC -> TBC for ctc + probs = probs.transpose(0, 1) + return probs + + def normalize(self, dense_x): + + bsz, tsz, csz = dense_x.shape + + if dense_x.numel() == 0: + raise Exception(dense_x.shape) + _, k = dense_x.max(-1) + hard_x = ( + dense_x.new_zeros(bsz * tsz, csz) + .scatter_(-1, k.view(-1, 1), 1.0) + .view(-1, csz) + ) + hard_probs = torch.mean(hard_x.float(), dim=0) + code_perplexity = torch.exp( + -torch.sum(hard_probs * torch.log(hard_probs + 1e-7), dim=-1) + ) + + avg_probs = torch.softmax(dense_x.reshape(-1, csz).float(), dim=-1).mean(dim=0) + prob_perplexity = torch.exp( + -torch.sum(avg_probs * torch.log(avg_probs + 1e-7), dim=-1) + ) + + if not self.no_softmax: + if self.training and self.gumbel: + dense_x = F.gumbel_softmax( + dense_x.float(), tau=self.curr_temp, hard=self.hard_gumbel + ).type_as(dense_x) + else: + dense_x = dense_x.softmax(-1) + + return dense_x, code_perplexity, prob_perplexity + + def forward( + self, + features, + padding_mask, + random_label=None, + dense_x_only=False, + segment=True, + ): + if segment: + features, padding_mask = self.segmenter.pre_segment(features, padding_mask) + + orig_size = features.size(0) * features.size(1) - padding_mask.sum() + + gen_result = self.generator(features, random_label, padding_mask) + + orig_dense_x, token_x = gen_result["dense_x"], gen_result["token_x"] + orig_dense_padding_mask = gen_result["dense_padding_mask"] + + if segment: + dense_x, dense_padding_mask = self.segmenter.logit_segment( + orig_dense_x, orig_dense_padding_mask + ) + else: + dense_x = orig_dense_x + dense_padding_mask = orig_dense_padding_mask + + dense_logits = dense_x + prob_perplexity = None + code_perplexity = None + + if not (self.no_softmax and dense_x_only): + dense_x, code_perplexity, prob_perplexity = self.normalize(dense_logits) + + if dense_x_only or self.discriminator is None: + return { + "logits": dense_x, + "padding_mask": dense_padding_mask, + } + + token_padding_mask = random_label == self.pad + + dense_y = self.discriminator(dense_x, dense_padding_mask) + token_y = self.discriminator(token_x, token_padding_mask) + + sample_size = features.size(0) + + d_step = self.discrim_step(self.update_num) + + fake_smooth = self.smoothing + real_smooth = self.smoothing + if self.smoothing_one_sided: + fake_smooth = 0 + + zero_loss = None + smoothness_loss = None + code_pen = None + + if d_step: + loss_dense = F.binary_cross_entropy_with_logits( + dense_y, + dense_y.new_ones(dense_y.shape) - fake_smooth, + reduction="sum", + ) + loss_token = F.binary_cross_entropy_with_logits( + token_y, + token_y.new_zeros(token_y.shape) + real_smooth, + reduction="sum", + ) + if self.training and self.gradient_penalty > 0: + grad_pen = self.calc_gradient_penalty(token_x, dense_x) + grad_pen = grad_pen.sum() * self.gradient_penalty + else: + grad_pen = None + else: + grad_pen = None + loss_token = None + loss_dense = F.binary_cross_entropy_with_logits( + dense_y, + dense_y.new_zeros(dense_y.shape) + fake_smooth, + reduction="sum", + ) + num_vars = dense_x.size(-1) + if prob_perplexity is not None: + code_pen = (num_vars - prob_perplexity) / num_vars + code_pen = code_pen * sample_size * self.code_penalty + + if self.smoothness_weight > 0: + smoothness_loss = F.mse_loss( + dense_logits[:, :-1], dense_logits[:, 1:], reduction="none" + ) + smoothness_loss[dense_padding_mask[:, 1:]] = 0 + smoothness_loss = ( + smoothness_loss.mean() * sample_size * self.smoothness_weight + ) + + result = { + "losses": { + "grad_pen": grad_pen, + "code_pen": code_pen, + "smoothness": smoothness_loss, + }, + "temp": self.curr_temp, + "code_ppl": code_perplexity, + "prob_ppl": prob_perplexity, + "d_steps": int(d_step), + "sample_size": sample_size, + } + + suff = "_d" if d_step else "_g" + result["losses"]["dense" + suff] = loss_dense + result["losses"]["token" + suff] = loss_token + + return result diff --git a/examples/wav2vec/unsupervised/scripts/apply_pca.py b/examples/wav2vec/unsupervised/scripts/apply_pca.py new file mode 100644 index 0000000000000000000000000000000000000000..10ad6ce47cfdf0a87ba089b299fe9551b29fa167 --- /dev/null +++ b/examples/wav2vec/unsupervised/scripts/apply_pca.py @@ -0,0 +1,76 @@ +#!/usr/bin/env python3 -u +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import os +import os.path as osp +import math +import numpy as np +import tqdm +import torch +from shutil import copyfile + +from npy_append_array import NpyAppendArray + + +def get_parser(): + parser = argparse.ArgumentParser( + description="transforms features via a given pca and stored them in target dir" + ) + # fmt: off + parser.add_argument('source', help='directory with features') + parser.add_argument('--split', help='which split to read', required=True) + parser.add_argument('--save-dir', help='where to save the output', required=True) + parser.add_argument('--pca-path', type=str, help='pca location. will append _A.npy and _b.npy', required=True) + parser.add_argument('--batch-size', type=int, default=2048000, help='batch size') + parser.add_argument('--unfiltered', action='store_true', help='process the unfiltered version') + # fmt: on + + return parser + + +def main(): + parser = get_parser() + args = parser.parse_args() + + source_path = osp.join(args.source, args.split) + data_poth = source_path + "_unfiltered" if args.unfiltered else source_path + + print(f"data path: {data_poth}") + + features = np.load(data_poth + ".npy", mmap_mode="r") + pca_A = torch.from_numpy(np.load(args.pca_path + "_A.npy")).cuda() + pca_b = torch.from_numpy(np.load(args.pca_path + "_b.npy")).cuda() + + os.makedirs(args.save_dir, exist_ok=True) + save_path = osp.join(args.save_dir, args.split) + + copyfile(source_path + ".tsv", save_path + ".tsv") + copyfile(data_poth + ".lengths", save_path + ".lengths") + + if osp.exists(source_path + ".phn"): + copyfile(source_path + ".phn", save_path + ".phn") + + if osp.exists(source_path + ".wrd"): + copyfile(source_path + ".wrd", save_path + ".wrd") + + if osp.exists(save_path + ".npy"): + os.remove(save_path + ".npy") + npaa = NpyAppendArray(save_path + ".npy") + + batches = math.ceil(features.shape[0] / args.batch_size) + + with torch.no_grad(): + for b in tqdm.trange(batches): + start = b * args.batch_size + end = start + args.batch_size + x = torch.from_numpy(features[start:end]).cuda() + x = torch.matmul(x, pca_A) + pca_b + npaa.append(x.cpu().numpy()) + + +if __name__ == "__main__": + main() diff --git a/examples/wav2vec/unsupervised/scripts/copy_labels.py b/examples/wav2vec/unsupervised/scripts/copy_labels.py new file mode 100644 index 0000000000000000000000000000000000000000..989868388eefccc37c82d7602f709632035c7aa1 --- /dev/null +++ b/examples/wav2vec/unsupervised/scripts/copy_labels.py @@ -0,0 +1,10 @@ +#!/usr/bin/env python3 -u +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import sys + +for idx, line in enumerate(sys.stdin): + print(f"utt{idx:010d} {line}", end="") diff --git a/examples/wav2vec/unsupervised/scripts/filter_lexicon.py b/examples/wav2vec/unsupervised/scripts/filter_lexicon.py new file mode 100644 index 0000000000000000000000000000000000000000..5bf3e51e7a50ac3f07cc41739198cde946dc79aa --- /dev/null +++ b/examples/wav2vec/unsupervised/scripts/filter_lexicon.py @@ -0,0 +1,40 @@ +#!/usr/bin/env python3 -u +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import sys + +from fairseq.data import Dictionary + + +def get_parser(): + parser = argparse.ArgumentParser( + description="filters a lexicon given a unit dictionary" + ) + parser.add_argument("-d", "--unit-dict", help="unit dictionary", required=True) + return parser + + +def main(): + parser = get_parser() + args = parser.parse_args() + + d = Dictionary.load(args.unit_dict) + symbols = set(d.symbols) + + for line in sys.stdin: + items = line.rstrip().split() + skip = len(items) < 2 + for x in items[1:]: + if x not in symbols: + skip = True + break + if not skip: + print(line, end="") + + +if __name__ == "__main__": + main() diff --git a/examples/wav2vec/unsupervised/scripts/filter_tsv.py b/examples/wav2vec/unsupervised/scripts/filter_tsv.py new file mode 100644 index 0000000000000000000000000000000000000000..a09d79acf31414ea3eae82db59cf9f105aefcdf1 --- /dev/null +++ b/examples/wav2vec/unsupervised/scripts/filter_tsv.py @@ -0,0 +1,37 @@ +#!/usr/bin/env python3 -u +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import os +import argparse +import sys + + +parser = argparse.ArgumentParser() +parser.add_argument("--tsv", required=True, type=str) +parser.add_argument("--no-skip", action="store_true") +parser.add_argument("--keep", action="store_true") +params = parser.parse_args() + + +def get_fname(line): + p = os.path.basename(line.split("\t")[0]) + p = os.path.splitext(p)[0] + return p + + +# filenames to exclude +seen = set() +with open(params.tsv) as f: + if not params.no_skip: + root = next(f).rstrip() + for line in f: + seen.add(get_fname(line)) + +for i, line in enumerate(sys.stdin): + exists = get_fname(line) in seen + keep = (exists and params.keep) or (not exists and not params.keep) + if i == 0 or keep: + print(line, end="") diff --git a/examples/wav2vec/unsupervised/scripts/g2p_wrd_to_phn.py b/examples/wav2vec/unsupervised/scripts/g2p_wrd_to_phn.py new file mode 100644 index 0000000000000000000000000000000000000000..2e31c307bd67d10941150160c7fb8c9e085ac5d9 --- /dev/null +++ b/examples/wav2vec/unsupervised/scripts/g2p_wrd_to_phn.py @@ -0,0 +1,45 @@ +#!/usr/bin/env python3 -u +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import sys + +from g2p_en import G2p + + +def main(): + parser = argparse.ArgumentParser() + parser.add_argument( + "--compact", + action="store_true", + help="if set, compacts phones", + ) + args = parser.parse_args() + + compact = args.compact + + wrd_to_phn = {} + g2p = G2p() + for line in sys.stdin: + words = line.strip().split() + phones = [] + for w in words: + if w not in wrd_to_phn: + wrd_to_phn[w] = g2p(w) + if compact: + wrd_to_phn[w] = [ + p[:-1] if p[-1].isnumeric() else p for p in wrd_to_phn[w] + ] + phones.extend(wrd_to_phn[w]) + try: + print(" ".join(phones)) + except: + print(wrd_to_phn, words, phones, file=sys.stderr) + raise + + +if __name__ == "__main__": + main() diff --git a/examples/wav2vec/unsupervised/scripts/ltr_to_wrd.py b/examples/wav2vec/unsupervised/scripts/ltr_to_wrd.py new file mode 100644 index 0000000000000000000000000000000000000000..36c85d1e2f60487494a92207feb4685e78db8aa2 --- /dev/null +++ b/examples/wav2vec/unsupervised/scripts/ltr_to_wrd.py @@ -0,0 +1,16 @@ +#!/usr/bin/env python3 -u +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import sys + + +def main(): + for line in sys.stdin: + print(line.replace(" ", "").replace("|", " ").strip()) + + +if __name__ == "__main__": + main() diff --git a/examples/wav2vec/unsupervised/scripts/mean_pool.py b/examples/wav2vec/unsupervised/scripts/mean_pool.py new file mode 100644 index 0000000000000000000000000000000000000000..4eea048ef3455cb3c897e74c18778c78fdc9fcbf --- /dev/null +++ b/examples/wav2vec/unsupervised/scripts/mean_pool.py @@ -0,0 +1,99 @@ +#!/usr/bin/env python3 -u +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import os +import os.path as osp +import math +import numpy as np +import tqdm +import torch +import torch.nn.functional as F +from shutil import copyfile + +from npy_append_array import NpyAppendArray + + +def get_parser(): + parser = argparse.ArgumentParser( + description="mean pools representations by compressing uniform splits of the data" + ) + # fmt: off + parser.add_argument('source', help='directory with features') + parser.add_argument('--split', help='which split to read', required=True) + parser.add_argument('--save-dir', help='where to save the output', required=True) + parser.add_argument('--subsample-rate', type=float, default=0.5, help='size to subsample data to') + + parser.add_argument('--remove-extra', action='store_true', help='if true, removes extra states that cant be pooled, otherwise pads with 0s') + # fmt: on + + return parser + + +def main(): + parser = get_parser() + args = parser.parse_args() + + source_path = osp.join(args.source, args.split) + + print(f"data path: {source_path}") + + features = np.load(source_path + ".npy", mmap_mode="r") + + os.makedirs(args.save_dir, exist_ok=True) + save_path = osp.join(args.save_dir, args.split) + + copyfile(source_path + ".tsv", save_path + ".tsv") + + if os.path.exists(source_path + ".phn"): + copyfile(source_path + ".phn", save_path + ".phn") + if os.path.exists(source_path + ".wrd"): + copyfile(source_path + ".wrd", save_path + ".wrd") + + if os.path.exists(osp.join(args.source, "dict.phn.txt")): + copyfile( + osp.join(args.source, "dict.phn.txt"), + osp.join(args.save_dir, "dict.phn.txt"), + ) + + if osp.exists(save_path + ".npy"): + os.remove(save_path + ".npy") + npaa = NpyAppendArray(save_path + ".npy") + + with open(source_path + ".lengths", "r") as lf: + lengths = lf.readlines() + + fsz = features.shape[-1] + start = 0 + with torch.no_grad(): + with open(save_path + ".lengths", "w") as lengths_out: + for length in tqdm.tqdm(lengths): + length = int(length) + end = start + length + feats = features[start:end] + start += length + x = torch.from_numpy(feats).cuda() + target_num = math.ceil(length * args.subsample_rate) + rem = length % target_num + + if rem > 0: + if args.remove_extra: + to_rem = target_num - rem + target_num -= 1 + x = x[:-to_rem] + else: + to_add = target_num - rem + x = F.pad(x, [0, 0, 0, to_add]) + x[-to_add:] = x[-to_add - 1] + + x = x.view(target_num, -1, fsz) + x = x.mean(dim=-2) + print(target_num, file=lengths_out) + npaa.append(x.cpu().numpy()) + + +if __name__ == "__main__": + main() diff --git a/examples/wav2vec/unsupervised/scripts/merge_clusters.py b/examples/wav2vec/unsupervised/scripts/merge_clusters.py new file mode 100644 index 0000000000000000000000000000000000000000..2780f9d971d847b3ad0b59e9a33780553ebce902 --- /dev/null +++ b/examples/wav2vec/unsupervised/scripts/merge_clusters.py @@ -0,0 +1,114 @@ +#!/usr/bin/env python3 -u +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import os +import os.path as osp +import numpy as np +import tqdm +import torch +import random +from shutil import copyfile + +from npy_append_array import NpyAppendArray + + +def get_parser(): + parser = argparse.ArgumentParser( + description="transforms features via a given pca and stored them in target dir" + ) + # fmt: off + parser.add_argument('source', help='directory with features') + parser.add_argument('--split', help='which split to read', required=True) + parser.add_argument('--save-dir', help='where to save the output', required=True) + parser.add_argument('--cluster-dir', help='where the clusters are') + parser.add_argument('--pooling', type=str, default='mean', choices=['mean', 'sample'], help='how to pool') + # fmt: on + + return parser + + +def main(): + parser = get_parser() + args = parser.parse_args() + + source_path = osp.join(args.source, args.split) + cluster_path = osp.join(args.cluster_dir, args.split + ".src") + print(f"data path: {source_path}") + + features = np.load(source_path + ".npy", mmap_mode="r") + sizes = [] + offsets = [] + offset = 0 + with open(source_path + ".lengths", "r") as len_f: + for line in len_f: + length = int(line.rstrip()) + sizes.append(length) + offsets.append(offset) + offset += length + + clusters = [] + with open(cluster_path, "r") as cf: + for line in cf: + line = line.rstrip() + items = line.split() + items = list(map(int, items)) + clusters.append(items) + + os.makedirs(args.save_dir, exist_ok=True) + save_path = osp.join(args.save_dir, args.split) + + copyfile(source_path + ".tsv", save_path + ".tsv") + + if os.path.exists(source_path + ".phn"): + copyfile(source_path + ".phn", save_path + ".phn") + if os.path.exists(osp.join(args.source, "dict.phn.txt")): + copyfile( + osp.join(args.source, "dict.phn.txt"), + osp.join(args.save_dir, "dict.phn.txt"), + ) + if os.path.exists(source_path + ".wrd"): + copyfile(source_path + ".wrd", save_path + ".wrd") + + if osp.exists(save_path + ".npy"): + os.remove(save_path + ".npy") + npaa = NpyAppendArray(save_path + ".npy") + + def merge(feats, clust): + feats = torch.from_numpy(feats.copy()) + clust = torch.LongTensor(clust) + _, counts = clust.unique_consecutive(return_counts=True) + curr = 0 + + merged = [] + for c in counts: + c = c.item() + start = curr + end = curr + c + curr += c + if args.pooling == "mean": + new_x = feats[start:end].mean(dim=0) + elif args.pooling == "sample": + new_x = feats[start + int(random.random() * c)] + else: + raise NotImplementedError() + merged.append(new_x) + + return torch.stack(merged, dim=0).numpy() + + with open(save_path + ".lengths", "w") as l_f: + for size, offset, clust in tqdm.tqdm( + zip(sizes, offsets, clusters), total=len(sizes) + ): + end = size + offset + feats = features[offset:end] + feats = merge(feats, clust) + print(len(feats), file=l_f) + npaa.append(feats) + + +if __name__ == "__main__": + main() diff --git a/examples/wav2vec/unsupervised/scripts/normalize_and_filter_text.py b/examples/wav2vec/unsupervised/scripts/normalize_and_filter_text.py new file mode 100644 index 0000000000000000000000000000000000000000..c2bd16efb530af5af3f72ab0edb3044b4e9fcd5c --- /dev/null +++ b/examples/wav2vec/unsupervised/scripts/normalize_and_filter_text.py @@ -0,0 +1,72 @@ +#!/usr/bin/env python3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import fasttext as ft +import os +import regex +import sys + + +def get_parser(): + parser = argparse.ArgumentParser( + description="reads text from stdin and outputs normalized, lid-filtered version to stdout" + ) + parser.add_argument( + "--fasttext-model", + help="path to fasttext model", + default="lid.187.bin", + ) + parser.add_argument("--lang", help="language id", required=True) + parser.add_argument( + "--lid-threshold", + type=float, + help="threshold for this lang id probability", + default=0.4, + ) + + return parser + + +def main(): + parser = get_parser() + args = parser.parse_args() + filter_r = regex.compile(r"[^\p{L}\p{N}\p{M}\' \-]") + + lg = args.lang.lower() + lg_label = f"__label__{lg}" + thresh = args.lid_threshold + + if os.path.exists(args.fasttext_model): + model = ft.load_model(args.fasttext_model) + else: + print( + f"fasttext language id model {args.fasttext_model} not found. Proceeding without language filtering. " + f"To enable language filtering, please download the latest language id model " + f"from https://fasttext.cc/docs/en/language-identification.html", + file=sys.stderr, + ) + model = None + + for line in sys.stdin: + line = line.strip() + line = filter_r.sub(" ", line) + line = " ".join(line.split()) + + if model is not None: + lid, prob = model.predict(line, k=100) + try: + target_idx = lid.index(lg_label) + except ValueError: + continue + if target_idx == 0 or prob[target_idx] >= thresh: + print(line) + else: + print(line) + + +if __name__ == "__main__": + main() diff --git a/examples/wav2vec/unsupervised/scripts/normalize_text.py b/examples/wav2vec/unsupervised/scripts/normalize_text.py new file mode 100644 index 0000000000000000000000000000000000000000..9d0ffeb27d038a6b82aaf0f6bdf208af565663f6 --- /dev/null +++ b/examples/wav2vec/unsupervised/scripts/normalize_text.py @@ -0,0 +1,22 @@ +#!/usr/bin/env python3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import regex +import sys + + +def main(): + filter_r = regex.compile(r"[^\p{L}\p{N}\p{M}\' \-]") + + for line in sys.stdin: + line = line.strip() + line = filter_r.sub(" ", line) + line = " ".join(line.split()) + print(line) + + +if __name__ == "__main__": + main() diff --git a/examples/wav2vec/unsupervised/scripts/pca.py b/examples/wav2vec/unsupervised/scripts/pca.py new file mode 100644 index 0000000000000000000000000000000000000000..948cf5319fd86ba1bccff65270b2881048faf9b1 --- /dev/null +++ b/examples/wav2vec/unsupervised/scripts/pca.py @@ -0,0 +1,53 @@ +#!/usr/bin/env python3 -u +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import os +import os.path as osp +import numpy as np + +import faiss + + + +def get_parser(): + parser = argparse.ArgumentParser( + description="compute a pca matrix given an array of numpy features" + ) + # fmt: off + parser.add_argument('data', help='numpy file containing features') + parser.add_argument('--output', help='where to save the pca matrix', required=True) + parser.add_argument('--dim', type=int, help='dim for pca reduction', required=True) + parser.add_argument('--eigen-power', type=float, default=0, help='eigen power, -0.5 for whitening') + + return parser + + +def main(): + parser = get_parser() + args = parser.parse_args() + + print("Reading features") + x = np.load(args.data, mmap_mode="r") + + print("Computing PCA") + pca = faiss.PCAMatrix(x.shape[-1], args.dim, args.eigen_power) + pca.train(x) + b = faiss.vector_to_array(pca.b) + A = faiss.vector_to_array(pca.A).reshape(pca.d_out, pca.d_in) + + os.makedirs(args.output, exist_ok=True) + + prefix = str(args.dim) + if args.eigen_power != 0: + prefix += f"_{args.eigen_power}" + + np.save(osp.join(args.output, f"{prefix}_pca_A"), A.T) + np.save(osp.join(args.output, f"{prefix}_pca_b"), b) + + +if __name__ == "__main__": + main() diff --git a/examples/wav2vec/unsupervised/scripts/phonemize_with_sil.py b/examples/wav2vec/unsupervised/scripts/phonemize_with_sil.py new file mode 100644 index 0000000000000000000000000000000000000000..c6512d7322def67b27aba46e9e36da171db6963b --- /dev/null +++ b/examples/wav2vec/unsupervised/scripts/phonemize_with_sil.py @@ -0,0 +1,83 @@ +#!/usr/bin/env python3 -u +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import numpy as np +import sys + + +def get_parser(): + parser = argparse.ArgumentParser( + description="converts words to phones adding optional silences around in between words" + ) + parser.add_argument( + "--sil-prob", + "-s", + type=float, + default=0, + help="probability of inserting silence between each word", + ) + parser.add_argument( + "--surround", + action="store_true", + help="if set, surrounds each example with silence", + ) + parser.add_argument( + "--lexicon", + help="lexicon to convert to phones", + required=True, + ) + + return parser + + +def main(): + parser = get_parser() + args = parser.parse_args() + + sil_prob = args.sil_prob + surround = args.surround + sil = "<SIL>" + + wrd_to_phn = {} + + with open(args.lexicon, "r") as lf: + for line in lf: + items = line.rstrip().split() + assert len(items) > 1, line + assert items[0] not in wrd_to_phn, items + wrd_to_phn[items[0]] = items[1:] + + for line in sys.stdin: + words = line.strip().split() + + if not all(w in wrd_to_phn for w in words): + continue + + phones = [] + if surround: + phones.append(sil) + + sample_sil_probs = None + if sil_prob > 0 and len(words) > 1: + sample_sil_probs = np.random.random(len(words) - 1) + + for i, w in enumerate(words): + phones.extend(wrd_to_phn[w]) + if ( + sample_sil_probs is not None + and i < len(sample_sil_probs) + and sample_sil_probs[i] < sil_prob + ): + phones.append(sil) + + if surround: + phones.append(sil) + print(" ".join(phones)) + + +if __name__ == "__main__": + main() diff --git a/examples/wav2vec/unsupervised/scripts/prepare_audio.sh b/examples/wav2vec/unsupervised/scripts/prepare_audio.sh new file mode 100644 index 0000000000000000000000000000000000000000..013f7a9b055a7693a29f9c5ba1e4003a9a25850e --- /dev/null +++ b/examples/wav2vec/unsupervised/scripts/prepare_audio.sh @@ -0,0 +1,78 @@ +#!/usr/bin/env zsh +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +source_dir=$1 +tgt_dir=$2 +model=$3 + +if [ -z "$4" ] + then + dim=512 + else + dim=$4 +fi + +echo "using $dim dim for PCA" + +if [ -z "$5" ] + then + layer=14 + else + layer=$5 +fi + +echo "extracting from layer $layer" + +train_split=train +valid_split=valid +test_split=test + +all_splits=($train_split) + +if [[ -f "$source_dir/valid.tsv" ]]; then + all_splits+=('valid') +fi + +if [[ -f "$source_dir/test.tsv" ]]; then + all_splits+=('test') +fi + +echo "processing splits: $all_splits" + +mkdir -p $tgt_dir + +cp $source_dir/*.tsv $tgt_dir +cp $source_dir/*.wrd $tgt_dir +cp $source_dir/*.ltr $tgt_dir +cp $source_dir/*.phn $tgt_dir +cp $source_dir/dict* $tgt_dir + +setopt shwordsplit + +for split in $all_splits; do + python $FAIRSEQ_ROOT/examples/wav2vec/unsupervised/scripts/wav2vec_extract_features.py $source_dir --split $split \ + --save-dir $tgt_dir --checkpoint $model --layer $layer +done + +python $FAIRSEQ_ROOT/examples/wav2vec/unsupervised/scripts/wav2vec_cluster_faiss.py $tgt_dir/${train_split}.tsv \ +--checkpoint $model --save-dir $tgt_dir -f "CLUS128" --sample-pct 1.0 + +for split in $all_splits; do + python $FAIRSEQ_ROOT/examples/wav2vec/unsupervised/scripts/wav2vec_apply_cluster_faiss.py $tgt_dir \ + --checkpoint $model --path $tgt_dir/CLUS128 --split $split +done + +python $FAIRSEQ_ROOT/examples/wav2vec/unsupervised/scripts/pca.py $tgt_dir/${train_split}.npy --output $tgt_dir/pca --dim $dim + +for split in $all_splits; do + python $FAIRSEQ_ROOT/examples/wav2vec/unsupervised/scripts/apply_pca.py $tgt_dir --split $split --save-dir $tgt_dir/precompute_pca$dim --pca-path $tgt_dir/pca/${dim}_pca --batch-size 1048000 + + python $FAIRSEQ_ROOT/examples/wav2vec/unsupervised/scripts/merge_clusters.py $tgt_dir/precompute_pca$dim --cluster-dir $tgt_dir/CLUS128 \ + --split $split --save-dir $tgt_dir/precompute_pca${dim}_cls128_mean --pooling mean + + python $FAIRSEQ_ROOT/examples/wav2vec/unsupervised/scripts/mean_pool.py $tgt_dir/precompute_pca${dim}_cls128_mean \ + --save-dir $tgt_dir/precompute_pca${dim}_cls128_mean_pooled --split $split +done diff --git a/examples/wav2vec/unsupervised/scripts/prepare_text.sh b/examples/wav2vec/unsupervised/scripts/prepare_text.sh new file mode 100644 index 0000000000000000000000000000000000000000..1caf13cb6a2a0bd84e5322c92124b2fa37368f9a --- /dev/null +++ b/examples/wav2vec/unsupervised/scripts/prepare_text.sh @@ -0,0 +1,82 @@ +#!/usr/bin/env zsh +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +lg=$1 +text_path=$2 +target_dir=$3 +min_phones=$4 +phonemizer=$5 +lid_path=$6 + +if [ -z "$lid_path" ]; then + lid_path="lid.187.bin" +fi + +ph_lg=${lg:l} +if test "$lg" = 'fr'; then + ph_lg='fr-fr' +elif test "$lg" = 'en'; then + ph_lg='en-us' +elif test "$lg" = 'pt'; then + ph_lg='pt-br' +fi + +ESPEAK_PATH='' +if test "$phonemizer" = 'espeak'; then + ESPEAK_PATH=$(which espeak) +elif test "$phonemizer" = 'espeak-ng'; then + ESPEAK_PATH=$(which espeak-ng) +elif test "$phonemizer" = 'G2P'; then + ESPEAK_PATH='' +else + echo "Unknown phonemizer $phonemizer. Valid options are espeak, espean-ng and G2P" + exit 1 +fi + +echo $lg +echo $ph_lg +echo $text_path +echo $target_dir +echo "min phone seen threshold is $min_phones" + +mkdir -p $target_dir +python $FAIRSEQ_ROOT/examples/wav2vec/unsupervised/scripts/normalize_and_filter_text.py --lang $lg --fasttext-model $lid_path < $text_path | grep -v '\-\-\-' >! $target_dir/lm.upper.lid.txt +python $FAIRSEQ_ROOT/fairseq_cli/preprocess.py --dataset-impl mmap --trainpref $target_dir/lm.upper.lid.txt --only-source --destdir $target_dir --thresholdsrc 2 --padding-factor 1 --dict-only +cut -f1 -d' ' $target_dir/dict.txt | grep -v -x '[[:punct:]]*' | grep -Pv '\d\d\d\d\d+' >! $target_dir/words.txt + + +if [ -z "$ESPEAK_PATH" ]; then + python $FAIRSEQ_ROOT/examples/wav2vec/unsupervised/scripts/g2p_wrd_to_phn.py --compact < $target_dir/words.txt > $target_dir/phones.txt +else + # echoing 1 into corpus will prevent the mismatch lines between lexicon and phones in case the phonemizer fails + one=$(echo "1" | PHONEMIZER_ESPEAK_PATH=$ESPEAK_PATH phonemize -p ' ' -w '' -l $ph_lg --language-switch remove-flags) + sed 's/$/ 1/' $target_dir/words.txt | PHONEMIZER_ESPEAK_PATH=$ESPEAK_PATH phonemize -o $target_dir/phones.txt -p ' ' -w '' -l $ph_lg -j 70 --language-switch remove-flags + echo "one is ${one}" + sed -i "s/${one}$//" $target_dir/phones.txt +fi + +paste $target_dir/words.txt $target_dir/phones.txt >! $target_dir/lexicon.lst + +python $FAIRSEQ_ROOT/fairseq_cli/preprocess.py --dataset-impl mmap --trainpref $target_dir/phones.txt --only-source --destdir $target_dir/phones --thresholdsrc $min_phones --padding-factor 1 --dict-only + +python $FAIRSEQ_ROOT/examples/wav2vec/unsupervised/scripts/filter_lexicon.py -d $target_dir/phones/dict.txt < $target_dir/lexicon.lst >! $target_dir/lexicon_filtered.lst +python $FAIRSEQ_ROOT/examples/wav2vec/unsupervised/scripts/phonemize_with_sil.py -s 0.25 --surround --lexicon $target_dir/lexicon_filtered.lst < $target_dir/lm.upper.lid.txt >! $target_dir/phones/lm.phones.filtered.txt +cp $target_dir/phones/dict.txt $target_dir/phones/dict.phn.txt +echo "<SIL> 0" >> $target_dir/phones/dict.phn.txt +python $FAIRSEQ_ROOT/fairseq_cli/preprocess.py --dataset-impl mmap --trainpref $target_dir/phones/lm.phones.filtered.txt --workers 70 --only-source --destdir $target_dir/phones --srcdict $target_dir/phones/dict.phn.txt + +$KENLM_ROOT/lmplz -o 4 < $target_dir/lm.upper.lid.txt --discount_fallback --prune 0 0 0 3 >! $target_dir/kenlm.wrd.o40003.arpa +$KENLM_ROOT/build_binary $target_dir/kenlm.wrd.o40003.arpa $target_dir/kenlm.wrd.o40003.bin + +lg=$lg python $FAIRSEQ_ROOT/examples/speech_recognition/kaldi/kaldi_initializer.py kaldi_root=$KALDI_ROOT fst_dir=$target_dir/fst/phn_to_words_sil lm_arpa=$target_dir/kenlm.wrd.o40003.arpa wav2letter_lexicon=$target_dir/lexicon_filtered.lst data_dir=$target_dir/phones in_labels=phn "blank_symbol='<SIL>'" +lg=$lg python $FAIRSEQ_ROOT/examples/speech_recognition/kaldi/kaldi_initializer.py kaldi_root=$KALDI_ROOT fst_dir=$target_dir/fst/phn_to_words lm_arpa=$target_dir/kenlm.wrd.o40003.arpa wav2letter_lexicon=$target_dir/lexicon_filtered.lst data_dir=$target_dir/phones in_labels=phn + +$KENLM_ROOT/lmplz -o 4 < $target_dir/phones/lm.phones.filtered.txt --discount_fallback >! $target_dir/phones/lm.phones.filtered.04.arpa +$KENLM_ROOT/build_binary $target_dir/phones/lm.phones.filtered.04.arpa $target_dir/phones/lm.phones.filtered.04.bin +$KENLM_ROOT/lmplz -o 6 < $target_dir/phones/lm.phones.filtered.txt --discount_fallback >! $target_dir/phones/lm.phones.filtered.06.arpa +$KENLM_ROOT/build_binary $target_dir/phones/lm.phones.filtered.06.arpa $target_dir/phones/lm.phones.filtered.06.bin + +lg=$lg python $FAIRSEQ_ROOT/examples/speech_recognition/kaldi/kaldi_initializer.py kaldi_root=$KALDI_ROOT fst_dir=$target_dir/fst/phn_to_phn_sil lm_arpa=$target_dir/phones/lm.phones.filtered.06.arpa data_dir=$target_dir/phones in_labels=phn "blank_symbol='<SIL>'" diff --git a/examples/wav2vec/unsupervised/scripts/remove_silence.py b/examples/wav2vec/unsupervised/scripts/remove_silence.py new file mode 100644 index 0000000000000000000000000000000000000000..fac88b989703262a84b242b2761df621bf02c739 --- /dev/null +++ b/examples/wav2vec/unsupervised/scripts/remove_silence.py @@ -0,0 +1,63 @@ +#!/usr/bin/env python3 -u +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +""" +get intervals from .vads file, specify output data, and this script removes silences and saves the audio data in out path folder +paths=shards/train.tsv +vads=shards/train.vads +python remove_silence.py --paths $paths --vads $vads +""" + +import os +import argparse +import torch +import torchaudio +import tqdm + + +parser = argparse.ArgumentParser() +parser.add_argument("--tsv", default="", type=str) +parser.add_argument("--vads", default="", type=str) +parser.add_argument("--out", type=str) +params = parser.parse_args() + +# load paths +paths = [] +with open(params.tsv) as f: + root = next(f).rstrip() + for line in f: + paths.append(os.path.join(root, line.rstrip().split("\t")[0])) + +# load vads +list_intervals = [] +with open(params.vads) as f: + for line in f: + interval = [ + [int(w.split(":")[0]), int(w.split(":")[1])] for w in line.rstrip().split() + ] + list_intervals.append(interval) + + +# load audio and keep only intervals (i.e. remove silences) +for i in tqdm.trange(len(paths)): + data, _ = torchaudio.load(paths[i]) + if len(list_intervals[i]) > 0: + data_filtered = torch.cat( + [data[0][int(it[0]) : int(it[1])] for it in list_intervals[i]] + ).unsqueeze(0) + else: + data_filtered = data + + # YOU MAY NEED TO MODIFY THIS TO GET THE RIGHT SUBPATH + # outpath = params.out + '/'.join(paths[i].split('/')[-1]) + outpath = params.out + "/" + "/".join(paths[i].split("/")[-2:]) + + if not os.path.isdir("/".join(outpath.split("/")[:-1])): + os.makedirs("/".join(outpath.split("/")[:-1])) + if not os.path.exists(outpath): + torchaudio.save(outpath, data_filtered, sample_rate=16000) + else: + print(outpath, "exists!") diff --git a/examples/wav2vec/unsupervised/scripts/vads.py b/examples/wav2vec/unsupervised/scripts/vads.py new file mode 100644 index 0000000000000000000000000000000000000000..2398da97d8c44b8f3f270b22d5508a003482b4d6 --- /dev/null +++ b/examples/wav2vec/unsupervised/scripts/vads.py @@ -0,0 +1,98 @@ +#!/usr/bin/env python3 -u +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import sys + +from copy import deepcopy +from scipy.signal import lfilter + +import numpy as np +from tqdm import tqdm +import soundfile as sf +import os.path as osp + + +def get_parser(): + parser = argparse.ArgumentParser(description="compute vad segments") + parser.add_argument( + "--rvad-home", + "-r", + help="path to rvad home (see https://github.com/zhenghuatan/rVADfast)", + required=True, + ) + + return parser + + +def rvad(speechproc, path): + winlen, ovrlen, pre_coef, nfilter, nftt = 0.025, 0.01, 0.97, 20, 512 + ftThres = 0.5 + vadThres = 0.4 + opts = 1 + + data, fs = sf.read(path) + assert fs == 16_000, "sample rate must be 16khz" + ft, flen, fsh10, nfr10 = speechproc.sflux(data, fs, winlen, ovrlen, nftt) + + # --spectral flatness -- + pv01 = np.zeros(ft.shape[0]) + pv01[np.less_equal(ft, ftThres)] = 1 + pitch = deepcopy(ft) + + pvblk = speechproc.pitchblockdetect(pv01, pitch, nfr10, opts) + + # --filtering-- + ENERGYFLOOR = np.exp(-50) + b = np.array([0.9770, -0.9770]) + a = np.array([1.0000, -0.9540]) + fdata = lfilter(b, a, data, axis=0) + + # --pass 1-- + noise_samp, noise_seg, n_noise_samp = speechproc.snre_highenergy( + fdata, nfr10, flen, fsh10, ENERGYFLOOR, pv01, pvblk + ) + + # sets noisy segments to zero + for j in range(n_noise_samp): + fdata[range(int(noise_samp[j, 0]), int(noise_samp[j, 1]) + 1)] = 0 + + vad_seg = speechproc.snre_vad( + fdata, nfr10, flen, fsh10, ENERGYFLOOR, pv01, pvblk, vadThres + ) + return vad_seg, data + + +def main(): + parser = get_parser() + args = parser.parse_args() + + sys.path.append(args.rvad_home) + import speechproc + + stride = 160 + lines = sys.stdin.readlines() + root = lines[0].rstrip() + for fpath in tqdm(lines[1:]): + path = osp.join(root, fpath.split()[0]) + vads, wav = rvad(speechproc, path) + + start = None + vad_segs = [] + for i, v in enumerate(vads): + if start is None and v == 1: + start = i * stride + elif start is not None and v == 0: + vad_segs.append((start, i * stride)) + start = None + if start is not None: + vad_segs.append((start, len(wav))) + + print(" ".join(f"{v[0]}:{v[1]}" for v in vad_segs)) + + +if __name__ == "__main__": + main() diff --git a/examples/wav2vec/unsupervised/scripts/wav2vec_apply_cluster_faiss.py b/examples/wav2vec/unsupervised/scripts/wav2vec_apply_cluster_faiss.py new file mode 100644 index 0000000000000000000000000000000000000000..a5dd7ae6c15b358206e067385be260c94021bf20 --- /dev/null +++ b/examples/wav2vec/unsupervised/scripts/wav2vec_apply_cluster_faiss.py @@ -0,0 +1,128 @@ +#!/usr/bin/env python3 -u +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import os +import os.path as osp +import numpy as np +import tqdm +import torch +import sys + +import faiss +import torch.nn.functional as F + +from wav2vec_cluster_faiss import parse_faiss_specs, Wav2VecFeatureReader + + +def get_parser(): + parser = argparse.ArgumentParser(description="apply clusters") + # fmt: off + parser.add_argument('data', help='location of tsv files') + parser.add_argument('--split', help='split to process', required=True) + parser.add_argument('--labels', help='split to process', default="phn") + parser.add_argument('--path', help='path to pca and centroids', required=True) + parser.add_argument('--checkpoint', type=str, help='checkpoint for wav2vec model (if using wav2vec features)', required=True) + parser.add_argument('--layer', '-l', type=int, help='which layer to read', default=14) + parser.add_argument('--max-tsz', type=int, help='batch kmeans up to this much', default=14) + # fmt: on + + return parser + + +def get_iterator(args): + label_path = osp.join(args.data, f"{args.split}.{args.labels}") + if osp.exists(label_path): + lp = open(label_path, "r") + else: + lp = None + + with open(osp.join(args.data, f"{args.split}.tsv"), "r") as fp: + lines = fp.read().split("\n") + root = lines.pop(0).strip() + files = [line.rstrip() for line in lines if len(line) > 0] + + if lp is not None: + lbls = [line.rstrip() for line in lp] + else: + lbls = [None] * len(files) + + num = len(files) + reader = Wav2VecFeatureReader(args.checkpoint, args.layer) + + def iterate(): + for fname, lbl in zip(files, lbls): + file = osp.join(root, fname.split("\t")[0]) + feats = reader.get_feats(file) + yield feats.data, fname, lbl + + return iterate, num, root + + +def main(): + parser = get_parser() + args = parser.parse_args() + + spec = osp.basename(args.path) + + try: + faiss_spec = parse_faiss_specs(spec.rstrip("/"))[0] + except: + print(spec) + raise + + print("Faiss Spec:", faiss_spec, file=sys.stderr) + + if faiss_spec.pca: + A = torch.from_numpy(np.load(osp.join(args.path, "pca_A.npy"))).cuda() + b = torch.from_numpy(np.load(osp.join(args.path, "pca_b.npy"))).cuda() + print("Loaded PCA", file=sys.stderr) + + centroids = np.load(osp.join(args.path, "centroids.npy")) + print("Loaded centroids", centroids.shape, file=sys.stderr) + + res = faiss.StandardGpuResources() + index_flat = ( + faiss.IndexFlatL2(centroids.shape[1]) + if not faiss_spec.sphere + else faiss.IndexFlatIP(centroids.shape[1]) + ) + faiss_index = faiss.index_cpu_to_gpu(res, 0, index_flat) + faiss_index.add(centroids) + + generator, num, root = get_iterator(args) + iterator = generator() + + had_labels = False + label_path = osp.join(args.path, f"{args.split}.{args.labels}") + + with torch.no_grad(): + with open(osp.join(args.path, f"{args.split}.src"), "w") as fp, open( + osp.join(args.path, f"{args.split}.tsv"), "w" + ) as pp, open(label_path, "w") as lp: + print(root, file=pp) + for f, fname, lbl in tqdm.tqdm(iterator, total=num): + if faiss_spec.pca: + f = torch.mm(f, A) + b + if faiss_spec.norm: + f = F.normalize(f, p=2, dim=-1) + + f = f.cpu().numpy() + + _, z = faiss_index.search(f, 1) + + print(" ".join(str(x.item()) for x in z), file=fp) + print(fname, file=pp) + + if lbl is not None: + print(lbl, file=lp) + had_labels = True + if not had_labels: + os.remove(label_path) + + +if __name__ == "__main__": + main() diff --git a/examples/wav2vec/unsupervised/scripts/wav2vec_cluster_faiss.py b/examples/wav2vec/unsupervised/scripts/wav2vec_cluster_faiss.py new file mode 100644 index 0000000000000000000000000000000000000000..632a69e9f4bd98d33abb689c15557c818d0e35ea --- /dev/null +++ b/examples/wav2vec/unsupervised/scripts/wav2vec_cluster_faiss.py @@ -0,0 +1,210 @@ +#!/usr/bin/env python3 -u +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import gc +import os +import os.path as osp +import random +import numpy as np +import tqdm +import torch + +from collections import namedtuple + +import faiss + +import fairseq +import soundfile as sf + + +def get_parser(): + parser = argparse.ArgumentParser( + description="compute kmeans codebook from kaldi-computed feats" + ) + # fmt: off + parser.add_argument('data', help='location of tsv files') + parser.add_argument('--save-dir', help='where to save the output', required=True) + parser.add_argument('--checkpoint', type=str, help='checkpoint for wav2vec model (if using wav2vec features)', required=True) + parser.add_argument('--sample-pct', '-r', type=float, help='percentage of timesteps to sample', default=0) + parser.add_argument('--layer', '-l', type=int, help='which layer to read', default=14) + parser.add_argument('--faiss-specs', '-f', type=str, + help='faiss index specs; separated by space ' + 'format is: PCAx_NORM_CLUSx_SPHERICAL -> ' + 'PCAx if exists first apply PCA ' + 'NORM if exists, normalize the vector by L2 norm ' + 'CLUSx must exist, cluster to x clusters ' + 'SPEHRICAL if exists, apply spherical kmeans', + default='l2') + # fmt: on + + return parser + + +faiss_spec = namedtuple("faiss_spec", ["pca", "norm", "n_clus", "sphere", "spec_str"]) + + +def parse_faiss_specs(specs_str): + specs = [] + for ss in specs_str.split(): + comps = ss.split("_") + pca = 0 + norm = False + n_clus = 0 + sphere = False + for c in comps: + if c.startswith("PCA"): + pca = int(c[3:]) + elif c == "NORM": + norm = True + elif c.startswith("CLUS"): + n_clus = int(c[4:]) + elif c == "SPHERICAL": + sphere = True + assert n_clus > 0 + specs.append( + faiss_spec(pca=pca, norm=norm, n_clus=n_clus, sphere=sphere, spec_str=ss) + ) + return specs + + +class Wav2VecFeatureReader(object): + def __init__(self, cp_file, layer): + state = fairseq.checkpoint_utils.load_checkpoint_to_cpu(cp_file) + + self.layer = layer + + if "cfg" in state: + w2v_args = state["cfg"] + task = fairseq.tasks.setup_task(w2v_args.task) + model = task.build_model(w2v_args.model) + else: + w2v_args = state["args"] + task = fairseq.tasks.setup_task(w2v_args) + model = task.build_model(w2v_args) + model.load_state_dict(state["model"], strict=True) + model.eval() + model.cuda() + self.model = model + + def read_audio(self, fname): + """Load an audio file and return PCM along with the sample rate""" + wav, sr = sf.read(fname) + assert sr == 16e3 + + return wav + + def get_feats(self, loc): + x = self.read_audio(loc) + with torch.no_grad(): + source = torch.from_numpy(x).view(1, -1).float().cuda() + res = self.model( + source=source, mask=False, features_only=True, layer=self.layer + ) + return res["layer_results"][self.layer][0].squeeze(1) + + +def get_iterator(args): + with open(args.data, "r") as fp: + lines = fp.read().split("\n") + root = lines.pop(0).strip() + files = [osp.join(root, line.split("\t")[0]) for line in lines if len(line) > 0] + + if getattr(args, "sample_pct", 0) > 0: + files = random.sample(files, int(args.sample_pct * len(files))) + num = len(files) + reader = Wav2VecFeatureReader(args.checkpoint, args.layer) + + def iterate(): + for fname in files: + feats = reader.get_feats(fname) + yield feats.cpu().numpy() + + return iterate, num + + +def main(): + parser = get_parser() + args = parser.parse_args() + + faiss_specs = parse_faiss_specs(args.faiss_specs) + print("Faiss Specs:", faiss_specs) + + feat_path = osp.join(args.save_dir, "features") + if osp.exists(feat_path + ".npy"): + feats = np.load(feat_path + ".npy") + else: + generator, num = get_iterator(args) + iterator = generator() + + feats = [] + for f in tqdm.tqdm(iterator, total=num): + feats.append(f) + + del iterator + del generator + + feats = np.concatenate(feats) + + print(feats.shape) + + os.makedirs(args.save_dir, exist_ok=True) + # np.save(feat_path, feats) + + gc.collect() + torch.cuda.empty_cache() + + reload = False + for spec in faiss_specs: + print("Processing spec", spec) + + if reload: + print("Reloading...") + del feats + gc.collect() + feats = np.load(feat_path + ".npy") + + save_path = osp.join(args.save_dir, spec.spec_str) + os.makedirs(save_path, exist_ok=True) + d = feats.shape[-1] + x = feats + if spec.pca > 0: + print("Computing PCA") + pca = faiss.PCAMatrix(d, spec.pca) + pca.train(x) + d = spec.pca + b = faiss.vector_to_array(pca.b) + A = faiss.vector_to_array(pca.A).reshape(pca.d_out, pca.d_in) + np.save(osp.join(save_path, "pca_A"), A.T) + np.save(osp.join(save_path, "pca_b"), b) + print("Applying PCA") + x = pca.apply_py(x) + + if spec.norm: + reload = spec.pca <= 0 + print("Normalizing") + faiss.normalize_L2(x) + + print("Computing kmeans") + kmeans = faiss.Kmeans( + d, + spec.n_clus, + niter=50, + verbose=True, + spherical=spec.sphere, + max_points_per_centroid=feats.shape[0], + gpu=True, + nredo=3, + ) + kmeans.train(x) + np.save(osp.join(save_path, "centroids"), kmeans.centroids) + del kmeans + del x + gc.collect() + + +if __name__ == "__main__": + main() diff --git a/examples/wav2vec/unsupervised/scripts/wav2vec_extract_features.py b/examples/wav2vec/unsupervised/scripts/wav2vec_extract_features.py new file mode 100644 index 0000000000000000000000000000000000000000..b07e274d202414ce40d00aa64a27cf97bb49c1c3 --- /dev/null +++ b/examples/wav2vec/unsupervised/scripts/wav2vec_extract_features.py @@ -0,0 +1,119 @@ +#!/usr/bin/env python3 -u +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import os +import os.path as osp +import tqdm +import torch +import torch.nn.functional as F +from shutil import copyfile + +from npy_append_array import NpyAppendArray + +import fairseq +import soundfile as sf + + +def get_parser(): + parser = argparse.ArgumentParser( + description="compute kmeans codebook from kaldi-computed feats" + ) + # fmt: off + parser.add_argument('data', help='location of tsv files') + parser.add_argument('--split', help='which split to read', required=True) + parser.add_argument('--save-dir', help='where to save the output', required=True) + parser.add_argument('--checkpoint', type=str, help='checkpoint for wav2vec ctc model', required=True) + parser.add_argument('--layer', type=int, default=14, help='which layer to use') + # fmt: on + + return parser + + +class Wav2VecFeatureReader(object): + def __init__(self, cp_file, layer): + model, cfg, task = fairseq.checkpoint_utils.load_model_ensemble_and_task( + [cp_file] + ) + model = model[0] + model.eval() + model.cuda() + self.model = model + self.task = task + self.layer = layer + + def read_audio(self, fname): + """Load an audio file and return PCM along with the sample rate""" + wav, sr = sf.read(fname) + assert sr == 16e3 + + return wav + + def get_feats(self, loc): + x = self.read_audio(loc) + with torch.no_grad(): + source = torch.from_numpy(x).float().cuda() + if self.task.cfg.normalize: + assert source.dim() == 1, source.dim() + with torch.no_grad(): + source = F.layer_norm(source, source.shape) + source = source.view(1, -1) + + m_res = self.model(source=source, mask=False, features_only=True, layer=self.layer) + return m_res["x"].squeeze(0).cpu() + + +def get_iterator(args): + with open(osp.join(args.data, args.split) + ".tsv", "r") as fp: + lines = fp.read().split("\n") + root = lines.pop(0).strip() + files = [osp.join(root, line.split("\t")[0]) for line in lines if len(line) > 0] + + num = len(files) + reader = Wav2VecFeatureReader(args.checkpoint, args.layer) + + def iterate(): + for fname in files: + w2v_feats = reader.get_feats(fname) + yield w2v_feats + + return iterate, num + + +def main(): + parser = get_parser() + args = parser.parse_args() + + os.makedirs(args.save_dir, exist_ok=True) + + def create_files(dest): + copyfile(osp.join(args.data, args.split) + ".tsv", dest + ".tsv") + if osp.exists(osp.join(args.data, args.split) + ".wrd"): + copyfile(osp.join(args.data, args.split) + ".wrd", dest + ".wrd") + if osp.exists(osp.join(args.data, args.split) + ".phn"): + copyfile(osp.join(args.data, args.split) + ".phn", dest + ".phn") + + if osp.exists(dest + ".npy"): + os.remove(dest + ".npy") + npaa = NpyAppendArray(dest + ".npy") + return npaa + + save_path = osp.join(args.save_dir, args.split) + npaa = create_files(save_path) + + generator, num = get_iterator(args) + iterator = generator() + + with open(save_path + ".lengths", "w") as l_f: + for w2v_feats in tqdm.tqdm(iterator, total=num): + print(len(w2v_feats), file=l_f) + + if len(w2v_feats) > 0: + npaa.append(w2v_feats.numpy()) + + +if __name__ == "__main__": + main() diff --git a/examples/wav2vec/unsupervised/scripts/wer.py b/examples/wav2vec/unsupervised/scripts/wer.py new file mode 100644 index 0000000000000000000000000000000000000000..613ab50d39019f6edf67c56c2353646be2a2f17d --- /dev/null +++ b/examples/wav2vec/unsupervised/scripts/wer.py @@ -0,0 +1,82 @@ +#!/usr/bin/env python3 -u +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +""" +Implement unsupervised metric for decoding hyperparameter selection: + $$ alpha * LM_PPL + ViterbitUER(%) * 100 $$ +""" +import argparse +import logging +import sys + +import editdistance + +logging.root.setLevel(logging.INFO) +logging.basicConfig(stream=sys.stdout, level=logging.INFO) +logger = logging.getLogger(__name__) + + +def get_parser(): + parser = argparse.ArgumentParser() + parser.add_argument("-s", "--hypo", help="hypo transcription", required=True) + parser.add_argument( + "-r", "--reference", help="reference transcription", required=True + ) + return parser + + +def compute_wer(ref_uid_to_tra, hyp_uid_to_tra, g2p): + d_cnt = 0 + w_cnt = 0 + w_cnt_h = 0 + for uid in hyp_uid_to_tra: + ref = ref_uid_to_tra[uid].split() + if g2p is not None: + hyp = g2p(hyp_uid_to_tra[uid]) + hyp = [p for p in hyp if p != "'" and p != " "] + hyp = [p[:-1] if p[-1].isnumeric() else p for p in hyp] + else: + hyp = hyp_uid_to_tra[uid].split() + d_cnt += editdistance.eval(ref, hyp) + w_cnt += len(ref) + w_cnt_h += len(hyp) + wer = float(d_cnt) / w_cnt + logger.debug( + ( + f"wer = {wer * 100:.2f}%; num. of ref words = {w_cnt}; " + f"num. of hyp words = {w_cnt_h}; num. of sentences = {len(ref_uid_to_tra)}" + ) + ) + return wer + + +def main(): + args = get_parser().parse_args() + + errs = 0 + count = 0 + with open(args.hypo, "r") as hf, open(args.reference, "r") as rf: + for h, r in zip(hf, rf): + h = h.rstrip().split() + r = r.rstrip().split() + errs += editdistance.eval(r, h) + count += len(r) + + logger.info(f"UER: {errs / count * 100:.2f}%") + + +if __name__ == "__main__": + main() + + +def load_tra(tra_path): + with open(tra_path, "r") as f: + uid_to_tra = {} + for line in f: + uid, tra = line.split(None, 1) + uid_to_tra[uid] = tra + logger.debug(f"loaded {len(uid_to_tra)} utterances from {tra_path}") + return uid_to_tra diff --git a/examples/wav2vec/unsupervised/scripts/wrd_to_ltr.py b/examples/wav2vec/unsupervised/scripts/wrd_to_ltr.py new file mode 100644 index 0000000000000000000000000000000000000000..f83471409a434556cab70086ca9e2d72d4bdddd5 --- /dev/null +++ b/examples/wav2vec/unsupervised/scripts/wrd_to_ltr.py @@ -0,0 +1,16 @@ +#!/usr/bin/env python3 -u +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import sys + + +def main(): + for line in sys.stdin: + print(" ".join(list(line.strip().replace(" ", "|"))) + " |") + + +if __name__ == "__main__": + main() diff --git a/examples/wav2vec/unsupervised/tasks/__init__.py b/examples/wav2vec/unsupervised/tasks/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..6d7dd625e09451be671908578f93148f371f53cd --- /dev/null +++ b/examples/wav2vec/unsupervised/tasks/__init__.py @@ -0,0 +1,11 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .unpaired_audio_text import UnpairedAudioText + + +__all__ = [ + "UnpairedAudioText", +] diff --git a/examples/wav2vec/unsupervised/tasks/unpaired_audio_text.py b/examples/wav2vec/unsupervised/tasks/unpaired_audio_text.py new file mode 100644 index 0000000000000000000000000000000000000000..5f292528f80d6bb51f16a4324d97342d28fce942 --- /dev/null +++ b/examples/wav2vec/unsupervised/tasks/unpaired_audio_text.py @@ -0,0 +1,447 @@ +# Copyright (c) 2017-present, Facebook, Inc. +# All rights reserved. +# +# This source code is licensed under the license found in the LICENSE file in +# the root directory of this source tree. An additional grant of patent rights +# can be found in the PATENTS file in the same directory. + +from dataclasses import dataclass, field +import logging +import math +import os +from typing import Optional +import torch + +from fairseq.logging import metrics +from fairseq.tasks import FairseqTask, register_task +from ..data import ExtractedFeaturesDataset, RandomInputDataset + +from fairseq.data import ( + Dictionary, + data_utils, + StripTokenDataset, +) +from fairseq.dataclass import FairseqDataclass +from fairseq.distributed.utils import get_data_parallel_world_size +from omegaconf import MISSING + +from examples.speech_recognition.kaldi.kaldi_decoder import ( + KaldiDecoder, + KaldiDecoderConfig, +) + + +logger = logging.getLogger(__name__) + + +@dataclass +class DecodingConfig(FairseqDataclass): + kenlm_path: Optional[str] = None + lm_weight: float = 0 + blank_weight: float = 0 + + +@dataclass +class UnpairedAudioTextConfig(FairseqDataclass): + data: str = field( + default=MISSING, metadata={"help": "path to data directory containing audio"} + ) + text_data: str = field( + default=MISSING, metadata={"help": "path to data directory containing text"} + ) + max_length: Optional[int] = None + labels: Optional[str] = field( + default=None, + metadata={"help": "extension of the label file to load, used for fine-tuning"}, + ) + unfiltered: bool = field( + default=False, metadata={"help": "load data with _unfiltered suffix"} + ) + ctc_eval: bool = field( + default=False, metadata={"help": "eval UER as if computed by CTC"} + ) + sort_by_length: bool = field( + default=True, metadata={"help": "sort examples by length of audio timesteps"} + ) + shuffle: bool = field(default=True, metadata={"help": "shuffle examples"}) + append_eos: bool = field(default=False, metadata={"help": "append eos"}) + uppercase: Optional[bool] = field( + default=False, metadata={"help": "uppercase for LM score computation"} + ) + skipwords: Optional[str] = field( + default="", + metadata={ + "help": "comma-separated words to be removed for LM score computation" + }, + ) + kenlm_path: Optional[str] = None + vocab_usage_power: float = 2 + + word_decoder_config: Optional[KaldiDecoderConfig] = None + word_kenlm_path: Optional[str] = None + + decoding_config: DecodingConfig = DecodingConfig() + + +@register_task("unpaired_audio_text", dataclass=UnpairedAudioTextConfig) +class UnpairedAudioText(FairseqTask): + """ """ + + cfg: UnpairedAudioTextConfig + + def __init__( + self, + cfg: UnpairedAudioTextConfig, + source_dictionary=None, + target_dictionary=None, + ): + super().__init__(cfg) + + self._target_dictionary = target_dictionary + self._source_dictionary = source_dictionary + self.num_symbols = ( + len([s for s in target_dictionary.symbols if not s.startswith("madeup")]) + - target_dictionary.nspecial + ) + self.sil_id = ( + target_dictionary.index("<SIL>") if "<SIL>" in target_dictionary else -1 + ) + self.kenlm = None + if cfg.kenlm_path is not None: + import kenlm + + self.kenlm = kenlm.Model(cfg.kenlm_path) + + self.word_kenlm = None + if cfg.word_kenlm_path is not None: + import kenlm + + self.word_kenlm = kenlm.Model(cfg.word_kenlm_path) + + self.uppercase = cfg.uppercase + self.skipwords = set(cfg.skipwords.split(",")) + + def str_postprocess(s): + s = " ".join(w for w in s.split() if w not in self.skipwords) + s = s.upper() if self.uppercase else s + return s + + self.str_postprocess = str_postprocess + self.compute_lm_score = lambda s: self.kenlm.score(self.str_postprocess(s)) + + self.compute_word_score = None + if cfg.word_decoder_config is not None: + self.kaldi_decoder = KaldiDecoder(cfg.word_decoder_config, beam=10) + + def compute_word_score(logits, padding): + res = self.kaldi_decoder.decode(logits, padding) + for r in res: + r = r.result() + assert len(r) == 1 + r = r[0] + yield r["score"], r["words"] + + self.compute_word_score = compute_word_score + + @classmethod + def setup_task(cls, cfg: UnpairedAudioTextConfig, **kwargs): + """Setup the task (e.g., load dictionaries). + + Args: + cfg (AudioPretrainingConfig): configuration of this task + """ + + dict_path = os.path.join(cfg.text_data, "dict.txt") + if os.path.exists(dict_path): + target_dictionary = Dictionary.load(dict_path) + else: + dict_path = os.path.join(cfg.data, f"dict.{cfg.labels}.txt") + target_dictionary = Dictionary.load(dict_path) + + return cls(cfg, target_dictionary=target_dictionary) + + def optimizer_step(self, optimizer, model, update_num): + if hasattr(model, "get_groups_for_update"): + groups = model.get_groups_for_update(update_num) + optimizer.step(groups={groups}) + else: + optimizer.step() + + def valid_step(self, sample, model, criterion): + res = model( + **sample["net_input"], + dense_x_only=True, + ) + + dense_x = res["logits"] + padding_mask = res["padding_mask"] + + word_scores = None + if self.compute_word_score is not None: + word_scores = self.compute_word_score(dense_x.cpu(), padding_mask.cpu()) + + z = dense_x.argmax(-1) + z[padding_mask] = self.target_dictionary.pad() + + vocab_seen = torch.zeros(self.num_symbols, dtype=torch.bool) + + import editdistance + + c_err = 0 + c_len = 0 + pred_c_len = 0 + lm_score_sum = 0 + for i, (x, t, id) in enumerate( + zip( + z, + sample["target"] if "target" in sample else [None] * len(z), + sample["id"], + ) + ): + + if t is not None: + t = t[(t >= self.target_dictionary.nspecial)] + x = x[ + (x >= self.target_dictionary.nspecial) + & (x < (self.num_symbols + self.target_dictionary.nspecial)) + ] + if self.sil_id >= 0: + x = x[x != self.sil_id] + + vocab_seen[x - self.target_dictionary.nspecial] = True + + pred_units_arr = x + if self.cfg.ctc_eval: + pred_units_arr = pred_units_arr.unique_consecutive() + pred_units_arr = pred_units_arr[pred_units_arr != 0] + + if id == 0: + if t is not None: + logger.info(f"REF: {self.target_dictionary.string(t)}") + logger.info(f"HYP: {self.target_dictionary.string(pred_units_arr)}") + + if self.kenlm is not None: + if t is not None: + ref_lm_s = self.compute_lm_score( + self.target_dictionary.string(t) + ) + logger.info( + f"LM [REF]: {ref_lm_s}, {math.pow(10, -ref_lm_s / (len(t) + 1))}" + ) + + hyp_lm_s = self.compute_lm_score( + self.target_dictionary.string(pred_units_arr) + ) + logger.info( + f"LM [HYP]: {hyp_lm_s}, {math.pow(10, -hyp_lm_s / (len(pred_units_arr) + 1))}" + ) + + pred_units_arr = pred_units_arr.tolist() + + pred_c_len += len(pred_units_arr) + + if t is not None: + t = t.tolist() + c_err += editdistance.eval(pred_units_arr, t) + c_len += len(t) + else: + c_len = pred_c_len + + if self.kenlm is not None: + pred_str = self.target_dictionary.string(pred_units_arr) + lm_score = self.compute_lm_score(pred_str) + lm_score_sum += lm_score + + kaldi_score_sum = 0 + word_lm_sum = 0 + num_words = 0 + if word_scores is not None: + for score, words in word_scores: + kaldi_score_sum += score + num_words += len(words) + if self.word_kenlm is not None: + word_lm_sum += self.kenlm.score(" ".join(words)) + + try: + world_size = get_data_parallel_world_size() + except: + world_size = 1 + + logging_output = { + "loss": c_err, + "_num_char_errors": c_err, + "_num_chars": c_len, + "_num_pred_chars": pred_c_len, + "ntokens": c_len, + "nsentences": z.size(0), + "sample_size": c_len, + "_world_size": world_size, + "_lm_score_sum": lm_score_sum, + "_kaldi_score_sum": kaldi_score_sum, + "_word_lm_sum": word_lm_sum, + "_num_words": num_words, + "_vocab_seen": vocab_seen, + } + + return c_err, c_len, logging_output + + def load_dataset(self, split: str, task_cfg: FairseqDataclass = None, **kwargs): + data_path = self.cfg.data + task_cfg = task_cfg or self.cfg + + has_unpaired_text = os.path.exists( + os.path.join(self.cfg.text_data, f"{split}.idx") + ) + + self.datasets[split] = ExtractedFeaturesDataset( + path=data_path, + split=split, + min_length=3, + max_length=task_cfg.max_length, + labels=None if has_unpaired_text else task_cfg.labels, + label_dict=self.target_dictionary, + shuffle=getattr(task_cfg, "shuffle", True), + sort_by_length=task_cfg.sort_by_length, + ) + + logger.info(f"split {split} has unpaired text? {has_unpaired_text}") + if has_unpaired_text: + text_dataset = data_utils.load_indexed_dataset( + os.path.join(self.cfg.text_data, split), self.target_dictionary + ) + text_dataset = StripTokenDataset(text_dataset, self.target_dictionary.eos()) + self.datasets[split] = RandomInputDataset( + self.datasets[split], + text_dataset, + ["random_label"], + add_to_input=True, + pad_idx=self.target_dictionary.pad(), + ) + + @property + def source_dictionary(self): + return self._source_dictionary + + @property + def target_dictionary(self): + """Return the :class:`~fairseq.data.Dictionary` for the language + model.""" + return self._target_dictionary + + def max_positions(self): + """Maximum input length supported by the encoder.""" + return None + + def reduce_metrics(self, logging_outputs, criterion): + super().reduce_metrics(logging_outputs, criterion) + + zero = torch.scalar_tensor(0.0) + num_char_errors = sum( + log.get("_num_char_errors", zero) for log in logging_outputs + ) + num_chars = sum(log.get("_num_chars", zero) for log in logging_outputs) + num_word_errors = sum( + log.get("_num_word_errors", zero) for log in logging_outputs + ) + num_words = sum(log.get("_num_words", zero) for log in logging_outputs) + num_pred_chars = sum( + log.get("_num_pred_chars", zero) for log in logging_outputs + ) + + lm_score_sum = sum(log.get("_lm_score_sum", zero) for log in logging_outputs) + vocab_seen = ( + sum(log.get("_vocab_seen", zero) for log in logging_outputs) + .bool() + .sum() + .item() + ) + kaldi_score_sum = sum( + log.get("_kaldi_score_sum", zero) for log in logging_outputs + ) + word_lm_sum = sum(log.get("_word_lm_sum", zero) for log in logging_outputs) + + metrics.log_scalar_sum("_num_char_errors", num_char_errors) + metrics.log_scalar_sum("_num_chars", num_chars) + metrics.log_scalar_sum("_num_word_errors", num_word_errors) + metrics.log_scalar_sum("_num_words", num_words) + + metrics.log_scalar_sum("lm_score_sum", lm_score_sum) + metrics.log_scalar_sum("num_pred_chars", num_pred_chars) + + if self.cfg.word_kenlm_path is not None: + metrics.log_scalar_sum("kaldi_score_sum", kaldi_score_sum) + metrics.log_scalar_sum("word_lm_sum", word_lm_sum) + + if num_chars > 0: + metrics.log_derived( + "uer", + lambda meters: meters["_num_char_errors"].sum + * 100.0 + / meters["_num_chars"].sum + if meters["_num_chars"].sum > 0 + else float("nan"), + ) + + if lm_score_sum < 0 and vocab_seen > 0: + metrics.log_scalar("vocab_seen_pct", vocab_seen / self.num_symbols) + + metrics.log_derived( + "weighted_lm_ppl", + lambda meters: math.pow( + 10, + -meters["lm_score_sum"].sum + / ( + meters["num_pred_chars"].sum + meters["nsentences"].sum + ), # account for </s> + ) + / meters["vocab_seen_pct"].avg ** self.cfg.vocab_usage_power, + ) + + metrics.log_derived( + "lm_ppl", + lambda meters: math.pow( + 10, + -meters["lm_score_sum"].sum + / ( + meters["num_pred_chars"].sum + meters["nsentences"].sum + ), # account for </s> + ), + ) + else: + metrics.log_derived("weighted_lm_ppl", lambda meters: float("inf")) + + if num_words > 0: + if word_lm_sum != 0: + metrics.log_derived( + "word_lm_ppl", + lambda meters: math.pow( + 10, + -meters["word_lm_sum"].sum + / ( + meters["_num_words"].sum + meters["nsentences"].sum + ), # account for </s> + ), + ) + metrics.log_derived( + "weighted_word_lm_ppl", + lambda meters: math.pow( + 10, + -meters["word_lm_sum"].sum + / ( + meters["_num_words"].sum + meters["nsentences"].sum + ), # account for </s> + ) + / meters["vocab_seen_pct"].avg ** self.cfg.vocab_usage_power, + ) + + if self.cfg.word_kenlm_path is not None: + metrics.log_derived( + "kaldi_score", + lambda meters: meters["kaldi_score_sum"].sum + / meters["nsentences"].sum, + ) + + def build_model(self, cfg: FairseqDataclass): + model = super().build_model(cfg) + + return model diff --git a/examples/wav2vec/unsupervised/w2vu_generate.py b/examples/wav2vec/unsupervised/w2vu_generate.py new file mode 100644 index 0000000000000000000000000000000000000000..b1e126665fae757ebe06dc5f633eb889c6f6570e --- /dev/null +++ b/examples/wav2vec/unsupervised/w2vu_generate.py @@ -0,0 +1,709 @@ +#!/usr/bin/env python3 -u +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +""" +Run inference for pre-processed data with a trained model. +""" + +import ast +from collections import namedtuple +from dataclasses import dataclass, field +from enum import Enum, auto +import hydra +from hydra.core.config_store import ConfigStore +import logging +import math +import os +from omegaconf import OmegaConf +from typing import Optional +import sys + +import editdistance +import torch + +from hydra.core.hydra_config import HydraConfig + +from fairseq import checkpoint_utils, progress_bar, tasks, utils +from fairseq.data.data_utils import post_process +from fairseq.dataclass.configs import FairseqDataclass, FairseqConfig +from fairseq.logging.meters import StopwatchMeter +from omegaconf import open_dict + +from examples.speech_recognition.kaldi.kaldi_decoder import KaldiDecoderConfig + +logging.root.setLevel(logging.INFO) +logging.basicConfig(stream=sys.stdout, level=logging.INFO) +logger = logging.getLogger(__name__) + + +class DecoderType(Enum): + VITERBI = auto() + KENLM = auto() + FAIRSEQ = auto() + KALDI = auto() + + +@dataclass +class UnsupGenerateConfig(FairseqDataclass): + fairseq: FairseqConfig = FairseqConfig() + lm_weight: float = field( + default=2.0, + metadata={"help": "language model weight"}, + ) + w2l_decoder: DecoderType = field( + default=DecoderType.VITERBI, + metadata={"help": "type of decoder to use"}, + ) + kaldi_decoder_config: Optional[KaldiDecoderConfig] = None + lexicon: Optional[str] = field( + default=None, + metadata={ + "help": "path to lexicon. This is also used to 'phonemize' for unsupvised param tuning" + }, + ) + lm_model: Optional[str] = field( + default=None, + metadata={"help": "path to language model (kenlm or fairseq)"}, + ) + unit_lm: bool = field( + default=False, + metadata={"help": "whether to use unit lm"}, + ) + beam_threshold: float = field( + default=50.0, + metadata={"help": "beam score threshold"}, + ) + beam_size_token: float = field( + default=100.0, + metadata={"help": "max tokens per beam"}, + ) + beam: int = field( + default=5, + metadata={"help": "decoder beam size"}, + ) + nbest: int = field( + default=1, + metadata={"help": "number of results to return"}, + ) + word_score: float = field( + default=1.0, + metadata={"help": "word score to add at end of word"}, + ) + unk_weight: float = field( + default=-math.inf, + metadata={"help": "unknown token weight"}, + ) + sil_weight: float = field( + default=0.0, + metadata={"help": "silence token weight"}, + ) + targets: Optional[str] = field( + default=None, + metadata={"help": "extension of ground truth labels to compute UER"}, + ) + results_path: Optional[str] = field( + default=None, + metadata={"help": "where to store results"}, + ) + post_process: Optional[str] = field( + default=None, + metadata={"help": "how to post process results"}, + ) + vocab_usage_power: float = field( + default=2, + metadata={"help": "for unsupervised param tuning"}, + ) + + viterbi_transcript: Optional[str] = field( + default=None, + metadata={"help": "for unsupervised param tuning"}, + ) + min_lm_ppl: float = field( + default=0, + metadata={"help": "for unsupervised param tuning"}, + ) + min_vt_uer: float = field( + default=0, + metadata={"help": "for unsupervised param tuning"}, + ) + + blank_weight: float = field( + default=0, + metadata={"help": "value to add or set for blank emission"}, + ) + blank_mode: str = field( + default="set", + metadata={ + "help": "can be add or set, how to modify blank emission with blank weight" + }, + ) + sil_is_blank: bool = field( + default=False, + metadata={"help": "if true, <SIL> token is same as blank token"}, + ) + + unsupervised_tuning: bool = field( + default=False, + metadata={ + "help": "if true, returns a score based on unsupervised param selection metric instead of UER" + }, + ) + is_ax: bool = field( + default=False, + metadata={ + "help": "if true, assumes we are using ax for tuning and returns a tuple for ax to consume" + }, + ) + + +def get_dataset_itr(cfg, task): + return task.get_batch_iterator( + dataset=task.dataset(cfg.fairseq.dataset.gen_subset), + max_tokens=cfg.fairseq.dataset.max_tokens, + max_sentences=cfg.fairseq.dataset.batch_size, + max_positions=(sys.maxsize, sys.maxsize), + ignore_invalid_inputs=cfg.fairseq.dataset.skip_invalid_size_inputs_valid_test, + required_batch_size_multiple=cfg.fairseq.dataset.required_batch_size_multiple, + num_shards=cfg.fairseq.dataset.num_shards, + shard_id=cfg.fairseq.dataset.shard_id, + num_workers=cfg.fairseq.dataset.num_workers, + data_buffer_size=cfg.fairseq.dataset.data_buffer_size, + ).next_epoch_itr(shuffle=False) + + +def process_predictions( + cfg: UnsupGenerateConfig, + hypos, + tgt_dict, + target_tokens, + res_files, +): + retval = [] + word_preds = [] + transcriptions = [] + dec_scores = [] + + for i, hypo in enumerate(hypos[: min(len(hypos), cfg.nbest)]): + if torch.is_tensor(hypo["tokens"]): + tokens = hypo["tokens"].int().cpu() + tokens = tokens[tokens >= tgt_dict.nspecial] + hyp_pieces = tgt_dict.string(tokens) + else: + hyp_pieces = " ".join(hypo["tokens"]) + + if "words" in hypo and len(hypo["words"]) > 0: + hyp_words = " ".join(hypo["words"]) + else: + hyp_words = post_process(hyp_pieces, cfg.post_process) + + to_write = {} + if res_files is not None: + to_write[res_files["hypo.units"]] = hyp_pieces + to_write[res_files["hypo.words"]] = hyp_words + + tgt_words = "" + if target_tokens is not None: + if isinstance(target_tokens, str): + tgt_pieces = tgt_words = target_tokens + else: + tgt_pieces = tgt_dict.string(target_tokens) + tgt_words = post_process(tgt_pieces, cfg.post_process) + + if res_files is not None: + to_write[res_files["ref.units"]] = tgt_pieces + to_write[res_files["ref.words"]] = tgt_words + + if not cfg.fairseq.common_eval.quiet: + logger.info(f"HYPO {i}:" + hyp_words) + if tgt_words: + logger.info("TARGET:" + tgt_words) + + if "am_score" in hypo and "lm_score" in hypo: + logger.info( + f"DECODER AM SCORE: {hypo['am_score']}, DECODER LM SCORE: {hypo['lm_score']}, DECODER SCORE: {hypo['score']}" + ) + elif "score" in hypo: + logger.info(f"DECODER SCORE: {hypo['score']}") + + logger.info("___________________") + + hyp_words_arr = hyp_words.split() + tgt_words_arr = tgt_words.split() + + retval.append( + ( + editdistance.eval(hyp_words_arr, tgt_words_arr), + len(hyp_words_arr), + len(tgt_words_arr), + hyp_pieces, + hyp_words, + ) + ) + word_preds.append(hyp_words_arr) + transcriptions.append(to_write) + dec_scores.append(-hypo.get("score", 0)) # negate cuz kaldi returns NLL + + if len(retval) > 1: + best = None + for r, t in zip(retval, transcriptions): + if best is None or r[0] < best[0][0]: + best = r, t + for dest, tran in best[1].items(): + print(tran, file=dest) + dest.flush() + return best[0] + + assert len(transcriptions) == 1 + for dest, tran in transcriptions[0].items(): + print(tran, file=dest) + + return retval[0] + + +def prepare_result_files(cfg: UnsupGenerateConfig): + def get_res_file(file_prefix): + if cfg.fairseq.dataset.num_shards > 1: + file_prefix = f"{cfg.fairseq.dataset.shard_id}_{file_prefix}" + path = os.path.join( + cfg.results_path, + "{}{}.txt".format( + cfg.fairseq.dataset.gen_subset, + file_prefix, + ), + ) + return open(path, "w", buffering=1) + + if not cfg.results_path: + return None + + return { + "hypo.words": get_res_file(""), + "hypo.units": get_res_file("_units"), + "ref.words": get_res_file("_ref"), + "ref.units": get_res_file("_ref_units"), + "hypo.nbest.words": get_res_file("_nbest_words"), + } + + +def optimize_models(cfg: UnsupGenerateConfig, use_cuda, models): + """Optimize ensemble for generation""" + for model in models: + model.eval() + if cfg.fairseq.common.fp16: + model.half() + if use_cuda: + model.cuda() + + +GenResult = namedtuple( + "GenResult", + [ + "count", + "errs_t", + "gen_timer", + "lengths_hyp_unit_t", + "lengths_hyp_t", + "lengths_t", + "lm_score_t", + "num_feats", + "num_sentences", + "num_symbols", + "vt_err_t", + "vt_length_t", + ], +) + + +def generate(cfg: UnsupGenerateConfig, models, saved_cfg, use_cuda): + task = tasks.setup_task(cfg.fairseq.task) + saved_cfg.task.labels = cfg.fairseq.task.labels + task.load_dataset(cfg.fairseq.dataset.gen_subset, task_cfg=saved_cfg.task) + # Set dictionary + tgt_dict = task.target_dictionary + logger.info( + "| {} {} {} examples".format( + cfg.fairseq.task.data, + cfg.fairseq.dataset.gen_subset, + len(task.dataset(cfg.fairseq.dataset.gen_subset)), + ) + ) + # Load dataset (possibly sharded) + itr = get_dataset_itr(cfg, task) + # Initialize generator + gen_timer = StopwatchMeter() + + def build_generator(cfg: UnsupGenerateConfig): + w2l_decoder = cfg.w2l_decoder + if w2l_decoder == DecoderType.VITERBI: + from examples.speech_recognition.w2l_decoder import W2lViterbiDecoder + + return W2lViterbiDecoder(cfg, task.target_dictionary) + elif w2l_decoder == DecoderType.KENLM: + from examples.speech_recognition.w2l_decoder import W2lKenLMDecoder + + return W2lKenLMDecoder(cfg, task.target_dictionary) + elif w2l_decoder == DecoderType.FAIRSEQ: + from examples.speech_recognition.w2l_decoder import W2lFairseqLMDecoder + + return W2lFairseqLMDecoder(cfg, task.target_dictionary) + elif w2l_decoder == DecoderType.KALDI: + from examples.speech_recognition.kaldi.kaldi_decoder import KaldiDecoder + + assert cfg.kaldi_decoder_config is not None + + return KaldiDecoder( + cfg.kaldi_decoder_config, + cfg.beam, + ) + else: + raise NotImplementedError( + "only wav2letter decoders with (viterbi, kenlm, fairseqlm) options are supported at the moment but found " + + str(w2l_decoder) + ) + + generator = build_generator(cfg) + + kenlm = None + fairseq_lm = None + if cfg.lm_model is not None: + import kenlm + + kenlm = kenlm.Model(cfg.lm_model) + + num_sentences = 0 + if cfg.results_path is not None and not os.path.exists(cfg.results_path): + os.makedirs(cfg.results_path) + + res_files = prepare_result_files(cfg) + errs_t = 0 + lengths_hyp_t = 0 + lengths_hyp_unit_t = 0 + lengths_t = 0 + count = 0 + num_feats = 0 + all_hyp_pieces = [] + all_hyp_words = [] + + num_symbols = ( + len([s for s in tgt_dict.symbols if not s.startswith("madeup")]) + - tgt_dict.nspecial + ) + targets = None + if cfg.targets is not None: + tgt_path = os.path.join( + cfg.fairseq.task.data, cfg.fairseq.dataset.gen_subset + "." + cfg.targets + ) + if os.path.exists(tgt_path): + with open(tgt_path, "r") as f: + targets = f.read().splitlines() + viterbi_transcript = None + if cfg.viterbi_transcript is not None and len(cfg.viterbi_transcript) > 0: + logger.info(f"loading viterbi transcript from {cfg.viterbi_transcript}") + with open(cfg.viterbi_transcript, "r") as vf: + viterbi_transcript = vf.readlines() + viterbi_transcript = [v.rstrip().split() for v in viterbi_transcript] + + gen_timer.start() + + start = 0 + end = len(itr) + + hypo_futures = None + if cfg.w2l_decoder == DecoderType.KALDI: + logger.info("Extracting features") + hypo_futures = [] + samples = [] + with progress_bar.build_progress_bar(cfg.fairseq.common, itr) as t: + for i, sample in enumerate(t): + if "net_input" not in sample or i < start or i >= end: + continue + if "padding_mask" not in sample["net_input"]: + sample["net_input"]["padding_mask"] = None + + hypos, num_feats = gen_hypos( + generator, models, num_feats, sample, task, use_cuda + ) + hypo_futures.append(hypos) + samples.append(sample) + if cfg.debug: + break + itr = list(zip(hypo_futures, samples)) + start = 0 + end = len(itr) + logger.info("Finished extracting features") + + with progress_bar.build_progress_bar(cfg.fairseq.common, itr) as t: + for i, sample in enumerate(t): + if i < start or i >= end: + continue + + if hypo_futures is not None: + hypos, sample = sample + hypos = [h.result() for h in hypos] + else: + if "net_input" not in sample: + continue + + hypos, num_feats = gen_hypos( + generator, models, num_feats, sample, task, use_cuda + ) + + for i, sample_id in enumerate(sample["id"].tolist()): + if targets is not None: + target_tokens = targets[sample_id] + elif "target" in sample or "target_label" in sample: + toks = ( + sample["target"][i, :] + if "target_label" not in sample + else sample["target_label"][i, :] + ) + + target_tokens = utils.strip_pad(toks, tgt_dict.pad()).int().cpu() + else: + target_tokens = None + + # Process top predictions + ( + errs, + length_hyp, + length, + hyp_pieces, + hyp_words, + ) = process_predictions( + cfg, + hypos[i], + tgt_dict, + target_tokens, + res_files, + ) + errs_t += errs + lengths_hyp_t += length_hyp + lengths_hyp_unit_t += ( + len(hyp_pieces) if len(hyp_pieces) > 0 else len(hyp_words) + ) + lengths_t += length + count += 1 + all_hyp_pieces.append(hyp_pieces) + all_hyp_words.append(hyp_words) + + num_sentences += ( + sample["nsentences"] if "nsentences" in sample else sample["id"].numel() + ) + + lm_score_sum = 0 + if kenlm is not None: + + if cfg.unit_lm: + lm_score_sum = sum(kenlm.score(w) for w in all_hyp_pieces) + else: + lm_score_sum = sum(kenlm.score(w) for w in all_hyp_words) + elif fairseq_lm is not None: + lm_score_sum = sum(fairseq_lm.score([h.split() for h in all_hyp_words])[0]) + + vt_err_t = 0 + vt_length_t = 0 + if viterbi_transcript is not None: + unit_hyps = [] + if cfg.targets is not None and cfg.lexicon is not None: + lex = {} + with open(cfg.lexicon, "r") as lf: + for line in lf: + items = line.rstrip().split() + lex[items[0]] = items[1:] + for h in all_hyp_pieces: + hyp_ws = [] + for w in h.split(): + assert w in lex, w + hyp_ws.extend(lex[w]) + unit_hyps.append(hyp_ws) + + else: + unit_hyps.extend([h.split() for h in all_hyp_words]) + + vt_err_t = sum( + editdistance.eval(vt, h) for vt, h in zip(viterbi_transcript, unit_hyps) + ) + + vt_length_t = sum(len(h) for h in viterbi_transcript) + + if res_files is not None: + for r in res_files.values(): + r.close() + + gen_timer.stop(lengths_hyp_t) + + return GenResult( + count, + errs_t, + gen_timer, + lengths_hyp_unit_t, + lengths_hyp_t, + lengths_t, + lm_score_sum, + num_feats, + num_sentences, + num_symbols, + vt_err_t, + vt_length_t, + ) + + +def gen_hypos(generator, models, num_feats, sample, task, use_cuda): + sample = utils.move_to_cuda(sample) if use_cuda else sample + + if "features" in sample["net_input"]: + sample["net_input"]["dense_x_only"] = True + num_feats += ( + sample["net_input"]["features"].shape[0] + * sample["net_input"]["features"].shape[1] + ) + hypos = task.inference_step(generator, models, sample, None) + return hypos, num_feats + + +def main(cfg: UnsupGenerateConfig, model=None): + if ( + cfg.fairseq.dataset.max_tokens is None + and cfg.fairseq.dataset.batch_size is None + ): + cfg.fairseq.dataset.max_tokens = 1024000 + + use_cuda = torch.cuda.is_available() and not cfg.fairseq.common.cpu + + task = tasks.setup_task(cfg.fairseq.task) + + overrides = ast.literal_eval(cfg.fairseq.common_eval.model_overrides) + + if cfg.fairseq.task._name == "gan_audio_pretraining_feats": + overrides["model"] = { + "blank_weight": cfg.blank_weight, + "blank_mode": cfg.blank_mode, + "blank_is_sil": cfg.sil_is_blank, + "no_softmax": True, + "segmentation": { + "type": "NONE", + }, + } + else: + overrides["model"] = { + "blank_weight": cfg.blank_weight, + "blank_mode": cfg.blank_mode, + } + + if model is None: + # Load ensemble + logger.info("| loading model(s) from {}".format(cfg.fairseq.common_eval.path)) + models, saved_cfg = checkpoint_utils.load_model_ensemble( + cfg.fairseq.common_eval.path.split("\\"), + arg_overrides=overrides, + task=task, + suffix=cfg.fairseq.checkpoint.checkpoint_suffix, + strict=(cfg.fairseq.checkpoint.checkpoint_shard_count == 1), + num_shards=cfg.fairseq.checkpoint.checkpoint_shard_count, + ) + optimize_models(cfg, use_cuda, models) + else: + models = [model] + saved_cfg = cfg.fairseq + + with open_dict(saved_cfg.task): + saved_cfg.task.shuffle = False + saved_cfg.task.sort_by_length = False + + gen_result = generate(cfg, models, saved_cfg, use_cuda) + + wer = None + if gen_result.lengths_t > 0: + wer = gen_result.errs_t * 100.0 / gen_result.lengths_t + logger.info(f"WER: {wer}") + + lm_ppl = float("inf") + + if gen_result.lm_score_t != 0 and gen_result.lengths_hyp_t > 0: + hyp_len = gen_result.lengths_hyp_t + lm_ppl = math.pow( + 10, -gen_result.lm_score_t / (hyp_len + gen_result.num_sentences) + ) + logger.info(f"LM PPL: {lm_ppl}") + + logger.info( + "| Processed {} sentences ({} tokens) in {:.1f}s ({:.2f}" + " sentences/s, {:.2f} tokens/s)".format( + gen_result.num_sentences, + gen_result.gen_timer.n, + gen_result.gen_timer.sum, + gen_result.num_sentences / gen_result.gen_timer.sum, + 1.0 / gen_result.gen_timer.avg, + ) + ) + + vt_diff = None + if gen_result.vt_length_t > 0: + vt_diff = gen_result.vt_err_t / gen_result.vt_length_t + vt_diff = max(cfg.min_vt_uer, vt_diff) + + lm_ppl = max(cfg.min_lm_ppl, lm_ppl) + + if not cfg.unsupervised_tuning == 0: + weighted_score = wer + else: + weighted_score = math.log(lm_ppl) * (vt_diff or 1.0) + + res = ( + f"| Generate {cfg.fairseq.dataset.gen_subset} with beam={cfg.beam}, " + f"lm_weight={cfg.kaldi_decoder_config.acoustic_scale if cfg.kaldi_decoder_config else cfg.lm_weight}, " + f"word_score={cfg.word_score}, sil_weight={cfg.sil_weight}, blank_weight={cfg.blank_weight}, " + f"WER: {wer}, LM_PPL: {lm_ppl}, num feats: {gen_result.num_feats}, " + f"length: {gen_result.lengths_hyp_t}, UER to viterbi: {(vt_diff or 0) * 100}, score: {weighted_score}" + ) + + logger.info(res) + # print(res) + + return task, weighted_score + + +@hydra.main( + config_path=os.path.join("../../..", "fairseq", "config"), config_name="config" +) +def hydra_main(cfg): + with open_dict(cfg): + # make hydra logging work with ddp (see # see https://github.com/facebookresearch/hydra/issues/1126) + cfg.job_logging_cfg = OmegaConf.to_container( + HydraConfig.get().job_logging, resolve=True + ) + + cfg = OmegaConf.create( + OmegaConf.to_container(cfg, resolve=False, enum_to_str=False) + ) + OmegaConf.set_struct(cfg, True) + logger.info(cfg) + + utils.import_user_module(cfg.fairseq.common) + + _, score = main(cfg) + + if cfg.is_ax: + return score, None + return score + + +def cli_main(): + try: + from hydra._internal.utils import get_args + + cfg_name = get_args().config_name or "config" + except: + logger.warning("Failed to get config name from hydra args") + cfg_name = "config" + + cs = ConfigStore.instance() + cs.store(name=cfg_name, node=UnsupGenerateConfig) + hydra_main() + + +if __name__ == "__main__": + cli_main() diff --git a/examples/wav2vec/vq-wav2vec_featurize.py b/examples/wav2vec/vq-wav2vec_featurize.py new file mode 100644 index 0000000000000000000000000000000000000000..627072ee174c22831209e00984b945eb9dc2c279 --- /dev/null +++ b/examples/wav2vec/vq-wav2vec_featurize.py @@ -0,0 +1,250 @@ +#!/usr/bin/env python3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +""" +Helper script to pre-compute embeddings for a flashlight (previously called wav2letter++) dataset +""" + +import argparse +import glob +import os +import os.path as osp +import pprint + +import soundfile as sf +import torch +import fairseq +from torch import nn +from torch.utils.data import DataLoader + + +try: + import tqdm +except: + print("Install tqdm to use --log-format=tqdm") + + +class FilesDataset: + def __init__(self, files, labels): + self.files = files + if labels and osp.exists(labels): + with open(labels, "r") as lbl_f: + self.labels = [line.rstrip() for line in lbl_f] + else: + self.labels = labels + + def __len__(self): + return len(self.files) + + def __getitem__(self, index): + fname = self.files[index] + + wav, sr = sf.read(fname) + assert sr == 16000 + + wav = torch.from_numpy(wav).float() + lbls = None + if self.labels: + if isinstance(self.labels, str): + lbl_file = osp.splitext(fname)[0] + "." + self.labels + with open(lbl_file, "r") as lblf: + lbls = lblf.readline() + assert lbls is not None + else: + lbls = self.labels[index] + return wav, lbls + + def collate(self, batch): + return batch + + +class ArgTypes: + @staticmethod + def existing_path(arg): + arg = str(arg) + assert osp.exists(arg), f"File {arg} does not exist" + return arg + + @staticmethod + def mkdir(arg): + arg = str(arg) + os.makedirs(arg, exist_ok=True) + return arg + + +class DatasetWriter: + def __init__(self): + + self.args = self.load_config() + pprint.pprint(self.args.__dict__) + + self.model = self.load_model() + + def __getattr__(self, attr): + return getattr(self.args, attr) + + def read_manifest(self, fname): + + with open(fname, "r") as fp: + lines = fp.read().split("\n") + root = lines.pop(0).strip() + fnames = [ + osp.join(root, line.split("\t")[0]) for line in lines if len(line) > 0 + ] + + return fnames + + def process_splits(self): + + if self.args.shard is not None or self.args.num_shards is not None: + assert self.args.shard is not None and self.args.num_shards is not None + + for split in self.splits: + print(split) + + if self.extension == "tsv": + datadir = osp.join(self.data_dir, f"{split}.{self.extension}") + print("Reading manifest file: ", datadir) + files = self.read_manifest(datadir) + else: + datadir = osp.join(self.data_dir, split, f"**/*.{self.extension}") + files = glob.glob(datadir, recursive=True) + + assert len(files) > 0 + + if self.args.shard is not None: + files = files[self.args.shard :: self.args.num_shards] + + lbls = [] + with open(self.data_file(split), "w") as srcf: + for line, lbl in self.iterate(files): + print(line, file=srcf) + if self.args.labels: + lbls.append(lbl + "\n") + + if self.args.labels: + assert all(a is not None for a in lbls) + with open(self.lbl_file(split), "w") as lblf: + lblf.writelines(lbls) + + def iterate(self, files): + + data = self.load_data(files) + for samples in tqdm.tqdm(data, total=len(files) // 32): + + for wav, lbl in samples: + x = wav.unsqueeze(0).float().cuda() + + div = 1 + while x.size(-1) // div > self.args.max_size: + div += 1 + + xs = x.chunk(div, dim=-1) + + result = [] + for x in xs: + torch.cuda.empty_cache() + x = self.model.feature_extractor(x) + if self.quantize_location == "encoder": + with torch.no_grad(): + _, idx = self.model.vector_quantizer.forward_idx(x) + idx = idx.squeeze(0).cpu() + else: + with torch.no_grad(): + z = self.model.feature_aggregator(x) + _, idx = self.model.vector_quantizer.forward_idx(z) + idx = idx.squeeze(0).cpu() + result.append(idx) + + idx = torch.cat(result, dim=0) + yield " ".join("-".join(map(str, a.tolist())) for a in idx), lbl + + def lbl_file(self, name): + shard_part = "" if self.args.shard is None else f".{self.args.shard}" + return osp.join(self.output_dir, f"{name}.lbl{shard_part}") + + def data_file(self, name): + shard_part = "" if self.args.shard is None else f".{self.args.shard}" + return osp.join(self.output_dir, f"{name}.src{shard_part}") + + def var_file(self): + return osp.join(self.output_dir, f"vars.pt") + + def load_config(self): + + parser = argparse.ArgumentParser("Vector Quantized wav2vec features") + + # Model Arguments + parser.add_argument("--checkpoint", type=ArgTypes.existing_path, required=True) + parser.add_argument("--data-parallel", action="store_true") + + # Output Arguments + parser.add_argument("--output-dir", type=ArgTypes.mkdir, required=True) + + # Data Arguments + parser.add_argument("--data-dir", type=ArgTypes.existing_path, required=True) + parser.add_argument("--splits", type=str, nargs="+", required=True) + parser.add_argument("--extension", type=str, required=True) + parser.add_argument("--labels", type=str, required=False) + + parser.add_argument("--shard", type=int, default=None) + parser.add_argument("--num-shards", type=int, default=None) + parser.add_argument("--max-size", type=int, default=1300000) + + # Logger Arguments + parser.add_argument( + "--log-format", type=str, choices=["none", "simple", "tqdm"] + ) + + return parser.parse_args() + + def load_data(self, fnames): + + dataset = FilesDataset(fnames, self.args.labels) + loader = DataLoader( + dataset, batch_size=32, collate_fn=dataset.collate, num_workers=8 + ) + return loader + + def load_model(self): + model, cfg, task = fairseq.checkpoint_utils.load_model_ensemble_and_task([self.checkpoint]) + model = model[0] + + self.quantize_location = getattr(cfg.model, "vq", "encoder") + + model.eval().float() + model.cuda() + + if self.data_parallel: + model = nn.DataParallel(model) + + return model + + def __call__(self): + + self.process_splits() + + if hasattr(self.model.feature_extractor, "vars") and ( + self.args.shard is None or self.args.shard == 0 + ): + vars = ( + self.model.feature_extractor.vars.view( + self.model.feature_extractor.banks, + self.model.feature_extractor.num_vars, + -1, + ) + .cpu() + .detach() + ) + print("writing learned latent variable embeddings: ", vars.shape) + torch.save(vars, self.var_file()) + + +if __name__ == "__main__": + write_data = DatasetWriter() + + write_data() + print("Done.") diff --git a/examples/wav2vec/wav2vec_featurize.py b/examples/wav2vec/wav2vec_featurize.py new file mode 100644 index 0000000000000000000000000000000000000000..588268b7080cbd3400ac144604b2d75cef2876dd --- /dev/null +++ b/examples/wav2vec/wav2vec_featurize.py @@ -0,0 +1,249 @@ +#!/usr/bin/env python3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +""" +Helper script to pre-compute embeddings for a flashlight (previously called wav2letter++) dataset +""" + +import argparse +import glob +import os +from shutil import copy + +import h5py +import numpy as np +import soundfile as sf +import torch +import tqdm +import fairseq +from torch import nn + + +def read_audio(fname): + """ Load an audio file and return PCM along with the sample rate """ + + wav, sr = sf.read(fname) + assert sr == 16e3 + + return wav, 16e3 + + +class PretrainedWav2VecModel(nn.Module): + def __init__(self, fname): + super().__init__() + + model, cfg, task = fairseq.checkpoint_utils.load_model_ensemble_and_task([fname]) + model = model[0] + model.eval() + + self.model = model + + def forward(self, x): + with torch.no_grad(): + z = self.model.feature_extractor(x) + if isinstance(z, tuple): + z = z[0] + c = self.model.feature_aggregator(z) + return z, c + + +class EmbeddingWriterConfig(argparse.ArgumentParser): + def __init__(self): + super().__init__("Pre-compute embeddings for flashlight datasets") + + kwargs = {"action": "store", "type": str, "required": True} + + self.add_argument("--input", "-i", help="Input Directory", **kwargs) + self.add_argument("--output", "-o", help="Output Directory", **kwargs) + self.add_argument("--model", help="Path to model checkpoint", **kwargs) + self.add_argument("--split", help="Dataset Splits", nargs="+", **kwargs) + self.add_argument( + "--ext", default="wav", required=False, help="Audio file extension" + ) + + self.add_argument( + "--no-copy-labels", + action="store_true", + help="Do not copy label files. Useful for large datasets, use --targetdir in flashlight then.", + ) + self.add_argument( + "--use-feat", + action="store_true", + help="Use the feature vector ('z') instead of context vector ('c') for features", + ) + self.add_argument("--gpu", help="GPU to use", default=0, type=int) + + +class Prediction: + """ Lightweight wrapper around a fairspeech embedding model """ + + def __init__(self, fname, gpu=0): + self.gpu = gpu + self.model = PretrainedWav2VecModel(fname).cuda(gpu) + + def __call__(self, x): + x = torch.from_numpy(x).float().cuda(self.gpu) + with torch.no_grad(): + z, c = self.model(x.unsqueeze(0)) + + return z.squeeze(0).cpu().numpy(), c.squeeze(0).cpu().numpy() + + +class H5Writer: + """ Write features as hdf5 file in flashlight compatible format """ + + def __init__(self, fname): + self.fname = fname + os.makedirs(os.path.dirname(self.fname), exist_ok=True) + + def write(self, data): + channel, T = data.shape + + with h5py.File(self.fname, "w") as out_ds: + data = data.T.flatten() + out_ds["features"] = data + out_ds["info"] = np.array([16e3 // 160, T, channel]) + + +class EmbeddingDatasetWriter(object): + """Given a model and a flashlight dataset, pre-compute and store embeddings + + Args: + input_root, str : + Path to the flashlight dataset + output_root, str : + Desired output directory. Will be created if non-existent + split, str : + Dataset split + """ + + def __init__( + self, + input_root, + output_root, + split, + model_fname, + extension="wav", + gpu=0, + verbose=False, + use_feat=False, + ): + + assert os.path.exists(model_fname) + + self.model_fname = model_fname + self.model = Prediction(self.model_fname, gpu) + + self.input_root = input_root + self.output_root = output_root + self.split = split + self.verbose = verbose + self.extension = extension + self.use_feat = use_feat + + assert os.path.exists(self.input_path), "Input path '{}' does not exist".format( + self.input_path + ) + + def _progress(self, iterable, **kwargs): + if self.verbose: + return tqdm.tqdm(iterable, **kwargs) + return iterable + + def require_output_path(self, fname=None): + path = self.get_output_path(fname) + os.makedirs(path, exist_ok=True) + + @property + def input_path(self): + return self.get_input_path() + + @property + def output_path(self): + return self.get_output_path() + + def get_input_path(self, fname=None): + if fname is None: + return os.path.join(self.input_root, self.split) + return os.path.join(self.get_input_path(), fname) + + def get_output_path(self, fname=None): + if fname is None: + return os.path.join(self.output_root, self.split) + return os.path.join(self.get_output_path(), fname) + + def copy_labels(self): + self.require_output_path() + + labels = list( + filter( + lambda x: self.extension not in x, glob.glob(self.get_input_path("*")) + ) + ) + for fname in tqdm.tqdm(labels): + copy(fname, self.output_path) + + @property + def input_fnames(self): + return sorted(glob.glob(self.get_input_path("*.{}".format(self.extension)))) + + def __len__(self): + return len(self.input_fnames) + + def write_features(self): + + paths = self.input_fnames + + fnames_context = map( + lambda x: os.path.join( + self.output_path, x.replace("." + self.extension, ".h5context") + ), + map(os.path.basename, paths), + ) + + for name, target_fname in self._progress( + zip(paths, fnames_context), total=len(self) + ): + wav, sr = read_audio(name) + z, c = self.model(wav) + feat = z if self.use_feat else c + writer = H5Writer(target_fname) + writer.write(feat) + + def __repr__(self): + + return "EmbeddingDatasetWriter ({n_files} files)\n\tinput:\t{input_root}\n\toutput:\t{output_root}\n\tsplit:\t{split})".format( + n_files=len(self), **self.__dict__ + ) + + +if __name__ == "__main__": + + args = EmbeddingWriterConfig().parse_args() + + for split in args.split: + + writer = EmbeddingDatasetWriter( + input_root=args.input, + output_root=args.output, + split=split, + model_fname=args.model, + gpu=args.gpu, + extension=args.ext, + use_feat=args.use_feat, + ) + + print(writer) + writer.require_output_path() + + print("Writing Features...") + writer.write_features() + print("Done.") + + if not args.no_copy_labels: + print("Copying label data...") + writer.copy_labels() + print("Done.") diff --git a/examples/wav2vec/wav2vec_manifest.py b/examples/wav2vec/wav2vec_manifest.py new file mode 100644 index 0000000000000000000000000000000000000000..9b8aa180e88d9ee98bdca7089aed5046ec0d9cb9 --- /dev/null +++ b/examples/wav2vec/wav2vec_manifest.py @@ -0,0 +1,87 @@ +#!/usr/bin/env python3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +Data pre-processing: build vocabularies and binarize training data. +""" + +import argparse +import glob +import os +import random + +import soundfile + + +def get_parser(): + parser = argparse.ArgumentParser() + parser.add_argument( + "root", metavar="DIR", help="root directory containing flac files to index" + ) + parser.add_argument( + "--valid-percent", + default=0.01, + type=float, + metavar="D", + help="percentage of data to use as validation set (between 0 and 1)", + ) + parser.add_argument( + "--dest", default=".", type=str, metavar="DIR", help="output directory" + ) + parser.add_argument( + "--ext", default="flac", type=str, metavar="EXT", help="extension to look for" + ) + parser.add_argument("--seed", default=42, type=int, metavar="N", help="random seed") + parser.add_argument( + "--path-must-contain", + default=None, + type=str, + metavar="FRAG", + help="if set, path must contain this substring for a file to be included in the manifest", + ) + return parser + + +def main(args): + assert args.valid_percent >= 0 and args.valid_percent <= 1.0 + + if not os.path.exists(args.dest): + os.makedirs(args.dest) + + dir_path = os.path.realpath(args.root) + search_path = os.path.join(dir_path, "**/*." + args.ext) + rand = random.Random(args.seed) + + valid_f = ( + open(os.path.join(args.dest, "valid.tsv"), "w") + if args.valid_percent > 0 + else None + ) + + with open(os.path.join(args.dest, "train.tsv"), "w") as train_f: + print(dir_path, file=train_f) + + if valid_f is not None: + print(dir_path, file=valid_f) + + for fname in glob.iglob(search_path, recursive=True): + file_path = os.path.realpath(fname) + + if args.path_must_contain and args.path_must_contain not in file_path: + continue + + frames = soundfile.info(fname).frames + dest = train_f if rand.random() > args.valid_percent else valid_f + print( + "{}\t{}".format(os.path.relpath(file_path, dir_path), frames), file=dest + ) + if valid_f is not None: + valid_f.close() + + +if __name__ == "__main__": + parser = get_parser() + args = parser.parse_args() + main(args) diff --git a/examples/wmt19/README.md b/examples/wmt19/README.md new file mode 100644 index 0000000000000000000000000000000000000000..5c90d0e6c4ae8d043ca622e70c5828dca6f9c2f2 --- /dev/null +++ b/examples/wmt19/README.md @@ -0,0 +1,85 @@ +# WMT 19 + +This page provides pointers to the models of Facebook-FAIR's WMT'19 news translation task submission [(Ng et al., 2019)](https://arxiv.org/abs/1907.06616). + +## Pre-trained models + +Model | Description | Download +---|---|--- +`transformer.wmt19.en-de` | En->De Ensemble | [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/wmt19.en-de.joined-dict.ensemble.tar.gz) +`transformer.wmt19.de-en` | De->En Ensemble | [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/wmt19.de-en.joined-dict.ensemble.tar.gz) +`transformer.wmt19.en-ru` | En->Ru Ensemble | [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/wmt19.en-ru.ensemble.tar.gz) +`transformer.wmt19.ru-en` | Ru->En Ensemble | [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/wmt19.ru-en.ensemble.tar.gz) +`transformer_lm.wmt19.en` | En Language Model | [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/lm/wmt19.en.tar.gz) +`transformer_lm.wmt19.de` | De Language Model | [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/lm/wmt19.de.tar.gz) +`transformer_lm.wmt19.ru` | Ru Language Model | [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/lm/wmt19.ru.tar.gz) + +## Pre-trained single models before finetuning + +Model | Description | Download +---|---|--- +`transformer.wmt19.en-de` | En->De Single, no finetuning | [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/wmt19.en-de.ffn8192.tar.gz) +`transformer.wmt19.de-en` | De->En Single, no finetuning | [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/wmt19.de-en.ffn8192.tar.gz) +`transformer.wmt19.en-ru` | En->Ru Single, no finetuning | [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/wmt19.en-ru.ffn8192.tar.gz) +`transformer.wmt19.ru-en` | Ru->En Single, no finetuning | [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/wmt19.ru-en.ffn8192.tar.gz) + +## Example usage (torch.hub) + +#### Requirements + +We require a few additional Python dependencies for preprocessing: +```bash +pip install fastBPE sacremoses +``` + +#### Translation + +```python +import torch + +# English to German translation +en2de = torch.hub.load('pytorch/fairseq', 'transformer.wmt19.en-de', checkpoint_file='model1.pt:model2.pt:model3.pt:model4.pt', + tokenizer='moses', bpe='fastbpe') +en2de.translate("Machine learning is great!") # 'Maschinelles Lernen ist großartig!' + +# German to English translation +de2en = torch.hub.load('pytorch/fairseq', 'transformer.wmt19.de-en', checkpoint_file='model1.pt:model2.pt:model3.pt:model4.pt', + tokenizer='moses', bpe='fastbpe') +de2en.translate("Maschinelles Lernen ist großartig!") # 'Machine learning is great!' + +# English to Russian translation +en2ru = torch.hub.load('pytorch/fairseq', 'transformer.wmt19.en-ru', checkpoint_file='model1.pt:model2.pt:model3.pt:model4.pt', + tokenizer='moses', bpe='fastbpe') +en2ru.translate("Machine learning is great!") # 'Машинное обучение - это здорово!' + +# Russian to English translation +ru2en = torch.hub.load('pytorch/fairseq', 'transformer.wmt19.ru-en', checkpoint_file='model1.pt:model2.pt:model3.pt:model4.pt', + tokenizer='moses', bpe='fastbpe') +ru2en.translate("Машинное обучение - это здорово!") # 'Machine learning is great!' +``` + +#### Language Modeling + +```python +# Sample from the English LM +en_lm = torch.hub.load('pytorch/fairseq', 'transformer_lm.wmt19.en', tokenizer='moses', bpe='fastbpe') +en_lm.sample("Machine learning is") # 'Machine learning is the future of computing, says Microsoft boss Satya Nadella ...' + +# Sample from the German LM +de_lm = torch.hub.load('pytorch/fairseq', 'transformer_lm.wmt19.de', tokenizer='moses', bpe='fastbpe') +de_lm.sample("Maschinelles lernen ist") # 'Maschinelles lernen ist das A und O (neues-deutschland.de) Die Arbeitsbedingungen für Lehrerinnen und Lehrer sind seit Jahren verbesserungswürdig ...' + +# Sample from the Russian LM +ru_lm = torch.hub.load('pytorch/fairseq', 'transformer_lm.wmt19.ru', tokenizer='moses', bpe='fastbpe') +ru_lm.sample("машинное обучение это") # 'машинное обучение это то, что мы называем "искусственным интеллектом".' +``` + +## Citation +```bibtex +@inproceedings{ng2019facebook}, + title = {Facebook FAIR's WMT19 News Translation Task Submission}, + author = {Ng, Nathan and Yee, Kyra and Baevski, Alexei and Ott, Myle and Auli, Michael and Edunov, Sergey}, + booktitle = {Proc. of WMT}, + year = 2019, +} +``` diff --git a/examples/wmt20/README.md b/examples/wmt20/README.md new file mode 100644 index 0000000000000000000000000000000000000000..b4f2874652f8be19998a65faa1d9276d8017ec59 --- /dev/null +++ b/examples/wmt20/README.md @@ -0,0 +1,72 @@ +# WMT 20 + +This page provides pointers to the models of Facebook-FAIR's WMT'20 news translation task submission [(Chen et al., 2020)](https://arxiv.org/abs/2011.08298). + +## Single best MT models (after finetuning on part of WMT20 news dev set) + +Model | Description | Download +---|---|--- +`transformer.wmt20.ta-en` | Ta->En | [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/wmt20.ta-en.single.tar.gz) +`transformer.wmt20.en-ta` | En->Ta | [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/wmt20.en-ta.single.tar.gz) +`transformer.wmt20.iu-en.news` | Iu->En (News domain) | [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/wmt20.iu-en.news.single.tar.gz) +`transformer.wmt20.en-iu.news` | En->Iu (News domain) | [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/wmt20.en-iu.news.single.tar.gz) +`transformer.wmt20.iu-en.nh` | Iu->En (Nunavut Hansard domain) | [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/wmt20.iu-en.nh.single.tar.gz) +`transformer.wmt20.en-iu.nh` | En->Iu (Nunavut Hansard domain) | [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/wmt20.en-iu.nh.single.tar.gz) + +## Language models +Model | Description | Download +---|---|--- +`transformer_lm.wmt20.en` | En Language Model | [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/wmt20.en.tar.gz) +`transformer_lm.wmt20.ta` | Ta Language Model | [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/wmt20.ta.tar.gz) +`transformer_lm.wmt20.iu.news` | Iu Language Model (News domain) | [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/wmt20.iu.news.tar.gz) +`transformer_lm.wmt20.iu.nh` | Iu Language Model (Nunavut Hansard domain) | [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/wmt20.iu.nh.tar.gz) + +## Example usage (torch.hub) + +#### Translation + +```python +import torch + +# English to Tamil translation +en2ta = torch.hub.load('pytorch/fairseq', 'transformer.wmt20.en-ta') +en2ta.translate("Machine learning is great!") # 'இயந்திரக் கற்றல் அருமை!' + +# Tamil to English translation +ta2en = torch.hub.load('pytorch/fairseq', 'transformer.wmt20.ta-en') +ta2en.translate("இயந்திரக் கற்றல் அருமை!") # 'Machine learning is great!' + +# English to Inuktitut translation +en2iu = torch.hub.load('pytorch/fairseq', 'transformer.wmt20.en-iu.news') +en2iu.translate("machine learning is great!") # 'ᖃᒧᑕᐅᔭᓄᑦ ᐃᓕᓐᓂᐊᕐᓂᖅ ᐱᐅᔪᒻᒪᕆᒃ!' + +# Inuktitut to English translation +iu2en = torch.hub.load('pytorch/fairseq', 'transformer.wmt20.iu-en.news') +iu2en.translate("ᖃᒧᑕᐅᔭᓄᑦ ᐃᓕᓐᓂᐊᕐᓂᖅ ᐱᐅᔪᒻᒪᕆᒃ!") # 'Machine learning excellence!' +``` + +#### Language Modeling + +```python +# Sample from the English LM +en_lm = torch.hub.load('pytorch/fairseq', 'transformer_lm.wmt20.en') +en_lm.sample("Machine learning is") # 'Machine learning is a type of artificial intelligence that uses machine learning to learn from data and make predictions.' + +# Sample from the Tamil LM +ta_lm = torch.hub.load('pytorch/fairseq', 'transformer_lm.wmt20.ta') +ta_lm.sample("இயந்திரக் கற்றல் என்பது செயற்கை நுண்ணறிவின்") # 'இயந்திரக் கற்றல் என்பது செயற்கை நுண்ணறிவின் ஒரு பகுதியாகும்.' + +# Sample from the Inuktitut LM +iu_lm = torch.hub.load('pytorch/fairseq', 'transformer_lm.wmt20.iu.news') +iu_lm.sample("ᖃᒧᑕᐅᔭᓄᑦ ᐃᓕᓐᓂᐊᕐᓂᖅ") # 'ᖃᒧᑕᐅᔭᓄᑦ ᐃᓕᓐᓂᐊᕐᓂᖅ, ᐊᒻᒪᓗ ᓯᓚᐅᑉ ᐊᓯᙳᖅᐸᓪᓕᐊᓂᖓᓄᑦ ᖃᓄᐃᓕᐅᕈᑎᒃᓴᑦ, ᐃᓚᖃᖅᖢᑎᒃ ᐅᑯᓂᖓ:' +``` + +## Citation +```bibtex +@inproceedings{chen2020facebook + title={Facebook AI's WMT20 News Translation Task Submission}, + author={Peng-Jen Chen and Ann Lee and Changhan Wang and Naman Goyal and Angela Fan and Mary Williamson and Jiatao Gu}, + booktitle={Proc. of WMT}, + year={2020}, +} +``` diff --git a/examples/xlmr/README.md b/examples/xlmr/README.md new file mode 100644 index 0000000000000000000000000000000000000000..b95bfe15d3fe6d03951453679135c2e9187d73c7 --- /dev/null +++ b/examples/xlmr/README.md @@ -0,0 +1,144 @@ +# Unsupervised Cross-lingual Representation Learning at Scale (XLM-RoBERTa) +https://arxiv.org/pdf/1911.02116.pdf + +# Larger-Scale Transformers for Multilingual Masked Language Modeling +https://arxiv.org/pdf/2105.00572.pdf + + +## What's New: +- June 2021: `XLMR-XL` AND `XLMR-XXL` models released. + +## Introduction + +`XLM-R` (`XLM-RoBERTa`) is a generic cross lingual sentence encoder that obtains state-of-the-art results on many cross-lingual understanding (XLU) benchmarks. It is trained on `2.5T` of filtered CommonCrawl data in 100 languages (list below). + + Language | Language|Language |Language | Language +---|---|---|---|--- +Afrikaans | Albanian | Amharic | Arabic | Armenian +Assamese | Azerbaijani | Basque | Belarusian | Bengali +Bengali Romanize | Bosnian | Breton | Bulgarian | Burmese +Burmese zawgyi font | Catalan | Chinese (Simplified) | Chinese (Traditional) | Croatian +Czech | Danish | Dutch | English | Esperanto +Estonian | Filipino | Finnish | French | Galician +Georgian | German | Greek | Gujarati | Hausa +Hebrew | Hindi | Hindi Romanize | Hungarian | Icelandic +Indonesian | Irish | Italian | Japanese | Javanese +Kannada | Kazakh | Khmer | Korean | Kurdish (Kurmanji) +Kyrgyz | Lao | Latin | Latvian | Lithuanian +Macedonian | Malagasy | Malay | Malayalam | Marathi +Mongolian | Nepali | Norwegian | Oriya | Oromo +Pashto | Persian | Polish | Portuguese | Punjabi +Romanian | Russian | Sanskrit | Scottish Gaelic | Serbian +Sindhi | Sinhala | Slovak | Slovenian | Somali +Spanish | Sundanese | Swahili | Swedish | Tamil +Tamil Romanize | Telugu | Telugu Romanize | Thai | Turkish +Ukrainian | Urdu | Urdu Romanize | Uyghur | Uzbek +Vietnamese | Welsh | Western Frisian | Xhosa | Yiddish + +## Pre-trained models + +Model | Description | #params | vocab size | Download +---|---|---|---|--- +`xlmr.base` | XLM-R using the BERT-base architecture | 250M | 250k | [xlm.base.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/xlmr.base.tar.gz) +`xlmr.large` | XLM-R using the BERT-large architecture | 560M | 250k | [xlm.large.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/xlmr.large.tar.gz) +`xlmr.xl` | XLM-R (`layers=36, model_dim=2560`) | 3.5B | 250k | [xlm.xl.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/xlmr/xlmr.xl.tar.gz) +`xlmr.xxl` | XLM-R (`layers=48, model_dim=4096`) | 10.7B | 250k | [xlm.xxl.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/xlmr/xlmr.xxl.tar.gz) + +## Results + +**[XNLI (Conneau et al., 2018)](https://arxiv.org/abs/1809.05053)** + +Model | average | en | fr | es | de | el | bg | ru | tr | ar | vi | th | zh | hi | sw | ur +---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|--- +`roberta.large.mnli` _(TRANSLATE-TEST)_ | 77.8 | 91.3 | 82.9 | 84.3 | 81.2 | 81.7 | 83.1 | 78.3 | 76.8 | 76.6 | 74.2 | 74.1 | 77.5 | 70.9 | 66.7 | 66.8 +`xlmr.large` _(TRANSLATE-TRAIN-ALL)_ | 83.6 | 89.1 | 85.1 | 86.6 | 85.7 | 85.3 | 85.9 | 83.5 | 83.2 | 83.1 | 83.7 | 81.5 | 83.7 | 81.6 | 78.0 | 78.1 +`xlmr.xl` _(TRANSLATE-TRAIN-ALL)_ | 85.4 | 91.1 | 87.2 | 88.1 | 87.0 | 87.4 | 87.8 | 85.3 | 85.2 | 85.3 | 86.2 | 83.8 | 85.3 | 83.1 | 79.8 | 78.2 | 85.4 +`xlmr.xxl` _(TRANSLATE-TRAIN-ALL)_ | 86.0 | 91.5 | 87.6 | 88.7 | 87.8 | 87.4 | 88.2 | 85.6 | 85.1 | 85.8 | 86.3 | 83.9 | 85.6 | 84.6 | 81.7 | 80.6 + +**[MLQA (Lewis et al., 2018)](https://arxiv.org/abs/1910.07475)** + +Model | average | en | es | de | ar | hi | vi | zh +---|---|---|---|---|---|---|---|--- +`BERT-large` | - | 80.2/67.4 | - | - | - | - | - | - +`mBERT` | 57.7 / 41.6 | 77.7 / 65.2 | 64.3 / 46.6 | 57.9 / 44.3 | 45.7 / 29.8| 43.8 / 29.7 | 57.1 / 38.6 | 57.5 / 37.3 +`xlmr.large` | 70.7 / 52.7 | 80.6 / 67.8 | 74.1 / 56.0 | 68.5 / 53.6 | 63.1 / 43.5 | 69.2 / 51.6 | 71.3 / 50.9 | 68.0 / 45.4 +`xlmr.xl` | 73.4 / 55.3 | 85.1 / 72.6 | 66.7 / 46.2 | 70.5 / 55.5 | 74.3 / 56.9 | 72.2 / 54.7 | 74.4 / 52.9 | 70.9 / 48.5 +`xlmr.xxl` | 74.8 / 56.6 | 85.5 / 72.4 | 68.6 / 48.4 | 72.7 / 57.8 | 75.4 / 57.6 | 73.7 / 55.8 | 76.0 / 55.0 | 71.7 / 48.9 + + +## Example usage + +##### Load XLM-R from torch.hub (PyTorch >= 1.1): +```python +import torch +xlmr = torch.hub.load('pytorch/fairseq', 'xlmr.large') +xlmr.eval() # disable dropout (or leave in train mode to finetune) +``` + +##### Load XLM-R (for PyTorch 1.0 or custom models): +```python +# Download xlmr.large model +wget https://dl.fbaipublicfiles.com/fairseq/models/xlmr.large.tar.gz +tar -xzvf xlmr.large.tar.gz + +# Load the model in fairseq +from fairseq.models.roberta import XLMRModel +xlmr = XLMRModel.from_pretrained('/path/to/xlmr.large', checkpoint_file='model.pt') +xlmr.eval() # disable dropout (or leave in train mode to finetune) +``` + +##### Apply sentence-piece-model (SPM) encoding to input text: +```python +en_tokens = xlmr.encode('Hello world!') +assert en_tokens.tolist() == [0, 35378, 8999, 38, 2] +xlmr.decode(en_tokens) # 'Hello world!' + +zh_tokens = xlmr.encode('你好，世界') +assert zh_tokens.tolist() == [0, 6, 124084, 4, 3221, 2] +xlmr.decode(zh_tokens) # '你好，世界' + +hi_tokens = xlmr.encode('नमस्ते दुनिया') +assert hi_tokens.tolist() == [0, 68700, 97883, 29405, 2] +xlmr.decode(hi_tokens) # 'नमस्ते दुनिया' + +ar_tokens = xlmr.encode('مرحبا بالعالم') +assert ar_tokens.tolist() == [0, 665, 193478, 258, 1705, 77796, 2] +xlmr.decode(ar_tokens) # 'مرحبا بالعالم' + +fr_tokens = xlmr.encode('Bonjour le monde') +assert fr_tokens.tolist() == [0, 84602, 95, 11146, 2] +xlmr.decode(fr_tokens) # 'Bonjour le monde' +``` + +##### Extract features from XLM-R: +```python +# Extract the last layer's features +last_layer_features = xlmr.extract_features(zh_tokens) +assert last_layer_features.size() == torch.Size([1, 6, 1024]) + +# Extract all layer's features (layer 0 is the embedding layer) +all_layers = xlmr.extract_features(zh_tokens, return_all_hiddens=True) +assert len(all_layers) == 25 +assert torch.all(all_layers[-1] == last_layer_features) +``` + +## Citation + +```bibtex +@article{conneau2019unsupervised, + title={Unsupervised Cross-lingual Representation Learning at Scale}, + author={Conneau, Alexis and Khandelwal, Kartikay and Goyal, Naman and Chaudhary, Vishrav and Wenzek, Guillaume and Guzm{\'a}n, Francisco and Grave, Edouard and Ott, Myle and Zettlemoyer, Luke and Stoyanov, Veselin}, + journal={arXiv preprint arXiv:1911.02116}, + year={2019} +} +``` + + +```bibtex +@article{goyal2021larger, + title={Larger-Scale Transformers for Multilingual Masked Language Modeling}, + author={Goyal, Naman and Du, Jingfei and Ott, Myle and Anantharaman, Giri and Conneau, Alexis}, + journal={arXiv preprint arXiv:2105.00572}, + year={2021} +} +``` diff --git a/fairseq/__init__.py b/fairseq/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..dc9fd1886d55756b5bdfeccf1ad329bd419a706e --- /dev/null +++ b/fairseq/__init__.py @@ -0,0 +1,44 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +"""isort:skip_file""" + +import os +import sys + +try: + from .version import __version__ # noqa +except ImportError: + version_txt = os.path.join(os.path.dirname(__file__), "version.txt") + with open(version_txt) as f: + __version__ = f.read().strip() + +__all__ = ["pdb"] + +# backwards compatibility to support `from fairseq.X import Y` +from fairseq.distributed import utils as distributed_utils +from fairseq.logging import meters, metrics, progress_bar # noqa + +sys.modules["fairseq.distributed_utils"] = distributed_utils +sys.modules["fairseq.meters"] = meters +sys.modules["fairseq.metrics"] = metrics +sys.modules["fairseq.progress_bar"] = progress_bar + +# initialize hydra +from fairseq.dataclass.initialize import hydra_init +hydra_init() + +import fairseq.criterions # noqa +import fairseq.distributed # noqa +import fairseq.models # noqa +import fairseq.modules # noqa +import fairseq.optim # noqa +import fairseq.optim.lr_scheduler # noqa +import fairseq.pdb # noqa +import fairseq.scoring # noqa +import fairseq.tasks # noqa +import fairseq.token_generation_constraints # noqa + +import fairseq.benchmark # noqa +import fairseq.model_parallel # noqa diff --git a/fairseq/benchmark/__init__.py b/fairseq/benchmark/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..0317d5c623778fe40b7bf07b77769cd10c243244 --- /dev/null +++ b/fairseq/benchmark/__init__.py @@ -0,0 +1,7 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +# import models/tasks to register them +from . import dummy_dataset, dummy_lm, dummy_masked_lm, dummy_model, dummy_mt # noqa diff --git a/fairseq/benchmark/dummy_dataset.py b/fairseq/benchmark/dummy_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..2f051754af55966e26850e94c121e0ff439bfd28 --- /dev/null +++ b/fairseq/benchmark/dummy_dataset.py @@ -0,0 +1,36 @@ +import numpy as np +from fairseq.data import FairseqDataset + + +class DummyDataset(FairseqDataset): + def __init__(self, batch, num_items, item_size): + super().__init__() + self.batch = batch + self.num_items = num_items + self.item_size = item_size + + def __getitem__(self, index): + return index + + def __len__(self): + return self.num_items + + def collater(self, samples): + return self.batch + + @property + def sizes(self): + return np.array([self.item_size] * self.num_items) + + def num_tokens(self, index): + return self.item_size + + def size(self, index): + return self.item_size + + def ordered_indices(self): + return np.arange(self.num_items) + + @property + def supports_prefetch(self): + return False diff --git a/fairseq/benchmark/dummy_lm.py b/fairseq/benchmark/dummy_lm.py new file mode 100644 index 0000000000000000000000000000000000000000..c6246a0c0e338fa36244b3aa4fb57f189fbffcb6 --- /dev/null +++ b/fairseq/benchmark/dummy_lm.py @@ -0,0 +1,83 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +from dataclasses import dataclass, field +from typing import Optional + +import torch +from .dummy_dataset import DummyDataset +from fairseq.data import Dictionary +from fairseq.dataclass import FairseqDataclass +from fairseq.tasks import FairseqTask, register_task +from omegaconf import II + + +logger = logging.getLogger(__name__) + + +@dataclass +class DummyLMConfig(FairseqDataclass): + dict_size: int = 49996 + dataset_size: int = 100000 + tokens_per_sample: int = field( + default=512, metadata={"help": "max sequence length"} + ) + add_bos_token: bool = False + batch_size: Optional[int] = II("dataset.batch_size") + max_tokens: Optional[int] = II("dataset.max_tokens") + max_target_positions: int = II("task.tokens_per_sample") + + +@register_task("dummy_lm", dataclass=DummyLMConfig) +class DummyLMTask(FairseqTask): + def __init__(self, cfg: DummyLMConfig): + super().__init__(cfg) + + # load dictionary + self.dictionary = Dictionary() + for i in range(cfg.dict_size): + self.dictionary.add_symbol("word{}".format(i)) + self.dictionary.pad_to_multiple_(8) # often faster if divisible by 8 + logger.info("dictionary: {} types".format(len(self.dictionary))) + + seq = torch.arange(cfg.tokens_per_sample + 1) + self.dictionary.pad() + 1 + + self.dummy_src = seq[:-1] + self.dummy_tgt = seq[1:] + + def load_dataset(self, split, epoch=1, combine=False, **kwargs): + """Load a given dataset split. + Args: + split (str): name of the split (e.g., train, valid, test) + """ + if self.cfg.batch_size is not None: + bsz = self.cfg.batch_size + else: + bsz = max(1, self.cfg.max_tokens // self.cfg.tokens_per_sample) + self.datasets[split] = DummyDataset( + { + "id": 1, + "net_input": { + "src_tokens": torch.stack([self.dummy_src for _ in range(bsz)]), + "src_lengths": torch.full( + (bsz,), self.cfg.tokens_per_sample, dtype=torch.long + ), + }, + "target": torch.stack([self.dummy_tgt for _ in range(bsz)]), + "nsentences": bsz, + "ntokens": bsz * self.cfg.tokens_per_sample, + }, + num_items=self.cfg.dataset_size, + item_size=self.cfg.tokens_per_sample, + ) + + @property + def source_dictionary(self): + return self.dictionary + + @property + def target_dictionary(self): + return self.dictionary diff --git a/fairseq/benchmark/dummy_masked_lm.py b/fairseq/benchmark/dummy_masked_lm.py new file mode 100644 index 0000000000000000000000000000000000000000..12b9c5d0f55993bf8750564882a351fc3f8055f0 --- /dev/null +++ b/fairseq/benchmark/dummy_masked_lm.py @@ -0,0 +1,94 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +from dataclasses import dataclass, field +from typing import Optional + +import torch +from omegaconf import II + +from .dummy_dataset import DummyDataset +from fairseq.data import Dictionary +from fairseq.dataclass import FairseqDataclass +from fairseq.tasks import FairseqTask, register_task + +logger = logging.getLogger(__name__) + + +@dataclass +class DummyMaskedLMConfig(FairseqDataclass): + dict_size: int = 49996 + dataset_size: int = 100000 + tokens_per_sample: int = field( + default=512, + metadata={ + "help": "max number of total tokens over all" + " segments per sample for BERT dataset" + }, + ) + batch_size: Optional[int] = II("dataset.batch_size") + max_tokens: Optional[int] = II("dataset.max_tokens") + max_target_positions: int = II("task.tokens_per_sample") + + +@register_task("dummy_masked_lm", dataclass=DummyMaskedLMConfig) +class DummyMaskedLMTask(FairseqTask): + def __init__(self, cfg: DummyMaskedLMConfig): + super().__init__(cfg) + + self.dictionary = Dictionary() + for i in range(cfg.dict_size): + self.dictionary.add_symbol("word{}".format(i)) + logger.info("dictionary: {} types".format(len(self.dictionary))) + # add mask token + self.mask_idx = self.dictionary.add_symbol("<mask>") + self.dictionary.pad_to_multiple_(8) # often faster if divisible by 8 + + mask_idx = 0 + pad_idx = 1 + seq = torch.arange(cfg.tokens_per_sample) + pad_idx + 1 + mask = torch.arange(2, cfg.tokens_per_sample, 7) # ~15% + src = seq.clone() + src[mask] = mask_idx + tgt = torch.full_like(seq, pad_idx) + tgt[mask] = seq[mask] + + self.dummy_src = src + self.dummy_tgt = tgt + + def load_dataset(self, split, epoch=1, combine=False, **kwargs): + """Load a given dataset split. + Args: + split (str): name of the split (e.g., train, valid, test) + """ + if self.cfg.batch_size is not None: + bsz = self.cfg.batch_size + else: + bsz = max(1, self.cfg.max_tokens // self.cfg.tokens_per_sample) + self.datasets[split] = DummyDataset( + { + "id": 1, + "net_input": { + "src_tokens": torch.stack([self.dummy_src for _ in range(bsz)]), + "src_lengths": torch.full( + (bsz,), self.cfg.tokens_per_sample, dtype=torch.long + ), + }, + "target": torch.stack([self.dummy_tgt for _ in range(bsz)]), + "nsentences": bsz, + "ntokens": bsz * self.cfg.tokens_per_sample, + }, + num_items=self.cfg.dataset_size, + item_size=self.cfg.tokens_per_sample, + ) + + @property + def source_dictionary(self): + return self.dictionary + + @property + def target_dictionary(self): + return self.dictionary diff --git a/fairseq/benchmark/dummy_model.py b/fairseq/benchmark/dummy_model.py new file mode 100644 index 0000000000000000000000000000000000000000..ff26e4fe655d8e8d7f9942c4bd3df7cd267405fb --- /dev/null +++ b/fairseq/benchmark/dummy_model.py @@ -0,0 +1,96 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch.nn as nn +import torch.nn.functional as F +from fairseq.data import Dictionary +from fairseq.models import ( + FairseqDecoder, + FairseqLanguageModel, + register_model, + register_model_architecture, +) + + +@register_model("dummy_model") +class DummyModel(FairseqLanguageModel): + def __init__(self, args, encoder): + super().__init__(encoder) + self.args = args + + @staticmethod + def add_args(parser): + parser.add_argument("--num-layers", type=int, default=24) + parser.add_argument("--embed-dim", type=int, default=1024) + + @classmethod + def build_model(cls, args, task): + encoder = DummyEncoder( + num_embed=len(task.target_dictionary), + embed_dim=args.embed_dim, + num_layers=args.num_layers, + ) + return cls(args, encoder) + + def forward(self, src_tokens, masked_tokens=None, **kwargs): + return self.decoder(src_tokens, masked_tokens=masked_tokens) + + +class DummyEncoder(FairseqDecoder): + def __init__(self, num_embed=50000, embed_dim=1024, num_layers=24): + super().__init__(Dictionary()) + self.embed = nn.Embedding( + num_embeddings=num_embed, embedding_dim=embed_dim, padding_idx=0 + ) + self.layers_a = nn.ModuleList( + [ + nn.Sequential( + nn.LayerNorm(embed_dim), + nn.Linear(embed_dim, 3 * embed_dim), # q, k, v input projection + nn.Linear(3 * embed_dim, embed_dim), # skip self-attention + nn.Linear(embed_dim, embed_dim), # output projection + nn.Dropout(), + ) + for i in range(num_layers) + ] + ) + self.layers_b = nn.ModuleList( + [ + nn.Sequential( + nn.LayerNorm(embed_dim), + nn.Linear(embed_dim, 4 * embed_dim), # FFN + nn.ReLU(), + nn.Linear(4 * embed_dim, embed_dim), # FFN + nn.Dropout(0.1), + ) + for i in range(num_layers) + ] + ) + self.out_proj = nn.Linear(embed_dim, num_embed) + + def forward(self, tokens, masked_tokens=None): + x = self.embed(tokens) + for layer_a, layer_b in zip(self.layers_a, self.layers_b): + x = x + layer_a(x) + x = x + layer_b(x) + x = self.out_proj(x) + if masked_tokens is not None: + x = x[masked_tokens] + return (x,) + + def max_positions(self): + return 1024 + + def get_normalized_probs(self, net_output, log_probs, sample=None): + logits = net_output[0].float() + if log_probs: + return F.log_softmax(logits, dim=-1) + else: + return F.softmax(logits, dim=-1) + + +@register_model_architecture("dummy_model", "dummy_model") +def base_architecture(args): + pass diff --git a/fairseq/benchmark/dummy_mt.py b/fairseq/benchmark/dummy_mt.py new file mode 100644 index 0000000000000000000000000000000000000000..4ca7be93a38d8d2b47685b74b4f8b8f9dcb03d2e --- /dev/null +++ b/fairseq/benchmark/dummy_mt.py @@ -0,0 +1,119 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging + +import numpy as np +import torch +from fairseq.data import Dictionary, FairseqDataset +from fairseq.tasks import LegacyFairseqTask, register_task + + +logger = logging.getLogger(__name__) + + +@register_task("dummy_mt") +class DummyMTTask(LegacyFairseqTask): + @staticmethod + def add_args(parser): + """Add task-specific arguments to the parser.""" + parser.add_argument("--dict-size", default=49996, type=int) + parser.add_argument("--dataset-size", default=100000, type=int) + parser.add_argument("--src-len", default=30, type=int) + parser.add_argument("--tgt-len", default=30, type=int) + + def __init__(self, args, dictionary): + super().__init__(args) + self.dictionary = dictionary + self.seed = args.seed + + dictionary.pad_to_multiple_(8) # often faster if divisible by 8 + + self.dummy_src = torch.arange(args.src_len + 1) + dictionary.pad() + 1 + self.dummy_tgt = torch.arange(args.tgt_len + 1) + dictionary.pad() + 1 + + @classmethod + def setup_task(cls, args, **kwargs): + """Setup the task. """ + dictionary = Dictionary() + for i in range(args.dict_size): + dictionary.add_symbol("word{}".format(i)) + logger.info("dictionary: {} types".format(len(dictionary))) + + args.max_source_positions = args.src_len + dictionary.pad() + 2 + args.max_target_positions = args.tgt_len + dictionary.pad() + 2 + + return cls(args, dictionary) + + def load_dataset(self, split, epoch=1, combine=False, **kwargs): + """Load a given dataset split. + Args: + split (str): name of the split (e.g., train, valid, test) + """ + item_size = max(self.args.src_len, self.args.tgt_len) + if self.args.batch_size is not None: + bsz = self.args.batch_size + else: + bsz = max(1, self.args.max_tokens // item_size) + tgt = torch.stack([self.dummy_tgt for _ in range(bsz)]) + self.datasets[split] = DummyDataset( + { + "id": 1, + "net_input": { + "src_tokens": torch.stack([self.dummy_src for _ in range(bsz)]), + "src_lengths": torch.full( + (bsz,), self.args.src_len, dtype=torch.long + ), + "prev_output_tokens": tgt.clone(), + }, + "target": tgt, + "nsentences": bsz, + "ntokens": bsz * self.args.tgt_len, + }, + num_items=self.args.dataset_size, + item_size=item_size, + ) + + @property + def source_dictionary(self): + return self.dictionary + + @property + def target_dictionary(self): + return self.dictionary + + +class DummyDataset(FairseqDataset): + def __init__(self, batch, num_items, item_size): + super().__init__() + self.batch = batch + self.num_items = num_items + self.item_size = item_size + + def __getitem__(self, index): + return index + + def __len__(self): + return self.num_items + + def collater(self, samples): + return self.batch + + @property + def sizes(self): + return np.array([self.item_size] * self.num_items) + + def num_tokens(self, index): + return self.item_size + + def size(self, index): + return self.item_size + + def ordered_indices(self): + return np.arange(self.num_items) + + @property + def supports_prefetch(self): + return False diff --git a/fairseq/binarizer.py b/fairseq/binarizer.py new file mode 100644 index 0000000000000000000000000000000000000000..18ae67bf25868095e101e7068962c78ee5d12aca --- /dev/null +++ b/fairseq/binarizer.py @@ -0,0 +1,114 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import os +from collections import Counter + +import torch +from fairseq.file_io import PathManager +from fairseq.tokenizer import tokenize_line +from typing import List, Dict + + +def safe_readline(f): + pos = f.tell() + while True: + try: + return f.readline() + except UnicodeDecodeError: + pos -= 1 + f.seek(pos) # search where this character begins + + +class Binarizer: + @staticmethod + def binarize( + filename, + dict, + consumer, + tokenize=tokenize_line, + append_eos=True, + reverse_order=False, + offset=0, + end=-1, + already_numberized=False, + ) -> Dict[str, int]: + nseq, ntok = 0, 0 + replaced = Counter() + + def replaced_consumer(word, idx): + if idx == dict.unk_index and word != dict.unk_word: + replaced.update([word]) + + with open(PathManager.get_local_path(filename), "r", encoding="utf-8") as f: + f.seek(offset) + # next(f) breaks f.tell(), hence readline() must be used + line = safe_readline(f) + while line: + # f.tell() does not always give the byte position in the file + # sometimes it skips to a very large number + # it is unlikely that through a normal read we go from + # end bytes to end + 2**32 bytes (4 GB) and this makes it unlikely + # that the procedure breaks by the undeterministic behavior of + # f.tell() + if end > 0 and f.tell() > end and f.tell() < end + 2 ** 32: + break + if already_numberized: + id_strings = line.strip().split() + id_list = [int(id_string) for id_string in id_strings] + if reverse_order: + id_list.reverse() + if append_eos: + id_list.append(dict.eos()) + ids = torch.IntTensor(id_list) + else: + ids = dict.encode_line( + line=line, + line_tokenizer=tokenize, + add_if_not_exist=False, + consumer=replaced_consumer, + append_eos=append_eos, + reverse_order=reverse_order, + ) + nseq += 1 + ntok += len(ids) + consumer(ids) + line = f.readline() + return { + "nseq": nseq, + "nunk": sum(replaced.values()), + "ntok": ntok, + "replaced": replaced, + } + + @staticmethod + def binarize_alignments( + filename, alignment_parser, consumer, offset=0, end=-1 + ) -> Dict[str, int]: + nseq = 0 + + with open(PathManager.get_local_path(filename), "r") as f: + f.seek(offset) + line = safe_readline(f) + while line: + if end > 0 and f.tell() > end: + break + ids = alignment_parser(line) + nseq += 1 + consumer(ids) + line = f.readline() + return {"nseq": nseq} + + @staticmethod + def find_offsets(filename, num_chunks) -> List[int]: + with open(PathManager.get_local_path(filename), "r", encoding="utf-8") as f: + size = os.fstat(f.fileno()).st_size + chunk_size = size // num_chunks + offsets = [0 for _ in range(num_chunks + 1)] + for i in range(1, num_chunks): + f.seek(chunk_size * i) + safe_readline(f) + offsets[i] = f.tell() + return offsets diff --git a/fairseq/checkpoint_utils.py b/fairseq/checkpoint_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..627f14160d2e4040f4dfe4e793f0986f53d8d39b --- /dev/null +++ b/fairseq/checkpoint_utils.py @@ -0,0 +1,798 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import ast +import collections +import contextlib +import logging +import os +import re +import time +import traceback +from collections import OrderedDict +from typing import Any, Dict, Optional, Union +from random import randint + +import torch +from fairseq.dataclass.configs import CheckpointConfig +from fairseq.dataclass.utils import ( + convert_namespace_to_omegaconf, + overwrite_args_by_name, +) +from fairseq.distributed.fully_sharded_data_parallel import FSDP, has_FSDP +from fairseq.file_io import PathManager +from fairseq.models import FairseqDecoder, FairseqEncoder +from omegaconf import DictConfig, open_dict, OmegaConf + + +logger = logging.getLogger(__name__) + + +def save_checkpoint(cfg: CheckpointConfig, trainer, epoch_itr, val_loss): + from fairseq import meters + + # only one worker should attempt to create the required dir + if trainer.data_parallel_rank == 0: + os.makedirs(cfg.save_dir, exist_ok=True) + + prev_best = getattr(save_checkpoint, "best", val_loss) + if val_loss is not None: + best_function = max if cfg.maximize_best_checkpoint_metric else min + save_checkpoint.best = best_function(val_loss, prev_best) + + if cfg.no_save: + return + + trainer.consolidate_optimizer() # TODO(SS): do we need this if no_save_optimizer_state + + if not trainer.should_save_checkpoint_on_current_rank: + if trainer.always_call_state_dict_during_save_checkpoint: + trainer.state_dict() + return + + write_timer = meters.StopwatchMeter() + write_timer.start() + + epoch = epoch_itr.epoch + end_of_epoch = epoch_itr.end_of_epoch() + updates = trainer.get_num_updates() + + logger.info(f"Preparing to save checkpoint for epoch {epoch} @ {updates} updates") + + def is_better(a, b): + return a >= b if cfg.maximize_best_checkpoint_metric else a <= b + + suffix = trainer.checkpoint_suffix + checkpoint_conds = collections.OrderedDict() + checkpoint_conds["checkpoint{}{}.pt".format(epoch, suffix)] = ( + end_of_epoch and not cfg.no_epoch_checkpoints and epoch % cfg.save_interval == 0 + ) + checkpoint_conds["checkpoint_{}_{}{}.pt".format(epoch, updates, suffix)] = ( + not end_of_epoch + and cfg.save_interval_updates > 0 + and updates % cfg.save_interval_updates == 0 + ) + checkpoint_conds["checkpoint_best{}.pt".format(suffix)] = val_loss is not None and ( + not hasattr(save_checkpoint, "best") + or is_better(val_loss, save_checkpoint.best) + ) + if val_loss is not None and cfg.keep_best_checkpoints > 0: + worst_best = getattr(save_checkpoint, "best", None) + chkpts = checkpoint_paths( + cfg.save_dir, + pattern=r"checkpoint\.best_{}_(\d+\.?\d*)\.pt".format( + cfg.best_checkpoint_metric + ), + ) + if len(chkpts) > 0: + p = chkpts[-1] if cfg.maximize_best_checkpoint_metric else chkpts[0] + worst_best = float(p.rsplit("_")[-1].replace(".pt", "")) + # add random digits to resolve ties + rand_sfx = randint(0, cfg.keep_best_checkpoints) + checkpoint_conds[ + "checkpoint.best_{}_{:.3f}{}.pt".format(cfg.best_checkpoint_metric, + val_loss, rand_sfx) + ] = worst_best is None or is_better(val_loss, worst_best) + checkpoint_conds[ + "checkpoint_last{}.pt".format(suffix) + ] = not cfg.no_last_checkpoints + + extra_state = {"train_iterator": epoch_itr.state_dict(), "val_loss": val_loss} + if hasattr(save_checkpoint, "best"): + extra_state.update({"best": save_checkpoint.best}) + + checkpoints = [ + os.path.join(cfg.save_dir, fn) for fn, cond in checkpoint_conds.items() if cond + ] + if len(checkpoints) > 0: + trainer.save_checkpoint(checkpoints[0], extra_state) + for cp in checkpoints[1:]: + if cfg.write_checkpoints_asynchronously: + # TODO[ioPath]: Need to implement a delayed asynchronous + # file copying/moving feature. + logger.warning( + f"ioPath is not copying {checkpoints[0]} to {cp} " + "since async write mode is on." + ) + else: + assert PathManager.copy( + checkpoints[0], cp, overwrite=True + ), f"Failed to copy {checkpoints[0]} to {cp}" + + write_timer.stop() + logger.info( + "Saved checkpoint {} (epoch {} @ {} updates, score {}) (writing took {} seconds)".format( + checkpoints[0], epoch, updates, val_loss, write_timer.sum + ) + ) + + if not end_of_epoch and cfg.keep_interval_updates > 0: + # remove old checkpoints; checkpoints are sorted in descending order + if cfg.keep_interval_updates_pattern == -1: + checkpoints = checkpoint_paths( + cfg.save_dir, pattern=r"checkpoint_\d+_(\d+){}\.pt".format(suffix) + ) + else: + checkpoints = checkpoint_paths( + cfg.save_dir, + pattern=r"checkpoint_\d+_(\d+){}\.pt".format(suffix), + keep_match=True, + ) + checkpoints = [ + x[0] + for x in checkpoints + if x[1] % cfg.keep_interval_updates_pattern != 0 + ] + + for old_chk in checkpoints[cfg.keep_interval_updates :]: + if os.path.lexists(old_chk): + os.remove(old_chk) + elif PathManager.exists(old_chk): + PathManager.rm(old_chk) + + if cfg.keep_last_epochs > 0: + # remove old epoch checkpoints; checkpoints are sorted in descending order + checkpoints = checkpoint_paths( + cfg.save_dir, pattern=r"checkpoint(\d+){}\.pt".format(suffix) + ) + for old_chk in checkpoints[cfg.keep_last_epochs :]: + if os.path.lexists(old_chk): + os.remove(old_chk) + + if cfg.keep_best_checkpoints > 0: + # only keep the best N checkpoints according to validation metric + checkpoints = checkpoint_paths( + cfg.save_dir, + pattern=r"checkpoint\.best_{}_(\d+\.?\d*){}\.pt".format( + cfg.best_checkpoint_metric, suffix + ), + ) + if not cfg.maximize_best_checkpoint_metric: + checkpoints = checkpoints[::-1] + for old_chk in checkpoints[cfg.keep_best_checkpoints :]: + if os.path.lexists(old_chk): + os.remove(old_chk) + + +def load_checkpoint(cfg: CheckpointConfig, trainer, **passthrough_args): + """ + Load a checkpoint and restore the training iterator. + + *passthrough_args* will be passed through to + ``trainer.get_train_iterator``. + """ + + reset_optimizer = cfg.reset_optimizer + reset_lr_scheduler = cfg.reset_lr_scheduler + optimizer_overrides = ast.literal_eval(cfg.optimizer_overrides) + reset_meters = cfg.reset_meters + reset_dataloader = cfg.reset_dataloader + + if cfg.finetune_from_model is not None and ( + reset_optimizer or reset_lr_scheduler or reset_meters or reset_dataloader + ): + raise ValueError( + "--finetune-from-model can not be set together with either --reset-optimizer" + " or reset_lr_scheduler or reset_meters or reset_dataloader" + ) + + suffix = trainer.checkpoint_suffix + if ( + cfg.restore_file == "checkpoint_last.pt" + ): # default value of restore_file is 'checkpoint_last.pt' + checkpoint_path = os.path.join( + cfg.save_dir, "checkpoint_last{}.pt".format(suffix) + ) + first_launch = not PathManager.exists(checkpoint_path) + if cfg.finetune_from_model is not None and first_launch: + # if there is no last checkpoint to restore, start the finetune from pretrained model + # else just use usual logic to load checkpoint, e.g. restart from last checkpoint and etc. + if PathManager.exists(cfg.finetune_from_model): + checkpoint_path = cfg.finetune_from_model + reset_optimizer = True + reset_lr_scheduler = True + reset_meters = True + reset_dataloader = True + logger.info( + f"loading pretrained model from {checkpoint_path}: " + "optimizer, lr scheduler, meters, dataloader will be reset" + ) + else: + raise ValueError( + f"--funetune-from-model {cfg.finetune_from_model} does not exist" + ) + elif suffix is not None: + checkpoint_path = cfg.restore_file.replace(".pt", suffix + ".pt") + else: + checkpoint_path = cfg.restore_file + + if cfg.restore_file != "checkpoint_last.pt" and cfg.finetune_from_model: + raise ValueError( + "--finetune-from-model and --restore-file (non-default value) " + "can not be specified together: " + str(cfg) + ) + + extra_state = trainer.load_checkpoint( + checkpoint_path, + reset_optimizer, + reset_lr_scheduler, + optimizer_overrides, + reset_meters=reset_meters, + ) + + if ( + extra_state is not None + and "best" in extra_state + and not reset_optimizer + and not reset_meters + ): + save_checkpoint.best = extra_state["best"] + + if extra_state is not None and not reset_dataloader: + # restore iterator from checkpoint + itr_state = extra_state["train_iterator"] + epoch_itr = trainer.get_train_iterator( + epoch=itr_state["epoch"], load_dataset=True, **passthrough_args + ) + epoch_itr.load_state_dict(itr_state) + else: + epoch_itr = trainer.get_train_iterator( + epoch=1, load_dataset=True, **passthrough_args + ) + + trainer.lr_step(epoch_itr.epoch) + + return extra_state, epoch_itr + + +def load_checkpoint_to_cpu(path, arg_overrides=None, load_on_all_ranks=False): + """Loads a checkpoint to CPU (with upgrading for backward compatibility). + + If doing single-GPU training or if the checkpoint is only being loaded by at + most one process on each node (current default behavior is for only rank 0 + to read the checkpoint from disk), load_on_all_ranks should be False to + avoid errors from torch.distributed not having been initialized or + torch.distributed.barrier() hanging. + + If all processes on each node may be loading the checkpoint + simultaneously, load_on_all_ranks should be set to True to avoid I/O + conflicts. + + There's currently no support for > 1 but < all processes loading the + checkpoint on each node. + """ + local_path = PathManager.get_local_path(path) + # The locally cached file returned by get_local_path() may be stale for + # remote files that are periodically updated/overwritten (ex: + # checkpoint_last.pt) - so we remove the local copy, sync across processes + # (if needed), and then download a fresh copy. + if local_path != path and PathManager.path_requires_pathmanager(path): + try: + os.remove(local_path) + except FileNotFoundError: + # With potentially multiple processes removing the same file, the + # file being missing is benign (missing_ok isn't available until + # Python 3.8). + pass + if load_on_all_ranks: + torch.distributed.barrier() + local_path = PathManager.get_local_path(path) + + with open(local_path, "rb") as f: + state = torch.load(f, map_location=torch.device("cpu")) + + if "args" in state and state["args"] is not None and arg_overrides is not None: + args = state["args"] + for arg_name, arg_val in arg_overrides.items(): + setattr(args, arg_name, arg_val) + + if "cfg" in state and state["cfg"] is not None: + + # hack to be able to set Namespace in dict config. this should be removed when we update to newer + # omegaconf version that supports object flags, or when we migrate all existing models + from omegaconf import _utils + + old_primitive = _utils.is_primitive_type + _utils.is_primitive_type = lambda _: True + + state["cfg"] = OmegaConf.create(state["cfg"]) + + _utils.is_primitive_type = old_primitive + OmegaConf.set_struct(state["cfg"], True) + + if arg_overrides is not None: + overwrite_args_by_name(state["cfg"], arg_overrides) + + state = _upgrade_state_dict(state) + return state + + +def load_model_ensemble( + filenames, + arg_overrides: Optional[Dict[str, Any]] = None, + task=None, + strict=True, + suffix="", + num_shards=1, + state=None, +): + """Loads an ensemble of models. + + Args: + filenames (List[str]): checkpoint files to load + arg_overrides (Dict[str,Any], optional): override model args that + were used during model training + task (fairseq.tasks.FairseqTask, optional): task to use for loading + """ + assert not ( + strict and num_shards > 1 + ), "Cannot load state dict with strict=True and checkpoint shards > 1" + ensemble, args, _task = load_model_ensemble_and_task( + filenames, + arg_overrides, + task, + strict, + suffix, + num_shards, + state, + ) + return ensemble, args + + +def get_maybe_sharded_checkpoint_filename( + filename: str, suffix: str, shard_idx: int, num_shards: int +) -> str: + orig_filename = filename + filename = filename.replace(".pt", suffix + ".pt") + fsdp_filename = filename[:-3] + f"-shard{shard_idx}.pt" + model_parallel_filename = orig_filename[:-3] + f"_part{shard_idx}.pt" + if PathManager.exists(fsdp_filename): + return fsdp_filename + elif num_shards > 1: + return model_parallel_filename + else: + return filename + + +def load_model_ensemble_and_task( + filenames, + arg_overrides: Optional[Dict[str, Any]] = None, + task=None, + strict=True, + suffix="", + num_shards=1, + state=None, +): + assert state is None or len(filenames) == 1 + + from fairseq import tasks + + assert not ( + strict and num_shards > 1 + ), "Cannot load state dict with strict=True and checkpoint shards > 1" + ensemble = [] + cfg = None + for filename in filenames: + orig_filename = filename + model_shard_state = {"shard_weights": [], "shard_metadata": []} + assert num_shards > 0 + st = time.time() + for shard_idx in range(num_shards): + filename = get_maybe_sharded_checkpoint_filename( + orig_filename, suffix, shard_idx, num_shards + ) + + if not PathManager.exists(filename): + raise IOError("Model file not found: {}".format(filename)) + if state is None: + state = load_checkpoint_to_cpu(filename, arg_overrides) + if "args" in state and state["args"] is not None: + cfg = convert_namespace_to_omegaconf(state["args"]) + elif "cfg" in state and state["cfg"] is not None: + cfg = state["cfg"] + else: + raise RuntimeError( + f"Neither args nor cfg exist in state keys = {state.keys()}" + ) + + if task is None: + task = tasks.setup_task(cfg.task) + + if "task_state" in state: + task.load_state_dict(state["task_state"]) + + if "fsdp_metadata" in state and num_shards > 1: + model_shard_state["shard_weights"].append(state["model"]) + model_shard_state["shard_metadata"].append(state["fsdp_metadata"]) + # check FSDP import before the code goes too far + if not has_FSDP: + raise ImportError( + "Cannot find FullyShardedDataParallel. " + "Please install fairscale with: pip install fairscale" + ) + if shard_idx == num_shards - 1: + consolidated_model_state = FSDP.consolidate_shard_weights( + shard_weights=model_shard_state["shard_weights"], + shard_metadata=model_shard_state["shard_metadata"], + ) + model = task.build_model(cfg.model) + model.load_state_dict( + consolidated_model_state, strict=strict, model_cfg=cfg.model + ) + else: + # model parallel checkpoint or unsharded checkpoint + model = task.build_model(cfg.model) + model.load_state_dict( + state["model"], strict=strict, model_cfg=cfg.model + ) + + # reset state so it gets loaded for the next model in ensemble + state = None + if shard_idx % 10 == 0 and shard_idx > 0: + elapsed = time.time() - st + logger.info(f"Loaded {shard_idx} shards in {elapsed:.2f}s, {elapsed / (shard_idx+1):.2f}s/shard") + + # build model for ensemble + ensemble.append(model) + return ensemble, cfg, task + + +def checkpoint_paths(path, pattern=r"checkpoint(\d+)\.pt", keep_match=False): + """Retrieves all checkpoints found in `path` directory. + + Checkpoints are identified by matching filename to the specified pattern. If + the pattern contains groups, the result will be sorted by the first group in + descending order. + """ + pt_regexp = re.compile(pattern) + files = PathManager.ls(path) + + entries = [] + for i, f in enumerate(files): + m = pt_regexp.fullmatch(f) + if m is not None: + idx = float(m.group(1)) if len(m.groups()) > 0 else i + entries.append((idx, m.group(0))) + if keep_match: + return [(os.path.join(path, x[1]), x[0]) for x in sorted(entries, reverse=True)] + else: + return [os.path.join(path, x[1]) for x in sorted(entries, reverse=True)] + + +def torch_persistent_save(obj, filename, async_write: bool = False): + if async_write: + with PathManager.opena(filename, "wb") as f: + _torch_persistent_save(obj, f) + else: + if PathManager.supports_rename(filename): + # do atomic save + with PathManager.open(filename + ".tmp", "wb") as f: + _torch_persistent_save(obj, f) + PathManager.rename(filename + ".tmp", filename) + else: + # fallback to non-atomic save + with PathManager.open(filename, "wb") as f: + _torch_persistent_save(obj, f) + + +def _torch_persistent_save(obj, f): + if isinstance(f, str): + with PathManager.open(f, "wb") as h: + torch_persistent_save(obj, h) + return + for i in range(3): + try: + return torch.save(obj, f) + except Exception: + if i == 2: + logger.error(traceback.format_exc()) + + +def _upgrade_state_dict(state): + """Helper for upgrading old model checkpoints.""" + + # add optimizer_history + if "optimizer_history" not in state: + state["optimizer_history"] = [ + {"criterion_name": "CrossEntropyCriterion", "best_loss": state["best_loss"]} + ] + state["last_optimizer_state"] = state["optimizer"] + del state["optimizer"] + del state["best_loss"] + # move extra_state into sub-dictionary + if "epoch" in state and "extra_state" not in state: + state["extra_state"] = { + "epoch": state["epoch"], + "batch_offset": state["batch_offset"], + "val_loss": state["val_loss"], + } + del state["epoch"] + del state["batch_offset"] + del state["val_loss"] + # reduce optimizer history's memory usage (only keep the last state) + if "optimizer" in state["optimizer_history"][-1]: + state["last_optimizer_state"] = state["optimizer_history"][-1]["optimizer"] + for optim_hist in state["optimizer_history"]: + del optim_hist["optimizer"] + # record the optimizer class name + if "optimizer_name" not in state["optimizer_history"][-1]: + state["optimizer_history"][-1]["optimizer_name"] = "FairseqNAG" + # move best_loss into lr_scheduler_state + if "lr_scheduler_state" not in state["optimizer_history"][-1]: + state["optimizer_history"][-1]["lr_scheduler_state"] = { + "best": state["optimizer_history"][-1]["best_loss"] + } + del state["optimizer_history"][-1]["best_loss"] + # keep track of number of updates + if "num_updates" not in state["optimizer_history"][-1]: + state["optimizer_history"][-1]["num_updates"] = 0 + # old model checkpoints may not have separate source/target positions + if ( + "args" in state + and hasattr(state["args"], "max_positions") + and not hasattr(state["args"], "max_source_positions") + ): + state["args"].max_source_positions = state["args"].max_positions + state["args"].max_target_positions = state["args"].max_positions + # use stateful training data iterator + if "train_iterator" not in state["extra_state"]: + state["extra_state"]["train_iterator"] = { + "epoch": state["extra_state"]["epoch"], + "iterations_in_epoch": state["extra_state"].get("batch_offset", 0), + } + + # backward compatibility, cfg updates + if "args" in state and state["args"] is not None: + # default to translation task + if not hasattr(state["args"], "task"): + state["args"].task = "translation" + # --raw-text and --lazy-load are deprecated + if getattr(state["args"], "raw_text", False): + state["args"].dataset_impl = "raw" + elif getattr(state["args"], "lazy_load", False): + state["args"].dataset_impl = "lazy" + # epochs start at 1 + if state["extra_state"]["train_iterator"] is not None: + state["extra_state"]["train_iterator"]["epoch"] = max( + state["extra_state"]["train_iterator"].get("epoch", 1), 1 + ) + # --remove-bpe ==> --postprocess + if hasattr(state["args"], "remove_bpe"): + state["args"].post_process = state["args"].remove_bpe + # --min-lr ==> --stop-min-lr + if hasattr(state["args"], "min_lr"): + state["args"].stop_min_lr = state["args"].min_lr + del state["args"].min_lr + # binary_cross_entropy / kd_binary_cross_entropy => wav2vec criterion + if ( + hasattr(state["args"], "criterion") + and state["args"].criterion in [ + "binary_cross_entropy", + "kd_binary_cross_entropy", + ] + ): + state["args"].criterion = "wav2vec" + # remove log_keys if it's None (criteria will supply a default value of []) + if hasattr(state["args"], "log_keys") and state["args"].log_keys is None: + delattr(state["args"], "log_keys") + # speech_pretraining => audio pretraining + if ( + hasattr(state["args"], "task") + and state["args"].task == "speech_pretraining" + ): + state["args"].task = "audio_pretraining" + # audio_cpc => wav2vec + if hasattr(state["args"], "arch") and state["args"].arch == "audio_cpc": + state["args"].arch = "wav2vec" + # convert legacy float learning rate to List[float] + if hasattr(state["args"], "lr") and isinstance(state["args"].lr, float): + state["args"].lr = [state["args"].lr] + # convert task data arg to a string instead of List[string] + if ( + hasattr(state["args"], "data") + and isinstance(state["args"].data, list) + and len(state["args"].data) > 0 + ): + state["args"].data = state["args"].data[0] + # remove keys in state["args"] related to teacher-student learning + for key in [ + "static_teachers", + "static_teacher_weights", + "dynamic_teachers", + "dynamic_teacher_weights", + ]: + if key in state["args"]: + delattr(state["args"], key) + + state["cfg"] = convert_namespace_to_omegaconf(state["args"]) + + if "cfg" in state and state["cfg"] is not None: + cfg = state["cfg"] + with open_dict(cfg): + # any upgrades for Hydra-based configs + if ( + "task" in cfg + and "eval_wer_config" in cfg.task + and isinstance(cfg.task.eval_wer_config.print_alignment, bool) + ): + cfg.task.eval_wer_config.print_alignment = "hard" + if "generation" in cfg and isinstance(cfg.generation.print_alignment, bool): + cfg.generation.print_alignment = "hard" + if ( + "model" in cfg + and "w2v_args" in cfg.model + and cfg.model.w2v_args is not None + and ( + hasattr(cfg.model.w2v_args, "task") or "task" in cfg.model.w2v_args + ) + and hasattr(cfg.model.w2v_args.task, "eval_wer_config") + and cfg.model.w2v_args.task.eval_wer_config is not None + and isinstance( + cfg.model.w2v_args.task.eval_wer_config.print_alignment, bool + ) + ): + cfg.model.w2v_args.task.eval_wer_config.print_alignment = "hard" + + return state + + +def prune_state_dict(state_dict, model_cfg: Optional[DictConfig]): + """Prune the given state_dict if desired for LayerDrop + (https://arxiv.org/abs/1909.11556). + + Training with LayerDrop allows models to be robust to pruning at inference + time. This function prunes state_dict to allow smaller models to be loaded + from a larger model and re-maps the existing state_dict for this to occur. + + It's called by functions that load models from checkpoints and does not + need to be called directly. + """ + arch = None + if model_cfg is not None: + arch = ( + model_cfg._name + if isinstance(model_cfg, DictConfig) + else getattr(model_cfg, "arch", None) + ) + + if not model_cfg or arch is None or arch == "ptt_transformer": + # args should not be none, but don't crash if it is. + return state_dict + + encoder_layers_to_keep = getattr(model_cfg, "encoder_layers_to_keep", None) + decoder_layers_to_keep = getattr(model_cfg, "decoder_layers_to_keep", None) + + if not encoder_layers_to_keep and not decoder_layers_to_keep: + return state_dict + + # apply pruning + logger.info( + "Pruning model to specified layer configuration - this works best if the model was trained with LayerDrop" + ) + + def create_pruning_pass(layers_to_keep, layer_name): + keep_layers = sorted( + int(layer_string) for layer_string in layers_to_keep.split(",") + ) + mapping_dict = {} + for i in range(len(keep_layers)): + mapping_dict[str(keep_layers[i])] = str(i) + + regex = re.compile(r"^{layer}.*\.layers\.(\d+)".format(layer=layer_name)) + return {"substitution_regex": regex, "mapping_dict": mapping_dict} + + pruning_passes = [] + if encoder_layers_to_keep: + pruning_passes.append(create_pruning_pass(encoder_layers_to_keep, "encoder")) + if decoder_layers_to_keep: + pruning_passes.append(create_pruning_pass(decoder_layers_to_keep, "decoder")) + + new_state_dict = {} + for layer_name in state_dict.keys(): + match = re.search(r"\.layers\.(\d+)\.", layer_name) + # if layer has no number in it, it is a supporting layer, such as an + # embedding + if not match: + new_state_dict[layer_name] = state_dict[layer_name] + continue + + # otherwise, layer should be pruned. + original_layer_number = match.group(1) + # figure out which mapping dict to replace from + for pruning_pass in pruning_passes: + if original_layer_number in pruning_pass["mapping_dict"] and pruning_pass[ + "substitution_regex" + ].search(layer_name): + new_layer_number = pruning_pass["mapping_dict"][original_layer_number] + substitution_match = pruning_pass["substitution_regex"].search( + layer_name + ) + new_state_key = ( + layer_name[: substitution_match.start(1)] + + new_layer_number + + layer_name[substitution_match.end(1) :] + ) + new_state_dict[new_state_key] = state_dict[layer_name] + + # Since layers are now pruned, *_layers_to_keep are no longer needed. + # This is more of "It would make it work fix" rather than a proper fix. + if isinstance(model_cfg, DictConfig): + context = open_dict(model_cfg) + else: + context = contextlib.ExitStack() + with context: + if hasattr(model_cfg, "encoder_layers_to_keep"): + model_cfg.encoder_layers_to_keep = None + if hasattr(model_cfg, "decoder_layers_to_keep"): + model_cfg.decoder_layers_to_keep = None + + return new_state_dict + + +def load_pretrained_component_from_model( + component: Union[FairseqEncoder, FairseqDecoder], checkpoint: str +): + """ + Load a pretrained FairseqEncoder or FairseqDecoder from checkpoint into the + provided `component` object. If state_dict fails to load, there may be a + mismatch in the architecture of the corresponding `component` found in the + `checkpoint` file. + """ + if not PathManager.exists(checkpoint): + raise IOError("Model file not found: {}".format(checkpoint)) + state = load_checkpoint_to_cpu(checkpoint) + if isinstance(component, FairseqEncoder): + component_type = "encoder" + elif isinstance(component, FairseqDecoder): + component_type = "decoder" + else: + raise ValueError( + "component to load must be either a FairseqEncoder or " + "FairseqDecoder. Loading other component types are not supported." + ) + component_state_dict = OrderedDict() + for key in state["model"].keys(): + if key.startswith(component_type): + # encoder.input_layers.0.0.weight --> input_layers.0.0.weight + component_subkey = key[len(component_type) + 1 :] + component_state_dict[component_subkey] = state["model"][key] + component.load_state_dict(component_state_dict, strict=True) + return component + + +def verify_checkpoint_directory(save_dir: str) -> None: + if not os.path.exists(save_dir): + os.makedirs(save_dir, exist_ok=True) + temp_file_path = os.path.join(save_dir, "dummy") + try: + with open(temp_file_path, "w"): + pass + except OSError as e: + logger.warning( + "Unable to access checkpoint save directory: {}".format(save_dir) + ) + raise e + else: + os.remove(temp_file_path) diff --git a/fairseq/clib/cuda/ngram_repeat_block_cuda.cpp b/fairseq/clib/cuda/ngram_repeat_block_cuda.cpp new file mode 100644 index 0000000000000000000000000000000000000000..4199cd6ea86b019cb688b20c07e85905b2244fa0 --- /dev/null +++ b/fairseq/clib/cuda/ngram_repeat_block_cuda.cpp @@ -0,0 +1,47 @@ +/* +Copyright (c) Microsoft Corporation. +Licensed under the MIT License. +*/ + +#include <torch/extension.h> +#include <vector> + +/* +CPP Binding for CUDA OP +*/ + +// CUDA forward declarations +torch::Tensor ngram_repeat_block_cuda_forward(torch::Tensor tokens, + torch::Tensor lprobs, int bsz, + int step, int beam_size, + int no_repeat_ngram_size); + +#define CHECK_CUDA(x) \ + TORCH_CHECK(x.type().is_cuda(), #x " must be a CUDA tensor") +#define CHECK_CONTIGUOUS(x) \ + TORCH_CHECK(x.is_contiguous(), #x " must be contiguous") +#define CHECK_INPUT(x) \ + CHECK_CUDA(x); \ + CHECK_CONTIGUOUS(x) + +// Input check and call to CUDA OP +// Backward method not required +torch::Tensor ngram_repeat_block_forward(torch::Tensor tokens, + torch::Tensor lprobs, int bsz, + int step, int beam_size, + int no_repeat_ngram_size) { + CHECK_INPUT(tokens); + CHECK_INPUT(lprobs); + assert(bsz > 0); + assert(step >= 0); + assert(beam_size > 0); + assert(no_repeat_ngram_size > 0); + + return ngram_repeat_block_cuda_forward(tokens, lprobs, bsz, step, beam_size, + no_repeat_ngram_size); +} + +PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { + m.def("forward", &ngram_repeat_block_forward, + "No Repeat Ngram Block forward (CUDA)"); +} diff --git a/fairseq/clib/cuda/ngram_repeat_block_cuda_kernel.cu b/fairseq/clib/cuda/ngram_repeat_block_cuda_kernel.cu new file mode 100644 index 0000000000000000000000000000000000000000..b458b0916a7b38d1662377ce95559e7562e4c65d --- /dev/null +++ b/fairseq/clib/cuda/ngram_repeat_block_cuda_kernel.cu @@ -0,0 +1,76 @@ +/* +Copyright (c) Microsoft Corporation. +Licensed under the MIT License. +*/ + +/* +Kernel implementation for blocking repeated n-grams. +*/ + +#include <cuda.h> +#include <cuda_runtime.h> +#include <math.h> +#include <torch/extension.h> +#include <vector> + +// Ban repeated ngrams of length = 'no_repeat_ngram_size' +__global__ void banRepeatedTokens(long* __restrict__ tokens, + float* __restrict__ lprobs, + int max_predict_len, int vocab_size, + int no_repeat_ngram_size) { + auto row = blockIdx.x; + auto col = threadIdx.x; + auto start = row * (max_predict_len) + col; + // Each thread compares ngram starting from + // thread index with final ngram starting from + // step - no_repeat_ngram_size +2 + auto check_start_pos = blockDim.x; + auto lprob_start = row * vocab_size; + bool is_banned = true; + extern __shared__ long tokens_shm[]; + tokens_shm[col] = tokens[start]; + if (col == blockDim.x - 1) { + for (int i=1; i<no_repeat_ngram_size; i++){ + if (col+i < max_predict_len){ + tokens_shm[col + i] = tokens[start + i]; + } + } + } + __syncthreads(); + + for (int k = 0; k < no_repeat_ngram_size - 1; k++) { + if (tokens_shm[col + k] != tokens_shm[check_start_pos + k]) { + is_banned = false; + } + } + if (is_banned == true) { + auto token_to_be_banned = tokens_shm[col + no_repeat_ngram_size - 1]; + lprobs[lprob_start + token_to_be_banned] = -INFINITY; + } +} + +// Allocate blocks and threads based on +// batch size and sequence length and launch +// kernel +torch::Tensor ngram_repeat_block_cuda_forward(const torch::Tensor tokens, + torch::Tensor lprobs, int bsz, + int step, int beam_size, + int no_repeat_ngram_size) { + int threads = step - no_repeat_ngram_size + 2; + if (threads <= 0) return lprobs; + int max_predict_len = tokens.size(1); + int vocab_size = lprobs.size(1); + auto token_ptr = tokens.data_ptr<long>(); + auto lprob_ptr = lprobs.data_ptr<float>(); + int blocks = bsz * beam_size; + int shared_mem_size = (step + 1) * sizeof(long); + + // Launching N blocks where N is number of samples in a batch (beams*bsz) + // Launching T threads where T is number of previous ngrams in a sample + // Allocating shared mem per block for fastser access of input tokens since + // each token will be accessed N times to compare with current Ngram where + // N is Ngram size. + banRepeatedTokens<<<blocks, threads, shared_mem_size>>>( + token_ptr, lprob_ptr, max_predict_len, vocab_size, no_repeat_ngram_size); + return lprobs; +} diff --git a/fairseq/clib/libbase/balanced_assignment.cpp b/fairseq/clib/libbase/balanced_assignment.cpp new file mode 100644 index 0000000000000000000000000000000000000000..296f03b6aeb87a11db92e5342d8dab90f1fc3867 --- /dev/null +++ b/fairseq/clib/libbase/balanced_assignment.cpp @@ -0,0 +1,95 @@ +/** + * Copyright 2017-present, Facebook, Inc. + * All rights reserved. + * + * This source code is licensed under the license found in the + * LICENSE file in the root directory of this source tree. + */ + +/* +C++ code for solving the linear assignment problem. +Based on the Auction Algorithm from https://dspace.mit.edu/bitstream/handle/1721.1/3265/P-2108-26912652.pdf and the implementation from: +https://github.com/bkj/auction-lap +Adapted to be more efficient when each worker is looking for k jobs instead of 1. +*/ +#include <torch/extension.h> +#include <iostream> +using namespace torch::indexing; +torch::Tensor balanced_assignment(torch::Tensor job_and_worker_to_score) { + int max_iterations = 100; + torch::Tensor epsilon = (job_and_worker_to_score.max() - job_and_worker_to_score.min()) / 50; + epsilon.clamp_min_(1e-04); + torch::Tensor worker_and_job_to_score = job_and_worker_to_score.detach().transpose(0,1).contiguous(); + int num_workers = worker_and_job_to_score.size(0); + int num_jobs = worker_and_job_to_score.size(1); + auto device = worker_and_job_to_score.device(); + int jobs_per_worker = num_jobs / num_workers; + torch::Tensor value = worker_and_job_to_score.clone(); + int counter = 0; + torch::Tensor max_value = worker_and_job_to_score.max(); + + torch::Tensor bid_indices; + torch::Tensor cost = worker_and_job_to_score.new_zeros({1, num_jobs}); + torch::Tensor bids = worker_and_job_to_score.new_empty({num_workers, num_jobs}); + torch::Tensor bid_increments = worker_and_job_to_score.new_empty({num_workers, jobs_per_worker}); + torch::Tensor top_values = worker_and_job_to_score.new_empty({num_workers, jobs_per_worker + 1}); + torch::Tensor high_bids = worker_and_job_to_score.new_empty({num_jobs}); + + torch::Tensor top_index = top_values.to(torch::kLong); + torch::Tensor high_bidders = top_index.new_empty({num_jobs}); + torch::Tensor have_bids = high_bidders.to(torch::kBool); + torch::Tensor jobs_indices = torch::arange({num_jobs}, torch::dtype(torch::kLong).device(device)); + torch::Tensor true_tensor = torch::ones({1}, torch::dtype(torch::kBool).device(device)); + + while (true) { + bids.zero_(); + torch::topk_out(top_values, top_index, value, jobs_per_worker + 1, 1); + + // Each worker bids the difference in value between that job and the k+1th job + torch::sub_out(bid_increments, + top_values.index({Slice(None, None), Slice(0, jobs_per_worker)}), + top_values.index({Slice(None, None), jobs_per_worker}).unsqueeze(1)); + + bid_increments.add_(epsilon); + bids.scatter_(1, + top_index.index({Slice(None, None),Slice(0, jobs_per_worker)}), + bid_increments); + + if (counter < max_iterations && counter > 0) { + // Put in a minimal bid to retain items from the last round if no-one else bids for them this round + bids.view(-1).index_put_({bid_indices}, epsilon); + } + + // Find the highest bidding worker per job + torch::max_out(high_bids, high_bidders, bids, 0); + torch::gt_out(have_bids, high_bids, 0); + + if (have_bids.all().item<bool>()) { + // All jobs were bid for + break; + } + + // Make popular items more expensive + cost.add_(high_bids); + torch::sub_out(value, worker_and_job_to_score, cost); + + bid_indices = ((high_bidders * num_jobs) + jobs_indices).index({have_bids}); + + if (counter < max_iterations) { + // Make sure that this item will be in the winning worker's top-k next time. + value.view(-1).index_put_({bid_indices}, max_value); + } + else { + // Suboptimal approximation that converges quickly from current solution + value.view(-1).index_put_({bid_indices}, worker_and_job_to_score.view(-1).index({bid_indices})); + } + + counter += 1; + } + + return top_index.index({Slice(None, None), Slice(0, jobs_per_worker)}).reshape(-1); +} + +PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { + m.def("balanced_assignment", &balanced_assignment, "Balanced Assignment"); +} diff --git a/fairseq/clib/libbleu/libbleu.cpp b/fairseq/clib/libbleu/libbleu.cpp new file mode 100644 index 0000000000000000000000000000000000000000..3cf2d65b6d16e19ea299ebe43c9c25e3481d4524 --- /dev/null +++ b/fairseq/clib/libbleu/libbleu.cpp @@ -0,0 +1,141 @@ +/** + * Copyright 2017-present, Facebook, Inc. + * All rights reserved. + * + * This source code is licensed under the license found in the + * LICENSE file in the root directory of this source tree. + */ + +#include <map> +#include <array> +#include <cstring> +#include <cstdio> + +typedef struct +{ + size_t reflen; + size_t predlen; + size_t match1; + size_t count1; + size_t match2; + size_t count2; + size_t match3; + size_t count3; + size_t match4; + size_t count4; +} bleu_stat; + +// left trim (remove pad) +void bleu_ltrim(size_t* len, int** sent, int pad) { + size_t start = 0; + while(start < *len) { + if (*(*sent + start) != pad) { break; } + start++; + } + *sent += start; + *len -= start; +} + +// right trim remove (eos) +void bleu_rtrim(size_t* len, int** sent, int pad, int eos) { + size_t end = *len - 1; + while (end > 0) { + if (*(*sent + end) != eos && *(*sent + end) != pad) { break; } + end--; + } + *len = end + 1; +} + +// left and right trim +void bleu_trim(size_t* len, int** sent, int pad, int eos) { + bleu_ltrim(len, sent, pad); + bleu_rtrim(len, sent, pad, eos); +} + +size_t bleu_hash(int len, int* data) { + size_t h = 14695981039346656037ul; + size_t prime = 0x100000001b3; + char* b = (char*) data; + size_t blen = sizeof(int) * len; + + while (blen-- > 0) { + h ^= *b++; + h *= prime; + } + + return h; +} + +void bleu_addngram( + size_t *ntotal, size_t *nmatch, size_t n, + size_t reflen, int* ref, size_t predlen, int* pred) { + + if (predlen < n) { return; } + + predlen = predlen - n + 1; + (*ntotal) += predlen; + + if (reflen < n) { return; } + + reflen = reflen - n + 1; + + std::map<size_t, size_t> count; + while (predlen > 0) { + size_t w = bleu_hash(n, pred++); + count[w]++; + predlen--; + } + + while (reflen > 0) { + size_t w = bleu_hash(n, ref++); + if (count[w] > 0) { + (*nmatch)++; + count[w] -=1; + } + reflen--; + } +} + +extern "C" { + +#ifdef _WIN64 +__declspec(dllexport) +#endif +void bleu_zero_init(bleu_stat* stat) { + std::memset(stat, 0, sizeof(bleu_stat)); +} + +#ifdef _WIN64 +__declspec(dllexport) +#endif +void bleu_one_init(bleu_stat* stat) { + bleu_zero_init(stat); + stat->count1 = 0; + stat->count2 = 1; + stat->count3 = 1; + stat->count4 = 1; + stat->match1 = 0; + stat->match2 = 1; + stat->match3 = 1; + stat->match4 = 1; +} + +#ifdef _WIN64 +__declspec(dllexport) +#endif +void bleu_add( + bleu_stat* stat, + size_t reflen, int* ref, size_t predlen, int* pred, int pad, int eos) { + + bleu_trim(&reflen, &ref, pad, eos); + bleu_trim(&predlen, &pred, pad, eos); + stat->reflen += reflen; + stat->predlen += predlen; + + bleu_addngram(&stat->count1, &stat->match1, 1, reflen, ref, predlen, pred); + bleu_addngram(&stat->count2, &stat->match2, 2, reflen, ref, predlen, pred); + bleu_addngram(&stat->count3, &stat->match3, 3, reflen, ref, predlen, pred); + bleu_addngram(&stat->count4, &stat->match4, 4, reflen, ref, predlen, pred); +} + +} diff --git a/fairseq/clib/libbleu/module.cpp b/fairseq/clib/libbleu/module.cpp new file mode 100644 index 0000000000000000000000000000000000000000..8ed9a84b1c028bfe9ed1d45be6857b6e79b3459f --- /dev/null +++ b/fairseq/clib/libbleu/module.cpp @@ -0,0 +1,37 @@ +/** + * Copyright 2017-present, Facebook, Inc. + * All rights reserved. + * + * This source code is licensed under the license found in the + * LICENSE file in the root directory of this source tree. + */ + +#include <Python.h> + + +static PyMethodDef method_def[] = { + {NULL, NULL, 0, NULL} +}; + +static struct PyModuleDef module_def = { + PyModuleDef_HEAD_INIT, + "libbleu", /* name of module */ + NULL, /* module documentation, may be NULL */ + -1, /* size of per-interpreter state of the module, + or -1 if the module keeps state in global variables. */ + method_def +}; + + +#if PY_MAJOR_VERSION == 2 +PyMODINIT_FUNC init_libbleu() +#else +PyMODINIT_FUNC PyInit_libbleu() +#endif +{ + PyObject *m = PyModule_Create(&module_def); + if (!m) { + return NULL; + } + return m; +} diff --git a/fairseq/clib/libnat/edit_dist.cpp b/fairseq/clib/libnat/edit_dist.cpp new file mode 100644 index 0000000000000000000000000000000000000000..6bc6a937d6abde0cd49769c4def69ac0560096bc --- /dev/null +++ b/fairseq/clib/libnat/edit_dist.cpp @@ -0,0 +1,231 @@ +/** + * Copyright 2017-present, Facebook, Inc. + * All rights reserved. + * + * This source code is licensed under the license found in the + * LICENSE file in the root directory of this source tree. + */ + +#include <torch/torch.h> // @manual=//caffe2:torch_extension +#include <pybind11/detail/common.h> +#include <pybind11/pybind11.h> +#include <vector> +#include <algorithm> +#include <cstdint> +#include <iosfwd> +#include <memory> +#include <new> +#include <string> +#include <utility> + +using namespace ::std; + +vector<vector<uint32_t>> edit_distance2_with_dp( + vector<uint32_t>& x, + vector<uint32_t>& y) { + uint32_t lx = x.size(); + uint32_t ly = y.size(); + vector<vector<uint32_t>> d(lx + 1, vector<uint32_t>(ly + 1)); + for (uint32_t i = 0; i < lx + 1; i++) { + d[i][0] = i; + } + for (uint32_t j = 0; j < ly + 1; j++) { + d[0][j] = j; + } + for (uint32_t i = 1; i < lx + 1; i++) { + for (uint32_t j = 1; j < ly + 1; j++) { + d[i][j] = + min(min(d[i - 1][j], d[i][j - 1]) + 1, + d[i - 1][j - 1] + 2 * (x.at(i - 1) == y.at(j - 1) ? 0 : 1)); + } + } + return d; +} + +vector<vector<uint32_t>> edit_distance2_backtracking( + vector<vector<uint32_t>>& d, + vector<uint32_t>& x, + vector<uint32_t>& y, + uint32_t terminal_symbol) { + vector<uint32_t> seq; + vector<vector<uint32_t>> edit_seqs(x.size() + 2, vector<uint32_t>()); + /* + edit_seqs: + 0~x.size() cell is the insertion sequences + last cell is the delete sequence + */ + + if (x.size() == 0) { + edit_seqs.at(0) = y; + return edit_seqs; + } + + uint32_t i = d.size() - 1; + uint32_t j = d.at(0).size() - 1; + + while ((i >= 0) && (j >= 0)) { + if ((i == 0) && (j == 0)) { + break; + } + + if ((j > 0) && (d.at(i).at(j - 1) < d.at(i).at(j))) { + seq.push_back(1); // insert + seq.push_back(y.at(j - 1)); + j--; + } else if ((i > 0) && (d.at(i - 1).at(j) < d.at(i).at(j))) { + seq.push_back(2); // delete + seq.push_back(x.at(i - 1)); + i--; + } else { + seq.push_back(3); // keep + seq.push_back(x.at(i - 1)); + i--; + j--; + } + } + + uint32_t prev_op, op, s, word; + prev_op = 0, s = 0; + for (uint32_t k = 0; k < seq.size() / 2; k++) { + op = seq.at(seq.size() - 2 * k - 2); + word = seq.at(seq.size() - 2 * k - 1); + if (prev_op != 1) { + s++; + } + if (op == 1) // insert + { + edit_seqs.at(s - 1).push_back(word); + } else if (op == 2) // delete + { + edit_seqs.at(x.size() + 1).push_back(1); + } else { + edit_seqs.at(x.size() + 1).push_back(0); + } + + prev_op = op; + } + + for (uint32_t k = 0; k < edit_seqs.size(); k++) { + if (edit_seqs[k].size() == 0) { + edit_seqs[k].push_back(terminal_symbol); + } + } + return edit_seqs; +} + +vector<vector<uint32_t>> edit_distance2_backtracking_with_delete( + vector<vector<uint32_t>>& d, + vector<uint32_t>& x, + vector<uint32_t>& y, + uint32_t terminal_symbol, + uint32_t deletion_symbol) { + vector<uint32_t> seq; + vector<vector<uint32_t>> edit_seqs(x.size() + 1, vector<uint32_t>()); + /* + edit_seqs: + 0~x.size() cell is the insertion sequences + last cell is the delete sequence + */ + + if (x.size() == 0) { + edit_seqs.at(0) = y; + return edit_seqs; + } + + uint32_t i = d.size() - 1; + uint32_t j = d.at(0).size() - 1; + + while ((i >= 0) && (j >= 0)) { + if ((i == 0) && (j == 0)) { + break; + } + + if ((j > 0) && (d.at(i).at(j - 1) < d.at(i).at(j))) { + seq.push_back(1); // insert + seq.push_back(y.at(j - 1)); + j--; + } else if ((i > 0) && (d.at(i - 1).at(j) < d.at(i).at(j))) { + seq.push_back(2); // delete + seq.push_back(x.at(i - 1)); + i--; + } else { + seq.push_back(3); // keep + seq.push_back(x.at(i - 1)); + i--; + j--; + } + } + + uint32_t prev_op, op, s, word; + prev_op = 0, s = 0; + for (uint32_t k = 0; k < seq.size() / 2; k++) { + op = seq.at(seq.size() - 2 * k - 2); + word = seq.at(seq.size() - 2 * k - 1); + if (prev_op != 1) { + s++; + } + if (op == 1) // insert + { + edit_seqs.at(s - 1).push_back(word); + } else if (op == 2) // delete + { + edit_seqs.at(s - 1).push_back(deletion_symbol); + } + + prev_op = op; + } + + for (uint32_t k = 0; k < edit_seqs.size(); k++) { + if (edit_seqs.at(k).size() == 0) { + edit_seqs.at(k).push_back(terminal_symbol); + } + } + return edit_seqs; +} + +vector<uint32_t> compute_ed2( + vector<vector<uint32_t>>& xs, + vector<vector<uint32_t>>& ys) { + vector<uint32_t> distances(xs.size()); + for (uint32_t i = 0; i < xs.size(); i++) { + vector<vector<uint32_t>> d = edit_distance2_with_dp(xs.at(i), ys.at(i)); + distances.at(i) = d.at(xs.at(i).size()).at(ys.at(i).size()); + } + return distances; +} + +vector<vector<vector<uint32_t>>> suggested_ed2_path( + vector<vector<uint32_t>>& xs, + vector<vector<uint32_t>>& ys, + uint32_t terminal_symbol) { + vector<vector<vector<uint32_t>>> seq(xs.size()); + for (uint32_t i = 0; i < xs.size(); i++) { + vector<vector<uint32_t>> d = edit_distance2_with_dp(xs.at(i), ys.at(i)); + seq.at(i) = + edit_distance2_backtracking(d, xs.at(i), ys.at(i), terminal_symbol); + } + return seq; +} + +vector<vector<vector<uint32_t>>> suggested_ed2_path_with_delete( + vector<vector<uint32_t>>& xs, + vector<vector<uint32_t>>& ys, + uint32_t terminal_symbol, + uint32_t deletion_symbol) { + vector<vector<vector<uint32_t>>> seq(xs.size()); + for (uint32_t i = 0; i < xs.size(); i++) { + vector<vector<uint32_t>> d = edit_distance2_with_dp(xs.at(i), ys.at(i)); + seq.at(i) = edit_distance2_backtracking_with_delete( + d, xs.at(i), ys.at(i), terminal_symbol, deletion_symbol); + } + return seq; +} + +PYBIND11_MODULE(libnat, m) { + m.def("compute_ed2", &compute_ed2, "compute_ed2"); + m.def("suggested_ed2_path", &suggested_ed2_path, "suggested_ed2_path"); + m.def( + "suggested_ed2_path_with_delete", + &suggested_ed2_path_with_delete, + "suggested_ed2_path_with_delete"); +} diff --git a/fairseq/clib/libnat_cuda/binding.cpp b/fairseq/clib/libnat_cuda/binding.cpp new file mode 100644 index 0000000000000000000000000000000000000000..aaa6244d5c6819acfae5f408280205661a3389ae --- /dev/null +++ b/fairseq/clib/libnat_cuda/binding.cpp @@ -0,0 +1,60 @@ +/** + * Copyright 2017-present, Facebook, Inc. + * All rights reserved. + * + * This source code is licensed under the license found in the + * LICENSE file in the root directory of this source tree. + */ + +/* + This code is partially adpoted from https://github.com/1ytic/pytorch-edit-distance + */ + +#include "edit_dist.h" +#include <torch/types.h> + +#ifndef TORCH_CHECK +#define TORCH_CHECK AT_CHECK +#endif + +#define CHECK_CUDA(x) TORCH_CHECK(x.type().is_cuda(), #x " must be a CUDA tensor") +#define CHECK_CONTIGUOUS(x) TORCH_CHECK(x.is_contiguous(), #x " must be contiguous") +#define CHECK_INPUT(x) CHECK_CUDA(x); CHECK_CONTIGUOUS(x) + + +torch::Tensor LevenshteinDistance( + torch::Tensor source, + torch::Tensor target, + torch::Tensor source_length, + torch::Tensor target_length) { + + CHECK_INPUT(source); + CHECK_INPUT(target); + CHECK_INPUT(source_length); + CHECK_INPUT(target_length); + return LevenshteinDistanceCuda(source, target, source_length, target_length); +} + +torch::Tensor GenerateDeletionLabel( + torch::Tensor source, + torch::Tensor operations) { + + CHECK_INPUT(source); + CHECK_INPUT(operations); + return GenerateDeletionLabelCuda(source, operations); +} + +std::pair<torch::Tensor, torch::Tensor> GenerateInsertionLabel( + torch::Tensor target, + torch::Tensor operations) { + + CHECK_INPUT(target); + CHECK_INPUT(operations); + return GenerateInsertionLabelCuda(target, operations); +} + +PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { + m.def("levenshtein_distance", &LevenshteinDistance, "Levenshtein distance"); + m.def("generate_deletion_labels", &GenerateDeletionLabel, "Generate Deletion Label"); + m.def("generate_insertion_labels", &GenerateInsertionLabel, "Generate Insertion Label"); +} diff --git a/fairseq/clib/libnat_cuda/edit_dist.cu b/fairseq/clib/libnat_cuda/edit_dist.cu new file mode 100644 index 0000000000000000000000000000000000000000..22de16b270851227348c43d6adfb763c8a325df6 --- /dev/null +++ b/fairseq/clib/libnat_cuda/edit_dist.cu @@ -0,0 +1,332 @@ +/** +* Copyright 2017-present, Facebook, Inc. +* All rights reserved. +* +* This source code is licensed under the license found in the +* LICENSE file in the root directory of this source tree. +*/ + +#include "edit_dist.h" +#include <THC/THC.h> +#include <cuda.h> +#include <cuda_runtime.h> +#include <device_launch_parameters.h> +#include <utility> // std::pair + +template <typename scalar_t> +__global__ void generate_deletion_label_kernel( + const scalar_t* __restrict__ source, + const size_t source_size, + const size_t operation_size, + int* __restrict__ operations, + int* __restrict__ labels) { + + const int index = blockIdx.x; + const int offset = index * operation_size; + const int offset_label = index * source_size; + + for (int i = 0; i < source_size; i++) { + labels[offset_label + i] = 0; + } + + int k = 0; + for (int i = 0; i < operation_size; i++){ + if (operations[offset + i] == 0){ + break; + } else if (operations[offset + i] == 1){ + continue; + } else { + labels[offset_label + k] = 3 - operations[offset + i]; + k++; + } + } +} + +template <typename scalar_t> +__global__ void generate_insertion_label_kernel( + const scalar_t* __restrict__ target, + const size_t target_size, + const size_t operation_size, + int* __restrict__ operations, + int* __restrict__ labels, + int* __restrict__ masks) { + + const int index = blockIdx.x; + const int offset = index * operation_size; + const int offset_label = index * target_size; + + int k = 0; + int u = 0; + int m = 0; + + for (int i = 0; i < target_size; i++) { + labels[offset_label + i] = 0; + masks[offset_label + i] = 0; + } + + for (int i = 0; i < operation_size-1; i++){ + if (operations[offset + i] == 0){ + break; + } else if (operations[offset + i] == 2){ + continue; + } else if (operations[offset + i] == 1){ + masks[offset_label + m] = 1; + u++; m++; + } else { + labels[offset_label + k] = u; + masks[offset_label + m] = 0; + k++; m++; + u = 0; + } + } +} + +template <typename scalar_t> +__global__ void levenshtein_distance_kernel( + const scalar_t* __restrict__ source, + const scalar_t* __restrict__ target, + const int* __restrict__ source_length, + const int* __restrict__ target_length, + const size_t source_size, + const size_t target_size, + int* __restrict__ operations, + int* __restrict__ errors_curr) { + + const int index = blockIdx.x; + const int offset = index * (source_size + target_size); + const int d = index * (source_size + 1) * (target_size + 1); + const int t = target_size + 1; + + auto err_idx = [d, t](int i, int j) { return d + i * t + j; }; + auto opt_idx = [offset](int k) { return offset + k; }; + + const int hyp_len = source_length[index]; + const int ref_len = target_length[index]; + const scalar_t* hyp_begin = source + index * source_size; + const scalar_t* ref_begin = target + index * target_size; + + // dynamic programming + for (int i = 0; i <= hyp_len; i++){ + errors_curr[err_idx(i, 0)] = i; + } + for (int j = 0; j <= ref_len; j++){ + errors_curr[err_idx(0, j)] = j; + } + for (int i = 1; i <= hyp_len; i++){ + for (int j = 1; j <= ref_len; j++){ + errors_curr[err_idx(i, j)] = min( + min( + errors_curr[err_idx(i-1, j)], + errors_curr[err_idx(i, j-1)] + ) + 1, + errors_curr[err_idx(i-1, j-1)] + 2 * ( + *(hyp_begin+i-1) == *(ref_begin+j-1) ? 0 : 1 + ) + ); + } + } + + // back-tracing + int i = hyp_len; + int j = ref_len; + int o = hyp_len + ref_len; + + for (int k = 0; k < source_size + target_size; k++) { + operations[opt_idx(k)] = 0; + } + + while ((i >= 0) && (j >= 0)) { + if ((i == 0) && (j == 0)) { + break; + } + + if ((j > 0) && (errors_curr[err_idx(i, j-1)] < errors_curr[err_idx(i, j)])) { + o--; operations[opt_idx(o)] = 1; j--; // insertion + } else if ((i > 0) && (errors_curr[err_idx(i-1, j)] < errors_curr[err_idx(i, j)])) { + o--; operations[opt_idx(o)] = 2; i--; // deletion + } else { + o--; operations[opt_idx(o)] = 3; i--; j--; // do nothing + } + } + + // moving to the left + for (int k = 0; k < hyp_len + ref_len; k++) { + if (k + o < hyp_len + ref_len){ + operations[opt_idx(k)] = operations[opt_idx(k+o)]; + } else{ + operations[opt_idx(k)] = 0; // padding + } + } + +} + +template <typename scalar_t> +__global__ void faster_levenshtein_distance_kernel( + const scalar_t* __restrict__ source, + const scalar_t* __restrict__ target, + const int* __restrict__ source_length, + const int* __restrict__ target_length, + const size_t source_size, + const size_t target_size, + int* __restrict__ operations) { + + extern __shared__ short errors[]; + auto errors_curr = errors; + + const int index = blockIdx.x; + const int offset = index * (source_size + target_size); + const int t = target_size + 1; + + auto err_idx = [t](int i, int j) { return i * t + j; }; + auto opt_idx = [offset](int k) { return offset + k; }; + + const int hyp_len = source_length[index]; + const int ref_len = target_length[index]; + const scalar_t* hyp_begin = source + index * source_size; + const scalar_t* ref_begin = target + index * target_size; + + // dynamic programming + for (int i = 0; i <= hyp_len; i++){ + errors_curr[err_idx(i, 0)] = i; + } + for (int j = 0; j <= ref_len; j++){ + errors_curr[err_idx(0, j)] = j; + } + for (int i = 1; i <= hyp_len; i++){ + for (int j = 1; j <= ref_len; j++){ + errors_curr[err_idx(i, j)] = min( + min( + errors_curr[err_idx(i-1, j)], + errors_curr[err_idx(i, j-1)] + ) + 1, + errors_curr[err_idx(i-1, j-1)] + 2 * ( + *(hyp_begin+i-1) == *(ref_begin+j-1) ? 0 : 1 + ) + ); + } + } + + // back-tracing + int i = hyp_len; + int j = ref_len; + int o = hyp_len + ref_len; + + for (int k = 0; k < source_size + target_size; k++) { + operations[opt_idx(k)] = 0; + } + + while ((i >= 0) && (j >= 0)) { + if ((i == 0) && (j == 0)) { + break; + } + + if ((j > 0) && (errors_curr[err_idx(i, j-1)] < errors_curr[err_idx(i, j)])) { + o--; operations[opt_idx(o)] = 1; j--; // insertion + } else if ((i > 0) && (errors_curr[err_idx(i-1, j)] < errors_curr[err_idx(i, j)])) { + o--; operations[opt_idx(o)] = 2; i--; // deletion + } else { + o--; operations[opt_idx(o)] = 3; i--; j--; // do nothing + } + } + + // moving to the left + for (int k = 0; k < hyp_len + ref_len; k++) { + if (k + o < hyp_len + ref_len){ + operations[opt_idx(k)] = operations[opt_idx(k+o)]; + } else{ + operations[opt_idx(k)] = 0; // padding + } + } + +} + + +torch::Tensor GenerateDeletionLabelCuda( + torch::Tensor source, + torch::Tensor operations) { + + const auto batch_size = source.size(0); + at::TensorOptions options(source.device()); + options = options.dtype(at::ScalarType::Int); + auto labels = torch::empty({batch_size, source.size(1)}, options); + auto stream = at::cuda::getCurrentCUDAStream(source.device().index()); + + AT_DISPATCH_ALL_TYPES(source.scalar_type(), "generate_deletion_labels", ([&] { + generate_deletion_label_kernel<scalar_t><<<batch_size, 1, 0, stream>>>( + source.data_ptr<scalar_t>(), + source.size(1), + operations.size(1), + operations.data_ptr<int>(), + labels.data_ptr<int>()); + })); + + return labels; +} + +std::pair<torch::Tensor, torch::Tensor> GenerateInsertionLabelCuda( + torch::Tensor target, + torch::Tensor operations) { + +const auto batch_size = target.size(0); +at::TensorOptions options(target.device()); +options = options.dtype(at::ScalarType::Int); +auto labels = torch::empty({batch_size, target.size(1)}, options); +auto masks = torch::empty({batch_size, target.size(1)}, options); +auto stream = at::cuda::getCurrentCUDAStream(target.device().index()); + +AT_DISPATCH_ALL_TYPES(target.scalar_type(), "generate_insertion_labels", ([&] { + generate_insertion_label_kernel<scalar_t><<<batch_size, 1, 0, stream>>>( + target.data_ptr<scalar_t>(), + target.size(1), + operations.size(1), + operations.data_ptr<int>(), + labels.data_ptr<int>(), + masks.data_ptr<int>()); +})); + +return std::make_pair(labels, masks); +} + + +torch::Tensor LevenshteinDistanceCuda( + torch::Tensor source, + torch::Tensor target, + torch::Tensor source_length, + torch::Tensor target_length) { + + const auto batch_size = source.size(0); + const auto shared_size = (source.size(1) + 1) * (target.size(1) + 1) * sizeof(short); + + at::TensorOptions options(source.device()); + options = options.dtype(at::ScalarType::Int); + auto operations = torch::empty({batch_size, source.size(1) + target.size(1)}, options); + auto stream = at::cuda::getCurrentCUDAStream(source.device().index()); + + if (shared_size > 40000) { + auto distances = torch::empty({batch_size, (source.size(1) + 1) * (target.size(1) + 1)}, options); + AT_DISPATCH_ALL_TYPES(source.scalar_type(), "levenshtein_distance", ([&] { + levenshtein_distance_kernel<scalar_t><<<batch_size, 1, 0, stream>>>( + source.data_ptr<scalar_t>(), + target.data_ptr<scalar_t>(), + source_length.data_ptr<int>(), + target_length.data_ptr<int>(), + source.size(1), + target.size(1), + operations.data_ptr<int>(), + distances.data_ptr<int>()); + })); + } else { + AT_DISPATCH_ALL_TYPES(source.scalar_type(), "faster_levenshtein_distance", ([&] { + faster_levenshtein_distance_kernel<scalar_t><<<batch_size, 1, shared_size, stream>>>( + source.data_ptr<scalar_t>(), + target.data_ptr<scalar_t>(), + source_length.data_ptr<int>(), + target_length.data_ptr<int>(), + source.size(1), + target.size(1), + operations.data_ptr<int>()); + })); + } + + return operations; +} diff --git a/fairseq/clib/libnat_cuda/edit_dist.h b/fairseq/clib/libnat_cuda/edit_dist.h new file mode 100644 index 0000000000000000000000000000000000000000..e3506cd34ddaa35bb724fe64a459bad8046b9a34 --- /dev/null +++ b/fairseq/clib/libnat_cuda/edit_dist.h @@ -0,0 +1,25 @@ +/** + * Copyright 2017-present, Facebook, Inc. + * All rights reserved. + * + * This source code is licensed under the license found in the + * LICENSE file in the root directory of this source tree. + */ + +#pragma once + +#include <torch/extension.h> + +torch::Tensor LevenshteinDistanceCuda( + torch::Tensor source, + torch::Tensor target, + torch::Tensor source_length, + torch::Tensor target_length); + +torch::Tensor GenerateDeletionLabelCuda( + torch::Tensor source, + torch::Tensor operations); + +std::pair<torch::Tensor, torch::Tensor> GenerateInsertionLabelCuda( + torch::Tensor source, + torch::Tensor operations); diff --git a/fairseq/config/__init__.py b/fairseq/config/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..6264236915a7269a4d920ee8213004374dd86a9a --- /dev/null +++ b/fairseq/config/__init__.py @@ -0,0 +1,4 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. diff --git a/fairseq/config/config.yaml b/fairseq/config/config.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e20d914b9b1620b21e702f1114aaf1131c3f6c55 --- /dev/null +++ b/fairseq/config/config.yaml @@ -0,0 +1,18 @@ +# @package _group_ + +hydra: + run: + dir: . + +defaults: + - task: null + - model: null + - criterion: cross_entropy + - optimizer: null + - lr_scheduler: fixed + - bpe: null + - tokenizer: null + - scoring: null + - generation: null + - common_eval: null + - eval_lm: null diff --git a/fairseq/config/model/transformer_lm/transformer_lm_baevski_gbw.yaml b/fairseq/config/model/transformer_lm/transformer_lm_baevski_gbw.yaml new file mode 100644 index 0000000000000000000000000000000000000000..30b1a4f1e0f5e7f7c2671ff8ec995cc32363f10f --- /dev/null +++ b/fairseq/config/model/transformer_lm/transformer_lm_baevski_gbw.yaml @@ -0,0 +1,36 @@ +# @package _group_ +activation_fn: "relu" +dropout: 0.1 +attention_dropout: 0.1 +activation_dropout: 0.0 +relu_dropout: 0.0 +decoder_embed_dim: 512 +decoder_output_dim: 512 +decoder_input_dim: 512 +decoder_ffn_embed_dim: 4096 +decoder_layers: 12 +decoder_attention_heads: 16 +decoder_normalize_before: true +no_decoder_final_norm: true +adaptive_softmax_cutoff: null +adaptive_softmax_dropout: 0 +adaptive_softmax_factor: 4 +no_token_positional_embeddings: false +share_decoder_input_output_embed: false +character_embeddings: false +character_filters: "[(1, 64), (2, 128), (3, 192), (4, 256), (5, 256), (6, 256), (7, 256)]" +character_embedding_dim: 4 +char_embedder_highway_layers: 2 +adaptive_input: false +adaptive_input_factor: 4 +adaptive_input_cutoff: null +tie_adaptive_weights: false +tie_adaptive_proj: false +decoder_learned_pos: false +decoder_layerdrop: 0 +decoder_layers_to_keep: null +layernorm_embedding: false +no_scale_embedding: false +quant_noise_pq: 0 +quant_noise_pq_block_size: 8 +quant_noise_scalar: 0 diff --git a/fairseq/config/model/transformer_lm/transformer_lm_baevski_wiki103.yaml b/fairseq/config/model/transformer_lm/transformer_lm_baevski_wiki103.yaml new file mode 100644 index 0000000000000000000000000000000000000000..1154cfa660ee5ce6a272cd1a0049eead1e92c117 --- /dev/null +++ b/fairseq/config/model/transformer_lm/transformer_lm_baevski_wiki103.yaml @@ -0,0 +1,36 @@ +# @package _group_ +activation_fn: "relu" +dropout: 0.3 +attention_dropout: 0.1 +activation_dropout: 0.1 +relu_dropout: 0.1 +decoder_embed_dim: 1024 +decoder_output_dim: 1024 +decoder_input_dim: 1024 +decoder_ffn_embed_dim: 4096 +decoder_layers: 16 +decoder_attention_heads: 8 +decoder_normalize_before: true +no_decoder_final_norm: true +adaptive_softmax_cutoff: "20000,60000" +adaptive_softmax_dropout: 0.2 +adaptive_softmax_factor: 4 +no_token_positional_embeddings: false +share_decoder_input_output_embed: false +character_embeddings: false +character_filters: "[(1, 64), (2, 128), (3, 192), (4, 256), (5, 256), (6, 256), (7, 256)]" +character_embedding_dim: 4 +char_embedder_highway_layers: 2 +adaptive_input: true +adaptive_input_factor: 4 +adaptive_input_cutoff: "20000,60000" +tie_adaptive_weights: true +tie_adaptive_proj: true +decoder_learned_pos: false +decoder_layerdrop: 0 +decoder_layers_to_keep: null +layernorm_embedding: false +no_scale_embedding: false +quant_noise_pq: 0 +quant_noise_pq_block_size: 8 +quant_noise_scalar: 0 diff --git a/fairseq/config/model/transformer_lm/transformer_lm_big.yaml b/fairseq/config/model/transformer_lm/transformer_lm_big.yaml new file mode 100644 index 0000000000000000000000000000000000000000..309575310bfc5d9c5cde31563073bef18abc646e --- /dev/null +++ b/fairseq/config/model/transformer_lm/transformer_lm_big.yaml @@ -0,0 +1,36 @@ +# @package _group_ +activation_fn: "relu" +dropout: 0.1 +attention_dropout: 0.0 +activation_dropout: 0.0 +relu_dropout: 0.0 +decoder_embed_dim: 1024 +decoder_output_dim: 1024 +decoder_input_dim: 1024 +decoder_ffn_embed_dim: 4096 +decoder_layers: 12 +decoder_attention_heads: 16 +decoder_normalize_before: true +no_decoder_final_norm: false +adaptive_softmax_cutoff: null +adaptive_softmax_dropout: 0 +adaptive_softmax_factor: 4 +no_token_positional_embeddings: false +share_decoder_input_output_embed: false +character_embeddings: false +character_filters: "[(1, 64), (2, 128), (3, 192), (4, 256), (5, 256), (6, 256), (7, 256)]" +character_embedding_dim: 4 +char_embedder_highway_layers: 2 +adaptive_input: false +adaptive_input_factor: 4 +adaptive_input_cutoff: null +tie_adaptive_weights: false +tie_adaptive_proj: false +decoder_learned_pos: false +decoder_layerdrop: 0 +decoder_layers_to_keep: null +layernorm_embedding: false +no_scale_embedding: false +quant_noise_pq: 0 +quant_noise_pq_block_size: 8 +quant_noise_scalar: 0 diff --git a/fairseq/config/model/transformer_lm/transformer_lm_gbw.yaml b/fairseq/config/model/transformer_lm/transformer_lm_gbw.yaml new file mode 100644 index 0000000000000000000000000000000000000000..30b1a4f1e0f5e7f7c2671ff8ec995cc32363f10f --- /dev/null +++ b/fairseq/config/model/transformer_lm/transformer_lm_gbw.yaml @@ -0,0 +1,36 @@ +# @package _group_ +activation_fn: "relu" +dropout: 0.1 +attention_dropout: 0.1 +activation_dropout: 0.0 +relu_dropout: 0.0 +decoder_embed_dim: 512 +decoder_output_dim: 512 +decoder_input_dim: 512 +decoder_ffn_embed_dim: 4096 +decoder_layers: 12 +decoder_attention_heads: 16 +decoder_normalize_before: true +no_decoder_final_norm: true +adaptive_softmax_cutoff: null +adaptive_softmax_dropout: 0 +adaptive_softmax_factor: 4 +no_token_positional_embeddings: false +share_decoder_input_output_embed: false +character_embeddings: false +character_filters: "[(1, 64), (2, 128), (3, 192), (4, 256), (5, 256), (6, 256), (7, 256)]" +character_embedding_dim: 4 +char_embedder_highway_layers: 2 +adaptive_input: false +adaptive_input_factor: 4 +adaptive_input_cutoff: null +tie_adaptive_weights: false +tie_adaptive_proj: false +decoder_learned_pos: false +decoder_layerdrop: 0 +decoder_layers_to_keep: null +layernorm_embedding: false +no_scale_embedding: false +quant_noise_pq: 0 +quant_noise_pq_block_size: 8 +quant_noise_scalar: 0 diff --git a/fairseq/config/model/transformer_lm/transformer_lm_gpt.yaml b/fairseq/config/model/transformer_lm/transformer_lm_gpt.yaml new file mode 100644 index 0000000000000000000000000000000000000000..2c6cb7be3801115371566932ffc78651c9ac6c0f --- /dev/null +++ b/fairseq/config/model/transformer_lm/transformer_lm_gpt.yaml @@ -0,0 +1,36 @@ +# @package _group_ +activation_fn: "gelu" +dropout: 0.1 +attention_dropout: 0.1 +activation_dropout: 0.0 +relu_dropout: 0.0 +decoder_embed_dim: 768 +decoder_output_dim: 768 +decoder_input_dim: 768 +decoder_ffn_embed_dim: 3072 +decoder_layers: 12 +decoder_attention_heads: 12 +decoder_normalize_before: true +no_decoder_final_norm: false +adaptive_softmax_cutoff: null +adaptive_softmax_dropout: 0 +adaptive_softmax_factor: 4 +no_token_positional_embeddings: false +share_decoder_input_output_embed: false +character_embeddings: false +character_filters: "[(1, 64), (2, 128), (3, 192), (4, 256), (5, 256), (6, 256), (7, 256)]" +character_embedding_dim: 4 +char_embedder_highway_layers: 2 +adaptive_input: false +adaptive_input_factor: 4 +adaptive_input_cutoff: null +tie_adaptive_weights: false +tie_adaptive_proj: false +decoder_learned_pos: false +decoder_layerdrop: 0 +decoder_layers_to_keep: null +layernorm_embedding: false +no_scale_embedding: false +quant_noise_pq: 0 +quant_noise_pq_block_size: 8 +quant_noise_scalar: 0 diff --git a/fairseq/config/model/transformer_lm/transformer_lm_gpt2_big.yaml b/fairseq/config/model/transformer_lm/transformer_lm_gpt2_big.yaml new file mode 100644 index 0000000000000000000000000000000000000000..a08769a1781abdb13302bf57bf1338bcaf68a0ec --- /dev/null +++ b/fairseq/config/model/transformer_lm/transformer_lm_gpt2_big.yaml @@ -0,0 +1,36 @@ +# @package _group_ +activation_fn: "gelu" +dropout: 0.1 +attention_dropout: 0.1 +activation_dropout: 0.0 +relu_dropout: 0.0 +decoder_embed_dim: 1600 +decoder_output_dim: 1600 +decoder_input_dim: 1600 +decoder_ffn_embed_dim: 6400 +decoder_layers: 48 +decoder_attention_heads: 25 +decoder_normalize_before: true +no_decoder_final_norm: false +adaptive_softmax_cutoff: null +adaptive_softmax_dropout: 0 +adaptive_softmax_factor: 4 +no_token_positional_embeddings: false +share_decoder_input_output_embed: false +character_embeddings: false +character_filters: "[(1, 64), (2, 128), (3, 192), (4, 256), (5, 256), (6, 256), (7, 256)]" +character_embedding_dim: 4 +char_embedder_highway_layers: 2 +adaptive_input: false +adaptive_input_factor: 4 +adaptive_input_cutoff: null +tie_adaptive_weights: false +tie_adaptive_proj: false +decoder_learned_pos: false +decoder_layerdrop: 0 +decoder_layers_to_keep: null +layernorm_embedding: false +no_scale_embedding: false +quant_noise_pq: 0 +quant_noise_pq_block_size: 8 +quant_noise_scalar: 0 diff --git a/fairseq/config/model/transformer_lm/transformer_lm_gpt2_medium.yaml b/fairseq/config/model/transformer_lm/transformer_lm_gpt2_medium.yaml new file mode 100644 index 0000000000000000000000000000000000000000..64261d793c0f1ae091c9bf5c8c77093a07326137 --- /dev/null +++ b/fairseq/config/model/transformer_lm/transformer_lm_gpt2_medium.yaml @@ -0,0 +1,36 @@ +# @package _group_ +activation_fn: "gelu" +dropout: 0.1 +attention_dropout: 0.1 +activation_dropout: 0.0 +relu_dropout: 0.0 +decoder_embed_dim: 1280 +decoder_output_dim: 1280 +decoder_input_dim: 1280 +decoder_ffn_embed_dim: 5120 +decoder_layers: 36 +decoder_attention_heads: 20 +decoder_normalize_before: true +no_decoder_final_norm: false +adaptive_softmax_cutoff: null +adaptive_softmax_dropout: 0 +adaptive_softmax_factor: 4 +no_token_positional_embeddings: false +share_decoder_input_output_embed: false +character_embeddings: false +character_filters: "[(1, 64), (2, 128), (3, 192), (4, 256), (5, 256), (6, 256), (7, 256)]" +character_embedding_dim: 4 +char_embedder_highway_layers: 2 +adaptive_input: false +adaptive_input_factor: 4 +adaptive_input_cutoff: null +tie_adaptive_weights: false +tie_adaptive_proj: false +decoder_learned_pos: false +decoder_layerdrop: 0 +decoder_layers_to_keep: null +layernorm_embedding: false +no_scale_embedding: false +quant_noise_pq: 0 +quant_noise_pq_block_size: 8 +quant_noise_scalar: 0 diff --git a/fairseq/config/model/transformer_lm/transformer_lm_gpt2_small.yaml b/fairseq/config/model/transformer_lm/transformer_lm_gpt2_small.yaml new file mode 100644 index 0000000000000000000000000000000000000000..702e81f466c82edf40433589d389edbe0a7b96db --- /dev/null +++ b/fairseq/config/model/transformer_lm/transformer_lm_gpt2_small.yaml @@ -0,0 +1,36 @@ +# @package _group_ +activation_fn: "gelu" +dropout: 0.1 +attention_dropout: 0.1 +activation_dropout: 0.0 +relu_dropout: 0.0 +decoder_embed_dim: 1024 +decoder_output_dim: 1024 +decoder_input_dim: 1024 +decoder_ffn_embed_dim: 4096 +decoder_layers: 24 +decoder_attention_heads: 16 +decoder_normalize_before: true +no_decoder_final_norm: false +adaptive_softmax_cutoff: null +adaptive_softmax_dropout: 0 +adaptive_softmax_factor: 4 +no_token_positional_embeddings: false +share_decoder_input_output_embed: false +character_embeddings: false +character_filters: "[(1, 64), (2, 128), (3, 192), (4, 256), (5, 256), (6, 256), (7, 256)]" +character_embedding_dim: 4 +char_embedder_highway_layers: 2 +adaptive_input: false +adaptive_input_factor: 4 +adaptive_input_cutoff: null +tie_adaptive_weights: false +tie_adaptive_proj: false +decoder_learned_pos: false +decoder_layerdrop: 0 +decoder_layers_to_keep: null +layernorm_embedding: false +no_scale_embedding: false +quant_noise_pq: 0 +quant_noise_pq_block_size: 8 +quant_noise_scalar: 0 diff --git a/fairseq/config/model/transformer_lm/transformer_lm_wiki103.yaml b/fairseq/config/model/transformer_lm/transformer_lm_wiki103.yaml new file mode 100644 index 0000000000000000000000000000000000000000..1154cfa660ee5ce6a272cd1a0049eead1e92c117 --- /dev/null +++ b/fairseq/config/model/transformer_lm/transformer_lm_wiki103.yaml @@ -0,0 +1,36 @@ +# @package _group_ +activation_fn: "relu" +dropout: 0.3 +attention_dropout: 0.1 +activation_dropout: 0.1 +relu_dropout: 0.1 +decoder_embed_dim: 1024 +decoder_output_dim: 1024 +decoder_input_dim: 1024 +decoder_ffn_embed_dim: 4096 +decoder_layers: 16 +decoder_attention_heads: 8 +decoder_normalize_before: true +no_decoder_final_norm: true +adaptive_softmax_cutoff: "20000,60000" +adaptive_softmax_dropout: 0.2 +adaptive_softmax_factor: 4 +no_token_positional_embeddings: false +share_decoder_input_output_embed: false +character_embeddings: false +character_filters: "[(1, 64), (2, 128), (3, 192), (4, 256), (5, 256), (6, 256), (7, 256)]" +character_embedding_dim: 4 +char_embedder_highway_layers: 2 +adaptive_input: true +adaptive_input_factor: 4 +adaptive_input_cutoff: "20000,60000" +tie_adaptive_weights: true +tie_adaptive_proj: true +decoder_learned_pos: false +decoder_layerdrop: 0 +decoder_layers_to_keep: null +layernorm_embedding: false +no_scale_embedding: false +quant_noise_pq: 0 +quant_noise_pq_block_size: 8 +quant_noise_scalar: 0 diff --git a/fairseq/config/model/wav2vec/vq_wav2vec_gumbel.yaml b/fairseq/config/model/wav2vec/vq_wav2vec_gumbel.yaml new file mode 100644 index 0000000000000000000000000000000000000000..ee1329bf4612d8bb295c6cc3d8bc0a3bcef1777d --- /dev/null +++ b/fairseq/config/model/wav2vec/vq_wav2vec_gumbel.yaml @@ -0,0 +1,5 @@ +# @package _group_ +activation: gelu +vq_type: gumbel +vq_depth: 2 +combine_groups: true diff --git a/fairseq/config/model/wav2vec2/wav2vec2_base.yaml b/fairseq/config/model/wav2vec2/wav2vec2_base.yaml new file mode 100644 index 0000000000000000000000000000000000000000..ce65499b808b9a3821cee4ca87c36e84d09005a1 --- /dev/null +++ b/fairseq/config/model/wav2vec2/wav2vec2_base.yaml @@ -0,0 +1,8 @@ +# @package _group_ + +quantize_targets: true +final_dim: 256 +encoder_layerdrop: 0.05 +dropout_input: 0.1 +dropout_features: 0.1 +feature_grad_mult: 0.1 diff --git a/fairseq/config/model/wav2vec2/wav2vec2_large.yaml b/fairseq/config/model/wav2vec2/wav2vec2_large.yaml new file mode 100644 index 0000000000000000000000000000000000000000..5846f75243f27f201c85bfe6820815c015971275 --- /dev/null +++ b/fairseq/config/model/wav2vec2/wav2vec2_large.yaml @@ -0,0 +1,20 @@ +# @package _group_ + +quantize_targets: true +extractor_mode: layer_norm +layer_norm_first: true +final_dim: 768 +latent_temp: [2.0,0.1,0.999995] +encoder_layerdrop: 0.0 +dropout_input: 0.0 +dropout_features: 0.0 +dropout: 0.0 +attention_dropout: 0.0 +conv_bias: true + +encoder_layers: 24 +encoder_embed_dim: 1024 +encoder_ffn_embed_dim: 4096 +encoder_attention_heads: 16 + +feature_grad_mult: 1.0 diff --git a/fairseq/criterions/__init__.py b/fairseq/criterions/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..4dbf46a1cb31ce65c4224ae79cbc2d7cf9e4d111 --- /dev/null +++ b/fairseq/criterions/__init__.py @@ -0,0 +1,36 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +"""isort:skip_file""" + +import importlib +import os + +from fairseq import registry +from fairseq.criterions.fairseq_criterion import ( # noqa + FairseqCriterion, + LegacyFairseqCriterion, +) +from omegaconf import DictConfig + + +( + build_criterion_, + register_criterion, + CRITERION_REGISTRY, + CRITERION_DATACLASS_REGISTRY, +) = registry.setup_registry( + "--criterion", base_class=FairseqCriterion, default="cross_entropy" +) + + +def build_criterion(cfg: DictConfig, task): + return build_criterion_(cfg, task) + + +# automatically import any Python files in the criterions/ directory +for file in sorted(os.listdir(os.path.dirname(__file__))): + if file.endswith(".py") and not file.startswith("_"): + file_name = file[: file.find(".py")] + importlib.import_module("fairseq.criterions." + file_name) diff --git a/fairseq/criterions/adaptive_loss.py b/fairseq/criterions/adaptive_loss.py new file mode 100644 index 0000000000000000000000000000000000000000..6209ceaedb6d8120ad820c11b55c13596447933c --- /dev/null +++ b/fairseq/criterions/adaptive_loss.py @@ -0,0 +1,123 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +from dataclasses import dataclass + +import torch.nn.functional as F +from fairseq import metrics, utils +from fairseq.criterions import FairseqCriterion, register_criterion +from fairseq.dataclass import FairseqDataclass +from fairseq.dataclass.constants import DDP_BACKEND_CHOICES +from omegaconf import II + + +@dataclass +class AdaptiveLossConfig(FairseqDataclass): + sentence_avg: bool = II("optimization.sentence_avg") + ddp_backend: DDP_BACKEND_CHOICES = II("distributed_training.ddp_backend") + + +@register_criterion("adaptive_loss", dataclass=AdaptiveLossConfig) +class AdaptiveLoss(FairseqCriterion): + """This is an implementation of the loss function accompanying the adaptive softmax approximation for + graphical processing units (GPU), described in the paper "Efficient softmax approximation for GPUs" + (http://arxiv.org/abs/1609.04309).""" + + def __init__(self, task, sentence_avg): + super().__init__(task) + self.sentence_avg = sentence_avg + + @classmethod + def build_criterion(cls, cfg: AdaptiveLossConfig, task): + if cfg.ddp_backend in {"c10d", "pytorch_ddp"}: + raise Exception( + "AdaptiveLoss is not compatible with the PyTorch " + "version of DistributedDataParallel. Please use " + "`--ddp-backend=legacy_ddp` instead." + ) + return cls(task, cfg.sentence_avg) + + def forward(self, model, sample, reduce=True): + """Compute the loss for the given sample. + + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + + assert ( + hasattr(model.decoder, "adaptive_softmax") + and model.decoder.adaptive_softmax is not None + ) + adaptive_softmax = model.decoder.adaptive_softmax + + net_output = model(**sample["net_input"]) + orig_target = model.get_targets(sample, net_output) + + nsentences = orig_target.size(0) + orig_target = orig_target.view(-1) + + bsz = orig_target.size(0) + + logits, target = adaptive_softmax(net_output[0], orig_target) + assert len(target) == len(logits) + + loss = net_output[0].new(1 if reduce else bsz).zero_() + + for i in range(len(target)): + if target[i] is not None: + assert target[i].min() >= 0 and target[i].max() <= logits[i].size(1) + loss += F.cross_entropy( + logits[i], + target[i], + ignore_index=self.padding_idx, + reduction="sum" if reduce else "none", + ) + + orig = utils.strip_pad(orig_target, self.padding_idx) + ntokens = orig.numel() + sample_size = sample["target"].size(0) if self.sentence_avg else ntokens + logging_output = { + "loss": loss.data, + "ntokens": ntokens, + "nsentences": nsentences, + "sample_size": sample_size, + } + return loss, sample_size, logging_output + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + """Aggregate logging outputs from data parallel training.""" + loss_sum = utils.item(sum(log.get("loss", 0) for log in logging_outputs)) + ntokens = utils.item(sum(log.get("ntokens", 0) for log in logging_outputs)) + sample_size = utils.item( + sum(log.get("sample_size", 0) for log in logging_outputs) + ) + + metrics.log_scalar( + "loss", loss_sum / sample_size / math.log(2), sample_size, round=3 + ) + if sample_size != ntokens: + metrics.log_scalar( + "nll_loss", loss_sum / ntokens / math.log(2), ntokens, round=3 + ) + metrics.log_derived( + "ppl", lambda meters: utils.get_perplexity(meters["nll_loss"].avg) + ) + else: + metrics.log_derived( + "ppl", lambda meters: utils.get_perplexity(meters["loss"].avg) + ) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return True diff --git a/fairseq/criterions/composite_loss.py b/fairseq/criterions/composite_loss.py new file mode 100644 index 0000000000000000000000000000000000000000..98e835fa6e4c0bcad062df9c519701bf795c98be --- /dev/null +++ b/fairseq/criterions/composite_loss.py @@ -0,0 +1,100 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq import utils +from fairseq.criterions import LegacyFairseqCriterion, register_criterion +from torch import nn + + +@register_criterion("composite_loss") +class CompositeLoss(LegacyFairseqCriterion): + """This is a composite loss that, given a list of model outputs and a list of targets, + computes an average of losses for each output-target pair""" + + def __init__(self, args, task): + super().__init__(args, task) + self.underlying_criterion = args.underlying_criterion + + @staticmethod + def add_args(parser): + """Add criterion-specific arguments to the parser.""" + # fmt: off + parser.add_argument('--underlying-criterion', type=str, metavar='VAL', required=True, + help='underlying criterion to use for the composite loss') + # fmt: on + + @staticmethod + def build_underlying_criterion(args, task): + saved_criterion = args.criterion + args.criterion = args.underlying_criterion + assert saved_criterion != args.underlying_criterion + underlying_criterion = task.build_criterion(args) + args.criterion = saved_criterion + return underlying_criterion + + @classmethod + def build_criterion(cls, args, task): + underlying_criterion = CompositeLoss.build_underlying_criterion(args, task) + + class FakeModel(nn.Module): + def __init__(self, model, net_out, target): + super().__init__() + self.model = model + self.net_out = net_out + self.target = target + + def forward(self, **unused): + return self.net_out + + def get_normalized_probs(self, net_output, log_probs, sample=None): + return self.model.get_normalized_probs( + net_output, log_probs, sample=sample + ) + + def get_targets(self, *unused): + return self.target + + @property + def decoder(self): + return self.model.decoder + + class _CompositeLoss(LegacyFairseqCriterion): + def __init__(self, args, task, underlying_criterion): + super().__init__(args, task) + self.underlying_criterion = underlying_criterion + + def forward(self, model, sample, reduce=True): + net_outputs = model(**sample["net_input"]) + targets = sample["target"] + + bsz = targets[0].size(0) + loss = net_outputs[0][0].new(1 if reduce else bsz).float().zero_() + + sample_size = 0 + logging_output = {} + for o, t in zip(net_outputs[0], targets): + m = FakeModel(model, (o, net_outputs[1]), t) + sample["target"] = t + l, ss, logging_output = self.underlying_criterion(m, sample, reduce) + loss += l + sample_size += ss + + loss.div_(len(targets)) + sample_size /= len(targets) + + logging_output["loss"] = utils.item(loss.data) if reduce else loss.data + return loss, sample_size, logging_output + + @staticmethod + def aggregate_logging_outputs(logging_outputs): + return underlying_criterion.__class__.aggregate_logging_outputs( + logging_outputs + ) + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + underlying_criterion.__class__.reduce_metrics(logging_outputs) + + return _CompositeLoss(args, task, underlying_criterion) diff --git a/fairseq/criterions/cross_entropy.py b/fairseq/criterions/cross_entropy.py new file mode 100644 index 0000000000000000000000000000000000000000..fe461064716b38ecf2eb610daddbb609a1884e6b --- /dev/null +++ b/fairseq/criterions/cross_entropy.py @@ -0,0 +1,90 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +from dataclasses import dataclass + +import torch.nn.functional as F +from fairseq import metrics, utils +from fairseq.criterions import FairseqCriterion, register_criterion +from fairseq.dataclass import FairseqDataclass +from omegaconf import II + + +@dataclass +class CrossEntropyCriterionConfig(FairseqDataclass): + sentence_avg: bool = II("optimization.sentence_avg") + + +@register_criterion("cross_entropy", dataclass=CrossEntropyCriterionConfig) +class CrossEntropyCriterion(FairseqCriterion): + def __init__(self, task, sentence_avg): + super().__init__(task) + self.sentence_avg = sentence_avg + + def forward(self, model, sample, reduce=True): + """Compute the loss for the given sample. + + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + net_output = model(**sample["net_input"]) + loss, _ = self.compute_loss(model, net_output, sample, reduce=reduce) + sample_size = ( + sample["target"].size(0) if self.sentence_avg else sample["ntokens"] + ) + logging_output = { + "loss": loss.data, + "ntokens": sample["ntokens"], + "nsentences": sample["target"].size(0), + "sample_size": sample_size, + } + return loss, sample_size, logging_output + + def compute_loss(self, model, net_output, sample, reduce=True): + lprobs = model.get_normalized_probs(net_output, log_probs=True) + lprobs = lprobs.view(-1, lprobs.size(-1)) + target = model.get_targets(sample, net_output).view(-1) + loss = F.nll_loss( + lprobs, + target, + ignore_index=self.padding_idx, + reduction="sum" if reduce else "none", + ) + return loss, loss + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + """Aggregate logging outputs from data parallel training.""" + loss_sum = sum(log.get("loss", 0) for log in logging_outputs) + ntokens = sum(log.get("ntokens", 0) for log in logging_outputs) + sample_size = sum(log.get("sample_size", 0) for log in logging_outputs) + + # we divide by log(2) to convert the loss from base e to base 2 + metrics.log_scalar( + "loss", loss_sum / sample_size / math.log(2), sample_size, round=3 + ) + if sample_size != ntokens: + metrics.log_scalar( + "nll_loss", loss_sum / ntokens / math.log(2), ntokens, round=3 + ) + metrics.log_derived( + "ppl", lambda meters: utils.get_perplexity(meters["nll_loss"].avg) + ) + else: + metrics.log_derived( + "ppl", lambda meters: utils.get_perplexity(meters["loss"].avg) + ) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return True diff --git a/fairseq/criterions/ctc.py b/fairseq/criterions/ctc.py new file mode 100644 index 0000000000000000000000000000000000000000..10e3618382c86a84466cb4264d62f31537980251 --- /dev/null +++ b/fairseq/criterions/ctc.py @@ -0,0 +1,295 @@ +# All rights reserved. +# +# This source code is licensed under the license found in the LICENSE file in +# the root directory of this source tree. An additional grant of patent rights +# can be found in the PATENTS file in the same directory. + +import math +from argparse import Namespace +from dataclasses import dataclass, field +from omegaconf import II +from typing import Optional + +import torch +import torch.nn.functional as F +from fairseq import metrics, utils +from fairseq.criterions import FairseqCriterion, register_criterion +from fairseq.dataclass import FairseqDataclass +from fairseq.data.data_utils import post_process +from fairseq.tasks import FairseqTask +from fairseq.logging.meters import safe_round + + +@dataclass +class CtcCriterionConfig(FairseqDataclass): + zero_infinity: bool = field( + default=False, + metadata={"help": "zero inf loss when source length <= target length"}, + ) + sentence_avg: bool = II("optimization.sentence_avg") + post_process: str = field( + default="letter", + metadata={ + "help": "how to post process predictions into words. can be letter, " + "wordpiece, BPE symbols, etc. " + "See fairseq.data.data_utils.post_process() for full list of options" + }, + ) + wer_kenlm_model: Optional[str] = field( + default=None, + metadata={ + "help": "if this is provided, use kenlm to compute wer (along with other wer_* args)" + }, + ) + wer_lexicon: Optional[str] = field( + default=None, + metadata={"help": "lexicon to use with wer_kenlm_model"}, + ) + wer_lm_weight: float = field( + default=2.0, + metadata={"help": "lm weight to use with wer_kenlm_model"}, + ) + wer_word_score: float = field( + default=-1.0, + metadata={"help": "lm word score to use with wer_kenlm_model"}, + ) + + wer_args: Optional[str] = field( + default=None, + metadata={ + "help": "DEPRECATED: tuple of (wer_kenlm_model, wer_lexicon, wer_lm_weight, wer_word_score)" + }, + ) + + +@register_criterion("ctc", dataclass=CtcCriterionConfig) +class CtcCriterion(FairseqCriterion): + def __init__(self, cfg: CtcCriterionConfig, task: FairseqTask): + super().__init__(task) + self.blank_idx = ( + task.target_dictionary.index(task.blank_symbol) + if hasattr(task, "blank_symbol") + else 0 + ) + self.pad_idx = task.target_dictionary.pad() + self.eos_idx = task.target_dictionary.eos() + self.post_process = cfg.post_process + + if cfg.wer_args is not None: + ( + cfg.wer_kenlm_model, + cfg.wer_lexicon, + cfg.wer_lm_weight, + cfg.wer_word_score, + ) = eval(cfg.wer_args) + + if cfg.wer_kenlm_model is not None: + from examples.speech_recognition.w2l_decoder import W2lKenLMDecoder + + dec_args = Namespace() + dec_args.nbest = 1 + dec_args.criterion = "ctc" + dec_args.kenlm_model = cfg.wer_kenlm_model + dec_args.lexicon = cfg.wer_lexicon + dec_args.beam = 50 + dec_args.beam_size_token = min(50, len(task.target_dictionary)) + dec_args.beam_threshold = min(50, len(task.target_dictionary)) + dec_args.lm_weight = cfg.wer_lm_weight + dec_args.word_score = cfg.wer_word_score + dec_args.unk_weight = -math.inf + dec_args.sil_weight = 0 + + self.w2l_decoder = W2lKenLMDecoder(dec_args, task.target_dictionary) + else: + self.w2l_decoder = None + + self.zero_infinity = cfg.zero_infinity + self.sentence_avg = cfg.sentence_avg + + def forward(self, model, sample, reduce=True): + net_output = model(**sample["net_input"]) + lprobs = model.get_normalized_probs( + net_output, log_probs=True + ).contiguous() # (T, B, C) from the encoder + + if "src_lengths" in sample["net_input"]: + input_lengths = sample["net_input"]["src_lengths"] + else: + if net_output["padding_mask"] is not None: + non_padding_mask = ~net_output["padding_mask"] + input_lengths = non_padding_mask.long().sum(-1) + else: + input_lengths = lprobs.new_full( + (lprobs.size(1),), lprobs.size(0), dtype=torch.long + ) + + pad_mask = (sample["target"] != self.pad_idx) & ( + sample["target"] != self.eos_idx + ) + targets_flat = sample["target"].masked_select(pad_mask) + if "target_lengths" in sample: + target_lengths = sample["target_lengths"] + else: + target_lengths = pad_mask.sum(-1) + + with torch.backends.cudnn.flags(enabled=False): + loss = F.ctc_loss( + lprobs, + targets_flat, + input_lengths, + target_lengths, + blank=self.blank_idx, + reduction="sum", + zero_infinity=self.zero_infinity, + ) + + ntokens = ( + sample["ntokens"] if "ntokens" in sample else target_lengths.sum().item() + ) + + sample_size = sample["target"].size(0) if self.sentence_avg else ntokens + logging_output = { + "loss": utils.item(loss.data), # * sample['ntokens'], + "ntokens": ntokens, + "nsentences": sample["id"].numel(), + "sample_size": sample_size, + } + + if not model.training: + import editdistance + + with torch.no_grad(): + lprobs_t = lprobs.transpose(0, 1).float().contiguous().cpu() + + c_err = 0 + c_len = 0 + w_errs = 0 + w_len = 0 + wv_errs = 0 + for lp, t, inp_l in zip( + lprobs_t, + sample["target_label"] + if "target_label" in sample + else sample["target"], + input_lengths, + ): + lp = lp[:inp_l].unsqueeze(0) + + decoded = None + if self.w2l_decoder is not None: + decoded = self.w2l_decoder.decode(lp) + if len(decoded) < 1: + decoded = None + else: + decoded = decoded[0] + if len(decoded) < 1: + decoded = None + else: + decoded = decoded[0] + + p = (t != self.task.target_dictionary.pad()) & ( + t != self.task.target_dictionary.eos() + ) + targ = t[p] + targ_units = self.task.target_dictionary.string(targ) + targ_units_arr = targ.tolist() + + toks = lp.argmax(dim=-1).unique_consecutive() + pred_units_arr = toks[toks != self.blank_idx].tolist() + + c_err += editdistance.eval(pred_units_arr, targ_units_arr) + c_len += len(targ_units_arr) + + targ_words = post_process(targ_units, self.post_process).split() + + pred_units = self.task.target_dictionary.string(pred_units_arr) + pred_words_raw = post_process(pred_units, self.post_process).split() + + if decoded is not None and "words" in decoded: + pred_words = decoded["words"] + w_errs += editdistance.eval(pred_words, targ_words) + wv_errs += editdistance.eval(pred_words_raw, targ_words) + else: + dist = editdistance.eval(pred_words_raw, targ_words) + w_errs += dist + wv_errs += dist + + w_len += len(targ_words) + + logging_output["wv_errors"] = wv_errs + logging_output["w_errors"] = w_errs + logging_output["w_total"] = w_len + logging_output["c_errors"] = c_err + logging_output["c_total"] = c_len + + return loss, sample_size, logging_output + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + """Aggregate logging outputs from data parallel training.""" + + loss_sum = utils.item(sum(log.get("loss", 0) for log in logging_outputs)) + ntokens = utils.item(sum(log.get("ntokens", 0) for log in logging_outputs)) + nsentences = utils.item( + sum(log.get("nsentences", 0) for log in logging_outputs) + ) + sample_size = utils.item( + sum(log.get("sample_size", 0) for log in logging_outputs) + ) + + metrics.log_scalar( + "loss", loss_sum / sample_size / math.log(2), sample_size, round=3 + ) + metrics.log_scalar("ntokens", ntokens) + metrics.log_scalar("nsentences", nsentences) + if sample_size != ntokens: + metrics.log_scalar( + "nll_loss", loss_sum / ntokens / math.log(2), ntokens, round=3 + ) + + c_errors = sum(log.get("c_errors", 0) for log in logging_outputs) + metrics.log_scalar("_c_errors", c_errors) + c_total = sum(log.get("c_total", 0) for log in logging_outputs) + metrics.log_scalar("_c_total", c_total) + w_errors = sum(log.get("w_errors", 0) for log in logging_outputs) + metrics.log_scalar("_w_errors", w_errors) + wv_errors = sum(log.get("wv_errors", 0) for log in logging_outputs) + metrics.log_scalar("_wv_errors", wv_errors) + w_total = sum(log.get("w_total", 0) for log in logging_outputs) + metrics.log_scalar("_w_total", w_total) + + if c_total > 0: + metrics.log_derived( + "uer", + lambda meters: safe_round( + meters["_c_errors"].sum * 100.0 / meters["_c_total"].sum, 3 + ) + if meters["_c_total"].sum > 0 + else float("nan"), + ) + if w_total > 0: + metrics.log_derived( + "wer", + lambda meters: safe_round( + meters["_w_errors"].sum * 100.0 / meters["_w_total"].sum, 3 + ) + if meters["_w_total"].sum > 0 + else float("nan"), + ) + metrics.log_derived( + "raw_wer", + lambda meters: safe_round( + meters["_wv_errors"].sum * 100.0 / meters["_w_total"].sum, 3 + ) + if meters["_w_total"].sum > 0 + else float("nan"), + ) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return True diff --git a/fairseq/criterions/fairseq_criterion.py b/fairseq/criterions/fairseq_criterion.py new file mode 100644 index 0000000000000000000000000000000000000000..ff4beb02503ea48a6c09596630aad4c710be94b6 --- /dev/null +++ b/fairseq/criterions/fairseq_criterion.py @@ -0,0 +1,120 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import inspect +from typing import Any, Dict, List + +from fairseq import metrics, utils +from fairseq.dataclass import FairseqDataclass +from fairseq.dataclass.utils import gen_parser_from_dataclass +from torch.nn.modules.loss import _Loss + + +class FairseqCriterion(_Loss): + def __init__(self, task): + super().__init__() + self.task = task + if hasattr(task, "target_dictionary"): + tgt_dict = task.target_dictionary + self.padding_idx = tgt_dict.pad() if tgt_dict is not None else -100 + + @classmethod + def add_args(cls, parser): + """Add criterion-specific arguments to the parser.""" + dc = getattr(cls, "__dataclass", None) + if dc is not None: + gen_parser_from_dataclass(parser, dc()) + + @classmethod + def build_criterion(cls, cfg: FairseqDataclass, task): + """Construct a criterion from command-line args.""" + # arguments in the __init__. + init_args = {} + for p in inspect.signature(cls).parameters.values(): + if ( + p.kind == p.POSITIONAL_ONLY + or p.kind == p.VAR_POSITIONAL + or p.kind == p.VAR_KEYWORD + ): + # we haven't implemented inference for these argument types, + # but PRs welcome :) + raise NotImplementedError("{} not supported".format(p.kind)) + + assert p.kind in {p.POSITIONAL_OR_KEYWORD, p.KEYWORD_ONLY} + + if p.name == "task": + init_args["task"] = task + elif p.name == "cfg": + init_args["cfg"] = cfg + elif hasattr(cfg, p.name): + init_args[p.name] = getattr(cfg, p.name) + elif p.default != p.empty: + pass # we'll use the default value + else: + raise NotImplementedError( + "Unable to infer Criterion arguments, please implement " + "{}.build_criterion".format(cls.__name__) + ) + return cls(**init_args) + + def forward(self, model, sample, reduce=True): + """Compute the loss for the given sample. + + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + raise NotImplementedError + + @staticmethod + def aggregate_logging_outputs( + logging_outputs: List[Dict[str, Any]] + ) -> Dict[str, Any]: + """Aggregate logging outputs from data parallel training.""" + utils.deprecation_warning( + "The aggregate_logging_outputs API is deprecated. " + "Please use the reduce_metrics API instead." + ) + raise NotImplementedError + + @classmethod + def reduce_metrics(cls, logging_outputs: List[Dict[str, Any]]) -> None: + """Aggregate logging outputs from data parallel training.""" + utils.deprecation_warning( + "Criterions should implement the reduce_metrics API. " + "Falling back to deprecated aggregate_logging_outputs API." + ) + agg_logging_outputs = cls.aggregate_logging_outputs(logging_outputs) + for k, v in agg_logging_outputs.items(): + if k in {"nsentences", "ntokens", "sample_size"}: + continue + metrics.log_scalar(k, v) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return False + + +class LegacyFairseqCriterion(FairseqCriterion): + def __init__(self, args, task): + super().__init__(task=task) + self.args = args + + utils.deprecation_warning( + "Criterions should take explicit arguments instead of an " + "argparse.Namespace object, please update your criterion by " + "extending FairseqCriterion instead of LegacyFairseqCriterion." + ) + + @classmethod + def build_criterion(cls, args, task): + """Construct a criterion from command-line args.""" + return cls(args, task) diff --git a/fairseq/criterions/hubert_criterion.py b/fairseq/criterions/hubert_criterion.py new file mode 100644 index 0000000000000000000000000000000000000000..68cb24e6f142c46e108c53479fd4027a741f5f92 --- /dev/null +++ b/fairseq/criterions/hubert_criterion.py @@ -0,0 +1,177 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +import re +from dataclasses import dataclass, field +from typing import List, Optional + +import torch +import torch.nn.functional as F +from fairseq import metrics, utils +from fairseq.criterions import FairseqCriterion, register_criterion +from fairseq.dataclass import FairseqDataclass + + +@dataclass +class HubertCriterionConfig(FairseqDataclass): + pred_masked_weight: float = field( + default=1.0, + metadata={"help": "weight for predictive loss for masked frames"}, + ) + pred_nomask_weight: float = field( + default=0.0, + metadata={"help": "weight for predictive loss for unmasked frames"}, + ) + loss_weights: Optional[List[float]] = field( + default=None, + metadata={"help": "weights for additional loss terms (not first one)"}, + ) + log_keys: List[str] = field( + default_factory=lambda: [], + metadata={"help": "output keys to log"}, + ) + + +@register_criterion("hubert", dataclass=HubertCriterionConfig) +class HubertCriterion(FairseqCriterion): + def __init__(self, task, pred_masked_weight, pred_nomask_weight, loss_weights=None, log_keys=None): + super().__init__(task) + self.pred_masked_weight = pred_masked_weight + self.pred_nomask_weight = pred_nomask_weight + self.loss_weights = loss_weights + self.log_keys = [] if log_keys is None else log_keys + + def forward(self, model, sample, reduce=True, log_pred=False): + """Compute the loss for the given sample. + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + net_output = model(target_list=sample["target_list"], **sample["net_input"]) + loss = 0. + sample_size = 0 + logging_output = {} + reduction = "sum" if reduce else "none" + + loss_m_list = [] + logp_m_list = model.get_logits(net_output, True) + targ_m_list = model.get_targets(net_output, True) + assert self.pred_masked_weight == 0 or len(logp_m_list) > 0 + for i, (logp_m, targ_m) in enumerate(zip(logp_m_list, targ_m_list)): + loss_m = F.cross_entropy(logp_m, targ_m, reduction=reduction) + loss_m_list.append(loss_m) + logging_output[f"loss_m_{i}"] = loss_m.detach().item() + if self.pred_masked_weight > 0: + loss += self.pred_masked_weight * sum(loss_m_list) + sample_size += targ_m_list[0].numel() + + loss_u_list = [] + logp_u_list = model.get_logits(net_output, False) + targ_u_list = model.get_targets(net_output, False) + assert self.pred_nomask_weight == 0 or len(logp_u_list) > 0 + for i, (logp_u, targ_u) in enumerate(zip(logp_u_list, targ_u_list)): + loss_u = F.cross_entropy(logp_u, targ_u, reduction=reduction) + loss_u_list.append(loss_u) + logging_output[f"loss_u_{i}"] = loss_u.detach().item() + if self.pred_nomask_weight > 0: + loss += self.pred_nomask_weight * sum(loss_u_list) + sample_size += targ_u_list[0].numel() + + if self.loss_weights is not None: + assert hasattr(model, "get_extra_losses") + extra_losses, names = model.get_extra_losses(net_output) + if torch.is_tensor(extra_losses): + extra_losses = [extra_losses] + names = [names] + if len(self.loss_weights) == 1 and len(extra_losses) != 1: + self.loss_weights = [self.loss_weights[0]] * len(extra_losses) + assert len(extra_losses) == len(self.loss_weights), f"{len(extra_losses)}, {len(self.loss_weights)}" + for p, n, coef in zip(extra_losses, names, self.loss_weights): + if coef != 0 and p is not None: + p = coef * p.float() * sample_size + loss += p + logging_output[f"loss_{n}"] = p.item() + + logging_output = { + "loss": loss.item() if reduce else loss, + "ntokens": sample_size, + "nsentences": sample["id"].numel(), + "sample_size": sample_size, + **logging_output, + } + + for lk in self.log_keys: + if lk in net_output: + logging_output[lk] = float((net_output[lk])) + + def compute_correct(logits): + if logits.numel() == 0: + return 0, 0 + else: + assert logits.dim() > 1, logits.shape + max = logits.argmax(-1) == 0 + min = logits.argmin(-1) == 0 + both = max & min + corr = max.long().sum().item() - both.long().sum().item() + count = max.numel() + return corr, count + + with torch.no_grad(): + for i, logp_m in enumerate(logp_m_list): + corr_m, count_m = compute_correct(logp_m) + logging_output[f"correct_m_{i}"] = corr_m + logging_output[f"count_m_{i}"] = count_m + + for i, logp_u in enumerate(logp_u_list): + corr_u, count_u = compute_correct(logp_u) + logging_output[f"correct_u_{i}"] = corr_u + logging_output[f"count_u_{i}"] = count_u + + return loss, sample_size, logging_output + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + """Aggregate logging outputs from data parallel training (copied from normal cross entropy).""" + loss_sum = sum(log.get("loss", 0) for log in logging_outputs) + ntokens = sum(log.get("ntokens", 0) for log in logging_outputs) + sample_size = sum(log.get("sample_size", 0) for log in logging_outputs) + + metrics.log_scalar("loss", loss_sum / sample_size / math.log(2), sample_size, round=3) + if sample_size != ntokens: + metrics.log_scalar("nll_loss", loss_sum / ntokens / math.log(2), ntokens, round=3) + metrics.log_derived("ppl", lambda meters: utils.get_perplexity(meters["nll_loss"].avg)) + else: + metrics.log_derived("ppl", lambda meters: utils.get_perplexity(meters["loss"].avg)) + + counts = {} + for lk in logging_outputs[0].keys(): + if lk.startswith("count_"): + val = sum(log[lk] for log in logging_outputs) + metrics.log_scalar(lk, val) + counts[lk] = val + + for lk in logging_outputs[0].keys(): + if lk.startswith("loss_"): + val = sum(log[lk] for log in logging_outputs) + metrics.log_scalar(lk, val / sample_size / math.log(2), round=3) + elif lk.startswith("correct_"): + val = sum(log[lk] for log in logging_outputs) + metrics.log_scalar(lk, val / counts[re.sub("correct", "count", lk)]) + + @staticmethod + def aggregate_logging_outputs(logging_outputs): + """Aggregate logging outputs from data parallel training.""" + raise NotImplementedError() + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return False diff --git a/fairseq/criterions/label_smoothed_cross_entropy.py b/fairseq/criterions/label_smoothed_cross_entropy.py new file mode 100644 index 0000000000000000000000000000000000000000..56d63e3e1b5a036e0adf32480e2b66f371738013 --- /dev/null +++ b/fairseq/criterions/label_smoothed_cross_entropy.py @@ -0,0 +1,170 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +from dataclasses import dataclass, field + +import torch +from fairseq import metrics, utils +from fairseq.criterions import FairseqCriterion, register_criterion +from fairseq.dataclass import FairseqDataclass +from omegaconf import II + + +@dataclass +class LabelSmoothedCrossEntropyCriterionConfig(FairseqDataclass): + label_smoothing: float = field( + default=0.0, + metadata={"help": "epsilon for label smoothing, 0 means no label smoothing"}, + ) + report_accuracy: bool = field( + default=False, + metadata={"help": "report accuracy metric"}, + ) + ignore_prefix_size: int = field( + default=0, + metadata={"help": "Ignore first N tokens"}, + ) + sentence_avg: bool = II("optimization.sentence_avg") + + +def label_smoothed_nll_loss(lprobs, target, epsilon, ignore_index=None, reduce=True): + if target.dim() == lprobs.dim() - 1: + target = target.unsqueeze(-1) + nll_loss = -lprobs.gather(dim=-1, index=target) + smooth_loss = -lprobs.sum(dim=-1, keepdim=True) + if ignore_index is not None: + pad_mask = target.eq(ignore_index) + nll_loss.masked_fill_(pad_mask, 0.0) + smooth_loss.masked_fill_(pad_mask, 0.0) + else: + nll_loss = nll_loss.squeeze(-1) + smooth_loss = smooth_loss.squeeze(-1) + if reduce: + nll_loss = nll_loss.sum() + smooth_loss = smooth_loss.sum() + eps_i = epsilon / (lprobs.size(-1) - 1) + loss = (1.0 - epsilon - eps_i) * nll_loss + eps_i * smooth_loss + return loss, nll_loss + + +@register_criterion( + "label_smoothed_cross_entropy", dataclass=LabelSmoothedCrossEntropyCriterionConfig +) +class LabelSmoothedCrossEntropyCriterion(FairseqCriterion): + def __init__( + self, + task, + sentence_avg, + label_smoothing, + ignore_prefix_size=0, + report_accuracy=False, + ): + super().__init__(task) + self.sentence_avg = sentence_avg + self.eps = label_smoothing + self.ignore_prefix_size = ignore_prefix_size + self.report_accuracy = report_accuracy + + def forward(self, model, sample, reduce=True): + """Compute the loss for the given sample. + + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + net_output = model(**sample["net_input"]) + loss, nll_loss = self.compute_loss(model, net_output, sample, reduce=reduce) + sample_size = ( + sample["target"].size(0) if self.sentence_avg else sample["ntokens"] + ) + logging_output = { + "loss": loss.data, + "nll_loss": nll_loss.data, + "ntokens": sample["ntokens"], + "nsentences": sample["target"].size(0), + "sample_size": sample_size, + } + if self.report_accuracy: + n_correct, total = self.compute_accuracy(model, net_output, sample) + logging_output["n_correct"] = utils.item(n_correct.data) + logging_output["total"] = utils.item(total.data) + return loss, sample_size, logging_output + + def get_lprobs_and_target(self, model, net_output, sample): + lprobs = model.get_normalized_probs(net_output, log_probs=True) + target = model.get_targets(sample, net_output) + if self.ignore_prefix_size > 0: + if getattr(lprobs, "batch_first", False): + lprobs = lprobs[:, self.ignore_prefix_size :, :].contiguous() + target = target[:, self.ignore_prefix_size :].contiguous() + else: + lprobs = lprobs[self.ignore_prefix_size :, :, :].contiguous() + target = target[self.ignore_prefix_size :, :].contiguous() + return lprobs.view(-1, lprobs.size(-1)), target.view(-1) + + def compute_loss(self, model, net_output, sample, reduce=True): + lprobs, target = self.get_lprobs_and_target(model, net_output, sample) + loss, nll_loss = label_smoothed_nll_loss( + lprobs, + target, + self.eps, + ignore_index=self.padding_idx, + reduce=reduce, + ) + return loss, nll_loss + + def compute_accuracy(self, model, net_output, sample): + lprobs, target = self.get_lprobs_and_target(model, net_output, sample) + mask = target.ne(self.padding_idx) + n_correct = torch.sum( + lprobs.argmax(1).masked_select(mask).eq(target.masked_select(mask)) + ) + total = torch.sum(mask) + return n_correct, total + + @classmethod + def reduce_metrics(cls, logging_outputs) -> None: + """Aggregate logging outputs from data parallel training.""" + loss_sum = sum(log.get("loss", 0) for log in logging_outputs) + nll_loss_sum = sum(log.get("nll_loss", 0) for log in logging_outputs) + ntokens = sum(log.get("ntokens", 0) for log in logging_outputs) + sample_size = sum(log.get("sample_size", 0) for log in logging_outputs) + + metrics.log_scalar( + "loss", loss_sum / sample_size / math.log(2), sample_size, round=3 + ) + metrics.log_scalar( + "nll_loss", nll_loss_sum / ntokens / math.log(2), ntokens, round=3 + ) + metrics.log_derived( + "ppl", lambda meters: utils.get_perplexity(meters["nll_loss"].avg) + ) + + total = utils.item(sum(log.get("total", 0) for log in logging_outputs)) + if total > 0: + metrics.log_scalar("total", total) + n_correct = utils.item( + sum(log.get("n_correct", 0) for log in logging_outputs) + ) + metrics.log_scalar("n_correct", n_correct) + metrics.log_derived( + "accuracy", + lambda meters: round( + meters["n_correct"].sum * 100.0 / meters["total"].sum, 3 + ) + if meters["total"].sum > 0 + else float("nan"), + ) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return True diff --git a/fairseq/criterions/label_smoothed_cross_entropy_latency_augmented.py b/fairseq/criterions/label_smoothed_cross_entropy_latency_augmented.py new file mode 100644 index 0000000000000000000000000000000000000000..051785238fdc4d18230de49ddd735f154ed5a3e7 --- /dev/null +++ b/fairseq/criterions/label_smoothed_cross_entropy_latency_augmented.py @@ -0,0 +1,108 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq.criterions import register_criterion +from fairseq.criterions.label_smoothed_cross_entropy import ( + LabelSmoothedCrossEntropyCriterion, +) + + +@register_criterion("latency_augmented_label_smoothed_cross_entropy") +class LatencyAugmentedLabelSmoothedCrossEntropyCriterion( + LabelSmoothedCrossEntropyCriterion +): + def __init__( + self, + task, + sentence_avg, + label_smoothing, + ignore_prefix_size, + report_accuracy, + latency_weight_avg, + latency_weight_avg_type, + latency_weight_var, + latency_weight_var_type, + mass_preservation, + average_method, + ): + super().__init__( + task, sentence_avg, label_smoothing, ignore_prefix_size, report_accuracy + ) + from examples.simultaneous_translation.utils.latency import LatencyTraining + self.eps = label_smoothing + self.latency_weight_avg = latency_weight_avg + self.latency_weight_avg_type = latency_weight_avg_type + self.latency_weight_var = latency_weight_var + self.latency_weight_var_type = latency_weight_var_type + self.mass_preservation = mass_preservation + self.average_method = average_method + self.latency_train = LatencyTraining( + self.latency_weight_avg, + self.latency_weight_var, + self.latency_weight_avg_type, + self.latency_weight_var_type, + self.mass_preservation, + self.average_method, + ) + + @staticmethod + def add_args(parser): + super( + LatencyAugmentedLabelSmoothedCrossEntropyCriterion, + LatencyAugmentedLabelSmoothedCrossEntropyCriterion, + ).add_args(parser) + # fmt: off + + """Add criterion-specific arguments to the parser.""" + parser.add_argument( + "--label-smoothing", + default=0.0, + type=float, + metavar="D", + help="epsilon for label smoothing, 0 means no label smoothing", + ) + parser.add_argument( + "--ignore_prefix_size", + default=0, + type=int, + help="ignore first N tokens", + ) + parser.add_argument( + "--report-accuracy", + default=False, + type=bool, + help="report accuracy metric", + ) + parser.add_argument("--latency-weight-avg", default=0., type=float, metavar='D', + help="Average loss weight") + parser.add_argument("--latency-weight-var", default=0., type=float, metavar='D', + help="Variance loss weight") + parser.add_argument("--latency-weight-avg-type", default="differentiable_average_lagging", + help="Statistics for Average loss type") + parser.add_argument("--latency-weight-var-type", default="variance_delay", + help="Statistics for variance loss type") + parser.add_argument("--average-method", default="weighted_average", + help="Average loss type") + # fmt: on + + def compute_loss(self, model, net_output, sample, reduce=True): + # Compute cross entropy loss first + loss, nll_loss = super().compute_loss(model, net_output, sample, reduce) + + # Obtain the expected alignment + attn_list = [item["alpha"] for item in net_output[-1]["attn_list"]] + + target_padding_mask = model.get_targets(sample, net_output).eq(self.padding_idx) + + source_padding_mask = net_output[-1].get("encoder_padding_mask", None) + + # Get latency loss + latency_loss = self.latency_train.loss( + attn_list, source_padding_mask, target_padding_mask + ) + + loss += latency_loss + + return loss, nll_loss diff --git a/fairseq/criterions/label_smoothed_cross_entropy_with_alignment.py b/fairseq/criterions/label_smoothed_cross_entropy_with_alignment.py new file mode 100644 index 0000000000000000000000000000000000000000..73cfa05310e51d9a5f349cc30b8406002d25861b --- /dev/null +++ b/fairseq/criterions/label_smoothed_cross_entropy_with_alignment.py @@ -0,0 +1,125 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +from fairseq import metrics, utils +from fairseq.criterions import register_criterion + +from .label_smoothed_cross_entropy import LabelSmoothedCrossEntropyCriterion + + +@register_criterion("label_smoothed_cross_entropy_with_alignment") +class LabelSmoothedCrossEntropyCriterionWithAlignment( + LabelSmoothedCrossEntropyCriterion +): + def __init__(self, task, sentence_avg, label_smoothing, alignment_lambda): + super().__init__(task, sentence_avg, label_smoothing) + self.alignment_lambda = alignment_lambda + + @staticmethod + def add_args(parser): + """Add criterion-specific arguments to the parser.""" + LabelSmoothedCrossEntropyCriterion.add_args(parser) + parser.add_argument( + "--alignment-lambda", + default=0.05, + type=float, + metavar="D", + help="weight for the alignment loss", + ) + + def forward(self, model, sample, reduce=True): + """Compute the loss for the given sample. + + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + net_output = model(**sample["net_input"]) + loss, nll_loss = self.compute_loss(model, net_output, sample, reduce=reduce) + sample_size = ( + sample["target"].size(0) if self.sentence_avg else sample["ntokens"] + ) + logging_output = { + "loss": utils.item(loss.data) if reduce else loss.data, + "nll_loss": utils.item(nll_loss.data) if reduce else nll_loss.data, + "ntokens": sample["ntokens"], + "nsentences": sample["target"].size(0), + "sample_size": sample_size, + } + + alignment_loss = None + + # Compute alignment loss only for training set and non dummy batches. + if "alignments" in sample and sample["alignments"] is not None: + alignment_loss = self.compute_alignment_loss(sample, net_output) + + if alignment_loss is not None: + logging_output["alignment_loss"] = utils.item(alignment_loss.data) + loss += self.alignment_lambda * alignment_loss + + return loss, sample_size, logging_output + + def compute_alignment_loss(self, sample, net_output): + attn_prob = net_output[1]["attn"][0] + bsz, tgt_sz, src_sz = attn_prob.shape + attn = attn_prob.view(bsz * tgt_sz, src_sz) + + align = sample["alignments"] + align_weights = sample["align_weights"].float() + + if len(align) > 0: + # Alignment loss computation. align (shape [:, 2]) contains the src-tgt index pairs corresponding to + # the alignments. align_weights (shape [:]) contains the 1 / frequency of a tgt index for normalizing. + loss = -( + (attn[align[:, 1][:, None], align[:, 0][:, None]]).log() + * align_weights[:, None] + ).sum() + else: + return None + + return loss + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + """Aggregate logging outputs from data parallel training.""" + loss_sum = utils.item(sum(log.get("loss", 0) for log in logging_outputs)) + nll_loss_sum = utils.item( + sum(log.get("nll_loss", 0) for log in logging_outputs) + ) + alignment_loss_sum = utils.item( + sum(log.get("alignment_loss", 0) for log in logging_outputs) + ) + ntokens = utils.item(sum(log.get("ntokens", 0) for log in logging_outputs)) + sample_size = utils.item( + sum(log.get("sample_size", 0) for log in logging_outputs) + ) + + metrics.log_scalar( + "loss", loss_sum / sample_size / math.log(2), sample_size, round=3 + ) + metrics.log_scalar( + "nll_loss", nll_loss_sum / ntokens / math.log(2), ntokens, round=3 + ) + metrics.log_scalar( + "alignment_loss", + alignment_loss_sum / sample_size / math.log(2), + sample_size, + round=3, + ) + metrics.log_derived( + "ppl", lambda meters: utils.get_perplexity(meters["nll_loss"].avg) + ) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return True diff --git a/fairseq/criterions/legacy_masked_lm.py b/fairseq/criterions/legacy_masked_lm.py new file mode 100644 index 0000000000000000000000000000000000000000..c70608c5a143b7b4fbd8c58dfcf9f873639d379c --- /dev/null +++ b/fairseq/criterions/legacy_masked_lm.py @@ -0,0 +1,177 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +import torch +import torch.nn.functional as F +from fairseq import metrics, utils +from fairseq.criterions import FairseqCriterion, register_criterion + + +def compute_cross_entropy_loss(logits, targets, ignore_index=-100): + """ + Function to compute the cross entropy loss. The default value of + ignore_index is the same as the default value for F.cross_entropy in + pytorch. + """ + assert logits.size(0) == targets.size( + -1 + ), "Logits and Targets tensor shapes don't match up" + + loss = F.nll_loss( + F.log_softmax(logits, -1, dtype=torch.float32), + targets, + reduction="sum", + ignore_index=ignore_index, + ) + return loss + + +@register_criterion("legacy_masked_lm_loss") +class LegacyMaskedLmLoss(FairseqCriterion): + """ + Implementation for the loss used in masked language model (MLM) training. + This optionally also computes the next sentence prediction (NSP) loss and + adds it to the overall loss based on the specified args. There are three + cases to consider: + 1) Generic MLM training without NSP loss. In this case sentence_targets + and sentence_logits are both None. + 2) BERT training without NSP loss. In this case sentence_targets is + not None but sentence_logits is None and we should not be computing + a sentence level loss. + 3) BERT training with NSP loss. In this case both sentence_targets and + sentence_logits are not None and we should be computing a sentence + level loss. The weight of the sentence level loss is specified as + an argument. + """ + + def __init__(self, task, masked_lm_only, nsp_loss_weight): + super().__init__(task) + self.masked_lm_only = masked_lm_only + self.nsp_loss_weight = nsp_loss_weight + + @staticmethod + def add_args(parser): + """Args for MaskedLM Loss""" + # Default for masked_lm_only is False so as to not break BERT training + parser.add_argument( + "--masked-lm-only", + default=False, + action="store_true", + help="compute MLM loss only", + ) + parser.add_argument( + "--nsp-loss-weight", + default=1.0, + type=float, + help="weight for next sentence prediction" " loss (default 1)", + ) + + def forward(self, model, sample, reduce=True): + """Compute the loss for the given sample. + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + lm_logits, output_metadata = model(**sample["net_input"]) + + # reshape lm_logits from (N,T,C) to (N*T,C) + lm_logits = lm_logits.view(-1, lm_logits.size(-1)) + lm_targets = sample["lm_target"].view(-1) + lm_loss = compute_cross_entropy_loss(lm_logits, lm_targets, self.padding_idx) + + # compute the number of tokens for which loss is computed. This is used + # to normalize the loss + ntokens = utils.strip_pad(lm_targets, self.padding_idx).numel() + loss = lm_loss / ntokens + nsentences = sample["nsentences"] + # nsentences = 0 + + # Compute sentence loss if masked_lm_only is False + sentence_loss = None + if not self.masked_lm_only: + sentence_logits = output_metadata["sentence_logits"] + sentence_targets = sample["sentence_target"].view(-1) + # This needs to be recomputed due to some differences between + # TokenBlock and BlockPair dataset. This can be resolved with a + # refactor of BERTModel which we will do in the future. + # TODO: Remove this after refactor of BERTModel + nsentences = sentence_targets.size(0) + + # Check for logits being none which can happen when remove_heads + # is set to true in the BERT model. Ideally we should set + # masked_lm_only to true in this case, but that requires some + # refactor in the BERT model. + if sentence_logits is not None: + sentence_loss = compute_cross_entropy_loss( + sentence_logits, sentence_targets + ) + + loss += self.nsp_loss_weight * (sentence_loss / nsentences) + + # NOTE: as we are summing up per token mlm loss and per sentence nsp loss + # we don't need to use sample_size as denominator for the gradient + # here sample_size is just used for logging + sample_size = 1 + logging_output = { + "loss": utils.item(loss.data) if reduce else loss.data, + "lm_loss": utils.item(lm_loss.data) if reduce else lm_loss.data, + # sentence loss is not always computed + "sentence_loss": ( + (utils.item(sentence_loss.data) if reduce else sentence_loss.data) + if sentence_loss is not None + else 0.0 + ), + "ntokens": ntokens, + "nsentences": nsentences, + "sample_size": sample_size, + } + return loss, sample_size, logging_output + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + """Aggregate logging outputs from data parallel training.""" + lm_loss_sum = sum(log.get("lm_loss", 0) for log in logging_outputs) + sentence_loss_sum = sum(log.get("sentence_loss", 0) for log in logging_outputs) + ntokens = sum(log.get("ntokens", 0) for log in logging_outputs) + nsentences = sum(log.get("nsentences", 0) for log in logging_outputs) + sample_size = sum(log.get("sample_size", 0) for log in logging_outputs) + agg_loss = sum(log.get("loss", 0) for log in logging_outputs) + + metrics.log_scalar( + "loss", + agg_loss / sample_size / math.log(2) if sample_size > 0 else 0.0, + sample_size, + round=3, + ) + metrics.log_scalar( + "lm_loss", + lm_loss_sum / ntokens / math.log(2) if ntokens > 0 else 0.0, + ntokens, + round=3, + ) + metrics.log_scalar( + "sentence_loss", + sentence_loss_sum / nsentences / math.log(2) if nsentences > 0 else 0.0, + nsentences, + round=3, + ) + metrics.log_scalar( + "nll_loss", + lm_loss_sum / ntokens / math.log(2) if ntokens > 0 else 0.0, + ntokens, + round=3, + ) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return True diff --git a/fairseq/criterions/masked_lm.py b/fairseq/criterions/masked_lm.py new file mode 100644 index 0000000000000000000000000000000000000000..b04cfbff6dcbfacb91156bb10a7c8cdbb9e76d37 --- /dev/null +++ b/fairseq/criterions/masked_lm.py @@ -0,0 +1,91 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +import torch +import torch.nn.functional as F +from fairseq import metrics, modules, utils +from fairseq.criterions import FairseqCriterion, register_criterion + + +@register_criterion("masked_lm") +class MaskedLmLoss(FairseqCriterion): + """ + Implementation for the loss used in masked language model (MLM) training. + """ + + def __init__(self, task, tpu=False): + super().__init__(task) + self.tpu = tpu + + def forward(self, model, sample, reduce=True): + """Compute the loss for the given sample. + + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + masked_tokens = sample["target"].ne(self.padding_idx) + sample_size = masked_tokens.int().sum() + + # Rare: when all tokens are masked, project all tokens. + # We use torch.where to avoid device-to-host transfers, + # except on CPU where torch.where is not well supported + # (see github.com/pytorch/pytorch/issues/26247). + if self.tpu: + masked_tokens = None # always project all tokens on TPU + elif masked_tokens.device == torch.device("cpu"): + if not masked_tokens.any(): + masked_tokens = None + else: + masked_tokens = torch.where( + masked_tokens.any(), + masked_tokens, + masked_tokens.new([True]), + ) + + logits = model(**sample["net_input"], masked_tokens=masked_tokens)[0] + targets = model.get_targets(sample, [logits]) + if masked_tokens is not None: + targets = targets[masked_tokens] + + loss = modules.cross_entropy( + logits.view(-1, logits.size(-1)), + targets.view(-1), + reduction="sum", + ignore_index=self.padding_idx, + ) + + logging_output = { + "loss": loss if self.tpu else loss.data, + "ntokens": sample["ntokens"], + "nsentences": sample["nsentences"], + "sample_size": sample_size, + } + return loss, sample_size, logging_output + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + """Aggregate logging outputs from data parallel training.""" + loss_sum = sum(log.get("loss", 0) for log in logging_outputs) + sample_size = sum(log.get("sample_size", 0) for log in logging_outputs) + + metrics.log_scalar( + "loss", loss_sum / sample_size / math.log(2), sample_size, round=3 + ) + metrics.log_derived( + "ppl", lambda meters: utils.get_perplexity(meters["loss"].avg) + ) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return True diff --git a/fairseq/criterions/model_criterion.py b/fairseq/criterions/model_criterion.py new file mode 100644 index 0000000000000000000000000000000000000000..30350f13b1c00498de6784579250d6b342ced7dd --- /dev/null +++ b/fairseq/criterions/model_criterion.py @@ -0,0 +1,138 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +from dataclasses import dataclass, field +from typing import Dict, List + +from fairseq import metrics, utils +from fairseq.criterions import FairseqCriterion, register_criterion +from fairseq.dataclass import FairseqDataclass + + +logger = logging.getLogger(__name__) + + +@dataclass +class ModelCriterionConfig(FairseqDataclass): + loss_weights: Dict[str, float] = field( + default_factory=dict, + metadata={"help": "weights for the loss terms"}, + ) + log_keys: List[str] = field( + default_factory=list, + metadata={"help": "additional output keys to log"}, + ) + + +@register_criterion("model", dataclass=ModelCriterionConfig) +class ModelCriterion(FairseqCriterion): + """ + This criterion relies on the model to supply losses. + The losses should be a dictionary of name -> scalar returned by + the model either by including it in the net_output dict or by + implementing a get_losses(net_output, sample) method. The final loss is + a scaled sum of all losses according to weights in loss_weights. + If no weights are provided, then all losses are scaled by 1.0. + + The losses will be automatically logged. Additional keys from + net_output dict can be logged via the log_keys parameter. + """ + + def __init__(self, task, loss_weights=None, log_keys=None): + super().__init__(task) + self.loss_weights = loss_weights + self.log_keys = log_keys + + def forward(self, model, sample, reduce=True): + net_output = model(**sample["net_input"]) + + sample_size = net_output["sample_size"] + scaled_losses = {} + + if hasattr(model, "get_losses"): + losses = model.get_losses(net_output, sample) + elif isinstance(net_output, dict) and "losses" in net_output: + losses = net_output["losses"] + else: + raise Exception("Could not retrieve losses") + + for lk, p in losses.items(): + try: + coef = 1.0 if len(self.loss_weights) == 0 else self.loss_weights[lk] + except KeyError: + logger.error( + f"weight for loss {lk} is not in loss_weights ({self.loss_weights})" + ) + raise + if coef != 0 and p is not None: + scaled_losses[lk] = coef * p.float() + + loss = sum(scaled_losses.values()) + if reduce and loss.numel() > 1: + loss = loss.sum() + + logging_output = { + "loss": loss.data, + "ntokens": sample_size, + "nsentences": sample["id"].numel(), + "sample_size": sample_size, + "_world_size": 1, + } + + for lk in self.log_keys: + if lk in net_output and net_output[lk] is not None: + logging_output[lk] = float(net_output[lk]) + + if len(scaled_losses) > 1: + for lk, l in scaled_losses.items(): + logging_output[f"loss_{lk}"] = l.item() + + return loss, sample_size, logging_output + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + """Aggregate logging outputs from data parallel training.""" + loss_sum = utils.item(sum(log.get("loss", 0) for log in logging_outputs)) + ntokens = utils.item(sum(log.get("ntokens", 0) for log in logging_outputs)) + nsentences = utils.item( + sum(log.get("nsentences", 0) for log in logging_outputs) + ) + sample_size = utils.item( + sum(log.get("sample_size", 0) for log in logging_outputs) + ) + + metrics.log_scalar("loss", loss_sum / sample_size, sample_size, round=3) + metrics.log_scalar("ntokens", ntokens) + metrics.log_scalar("nsentences", nsentences) + + builtin_keys = { + "loss", + "ntokens", + "nsentences", + "sample_size", + "_world_size", + } + + world_size = utils.item( + sum(log.get("_world_size", 0) for log in logging_outputs) + ) + + for k in logging_outputs[0]: + if k not in builtin_keys: + val = sum(log.get(k, 0) for log in logging_outputs) + if k.startswith("loss_"): + metrics.log_scalar(k, val / sample_size, sample_size, round=3) + else: + metrics.log_scalar(k, val / world_size, round=3) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return True diff --git a/fairseq/criterions/nat_loss.py b/fairseq/criterions/nat_loss.py new file mode 100644 index 0000000000000000000000000000000000000000..cdc7da861d7d5d5af183a78fdde51f49eb0cf5e7 --- /dev/null +++ b/fairseq/criterions/nat_loss.py @@ -0,0 +1,180 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +import torch +import torch.nn.functional as F +from fairseq import metrics, utils +from fairseq.criterions import FairseqCriterion, register_criterion +from torch import Tensor + + +@register_criterion("nat_loss") +class LabelSmoothedDualImitationCriterion(FairseqCriterion): + def __init__(self, task, label_smoothing): + super().__init__(task) + self.label_smoothing = label_smoothing + + @staticmethod + def add_args(parser): + """Add criterion-specific arguments to the parser.""" + parser.add_argument( + "--label-smoothing", + default=0.0, + type=float, + metavar="D", + help="epsilon for label smoothing, 0 means no label smoothing", + ) + + def _compute_loss( + self, outputs, targets, masks=None, label_smoothing=0.0, name="loss", factor=1.0 + ): + """ + outputs: batch x len x d_model + targets: batch x len + masks: batch x len + + policy_logprob: if there is some policy + depends on the likelihood score as rewards. + """ + + def mean_ds(x: Tensor, dim=None) -> Tensor: + return ( + x.float().mean().type_as(x) + if dim is None + else x.float().mean(dim).type_as(x) + ) + + if masks is not None: + outputs, targets = outputs[masks], targets[masks] + + if masks is not None and not masks.any(): + nll_loss = torch.tensor(0) + loss = nll_loss + else: + logits = F.log_softmax(outputs, dim=-1) + if targets.dim() == 1: + losses = F.nll_loss(logits, targets.to(logits.device), reduction="none") + + else: # soft-labels + losses = F.kl_div(logits, targets.to(logits.device), reduction="none") + losses = losses.sum(-1) + + nll_loss = mean_ds(losses) + if label_smoothing > 0: + loss = ( + nll_loss * (1 - label_smoothing) - mean_ds(logits) * label_smoothing + ) + else: + loss = nll_loss + + loss = loss * factor + return {"name": name, "loss": loss, "nll_loss": nll_loss, "factor": factor} + + def _custom_loss(self, loss, name="loss", factor=1.0): + return {"name": name, "loss": loss, "factor": factor} + + def forward(self, model, sample, reduce=True): + """Compute the loss for the given sample. + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + nsentences, ntokens = sample["nsentences"], sample["ntokens"] + + # B x T + src_tokens, src_lengths = ( + sample["net_input"]["src_tokens"], + sample["net_input"]["src_lengths"], + ) + tgt_tokens, prev_output_tokens = sample["target"], sample["prev_target"] + + outputs = model(src_tokens, src_lengths, prev_output_tokens, tgt_tokens) + losses, nll_loss = [], [] + + for obj in outputs: + if outputs[obj].get("loss", None) is None: + _losses = self._compute_loss( + outputs[obj].get("out"), + outputs[obj].get("tgt"), + outputs[obj].get("mask", None), + outputs[obj].get("ls", 0.0), + name=obj + "-loss", + factor=outputs[obj].get("factor", 1.0), + ) + else: + _losses = self._custom_loss( + outputs[obj].get("loss"), + name=obj + "-loss", + factor=outputs[obj].get("factor", 1.0), + ) + + losses += [_losses] + if outputs[obj].get("nll_loss", False): + nll_loss += [_losses.get("nll_loss", 0.0)] + + loss = sum(l["loss"] for l in losses) + nll_loss = sum(l for l in nll_loss) if len(nll_loss) > 0 else loss.new_tensor(0) + + # NOTE: + # we don't need to use sample_size as denominator for the gradient + # here sample_size is just used for logging + sample_size = 1 + logging_output = { + "loss": loss.data, + "nll_loss": nll_loss.data, + "ntokens": ntokens, + "nsentences": nsentences, + "sample_size": sample_size, + } + + for l in losses: + logging_output[l["name"]] = ( + utils.item(l["loss"].data / l["factor"]) + if reduce + else l[["loss"]].data / l["factor"] + ) + + return loss, sample_size, logging_output + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + """Aggregate logging outputs from data parallel training.""" + sample_size = utils.item( + sum(log.get("sample_size", 0) for log in logging_outputs) + ) + loss = utils.item(sum(log.get("loss", 0) for log in logging_outputs)) + nll_loss = utils.item(sum(log.get("nll_loss", 0) for log in logging_outputs)) + + metrics.log_scalar( + "loss", loss / sample_size / math.log(2), sample_size, round=3 + ) + metrics.log_scalar( + "nll_loss", nll_loss / sample_size / math.log(2), sample_size, round=3 + ) + metrics.log_derived( + "ppl", lambda meters: utils.get_perplexity(meters["loss"].avg) + ) + + for key in logging_outputs[0]: + if key[-5:] == "-loss": + val = sum(log.get(key, 0) for log in logging_outputs) + metrics.log_scalar( + key[:-5], + val / sample_size / math.log(2) if sample_size > 0 else 0.0, + sample_size, + round=3, + ) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return True diff --git a/fairseq/criterions/sentence_prediction.py b/fairseq/criterions/sentence_prediction.py new file mode 100644 index 0000000000000000000000000000000000000000..9519fdc56d7de86b727f74ef5b18db520382e562 --- /dev/null +++ b/fairseq/criterions/sentence_prediction.py @@ -0,0 +1,99 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +import torch +import torch.nn.functional as F +from fairseq import metrics, utils +from fairseq.criterions import FairseqCriterion, register_criterion + + +@register_criterion("sentence_prediction") +class SentencePredictionCriterion(FairseqCriterion): + def __init__(self, task, classification_head_name, regression_target): + super().__init__(task) + self.classification_head_name = classification_head_name + self.regression_target = regression_target + + @staticmethod + def add_args(parser): + # fmt: off + parser.add_argument('--classification-head-name', + default='sentence_classification_head', + help='name of the classification head to use') + # fmt: on + + def forward(self, model, sample, reduce=True): + """Compute the loss for the given sample. + + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + assert ( + hasattr(model, "classification_heads") + and self.classification_head_name in model.classification_heads + ), "model must provide sentence classification head for --criterion=sentence_prediction" + + logits, _ = model( + **sample["net_input"], + features_only=True, + classification_head_name=self.classification_head_name, + ) + targets = model.get_targets(sample, [logits]).view(-1) + sample_size = targets.numel() + + if not self.regression_target: + lprobs = F.log_softmax(logits, dim=-1, dtype=torch.float32) + loss = F.nll_loss(lprobs, targets, reduction="sum") + else: + logits = logits.view(-1).float() + targets = targets.float() + loss = F.mse_loss(logits, targets, reduction="sum") + + logging_output = { + "loss": loss.data, + "ntokens": sample["ntokens"], + "nsentences": sample_size, + "sample_size": sample_size, + } + if not self.regression_target: + preds = logits.argmax(dim=1) + logging_output["ncorrect"] = (preds == targets).sum() + + return loss, sample_size, logging_output + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + """Aggregate logging outputs from data parallel training.""" + loss_sum = sum(log.get("loss", 0) for log in logging_outputs) + ntokens = sum(log.get("ntokens", 0) for log in logging_outputs) + nsentences = sum(log.get("nsentences", 0) for log in logging_outputs) + sample_size = sum(log.get("sample_size", 0) for log in logging_outputs) + + metrics.log_scalar( + "loss", loss_sum / sample_size / math.log(2), sample_size, round=3 + ) + if sample_size != ntokens: + metrics.log_scalar( + "nll_loss", loss_sum / ntokens / math.log(2), ntokens, round=3 + ) + + if len(logging_outputs) > 0 and "ncorrect" in logging_outputs[0]: + ncorrect = sum(log.get("ncorrect", 0) for log in logging_outputs) + metrics.log_scalar( + "accuracy", 100.0 * ncorrect / nsentences, nsentences, round=1 + ) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return True diff --git a/fairseq/criterions/sentence_ranking.py b/fairseq/criterions/sentence_ranking.py new file mode 100644 index 0000000000000000000000000000000000000000..d4c76341d4d87e6d0da21ac89e833ce0bda13a0c --- /dev/null +++ b/fairseq/criterions/sentence_ranking.py @@ -0,0 +1,120 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +import torch +import torch.nn.functional as F +from fairseq import metrics, utils +from fairseq.criterions import FairseqCriterion, register_criterion + + +@register_criterion("sentence_ranking") +class SentenceRankingCriterion(FairseqCriterion): + def __init__(self, task, ranking_head_name, save_predictions, num_classes): + super().__init__(task) + self.ranking_head_name = ranking_head_name + if save_predictions is not None: + self.prediction_h = open(save_predictions, "w") + else: + self.prediction_h = None + self.num_classes = num_classes + + def __del__(self): + if self.prediction_h is not None: + self.prediction_h.close() + + @staticmethod + def add_args(parser): + # fmt: off + parser.add_argument('--save-predictions', metavar='FILE', + help='file to save predictions to') + parser.add_argument('--ranking-head-name', + default='sentence_classification_head', + help='name of the ranking head to use') + # fmt: on + + def forward(self, model, sample, reduce=True): + """Compute ranking loss for the given sample. + + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + assert ( + hasattr(model, "classification_heads") + and self.ranking_head_name in model.classification_heads + ), "model must provide sentence ranking head for --criterion=sentence_ranking" + + scores = [] + for idx in range(self.num_classes): + score, _ = model( + **sample["net_input{idx}".format(idx=idx + 1)], + classification_head_name=self.ranking_head_name, + ) + scores.append(score) + + logits = torch.cat(scores, dim=1) + sample_size = logits.size(0) + + if "target" in sample: + targets = model.get_targets(sample, [logits]).view(-1) + lprobs = F.log_softmax(logits, dim=-1, dtype=torch.float32) + loss = F.nll_loss(lprobs, targets, reduction="sum") + else: + targets = None + loss = torch.tensor(0.0, requires_grad=True) + + if self.prediction_h is not None: + preds = logits.argmax(dim=1) + for i, (id, pred) in enumerate(zip(sample["id"].tolist(), preds.tolist())): + if targets is not None: + label = targets[i].item() + print("{}\t{}\t{}".format(id, pred, label), file=self.prediction_h) + else: + print("{}\t{}".format(id, pred), file=self.prediction_h) + + logging_output = { + "loss": loss.data, + "ntokens": sample["ntokens"], + "nsentences": sample_size, + "sample_size": sample_size, + } + if targets is not None: + logging_output["ncorrect"] = (logits.argmax(dim=1) == targets).sum() + + return loss, sample_size, logging_output + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + """Aggregate logging outputs from data parallel training.""" + loss_sum = sum(log.get("loss", 0) for log in logging_outputs) + ntokens = sum(log.get("ntokens", 0) for log in logging_outputs) + nsentences = sum(log.get("nsentences", 0) for log in logging_outputs) + sample_size = sum(log.get("sample_size", 0) for log in logging_outputs) + + metrics.log_scalar( + "loss", loss_sum / sample_size / math.log(2), sample_size, round=3 + ) + if sample_size != ntokens: + metrics.log_scalar( + "nll_loss", loss_sum / ntokens / math.log(2), ntokens, round=3 + ) + + if len(logging_outputs) > 0 and "ncorrect" in logging_outputs[0]: + ncorrect = sum(log.get("ncorrect", 0) for log in logging_outputs) + metrics.log_scalar( + "accuracy", 100.0 * ncorrect / nsentences, nsentences, round=1 + ) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return True diff --git a/fairseq/criterions/wav2vec_criterion.py b/fairseq/criterions/wav2vec_criterion.py new file mode 100644 index 0000000000000000000000000000000000000000..e04786cc3b75517cefd06303f98f8536f9279311 --- /dev/null +++ b/fairseq/criterions/wav2vec_criterion.py @@ -0,0 +1,229 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +from dataclasses import dataclass, field +from typing import List, Optional + +import torch +import torch.nn.functional as F +from fairseq import metrics, utils +from fairseq.criterions import FairseqCriterion, register_criterion +from fairseq.dataclass import FairseqDataclass +from fairseq.logging.meters import safe_round +from fairseq.utils import is_xla_tensor + + +@dataclass +class Wav2VecCriterionConfig(FairseqDataclass): + infonce: bool = field( + default=False, + metadata={ + "help": "if set, uses cross entropy instead of binary cross entropy (i.e. InfoNCE loss)" + }, + ) + loss_weights: Optional[List[float]] = field( + default=None, + metadata={"help": "weights for additional loss terms (not first one)"}, + ) + log_keys: List[str] = field( + default_factory=lambda: [], + metadata={"help": "output keys to log"}, + ) + +@register_criterion("wav2vec", dataclass=Wav2VecCriterionConfig) +class Wav2vecCriterion(FairseqCriterion): + def __init__(self, task, infonce=False, loss_weights=None, log_keys=None): + super().__init__(task) + self.infonce = infonce + self.loss_weights = loss_weights + self.log_keys = [] if log_keys is None else log_keys + + def forward(self, model, sample, reduce=True): + """Compute the loss for the given sample. + + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + net_output = model(**sample["net_input"]) + logits = model.get_logits(net_output).float() + target = model.get_targets(sample, net_output) + self.xla = is_xla_tensor(logits) + + # XXX: handle weights on xla. + weights = None + if hasattr(model, "get_target_weights") and not self.infonce: + weights = model.get_target_weights(target, net_output) + if torch.is_tensor(weights): + weights = weights.float() + + losses = [] + + reduction = "none" if ((not reduce) or self.xla) else "sum" + if self.infonce: + loss = F.cross_entropy(logits, target, reduction=reduction) + else: + loss = F.binary_cross_entropy_with_logits( + logits, target.float(), weights, reduction=reduction + ) + + if self.xla: + # tpu-comment: since dynamic shapes lead to recompilations on xla, + # we don't shrink tensors using mask_indices. + # Instead, we use mask indices to adjust loss. + mi = ( + sample['net_input']['mask_indices'] + .transpose(0, 1) # logits are transposed in `model.get_logits` + .reshape(logits.size(0)) + ) + loss = (loss * mi).sum() if reduce else (loss * mi) + + if 'sample_size' in sample: + sample_size = sample['sample_size'] + elif 'mask_indices' in sample['net_input']: + sample_size = sample['net_input']['mask_indices'].sum() + else: + sample_size = target.numel() if self.infonce else target.long().sum().item() + losses.append(loss.detach().clone()) + + if self.loss_weights is not None: + assert hasattr(model, "get_extra_losses") + extra_losses = model.get_extra_losses(net_output) + if torch.is_tensor(extra_losses): + extra_losses = [extra_losses] + if len(self.loss_weights) == 1 and len(extra_losses) != 1: + self.loss_weights = [self.loss_weights[0]] * len(extra_losses) + assert len(extra_losses) == len( + self.loss_weights + ), f"{len(extra_losses)}, {len(self.loss_weights)}" + for p, coef in zip(extra_losses, self.loss_weights): + if coef != 0 and p is not None: + p = coef * p.float() * sample_size + loss += p + losses.append(p) + + logging_output = { + "loss": loss.item() if (reduce and not self.xla) else loss.detach(), + "ntokens": sample_size, + "nsentences": sample["id"].numel(), + "sample_size": sample_size, + } + + for lk in self.log_keys: + # Only store "logits" and "target" for computing MAP and MAUC + # during validation + if lk == "logits": + if not self.training: + logging_output["logits"] = logits.cpu().numpy() + elif lk == "target": + if not self.training: + # If the targets have been mixed with the predictions of + # teacher models, find the original targets + if hasattr(model, "get_original_targets"): + original_target = model.get_original_targets(sample, net_output) + else: + original_target = target + logging_output["target"] = original_target.cpu().numpy() + elif lk in net_output: + value = net_output[lk] + if not is_xla_tensor(value): + value = float(value) + logging_output[lk] = value + + if len(losses) > 1: + for i, l in enumerate(losses): + logging_output[f"loss_{i}"] = l.item() if not self.xla else l.detach() + + if self.infonce: + with torch.no_grad(): + if logits.numel() == 0: + corr = 0 + count = 0 + else: + assert logits.dim() > 1, logits.shape + max = logits.argmax(-1) == 0 + min = logits.argmin(-1) == 0 + if is_xla_tensor(logits): + max, min = max * mi, min * mi + both = max & min + corr = max.long().sum() - both.long().sum() + count = mi.sum() + else: + both = max & min + corr = max.long().sum().item() - both.long().sum().item() + count = float(max.numel()) + + logging_output["correct"] = corr + logging_output["count"] = count + + return loss, sample_size, logging_output + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + """Aggregate logging outputs from data parallel training.""" + loss_sum = utils.item(sum(log.get("loss", 0) for log in logging_outputs)) + ntokens = utils.item(sum(log.get("ntokens", 0) for log in logging_outputs)) + nsentences = utils.item( + sum(log.get("nsentences", 0) for log in logging_outputs) + ) + sample_size = utils.item( + sum(log.get("sample_size", 0) for log in logging_outputs) + ) + + metrics.log_scalar( + "loss", loss_sum / (sample_size or 1) / math.log(2), sample_size, round=3 + ) + metrics.log_scalar("ntokens", ntokens) + metrics.log_scalar("nsentences", nsentences) + + correct = sum(log.get("correct", 0) for log in logging_outputs) + metrics.log_scalar("_correct", correct) + + total = sum(log.get("count", 0) for log in logging_outputs) + metrics.log_scalar("_total", total) + + if total > 0: + metrics.log_derived( + "accuracy", + lambda meters: safe_round( + meters["_correct"].sum / meters["_total"].sum, 5 + ) + if meters["_total"].sum > 0 + else float("nan"), + ) + + builtin_keys = { + "loss", + "ntokens", + "nsentences", + "sample_size", + "correct", + "count", + } + + for k in logging_outputs[0]: + if k not in builtin_keys: + val = sum(log.get(k, 0) for log in logging_outputs) + if k.startswith("loss"): + metrics.log_scalar( + k, val / (sample_size or 1) / math.log(2), sample_size, round=3 + ) + else: + metrics.log_scalar(k, val / len(logging_outputs), round=3) + + # FIXME: revert when gather based xla reduction is implemented + #@staticmethod + #def logging_outputs_can_be_summed() -> bool: + def logging_outputs_can_be_summed(self) -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + # XXX: Gather based reduction not implemented for xla yet. + # So we fall to sum based reduction for xla. + return self.xla diff --git a/fairseq/data/__init__.py b/fairseq/data/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..8b7eb2ec4fc5190c4dcdfe34b0259e6f448e18a9 --- /dev/null +++ b/fairseq/data/__init__.py @@ -0,0 +1,128 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +"""isort:skip_file""" + +from .dictionary import Dictionary, TruncatedDictionary + +from .fairseq_dataset import FairseqDataset, FairseqIterableDataset + +from .base_wrapper_dataset import BaseWrapperDataset + +from .add_target_dataset import AddTargetDataset +from .append_token_dataset import AppendTokenDataset +from .audio.raw_audio_dataset import BinarizedAudioDataset, FileAudioDataset +from .audio.hubert_dataset import HubertDataset +from .backtranslation_dataset import BacktranslationDataset +from .bucket_pad_length_dataset import BucketPadLengthDataset +from .colorize_dataset import ColorizeDataset +from .concat_dataset import ConcatDataset +from .concat_sentences_dataset import ConcatSentencesDataset +from .denoising_dataset import DenoisingDataset +from .id_dataset import IdDataset +from .indexed_dataset import ( + IndexedCachedDataset, + IndexedDataset, + IndexedRawTextDataset, + MMapIndexedDataset, +) +from .language_pair_dataset import LanguagePairDataset +from .list_dataset import ListDataset +from .lm_context_window_dataset import LMContextWindowDataset +from .lru_cache_dataset import LRUCacheDataset +from .mask_tokens_dataset import MaskTokensDataset +from .monolingual_dataset import MonolingualDataset +from .multi_corpus_sampled_dataset import MultiCorpusSampledDataset +from .nested_dictionary_dataset import NestedDictionaryDataset +from .noising import NoisingDataset +from .numel_dataset import NumelDataset +from .num_samples_dataset import NumSamplesDataset +from .offset_tokens_dataset import OffsetTokensDataset +from .pad_dataset import LeftPadDataset, PadDataset, RightPadDataset +from .prepend_dataset import PrependDataset +from .prepend_token_dataset import PrependTokenDataset +from .raw_label_dataset import RawLabelDataset +from .replace_dataset import ReplaceDataset +from .resampling_dataset import ResamplingDataset +from .roll_dataset import RollDataset +from .round_robin_zip_datasets import RoundRobinZipDatasets +from .sort_dataset import SortDataset +from .strip_token_dataset import StripTokenDataset +from .subsample_dataset import SubsampleDataset +from .token_block_dataset import TokenBlockDataset +from .transform_eos_dataset import TransformEosDataset +from .transform_eos_lang_pair_dataset import TransformEosLangPairDataset +from .shorten_dataset import TruncateDataset, RandomCropDataset +from .multilingual.sampled_multi_dataset import SampledMultiDataset +from .multilingual.sampled_multi_epoch_dataset import SampledMultiEpochDataset +from .fasta_dataset import FastaDataset, EncodedFastaDataset + +from .iterators import ( + CountingIterator, + EpochBatchIterator, + GroupedIterator, + ShardedIterator, +) + +__all__ = [ + "AddTargetDataset", + "AppendTokenDataset", + "BacktranslationDataset", + "BaseWrapperDataset", + "BinarizedAudioDataset", + "BucketPadLengthDataset", + "ColorizeDataset", + "ConcatDataset", + "ConcatSentencesDataset", + "CountingIterator", + "DenoisingDataset", + "Dictionary", + "EncodedFastaDataset", + "EpochBatchIterator", + "FairseqDataset", + "FairseqIterableDataset", + "FastaDataset", + "FileAudioDataset", + "GroupedIterator", + "HubertDataset", + "IdDataset", + "IndexedCachedDataset", + "IndexedDataset", + "IndexedRawTextDataset", + "LanguagePairDataset", + "LeftPadDataset", + "ListDataset", + "LMContextWindowDataset", + "LRUCacheDataset", + "MaskTokensDataset", + "MMapIndexedDataset", + "MonolingualDataset", + "MultiCorpusSampledDataset", + "NestedDictionaryDataset", + "NoisingDataset", + "NumelDataset", + "NumSamplesDataset", + "OffsetTokensDataset", + "PadDataset", + "PrependDataset", + "PrependTokenDataset", + "RandomCropDataset", + "RawLabelDataset", + "ResamplingDataset", + "ReplaceDataset", + "RightPadDataset", + "RollDataset", + "RoundRobinZipDatasets", + "SampledMultiDataset", + "SampledMultiEpochDataset", + "ShardedIterator", + "SortDataset", + "StripTokenDataset", + "SubsampleDataset", + "TokenBlockDataset", + "TransformEosDataset", + "TransformEosLangPairDataset", + "TruncateDataset", + "TruncatedDictionary", +] diff --git a/fairseq/data/add_target_dataset.py b/fairseq/data/add_target_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..9ef467058b89d9d74f703acbe5b45cb5ef9b2b69 --- /dev/null +++ b/fairseq/data/add_target_dataset.py @@ -0,0 +1,70 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch + +from . import BaseWrapperDataset, data_utils + + +class AddTargetDataset(BaseWrapperDataset): + def __init__( + self, + dataset, + labels, + pad, + eos, + batch_targets, + process_label=None, + add_to_input=False, + ): + super().__init__(dataset) + self.labels = labels + self.batch_targets = batch_targets + self.pad = pad + self.eos = eos + self.process_label = process_label + self.add_to_input = add_to_input + + def get_label(self, index): + return ( + self.labels[index] + if self.process_label is None + else self.process_label(self.labels[index]) + ) + + def __getitem__(self, index): + item = self.dataset[index] + item["label"] = self.get_label(index) + return item + + def size(self, index): + sz = self.dataset.size(index) + own_sz = len(self.get_label(index)) + return (sz, own_sz) + + def collater(self, samples): + collated = self.dataset.collater(samples) + if len(collated) == 0: + return collated + indices = set(collated["id"].tolist()) + target = [s["label"] for s in samples if s["id"] in indices] + + if self.batch_targets: + collated["target_lengths"] = torch.LongTensor([len(t) for t in target]) + target = data_utils.collate_tokens(target, pad_idx=self.pad, left_pad=False) + collated["ntokens"] = collated["target_lengths"].sum().item() + else: + collated["ntokens"] = sum([len(t) for t in target]) + + collated["target"] = target + + if self.add_to_input: + eos = target.new_full((target.size(0), 1), self.eos) + collated["target"] = torch.cat([target, eos], dim=-1).long() + collated["net_input"]["prev_output_tokens"] = torch.cat( + [eos, target], dim=-1 + ).long() + collated["ntokens"] += target.size(0) + return collated diff --git a/fairseq/data/append_token_dataset.py b/fairseq/data/append_token_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..87695bd0f5fcb6b10247e3b743340623e6438cc1 --- /dev/null +++ b/fairseq/data/append_token_dataset.py @@ -0,0 +1,41 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import numpy as np +import torch + +from . import BaseWrapperDataset + + +class AppendTokenDataset(BaseWrapperDataset): + def __init__(self, dataset, token=None): + super().__init__(dataset) + self.token = token + if token is not None: + self._sizes = np.array(dataset.sizes) + 1 + else: + self._sizes = dataset.sizes + + def __getitem__(self, idx): + item = self.dataset[idx] + if self.token is not None: + item = torch.cat([item, item.new([self.token])]) + return item + + @property + def sizes(self): + return self._sizes + + def num_tokens(self, index): + n = self.dataset.num_tokens(index) + if self.token is not None: + n += 1 + return n + + def size(self, index): + n = self.dataset.size(index) + if self.token is not None: + n += 1 + return n diff --git a/fairseq/data/audio/__init__.py b/fairseq/data/audio/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/fairseq/data/audio/audio_utils.py b/fairseq/data/audio/audio_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..f51cb0cddc29c732a4573b7b3e915844a80cd2f3 --- /dev/null +++ b/fairseq/data/audio/audio_utils.py @@ -0,0 +1,174 @@ +from pathlib import Path +from typing import BinaryIO, Optional, Tuple, Union, List + +import numpy as np +import torch + + +SF_AUDIO_FILE_EXTENSIONS = {".wav", ".flac", ".ogg"} +FEATURE_OR_SF_AUDIO_FILE_EXTENSIONS = {".npy", ".wav", ".flac", ".ogg"} + + +def _convert_to_mono( + waveform: torch.FloatTensor, sample_rate: int +) -> torch.FloatTensor: + if waveform.shape[0] > 1: + try: + import torchaudio.sox_effects as ta_sox + except ImportError: + raise ImportError( + "Please install torchaudio to convert multi-channel audios" + ) + effects = [['channels', '1']] + return ta_sox.apply_effects_tensor(waveform, sample_rate, effects)[0] + return waveform + + +def convert_to_mono(waveform: np.ndarray, sample_rate: int) -> np.ndarray: + if waveform.shape[0] > 1: + _waveform = torch.from_numpy(waveform) + return _convert_to_mono(_waveform, sample_rate).numpy() + return waveform + + +def get_waveform( + path_or_fp: Union[str, BinaryIO], normalization=True, mono=True, + frames=-1, start=0, always_2d=True +) -> Tuple[np.ndarray, int]: + """Get the waveform and sample rate of a 16-bit WAV/FLAC/OGG Vorbis audio. + + Args: + path_or_fp (str or BinaryIO): the path or file-like object + normalization (bool): Normalize values to [-1, 1] (Default: True) + mono (bool): convert multi-channel audio to mono-channel one + frames (int): the number of frames to read. (-1 for reading all) + start (int): Where to start reading. A negative value counts from the end. + always_2d (bool): always return 2D array even for mono-channel audios + Returns: + waveform (numpy.ndarray): 1D or 2D waveform (channels x length) + sample_rate (float): sample rate + """ + if isinstance(path_or_fp, str): + ext = Path(path_or_fp).suffix + if ext not in SF_AUDIO_FILE_EXTENSIONS: + raise ValueError(f"Unsupported audio format: {ext}") + + try: + import soundfile as sf + except ImportError: + raise ImportError( + "Please install soundfile to load WAV/FLAC/OGG Vorbis audios" + ) + + waveform, sample_rate = sf.read( + path_or_fp, dtype="float32", always_2d=True, frames=frames, start=start + ) + waveform = waveform.T # T x C -> C x T + if mono and waveform.shape[0] > 1: + waveform = convert_to_mono(waveform, sample_rate) + if not normalization: + waveform *= 2 ** 15 # denormalized to 16-bit signed integers + if not always_2d: + waveform = waveform.squeeze(axis=0) + return waveform, sample_rate + + +def _get_kaldi_fbank( + waveform: np.ndarray, sample_rate: int, n_bins=80 +) -> Optional[np.ndarray]: + """Get mel-filter bank features via PyKaldi.""" + try: + from kaldi.feat.mel import MelBanksOptions + from kaldi.feat.fbank import FbankOptions, Fbank + from kaldi.feat.window import FrameExtractionOptions + from kaldi.matrix import Vector + + mel_opts = MelBanksOptions() + mel_opts.num_bins = n_bins + frame_opts = FrameExtractionOptions() + frame_opts.samp_freq = sample_rate + opts = FbankOptions() + opts.mel_opts = mel_opts + opts.frame_opts = frame_opts + fbank = Fbank(opts=opts) + features = fbank.compute(Vector(waveform.squeeze()), 1.0).numpy() + return features + except ImportError: + return None + + +def _get_torchaudio_fbank( + waveform: np.ndarray, sample_rate, n_bins=80 +) -> Optional[np.ndarray]: + """Get mel-filter bank features via TorchAudio.""" + try: + import torchaudio.compliance.kaldi as ta_kaldi + waveform = torch.from_numpy(waveform) + features = ta_kaldi.fbank( + waveform, num_mel_bins=n_bins, sample_frequency=sample_rate + ) + return features.numpy() + except ImportError: + return None + + +def get_fbank(path_or_fp: Union[str, BinaryIO], n_bins=80) -> np.ndarray: + """Get mel-filter bank features via PyKaldi or TorchAudio. Prefer PyKaldi + (faster CPP implementation) to TorchAudio (Python implementation). Note that + Kaldi/TorchAudio requires 16-bit signed integers as inputs and hence the + waveform should not be normalized.""" + waveform, sample_rate = get_waveform(path_or_fp, normalization=False) + + features = _get_kaldi_fbank(waveform, sample_rate, n_bins) + if features is None: + features = _get_torchaudio_fbank(waveform, sample_rate, n_bins) + if features is None: + raise ImportError( + "Please install pyKaldi or torchaudio to enable " + "online filterbank feature extraction" + ) + + return features + + +def is_npy_data(data: bytes) -> bool: + return data[0] == 147 and data[1] == 78 + + +def is_sf_audio_data(data: bytes) -> bool: + is_wav = (data[0] == 82 and data[1] == 73 and data[2] == 70) + is_flac = (data[0] == 102 and data[1] == 76 and data[2] == 97) + is_ogg = (data[0] == 79 and data[1] == 103 and data[2] == 103) + return is_wav or is_flac or is_ogg + + +def read_from_stored_zip(zip_path: str, offset: int, file_size: int) -> bytes: + with open(zip_path, "rb") as f: + f.seek(offset) + data = f.read(file_size) + return data + + +def parse_path(path: str) -> Tuple[str, List[int]]: + """Parse data path which is either a path to + 1. a .npy/.wav/.flac/.ogg file + 2. a stored ZIP file with slicing info: "[zip_path]:[offset]:[length]" + + Args: + path (str): the data path to parse + + Returns: + file_path (str): the file path + slice_ptr (list of int): empty in case 1; + byte offset and length for the slice in case 2 + """ + + if Path(path).suffix in FEATURE_OR_SF_AUDIO_FILE_EXTENSIONS: + _path, slice_ptr = path, [] + else: + _path, *slice_ptr = path.split(":") + if not Path(_path).is_file(): + raise FileNotFoundError(f"File not found: {_path}") + assert len(slice_ptr) in {0, 2}, f"Invalid path: {path}" + slice_ptr = [int(i) for i in slice_ptr] + return _path, slice_ptr diff --git a/fairseq/data/audio/feature_transforms/__init__.py b/fairseq/data/audio/feature_transforms/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..359fa069716cba0dd615ce0959368b20828c31f7 --- /dev/null +++ b/fairseq/data/audio/feature_transforms/__init__.py @@ -0,0 +1,82 @@ +import importlib +import os +from abc import ABC, abstractmethod +from typing import Dict, Optional + + +class AudioFeatureTransform(ABC): + @classmethod + @abstractmethod + def from_config_dict(cls, config: Optional[Dict] = None): + pass + + +AUDIO_FEATURE_TRANSFORM_REGISTRY = {} +AUDIO_FEATURE_TRANSFORM_CLASS_NAMES = set() + + +def register_audio_feature_transform(name): + def register_audio_feature_transform_cls(cls): + if name in AUDIO_FEATURE_TRANSFORM_REGISTRY: + raise ValueError(f"Cannot register duplicate transform ({name})") + if not issubclass(cls, AudioFeatureTransform): + raise ValueError( + f"Transform ({name}: {cls.__name__}) must extend " + "AudioFeatureTransform" + ) + if cls.__name__ in AUDIO_FEATURE_TRANSFORM_CLASS_NAMES: + raise ValueError( + f"Cannot register audio feature transform with duplicate " + f"class name ({cls.__name__})" + ) + AUDIO_FEATURE_TRANSFORM_REGISTRY[name] = cls + AUDIO_FEATURE_TRANSFORM_CLASS_NAMES.add(cls.__name__) + return cls + + return register_audio_feature_transform_cls + + +def get_audio_feature_transform(name): + return AUDIO_FEATURE_TRANSFORM_REGISTRY[name] + + +transforms_dir = os.path.dirname(__file__) +for file in os.listdir(transforms_dir): + path = os.path.join(transforms_dir, file) + if ( + not file.startswith("_") + and not file.startswith(".") + and (file.endswith(".py") or os.path.isdir(path)) + ): + name = file[: file.find(".py")] if file.endswith(".py") else file + importlib.import_module("fairseq.data.audio.feature_transforms." + name) + + +class CompositeAudioFeatureTransform(AudioFeatureTransform): + @classmethod + def from_config_dict(cls, config=None): + _config = {} if config is None else config + _transforms = _config.get("transforms") + if _transforms is None: + return None + transforms = [ + get_audio_feature_transform(_t).from_config_dict(_config.get(_t)) + for _t in _transforms + ] + return CompositeAudioFeatureTransform(transforms) + + def __init__(self, transforms): + self.transforms = [t for t in transforms if t is not None] + + def __call__(self, x): + for t in self.transforms: + x = t(x) + return x + + def __repr__(self): + format_string = ( + [self.__class__.__name__ + "("] + + [f" {t.__repr__()}" for t in self.transforms] + + [")"] + ) + return "\n".join(format_string) diff --git a/fairseq/data/audio/feature_transforms/global_cmvn.py b/fairseq/data/audio/feature_transforms/global_cmvn.py new file mode 100644 index 0000000000000000000000000000000000000000..e457ff176fee3b996da11f47e7dc61b81c445ba3 --- /dev/null +++ b/fairseq/data/audio/feature_transforms/global_cmvn.py @@ -0,0 +1,29 @@ +import numpy as np +from fairseq.data.audio.feature_transforms import ( + AudioFeatureTransform, + register_audio_feature_transform, +) + + +@register_audio_feature_transform("global_cmvn") +class GlobalCMVN(AudioFeatureTransform): + """Global CMVN (cepstral mean and variance normalization). The global mean + and variance need to be pre-computed and stored in NumPy format (.npz).""" + + @classmethod + def from_config_dict(cls, config=None): + _config = {} if config is None else config + return GlobalCMVN(_config.get("stats_npz_path")) + + def __init__(self, stats_npz_path): + self.stats_npz_path = stats_npz_path + stats = np.load(stats_npz_path) + self.mean, self.std = stats["mean"], stats["std"] + + def __repr__(self): + return self.__class__.__name__ + f'(stats_npz_path="{self.stats_npz_path}")' + + def __call__(self, x): + x = np.subtract(x, self.mean) + x = np.divide(x, self.std) + return x diff --git a/fairseq/data/audio/feature_transforms/specaugment.py b/fairseq/data/audio/feature_transforms/specaugment.py new file mode 100644 index 0000000000000000000000000000000000000000..ce5802b41a903ea8f3e3e8a169d5048b4e908f99 --- /dev/null +++ b/fairseq/data/audio/feature_transforms/specaugment.py @@ -0,0 +1,131 @@ +import math +import numbers +from typing import Optional + +import numpy as np +from fairseq.data.audio.feature_transforms import ( + AudioFeatureTransform, + register_audio_feature_transform, +) + + +@register_audio_feature_transform("specaugment") +class SpecAugmentTransform(AudioFeatureTransform): + """SpecAugment (https://arxiv.org/abs/1904.08779)""" + + @classmethod + def from_config_dict(cls, config=None): + _config = {} if config is None else config + return SpecAugmentTransform( + _config.get("time_warp_W", 0), + _config.get("freq_mask_N", 0), + _config.get("freq_mask_F", 0), + _config.get("time_mask_N", 0), + _config.get("time_mask_T", 0), + _config.get("time_mask_p", 0.0), + _config.get("mask_value", None), + ) + + def __init__( + self, + time_warp_w: int = 0, + freq_mask_n: int = 0, + freq_mask_f: int = 0, + time_mask_n: int = 0, + time_mask_t: int = 0, + time_mask_p: float = 0.0, + mask_value: Optional[float] = 0.0, + ): + # Sanity checks + assert mask_value is None or isinstance( + mask_value, numbers.Number + ), f"mask_value (type: {type(mask_value)}) must be None or a number" + if freq_mask_n > 0: + assert freq_mask_f > 0, ( + f"freq_mask_F ({freq_mask_f}) " + f"must be larger than 0 when doing freq masking." + ) + if time_mask_n > 0: + assert time_mask_t > 0, ( + f"time_mask_T ({time_mask_t}) must be larger than 0 when " + f"doing time masking." + ) + + self.time_warp_w = time_warp_w + self.freq_mask_n = freq_mask_n + self.freq_mask_f = freq_mask_f + self.time_mask_n = time_mask_n + self.time_mask_t = time_mask_t + self.time_mask_p = time_mask_p + self.mask_value = mask_value + + def __repr__(self): + return ( + self.__class__.__name__ + + "(" + + ", ".join( + [ + f"time_warp_w={self.time_warp_w}", + f"freq_mask_n={self.freq_mask_n}", + f"freq_mask_f={self.freq_mask_f}", + f"time_mask_n={self.time_mask_n}", + f"time_mask_t={self.time_mask_t}", + f"time_mask_p={self.time_mask_p}", + ] + ) + + ")" + ) + + def __call__(self, spectrogram): + assert len(spectrogram.shape) == 2, "spectrogram must be a 2-D tensor." + + distorted = spectrogram.copy() # make a copy of input spectrogram. + num_frames = spectrogram.shape[0] # or 'tau' in the paper. + num_freqs = spectrogram.shape[1] # or 'miu' in the paper. + mask_value = self.mask_value + + if mask_value is None: # if no value was specified, use local mean. + mask_value = spectrogram.mean() + + if num_frames == 0: + return spectrogram + + if num_freqs < self.freq_mask_f: + return spectrogram + + if self.time_warp_w > 0: + if 2 * self.time_warp_w < num_frames: + import cv2 + + w0 = np.random.randint(self.time_warp_w, num_frames - self.time_warp_w) + w = np.random.randint(-self.time_warp_w + 1, self.time_warp_w) + upper, lower = distorted[:w0, :], distorted[w0:, :] + upper = cv2.resize( + upper, dsize=(num_freqs, w0 + w), interpolation=cv2.INTER_LINEAR + ) + lower = cv2.resize( + lower, + dsize=(num_freqs, num_frames - w0 - w), + interpolation=cv2.INTER_LINEAR, + ) + distorted = np.concatenate((upper, lower), axis=0) + + for _i in range(self.freq_mask_n): + f = np.random.randint(0, self.freq_mask_f) + f0 = np.random.randint(0, num_freqs - f) + if f != 0: + distorted[:, f0 : f0 + f] = mask_value + + max_time_mask_t = min( + self.time_mask_t, math.floor(num_frames * self.time_mask_p) + ) + if max_time_mask_t < 1: + return distorted + + for _i in range(self.time_mask_n): + t = np.random.randint(0, max_time_mask_t) + t0 = np.random.randint(0, num_frames - t) + if t != 0: + distorted[t0 : t0 + t, :] = mask_value + + return distorted diff --git a/fairseq/data/audio/feature_transforms/utterance_cmvn.py b/fairseq/data/audio/feature_transforms/utterance_cmvn.py new file mode 100644 index 0000000000000000000000000000000000000000..6bbd0ae821b42ab693f4141e7c161d6d7cb0b15a --- /dev/null +++ b/fairseq/data/audio/feature_transforms/utterance_cmvn.py @@ -0,0 +1,40 @@ +import numpy as np +from fairseq.data.audio.feature_transforms import ( + AudioFeatureTransform, + register_audio_feature_transform, +) + + +@register_audio_feature_transform("utterance_cmvn") +class UtteranceCMVN(AudioFeatureTransform): + """Utterance-level CMVN (cepstral mean and variance normalization)""" + + @classmethod + def from_config_dict(cls, config=None): + _config = {} if config is None else config + return UtteranceCMVN( + _config.get("norm_means", True), + _config.get("norm_vars", True), + ) + + def __init__(self, norm_means=True, norm_vars=True): + self.norm_means, self.norm_vars = norm_means, norm_vars + + def __repr__(self): + return ( + self.__class__.__name__ + + f"(norm_means={self.norm_means}, norm_vars={self.norm_vars})" + ) + + def __call__(self, x): + mean = x.mean(axis=0) + square_sums = (x ** 2).sum(axis=0) + + if self.norm_means: + x = np.subtract(x, mean) + if self.norm_vars: + var = square_sums / x.shape[0] - mean ** 2 + std = np.sqrt(np.maximum(var, 1e-10)) + x = np.divide(x, std) + + return x diff --git a/fairseq/data/audio/hubert_dataset.py b/fairseq/data/audio/hubert_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..f00fe301a64a8740ed3ce07e44f6774edb933926 --- /dev/null +++ b/fairseq/data/audio/hubert_dataset.py @@ -0,0 +1,358 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import itertools +import logging +import os +import sys +from typing import Any, List, Optional, Union + +import numpy as np + +import torch +import torch.nn.functional as F +from fairseq.data import data_utils +from fairseq.data.fairseq_dataset import FairseqDataset + +logger = logging.getLogger(__name__) + + +def load_audio(manifest_path, max_keep, min_keep): + n_long, n_short = 0, 0 + names, inds, sizes = [], [], [] + with open(manifest_path) as f: + root = f.readline().strip() + for ind, line in enumerate(f): + items = line.strip().split("\t") + assert len(items) == 2, line + sz = int(items[1]) + if min_keep is not None and sz < min_keep: + n_short += 1 + elif max_keep is not None and sz > max_keep: + n_long += 1 + else: + names.append(items[0]) + inds.append(ind) + sizes.append(sz) + tot = ind + 1 + logger.info( + ( + f"max_keep={max_keep}, min_keep={min_keep}, " + f"loaded {len(names)}, skipped {n_short} short and {n_long} long, " + f"longest-loaded={max(sizes)}, shortest-loaded={min(sizes)}" + ) + ) + return root, names, inds, tot, sizes + + +def load_label(label_path, inds, tot): + with open(label_path) as f: + labels = [line.rstrip() for line in f] + assert ( + len(labels) == tot + ), f"number of labels does not match ({len(labels)} != {tot})" + labels = [labels[i] for i in inds] + return labels + + +def load_label_offset(label_path, inds, tot): + with open(label_path) as f: + code_lengths = [len(line.encode("utf-8")) for line in f] + assert ( + len(code_lengths) == tot + ), f"number of labels does not match ({len(code_lengths)} != {tot})" + offsets = list(itertools.accumulate([0] + code_lengths)) + offsets = [(offsets[i], offsets[i + 1]) for i in inds] + return offsets + + +def verify_label_lengths( + audio_sizes, + audio_rate, + label_path, + label_rate, + inds, + tot, + tol=0.1, # tolerance in seconds +): + if label_rate < 0: + logger.info(f"{label_path} is sequence label. skipped") + return + + with open(label_path) as f: + lengths = [len(line.rstrip().split()) for line in f] + assert len(lengths) == tot + lengths = [lengths[i] for i in inds] + num_invalid = 0 + for i, ind in enumerate(inds): + dur_from_audio = audio_sizes[i] / audio_rate + dur_from_label = lengths[i] / label_rate + if abs(dur_from_audio - dur_from_label) > tol: + logger.warning( + ( + f"audio and label duration differ too much " + f"(|{dur_from_audio} - {dur_from_label}| > {tol}) " + f"in line {ind+1} of {label_path}. Check if `label_rate` " + f"is correctly set (currently {label_rate}). " + f"num. of samples = {audio_sizes[i]}; " + f"label length = {lengths[i]}" + ) + ) + num_invalid += 1 + if num_invalid > 0: + logger.warning( + f"total {num_invalid} (audio, label) pairs with mismatched lengths" + ) + + +class HubertDataset(FairseqDataset): + def __init__( + self, + manifest_path: str, + sample_rate: float, + label_paths: List[str], + label_rates: Union[List[float], float], # -1 for sequence labels + pad_list: List[str], + eos_list: List[str], + label_processors: Optional[List[Any]] = None, + max_keep_sample_size: Optional[int] = None, + min_keep_sample_size: Optional[int] = None, + max_sample_size: Optional[int] = None, + shuffle: bool = True, + pad_audio: bool = False, + normalize: bool = False, + store_labels: bool = True, + random_crop: bool = False, + single_target: bool = False, + ): + self.audio_root, self.audio_names, inds, tot, self.sizes = load_audio( + manifest_path, max_keep_sample_size, min_keep_sample_size + ) + self.sample_rate = sample_rate + self.shuffle = shuffle + self.random_crop = random_crop + + self.num_labels = len(label_paths) + self.pad_list = pad_list + self.eos_list = eos_list + self.label_processors = label_processors + self.single_target = single_target + self.label_rates = ( + [label_rates for _ in range(len(label_paths))] + if isinstance(label_rates, int) + else label_rates + ) + self.store_labels = store_labels + if store_labels: + self.label_list = [load_label(p, inds, tot) for p in label_paths] + else: + self.label_paths = label_paths + self.label_offsets_list = [ + load_label_offset(p, inds, tot) for p in label_paths + ] + assert ( + label_processors is None + or len(label_processors) == self.num_labels + ) + for label_path, label_rate in zip(label_paths, self.label_rates): + verify_label_lengths( + self.sizes, sample_rate, label_path, label_rate, inds, tot + ) + + self.max_sample_size = ( + max_sample_size if max_sample_size is not None else sys.maxsize + ) + self.pad_audio = pad_audio + self.normalize = normalize + logger.info( + f"pad_audio={pad_audio}, random_crop={random_crop}, " + f"normalize={normalize}, max_sample_size={self.max_sample_size}" + ) + + def get_audio(self, index): + import soundfile as sf + + wav_path = os.path.join(self.audio_root, self.audio_names[index]) + wav, cur_sample_rate = sf.read(wav_path) + wav = torch.from_numpy(wav).float() + wav = self.postprocess(wav, cur_sample_rate) + return wav + + def get_label(self, index, label_idx): + if self.store_labels: + label = self.label_list[label_idx][index] + else: + with open(self.label_paths[label_idx]) as f: + offset_s, offset_e = self.label_offsets_list[label_idx][index] + f.seek(offset_s) + label = f.read(offset_e - offset_s) + + if self.label_processors is not None: + label = self.label_processors[label_idx](label) + return label + + def get_labels(self, index): + return [self.get_label(index, i) for i in range(self.num_labels)] + + def __getitem__(self, index): + wav = self.get_audio(index) + labels = self.get_labels(index) + return {"id": index, "source": wav, "label_list": labels} + + def __len__(self): + return len(self.sizes) + + def crop_to_max_size(self, wav, target_size): + size = len(wav) + diff = size - target_size + if diff <= 0: + return wav, 0 + + start, end = 0, target_size + if self.random_crop: + start = np.random.randint(0, diff + 1) + end = size - diff + start + return wav[start:end], start + + def collater(self, samples): + # target = max(sizes) -> random_crop not used + # target = max_sample_size -> random_crop used for long + samples = [s for s in samples if s["source"] is not None] + if len(samples) == 0: + return {} + + audios = [s["source"] for s in samples] + audio_sizes = [len(s) for s in audios] + if self.pad_audio: + audio_size = min(max(audio_sizes), self.max_sample_size) + else: + audio_size = min(min(audio_sizes), self.max_sample_size) + collated_audios, padding_mask, audio_starts = self.collater_audio( + audios, audio_size + ) + + targets_by_label = [ + [s["label_list"][i] for s in samples] + for i in range(self.num_labels) + ] + targets_list, lengths_list, ntokens_list = self.collater_label( + targets_by_label, audio_size, audio_starts + ) + + net_input = {"source": collated_audios, "padding_mask": padding_mask} + batch = { + "id": torch.LongTensor([s["id"] for s in samples]), + "net_input": net_input, + } + + if self.single_target: + batch["target_lengths"] = lengths_list[0] + batch["ntokens"] = ntokens_list[0] + batch["target"] = targets_list[0] + else: + batch["target_lengths_list"] = lengths_list + batch["ntokens_list"] = ntokens_list + batch["target_list"] = targets_list + return batch + + def collater_audio(self, audios, audio_size): + collated_audios = audios[0].new_zeros(len(audios), audio_size) + padding_mask = ( + torch.BoolTensor(collated_audios.shape).fill_(False) + # if self.pad_audio else None + ) + audio_starts = [0 for _ in audios] + for i, audio in enumerate(audios): + diff = len(audio) - audio_size + if diff == 0: + collated_audios[i] = audio + elif diff < 0: + assert self.pad_audio + collated_audios[i] = torch.cat( + [audio, audio.new_full((-diff,), 0.0)] + ) + padding_mask[i, diff:] = True + else: + collated_audios[i], audio_starts[i] = self.crop_to_max_size( + audio, audio_size + ) + return collated_audios, padding_mask, audio_starts + + def collater_frm_label( + self, targets, audio_size, audio_starts, label_rate, pad + ): + assert label_rate > 0 + s2f = label_rate / self.sample_rate + frm_starts = [int(round(s * s2f)) for s in audio_starts] + frm_size = int(round(audio_size * s2f)) + if not self.pad_audio: + rem_size = [len(t) - s for t, s in zip(targets, frm_starts)] + frm_size = min(frm_size, *rem_size) + targets = [t[s: s + frm_size] for t, s in zip(targets, frm_starts)] + logger.debug(f"audio_starts={audio_starts}") + logger.debug(f"frame_starts={frm_starts}") + logger.debug(f"frame_size={frm_size}") + + lengths = torch.LongTensor([len(t) for t in targets]) + ntokens = lengths.sum().item() + targets = data_utils.collate_tokens( + targets, pad_idx=pad, left_pad=False + ) + return targets, lengths, ntokens + + def collater_seq_label(self, targets, pad): + lengths = torch.LongTensor([len(t) for t in targets]) + ntokens = lengths.sum().item() + targets = data_utils.collate_tokens( + targets, pad_idx=pad, left_pad=False + ) + return targets, lengths, ntokens + + def collater_label(self, targets_by_label, audio_size, audio_starts): + targets_list, lengths_list, ntokens_list = [], [], [] + itr = zip(targets_by_label, self.label_rates, self.pad_list) + for targets, label_rate, pad in itr: + if label_rate == -1: + targets, lengths, ntokens = self.collater_seq_label( + targets, pad + ) + else: + targets, lengths, ntokens = self.collater_frm_label( + targets, audio_size, audio_starts, label_rate, pad + ) + targets_list.append(targets) + lengths_list.append(lengths) + ntokens_list.append(ntokens) + return targets_list, lengths_list, ntokens_list + + def num_tokens(self, index): + return self.size(index) + + def size(self, index): + if self.pad_audio: + return self.sizes[index] + return min(self.sizes[index], self.max_sample_size) + + def ordered_indices(self): + if self.shuffle: + order = [np.random.permutation(len(self))] + else: + order = [np.arange(len(self))] + + order.append(self.sizes) + return np.lexsort(order)[::-1] + + def postprocess(self, wav, cur_sample_rate): + if wav.dim() == 2: + wav = wav.mean(-1) + assert wav.dim() == 1, wav.dim() + + if cur_sample_rate != self.sample_rate: + raise Exception(f"sr {cur_sample_rate} != {self.sample_rate}") + + if self.normalize: + with torch.no_grad(): + wav = F.layer_norm(wav, wav.shape) + return wav diff --git a/fairseq/data/audio/raw_audio_dataset.py b/fairseq/data/audio/raw_audio_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..9ce3f7e39d55860f38b3332fe79917c8d38724fe --- /dev/null +++ b/fairseq/data/audio/raw_audio_dataset.py @@ -0,0 +1,386 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +import logging +import os +import sys +import io + +import numpy as np +import torch +import torch.nn.functional as F + +from .. import FairseqDataset +from ..data_utils import compute_mask_indices, get_buckets, get_bucketed_sizes +from fairseq.data.audio.audio_utils import ( + parse_path, + read_from_stored_zip, + is_sf_audio_data, +) + + +logger = logging.getLogger(__name__) + + +class RawAudioDataset(FairseqDataset): + def __init__( + self, + sample_rate, + max_sample_size=None, + min_sample_size=0, + shuffle=True, + pad=False, + normalize=False, + compute_mask_indices=False, + **mask_compute_kwargs, + ): + super().__init__() + + self.sample_rate = sample_rate + self.sizes = [] + self.max_sample_size = ( + max_sample_size if max_sample_size is not None else sys.maxsize + ) + self.min_sample_size = min_sample_size + self.pad = pad + self.shuffle = shuffle + self.normalize = normalize + self.compute_mask_indices = compute_mask_indices + if self.compute_mask_indices: + self.mask_compute_kwargs = mask_compute_kwargs + self._features_size_map = {} + self._C = mask_compute_kwargs["encoder_embed_dim"] + self._conv_feature_layers = eval(mask_compute_kwargs["conv_feature_layers"]) + + def __getitem__(self, index): + raise NotImplementedError() + + def __len__(self): + return len(self.sizes) + + def postprocess(self, feats, curr_sample_rate): + if feats.dim() == 2: + feats = feats.mean(-1) + + if curr_sample_rate != self.sample_rate: + raise Exception(f"sample rate: {curr_sample_rate}, need {self.sample_rate}") + + assert feats.dim() == 1, feats.dim() + + if self.normalize: + with torch.no_grad(): + feats = F.layer_norm(feats, feats.shape) + return feats + + def crop_to_max_size(self, wav, target_size): + size = len(wav) + diff = size - target_size + if diff <= 0: + return wav + + start = np.random.randint(0, diff + 1) + end = size - diff + start + return wav[start:end] + + def _compute_mask_indices(self, dims, padding_mask): + B, T, C = dims + mask_indices, mask_channel_indices = None, None + if self.mask_compute_kwargs["mask_prob"] > 0: + mask_indices = compute_mask_indices( + (B, T), + padding_mask, + self.mask_compute_kwargs["mask_prob"], + self.mask_compute_kwargs["mask_length"], + self.mask_compute_kwargs["mask_selection"], + self.mask_compute_kwargs["mask_other"], + min_masks=2, + no_overlap=self.mask_compute_kwargs["no_mask_overlap"], + min_space=self.mask_compute_kwargs["mask_min_space"], + ) + mask_indices = torch.from_numpy(mask_indices) + if self.mask_compute_kwargs["mask_channel_prob"] > 0: + mask_channel_indices = compute_mask_indices( + (B, C), + None, + self.mask_compute_kwargs["mask_channel_prob"], + self.mask_compute_kwargs["mask_channel_length"], + self.mask_compute_kwargs["mask_channel_selection"], + self.mask_compute_kwargs["mask_channel_other"], + no_overlap=self.mask_compute_kwargs["no_mask_channel_overlap"], + min_space=self.mask_compute_kwargs["mask_channel_min_space"], + ) + mask_channel_indices = ( + torch.from_numpy(mask_channel_indices).unsqueeze(1).expand(-1, T, -1) + ) + + return mask_indices, mask_channel_indices + + @staticmethod + def _bucket_tensor(tensor, num_pad, value): + return F.pad(tensor, (0, num_pad), value=value) + + def collater(self, samples): + samples = [s for s in samples if s["source"] is not None] + if len(samples) == 0: + return {} + + sources = [s["source"] for s in samples] + sizes = [len(s) for s in sources] + + if self.pad: + target_size = min(max(sizes), self.max_sample_size) + else: + target_size = min(min(sizes), self.max_sample_size) + + collated_sources = sources[0].new_zeros(len(sources), target_size) + padding_mask = ( + torch.BoolTensor(collated_sources.shape).fill_(False) if self.pad else None + ) + for i, (source, size) in enumerate(zip(sources, sizes)): + diff = size - target_size + if diff == 0: + collated_sources[i] = source + elif diff < 0: + assert self.pad + collated_sources[i] = torch.cat( + [source, source.new_full((-diff,), 0.0)] + ) + padding_mask[i, diff:] = True + else: + collated_sources[i] = self.crop_to_max_size(source, target_size) + + input = {"source": collated_sources} + out = {"id": torch.LongTensor([s["id"] for s in samples])} + if self.pad: + input["padding_mask"] = padding_mask + + if hasattr(self, "num_buckets") and self.num_buckets > 0: + assert self.pad, "Cannot bucket without padding first." + bucket = max(self._bucketed_sizes[s["id"]] for s in samples) + num_pad = bucket - collated_sources.size(-1) + if num_pad: + input["source"] = self._bucket_tensor(collated_sources, num_pad, 0) + input["padding_mask"] = self._bucket_tensor(padding_mask, num_pad, True) + + if self.compute_mask_indices: + B = input["source"].size(0) + T = self._get_mask_indices_dims(input["source"].size(-1)) + padding_mask_reshaped = input["padding_mask"].clone() + extra = padding_mask_reshaped.size(1) % T + if extra > 0: + padding_mask_reshaped = padding_mask_reshaped[:, :-extra] + padding_mask_reshaped = padding_mask_reshaped.view( + padding_mask_reshaped.size(0), T, -1 + ) + padding_mask_reshaped = padding_mask_reshaped.all(-1) + input["padding_count"] = padding_mask_reshaped.sum(-1).max().item() + mask_indices, mask_channel_indices = self._compute_mask_indices( + (B, T, self._C), + padding_mask_reshaped, + ) + input["mask_indices"] = mask_indices + input["mask_channel_indices"] = mask_channel_indices + out["sample_size"] = mask_indices.sum().item() + + out["net_input"] = input + return out + + def _get_mask_indices_dims(self, size, padding=0, dilation=1): + if size not in self._features_size_map: + L_in = size + for (_, kernel_size, stride) in self._conv_feature_layers: + L_out = L_in + 2 * padding - dilation * (kernel_size - 1) - 1 + L_out = 1 + L_out // stride + L_in = L_out + self._features_size_map[size] = L_out + return self._features_size_map[size] + + def num_tokens(self, index): + return self.size(index) + + def size(self, index): + """Return an example's size as a float or tuple. This value is used when + filtering a dataset with ``--max-positions``.""" + if self.pad: + return self.sizes[index] + return min(self.sizes[index], self.max_sample_size) + + def ordered_indices(self): + """Return an ordered list of indices. Batches will be constructed based + on this order.""" + + if self.shuffle: + order = [np.random.permutation(len(self))] + order.append( + np.minimum( + np.array(self.sizes), + self.max_sample_size, + ) + ) + return np.lexsort(order)[::-1] + else: + return np.arange(len(self)) + + def set_bucket_info(self, num_buckets): + self.num_buckets = num_buckets + if self.num_buckets > 0: + self._collated_sizes = np.minimum( + np.array(self.sizes), + self.max_sample_size, + ) + self.buckets = get_buckets( + self._collated_sizes, + self.num_buckets, + ) + self._bucketed_sizes = get_bucketed_sizes( + self._collated_sizes, self.buckets + ) + logger.info( + f"{len(self.buckets)} bucket(s) for the audio dataset: " + f"{self.buckets}" + ) + + +class FileAudioDataset(RawAudioDataset): + def __init__( + self, + manifest_path, + sample_rate, + max_sample_size=None, + min_sample_size=0, + shuffle=True, + pad=False, + normalize=False, + num_buckets=0, + compute_mask_indices=False, + **mask_compute_kwargs, + ): + super().__init__( + sample_rate=sample_rate, + max_sample_size=max_sample_size, + min_sample_size=min_sample_size, + shuffle=shuffle, + pad=pad, + normalize=normalize, + compute_mask_indices=compute_mask_indices, + **mask_compute_kwargs, + ) + + skipped = 0 + self.fnames = [] + sizes = [] + self.skipped_indices = set() + + with open(manifest_path, "r") as f: + self.root_dir = f.readline().strip() + for i, line in enumerate(f): + items = line.strip().split("\t") + assert len(items) == 2, line + sz = int(items[1]) + if min_sample_size is not None and sz < min_sample_size: + skipped += 1 + self.skipped_indices.add(i) + continue + self.fnames.append(items[0]) + sizes.append(sz) + logger.info(f"loaded {len(self.fnames)}, skipped {skipped} samples") + + self.sizes = np.array(sizes, dtype=np.int64) + + try: + import pyarrow + + self.fnames = pyarrow.array(self.fnames) + except: + logger.debug( + "Could not create a pyarrow array. Please install pyarrow for better performance" + ) + pass + + self.set_bucket_info(num_buckets) + + def __getitem__(self, index): + import soundfile as sf + + path_or_fp = os.path.join(self.root_dir, str(self.fnames[index])) + _path, slice_ptr = parse_path(path_or_fp) + if len(slice_ptr) == 2: + byte_data = read_from_stored_zip(_path, slice_ptr[0], slice_ptr[1]) + assert is_sf_audio_data(byte_data) + path_or_fp = io.BytesIO(byte_data) + + wav, curr_sample_rate = sf.read(path_or_fp, dtype="float32") + + feats = torch.from_numpy(wav).float() + feats = self.postprocess(feats, curr_sample_rate) + return {"id": index, "source": feats} + + +class BinarizedAudioDataset(RawAudioDataset): + def __init__( + self, + data_dir, + split, + sample_rate, + max_sample_size=None, + min_sample_size=0, + shuffle=True, + pad=False, + normalize=False, + num_buckets=0, + compute_mask_indices=False, + **mask_compute_kwargs, + ): + super().__init__( + sample_rate=sample_rate, + max_sample_size=max_sample_size, + min_sample_size=min_sample_size, + shuffle=shuffle, + pad=pad, + normalize=normalize, + compute_mask_indices=compute_mask_indices, + **mask_compute_kwargs, + ) + + from fairseq.data import data_utils, Dictionary + + self.fnames_dict = Dictionary.load(os.path.join(data_dir, "dict.txt")) + + root_path = os.path.join(data_dir, f"{split}.root") + if os.path.exists(root_path): + with open(root_path, "r") as f: + self.root_dir = next(f).strip() + else: + self.root_dir = None + + fnames_path = os.path.join(data_dir, split) + self.fnames = data_utils.load_indexed_dataset(fnames_path, self.fnames_dict) + lengths_path = os.path.join(data_dir, f"{split}.lengths") + + with open(lengths_path, "r") as f: + for line in f: + sz = int(line.rstrip()) + assert ( + sz >= min_sample_size + ), f"Min sample size is not supported for binarized dataset, but found a sample with size {sz}" + self.sizes.append(sz) + + self.sizes = np.array(self.sizes, dtype=np.int64) + + self.set_bucket_info(num_buckets) + logger.info(f"loaded {len(self.fnames)} samples") + + def __getitem__(self, index): + import soundfile as sf + + fname = self.fnames_dict.string(self.fnames[index], separator="") + if self.root_dir: + fname = os.path.join(self.root_dir, fname) + + wav, curr_sample_rate = sf.read(fname) + feats = torch.from_numpy(wav).float() + feats = self.postprocess(feats, curr_sample_rate) + return {"id": index, "source": feats} diff --git a/fairseq/data/audio/speech_to_text_dataset.py b/fairseq/data/audio/speech_to_text_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d4b5668d8f9d4bc93fcbda73d867554d8f1b3107 --- /dev/null +++ b/fairseq/data/audio/speech_to_text_dataset.py @@ -0,0 +1,511 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import csv +import io +import logging +import os.path as op +import re +from typing import Dict, List, Optional, Tuple + +import numpy as np +import torch +from fairseq.data import ( + ConcatDataset, + Dictionary, + FairseqDataset, + ResamplingDataset, + data_utils as fairseq_data_utils, +) +from fairseq.data.audio.audio_utils import ( + get_fbank, get_waveform, read_from_stored_zip, is_npy_data, + is_sf_audio_data, parse_path, FEATURE_OR_SF_AUDIO_FILE_EXTENSIONS +) +from fairseq.data.audio.feature_transforms import CompositeAudioFeatureTransform + + +logger = logging.getLogger(__name__) + + +class S2TDataConfig(object): + """Wrapper class for data config YAML""" + + def __init__(self, yaml_path): + try: + import yaml + except ImportError: + print("Please install PyYAML to load YAML files for " "S2T data config") + self.config = {} + if op.isfile(yaml_path): + try: + with open(yaml_path) as f: + self.config = yaml.load(f, Loader=yaml.FullLoader) + except Exception as e: + raise Exception(f"Failed to load config from {yaml_path}: {e}") + else: + raise FileNotFoundError(f"{yaml_path} not found") + + @property + def vocab_filename(self): + """fairseq vocabulary file under data root""" + return self.config.get("vocab_filename", "dict.txt") + + @property + def shuffle(self) -> bool: + """Shuffle dataset samples before batching""" + return self.config.get("shuffle", False) + + @property + def pre_tokenizer(self) -> Dict: + """Pre-tokenizer to apply before subword tokenization. Returning + a dictionary with `tokenizer` providing the tokenizer name and + the other items providing the tokenizer-specific arguments. + Tokenizers are defined in `fairseq.data.encoders.*`""" + return self.config.get("pre_tokenizer", {"tokenizer": None}) + + @property + def bpe_tokenizer(self) -> Dict: + """Subword tokenizer to apply after pre-tokenization. Returning + a dictionary with `bpe` providing the tokenizer name and + the other items providing the tokenizer-specific arguments. + Tokenizers are defined in `fairseq.data.encoders.*`""" + return self.config.get("bpe_tokenizer", {"bpe": None}) + + @property + def prepend_tgt_lang_tag(self) -> bool: + """Prepend target lang ID token as the target BOS (e.g. for to-many + multilingual setting). During inference, this requires `--prefix-size 1` + to force BOS to be lang ID token.""" + return self.config.get("prepend_tgt_lang_tag", False) + + @property + def input_feat_per_channel(self): + """The dimension of input features (per audio channel)""" + return self.config.get("input_feat_per_channel", 80) + + @property + def input_channels(self): + """The number of channels in the input audio""" + return self.config.get("input_channels", 1) + + @property + def sampling_alpha(self): + """Hyper-parameter alpha = 1/T for temperature-based resampling. + (alpha = 1 for no resampling)""" + return self.config.get("sampling_alpha", 1.0) + + @property + def use_audio_input(self): + """Needed by the dataset loader to see if the model requires + raw audio as inputs.""" + return self.config.get("use_audio_input", False) + + @property + def audio_root(self): + """Audio paths in the manifest TSV can be relative and this provides + the root path. Set this to empty string when using absolute paths.""" + return self.config.get("audio_root", "") + + def get_feature_transforms(self, split, is_train): + """Split-specific feature transforms. Allowing train set wildcard `_train`, + evaluation set wildcard `_eval` and general wildcard `*` for matching.""" + from copy import deepcopy + + cfg = deepcopy(self.config) + _cur = cfg.get("transforms", {}) + cur = _cur.get(split) + cur = _cur.get("_train") if cur is None and is_train else cur + cur = _cur.get("_eval") if cur is None and not is_train else cur + cur = _cur.get("*") if cur is None else cur + cfg["transforms"] = cur + return cfg + + +def get_features_from_npy_or_audio(path): + ext = op.splitext(op.basename(path))[1] + if ext not in FEATURE_OR_SF_AUDIO_FILE_EXTENSIONS: + raise ValueError(f'Unsupported file format for "{path}"') + return np.load(path) if ext == ".npy" else get_fbank(path) + + +def get_features_or_waveform_from_stored_zip( + path, byte_offset, byte_size, need_waveform=False +): + assert path.endswith(".zip") + data = read_from_stored_zip(path, byte_offset, byte_size) + f = io.BytesIO(data) + if is_npy_data(data): + features_or_waveform = np.load(f) + elif is_sf_audio_data(data): + features_or_waveform = \ + get_waveform(f, always_2d=False)[0] if need_waveform else get_fbank(f) + else: + raise ValueError(f'Unknown file format for "{path}"') + return features_or_waveform + + +def get_features_or_waveform(path: str, need_waveform=False): + """Get speech features from .npy file or waveform from .wav/.flac file. + The file may be inside an uncompressed ZIP file and is accessed via byte + offset and length. + + Args: + path (str): File path in the format of "<.npy/.wav/.flac path>" or + "<zip path>:<byte offset>:<byte length>". + need_waveform (bool): return waveform instead of features. + + Returns: + features_or_waveform (numpy.ndarray): speech features or waveform. + """ + _path, slice_ptr = parse_path(path) + if len(slice_ptr) == 0: + if need_waveform: + return get_waveform(_path, always_2d=False) + return get_features_from_npy_or_audio(_path) + elif len(slice_ptr) == 2: + features_or_waveform = get_features_or_waveform_from_stored_zip( + _path, slice_ptr[0], slice_ptr[1], need_waveform=need_waveform + ) + else: + raise ValueError(f"Invalid path: {path}") + + return features_or_waveform + + +def _collate_frames( + frames: List[torch.Tensor], is_audio_input: bool = False +) -> torch.Tensor: + """ + Convert a list of 2D frames into a padded 3D tensor + Args: + frames (list): list of 2D frames of size L[i]*f_dim. Where L[i] is + length of i-th frame and f_dim is static dimension of features + Returns: + 3D tensor of size len(frames)*len_max*f_dim where len_max is max of L[i] + """ + max_len = max(frame.size(0) for frame in frames) + if is_audio_input: + out = frames[0].new_zeros((len(frames), max_len)) + else: + out = frames[0].new_zeros((len(frames), max_len, frames[0].size(1))) + for i, v in enumerate(frames): + out[i, : v.size(0)] = v + return out + + +class SpeechToTextDataset(FairseqDataset): + LANG_TAG_TEMPLATE = "<lang:{}>" + + def __init__( + self, + split: str, + is_train_split: bool, + data_cfg: S2TDataConfig, + audio_paths: List[str], + n_frames: List[int], + src_texts: Optional[List[str]] = None, + tgt_texts: Optional[List[str]] = None, + speakers: Optional[List[str]] = None, + src_langs: Optional[List[str]] = None, + tgt_langs: Optional[List[str]] = None, + ids: Optional[List[str]] = None, + tgt_dict: Optional[Dictionary] = None, + pre_tokenizer=None, + bpe_tokenizer=None, + ): + self.split, self.is_train_split = split, is_train_split + self.data_cfg = data_cfg + self.audio_paths, self.n_frames = audio_paths, n_frames + self.n_samples = len(audio_paths) + assert len(n_frames) == self.n_samples > 0 + assert src_texts is None or len(src_texts) == self.n_samples + assert tgt_texts is None or len(tgt_texts) == self.n_samples + assert speakers is None or len(speakers) == self.n_samples + assert src_langs is None or len(src_langs) == self.n_samples + assert tgt_langs is None or len(tgt_langs) == self.n_samples + assert ids is None or len(ids) == self.n_samples + assert (tgt_dict is None and tgt_texts is None) or ( + tgt_dict is not None and tgt_texts is not None + ) + self.src_texts, self.tgt_texts = src_texts, tgt_texts + self.src_langs, self.tgt_langs = src_langs, tgt_langs + self.tgt_dict = tgt_dict + self.check_tgt_lang_tag() + self.ids = ids + self.shuffle = data_cfg.shuffle if is_train_split else False + + self.feature_transforms = CompositeAudioFeatureTransform.from_config_dict( + self.data_cfg.get_feature_transforms(split, is_train_split) + ) + + self.pre_tokenizer = pre_tokenizer + self.bpe_tokenizer = bpe_tokenizer + + logger.info(self.__repr__()) + + def __repr__(self): + return ( + self.__class__.__name__ + + f'(split="{self.split}", n_samples={self.n_samples}, ' + f"prepend_tgt_lang_tag={self.data_cfg.prepend_tgt_lang_tag}, " + f"shuffle={self.shuffle}, transforms={self.feature_transforms})" + ) + + @classmethod + def is_lang_tag(cls, token): + pattern = cls.LANG_TAG_TEMPLATE.replace("{}", "(.*)") + return re.match(pattern, token) + + def check_tgt_lang_tag(self): + if self.data_cfg.prepend_tgt_lang_tag: + assert self.tgt_langs is not None and self.tgt_dict is not None + tgt_lang_tags = [ + self.LANG_TAG_TEMPLATE.format(t) for t in set(self.tgt_langs) + ] + assert all(t in self.tgt_dict for t in tgt_lang_tags) + + def tokenize_text(self, text: str): + if self.pre_tokenizer is not None: + text = self.pre_tokenizer.encode(text) + if self.bpe_tokenizer is not None: + text = self.bpe_tokenizer.encode(text) + return text + + def __getitem__( + self, index: int + ) -> Tuple[int, torch.Tensor, Optional[torch.Tensor]]: + source = get_features_or_waveform( + self.audio_paths[index], need_waveform=self.data_cfg.use_audio_input + ) + if self.feature_transforms is not None: + assert not self.data_cfg.use_audio_input + source = self.feature_transforms(source) + source = torch.from_numpy(source).float() + + target = None + if self.tgt_texts is not None: + tokenized = self.tokenize_text(self.tgt_texts[index]) + target = self.tgt_dict.encode_line( + tokenized, add_if_not_exist=False, append_eos=True + ).long() + if self.data_cfg.prepend_tgt_lang_tag: + lang_tag = self.LANG_TAG_TEMPLATE.format(self.tgt_langs[index]) + lang_tag_idx = self.tgt_dict.index(lang_tag) + target = torch.cat((torch.LongTensor([lang_tag_idx]), target), 0) + return index, source, target + + def __len__(self): + return self.n_samples + + def collater(self, samples: List[Tuple[int, torch.Tensor, torch.Tensor]]) -> Dict: + if len(samples) == 0: + return {} + indices = torch.tensor([i for i, _, _ in samples], dtype=torch.long) + frames = _collate_frames( + [s for _, s, _ in samples], self.data_cfg.use_audio_input + ) + # sort samples by descending number of frames + n_frames = torch.tensor([s.size(0) for _, s, _ in samples], dtype=torch.long) + n_frames, order = n_frames.sort(descending=True) + indices = indices.index_select(0, order) + frames = frames.index_select(0, order) + + target, target_lengths = None, None + prev_output_tokens = None + ntokens = None + if self.tgt_texts is not None: + target = fairseq_data_utils.collate_tokens( + [t for _, _, t in samples], + self.tgt_dict.pad(), + self.tgt_dict.eos(), + left_pad=False, + move_eos_to_beginning=False, + ) + target = target.index_select(0, order) + target_lengths = torch.tensor( + [t.size(0) for _, _, t in samples], dtype=torch.long + ).index_select(0, order) + prev_output_tokens = fairseq_data_utils.collate_tokens( + [t for _, _, t in samples], + self.tgt_dict.pad(), + self.tgt_dict.eos(), + left_pad=False, + move_eos_to_beginning=True, + ) + prev_output_tokens = prev_output_tokens.index_select(0, order) + ntokens = sum(t.size(0) for _, _, t in samples) + + out = { + "id": indices, + "net_input": { + "src_tokens": frames, + "src_lengths": n_frames, + "prev_output_tokens": prev_output_tokens, + }, + "target": target, + "target_lengths": target_lengths, + "ntokens": ntokens, + "nsentences": len(samples), + } + return out + + def num_tokens(self, index): + return self.n_frames[index] + + def size(self, index): + t_len = 0 + if self.tgt_texts is not None: + tokenized = self.tokenize_text(self.tgt_texts[index]) + t_len = len(tokenized.split(" ")) + return self.n_frames[index], t_len + + @property + def sizes(self): + return np.array(self.n_frames) + + @property + def can_reuse_epoch_itr_across_epochs(self): + return True + + def ordered_indices(self): + if self.shuffle: + order = [np.random.permutation(len(self))] + else: + order = [np.arange(len(self))] + # first by descending order of # of frames then by original/random order + order.append([-n for n in self.n_frames]) + return np.lexsort(order) + + def prefetch(self, indices): + raise False + + +class SpeechToTextDatasetCreator(object): + # mandatory columns + KEY_ID, KEY_AUDIO, KEY_N_FRAMES = "id", "audio", "n_frames" + KEY_TGT_TEXT = "tgt_text" + # optional columns + KEY_SPEAKER, KEY_SRC_TEXT = "speaker", "src_text" + KEY_SRC_LANG, KEY_TGT_LANG = "src_lang", "tgt_lang" + # default values + DEFAULT_SPEAKER = DEFAULT_SRC_TEXT = DEFAULT_LANG = "" + + @classmethod + def _from_list( + cls, + split_name: str, + is_train_split, + samples: List[List[Dict]], + data_cfg: S2TDataConfig, + tgt_dict, + pre_tokenizer, + bpe_tokenizer, + ) -> SpeechToTextDataset: + audio_paths, n_frames, src_texts, tgt_texts, ids = [], [], [], [], [] + speakers, src_langs, tgt_langs = [], [], [] + for s in samples: + ids.extend([ss[cls.KEY_ID] for ss in s]) + audio_paths.extend( + [op.join(data_cfg.audio_root, ss[cls.KEY_AUDIO]) for ss in s] + ) + n_frames.extend([int(ss[cls.KEY_N_FRAMES]) for ss in s]) + tgt_texts.extend([ss[cls.KEY_TGT_TEXT] for ss in s]) + src_texts.extend( + [ss.get(cls.KEY_SRC_TEXT, cls.DEFAULT_SRC_TEXT) for ss in s] + ) + speakers.extend([ss.get(cls.KEY_SPEAKER, cls.DEFAULT_SPEAKER) for ss in s]) + src_langs.extend([ss.get(cls.KEY_SRC_LANG, cls.DEFAULT_LANG) for ss in s]) + tgt_langs.extend([ss.get(cls.KEY_TGT_LANG, cls.DEFAULT_LANG) for ss in s]) + return SpeechToTextDataset( + split_name, + is_train_split, + data_cfg, + audio_paths, + n_frames, + src_texts, + tgt_texts, + speakers, + src_langs, + tgt_langs, + ids, + tgt_dict, + pre_tokenizer, + bpe_tokenizer, + ) + + @classmethod + def _get_size_ratios(cls, ids: List[str], sizes: List[int], alpha: float = 1.0): + """Size ratios for temperature-based sampling + (https://arxiv.org/abs/1907.05019)""" + _sizes = np.array(sizes) + prob = _sizes / _sizes.sum() + smoothed_prob = prob ** alpha + smoothed_prob = smoothed_prob / smoothed_prob.sum() + size_ratio = (smoothed_prob * _sizes.sum()) / _sizes + + o_str = str({_i: f"{prob[i]:.3f}" for i, _i in enumerate(ids)}) + logger.info(f"original sampling probability: {o_str}") + p_str = str({_i: f"{smoothed_prob[i]:.3f}" for i, _i in enumerate(ids)}) + logger.info(f"balanced sampling probability: {p_str}") + sr_str = str({_id: f"{size_ratio[i]:.3f}" for i, _id in enumerate(ids)}) + logger.info(f"balanced sampling size ratio: {sr_str}") + return size_ratio.tolist() + + @classmethod + def from_tsv( + cls, + root: str, + data_cfg: S2TDataConfig, + splits: str, + tgt_dict, + pre_tokenizer, + bpe_tokenizer, + is_train_split: bool, + epoch: int, + seed: int, + ) -> SpeechToTextDataset: + samples = [] + _splits = splits.split(",") + for split in _splits: + tsv_path = op.join(root, f"{split}.tsv") + if not op.isfile(tsv_path): + raise FileNotFoundError(f"Dataset not found: {tsv_path}") + with open(tsv_path) as f: + reader = csv.DictReader( + f, + delimiter="\t", + quotechar=None, + doublequote=False, + lineterminator="\n", + quoting=csv.QUOTE_NONE, + ) + samples.append([dict(e) for e in reader]) + assert len(samples) > 0 + + datasets = [ + cls._from_list( + name, + is_train_split, + [s], + data_cfg, + tgt_dict, + pre_tokenizer, + bpe_tokenizer, + ) + for name, s in zip(_splits, samples) + ] + + if is_train_split and len(_splits) > 1 and data_cfg.sampling_alpha != 1.0: + # temperature-based sampling + size_ratios = cls._get_size_ratios( + _splits, [len(s) for s in samples], alpha=data_cfg.sampling_alpha + ) + datasets = [ + ResamplingDataset( + d, size_ratio=r, seed=seed, epoch=epoch, replace=(r >= 1.0) + ) + for d, r in zip(datasets, size_ratios) + ] + return ConcatDataset(datasets) diff --git a/fairseq/data/backtranslation_dataset.py b/fairseq/data/backtranslation_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..8f70c90df3d237077537993e125d366c95292f1a --- /dev/null +++ b/fairseq/data/backtranslation_dataset.py @@ -0,0 +1,165 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +from fairseq import utils + +from . import FairseqDataset + + +def backtranslate_samples(samples, collate_fn, generate_fn, cuda=True): + """Backtranslate a list of samples. + + Given an input (*samples*) of the form: + + [{'id': 1, 'source': 'hallo welt'}] + + this will return: + + [{'id': 1, 'source': 'hello world', 'target': 'hallo welt'}] + + Args: + samples (List[dict]): samples to backtranslate. Individual samples are + expected to have a 'source' key, which will become the 'target' + after backtranslation. + collate_fn (callable): function to collate samples into a mini-batch + generate_fn (callable): function to generate backtranslations + cuda (bool): use GPU for generation (default: ``True``) + + Returns: + List[dict]: an updated list of samples with a backtranslated source + """ + collated_samples = collate_fn(samples) + s = utils.move_to_cuda(collated_samples) if cuda else collated_samples + generated_sources = generate_fn(s) + + id_to_src = {sample["id"]: sample["source"] for sample in samples} + + # Go through each tgt sentence in batch and its corresponding best + # generated hypothesis and create a backtranslation data pair + # {id: id, source: generated backtranslation, target: original tgt} + return [ + { + "id": id.item(), + "target": id_to_src[id.item()], + "source": hypos[0]["tokens"].cpu(), + } + for id, hypos in zip(collated_samples["id"], generated_sources) + ] + + +class BacktranslationDataset(FairseqDataset): + """ + Sets up a backtranslation dataset which takes a tgt batch, generates + a src using a tgt-src backtranslation function (*backtranslation_fn*), + and returns the corresponding `{generated src, input tgt}` batch. + + Args: + tgt_dataset (~fairseq.data.FairseqDataset): the dataset to be + backtranslated. Only the source side of this dataset will be used. + After backtranslation, the source sentences in this dataset will be + returned as the targets. + src_dict (~fairseq.data.Dictionary): the dictionary of backtranslated + sentences. + tgt_dict (~fairseq.data.Dictionary, optional): the dictionary of + sentences to be backtranslated. + backtranslation_fn (callable, optional): function to call to generate + backtranslations. This is typically the `generate` method of a + :class:`~fairseq.sequence_generator.SequenceGenerator` object. + Pass in None when it is not available at initialization time, and + use set_backtranslation_fn function to set it when available. + output_collater (callable, optional): function to call on the + backtranslated samples to create the final batch + (default: ``tgt_dataset.collater``). + cuda: use GPU for generation + """ + + def __init__( + self, + tgt_dataset, + src_dict, + tgt_dict=None, + backtranslation_fn=None, + output_collater=None, + cuda=True, + **kwargs + ): + self.tgt_dataset = tgt_dataset + self.backtranslation_fn = backtranslation_fn + self.output_collater = ( + output_collater if output_collater is not None else tgt_dataset.collater + ) + self.cuda = cuda if torch.cuda.is_available() else False + self.src_dict = src_dict + self.tgt_dict = tgt_dict + + def __getitem__(self, index): + """ + Returns a single sample from *tgt_dataset*. Note that backtranslation is + not applied in this step; use :func:`collater` instead to backtranslate + a batch of samples. + """ + return self.tgt_dataset[index] + + def __len__(self): + return len(self.tgt_dataset) + + def set_backtranslation_fn(self, backtranslation_fn): + self.backtranslation_fn = backtranslation_fn + + def collater(self, samples): + """Merge and backtranslate a list of samples to form a mini-batch. + + Using the samples from *tgt_dataset*, load a collated target sample to + feed to the backtranslation model. Then take the backtranslation with + the best score as the source and the original input as the target. + + Note: we expect *tgt_dataset* to provide a function `collater()` that + will collate samples into the format expected by *backtranslation_fn*. + After backtranslation, we will feed the new list of samples (i.e., the + `(backtranslated source, original source)` pairs) to *output_collater* + and return the result. + + Args: + samples (List[dict]): samples to backtranslate and collate + + Returns: + dict: a mini-batch with keys coming from *output_collater* + """ + if samples[0].get("is_dummy", False): + return samples + samples = backtranslate_samples( + samples=samples, + collate_fn=self.tgt_dataset.collater, + generate_fn=(lambda net_input: self.backtranslation_fn(net_input)), + cuda=self.cuda, + ) + return self.output_collater(samples) + + def num_tokens(self, index): + """Just use the tgt dataset num_tokens""" + return self.tgt_dataset.num_tokens(index) + + def ordered_indices(self): + """Just use the tgt dataset ordered_indices""" + return self.tgt_dataset.ordered_indices() + + def size(self, index): + """Return an example's size as a float or tuple. This value is used + when filtering a dataset with ``--max-positions``. + + Note: we use *tgt_dataset* to approximate the length of the source + sentence, since we do not know the actual length until after + backtranslation. + """ + tgt_size = self.tgt_dataset.size(index)[0] + return (tgt_size, tgt_size) + + @property + def supports_prefetch(self): + return getattr(self.tgt_dataset, "supports_prefetch", False) + + def prefetch(self, indices): + return self.tgt_dataset.prefetch(indices) diff --git a/fairseq/data/base_wrapper_dataset.py b/fairseq/data/base_wrapper_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..134d398b47dc73c8807759188504aee205b3b34d --- /dev/null +++ b/fairseq/data/base_wrapper_dataset.py @@ -0,0 +1,78 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from torch.utils.data.dataloader import default_collate + +from . import FairseqDataset + + +class BaseWrapperDataset(FairseqDataset): + def __init__(self, dataset): + super().__init__() + self.dataset = dataset + + def __getitem__(self, index): + return self.dataset[index] + + def __len__(self): + return len(self.dataset) + + def collater(self, samples): + if hasattr(self.dataset, "collater"): + return self.dataset.collater(samples) + else: + return default_collate(samples) + + @property + def sizes(self): + return self.dataset.sizes + + def num_tokens(self, index): + return self.dataset.num_tokens(index) + + def size(self, index): + return self.dataset.size(index) + + def ordered_indices(self): + return self.dataset.ordered_indices() + + @property + def supports_prefetch(self): + return getattr(self.dataset, "supports_prefetch", False) + + def attr(self, attr: str, index: int): + return self.dataset.attr(attr, index) + + def prefetch(self, indices): + self.dataset.prefetch(indices) + + def get_batch_shapes(self): + return self.dataset.get_batch_shapes() + + def batch_by_size( + self, + indices, + max_tokens=None, + max_sentences=None, + required_batch_size_multiple=1, + ): + return self.dataset.batch_by_size( + indices, + max_tokens=max_tokens, + max_sentences=max_sentences, + required_batch_size_multiple=required_batch_size_multiple, + ) + + def filter_indices_by_size(self, indices, max_sizes): + return self.dataset.filter_indices_by_size(indices, max_sizes) + + @property + def can_reuse_epoch_itr_across_epochs(self): + return self.dataset.can_reuse_epoch_itr_across_epochs + + def set_epoch(self, epoch): + super().set_epoch(epoch) + if hasattr(self.dataset, "set_epoch"): + self.dataset.set_epoch(epoch) diff --git a/fairseq/data/bucket_pad_length_dataset.py b/fairseq/data/bucket_pad_length_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..0f9410014845873bb0344fca6478c231c88e9dea --- /dev/null +++ b/fairseq/data/bucket_pad_length_dataset.py @@ -0,0 +1,78 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import numpy as np +import torch.nn.functional as F +from fairseq.data import BaseWrapperDataset +from fairseq.data.data_utils import get_buckets, get_bucketed_sizes + + +class BucketPadLengthDataset(BaseWrapperDataset): + """ + Bucket and pad item lengths to the nearest bucket size. This can be used to + reduce the number of unique batch shapes, which is important on TPUs since + each new batch shape requires a recompilation. + + Args: + dataset (FairseqDatset): dataset to bucket + sizes (List[int]): all item sizes + num_buckets (int): number of buckets to create + pad_idx (int): padding symbol + left_pad (bool): if True, pad on the left; otherwise right pad + """ + + def __init__( + self, + dataset, + sizes, + num_buckets, + pad_idx, + left_pad, + tensor_key=None, + ): + super().__init__(dataset) + self.pad_idx = pad_idx + self.left_pad = left_pad + + assert num_buckets > 0 + self.buckets = get_buckets(sizes, num_buckets) + self._bucketed_sizes = get_bucketed_sizes(sizes, self.buckets) + self._tensor_key = tensor_key + + def _set_tensor(self, item, val): + if self._tensor_key is None: + return val + item[self._tensor_key] = val + return item + + def _get_tensor(self, item): + if self._tensor_key is None: + return item + return item[self._tensor_key] + + def _pad(self, tensor, bucket_size, dim=-1): + num_pad = bucket_size - tensor.size(dim) + return F.pad( + tensor, + (num_pad if self.left_pad else 0, 0 if self.left_pad else num_pad), + value=self.pad_idx, + ) + + def __getitem__(self, index): + item = self.dataset[index] + bucket_size = self._bucketed_sizes[index] + tensor = self._get_tensor(item) + padded = self._pad(tensor, bucket_size) + return self._set_tensor(item, padded) + + @property + def sizes(self): + return self._bucketed_sizes + + def num_tokens(self, index): + return self._bucketed_sizes[index] + + def size(self, index): + return self._bucketed_sizes[index] diff --git a/fairseq/data/colorize_dataset.py b/fairseq/data/colorize_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..6ef097bff1a013f4944b1cb55e1e7e4e2480b3a6 --- /dev/null +++ b/fairseq/data/colorize_dataset.py @@ -0,0 +1,25 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch + +from . import BaseWrapperDataset + + +class ColorizeDataset(BaseWrapperDataset): + """ Adds 'colors' property to net input that is obtained from the provided color getter for use by models """ + + def __init__(self, dataset, color_getter): + super().__init__(dataset) + self.color_getter = color_getter + + def collater(self, samples): + base_collate = super().collater(samples) + if len(base_collate) > 0: + base_collate["net_input"]["colors"] = torch.tensor( + list(self.color_getter(self.dataset, s["id"]) for s in samples), + dtype=torch.long, + ) + return base_collate diff --git a/fairseq/data/concat_dataset.py b/fairseq/data/concat_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..01a4078bb159fa44b2d1062b9a971fe7f1abd1c2 --- /dev/null +++ b/fairseq/data/concat_dataset.py @@ -0,0 +1,124 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import bisect + +import numpy as np +from torch.utils.data.dataloader import default_collate + +from . import FairseqDataset + + +class ConcatDataset(FairseqDataset): + @staticmethod + def cumsum(sequence, sample_ratios): + r, s = [], 0 + for e, ratio in zip(sequence, sample_ratios): + curr_len = int(ratio * len(e)) + r.append(curr_len + s) + s += curr_len + return r + + def __init__(self, datasets, sample_ratios=1): + super(ConcatDataset, self).__init__() + assert len(datasets) > 0, "datasets should not be an empty iterable" + self.datasets = list(datasets) + if isinstance(sample_ratios, int): + sample_ratios = [sample_ratios] * len(self.datasets) + self.sample_ratios = sample_ratios + self.cumulative_sizes = self.cumsum(self.datasets, sample_ratios) + self.real_sizes = [len(d) for d in self.datasets] + + def __len__(self): + return self.cumulative_sizes[-1] + + def __getitem__(self, idx): + dataset_idx, sample_idx = self._get_dataset_and_sample_index(idx) + return self.datasets[dataset_idx][sample_idx] + + def _get_dataset_and_sample_index(self, idx: int): + dataset_idx = bisect.bisect_right(self.cumulative_sizes, idx) + if dataset_idx == 0: + sample_idx = idx + else: + sample_idx = idx - self.cumulative_sizes[dataset_idx - 1] + sample_idx = sample_idx % self.real_sizes[dataset_idx] + return dataset_idx, sample_idx + + def collater(self, samples, **extra_args): + # For now only supports datasets with same underlying collater implementations + if hasattr(self.datasets[0], "collater"): + return self.datasets[0].collater(samples, **extra_args) + else: + return default_collate(samples, **extra_args) + + def size(self, idx: int): + """ + Return an example's size as a float or tuple. + """ + dataset_idx, sample_idx = self._get_dataset_and_sample_index(idx) + return self.datasets[dataset_idx].size(sample_idx) + + def num_tokens(self, index: int): + return np.max(self.size(index)) + + def attr(self, attr: str, index: int): + dataset_idx = bisect.bisect_right(self.cumulative_sizes, index) + return getattr(self.datasets[dataset_idx], attr, None) + + @property + def sizes(self): + _dataset_sizes = [] + for ds, sr in zip(self.datasets, self.sample_ratios): + if isinstance(ds.sizes, np.ndarray): + _dataset_sizes.append(np.tile(ds.sizes, sr)) + else: + # Only support underlying dataset with single size array. + assert isinstance(ds.sizes, list) + _dataset_sizes.append(np.tile(ds.sizes[0], sr)) + return np.concatenate(_dataset_sizes) + + @property + def supports_prefetch(self): + return all(d.supports_prefetch for d in self.datasets) + + def ordered_indices(self): + """ + Returns indices sorted by length. So less padding is needed. + """ + if isinstance(self.sizes, np.ndarray) and len(self.sizes.shape) > 1: + # special handling for concatenating lang_pair_datasets + indices = np.arange(len(self)) + sizes = self.sizes + tgt_sizes = ( + sizes[:, 1] if len(sizes.shape) > 0 and sizes.shape[1] > 1 else None + ) + src_sizes = ( + sizes[:, 0] if len(sizes.shape) > 0 and sizes.shape[1] > 1 else sizes + ) + # sort by target length, then source length + if tgt_sizes is not None: + indices = indices[np.argsort(tgt_sizes[indices], kind="mergesort")] + return indices[np.argsort(src_sizes[indices], kind="mergesort")] + else: + return np.argsort(self.sizes) + + def prefetch(self, indices): + frm = 0 + for to, ds in zip(self.cumulative_sizes, self.datasets): + real_size = len(ds) + if getattr(ds, "supports_prefetch", False): + ds.prefetch([(i - frm) % real_size for i in indices if frm <= i < to]) + frm = to + + @property + def can_reuse_epoch_itr_across_epochs(self): + return all(d.can_reuse_epoch_itr_across_epochs for d in self.datasets) + + def set_epoch(self, epoch): + super().set_epoch(epoch) + for ds in self.datasets: + if hasattr(ds, "set_epoch"): + ds.set_epoch(epoch) diff --git a/fairseq/data/concat_sentences_dataset.py b/fairseq/data/concat_sentences_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..625a29370e90f9d1d7274024afb902ed83a22325 --- /dev/null +++ b/fairseq/data/concat_sentences_dataset.py @@ -0,0 +1,54 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch + +from . import FairseqDataset + + +class ConcatSentencesDataset(FairseqDataset): + def __init__(self, *datasets): + super().__init__() + self.datasets = datasets + assert all( + len(ds) == len(datasets[0]) for ds in datasets + ), "datasets must have the same length" + + def __getitem__(self, index): + return torch.cat([ds[index] for ds in self.datasets]) + + def __len__(self): + return len(self.datasets[0]) + + def collater(self, samples): + return self.datasets[0].collater(samples) + + @property + def sizes(self): + return sum(ds.sizes for ds in self.datasets) + + def num_tokens(self, index): + return sum(ds.num_tokens(index) for ds in self.datasets) + + def size(self, index): + return sum(ds.size(index) for ds in self.datasets) + + def ordered_indices(self): + return self.datasets[0].ordered_indices() + + @property + def supports_prefetch(self): + return any(getattr(ds, "supports_prefetch", False) for ds in self.datasets) + + def prefetch(self, indices): + for ds in self.datasets: + if getattr(ds, "supports_prefetch", False): + ds.prefetch(indices) + + def set_epoch(self, epoch): + super().set_epoch(epoch) + for ds in self.datasets: + if hasattr(ds, "set_epoch"): + ds.set_epoch(epoch) diff --git a/fairseq/data/data_utils.py b/fairseq/data/data_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..b3de57681e0fb6b026003eff19f7745caf6799d3 --- /dev/null +++ b/fairseq/data/data_utils.py @@ -0,0 +1,595 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +try: + from collections.abc import Iterable +except ImportError: + from collections import Iterable +import contextlib +import itertools +import logging +import re +import warnings +from typing import Optional, Tuple + +import numpy as np +import torch + +from fairseq.file_io import PathManager +from fairseq import utils +import os + +logger = logging.getLogger(__name__) + + +def infer_language_pair(path): + """Infer language pair from filename: <split>.<lang1>-<lang2>.(...).idx""" + src, dst = None, None + for filename in PathManager.ls(path): + parts = filename.split(".") + if len(parts) >= 3 and len(parts[1].split("-")) == 2: + return parts[1].split("-") + return src, dst + + +def collate_tokens( + values, + pad_idx, + eos_idx=None, + left_pad=False, + move_eos_to_beginning=False, + pad_to_length=None, + pad_to_multiple=1, + pad_to_bsz=None, +): + """Convert a list of 1d tensors into a padded 2d tensor.""" + size = max(v.size(0) for v in values) + size = size if pad_to_length is None else max(size, pad_to_length) + if pad_to_multiple != 1 and size % pad_to_multiple != 0: + size = int(((size - 0.1) // pad_to_multiple + 1) * pad_to_multiple) + + batch_size = len(values) if pad_to_bsz is None else max(len(values), pad_to_bsz) + res = values[0].new(batch_size, size).fill_(pad_idx) + + def copy_tensor(src, dst): + assert dst.numel() == src.numel() + if move_eos_to_beginning: + if eos_idx is None: + # if no eos_idx is specified, then use the last token in src + dst[0] = src[-1] + else: + dst[0] = eos_idx + dst[1:] = src[:-1] + else: + dst.copy_(src) + + for i, v in enumerate(values): + copy_tensor(v, res[i][size - len(v) :] if left_pad else res[i][: len(v)]) + return res + +def load_indexed_dataset( + path, dictionary=None, dataset_impl=None, combine=False, default="cached" +): + """A helper function for loading indexed datasets. + + Args: + path (str): path to indexed dataset (e.g., 'data-bin/train') + dictionary (~fairseq.data.Dictionary): data dictionary + dataset_impl (str, optional): which dataset implementation to use. If + not provided, it will be inferred automatically. For legacy indexed + data we use the 'cached' implementation by default. + combine (bool, optional): automatically load and combine multiple + datasets. For example, if *path* is 'data-bin/train', then we will + combine 'data-bin/train', 'data-bin/train1', ... and return a + single ConcatDataset instance. + """ + import fairseq.data.indexed_dataset as indexed_dataset + from fairseq.data.concat_dataset import ConcatDataset + + datasets = [] + for k in itertools.count(): + path_k = path + (str(k) if k > 0 else "") + try: + path_k = indexed_dataset.get_indexed_dataset_to_local(path_k) + except Exception as e: + if "StorageException: [404] Path not found" in str(e): + logger.warning(f"path_k: {e} not found") + else: + raise e + + dataset_impl_k = dataset_impl + if dataset_impl_k is None: + dataset_impl_k = indexed_dataset.infer_dataset_impl(path_k) + dataset = indexed_dataset.make_dataset( + path_k, + impl=dataset_impl_k or default, + fix_lua_indexing=True, + dictionary=dictionary, + ) + if dataset is None: + break + logger.info("loaded {:,} examples from: {}".format(len(dataset), path_k)) + datasets.append(dataset) + if not combine: + break + if len(datasets) == 0: + return None + elif len(datasets) == 1: + return datasets[0] + else: + return ConcatDataset(datasets) + + +@contextlib.contextmanager +def numpy_seed(seed, *addl_seeds): + """Context manager which seeds the NumPy PRNG with the specified seed and + restores the state afterward""" + if seed is None: + yield + return + if len(addl_seeds) > 0: + seed = int(hash((seed, *addl_seeds)) % 1e6) + state = np.random.get_state() + np.random.seed(seed) + try: + yield + finally: + np.random.set_state(state) + + +def collect_filtered(function, iterable, filtered): + """ + Similar to :func:`filter` but collects filtered elements in ``filtered``. + + Args: + function (callable): function that returns ``False`` for elements that + should be filtered + iterable (iterable): iterable to filter + filtered (list): list to store filtered elements + """ + for el in iterable: + if function(el): + yield el + else: + filtered.append(el) + + +def _filter_by_size_dynamic(indices, size_fn, max_positions, raise_exception=False): + def compare_leq(a, b): + return a <= b if not isinstance(a, tuple) else max(a) <= b + + def check_size(idx): + if isinstance(max_positions, float) or isinstance(max_positions, int): + return size_fn(idx) <= max_positions + elif isinstance(max_positions, dict): + idx_size = size_fn(idx) + assert isinstance(idx_size, dict) + intersect_keys = set(max_positions.keys()) & set(idx_size.keys()) + return all( + all( + a is None or b is None or a <= b + for a, b in zip(idx_size[key], max_positions[key]) + ) + for key in intersect_keys + ) + else: + # For MultiCorpusSampledDataset, will generalize it later + if not isinstance(size_fn(idx), Iterable): + return all(size_fn(idx) <= b for b in max_positions) + return all( + a is None or b is None or a <= b + for a, b in zip(size_fn(idx), max_positions) + ) + + ignored = [] + itr = collect_filtered(check_size, indices, ignored) + indices = np.fromiter(itr, dtype=np.int64, count=-1) + return indices, ignored + + +def filter_by_size(indices, dataset, max_positions, raise_exception=False): + """ + [deprecated] Filter indices based on their size. + Use `FairseqDataset::filter_indices_by_size` instead. + + Args: + indices (List[int]): ordered list of dataset indices + dataset (FairseqDataset): fairseq dataset instance + max_positions (tuple): filter elements larger than this size. + Comparisons are done component-wise. + raise_exception (bool, optional): if ``True``, raise an exception if + any elements are filtered (default: False). + """ + warnings.warn( + "data_utils.filter_by_size is deprecated. " + "Use `FairseqDataset::filter_indices_by_size` instead.", + stacklevel=2, + ) + if isinstance(max_positions, float) or isinstance(max_positions, int): + if hasattr(dataset, "sizes") and isinstance(dataset.sizes, np.ndarray): + ignored = indices[dataset.sizes[indices] > max_positions].tolist() + indices = indices[dataset.sizes[indices] <= max_positions] + elif ( + hasattr(dataset, "sizes") + and isinstance(dataset.sizes, list) + and len(dataset.sizes) == 1 + ): + ignored = indices[dataset.sizes[0][indices] > max_positions].tolist() + indices = indices[dataset.sizes[0][indices] <= max_positions] + else: + indices, ignored = _filter_by_size_dynamic( + indices, dataset.size, max_positions + ) + else: + indices, ignored = _filter_by_size_dynamic(indices, dataset.size, max_positions) + + if len(ignored) > 0 and raise_exception: + raise Exception( + ( + "Size of sample #{} is invalid (={}) since max_positions={}, " + "skip this example with --skip-invalid-size-inputs-valid-test" + ).format(ignored[0], dataset.size(ignored[0]), max_positions) + ) + if len(ignored) > 0: + logger.warning( + ( + "{} samples have invalid sizes and will be skipped, " + "max_positions={}, first few sample ids={}" + ).format(len(ignored), max_positions, ignored[:10]) + ) + return indices + + +def filter_paired_dataset_indices_by_size(src_sizes, tgt_sizes, indices, max_sizes): + """Filter a list of sample indices. Remove those that are longer + than specified in max_sizes. + + Args: + indices (np.array): original array of sample indices + max_sizes (int or list[int] or tuple[int]): max sample size, + can be defined separately for src and tgt (then list or tuple) + + Returns: + np.array: filtered sample array + list: list of removed indices + """ + if max_sizes is None: + return indices, [] + if type(max_sizes) in (int, float): + max_src_size, max_tgt_size = max_sizes, max_sizes + else: + max_src_size, max_tgt_size = max_sizes + if tgt_sizes is None: + ignored = indices[src_sizes[indices] > max_src_size] + else: + ignored = indices[ + (src_sizes[indices] > max_src_size) | (tgt_sizes[indices] > max_tgt_size) + ] + if len(ignored) > 0: + if tgt_sizes is None: + indices = indices[src_sizes[indices] <= max_src_size] + else: + indices = indices[ + (src_sizes[indices] <= max_src_size) + & (tgt_sizes[indices] <= max_tgt_size) + ] + return indices, ignored.tolist() + + +def batch_by_size( + indices, + num_tokens_fn, + num_tokens_vec=None, + max_tokens=None, + max_sentences=None, + required_batch_size_multiple=1, + fixed_shapes=None, +): + """ + Yield mini-batches of indices bucketed by size. Batches may contain + sequences of different lengths. + + Args: + indices (List[int]): ordered list of dataset indices + num_tokens_fn (callable): function that returns the number of tokens at + a given index + num_tokens_vec (List[int], optional): precomputed vector of the number + of tokens for each index in indices (to enable faster batch generation) + max_tokens (int, optional): max number of tokens in each batch + (default: None). + max_sentences (int, optional): max number of sentences in each + batch (default: None). + required_batch_size_multiple (int, optional): require batch size to + be less than N or a multiple of N (default: 1). + fixed_shapes (List[Tuple[int, int]], optional): if given, batches will + only be created with the given shapes. *max_sentences* and + *required_batch_size_multiple* will be ignored (default: None). + """ + try: + from fairseq.data.data_utils_fast import ( + batch_by_size_fn, + batch_by_size_vec, + batch_fixed_shapes_fast, + ) + except ImportError: + raise ImportError( + "Please build Cython components with: " + "`python setup.py build_ext --inplace`" + ) + except ValueError: + raise ValueError( + "Please build (or rebuild) Cython components with `python setup.py build_ext --inplace`." + ) + + # added int() to avoid TypeError: an integer is required + max_tokens = ( + int(max_tokens) if max_tokens is not None else -1 + ) + max_sentences = max_sentences if max_sentences is not None else -1 + bsz_mult = required_batch_size_multiple + + if not isinstance(indices, np.ndarray): + indices = np.fromiter(indices, dtype=np.int64, count=-1) + + if num_tokens_vec is not None and not isinstance(num_tokens_vec, np.ndarray): + num_tokens_vec = np.fromiter(num_tokens_vec, dtype=np.int64, count=-1) + + if fixed_shapes is None: + if num_tokens_vec is None: + return batch_by_size_fn( + indices, + num_tokens_fn, + max_tokens, + max_sentences, + bsz_mult, + ) + else: + return batch_by_size_vec( + indices, + num_tokens_vec, + max_tokens, + max_sentences, + bsz_mult, + ) + + else: + fixed_shapes = np.array(fixed_shapes, dtype=np.int64) + sort_order = np.lexsort( + [ + fixed_shapes[:, 1].argsort(), # length + fixed_shapes[:, 0].argsort(), # bsz + ] + ) + fixed_shapes_sorted = fixed_shapes[sort_order] + return batch_fixed_shapes_fast(indices, num_tokens_fn, fixed_shapes_sorted) + + +def post_process(sentence: str, symbol: str): + if symbol == "sentencepiece": + sentence = sentence.replace(" ", "").replace("\u2581", " ").strip() + elif symbol == "wordpiece": + sentence = sentence.replace(" ", "").replace("_", " ").strip() + elif symbol == "letter": + sentence = sentence.replace(" ", "").replace("|", " ").strip() + elif symbol == "silence": + import re + sentence = sentence.replace("<SIL>", "") + sentence = re.sub(' +', ' ', sentence).strip() + elif symbol == "_EOW": + sentence = sentence.replace(" ", "").replace("_EOW", " ").strip() + elif symbol in {"subword_nmt", "@@ ", "@@"}: + if symbol == "subword_nmt": + symbol = "@@ " + sentence = (sentence + " ").replace(symbol, "").rstrip() + elif symbol == "none": + pass + elif symbol is not None: + raise NotImplementedError(f"Unknown post_process option: {symbol}") + return sentence + + +def compute_mask_indices( + shape: Tuple[int, int], + padding_mask: Optional[torch.Tensor], + mask_prob: float, + mask_length: int, + mask_type: str = "static", + mask_other: float = 0.0, + min_masks: int = 0, + no_overlap: bool = False, + min_space: int = 0, +) -> np.ndarray: + """ + Computes random mask spans for a given shape + + Args: + shape: the the shape for which to compute masks. + should be of size 2 where first element is batch size and 2nd is timesteps + padding_mask: optional padding mask of the same size as shape, which will prevent masking padded elements + mask_prob: probability for each token to be chosen as start of the span to be masked. this will be multiplied by + number of timesteps divided by length of mask span to mask approximately this percentage of all elements. + however due to overlaps, the actual number will be smaller (unless no_overlap is True) + mask_type: how to compute mask lengths + static = fixed size + uniform = sample from uniform distribution [mask_other, mask_length*2] + normal = sample from normal distribution with mean mask_length and stdev mask_other. mask is min 1 element + poisson = sample from possion distribution with lambda = mask length + min_masks: minimum number of masked spans + no_overlap: if false, will switch to an alternative recursive algorithm that prevents spans from overlapping + min_space: only used if no_overlap is True, this is how many elements to keep unmasked between spans + """ + + bsz, all_sz = shape + mask = np.full((bsz, all_sz), False) + + all_num_mask = int( + # add a random number for probabilistic rounding + mask_prob * all_sz / float(mask_length) + + np.random.rand() + ) + + all_num_mask = max(min_masks, all_num_mask) + + mask_idcs = [] + for i in range(bsz): + if padding_mask is not None: + sz = all_sz - padding_mask[i].long().sum().item() + num_mask = int( + # add a random number for probabilistic rounding + mask_prob * sz / float(mask_length) + + np.random.rand() + ) + num_mask = max(min_masks, num_mask) + else: + sz = all_sz + num_mask = all_num_mask + + if mask_type == "static": + lengths = np.full(num_mask, mask_length) + elif mask_type == "uniform": + lengths = np.random.randint(mask_other, mask_length * 2 + 1, size=num_mask) + elif mask_type == "normal": + lengths = np.random.normal(mask_length, mask_other, size=num_mask) + lengths = [max(1, int(round(x))) for x in lengths] + elif mask_type == "poisson": + lengths = np.random.poisson(mask_length, size=num_mask) + lengths = [int(round(x)) for x in lengths] + else: + raise Exception("unknown mask selection " + mask_type) + + if sum(lengths) == 0: + lengths[0] = min(mask_length, sz - 1) + + if no_overlap: + mask_idc = [] + + def arrange(s, e, length, keep_length): + span_start = np.random.randint(s, e - length) + mask_idc.extend(span_start + i for i in range(length)) + + new_parts = [] + if span_start - s - min_space >= keep_length: + new_parts.append((s, span_start - min_space + 1)) + if e - span_start - keep_length - min_space > keep_length: + new_parts.append((span_start + length + min_space, e)) + return new_parts + + parts = [(0, sz)] + min_length = min(lengths) + for length in sorted(lengths, reverse=True): + lens = np.fromiter( + (e - s if e - s >= length + min_space else 0 for s, e in parts), + np.int, + ) + l_sum = np.sum(lens) + if l_sum == 0: + break + probs = lens / np.sum(lens) + c = np.random.choice(len(parts), p=probs) + s, e = parts.pop(c) + parts.extend(arrange(s, e, length, min_length)) + mask_idc = np.asarray(mask_idc) + else: + min_len = min(lengths) + if sz - min_len <= num_mask: + min_len = sz - num_mask - 1 + + mask_idc = np.random.choice(sz - min_len, num_mask, replace=False) + + mask_idc = np.asarray( + [ + mask_idc[j] + offset + for j in range(len(mask_idc)) + for offset in range(lengths[j]) + ] + ) + + mask_idcs.append(np.unique(mask_idc[mask_idc < sz])) + + min_len = min([len(m) for m in mask_idcs]) + for i, mask_idc in enumerate(mask_idcs): + if len(mask_idc) > min_len: + mask_idc = np.random.choice(mask_idc, min_len, replace=False) + mask[i, mask_idc] = True + + return mask + + +def get_mem_usage(): + try: + import psutil + + mb = 1024 * 1024 + return f"used={psutil.virtual_memory().used / mb}Mb; avail={psutil.virtual_memory().available / mb}Mb" + except ImportError: + return "N/A" + + +# lens: torch.LongTensor +# returns: torch.BoolTensor +def lengths_to_padding_mask(lens): + bsz, max_lens = lens.size(0), torch.max(lens).item() + mask = torch.arange(max_lens).to(lens.device).view(1, max_lens) + mask = mask.expand(bsz, -1) >= lens.view(bsz, 1).expand(-1, max_lens) + return mask + + +# lens: torch.LongTensor +# returns: torch.BoolTensor +def lengths_to_mask(lens): + return ~lengths_to_padding_mask(lens) + + +def get_buckets(sizes, num_buckets): + buckets = np.unique( + np.percentile( + sizes, + np.linspace(0, 100, num_buckets + 1), + interpolation='lower', + )[1:] + ) + return buckets + + +def get_bucketed_sizes(orig_sizes, buckets): + sizes = np.copy(orig_sizes) + assert np.min(sizes) >= 0 + start_val = -1 + for end_val in buckets: + mask = (sizes > start_val) & (sizes <= end_val) + sizes[mask] = end_val + start_val = end_val + return sizes + + + +def _find_extra_valid_paths(dataset_path: str) -> set: + paths = utils.split_paths(dataset_path) + all_valid_paths = set() + for sub_dir in paths: + contents = PathManager.ls(sub_dir) + valid_paths = [c for c in contents if re.match("valid*[0-9].*", c) is not None] + all_valid_paths |= {os.path.basename(p) for p in valid_paths} + # Remove .bin, .idx etc + roots = {os.path.splitext(p)[0] for p in all_valid_paths} + return roots + + +def raise_if_valid_subsets_unintentionally_ignored(train_cfg) -> None: + """Raises if there are paths matching 'valid*[0-9].*' which are not combined or ignored.""" + if ( + train_cfg.dataset.ignore_unused_valid_subsets + or train_cfg.dataset.combine_valid_subsets + or train_cfg.dataset.disable_validation + or not hasattr(train_cfg.task, "data") + ): + return + other_paths = _find_extra_valid_paths(train_cfg.task.data) + specified_subsets = train_cfg.dataset.valid_subset.split(",") + ignored_paths = [p for p in other_paths if p not in specified_subsets] + if ignored_paths: + advice = "Set --combine-val to combine them or --ignore-unused-valid-subsets to ignore them." + msg = f"Valid paths {ignored_paths} will be ignored. {advice}" + raise ValueError(msg) diff --git a/fairseq/data/data_utils_fast.pyx b/fairseq/data/data_utils_fast.pyx new file mode 100644 index 0000000000000000000000000000000000000000..c61f31d6b2113d4c6a03d6553335997098ba0c20 --- /dev/null +++ b/fairseq/data/data_utils_fast.pyx @@ -0,0 +1,178 @@ +# cython: language_level=3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import numpy as np + +cimport cython +cimport numpy as np + +from libc.stdint cimport int32_t, int64_t +from libcpp cimport bool as bool_t + +ctypedef int64_t DTYPE_t + +@cython.cdivision(True) +@cython.boundscheck(False) +@cython.wraparound(False) +cpdef list batch_by_size_vec( + np.ndarray[int64_t, ndim=1] indices, + np.ndarray[int64_t, ndim=1] num_tokens_vec, + int64_t max_tokens, + int64_t max_sentences, + int32_t bsz_mult, +): + if indices.shape[0] == 0: + return [] + + assert max_tokens <= 0 or np.max(num_tokens_vec) <= max_tokens, ( + f"Sentences lengths should not exceed max_tokens={max_tokens}" + ) + + cdef int32_t indices_len = indices.shape[0] + cdef np.ndarray[int32_t, ndim=1] batches_ends = \ + np.zeros(indices_len, dtype=np.int32) + cdef int32_t[:] batches_ends_view = batches_ends + cdef int64_t[:] num_tokens_view = num_tokens_vec + + cdef int32_t pos = 0 + cdef int32_t new_batch_end = 0 + + cdef int64_t new_batch_max_tokens = 0 + cdef int32_t new_batch_sentences = 0 + cdef int64_t new_batch_num_tokens = 0 + + cdef bool_t overflow = False + cdef bool_t size_matches_with_bsz_mult = False + + cdef int32_t batches_count = 0 + cdef int32_t batch_start = 0 + cdef int64_t tail_max_tokens = 0 + cdef int64_t batch_max_tokens = 0 + + for pos in range(indices_len): + # At every pos we keep stats about the last complete batch [batch_start:batch_end), + # and tail [batch_end:pos]. + # 1) Every time when (batch + tail) forms a valid batch + # (according to max_tokens, max_sentences and bsz_mult) we append tail to batch. + # 2) When (batch+tail) violates max_tokens or max_sentences constraints + # we finalize running batch, and tail becomes a new batch. + # 3) There is a corner case when tail also violates constraints. + # In that situation [batch_end:pos-1] (tail without the current pos) + # gets added to the finalized batches, while [pos:pos] becomes a new tail. + # + # Important: For the sake of performance try to avoid using function calls within this loop. + + tail_max_tokens = tail_max_tokens \ + if tail_max_tokens > num_tokens_view[pos] \ + else num_tokens_view[pos] + new_batch_end = pos + 1 + new_batch_max_tokens = batch_max_tokens \ + if batch_max_tokens > tail_max_tokens \ + else tail_max_tokens + new_batch_sentences = new_batch_end - batch_start + new_batch_num_tokens = new_batch_sentences * new_batch_max_tokens + + overflow = (new_batch_sentences > max_sentences > 0 or + new_batch_num_tokens > max_tokens > 0) + size_matches_with_bsz_mult = (new_batch_sentences < bsz_mult or + new_batch_sentences % bsz_mult == 0) + + if overflow: + tail_num_tokens = tail_max_tokens * \ + (new_batch_end - batches_ends_view[batches_count]) + tail_overflow = tail_num_tokens > max_tokens > 0 + # In case of a tail overflow finalize two batches + if tail_overflow: + batches_count += 1 + batches_ends_view[batches_count] = pos + tail_max_tokens = num_tokens_view[pos] + batch_start = batches_ends_view[batches_count] + batches_count += 1 + new_batch_max_tokens = tail_max_tokens + + if overflow or size_matches_with_bsz_mult: + batches_ends_view[batches_count] = new_batch_end + batch_max_tokens = new_batch_max_tokens + tail_max_tokens = 0 + if batches_ends_view[batches_count] != indices_len: + batches_count += 1 + # Memory and time-efficient split + return np.split(indices, batches_ends[:batches_count]) + + +@cython.boundscheck(False) +@cython.wraparound(False) +cpdef list batch_by_size_fn( + np.ndarray[DTYPE_t, ndim=1] indices, + num_tokens_fn, + int64_t max_tokens, + int64_t max_sentences, + int32_t bsz_mult, +): + cdef int32_t indices_len = indices.shape[0] + cdef np.ndarray[int64_t, ndim=1] num_tokens_vec = np.zeros(indices_len, + dtype=np.int64) + cdef DTYPE_t[:] indices_view = indices + cdef DTYPE_t[:] num_tokens_vec_view = num_tokens_vec + cdef int64_t pos + for pos in range(indices_len): + num_tokens_vec[pos] = num_tokens_fn(indices_view[pos]) + return batch_by_size_vec(indices, num_tokens_vec, max_tokens, + max_sentences, bsz_mult,) + + +cdef _find_valid_shape( + DTYPE_t[:, :] shapes_view, + int64_t num_sentences, + int64_t num_tokens, +): + """Return index of first valid shape of -1 if none is found.""" + for i in range(shapes_view.shape[0]): + if num_sentences <= shapes_view[i][0] and num_tokens <= shapes_view[i][1]: + return i + return -1 + + +@cython.cdivision(True) +cpdef list batch_fixed_shapes_fast( + np.ndarray[DTYPE_t, ndim=1] indices, + num_tokens_fn, + np.ndarray[DTYPE_t, ndim=2] fixed_shapes_sorted, +): + cdef int64_t sample_len = 0 + cdef list sample_lens = [] + cdef list batch = [] + cdef list batches = [] + cdef int64_t mod_len + cdef int64_t i + cdef int64_t idx + cdef int64_t num_tokens + cdef DTYPE_t[:] indices_view = indices + cdef DTYPE_t[:, :] shapes_view = fixed_shapes_sorted + + for i in range(len(indices_view)): + idx = indices_view[i] + num_tokens = num_tokens_fn(idx) + sample_lens.append(num_tokens) + sample_len = max(sample_len, num_tokens) + + shape_idx = _find_valid_shape(shapes_view, len(batch) + 1, sample_len) + if shape_idx == -1: + batches.append(batch) + batch = [] + sample_lens = [] + sample_len = 0 + shapes_view = fixed_shapes_sorted + elif shape_idx > 0: + # small optimization for the next call to _find_valid_shape + shapes_view = shapes_view[shape_idx:] + + batch.append(idx) + + if len(batch) > 0: + batches.append(batch) + + return batches diff --git a/fairseq/data/denoising_dataset.py b/fairseq/data/denoising_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..bdb62c8d5db9c8755c72db4d0d8083c936f18dc8 --- /dev/null +++ b/fairseq/data/denoising_dataset.py @@ -0,0 +1,436 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +import numpy as np +import torch + +from . import FairseqDataset, data_utils + + +def collate( + samples, + pad_idx, + eos_idx, + vocab, + left_pad_source=False, + left_pad_target=False, + input_feeding=True, + pad_to_length=None, +): + assert input_feeding + if len(samples) == 0: + return {} + + def merge(key, left_pad, move_eos_to_beginning=False, pad_to_length=None): + return data_utils.collate_tokens( + [s[key] for s in samples], + pad_idx, + eos_idx=None, # use eos_idx of each sample instead of vocab.eos() + left_pad=left_pad, + move_eos_to_beginning=move_eos_to_beginning, + pad_to_length=pad_to_length, + ) + + id = torch.LongTensor([s["id"] for s in samples]) + src_tokens = merge( + "source", + left_pad=left_pad_source, + pad_to_length=pad_to_length["source"] if pad_to_length is not None else None, + ) + # sort by descending source length + src_lengths = torch.LongTensor([s["source"].numel() for s in samples]) + src_lengths, sort_order = src_lengths.sort(descending=True) + id = id.index_select(0, sort_order) + src_tokens = src_tokens.index_select(0, sort_order) + + prev_output_tokens = None + target = None + if samples[0].get("target", None) is not None: + target = merge( + "target", + left_pad=left_pad_target, + pad_to_length=pad_to_length["target"] + if pad_to_length is not None + else None, + ) + target = target.index_select(0, sort_order) + ntokens = sum(len(s["target"]) for s in samples) + + if input_feeding: + # we create a shifted version of targets for feeding the + # previous output token(s) into the next decoder step + prev_output_tokens = merge( + "target", + left_pad=left_pad_target, + move_eos_to_beginning=True, + pad_to_length=pad_to_length["target"] + if pad_to_length is not None + else None, + ) + prev_output_tokens = prev_output_tokens.index_select(0, sort_order) + else: + ntokens = sum(len(s["source"]) for s in samples) + + batch = { + "id": id, + "ntokens": ntokens, + "net_input": { + "src_tokens": src_tokens, + "src_lengths": src_lengths, + }, + "target": target, + "nsentences": samples[0]["source"].size(0), + "sort_order": sort_order, + } + if prev_output_tokens is not None: + batch["net_input"]["prev_output_tokens"] = prev_output_tokens + + return batch + + +class DenoisingDataset(FairseqDataset): + """ + A wrapper around TokenBlockDataset for BART dataset. + + Args: + dataset (TokenBlockDataset): dataset to wrap + sizes (List[int]): sentence lengths + vocab (~fairseq.data.Dictionary): vocabulary + mask_idx (int): dictionary index used for masked token + mask_whole_words: only mask whole words. This should be a byte mask + over vocab indices, indicating whether it is the beginning of a + word. We will extend any mask to encompass the whole word. + shuffle (bool, optional): shuffle the elements before batching. + Default: ``True`` + seed: Seed for random number generator for reproducibility. + args: argparse arguments. + """ + + def __init__( + self, + dataset, + sizes, + vocab, + mask_idx, + mask_whole_words, + shuffle, + seed, + args, + eos=None, + item_transform_func=None, + ): + self.dataset = dataset + + self.sizes = sizes + + self.vocab = vocab + self.shuffle = shuffle + self.seed = seed + self.mask_idx = mask_idx + self.mask_whole_word = mask_whole_words + self.mask_ratio = args.mask + self.random_ratio = args.mask_random + self.insert_ratio = args.insert + self.rotate_ratio = args.rotate + self.permute_sentence_ratio = args.permute_sentences + self.eos = eos if eos is not None else vocab.eos() + self.item_transform_func = item_transform_func + + if args.bpe != "gpt2": + self.full_stop_index = self.vocab.eos() + else: + assert args.bpe == "gpt2" + self.full_stop_index = self.vocab.index("13") + + self.replace_length = args.replace_length + if self.replace_length not in [-1, 0, 1]: + raise ValueError(f"invalid arg: replace_length={self.replace_length}") + if args.mask_length not in ["subword", "word", "span-poisson"]: + raise ValueError(f"invalid arg: mask-length={args.mask_length}") + if args.mask_length == "subword" and args.replace_length not in [0, 1]: + raise ValueError(f"if using subwords, use replace-length=1 or 0") + + self.mask_span_distribution = None + if args.mask_length == "span-poisson": + _lambda = args.poisson_lambda + + lambda_to_the_k = 1 + e_to_the_minus_lambda = math.exp(-_lambda) + k_factorial = 1 + ps = [] + for k in range(0, 128): + ps.append(e_to_the_minus_lambda * lambda_to_the_k / k_factorial) + lambda_to_the_k *= _lambda + k_factorial *= k + 1 + if ps[-1] < 0.0000001: + break + ps = torch.FloatTensor(ps) + self.mask_span_distribution = torch.distributions.Categorical(ps) + + self.epoch = 0 + + @property + def can_reuse_epoch_itr_across_epochs(self): + return True # only the noise changes, not item sizes + + def set_epoch(self, epoch, **unused): + self.epoch = epoch + + def __getitem__(self, index): + with data_utils.numpy_seed(self.seed, self.epoch, index): + tokens = self.dataset[index] + assert tokens[-1] == self.eos + source, target = tokens, tokens.clone() + + if self.permute_sentence_ratio > 0.0: + source = self.permute_sentences(source, self.permute_sentence_ratio) + + if self.mask_ratio > 0: + source = self.add_whole_word_mask(source, self.mask_ratio) + + if self.insert_ratio > 0: + source = self.add_insertion_noise(source, self.insert_ratio) + + if self.rotate_ratio > 0.0 and np.random.random() < self.rotate_ratio: + source = self.add_rolling_noise(source) + # there can additional changes to make: + if self.item_transform_func is not None: + source, target = self.item_transform_func(source, target) + + assert (source >= 0).all() + assert (source[1:-1] >= 1).all() + assert (source <= len(self.vocab)).all() + assert source[0] == self.vocab.bos() + assert source[-1] == self.eos + return { + "id": index, + "source": source, + "target": target, + } + + def __len__(self): + return len(self.dataset) + + def permute_sentences(self, source, p=1.0): + full_stops = source == self.full_stop_index + # Pretend it ends with a full stop so last span is a sentence + full_stops[-2] = 1 + + # Tokens that are full stops, where the previous token is not + sentence_ends = (full_stops[1:] * ~full_stops[:-1]).nonzero(as_tuple=False) + 2 + result = source.clone() + + num_sentences = sentence_ends.size(0) + num_to_permute = math.ceil((num_sentences * 2 * p) / 2.0) + substitutions = torch.randperm(num_sentences)[:num_to_permute] + ordering = torch.arange(0, num_sentences) + ordering[substitutions] = substitutions[torch.randperm(num_to_permute)] + + # Ignore <bos> at start + index = 1 + for i in ordering: + sentence = source[(sentence_ends[i - 1] if i > 0 else 1) : sentence_ends[i]] + result[index : index + sentence.size(0)] = sentence + index += sentence.size(0) + return result + + def word_starts(self, source): + if self.mask_whole_word is not None: + is_word_start = self.mask_whole_word.gather(0, source) + else: + is_word_start = torch.ones(source.size()) + is_word_start[0] = 0 + is_word_start[-1] = 0 + return is_word_start + + def add_whole_word_mask(self, source, p): + is_word_start = self.word_starts(source) + num_to_mask = int(math.ceil(is_word_start.float().sum() * p)) + num_inserts = 0 + if num_to_mask == 0: + return source + + if self.mask_span_distribution is not None: + lengths = self.mask_span_distribution.sample(sample_shape=(num_to_mask,)) + + # Make sure we have enough to mask + cum_length = torch.cumsum(lengths, 0) + while cum_length[-1] < num_to_mask: + lengths = torch.cat( + [ + lengths, + self.mask_span_distribution.sample(sample_shape=(num_to_mask,)), + ], + dim=0, + ) + cum_length = torch.cumsum(lengths, 0) + + # Trim to masking budget + i = 0 + while cum_length[i] < num_to_mask: + i += 1 + lengths[i] = num_to_mask - (0 if i == 0 else cum_length[i - 1]) + num_to_mask = i + 1 + lengths = lengths[:num_to_mask] + + # Handle 0-length mask (inserts) separately + lengths = lengths[lengths > 0] + num_inserts = num_to_mask - lengths.size(0) + num_to_mask -= num_inserts + if num_to_mask == 0: + return self.add_insertion_noise(source, num_inserts / source.size(0)) + + assert (lengths > 0).all() + else: + lengths = torch.ones((num_to_mask,)).long() + assert is_word_start[-1] == 0 + word_starts = is_word_start.nonzero(as_tuple=False) + indices = word_starts[ + torch.randperm(word_starts.size(0))[:num_to_mask] + ].squeeze(1) + mask_random = torch.FloatTensor(num_to_mask).uniform_() < self.random_ratio + + source_length = source.size(0) + assert source_length - 1 not in indices + to_keep = torch.ones(source_length, dtype=torch.bool) + is_word_start[ + -1 + ] = 255 # acts as a long length, so spans don't go over the end of doc + if self.replace_length == 0: + to_keep[indices] = 0 + else: + # keep index, but replace it with [MASK] + source[indices] = self.mask_idx + source[indices[mask_random]] = torch.randint( + 1, len(self.vocab), size=(mask_random.sum(),) + ) + + if self.mask_span_distribution is not None: + assert len(lengths.size()) == 1 + assert lengths.size() == indices.size() + lengths -= 1 + while indices.size(0) > 0: + assert lengths.size() == indices.size() + lengths -= is_word_start[indices + 1].long() + uncompleted = lengths >= 0 + indices = indices[uncompleted] + 1 + mask_random = mask_random[uncompleted] + lengths = lengths[uncompleted] + if self.replace_length != -1: + # delete token + to_keep[indices] = 0 + else: + # keep index, but replace it with [MASK] + source[indices] = self.mask_idx + source[indices[mask_random]] = torch.randint( + 1, len(self.vocab), size=(mask_random.sum(),) + ) + else: + # A bit faster when all lengths are 1 + while indices.size(0) > 0: + uncompleted = is_word_start[indices + 1] == 0 + indices = indices[uncompleted] + 1 + mask_random = mask_random[uncompleted] + if self.replace_length != -1: + # delete token + to_keep[indices] = 0 + else: + # keep index, but replace it with [MASK] + source[indices] = self.mask_idx + source[indices[mask_random]] = torch.randint( + 1, len(self.vocab), size=(mask_random.sum(),) + ) + + assert source_length - 1 not in indices + + source = source[to_keep] + + if num_inserts > 0: + source = self.add_insertion_noise(source, num_inserts / source.size(0)) + + return source + + def add_permuted_noise(self, tokens, p): + num_words = len(tokens) + num_to_permute = math.ceil(((num_words * 2) * p) / 2.0) + substitutions = torch.randperm(num_words - 2)[:num_to_permute] + 1 + tokens[substitutions] = tokens[substitutions[torch.randperm(num_to_permute)]] + return tokens + + def add_rolling_noise(self, tokens): + offset = np.random.randint(1, max(1, tokens.size(-1) - 1) + 1) + tokens = torch.cat( + (tokens[0:1], tokens[offset:-1], tokens[1:offset], tokens[-1:]), + dim=0, + ) + return tokens + + def add_insertion_noise(self, tokens, p): + if p == 0.0: + return tokens + + num_tokens = len(tokens) + n = int(math.ceil(num_tokens * p)) + + noise_indices = torch.randperm(num_tokens + n - 2)[:n] + 1 + noise_mask = torch.zeros(size=(num_tokens + n,), dtype=torch.bool) + noise_mask[noise_indices] = 1 + result = torch.LongTensor(n + len(tokens)).fill_(-1) + + num_random = int(math.ceil(n * self.random_ratio)) + result[noise_indices[num_random:]] = self.mask_idx + result[noise_indices[:num_random]] = torch.randint( + low=1, high=len(self.vocab), size=(num_random,) + ) + + result[~noise_mask] = tokens + + assert (result >= 0).all() + return result + + def collater(self, samples, pad_to_length=None): + """Merge a list of samples to form a mini-batch. + Args: + samples (List[dict]): samples to collate + Returns: + dict: a mini-batch of data + """ + return collate( + samples, self.vocab.pad(), self.eos, self.vocab, pad_to_length=pad_to_length + ) + + def num_tokens(self, index): + """Return the number of tokens in a sample. This value is used to + enforce ``--max-tokens`` during batching.""" + return self.sizes[index] + + def size(self, index): + """Return an example's size as a float or tuple. This value is used when + filtering a dataset with ``--max-positions``.""" + return self.sizes[index] + + def ordered_indices(self): + """Return an ordered list of indices. Batches will be constructed based + on this order.""" + if self.shuffle: + indices = np.random.permutation(len(self)) + else: + indices = np.arange(len(self)) + return indices[np.argsort(self.sizes[indices], kind="mergesort")] + + def prefetch(self, indices): + self.src.prefetch(indices) + self.tgt.prefetch(indices) + + @property + def supports_prefetch(self): + return ( + hasattr(self.src, "supports_prefetch") + and self.src.supports_prefetch + and hasattr(self.tgt, "supports_prefetch") + and self.tgt.supports_prefetch + ) diff --git a/fairseq/data/dictionary.py b/fairseq/data/dictionary.py new file mode 100644 index 0000000000000000000000000000000000000000..0d8308a811c5e558d9024a18d8545804dc0ecdfd --- /dev/null +++ b/fairseq/data/dictionary.py @@ -0,0 +1,395 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import os +from collections import Counter +from multiprocessing import Pool + +import torch +from fairseq import utils +from fairseq.binarizer import safe_readline +from fairseq.data import data_utils +from fairseq.file_io import PathManager +from fairseq.tokenizer import tokenize_line + + +class Dictionary: + """A mapping from symbols to consecutive integers""" + + def __init__( + self, + *, # begin keyword-only arguments + bos="<s>", + pad="<pad>", + eos="</s>", + unk="<unk>", + extra_special_symbols=None, + ): + self.bos_word, self.unk_word, self.pad_word, self.eos_word = bos, unk, pad, eos + self.symbols = [] + self.count = [] + self.indices = {} + self.bos_index = self.add_symbol(bos) + self.pad_index = self.add_symbol(pad) + self.eos_index = self.add_symbol(eos) + self.unk_index = self.add_symbol(unk) + if extra_special_symbols: + for s in extra_special_symbols: + self.add_symbol(s) + self.nspecial = len(self.symbols) + + def __eq__(self, other): + return self.indices == other.indices + + def __getitem__(self, idx): + if idx < len(self.symbols): + return self.symbols[idx] + return self.unk_word + + def __len__(self): + """Returns the number of symbols in the dictionary""" + return len(self.symbols) + + def __contains__(self, sym): + return sym in self.indices + + def index(self, sym): + """Returns the index of the specified symbol""" + assert isinstance(sym, str) + if sym in self.indices: + return self.indices[sym] + return self.unk_index + + def string( + self, + tensor, + bpe_symbol=None, + escape_unk=False, + extra_symbols_to_ignore=None, + unk_string=None, + include_eos=False, + separator=" ", + ): + """Helper for converting a tensor of token indices to a string. + + Can optionally remove BPE symbols or escape <unk> words. + """ + if torch.is_tensor(tensor) and tensor.dim() == 2: + return "\n".join( + self.string(t, bpe_symbol, escape_unk, extra_symbols_to_ignore, include_eos=include_eos) + for t in tensor + ) + + extra_symbols_to_ignore = set(extra_symbols_to_ignore or []) + extra_symbols_to_ignore.add(self.eos()) + + def token_string(i): + if i == self.unk(): + if unk_string is not None: + return unk_string + else: + return self.unk_string(escape_unk) + else: + return self[i] + + if hasattr(self, "bos_index"): + extra_symbols_to_ignore.add(self.bos()) + + sent = separator.join( + token_string(i) + for i in tensor + if utils.item(i) not in extra_symbols_to_ignore + ) + + return data_utils.post_process(sent, bpe_symbol) + + def unk_string(self, escape=False): + """Return unknown string, optionally escaped as: <<unk>>""" + if escape: + return "<{}>".format(self.unk_word) + else: + return self.unk_word + + def add_symbol(self, word, n=1, overwrite=False): + """Adds a word to the dictionary""" + if word in self.indices and not overwrite: + idx = self.indices[word] + self.count[idx] = self.count[idx] + n + return idx + else: + idx = len(self.symbols) + self.indices[word] = idx + self.symbols.append(word) + self.count.append(n) + return idx + + def update(self, new_dict): + """Updates counts from new dictionary.""" + for word in new_dict.symbols: + idx2 = new_dict.indices[word] + if word in self.indices: + idx = self.indices[word] + self.count[idx] = self.count[idx] + new_dict.count[idx2] + else: + idx = len(self.symbols) + self.indices[word] = idx + self.symbols.append(word) + self.count.append(new_dict.count[idx2]) + + def finalize(self, threshold=-1, nwords=-1, padding_factor=8): + """Sort symbols by frequency in descending order, ignoring special ones. + + Args: + - threshold defines the minimum word count + - nwords defines the total number of words in the final dictionary, + including special symbols + - padding_factor can be used to pad the dictionary size to be a + multiple of 8, which is important on some hardware (e.g., Nvidia + Tensor Cores). + """ + if nwords <= 0: + nwords = len(self) + + new_indices = dict(zip(self.symbols[: self.nspecial], range(self.nspecial))) + new_symbols = self.symbols[: self.nspecial] + new_count = self.count[: self.nspecial] + + c = Counter( + dict( + sorted(zip(self.symbols[self.nspecial :], self.count[self.nspecial :])) + ) + ) + for symbol, count in c.most_common(nwords - self.nspecial): + if count >= threshold: + new_indices[symbol] = len(new_symbols) + new_symbols.append(symbol) + new_count.append(count) + else: + break + + assert len(new_symbols) == len(new_indices) + + self.count = list(new_count) + self.symbols = list(new_symbols) + self.indices = new_indices + + self.pad_to_multiple_(padding_factor) + + def pad_to_multiple_(self, padding_factor): + """Pad Dictionary size to be a multiple of *padding_factor*.""" + if padding_factor > 1: + i = 0 + while len(self) % padding_factor != 0: + symbol = "madeupword{:04d}".format(i) + self.add_symbol(symbol, n=0) + i += 1 + + def bos(self): + """Helper to get index of beginning-of-sentence symbol""" + return self.bos_index + + def pad(self): + """Helper to get index of pad symbol""" + return self.pad_index + + def eos(self): + """Helper to get index of end-of-sentence symbol""" + return self.eos_index + + def unk(self): + """Helper to get index of unk symbol""" + return self.unk_index + + @classmethod + def load(cls, f): + """Loads the dictionary from a text file with the format: + + ``` + <symbol0> <count0> + <symbol1> <count1> + ... + ``` + """ + d = cls() + d.add_from_file(f) + return d + + def add_from_file(self, f): + """ + Loads a pre-existing dictionary from a text file and adds its symbols + to this instance. + """ + if isinstance(f, str): + try: + with open(PathManager.get_local_path(f), "r", encoding="utf-8") as fd: + self.add_from_file(fd) + except FileNotFoundError as fnfe: + raise fnfe + except UnicodeError: + raise Exception( + "Incorrect encoding detected in {}, please " + "rebuild the dataset".format(f) + ) + return + + lines = f.readlines() + indices_start_line = self._load_meta(lines) + + for line in lines[indices_start_line:]: + try: + line, field = line.rstrip().rsplit(" ", 1) + if field == "#fairseq:overwrite": + overwrite = True + line, field = line.rsplit(" ", 1) + else: + overwrite = False + count = int(field) + word = line + if word in self and not overwrite: + raise RuntimeError( + "Duplicate word found when loading Dictionary: '{}'. " + "Duplicate words can overwrite earlier ones by adding the " + "#fairseq:overwrite flag at the end of the corresponding row " + "in the dictionary file. If using the Camembert model, please " + "download an updated copy of the model file.".format(word) + ) + self.add_symbol(word, n=count, overwrite=overwrite) + except ValueError: + raise ValueError( + "Incorrect dictionary format, expected '<token> <cnt> [flags]'" + ) + + def _save(self, f, kv_iterator): + if isinstance(f, str): + PathManager.mkdirs(os.path.dirname(f)) + with PathManager.open(f, "w", encoding="utf-8") as fd: + return self.save(fd) + for k, v in kv_iterator: + print("{} {}".format(k, v), file=f) + + def _get_meta(self): + return [], [] + + def _load_meta(self, lines): + return 0 + + def save(self, f): + """Stores dictionary into a text file""" + ex_keys, ex_vals = self._get_meta() + self._save( + f, + zip( + ex_keys + self.symbols[self.nspecial :], + ex_vals + self.count[self.nspecial :], + ), + ) + + def dummy_sentence(self, length): + t = torch.Tensor(length).uniform_(self.nspecial + 1, len(self)).long() + t[-1] = self.eos() + return t + + def encode_line( + self, + line, + line_tokenizer=tokenize_line, + add_if_not_exist=True, + consumer=None, + append_eos=True, + reverse_order=False, + ) -> torch.IntTensor: + words = line_tokenizer(line) + if reverse_order: + words = list(reversed(words)) + nwords = len(words) + ids = torch.IntTensor(nwords + 1 if append_eos else nwords) + + for i, word in enumerate(words): + if add_if_not_exist: + idx = self.add_symbol(word) + else: + idx = self.index(word) + if consumer is not None: + consumer(word, idx) + ids[i] = idx + if append_eos: + ids[nwords] = self.eos_index + return ids + + @staticmethod + def _add_file_to_dictionary_single_worker( + filename, tokenize, eos_word, worker_id=0, num_workers=1 + ): + counter = Counter() + with open(PathManager.get_local_path(filename), "r", encoding="utf-8") as f: + size = os.fstat(f.fileno()).st_size + chunk_size = size // num_workers + offset = worker_id * chunk_size + end = offset + chunk_size + f.seek(offset) + if offset > 0: + safe_readline(f) # drop first incomplete line + line = f.readline() + while line: + for word in tokenize(line): + counter.update([word]) + counter.update([eos_word]) + # f.tell() returns only an opaque number which can + # return to the position in the file via f.seek() + # and does not necessarily represent a byte position + # in the file. However, f.tell() is faithful to the + # byte position _most of the time_. Thus we can just + # check against the file size to prevent early exit. + if f.tell() > end and f.tell() < size: + break + line = f.readline() + return counter + + @staticmethod + def add_file_to_dictionary(filename, dict, tokenize, num_workers): + def merge_result(counter): + for w, c in sorted(counter.items()): + dict.add_symbol(w, c) + + if num_workers > 1: + pool = Pool(processes=num_workers) + results = [] + for worker_id in range(num_workers): + results.append( + pool.apply_async( + Dictionary._add_file_to_dictionary_single_worker, + (filename, tokenize, dict.eos_word, worker_id, num_workers), + ) + ) + pool.close() + pool.join() + for r in results: + merge_result(r.get()) + else: + merge_result( + Dictionary._add_file_to_dictionary_single_worker( + filename, tokenize, dict.eos_word + ) + ) + + +class TruncatedDictionary(object): + def __init__(self, wrapped_dict, length): + self.__class__ = type( + wrapped_dict.__class__.__name__, + (self.__class__, wrapped_dict.__class__), + {}, + ) + self.__dict__ = wrapped_dict.__dict__ + self.wrapped_dict = wrapped_dict + self.length = min(len(self.wrapped_dict), length) + + def __len__(self): + return self.length + + def __getitem__(self, i): + if i < self.length: + return self.wrapped_dict[i] + return self.wrapped_dict.unk() diff --git a/fairseq/data/encoders/__init__.py b/fairseq/data/encoders/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..7cbe00a10520331709441e5e77991bd2edca8c06 --- /dev/null +++ b/fairseq/data/encoders/__init__.py @@ -0,0 +1,29 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +import importlib +import os + +from fairseq import registry + + +build_tokenizer, register_tokenizer, TOKENIZER_REGISTRY, _ = registry.setup_registry( + "--tokenizer", + default=None, +) + + +build_bpe, register_bpe, BPE_REGISTRY, _ = registry.setup_registry( + "--bpe", + default=None, +) + + +# automatically import any Python files in the encoders/ directory +for file in sorted(os.listdir(os.path.dirname(__file__))): + if file.endswith(".py") and not file.startswith("_"): + module = file[: file.find(".py")] + importlib.import_module("fairseq.data.encoders." + module) diff --git a/fairseq/data/encoders/byte_bpe.py b/fairseq/data/encoders/byte_bpe.py new file mode 100644 index 0000000000000000000000000000000000000000..31e3a0627827f19ca7f0b58da45e46d40a80c3bf --- /dev/null +++ b/fairseq/data/encoders/byte_bpe.py @@ -0,0 +1,48 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +from dataclasses import dataclass, field + +from fairseq import file_utils +from fairseq.data.encoders import register_bpe +from fairseq.data.encoders.byte_utils import ( + SPACE, + SPACE_ESCAPE, + byte_encode, + smart_byte_decode, +) +from fairseq.dataclass import FairseqDataclass + + +@dataclass +class ByteBpeConfig(FairseqDataclass): + sentencepiece_model_path: str = field( + default="???", metadata={"help": "path to sentencepiece model"} + ) + + +@register_bpe("byte_bpe", dataclass=ByteBpeConfig) +class ByteBPE(object): + def __init__(self, cfg): + vocab = file_utils.cached_path(cfg.sentencepiece_model_path) + try: + import sentencepiece as spm + + self.sp = spm.SentencePieceProcessor() + self.sp.Load(vocab) + except ImportError: + raise ImportError( + "Please install sentencepiece with: pip install sentencepiece" + ) + + def encode(self, x: str) -> str: + byte_encoded = byte_encode(x) + return SPACE.join(self.sp.EncodeAsPieces(byte_encoded)) + + @staticmethod + def decode(x: str) -> str: + unescaped = x.replace(SPACE, "").replace(SPACE_ESCAPE, SPACE) + return smart_byte_decode(unescaped) diff --git a/fairseq/data/encoders/byte_utils.py b/fairseq/data/encoders/byte_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..a305c080926c2d094b7e8ae48f5331da82025a75 --- /dev/null +++ b/fairseq/data/encoders/byte_utils.py @@ -0,0 +1,51 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import re + + +WHITESPACE_NORMALIZER = re.compile(r"\s+") +SPACE = chr(32) +SPACE_ESCAPE = chr(9601) +# excluding non-breaking space (160) here +PRINTABLE_LATIN = set( + list(range(32, 126 + 1)) + list(range(161, 172 + 1)) + list(range(174, 255 + 1)) +) +BYTE_TO_BCHAR = { + b: chr(b) if b in PRINTABLE_LATIN else chr(256 + b) for b in range(256) +} +BCHAR_TO_BYTE = {bc: b for b, bc in BYTE_TO_BCHAR.items()} + + +def byte_encode(x: str) -> str: + normalized = WHITESPACE_NORMALIZER.sub(SPACE, x) + return "".join([BYTE_TO_BCHAR[b] for b in normalized.encode("utf-8")]) + + +def byte_decode(x: str) -> str: + try: + return bytes([BCHAR_TO_BYTE[bc] for bc in x]).decode("utf-8") + except ValueError: + return "" + + +def smart_byte_decode(x: str) -> str: + output = byte_decode(x) + if output == "": + # DP the best recovery (max valid chars) if it's broken + n_bytes = len(x) + f = [0 for _ in range(n_bytes + 1)] + pt = [0 for _ in range(n_bytes + 1)] + for i in range(1, n_bytes + 1): + f[i], pt[i] = f[i - 1], i - 1 + for j in range(1, min(4, i) + 1): + if f[i - j] + 1 > f[i] and len(byte_decode(x[i - j : i])) > 0: + f[i], pt[i] = f[i - j] + 1, i - j + cur_pt = n_bytes + while cur_pt > 0: + if f[cur_pt] == f[pt[cur_pt]] + 1: + output = byte_decode(x[pt[cur_pt] : cur_pt]) + output + cur_pt = pt[cur_pt] + return output diff --git a/fairseq/data/encoders/bytes.py b/fairseq/data/encoders/bytes.py new file mode 100644 index 0000000000000000000000000000000000000000..f88f8f6929f5b6bdb0db470be9ebedf8fe1f752d --- /dev/null +++ b/fairseq/data/encoders/bytes.py @@ -0,0 +1,34 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +from fairseq.data.encoders import register_bpe +from fairseq.data.encoders.byte_utils import ( + SPACE, + SPACE_ESCAPE, + byte_encode, + smart_byte_decode, +) + + +@register_bpe("bytes") +class Bytes(object): + def __init__(self, *unused): + pass + + @staticmethod + def add_args(parser): + pass + + @staticmethod + def encode(x: str) -> str: + encoded = byte_encode(x) + escaped = encoded.replace(SPACE, SPACE_ESCAPE) + return SPACE.join(list(escaped)) + + @staticmethod + def decode(x: str) -> str: + unescaped = x.replace(SPACE, "").replace(SPACE_ESCAPE, SPACE) + return smart_byte_decode(unescaped) diff --git a/fairseq/data/encoders/characters.py b/fairseq/data/encoders/characters.py new file mode 100644 index 0000000000000000000000000000000000000000..494ea219392716dc75d2c1e19d71cd55b9b2f4ba --- /dev/null +++ b/fairseq/data/encoders/characters.py @@ -0,0 +1,30 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +from fairseq.data.encoders import register_bpe + + +SPACE = chr(32) +SPACE_ESCAPE = chr(9601) + + +@register_bpe("characters") +class Characters(object): + def __init__(self, *unused): + pass + + @staticmethod + def add_args(parser): + pass + + @staticmethod + def encode(x: str) -> str: + escaped = x.replace(SPACE, SPACE_ESCAPE) + return SPACE.join(list(escaped)) + + @staticmethod + def decode(x: str) -> str: + return x.replace(SPACE, "").replace(SPACE_ESCAPE, SPACE) diff --git a/fairseq/data/encoders/fastbpe.py b/fairseq/data/encoders/fastbpe.py new file mode 100644 index 0000000000000000000000000000000000000000..f7c21039549ea002e73d1ad7cde5735f215f11ee --- /dev/null +++ b/fairseq/data/encoders/fastbpe.py @@ -0,0 +1,36 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from dataclasses import dataclass, field + +from fairseq import file_utils +from fairseq.data.encoders import register_bpe +from fairseq.dataclass import FairseqDataclass + + +@dataclass +class fastBPEConfig(FairseqDataclass): + bpe_codes: str = field(default="???", metadata={"help": "path to fastBPE BPE"}) + + +@register_bpe("fastbpe", dataclass=fastBPEConfig) +class fastBPE(object): + def __init__(self, cfg): + if cfg.bpe_codes is None: + raise ValueError("--bpe-codes is required for --bpe=fastbpe") + codes = file_utils.cached_path(cfg.bpe_codes) + try: + import fastBPE + + self.bpe = fastBPE.fastBPE(codes) + self.bpe_symbol = "@@ " + except ImportError: + raise ImportError("Please install fastBPE with: pip install fastBPE") + + def encode(self, x: str) -> str: + return self.bpe.apply([x])[0] + + def decode(self, x: str) -> str: + return (x + " ").replace(self.bpe_symbol, "").rstrip() diff --git a/fairseq/data/encoders/gpt2_bpe.py b/fairseq/data/encoders/gpt2_bpe.py new file mode 100644 index 0000000000000000000000000000000000000000..e661426a73c7e735f7054bcb04281bf1649bb46c --- /dev/null +++ b/fairseq/data/encoders/gpt2_bpe.py @@ -0,0 +1,45 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from dataclasses import dataclass, field + +from fairseq import file_utils +from fairseq.data.encoders import register_bpe +from fairseq.dataclass import FairseqDataclass + +from .gpt2_bpe_utils import get_encoder + + +DEFAULT_ENCODER_JSON = "https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/encoder.json" +DEFAULT_VOCAB_BPE = "https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/vocab.bpe" + + +@dataclass +class GPT2BPEConfig(FairseqDataclass): + gpt2_encoder_json: str = field( + default=DEFAULT_ENCODER_JSON, metadata={"help": "path to encoder.json"} + ) + gpt2_vocab_bpe: str = field( + default=DEFAULT_VOCAB_BPE, metadata={"help": "path to vocab.bpe"} + ) + + +@register_bpe("gpt2", dataclass=GPT2BPEConfig) +class GPT2BPE(object): + def __init__(self, cfg): + encoder_json = file_utils.cached_path(cfg.gpt2_encoder_json) + vocab_bpe = file_utils.cached_path(cfg.gpt2_vocab_bpe) + self.bpe = get_encoder(encoder_json, vocab_bpe) + + def encode(self, x: str) -> str: + return " ".join(map(str, self.bpe.encode(x))) + + def decode(self, x: str) -> str: + return self.bpe.decode( + [int(tok) if tok not in {"<unk>", "<mask>"} else tok for tok in x.split()] + ) + + def is_beginning_of_word(self, x: str) -> bool: + return self.decode(x).startswith(" ") diff --git a/fairseq/data/encoders/gpt2_bpe_utils.py b/fairseq/data/encoders/gpt2_bpe_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..688d4e36e358df2dcc432d37d3e57bd81e2f1ed1 --- /dev/null +++ b/fairseq/data/encoders/gpt2_bpe_utils.py @@ -0,0 +1,140 @@ +""" +Byte pair encoding utilities from GPT-2. + +Original source: https://github.com/openai/gpt-2/blob/master/src/encoder.py +Original license: MIT +""" + +import json +from functools import lru_cache + + +@lru_cache() +def bytes_to_unicode(): + """ + Returns list of utf-8 byte and a corresponding list of unicode strings. + The reversible bpe codes work on unicode strings. + This means you need a large # of unicode characters in your vocab if you want to avoid UNKs. + When you're at something like a 10B token dataset you end up needing around 5K for decent coverage. + This is a signficant percentage of your normal, say, 32K bpe vocab. + To avoid that, we want lookup tables between utf-8 bytes and unicode strings. + And avoids mapping to whitespace/control characters the bpe code barfs on. + """ + bs = ( + list(range(ord("!"), ord("~") + 1)) + + list(range(ord("¡"), ord("¬") + 1)) + + list(range(ord("®"), ord("ÿ") + 1)) + ) + cs = bs[:] + n = 0 + for b in range(2 ** 8): + if b not in bs: + bs.append(b) + cs.append(2 ** 8 + n) + n += 1 + cs = [chr(n) for n in cs] + return dict(zip(bs, cs)) + + +def get_pairs(word): + """Return set of symbol pairs in a word. + Word is represented as tuple of symbols (symbols being variable-length strings). + """ + pairs = set() + prev_char = word[0] + for char in word[1:]: + pairs.add((prev_char, char)) + prev_char = char + return pairs + + +class Encoder: + def __init__(self, encoder, bpe_merges, errors="replace"): + self.encoder = encoder + self.decoder = {v: k for k, v in self.encoder.items()} + self.errors = errors # how to handle errors in decoding + self.byte_encoder = bytes_to_unicode() + self.byte_decoder = {v: k for k, v in self.byte_encoder.items()} + self.bpe_ranks = dict(zip(bpe_merges, range(len(bpe_merges)))) + self.cache = {} + + try: + import regex as re + + self.re = re + except ImportError: + raise ImportError("Please install regex with: pip install regex") + + # Should haved added re.IGNORECASE so BPE merges can happen for capitalized versions of contractions + self.pat = self.re.compile( + r"""'s|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+""" + ) + + def bpe(self, token): + if token in self.cache: + return self.cache[token] + word = tuple(token) + pairs = get_pairs(word) + + if not pairs: + return token + + while True: + bigram = min(pairs, key=lambda pair: self.bpe_ranks.get(pair, float("inf"))) + if bigram not in self.bpe_ranks: + break + first, second = bigram + new_word = [] + i = 0 + while i < len(word): + try: + j = word.index(first, i) + new_word.extend(word[i:j]) + i = j + except: + new_word.extend(word[i:]) + break + + if word[i] == first and i < len(word) - 1 and word[i + 1] == second: + new_word.append(first + second) + i += 2 + else: + new_word.append(word[i]) + i += 1 + new_word = tuple(new_word) + word = new_word + if len(word) == 1: + break + else: + pairs = get_pairs(word) + word = " ".join(word) + self.cache[token] = word + return word + + def encode(self, text): + bpe_tokens = [] + for token in self.re.findall(self.pat, text): + token = "".join(self.byte_encoder[b] for b in token.encode("utf-8")) + bpe_tokens.extend( + self.encoder[bpe_token] for bpe_token in self.bpe(token).split(" ") + ) + return bpe_tokens + + def decode(self, tokens): + text = "".join([self.decoder.get(token, token) for token in tokens]) + text = bytearray([self.byte_decoder[c] for c in text]).decode( + "utf-8", errors=self.errors + ) + return text + + +def get_encoder(encoder_json_path, vocab_bpe_path): + with open(encoder_json_path, "r") as f: + encoder = json.load(f) + with open(vocab_bpe_path, "r", encoding="utf-8") as f: + bpe_data = f.read() + bpe_merges = [tuple(merge_str.split()) for merge_str in bpe_data.split("\n")[1:-1]] + return Encoder( + encoder=encoder, + bpe_merges=bpe_merges, + ) diff --git a/fairseq/data/encoders/hf_bert_bpe.py b/fairseq/data/encoders/hf_bert_bpe.py new file mode 100644 index 0000000000000000000000000000000000000000..a41c059343ec7e2914b2c9d2f53f526c33f9659d --- /dev/null +++ b/fairseq/data/encoders/hf_bert_bpe.py @@ -0,0 +1,50 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from dataclasses import dataclass, field +from typing import Optional + +from fairseq.data.encoders import register_bpe +from fairseq.dataclass import FairseqDataclass + + +@dataclass +class BertBPEConfig(FairseqDataclass): + bpe_cased: bool = field(default=False, metadata={"help": "set for cased BPE"}) + bpe_vocab_file: Optional[str] = field( + default=None, metadata={"help": "bpe vocab file"} + ) + + +@register_bpe("bert", dataclass=BertBPEConfig) +class BertBPE(object): + def __init__(self, cfg): + try: + from transformers import BertTokenizer + except ImportError: + raise ImportError( + "Please install transformers with: pip install transformers" + ) + + if cfg.bpe_vocab_file: + self.bert_tokenizer = BertTokenizer( + cfg.bpe_vocab_file, do_lower_case=not cfg.bpe_cased + ) + else: + vocab_file_name = ( + "bert-base-cased" if cfg.bpe_cased else "bert-base-uncased" + ) + self.bert_tokenizer = BertTokenizer.from_pretrained(vocab_file_name) + + def encode(self, x: str) -> str: + return " ".join(self.bert_tokenizer.tokenize(x)) + + def decode(self, x: str) -> str: + return self.bert_tokenizer.clean_up_tokenization( + self.bert_tokenizer.convert_tokens_to_string(x.split(" ")) + ) + + def is_beginning_of_word(self, x: str) -> bool: + return not x.startswith("##") diff --git a/fairseq/data/encoders/hf_byte_bpe.py b/fairseq/data/encoders/hf_byte_bpe.py new file mode 100644 index 0000000000000000000000000000000000000000..c508578d41bf6b7ce0a847e0797d71b19beb393d --- /dev/null +++ b/fairseq/data/encoders/hf_byte_bpe.py @@ -0,0 +1,50 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from dataclasses import dataclass, field + +from fairseq.data.encoders import register_bpe +from fairseq.dataclass import FairseqDataclass +from fairseq import file_utils + + +@dataclass +class HuggingFaceByteLevelBPEConfig(FairseqDataclass): + bpe_merges: str = field(default="???", metadata={"help": "path to merges.txt"}) + bpe_vocab: str = field(default="???", metadata={"help": "path to vocab.json"}) + bpe_add_prefix_space: bool = field( + default=False, metadata={"help": "add prefix space before encoding"} + ) + + +@register_bpe("hf_byte_bpe", dataclass=HuggingFaceByteLevelBPEConfig) +class HuggingFaceByteLevelBPE(object): + def __init__(self, cfg): + try: + from tokenizers import ByteLevelBPETokenizer + except ImportError: + raise ImportError( + "Please install huggingface/tokenizers with: " "pip install tokenizers" + ) + + bpe_vocab = file_utils.cached_path(cfg.bpe_vocab) + bpe_merges = file_utils.cached_path(cfg.bpe_merges) + + self.bpe = ByteLevelBPETokenizer( + bpe_vocab, + bpe_merges, + add_prefix_space=cfg.bpe_add_prefix_space, + ) + + def encode(self, x: str) -> str: + return " ".join(map(str, self.bpe.encode(x).ids)) + + def decode(self, x: str) -> str: + return self.bpe.decode( + [int(tok) if tok not in {"<unk>", "<mask>"} else tok for tok in x.split()] + ) + + def is_beginning_of_word(self, x: str) -> bool: + return self.decode(x).startswith(" ") diff --git a/fairseq/data/encoders/moses_tokenizer.py b/fairseq/data/encoders/moses_tokenizer.py new file mode 100644 index 0000000000000000000000000000000000000000..e236dad167a037a8ed95f7fc8292b27b10d580b0 --- /dev/null +++ b/fairseq/data/encoders/moses_tokenizer.py @@ -0,0 +1,49 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from dataclasses import dataclass, field + +from fairseq.data.encoders import register_tokenizer +from fairseq.dataclass import FairseqDataclass + + +@dataclass +class MosesTokenizerConfig(FairseqDataclass): + source_lang: str = field(default="en", metadata={"help": "source language"}) + target_lang: str = field(default="en", metadata={"help": "target language"}) + moses_no_dash_splits: bool = field( + default=False, metadata={"help": "don't apply dash split rules"} + ) + moses_no_escape: bool = field( + default=False, + metadata={"help": "don't perform HTML escaping on apostrophe, quotes, etc."}, + ) + + +@register_tokenizer("moses", dataclass=MosesTokenizerConfig) +class MosesTokenizer(object): + def __init__(self, cfg: MosesTokenizerConfig): + self.cfg = cfg + + try: + from sacremoses import MosesTokenizer, MosesDetokenizer + + self.tok = MosesTokenizer(cfg.source_lang) + self.detok = MosesDetokenizer(cfg.target_lang) + except ImportError: + raise ImportError( + "Please install Moses tokenizer with: pip install sacremoses" + ) + + def encode(self, x: str) -> str: + return self.tok.tokenize( + x, + aggressive_dash_splits=(not self.cfg.moses_no_dash_splits), + return_str=True, + escape=(not self.cfg.moses_no_escape), + ) + + def decode(self, x: str) -> str: + return self.detok.detokenize(x.split()) diff --git a/fairseq/data/encoders/nltk_tokenizer.py b/fairseq/data/encoders/nltk_tokenizer.py new file mode 100644 index 0000000000000000000000000000000000000000..0ab92377b3a23bb48384c3f7acf299612e8b0775 --- /dev/null +++ b/fairseq/data/encoders/nltk_tokenizer.py @@ -0,0 +1,24 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq.data.encoders import register_tokenizer +from fairseq.dataclass import FairseqDataclass + + +@register_tokenizer("nltk", dataclass=FairseqDataclass) +class NLTKTokenizer(object): + def __init__(self, *unused): + try: + from nltk.tokenize import word_tokenize + + self.word_tokenize = word_tokenize + except ImportError: + raise ImportError("Please install nltk with: pip install nltk") + + def encode(self, x: str) -> str: + return " ".join(self.word_tokenize(x)) + + def decode(self, x: str) -> str: + return x diff --git a/fairseq/data/encoders/sentencepiece_bpe.py b/fairseq/data/encoders/sentencepiece_bpe.py new file mode 100644 index 0000000000000000000000000000000000000000..a76d46a2014e81eff72b19f6c13084a855fcd477 --- /dev/null +++ b/fairseq/data/encoders/sentencepiece_bpe.py @@ -0,0 +1,48 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from dataclasses import dataclass, field + +from fairseq import file_utils +from fairseq.data.encoders import register_bpe +from fairseq.dataclass import FairseqDataclass + + +@dataclass +class SentencepieceConfig(FairseqDataclass): + sentencepiece_model: str = field( + default="???", metadata={"help": "path to sentencepiece model"} + ) + + +@register_bpe("sentencepiece", dataclass=SentencepieceConfig) +class SentencepieceBPE(object): + def __init__(self, cfg): + sentencepiece_model = file_utils.cached_path(cfg.sentencepiece_model) + try: + import sentencepiece as spm + + self.sp = spm.SentencePieceProcessor() + self.sp.Load(sentencepiece_model) + except ImportError: + raise ImportError( + "Please install sentencepiece with: pip install sentencepiece" + ) + + def encode(self, x: str) -> str: + return " ".join(self.sp.EncodeAsPieces(x)) + + def decode(self, x: str) -> str: + return x.replace(" ", "").replace("\u2581", " ").strip() + + def is_beginning_of_word(self, x: str) -> bool: + if x in ["<unk>", "<s>", "</s>", "<pad>"]: + # special elements are always considered beginnings + # HACK: this logic is already present in fairseq/tasks/masked_lm.py + # but these special tokens are also contained in the sentencepiece + # vocabulary which causes duplicate special tokens. This hack makes + # sure that they are all taken into account. + return True + return x.startswith("\u2581") diff --git a/fairseq/data/encoders/space_tokenizer.py b/fairseq/data/encoders/space_tokenizer.py new file mode 100644 index 0000000000000000000000000000000000000000..925ad41b7c1aee6738c63938c36bd3ee16dca812 --- /dev/null +++ b/fairseq/data/encoders/space_tokenizer.py @@ -0,0 +1,21 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import re + +from fairseq.data.encoders import register_tokenizer +from fairseq.dataclass import FairseqDataclass + + +@register_tokenizer("space", dataclass=FairseqDataclass) +class SpaceTokenizer(object): + def __init__(self, *unused): + self.space_tok = re.compile(r"\s+") + + def encode(self, x: str) -> str: + return self.space_tok.sub(" ", x) + + def decode(self, x: str) -> str: + return x diff --git a/fairseq/data/encoders/subword_nmt_bpe.py b/fairseq/data/encoders/subword_nmt_bpe.py new file mode 100644 index 0000000000000000000000000000000000000000..5d724d2730a5895ca55af2998c2ced471625b516 --- /dev/null +++ b/fairseq/data/encoders/subword_nmt_bpe.py @@ -0,0 +1,54 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from dataclasses import dataclass, field + +from fairseq import file_utils +from fairseq.data.encoders import register_bpe +from fairseq.dataclass import FairseqDataclass + + +@dataclass +class SubwordNMTBPEConfig(FairseqDataclass): + bpe_codes: str = field(default="???", metadata={"help": "path to subword NMT BPE"}) + bpe_separator: str = field(default="@@", metadata={"help": "BPE separator"}) + + +@register_bpe("subword_nmt", dataclass=SubwordNMTBPEConfig) +class SubwordNMTBPE(object): + def __init__(self, cfg): + if cfg.bpe_codes is None: + raise ValueError("--bpe-codes is required for --bpe=subword_nmt") + codes = file_utils.cached_path(cfg.bpe_codes) + try: + from subword_nmt import apply_bpe + + bpe_parser = apply_bpe.create_parser() + bpe_args = bpe_parser.parse_args( + [ + "--codes", + codes, + "--separator", + cfg.bpe_separator, + ] + ) + self.bpe = apply_bpe.BPE( + bpe_args.codes, + bpe_args.merges, + bpe_args.separator, + None, + bpe_args.glossaries, + ) + self.bpe_symbol = bpe_args.separator + " " + except ImportError: + raise ImportError( + "Please install subword_nmt with: pip install subword-nmt" + ) + + def encode(self, x: str) -> str: + return self.bpe.process_line(x) + + def decode(self, x: str) -> str: + return (x + " ").replace(self.bpe_symbol, "").rstrip() diff --git a/fairseq/data/encoders/utils.py b/fairseq/data/encoders/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..d93eb532ef84f0e2bc708b777229ab2cb76ca14b --- /dev/null +++ b/fairseq/data/encoders/utils.py @@ -0,0 +1,30 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +from fairseq.data import encoders + + +def get_whole_word_mask(args, dictionary): + bpe = encoders.build_bpe(args) + if bpe is not None: + + def is_beginning_of_word(i): + if i < dictionary.nspecial: + # special elements are always considered beginnings + return True + tok = dictionary[i] + if tok.startswith("madeupword"): + return True + try: + return bpe.is_beginning_of_word(tok) + except ValueError: + return True + + mask_whole_words = torch.ByteTensor( + list(map(is_beginning_of_word, range(len(dictionary)))) + ) + return mask_whole_words + return None diff --git a/fairseq/data/fairseq_dataset.py b/fairseq/data/fairseq_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..23e6992dbaf34e52f2fdcd0c8fc418c93744ea4e --- /dev/null +++ b/fairseq/data/fairseq_dataset.py @@ -0,0 +1,205 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import numpy as np +import torch.utils.data +from fairseq.data import data_utils + +logger = logging.getLogger(__name__) + + +class EpochListening: + """Mixin for receiving updates whenever the epoch increments.""" + + @property + def can_reuse_epoch_itr_across_epochs(self): + """ + Whether we can reuse the :class:`fairseq.data.EpochBatchIterator` for + this dataset across epochs. + + This needs to return ``False`` if the sample sizes can change across + epochs, in which case we may need to regenerate batches at each epoch. + If your dataset relies in ``set_epoch`` then you should consider setting + this to ``False``. + """ + return True + + def set_epoch(self, epoch): + """Will receive the updated epoch number at the beginning of the epoch.""" + pass + + +class FairseqDataset(torch.utils.data.Dataset, EpochListening): + """A dataset that provides helpers for batching.""" + + def __getitem__(self, index): + raise NotImplementedError + + def __len__(self): + raise NotImplementedError + + def collater(self, samples): + """Merge a list of samples to form a mini-batch. + + Args: + samples (List[dict]): samples to collate + + Returns: + dict: a mini-batch suitable for forwarding with a Model + """ + raise NotImplementedError + + def num_tokens(self, index): + """Return the number of tokens in a sample. This value is used to + enforce ``--max-tokens`` during batching.""" + raise NotImplementedError + + def num_tokens_vec(self, indices): + """Return the number of tokens for a set of positions defined by indices. + This value is used to enforce ``--max-tokens`` during batching.""" + raise NotImplementedError + + def size(self, index): + """Return an example's size as a float or tuple. This value is used when + filtering a dataset with ``--max-positions``.""" + raise NotImplementedError + + def ordered_indices(self): + """Return an ordered list of indices. Batches will be constructed based + on this order.""" + return np.arange(len(self), dtype=np.int64) + + @property + def supports_prefetch(self): + """Whether this dataset supports prefetching.""" + return False + + def attr(self, attr: str, index: int): + return getattr(self, attr, None) + + def prefetch(self, indices): + """Prefetch the data required for this epoch.""" + raise NotImplementedError + + def get_batch_shapes(self): + """ + Return a list of valid batch shapes, for example:: + + [(8, 512), (16, 256), (32, 128)] + + The first dimension of each tuple is the batch size and can be ``None`` + to automatically infer the max batch size based on ``--max-tokens``. + The second dimension of each tuple is the max supported length as given + by :func:`fairseq.data.FairseqDataset.num_tokens`. + + This will be used by :func:`fairseq.data.FairseqDataset.batch_by_size` + to restrict batch shapes. This is useful on TPUs to avoid too many + dynamic shapes (and recompilations). + """ + return None + + def batch_by_size( + self, + indices, + max_tokens=None, + max_sentences=None, + required_batch_size_multiple=1, + ): + """ + Given an ordered set of indices, return batches according to + *max_tokens*, *max_sentences* and *required_batch_size_multiple*. + """ + from fairseq.data import data_utils + + fixed_shapes = self.get_batch_shapes() + if fixed_shapes is not None: + + def adjust_bsz(bsz, num_tokens): + if bsz is None: + assert max_tokens is not None, "Must specify --max-tokens" + bsz = max_tokens // num_tokens + if max_sentences is not None: + bsz = min(bsz, max_sentences) + elif ( + bsz >= required_batch_size_multiple + and bsz % required_batch_size_multiple != 0 + ): + bsz -= bsz % required_batch_size_multiple + return bsz + + fixed_shapes = np.array( + [ + [adjust_bsz(bsz, num_tokens), num_tokens] + for (bsz, num_tokens) in fixed_shapes + ] + ) + + try: + num_tokens_vec = self.num_tokens_vec(indices).astype('int64') + except NotImplementedError: + num_tokens_vec = None + + return data_utils.batch_by_size( + indices, + num_tokens_fn=self.num_tokens, + num_tokens_vec=num_tokens_vec, + max_tokens=max_tokens, + max_sentences=max_sentences, + required_batch_size_multiple=required_batch_size_multiple, + fixed_shapes=fixed_shapes, + ) + + def filter_indices_by_size(self, indices, max_sizes): + """ + Filter a list of sample indices. Remove those that are longer than + specified in *max_sizes*. + + WARNING: don't update, override method in child classes + + Args: + indices (np.array): original array of sample indices + max_sizes (int or list[int] or tuple[int]): max sample size, + can be defined separately for src and tgt (then list or tuple) + + Returns: + np.array: filtered sample array + list: list of removed indices + """ + if isinstance(max_sizes, float) or isinstance(max_sizes, int): + if hasattr(self, "sizes") and isinstance(self.sizes, np.ndarray): + ignored = indices[self.sizes[indices] > max_sizes].tolist() + indices = indices[self.sizes[indices] <= max_sizes] + elif ( + hasattr(self, "sizes") + and isinstance(self.sizes, list) + and len(self.sizes) == 1 + ): + ignored = indices[self.sizes[0][indices] > max_sizes].tolist() + indices = indices[self.sizes[0][indices] <= max_sizes] + else: + indices, ignored = data_utils._filter_by_size_dynamic( + indices, self.size, max_sizes + ) + else: + indices, ignored = data_utils._filter_by_size_dynamic( + indices, self.size, max_sizes + ) + return indices, ignored + + @property + def supports_fetch_outside_dataloader(self): + """Whether this dataset supports fetching outside the workers of the dataloader.""" + return True + + +class FairseqIterableDataset(torch.utils.data.IterableDataset, EpochListening): + """ + For datasets that need to be read sequentially, usually because the data is + being streamed or otherwise can't be manipulated on a single machine. + """ + + def __iter__(self): + raise NotImplementedError diff --git a/fairseq/data/fasta_dataset.py b/fairseq/data/fasta_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..007011974a997fd7446dd29d7eba097d7513bab0 --- /dev/null +++ b/fairseq/data/fasta_dataset.py @@ -0,0 +1,107 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import os +import subprocess +import threading +from pathlib import Path + +import numpy as np +import torch + + +def fasta_file_path(prefix_path): + return prefix_path + ".fasta" + + +class FastaDataset(torch.utils.data.Dataset): + """ + For loading protein sequence datasets in the common FASTA data format + """ + + def __init__(self, path: str, cache_indices=False): + self.fn = fasta_file_path(path) + self.threadlocal = threading.local() + self.cache = Path(f"{path}.fasta.idx.npy") + if cache_indices: + if self.cache.exists(): + self.offsets, self.sizes = np.load(self.cache) + else: + self.offsets, self.sizes = self._build_index(path) + np.save(self.cache, np.stack([self.offsets, self.sizes])) + else: + self.offsets, self.sizes = self._build_index(path) + + def _get_file(self): + if not hasattr(self.threadlocal, "f"): + self.threadlocal.f = open(self.fn, "r") + return self.threadlocal.f + + def __getitem__(self, idx): + f = self._get_file() + f.seek(self.offsets[idx]) + desc = f.readline().strip() + line = f.readline() + seq = "" + while line != "" and line[0] != ">": + seq += line.strip() + line = f.readline() + return desc, seq + + def __len__(self): + return self.offsets.size + + def _build_index(self, path: str): + # Use grep and awk to get 100M/s on local SSD. + # Should process your enormous 100G fasta in ~10 min single core... + path = fasta_file_path(path) + bytes_offsets = subprocess.check_output( + f"cat {path} | tqdm --bytes --total $(wc -c < {path})" + "| grep --byte-offset '^>' -o | cut -d: -f1", + shell=True, + ) + fasta_lengths = subprocess.check_output( + f"cat {path} | tqdm --bytes --total $(wc -c < {path})" + "| awk '/^>/ {print \"\";next;} { printf(\"%s\",$0);}' | tail -n+2 | awk '{print length($1)}'", + shell=True, + ) + bytes_np = np.fromstring(bytes_offsets, dtype=np.int64, sep=" ") + sizes_np = np.fromstring(fasta_lengths, dtype=np.int64, sep=" ") + return bytes_np, sizes_np + + def __setstate__(self, state): + self.__dict__ = state + self.threadlocal = threading.local() + + def __getstate__(self): + d = {} + for i, v in self.__dict__.items(): + if i != "threadlocal": + d[i] = v + return d + + def __del__(self): + if hasattr(self.threadlocal, "f"): + self.threadlocal.f.close() + del self.threadlocal.f + + @staticmethod + def exists(path): + return os.path.exists(fasta_file_path(path)) + + +class EncodedFastaDataset(FastaDataset): + """ + The FastaDataset returns raw sequences - this allows us to return + indices with a dictionary instead. + """ + + def __init__(self, path, dictionary): + super().__init__(path, cache_indices=True) + self.dictionary = dictionary + + def __getitem__(self, idx): + desc, seq = super().__getitem__(idx) + return self.dictionary.encode_line(seq, line_tokenizer=list).long() diff --git a/fairseq/data/id_dataset.py b/fairseq/data/id_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..3e4d7969cf2a26e852b466f165a6fadabae3b35f --- /dev/null +++ b/fairseq/data/id_dataset.py @@ -0,0 +1,19 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch + +from . import FairseqDataset + + +class IdDataset(FairseqDataset): + def __getitem__(self, index): + return index + + def __len__(self): + return 0 + + def collater(self, samples): + return torch.tensor(samples) diff --git a/fairseq/data/indexed_dataset.py b/fairseq/data/indexed_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..802e37a7ff849e435d4fa89ad7609c17cedd1980 --- /dev/null +++ b/fairseq/data/indexed_dataset.py @@ -0,0 +1,576 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import shutil +import struct +from functools import lru_cache + +import numpy as np +import torch +from fairseq.dataclass.constants import DATASET_IMPL_CHOICES +from fairseq.data.fasta_dataset import FastaDataset +from fairseq.file_io import PathManager + +from . import FairseqDataset + +from typing import Union + + +def best_fitting_int_dtype( + max_int_to_represent, +) -> Union[np.uint16, np.uint32, np.int64]: + + if max_int_to_represent is None: + return np.uint32 # Safe guess + elif max_int_to_represent < 65500: + return np.uint16 + elif max_int_to_represent < 4294967295: + return np.uint32 + else: + return np.int64 + # we avoid np.uint64 because it doesn't save space and its type promotion behaves unexpectedly + # https://github.com/numpy/numpy/issues/5745 + + +def get_available_dataset_impl(): + return list(map(str, DATASET_IMPL_CHOICES)) + + +def infer_dataset_impl(path): + if IndexedRawTextDataset.exists(path): + return "raw" + elif IndexedDataset.exists(path): + with open(index_file_path(path), "rb") as f: + magic = f.read(8) + if magic == IndexedDataset._HDR_MAGIC: + return "cached" + elif magic == MMapIndexedDataset.Index._HDR_MAGIC[:8]: + return "mmap" + else: + return None + elif FastaDataset.exists(path): + return "fasta" + else: + return None + + +def make_builder(out_file, impl, vocab_size=None): + if impl == "mmap": + return MMapIndexedDatasetBuilder( + out_file, dtype=best_fitting_int_dtype(vocab_size) + ) + elif impl == "fasta": + raise NotImplementedError + else: + return IndexedDatasetBuilder(out_file) + + +def make_dataset(path, impl, fix_lua_indexing=False, dictionary=None): + if impl == "raw" and IndexedRawTextDataset.exists(path): + assert dictionary is not None + return IndexedRawTextDataset(path, dictionary) + elif impl == "lazy" and IndexedDataset.exists(path): + return IndexedDataset(path, fix_lua_indexing=fix_lua_indexing) + elif impl == "cached" and IndexedDataset.exists(path): + return IndexedCachedDataset(path, fix_lua_indexing=fix_lua_indexing) + elif impl == "mmap" and MMapIndexedDataset.exists(path): + return MMapIndexedDataset(path) + elif impl == "fasta" and FastaDataset.exists(path): + from fairseq.data.fasta_dataset import EncodedFastaDataset + + return EncodedFastaDataset(path, dictionary) + return None + + +def dataset_exists(path, impl): + if impl == "raw": + return IndexedRawTextDataset.exists(path) + elif impl == "mmap": + return MMapIndexedDataset.exists(path) + else: + return IndexedDataset.exists(path) + + +def read_longs(f, n): + a = np.empty(n, dtype=np.int64) + f.readinto(a) + return a + + +def write_longs(f, a): + f.write(np.array(a, dtype=np.int64)) + + +_code_to_dtype = { + 1: np.uint8, + 2: np.int8, + 3: np.int16, + 4: np.int32, + 5: np.int64, + 6: np.float, + 7: np.double, + 8: np.uint16, + 9: np.uint32, + 10: np.uint64, +} + + +def _dtype_header_code(dtype) -> int: + for k in _code_to_dtype.keys(): + if _code_to_dtype[k] == dtype: + return k + raise ValueError(dtype) + + +def index_file_path(prefix_path): + return prefix_path + ".idx" + + +def data_file_path(prefix_path): + return prefix_path + ".bin" + + +class IndexedDataset(FairseqDataset): + """Loader for TorchNet IndexedDataset""" + + _HDR_MAGIC = b"TNTIDX\x00\x00" + + def __init__(self, path, fix_lua_indexing=False): + super().__init__() + self.path = path + self.fix_lua_indexing = fix_lua_indexing + self.data_file = None + self.read_index(path) + + def read_index(self, path): + with open(index_file_path(path), "rb") as f: + magic = f.read(8) + assert magic == self._HDR_MAGIC, ( + "Index file doesn't match expected format. " + "Make sure that --dataset-impl is configured properly." + ) + version = f.read(8) + assert struct.unpack("<Q", version) == (1,) + code, self.element_size = struct.unpack("<QQ", f.read(16)) + self.dtype = _code_to_dtype[code] + self._len, self.s = struct.unpack("<QQ", f.read(16)) + self.dim_offsets = read_longs(f, self._len + 1) + self.data_offsets = read_longs(f, self._len + 1) + self.sizes = read_longs(f, self.s) + + def read_data(self, path): + self.data_file = open(data_file_path(path), "rb", buffering=0) + + def check_index(self, i): + if i < 0 or i >= self._len: + raise IndexError("index out of range") + + def __del__(self): + if self.data_file: + self.data_file.close() + + @lru_cache(maxsize=8) + def __getitem__(self, i) -> torch.Tensor: + if not self.data_file: + self.read_data(self.path) + self.check_index(i) + tensor_size = self.sizes[self.dim_offsets[i] : self.dim_offsets[i + 1]] + a = np.empty(tensor_size, dtype=self.dtype) + self.data_file.seek(self.data_offsets[i] * self.element_size) + self.data_file.readinto(a) + item = torch.from_numpy(a).long() + if self.fix_lua_indexing: + item -= 1 # subtract 1 for 0-based indexing + return item + + def __len__(self): + return self._len + + def num_tokens(self, index): + return self.sizes[index] + + def size(self, index): + return self.sizes[index] + + @staticmethod + def exists(path): + return PathManager.exists(index_file_path(path)) and PathManager.exists( + data_file_path(path) + ) + + @property + def supports_prefetch(self): + return False # avoid prefetching to save memory + + +class IndexedCachedDataset(IndexedDataset): + def __init__(self, path, fix_lua_indexing=False): + super().__init__(path, fix_lua_indexing=fix_lua_indexing) + self.cache = None + self.cache_index = {} + + @property + def supports_prefetch(self): + return True + + def prefetch(self, indices): + if all(i in self.cache_index for i in indices): + return + if not self.data_file: + self.read_data(self.path) + indices = sorted(set(indices)) + total_size = 0 + for i in indices: + total_size += self.data_offsets[i + 1] - self.data_offsets[i] + self.cache = np.empty(total_size, dtype=self.dtype) + ptx = 0 + self.cache_index.clear() + for i in indices: + self.cache_index[i] = ptx + size = self.data_offsets[i + 1] - self.data_offsets[i] + a = self.cache[ptx : ptx + size] + self.data_file.seek(self.data_offsets[i] * self.element_size) + self.data_file.readinto(a) + ptx += size + if self.data_file: + # close and delete data file after prefetch so we can pickle + self.data_file.close() + self.data_file = None + + @lru_cache(maxsize=8) + def __getitem__(self, i): + self.check_index(i) + tensor_size = self.sizes[self.dim_offsets[i] : self.dim_offsets[i + 1]] + a = np.empty(tensor_size, dtype=self.dtype) + ptx = self.cache_index[i] + np.copyto(a, self.cache[ptx : ptx + a.size]) + item = torch.from_numpy(a).long() + if self.fix_lua_indexing: + item -= 1 # subtract 1 for 0-based indexing + return item + + +class IndexedRawTextDataset(FairseqDataset): + """Takes a text file as input and binarizes it in memory at instantiation. + Original lines are also kept in memory""" + + def __init__(self, path, dictionary, append_eos=True, reverse_order=False): + self.tokens_list = [] + self.lines = [] + self.sizes = [] + self.append_eos = append_eos + self.reverse_order = reverse_order + self.read_data(path, dictionary) + self.size = len(self.tokens_list) + + def read_data(self, path, dictionary): + with open(path, "r", encoding="utf-8") as f: + for line in f: + self.lines.append(line.strip("\n")) + tokens = dictionary.encode_line( + line, + add_if_not_exist=False, + append_eos=self.append_eos, + reverse_order=self.reverse_order, + ).long() + self.tokens_list.append(tokens) + self.sizes.append(len(tokens)) + self.sizes = np.array(self.sizes) + + def check_index(self, i): + if i < 0 or i >= self.size: + raise IndexError("index out of range") + + @lru_cache(maxsize=8) + def __getitem__(self, i): + self.check_index(i) + return self.tokens_list[i] + + def get_original_text(self, i): + self.check_index(i) + return self.lines[i] + + def __del__(self): + pass + + def __len__(self): + return self.size + + def num_tokens(self, index): + return self.sizes[index] + + def size(self, index): + return self.sizes[index] + + @staticmethod + def exists(path): + return PathManager.exists(path) + + +class IndexedDatasetBuilder: + element_sizes = { + np.uint8: 1, + np.int8: 1, + np.int16: 2, + np.int32: 4, + np.int64: 8, + np.float: 4, + np.double: 8, + } + + def __init__(self, out_file, dtype=np.int32): + self.out_file = open(out_file, "wb") + self.dtype = dtype + self.data_offsets = [0] + self.dim_offsets = [0] + self.sizes = [] + self.element_size = self.element_sizes[self.dtype] + + def add_item(self, tensor): + # +1 for Lua compatibility + bytes = self.out_file.write(np.array(tensor.numpy() + 1, dtype=self.dtype)) + self.data_offsets.append(self.data_offsets[-1] + bytes / self.element_size) + for s in tensor.size(): + self.sizes.append(s) + self.dim_offsets.append(self.dim_offsets[-1] + len(tensor.size())) + + def merge_file_(self, another_file): + index = IndexedDataset(another_file) + assert index.dtype == self.dtype + + begin = self.data_offsets[-1] + for offset in index.data_offsets[1:]: + self.data_offsets.append(begin + offset) + self.sizes.extend(index.sizes) + begin = self.dim_offsets[-1] + for dim_offset in index.dim_offsets[1:]: + self.dim_offsets.append(begin + dim_offset) + + with open(data_file_path(another_file), "rb") as f: + while True: + data = f.read(1024) + if data: + self.out_file.write(data) + else: + break + + def finalize(self, index_file): + self.out_file.close() + index = open(index_file, "wb") + index.write(b"TNTIDX\x00\x00") + index.write(struct.pack("<Q", 1)) + index.write( + struct.pack("<QQ", _dtype_header_code(self.dtype), self.element_size) + ) + index.write(struct.pack("<QQ", len(self.data_offsets) - 1, len(self.sizes))) + write_longs(index, self.dim_offsets) + write_longs(index, self.data_offsets) + write_longs(index, self.sizes) + index.close() + + +def _warmup_mmap_file(path): + with open(path, "rb") as stream: + while stream.read(100 * 1024 * 1024): + pass + + +class MMapIndexedDataset(torch.utils.data.Dataset): + class Index: + _HDR_MAGIC = b"MMIDIDX\x00\x00" + + @classmethod + def writer(cls, path, dtype): + class _Writer: + def __enter__(self): + self._file = open(path, "wb") + + self._file.write(cls._HDR_MAGIC) + self._file.write(struct.pack("<Q", 1)) + self._file.write(struct.pack("<B", _dtype_header_code(dtype))) + + return self + + @staticmethod + def _get_pointers(sizes): + dtype_size = dtype().itemsize + address = 0 + pointers = [] + + for size in sizes: + pointers.append(address) + address += size * dtype_size + + return pointers + + def write(self, sizes): + pointers = self._get_pointers(sizes) + + self._file.write(struct.pack("<Q", len(sizes))) + + sizes = np.array(sizes, dtype=np.int32) + self._file.write(sizes.tobytes(order="C")) + del sizes + + pointers = np.array(pointers, dtype=np.int64) + self._file.write(pointers.tobytes(order="C")) + del pointers + + def __exit__(self, exc_type, exc_val, exc_tb): + self._file.close() + + return _Writer() + + def __init__(self, path): + with open(path, "rb") as stream: + magic_test = stream.read(9) + assert self._HDR_MAGIC == magic_test, ( + "Index file doesn't match expected format. " + "Make sure that --dataset-impl is configured properly." + ) + version = struct.unpack("<Q", stream.read(8)) + assert (1,) == version + + (dtype_code,) = struct.unpack("<B", stream.read(1)) + self._dtype = _code_to_dtype[dtype_code] + self._dtype_size = self._dtype().itemsize + + self._len = struct.unpack("<Q", stream.read(8))[0] + offset = stream.tell() + + _warmup_mmap_file(path) + + self._bin_buffer_mmap = np.memmap(path, mode="r", order="C") + self._bin_buffer = memoryview(self._bin_buffer_mmap) + self._sizes = np.frombuffer( + self._bin_buffer, dtype=np.int32, count=self._len, offset=offset + ) + self._pointers = np.frombuffer( + self._bin_buffer, + dtype=np.int64, + count=self._len, + offset=offset + self._sizes.nbytes, + ) + + def __del__(self): + self._bin_buffer_mmap._mmap.close() + del self._bin_buffer_mmap + + @property + def dtype(self): + return self._dtype + + @property + def sizes(self): + return self._sizes + + @lru_cache(maxsize=8) + def __getitem__(self, i): + return self._pointers[i], self._sizes[i] + + def __len__(self): + return self._len + + def __init__(self, path): + super().__init__() + + self._path = None + self._index = None + self._bin_buffer = None + + self._do_init(path) + + def __getstate__(self): + return self._path + + def __setstate__(self, state): + self._do_init(state) + + def _do_init(self, path): + self._path = path + self._index = self.Index(index_file_path(self._path)) + + _warmup_mmap_file(data_file_path(self._path)) + self._bin_buffer_mmap = np.memmap( + data_file_path(self._path), mode="r", order="C" + ) + self._bin_buffer = memoryview(self._bin_buffer_mmap) + + def __del__(self): + self._bin_buffer_mmap._mmap.close() + del self._bin_buffer_mmap + del self._index + + def __len__(self): + return len(self._index) + + @lru_cache(maxsize=8) + def __getitem__(self, i): + ptr, size = self._index[i] + np_array = np.frombuffer( + self._bin_buffer, dtype=self._index.dtype, count=size, offset=ptr + ) + if self._index.dtype != np.int64: + np_array = np_array.astype(np.int64) + + return torch.from_numpy(np_array) + + @property + def sizes(self): + return self._index.sizes + + @property + def supports_prefetch(self): + return False + + @staticmethod + def exists(path): + return PathManager.exists(index_file_path(path)) and PathManager.exists( + data_file_path(path) + ) + + +def get_indexed_dataset_to_local(path) -> str: + local_index_path = PathManager.get_local_path(index_file_path(path)) + local_data_path = PathManager.get_local_path(data_file_path(path)) + + assert local_index_path.endswith(".idx") and local_data_path.endswith(".bin"), ( + "PathManager.get_local_path does not return files with expected patterns: " + f"{local_index_path} and {local_data_path}" + ) + + local_path = local_data_path[:-4] # stripping surfix ".bin" + assert local_path == local_index_path[:-4] # stripping surfix ".idx" + return local_path + + +class MMapIndexedDatasetBuilder: + def __init__(self, out_file, dtype=np.int64): + self._data_file = open(out_file, "wb") + self._dtype = dtype + self._sizes = [] + + def add_item(self, tensor): + np_array = np.array(tensor.numpy(), dtype=self._dtype) + self._data_file.write(np_array.tobytes(order="C")) + self._sizes.append(np_array.size) + + def merge_file_(self, another_file): + # Concatenate index + index = MMapIndexedDataset.Index(index_file_path(another_file)) + assert index.dtype == self._dtype + + for size in index.sizes: + self._sizes.append(size) + + # Concatenate data + with open(data_file_path(another_file), "rb") as f: + shutil.copyfileobj(f, self._data_file) + + def finalize(self, index_file): + self._data_file.close() + + with MMapIndexedDataset.Index.writer(index_file, self._dtype) as index: + index.write(self._sizes) diff --git a/fairseq/data/iterators.py b/fairseq/data/iterators.py new file mode 100644 index 0000000000000000000000000000000000000000..86f6d0553371c3195aaa780778e7e830e7e27e1a --- /dev/null +++ b/fairseq/data/iterators.py @@ -0,0 +1,640 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import itertools +import logging +import math +import operator +import os +import queue +import time +from threading import Thread + +import numpy as np +import torch +from fairseq.data import data_utils + + +logger = logging.getLogger(__name__) + +# Object used by _background_consumer to signal the source is exhausted +# to the main thread. +_sentinel = object() + + +class CountingIterator(object): + """Wrapper around an iterable that maintains the iteration count. + + Args: + iterable (iterable): iterable to wrap + start (int): starting iteration count. Note that this doesn't + actually advance the iterator. + total (int): override the iterator length returned by ``__len``. + This can be used to truncate *iterator*. + + Attributes: + n (int): number of elements consumed from this iterator + """ + + def __init__(self, iterable, start=None, total=None): + self._itr = iter(iterable) + self.n = start or getattr(iterable, "n", 0) + self.total = total or self.n + len(iterable) + + def __len__(self): + return self.total + + def __iter__(self): + return self + + def __next__(self): + if not self.has_next(): + raise StopIteration + try: + x = next(self._itr) + except StopIteration: + raise IndexError(f"Iterator expected to have length {self.total}, " + "but exhausted at position {self.n}.") + self.n += 1 + return x + + def has_next(self): + """Whether the iterator has been exhausted.""" + return self.n < self.total + + def skip(self, n): + """Fast-forward the iterator by skipping n elements.""" + for _ in range(n): + next(self) + return self + + def take(self, n): + """Truncate the iterator to n elements at most.""" + self.total = min(self.total, n) + # Propagate this change to the underlying iterator + if hasattr(self._itr, "take"): + self._itr.take(max(n - self.n, 0)) + return self + + +class EpochBatchIterating(object): + def __len__(self) -> int: + raise NotImplementedError + + @property + def next_epoch_idx(self): + raise NotImplementedError + + def next_epoch_itr( + self, shuffle=True, fix_batches_to_gpus=False, set_dataset_epoch=True + ): + """Return a new iterator over the dataset. + + Args: + shuffle (bool, optional): shuffle batches before returning the + iterator (default: True). + fix_batches_to_gpus (bool, optional): ensure that batches are always + allocated to the same shards across epochs. Requires + that :attr:`dataset` supports prefetching (default: False). + set_dataset_epoch (bool, optional): update the wrapped Dataset with + the new epoch number (default: True). + """ + raise NotImplementedError + + def end_of_epoch(self) -> bool: + """Returns whether the most recent epoch iterator has been exhausted""" + raise NotImplementedError + + @property + def iterations_in_epoch(self) -> int: + """The number of consumed batches in the current epoch.""" + raise NotImplementedError + + def state_dict(self): + """Returns a dictionary containing a whole state of the iterator.""" + raise NotImplementedError + + def load_state_dict(self, state_dict): + """Copies the state of the iterator from the given *state_dict*.""" + raise NotImplementedError + + @property + def first_batch(self): + return "DUMMY" + + +class StreamingEpochBatchIterator(EpochBatchIterating): + """A steaming-style iterator over a :class:`torch.utils.data.IterableDataset`. + + Args: + dataset (~torch.utils.data.Dataset): dataset from which to load the data + max_sentences: batch size + collate_fn (callable): merges a list of samples to form a mini-batch + num_workers (int, optional): how many subprocesses to use for data + loading. 0 means the data will be loaded in the main process + (default: 0). + epoch (int, optional): the epoch to start the iterator from + (default: 1). + buffer_size (int, optional): the number of batches to keep ready in the + queue. Helps speeding up dataloading. When buffer_size is zero, the + default torch.utils.data.DataLoader preloading is used. + timeout (int, optional): if positive, the timeout value for collecting a batch + from workers. Should always be non-negative (default: ``0``). + """ + + def __init__( + self, + dataset, + max_sentences=1, + collate_fn=None, + epoch=1, + num_workers=0, + buffer_size=0, + timeout=0, + ): + assert isinstance(dataset, torch.utils.data.IterableDataset) + self.dataset = dataset + self.max_sentences = max_sentences + self.collate_fn = collate_fn + self.epoch = max(epoch, 1) # we use 1-based indexing for epochs + self.num_workers = num_workers + # This upper limit here is to prevent people from abusing this feature + # in a shared computing environment. + self.buffer_size = min(buffer_size, 20) + self.timeout = timeout + + self._current_epoch_iterator = None + + @property + def next_epoch_idx(self): + """Return the epoch index after *next_epoch_itr* is called.""" + if self._current_epoch_iterator is not None and self.end_of_epoch(): + return self.epoch + 1 + else: + return self.epoch + + def next_epoch_itr( + self, shuffle=True, fix_batches_to_gpus=False, set_dataset_epoch=True + ): + self.epoch = self.next_epoch_idx + if set_dataset_epoch and hasattr(self.dataset, "set_epoch"): + self.dataset.set_epoch(self.epoch) + self._current_epoch_iterator = self._get_iterator_for_epoch(self.epoch, shuffle) + return self._current_epoch_iterator + + def end_of_epoch(self) -> bool: + return not self._current_epoch_iterator.has_next() + + @property + def iterations_in_epoch(self) -> int: + if self._current_epoch_iterator is not None: + return self._current_epoch_iterator.n + return 0 + + def state_dict(self): + return { + "epoch": self.epoch, + } + + def load_state_dict(self, state_dict): + self.epoch = state_dict["epoch"] + + def _get_iterator_for_epoch(self, epoch, shuffle, offset=0): + if self.num_workers > 0: + os.environ["PYTHONWARNINGS"] = "ignore:semaphore_tracker:UserWarning" + + # Create data loader + worker_init_fn = getattr(self.dataset, "worker_init_fn", None) + itr = torch.utils.data.DataLoader( + self.dataset, + batch_size=self.max_sentences, + collate_fn=self.collate_fn, + num_workers=self.num_workers, + timeout=self.timeout, + worker_init_fn=worker_init_fn, + pin_memory=True, + ) + + # Wrap with a BufferedIterator if needed + if self.buffer_size > 0: + itr = BufferedIterator(self.buffer_size, itr) + + # Wrap with CountingIterator + itr = CountingIterator(itr, start=offset) + + return itr + + +class EpochBatchIterator(EpochBatchIterating): + """A multi-epoch iterator over a :class:`torch.utils.data.Dataset`. + + Compared to :class:`torch.utils.data.DataLoader`, this iterator: + + - can be reused across multiple epochs with the :func:`next_epoch_itr` + method (optionally shuffled between epochs) + - can be serialized/deserialized with the :func:`state_dict` and + :func:`load_state_dict` methods + - supports sharding with the *num_shards* and *shard_id* arguments + + Args: + dataset (~torch.utils.data.Dataset): dataset from which to load the data + collate_fn (callable): merges a list of samples to form a mini-batch + batch_sampler (~torch.utils.data.Sampler or a callable): an iterator over batches of + indices, or a callable to create such an iterator (~torch.utils.data.Sampler). + A callable batch_sampler will be called for each epoch to enable per epoch dynamic + batch iterators defined by this callable batch_sampler. + seed (int, optional): seed for random number generator for + reproducibility (default: 1). + num_shards (int, optional): shard the data iterator into N + shards (default: 1). + shard_id (int, optional): which shard of the data iterator to + return (default: 0). + num_workers (int, optional): how many subprocesses to use for data + loading. 0 means the data will be loaded in the main process + (default: 0). + epoch (int, optional): the epoch to start the iterator from + (default: 1). + buffer_size (int, optional): the number of batches to keep ready in the + queue. Helps speeding up dataloading. When buffer_size is zero, the + default torch.utils.data.DataLoader preloading is used. + timeout (int, optional): if positive, the timeout value for collecting a batch + from workers. Should always be non-negative (default: ``0``). + disable_shuffling (bool, optional): force disable shuffling + (default: ``False``). + """ + + def __init__( + self, + dataset, + collate_fn, + batch_sampler, + seed=1, + num_shards=1, + shard_id=0, + num_workers=0, + epoch=1, + buffer_size=0, + timeout=0, + disable_shuffling=False, + ): + assert isinstance(dataset, torch.utils.data.Dataset) + self.dataset = dataset + self.collate_fn = collate_fn + self.batch_sampler = batch_sampler + self._frozen_batches = ( + tuple(batch_sampler) if not callable(batch_sampler) else None + ) + self.seed = seed + self.num_shards = num_shards + self.shard_id = shard_id + self.num_workers = num_workers + # This upper limit here is to prevent people from abusing this feature + # in a shared computing environment. + self.buffer_size = min(buffer_size, 20) + self.timeout = timeout + self.disable_shuffling = disable_shuffling + + self.epoch = max(epoch, 1) # we use 1-based indexing for epochs + self.shuffle = not disable_shuffling + self._cur_epoch_itr = None + self._next_epoch_itr = None + self._supports_prefetch = getattr(dataset, "supports_prefetch", False) + + @property + def frozen_batches(self): + if self._frozen_batches is None: + self._frozen_batches = tuple(self.batch_sampler(self.dataset, self.epoch)) + return self._frozen_batches + + @property + def first_batch(self): + if len(self.frozen_batches) == 0: + raise Exception( + "The dataset is empty. This could indicate " + "that all elements in the dataset have been skipped. " + "Try increasing the max number of allowed tokens or using " + "a larger dataset." + ) + + if getattr(self.dataset, "supports_fetch_outside_dataloader", True): + return self.collate_fn([self.dataset[i] for i in self.frozen_batches[0]]) + else: + return "DUMMY" + + def __len__(self): + return int(math.ceil(len(self.frozen_batches) / float(self.num_shards))) + + @property + def n(self): + return self.iterations_in_epoch + + @property + def next_epoch_idx(self): + """Return the epoch index after *next_epoch_itr* is called.""" + if self._next_epoch_itr is not None: + return self.epoch + elif self._cur_epoch_itr is not None and self.end_of_epoch(): + return self.epoch + 1 + else: + return self.epoch + + def next_epoch_itr( + self, shuffle=True, fix_batches_to_gpus=False, set_dataset_epoch=True + ): + """Return a new iterator over the dataset. + + Args: + shuffle (bool, optional): shuffle batches before returning the + iterator (default: True). + fix_batches_to_gpus (bool, optional): ensure that batches are always + allocated to the same shards across epochs. Requires + that :attr:`dataset` supports prefetching (default: False). + set_dataset_epoch (bool, optional): update the wrapped Dataset with + the new epoch number (default: True). + """ + if self.disable_shuffling: + shuffle = False + prev_epoch = self.epoch + self.epoch = self.next_epoch_idx + if set_dataset_epoch and hasattr(self.dataset, "set_epoch"): + self.dataset.set_epoch(self.epoch) + if self._next_epoch_itr is not None: + self._cur_epoch_itr = self._next_epoch_itr + self._next_epoch_itr = None + else: + if callable(self.batch_sampler) and prev_epoch != self.epoch: + # reset _frozen_batches to refresh the next epoch + self._frozen_batches = None + self._cur_epoch_itr = self._get_iterator_for_epoch( + self.epoch, + shuffle, + fix_batches_to_gpus=fix_batches_to_gpus, + ) + self.shuffle = shuffle + return self._cur_epoch_itr + + def end_of_epoch(self) -> bool: + """Returns whether the most recent epoch iterator has been exhausted""" + return not self._cur_epoch_itr.has_next() + + @property + def iterations_in_epoch(self): + """The number of consumed batches in the current epoch.""" + if self._cur_epoch_itr is not None: + return self._cur_epoch_itr.n + elif self._next_epoch_itr is not None: + return self._next_epoch_itr.n + return 0 + + def state_dict(self): + """Returns a dictionary containing a whole state of the iterator.""" + if self.end_of_epoch(): + epoch = self.epoch + 1 + iter_in_epoch = 0 + else: + epoch = self.epoch + iter_in_epoch = self.iterations_in_epoch + return { + "version": 2, + "epoch": epoch, + "iterations_in_epoch": iter_in_epoch, + "shuffle": self.shuffle, + } + + def load_state_dict(self, state_dict): + """Copies the state of the iterator from the given *state_dict*.""" + self.epoch = state_dict["epoch"] + itr_pos = state_dict.get("iterations_in_epoch", 0) + version = state_dict.get("version", 1) + if itr_pos > 0: + # fast-forward epoch iterator + self._next_epoch_itr = self._get_iterator_for_epoch( + self.epoch, + shuffle=state_dict.get("shuffle", True), + offset=itr_pos, + ) + if self._next_epoch_itr is None: + if version == 1: + # legacy behavior: we finished the epoch, increment epoch counter + self.epoch += 1 + else: + raise RuntimeError( + "Cannot resume training due to dataloader mismatch, please " + "report this to the fairseq developers. You can relaunch " + "training with `--reset-dataloader` and it should work." + ) + else: + self._next_epoch_itr = None + + def _get_iterator_for_epoch( + self, epoch, shuffle, fix_batches_to_gpus=False, offset=0 + ): + def shuffle_batches(batches, seed): + with data_utils.numpy_seed(seed): + np.random.shuffle(batches) + return batches + + if self._supports_prefetch: + batches = self.frozen_batches + + if shuffle and not fix_batches_to_gpus: + batches = shuffle_batches(list(batches), self.seed + epoch) + + batches = list( + ShardedIterator(batches, self.num_shards, self.shard_id, fill_value=[]) + ) + self.dataset.prefetch([i for s in batches for i in s]) + + if shuffle and fix_batches_to_gpus: + batches = shuffle_batches(batches, self.seed + epoch + self.shard_id) + else: + if shuffle: + batches = shuffle_batches(list(self.frozen_batches), self.seed + epoch) + else: + batches = self.frozen_batches + batches = list( + ShardedIterator(batches, self.num_shards, self.shard_id, fill_value=[]) + ) + + if offset > 0 and offset >= len(batches): + return None + + if self.num_workers > 0: + os.environ["PYTHONWARNINGS"] = "ignore:semaphore_tracker:UserWarning" + + # Create data loader + itr = torch.utils.data.DataLoader( + self.dataset, + collate_fn=self.collate_fn, + batch_sampler=batches[offset:], + num_workers=self.num_workers, + timeout=self.timeout, + pin_memory=True, + ) + + # Wrap with a BufferedIterator if needed + if self.buffer_size > 0: + itr = BufferedIterator(self.buffer_size, itr) + + # Wrap with CountingIterator + itr = CountingIterator(itr, start=offset) + return itr + + +class GroupedIterator(CountingIterator): + """Wrapper around an iterable that returns groups (chunks) of items. + + Args: + iterable (iterable): iterable to wrap + chunk_size (int): size of each chunk + + Attributes: + n (int): number of elements consumed from this iterator + """ + + def __init__(self, iterable, chunk_size): + itr = _chunk_iterator(iterable, chunk_size) + super().__init__( + itr, + start=int(math.ceil(getattr(iterable, "n", 0) / float(chunk_size))), + total=int(math.ceil(len(iterable) / float(chunk_size))), + ) + self.chunk_size = chunk_size + + +def _chunk_iterator(itr, chunk_size): + chunk = [] + for x in itr: + chunk.append(x) + if len(chunk) == chunk_size: + yield chunk + chunk = [] + if len(chunk) > 0: + yield chunk + + +class ShardedIterator(CountingIterator): + """A sharded wrapper around an iterable, padded to length. + + Args: + iterable (iterable): iterable to wrap + num_shards (int): number of shards to split the iterable into + shard_id (int): which shard to iterator over + fill_value (Any, optional): padding value when the iterable doesn't + evenly divide *num_shards* (default: None). + + Attributes: + n (int): number of elements consumed from this iterator + """ + + def __init__(self, iterable, num_shards, shard_id, fill_value=None): + if shard_id < 0 or shard_id >= num_shards: + raise ValueError("shard_id must be between 0 and num_shards") + sharded_len = int(math.ceil(len(iterable) / float(num_shards))) + itr = map( + operator.itemgetter(1), + itertools.zip_longest( + range(sharded_len), + itertools.islice(iterable, shard_id, len(iterable), num_shards), + fillvalue=fill_value, + ), + ) + super().__init__( + itr, + start=int(math.ceil(getattr(iterable, "n", 0) / float(num_shards))), + total=sharded_len, + ) + + +class BackgroundConsumer(Thread): + def __init__(self, queue, source, max_len, cuda_device): + Thread.__init__(self) + + self._queue = queue + self._source = source + self._max_len = max_len + self.count = 0 + self.cuda_device = cuda_device + + def run(self): + # set_device to avoid creation of GPU0 context when using pin_memory + if self.cuda_device is not None: + torch.cuda.set_device(self.cuda_device) + + try: + for item in self._source: + self._queue.put(item) + + # Stop if we reached the maximum length + self.count += 1 + if self._max_len is not None and self.count >= self._max_len: + break + + # Signal the consumer we are done. + self._queue.put(_sentinel) + except Exception as e: + self._queue.put(e) + + +class BufferedIterator(object): + def __init__(self, size, iterable): + self._queue = queue.Queue(size) + self._iterable = iterable + self._consumer = None + + self.start_time = time.time() + self.warning_time = None + + self.total = len(iterable) + + def _create_consumer(self): + self._consumer = BackgroundConsumer( + self._queue, + self._iterable, + self.total, + torch.cuda.current_device() if torch.cuda.is_available() else None + ) + self._consumer.daemon = True + self._consumer.start() + + def __iter__(self): + return self + + def __len__(self): + return self.total + + def take(self, n): + self.total = min(self.total, n) + # Propagate this change to the underlying iterator + if hasattr(self._iterable, "take"): + self._iterable.take(n) + return self + + def __next__(self): + # Create consumer if not created yet + if self._consumer is None: + self._create_consumer() + + # Notify the user if there is a data loading bottleneck + if self._queue.qsize() < min(2, max(1, self._queue.maxsize // 2)): + if time.time() - self.start_time > 5 * 60: + if ( + self.warning_time is None + or time.time() - self.warning_time > 15 * 60 + ): + logger.debug( + "Data loading buffer is empty or nearly empty. This may " + "indicate a data loading bottleneck, and increasing the " + "number of workers (--num-workers) may help." + ) + self.warning_time = time.time() + + # Get next example + item = self._queue.get(True) + if isinstance(item, Exception): + raise item + if item is _sentinel: + raise StopIteration() + return item diff --git a/fairseq/data/language_pair_dataset.py b/fairseq/data/language_pair_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..ff3e14bf14770638524ef6067b558e455dbe5f2b --- /dev/null +++ b/fairseq/data/language_pair_dataset.py @@ -0,0 +1,471 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging + +import numpy as np +import torch +from fairseq.data import FairseqDataset, data_utils + + +logger = logging.getLogger(__name__) + + +def collate( + samples, + pad_idx, + eos_idx, + left_pad_source=True, + left_pad_target=False, + input_feeding=True, + pad_to_length=None, + pad_to_multiple=1, +): + if len(samples) == 0: + return {} + + def merge(key, left_pad, move_eos_to_beginning=False, pad_to_length=None): + return data_utils.collate_tokens( + [s[key] for s in samples], + pad_idx, + eos_idx, + left_pad, + move_eos_to_beginning, + pad_to_length=pad_to_length, + pad_to_multiple=pad_to_multiple, + ) + + def check_alignment(alignment, src_len, tgt_len): + if alignment is None or len(alignment) == 0: + return False + if ( + alignment[:, 0].max().item() >= src_len - 1 + or alignment[:, 1].max().item() >= tgt_len - 1 + ): + logger.warning("alignment size mismatch found, skipping alignment!") + return False + return True + + def compute_alignment_weights(alignments): + """ + Given a tensor of shape [:, 2] containing the source-target indices + corresponding to the alignments, a weight vector containing the + inverse frequency of each target index is computed. + For e.g. if alignments = [[5, 7], [2, 3], [1, 3], [4, 2]], then + a tensor containing [1., 0.5, 0.5, 1] should be returned (since target + index 3 is repeated twice) + """ + align_tgt = alignments[:, 1] + _, align_tgt_i, align_tgt_c = torch.unique( + align_tgt, return_inverse=True, return_counts=True + ) + align_weights = align_tgt_c[align_tgt_i[np.arange(len(align_tgt))]] + return 1.0 / align_weights.float() + + id = torch.LongTensor([s["id"] for s in samples]) + src_tokens = merge( + "source", + left_pad=left_pad_source, + pad_to_length=pad_to_length["source"] if pad_to_length is not None else None, + ) + # sort by descending source length + src_lengths = torch.LongTensor( + [s["source"].ne(pad_idx).long().sum() for s in samples] + ) + src_lengths, sort_order = src_lengths.sort(descending=True) + id = id.index_select(0, sort_order) + src_tokens = src_tokens.index_select(0, sort_order) + + prev_output_tokens = None + target = None + if samples[0].get("target", None) is not None: + target = merge( + "target", + left_pad=left_pad_target, + pad_to_length=pad_to_length["target"] + if pad_to_length is not None + else None, + ) + target = target.index_select(0, sort_order) + tgt_lengths = torch.LongTensor( + [s["target"].ne(pad_idx).long().sum() for s in samples] + ).index_select(0, sort_order) + ntokens = tgt_lengths.sum().item() + + if samples[0].get("prev_output_tokens", None) is not None: + prev_output_tokens = merge("prev_output_tokens", left_pad=left_pad_target) + elif input_feeding: + # we create a shifted version of targets for feeding the + # previous output token(s) into the next decoder step + prev_output_tokens = merge( + "target", + left_pad=left_pad_target, + move_eos_to_beginning=True, + pad_to_length=pad_to_length["target"] + if pad_to_length is not None + else None, + ) + else: + ntokens = src_lengths.sum().item() + + batch = { + "id": id, + "nsentences": len(samples), + "ntokens": ntokens, + "net_input": {"src_tokens": src_tokens, "src_lengths": src_lengths,}, + "target": target, + } + if prev_output_tokens is not None: + batch["net_input"]["prev_output_tokens"] = prev_output_tokens.index_select( + 0, sort_order + ) + + if samples[0].get("alignment", None) is not None: + bsz, tgt_sz = batch["target"].shape + src_sz = batch["net_input"]["src_tokens"].shape[1] + + offsets = torch.zeros((len(sort_order), 2), dtype=torch.long) + offsets[:, 1] += torch.arange(len(sort_order), dtype=torch.long) * tgt_sz + if left_pad_source: + offsets[:, 0] += src_sz - src_lengths + if left_pad_target: + offsets[:, 1] += tgt_sz - tgt_lengths + + alignments = [ + alignment + offset + for align_idx, offset, src_len, tgt_len in zip( + sort_order, offsets, src_lengths, tgt_lengths + ) + for alignment in [samples[align_idx]["alignment"].view(-1, 2)] + if check_alignment(alignment, src_len, tgt_len) + ] + + if len(alignments) > 0: + alignments = torch.cat(alignments, dim=0) + align_weights = compute_alignment_weights(alignments) + + batch["alignments"] = alignments + batch["align_weights"] = align_weights + + if samples[0].get("constraints", None) is not None: + # Collate the packed constraints across the samples, padding to + # the length of the longest sample. + lens = [sample.get("constraints").size(0) for sample in samples] + max_len = max(lens) + constraints = torch.zeros((len(samples), max(lens))).long() + for i, sample in enumerate(samples): + constraints[i, 0 : lens[i]] = samples[i].get("constraints") + batch["constraints"] = constraints.index_select(0, sort_order) + + return batch + + +class LanguagePairDataset(FairseqDataset): + """ + A pair of torch.utils.data.Datasets. + + Args: + src (torch.utils.data.Dataset): source dataset to wrap + src_sizes (List[int]): source sentence lengths + src_dict (~fairseq.data.Dictionary): source vocabulary + tgt (torch.utils.data.Dataset, optional): target dataset to wrap + tgt_sizes (List[int], optional): target sentence lengths + tgt_dict (~fairseq.data.Dictionary, optional): target vocabulary + left_pad_source (bool, optional): pad source tensors on the left side + (default: True). + left_pad_target (bool, optional): pad target tensors on the left side + (default: False). + shuffle (bool, optional): shuffle dataset elements before batching + (default: True). + input_feeding (bool, optional): create a shifted version of the targets + to be passed into the model for teacher forcing (default: True). + remove_eos_from_source (bool, optional): if set, removes eos from end + of source if it's present (default: False). + append_eos_to_target (bool, optional): if set, appends eos to end of + target if it's absent (default: False). + align_dataset (torch.utils.data.Dataset, optional): dataset + containing alignments. + constraints (Tensor, optional): 2d tensor with a concatenated, zero- + delimited list of constraints for each sentence. + append_bos (bool, optional): if set, appends bos to the beginning of + source/target sentence. + num_buckets (int, optional): if set to a value greater than 0, then + batches will be bucketed into the given number of batch shapes. + src_lang_id (int, optional): source language ID, if set, the collated batch + will contain a field 'src_lang_id' in 'net_input' which indicates the + source language of the samples. + tgt_lang_id (int, optional): target language ID, if set, the collated batch + will contain a field 'tgt_lang_id' which indicates the target language + of the samples. + """ + + def __init__( + self, + src, + src_sizes, + src_dict, + tgt=None, + tgt_sizes=None, + tgt_dict=None, + left_pad_source=True, + left_pad_target=False, + shuffle=True, + input_feeding=True, + remove_eos_from_source=False, + append_eos_to_target=False, + align_dataset=None, + constraints=None, + append_bos=False, + eos=None, + num_buckets=0, + src_lang_id=None, + tgt_lang_id=None, + pad_to_multiple=1, + ): + if tgt_dict is not None: + assert src_dict.pad() == tgt_dict.pad() + assert src_dict.eos() == tgt_dict.eos() + assert src_dict.unk() == tgt_dict.unk() + if tgt is not None: + assert len(src) == len( + tgt + ), "Source and target must contain the same number of examples" + self.src = src + self.tgt = tgt + self.src_sizes = np.array(src_sizes) + self.tgt_sizes = np.array(tgt_sizes) if tgt_sizes is not None else None + self.sizes = ( + np.vstack((self.src_sizes, self.tgt_sizes)).T + if self.tgt_sizes is not None + else self.src_sizes + ) + self.src_dict = src_dict + self.tgt_dict = tgt_dict + self.left_pad_source = left_pad_source + self.left_pad_target = left_pad_target + self.shuffle = shuffle + self.input_feeding = input_feeding + self.remove_eos_from_source = remove_eos_from_source + self.append_eos_to_target = append_eos_to_target + self.align_dataset = align_dataset + if self.align_dataset is not None: + assert ( + self.tgt_sizes is not None + ), "Both source and target needed when alignments are provided" + self.constraints = constraints + self.append_bos = append_bos + self.eos = eos if eos is not None else src_dict.eos() + self.src_lang_id = src_lang_id + self.tgt_lang_id = tgt_lang_id + if num_buckets > 0: + from fairseq.data import BucketPadLengthDataset + + self.src = BucketPadLengthDataset( + self.src, + sizes=self.src_sizes, + num_buckets=num_buckets, + pad_idx=self.src_dict.pad(), + left_pad=self.left_pad_source, + ) + self.src_sizes = self.src.sizes + logger.info("bucketing source lengths: {}".format(list(self.src.buckets))) + if self.tgt is not None: + self.tgt = BucketPadLengthDataset( + self.tgt, + sizes=self.tgt_sizes, + num_buckets=num_buckets, + pad_idx=self.tgt_dict.pad(), + left_pad=self.left_pad_target, + ) + self.tgt_sizes = self.tgt.sizes + logger.info( + "bucketing target lengths: {}".format(list(self.tgt.buckets)) + ) + + # determine bucket sizes using self.num_tokens, which will return + # the padded lengths (thanks to BucketPadLengthDataset) + num_tokens = np.vectorize(self.num_tokens, otypes=[np.compat.long]) + self.bucketed_num_tokens = num_tokens(np.arange(len(self.src))) + self.buckets = [ + (None, num_tokens) for num_tokens in np.unique(self.bucketed_num_tokens) + ] + else: + self.buckets = None + self.pad_to_multiple = pad_to_multiple + + def get_batch_shapes(self): + return self.buckets + + def __getitem__(self, index): + tgt_item = self.tgt[index] if self.tgt is not None else None + src_item = self.src[index] + # Append EOS to end of tgt sentence if it does not have an EOS and remove + # EOS from end of src sentence if it exists. This is useful when we use + # use existing datasets for opposite directions i.e., when we want to + # use tgt_dataset as src_dataset and vice versa + if self.append_eos_to_target: + eos = self.tgt_dict.eos() if self.tgt_dict else self.src_dict.eos() + if self.tgt and self.tgt[index][-1] != eos: + tgt_item = torch.cat([self.tgt[index], torch.LongTensor([eos])]) + + if self.append_bos: + bos = self.tgt_dict.bos() if self.tgt_dict else self.src_dict.bos() + if self.tgt and self.tgt[index][0] != bos: + tgt_item = torch.cat([torch.LongTensor([bos]), self.tgt[index]]) + + bos = self.src_dict.bos() + if self.src[index][0] != bos: + src_item = torch.cat([torch.LongTensor([bos]), self.src[index]]) + + if self.remove_eos_from_source: + eos = self.src_dict.eos() + if self.src[index][-1] == eos: + src_item = self.src[index][:-1] + + example = { + "id": index, + "source": src_item, + "target": tgt_item, + } + if self.align_dataset is not None: + example["alignment"] = self.align_dataset[index] + if self.constraints is not None: + example["constraints"] = self.constraints[index] + return example + + def __len__(self): + return len(self.src) + + def collater(self, samples, pad_to_length=None): + """Merge a list of samples to form a mini-batch. + + Args: + samples (List[dict]): samples to collate + pad_to_length (dict, optional): a dictionary of + {'source': source_pad_to_length, 'target': target_pad_to_length} + to indicate the max length to pad to in source and target respectively. + + Returns: + dict: a mini-batch with the following keys: + + - `id` (LongTensor): example IDs in the original input order + - `ntokens` (int): total number of tokens in the batch + - `net_input` (dict): the input to the Model, containing keys: + + - `src_tokens` (LongTensor): a padded 2D Tensor of tokens in + the source sentence of shape `(bsz, src_len)`. Padding will + appear on the left if *left_pad_source* is ``True``. + - `src_lengths` (LongTensor): 1D Tensor of the unpadded + lengths of each source sentence of shape `(bsz)` + - `prev_output_tokens` (LongTensor): a padded 2D Tensor of + tokens in the target sentence, shifted right by one + position for teacher forcing, of shape `(bsz, tgt_len)`. + This key will not be present if *input_feeding* is + ``False``. Padding will appear on the left if + *left_pad_target* is ``True``. + - `src_lang_id` (LongTensor): a long Tensor which contains source + language IDs of each sample in the batch + + - `target` (LongTensor): a padded 2D Tensor of tokens in the + target sentence of shape `(bsz, tgt_len)`. Padding will appear + on the left if *left_pad_target* is ``True``. + - `tgt_lang_id` (LongTensor): a long Tensor which contains target language + IDs of each sample in the batch + """ + res = collate( + samples, + pad_idx=self.src_dict.pad(), + eos_idx=self.eos, + left_pad_source=self.left_pad_source, + left_pad_target=self.left_pad_target, + input_feeding=self.input_feeding, + pad_to_length=pad_to_length, + pad_to_multiple=self.pad_to_multiple, + ) + if self.src_lang_id is not None or self.tgt_lang_id is not None: + src_tokens = res["net_input"]["src_tokens"] + bsz = src_tokens.size(0) + if self.src_lang_id is not None: + res["net_input"]["src_lang_id"] = ( + torch.LongTensor([[self.src_lang_id]]).expand(bsz, 1).to(src_tokens) + ) + if self.tgt_lang_id is not None: + res["tgt_lang_id"] = ( + torch.LongTensor([[self.tgt_lang_id]]).expand(bsz, 1).to(src_tokens) + ) + return res + + def num_tokens(self, index): + """Return the number of tokens in a sample. This value is used to + enforce ``--max-tokens`` during batching.""" + return max( + self.src_sizes[index], + self.tgt_sizes[index] if self.tgt_sizes is not None else 0, + ) + + def num_tokens_vec(self, indices): + """Return the number of tokens for a set of positions defined by indices. + This value is used to enforce ``--max-tokens`` during batching.""" + sizes = self.src_sizes[indices] + if self.tgt_sizes is not None: + sizes = np.maximum(sizes, self.tgt_sizes[indices]) + return sizes + + def size(self, index): + """Return an example's size as a float or tuple. This value is used when + filtering a dataset with ``--max-positions``.""" + return ( + self.src_sizes[index], + self.tgt_sizes[index] if self.tgt_sizes is not None else 0, + ) + + def ordered_indices(self): + """Return an ordered list of indices. Batches will be constructed based + on this order.""" + if self.shuffle: + indices = np.random.permutation(len(self)).astype(np.int64) + else: + indices = np.arange(len(self), dtype=np.int64) + if self.buckets is None: + # sort by target length, then source length + if self.tgt_sizes is not None: + indices = indices[np.argsort(self.tgt_sizes[indices], kind="mergesort")] + return indices[np.argsort(self.src_sizes[indices], kind="mergesort")] + else: + # sort by bucketed_num_tokens, which is: + # max(padded_src_len, padded_tgt_len) + return indices[ + np.argsort(self.bucketed_num_tokens[indices], kind="mergesort") + ] + + @property + def supports_prefetch(self): + return getattr(self.src, "supports_prefetch", False) and ( + getattr(self.tgt, "supports_prefetch", False) or self.tgt is None + ) + + def prefetch(self, indices): + self.src.prefetch(indices) + if self.tgt is not None: + self.tgt.prefetch(indices) + if self.align_dataset is not None: + self.align_dataset.prefetch(indices) + + def filter_indices_by_size(self, indices, max_sizes): + """Filter a list of sample indices. Remove those that are longer + than specified in max_sizes. + + Args: + indices (np.array): original array of sample indices + max_sizes (int or list[int] or tuple[int]): max sample size, + can be defined separately for src and tgt (then list or tuple) + + Returns: + np.array: filtered sample array + list: list of removed indices + """ + return data_utils.filter_paired_dataset_indices_by_size( + self.src_sizes, self.tgt_sizes, indices, max_sizes, + ) diff --git a/fairseq/data/legacy/__init__.py b/fairseq/data/legacy/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..9bd5c72b5e9d7f67fb7e4ef10808d7ec08967ff4 --- /dev/null +++ b/fairseq/data/legacy/__init__.py @@ -0,0 +1,16 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .block_pair_dataset import BlockPairDataset +from .masked_lm_dataset import MaskedLMDataset +from .masked_lm_dictionary import BertDictionary, MaskedLMDictionary + + +__all__ = [ + "BertDictionary", + "BlockPairDataset", + "MaskedLMDataset", + "MaskedLMDictionary", +] diff --git a/fairseq/data/legacy/block_pair_dataset.py b/fairseq/data/legacy/block_pair_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..ba069b46052286c531b4f9706d96788732cd2ad2 --- /dev/null +++ b/fairseq/data/legacy/block_pair_dataset.py @@ -0,0 +1,311 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +import numpy as np +import torch +from fairseq.data import FairseqDataset + + +class BlockPairDataset(FairseqDataset): + """Break a Dataset of tokens into sentence pair blocks for next sentence + prediction as well as masked language model. + + High-level logics are: + 1. break input tensor to tensor blocks + 2. pair the blocks with 50% next sentence and 50% random sentence + 3. return paired blocks as well as related segment labels + + Args: + dataset (~torch.utils.data.Dataset): dataset to break into blocks + sizes: array of sentence lengths + dictionary: dictionary for the task + block_size: maximum block size + break_mode: mode for breaking copurs into block pairs. currently we support + 2 modes + doc: respect document boundaries and each part of the pair should belong to on document + none: don't respect any boundary and cut tokens evenly + short_seq_prob: probability for generating shorter block pairs + doc_break_size: Size for empty line separating documents. Typically 1 if + the sentences have eos, 0 otherwise. + """ + + def __init__( + self, + dataset, + dictionary, + sizes, + block_size, + break_mode="doc", + short_seq_prob=0.1, + doc_break_size=1, + ): + super().__init__() + self.dataset = dataset + self.pad = dictionary.pad() + self.eos = dictionary.eos() + self.cls = dictionary.cls() + self.mask = dictionary.mask() + self.sep = dictionary.sep() + self.break_mode = break_mode + self.dictionary = dictionary + self.short_seq_prob = short_seq_prob + self.block_indices = [] + + assert len(dataset) == len(sizes) + + if break_mode == "doc": + cur_doc = [] + for sent_id, sz in enumerate(sizes): + assert doc_break_size == 0 or sz != 0, ( + "when doc_break_size is non-zero, we expect documents to be" + "separated by a blank line with a single eos." + ) + # empty line as document separator + if sz == doc_break_size: + if len(cur_doc) == 0: + continue + self.block_indices.append(cur_doc) + cur_doc = [] + else: + cur_doc.append(sent_id) + max_num_tokens = block_size - 3 # Account for [CLS], [SEP], [SEP] + self.sent_pairs = [] + self.sizes = [] + for doc_id, doc in enumerate(self.block_indices): + self._generate_sentence_pair(doc, doc_id, max_num_tokens, sizes) + elif break_mode is None or break_mode == "none": + # each block should have half of the block size since we are constructing block pair + sent_length = (block_size - 3) // 2 + total_len = sum(dataset.sizes) + length = math.ceil(total_len / sent_length) + + def block_at(i): + start = i * sent_length + end = min(start + sent_length, total_len) + return (start, end) + + sent_indices = np.array([block_at(i) for i in range(length)]) + sent_sizes = np.array([e - s for s, e in sent_indices]) + dataset_index = self._sent_to_dataset_index(sent_sizes) + + # pair sentences + self._pair_sentences(dataset_index) + else: + raise ValueError("Invalid break_mode: " + break_mode) + + def _pair_sentences(self, dataset_index): + """ + Give a list of evenly cut blocks/sentences, pair these sentences with 50% + consecutive sentences and 50% random sentences. + This is used for none break mode + """ + # pair sentences + for sent_id, sent in enumerate(dataset_index): + next_sent_label = ( + 1 if np.random.rand() > 0.5 and sent_id != len(dataset_index) - 1 else 0 + ) + if next_sent_label: + next_sent = dataset_index[sent_id + 1] + else: + next_sent = dataset_index[ + self._skip_sampling(len(dataset_index), [sent_id, sent_id + 1]) + ] + self.sent_pairs.append((sent, next_sent, next_sent_label)) + + # The current blocks don't include the special tokens but the + # sizes already account for this + self.sizes.append(3 + sent[3] + next_sent[3]) + + def _sent_to_dataset_index(self, sent_sizes): + """ + Build index mapping block indices to the underlying dataset indices + """ + dataset_index = [] + ds_idx, ds_remaining = -1, 0 + for to_consume in sent_sizes: + sent_size = to_consume + if ds_remaining == 0: + ds_idx += 1 + ds_remaining = sent_sizes[ds_idx] + start_ds_idx = ds_idx + start_offset = sent_sizes[ds_idx] - ds_remaining + while to_consume > ds_remaining: + to_consume -= ds_remaining + ds_idx += 1 + ds_remaining = sent_sizes[ds_idx] + ds_remaining -= to_consume + dataset_index.append( + ( + start_ds_idx, # starting index in dataset + start_offset, # starting offset within starting index + ds_idx, # ending index in dataset + sent_size, # sentence length + ) + ) + assert ds_remaining == 0 + assert ds_idx == len(self.dataset) - 1 + return dataset_index + + def _generate_sentence_pair(self, doc, doc_id, max_num_tokens, sizes): + """ + Go through a single document and genrate sentence paris from it + """ + current_chunk = [] + current_length = 0 + curr = 0 + # To provide more randomness, we decrease target seq length for parts of + # samples (10% by default). Note that max_num_tokens is the hard threshold + # for batching and will never be changed. + target_seq_length = max_num_tokens + if np.random.random() < self.short_seq_prob: + target_seq_length = np.random.randint(2, max_num_tokens) + # loop through all sentences in document + while curr < len(doc): + sent_id = doc[curr] + current_chunk.append(sent_id) + current_length = sum(sizes[current_chunk]) + # split chunk and generate pair when exceed target_seq_length or + # finish the loop + if curr == len(doc) - 1 or current_length >= target_seq_length: + # split the chunk into 2 parts + a_end = 1 + if len(current_chunk) > 2: + a_end = np.random.randint(1, len(current_chunk) - 1) + sent_a = current_chunk[:a_end] + len_a = sum(sizes[sent_a]) + # generate next sentence label, note that if there is only 1 sentence + # in current chunk, label is always 0 + next_sent_label = ( + 1 if np.random.rand() > 0.5 and len(current_chunk) != 1 else 0 + ) + if not next_sent_label: + # if next sentence label is 0, sample sent_b from a random doc + target_b_length = target_seq_length - len_a + rand_doc_id = self._skip_sampling(len(self.block_indices), [doc_id]) + random_doc = self.block_indices[rand_doc_id] + random_start = np.random.randint(0, len(random_doc)) + sent_b = [] + len_b = 0 + for j in range(random_start, len(random_doc)): + sent_b.append(random_doc[j]) + len_b = sum(sizes[sent_b]) + if len_b >= target_b_length: + break + # return the second part of the chunk since it's not used + num_unused_segments = len(current_chunk) - a_end + curr -= num_unused_segments + else: + # if next sentence label is 1, use the second part of chunk as sent_B + sent_b = current_chunk[a_end:] + len_b = sum(sizes[sent_b]) + # currently sent_a and sent_B may be longer than max_num_tokens, + # truncate them and return block idx and offsets for them + sent_a, sent_b = self._truncate_sentences( + sent_a, sent_b, max_num_tokens + ) + self.sent_pairs.append((sent_a, sent_b, next_sent_label)) + self.sizes.append(3 + sent_a[3] + sent_b[3]) + current_chunk = [] + curr += 1 + + def _skip_sampling(self, total, skip_ids): + """ + Generate a random integer which is not in skip_ids. Sample range is [0, total) + TODO: ids in skip_ids should be consecutive, we can extend it to more generic version later + """ + rand_id = np.random.randint(total - len(skip_ids)) + return rand_id if rand_id < min(skip_ids) else rand_id + len(skip_ids) + + def _truncate_sentences(self, sent_a, sent_b, max_num_tokens): + """ + Trancate a pair of sentence to limit total length under max_num_tokens + Logics: + 1. Truncate longer sentence + 2. Tokens to be truncated could be at the beginning or the end of the sentnce + Returns: + Truncated sentences represented by dataset idx + """ + len_a, len_b = sum(self.dataset.sizes[sent_a]), sum(self.dataset.sizes[sent_b]) + front_cut_a = front_cut_b = end_cut_a = end_cut_b = 0 + + while True: + total_length = ( + len_a + len_b - front_cut_a - front_cut_b - end_cut_a - end_cut_b + ) + if total_length <= max_num_tokens: + break + + if len_a - front_cut_a - end_cut_a > len_b - front_cut_b - end_cut_b: + if np.random.rand() < 0.5: + front_cut_a += 1 + else: + end_cut_a += 1 + else: + if np.random.rand() < 0.5: + front_cut_b += 1 + else: + end_cut_b += 1 + + # calculate ds indices as well as offsets and return + truncated_sent_a = self._cut_sentence(sent_a, front_cut_a, end_cut_a) + truncated_sent_b = self._cut_sentence(sent_b, front_cut_b, end_cut_b) + return truncated_sent_a, truncated_sent_b + + def _cut_sentence(self, sent, front_cut, end_cut): + """ + Cut a sentence based on the numbers of tokens to be cut from beginning and end + Represent the sentence as dataset idx and return + """ + start_ds_idx, end_ds_idx, offset = sent[0], sent[-1], 0 + target_len = sum(self.dataset.sizes[sent]) - front_cut - end_cut + while front_cut > 0: + if self.dataset.sizes[start_ds_idx] > front_cut: + offset += front_cut + break + else: + front_cut -= self.dataset.sizes[start_ds_idx] + start_ds_idx += 1 + while end_cut > 0: + if self.dataset.sizes[end_ds_idx] > end_cut: + break + else: + end_cut -= self.dataset.sizes[end_ds_idx] + end_ds_idx -= 1 + return start_ds_idx, offset, end_ds_idx, target_len + + def _fetch_block(self, start_ds_idx, offset, end_ds_idx, length): + """ + Fetch a block of tokens based on its dataset idx + """ + buffer = torch.cat( + [self.dataset[idx] for idx in range(start_ds_idx, end_ds_idx + 1)] + ) + s, e = offset, offset + length + return buffer[s:e] + + def __getitem__(self, index): + block1, block2, next_sent_label = self.sent_pairs[index] + block1 = self._fetch_block(*block1) + block2 = self._fetch_block(*block2) + return block1, block2, next_sent_label + + def __len__(self): + return len(self.sizes) + + @property + def supports_prefetch(self): + return getattr(self.dataset, "supports_prefetch", False) + + def prefetch(self, indices): + prefetch_idx = set() + for index in indices: + for block1, block2, _ in [self.sent_pairs[index]]: + for ds_idx in range(block1[0], block1[2] + 1): + prefetch_idx.add(ds_idx) + for ds_idx in range(block2[0], block2[2] + 1): + prefetch_idx.add(ds_idx) + self.dataset.prefetch(prefetch_idx) diff --git a/fairseq/data/legacy/masked_lm_dataset.py b/fairseq/data/legacy/masked_lm_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..dd8ea2c60aff306ab3a756223a298a28d41a4991 --- /dev/null +++ b/fairseq/data/legacy/masked_lm_dataset.py @@ -0,0 +1,303 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +from typing import Dict, List, Tuple + +import numpy as np +import torch +from fairseq.data import Dictionary, FairseqDataset, data_utils +from fairseq.data.concat_dataset import ConcatDataset +from fairseq.data.legacy.block_pair_dataset import BlockPairDataset +from fairseq.data.token_block_dataset import TokenBlockDataset + + +class MaskedLMDataset(FairseqDataset): + """ + A wrapper Dataset for masked language modelling. The dataset + wraps around TokenBlockDataset or BlockedPairDataset and creates a batch + where the input blocks are masked according to the specified masking + probability. Additionally the batch can also contain sentence level targets + if this is specified. + + Args: + dataset: Dataset which generates blocks of data. Only BlockPairDataset + and TokenBlockDataset are supported. + sizes: Sentence lengths + vocab: Dictionary with the vocabulary and special tokens. + pad_idx: Id of padding token in dictionary + mask_idx: Id of mask token in dictionary + classif_token_idx: Id of classification token in dictionary. This is the + token associated with the sentence embedding (Eg: CLS for BERT) + sep_token_idx: Id of separator token in dictionary + (Eg: SEP in BERT) + seed: Seed for random number generator for reproducibility. + shuffle: Shuffle the elements before batching. + has_pairs: Specifies whether the underlying dataset + generates a pair of blocks along with a sentence_target or not. + Setting it to True assumes that the underlying dataset generates a + label for the pair of sentences which is surfaced as + sentence_target. The default value assumes a single block with no + sentence target. + segment_id: An optional segment id for filling in the segment labels + when we are in the single block setting (Eg: XLM). Default is 0. + masking_ratio: specifies what percentage of the blocks should be masked. + masking_prob: specifies the probability of a given token being + replaced with the "MASK" token. + random_token_prob: specifies the probability of a given token being + replaced by a random token from the vocabulary. + """ + + def __init__( + self, + dataset: FairseqDataset, + sizes: np.ndarray, + vocab: Dictionary, + pad_idx: int, + mask_idx: int, + classif_token_idx: int, + sep_token_idx: int, + seed: int = 1, + shuffle: bool = True, + has_pairs: bool = True, + segment_id: int = 0, + masking_ratio: float = 0.15, + masking_prob: float = 0.8, + random_token_prob: float = 0.1, + ): + # Make sure the input datasets are the ones supported + assert ( + isinstance(dataset, TokenBlockDataset) + or isinstance(dataset, BlockPairDataset) + or isinstance(dataset, ConcatDataset) + ), ( + "MaskedLMDataset only wraps TokenBlockDataset or BlockPairDataset or " + "ConcatDataset" + ) + + self.dataset = dataset + self.sizes = np.array(sizes) + self.vocab = vocab + self.pad_idx = pad_idx + self.mask_idx = mask_idx + self.classif_token_idx = classif_token_idx + self.sep_token_idx = sep_token_idx + self.shuffle = shuffle + self.seed = seed + self.has_pairs = has_pairs + self.segment_id = segment_id + self.masking_ratio = masking_ratio + self.masking_prob = masking_prob + self.random_token_prob = random_token_prob + + # If we have only one block then sizes needs to be updated to include + # the classification token + if not has_pairs: + self.sizes = self.sizes + 1 + + def __getitem__(self, index: int): + # if has_pairs, then expect 2 blocks and a sentence target + if self.has_pairs: + (block_one, block_two, sentence_target) = self.dataset[index] + else: + block_one = self.dataset[index] + + return { + "id": index, + "block_one": block_one, + "block_two": block_two if self.has_pairs else None, + "sentence_target": sentence_target if self.has_pairs else None, + } + + def __len__(self): + return len(self.dataset) + + def _mask_block( + self, + sentence: np.ndarray, + mask_idx: int, + pad_idx: int, + dictionary_token_range: Tuple, + ): + """ + Mask tokens for Masked Language Model training + Samples mask_ratio tokens that will be predicted by LM. + + Note:This function may not be efficient enough since we had multiple + conversions between np and torch, we can replace them with torch + operators later. + + Args: + sentence: 1d tensor to be masked + mask_idx: index to use for masking the sentence + pad_idx: index to use for masking the target for tokens we aren't + predicting + dictionary_token_range: range of indices in dictionary which can + be used for random word replacement + (e.g. without special characters) + Return: + masked_sent: masked sentence + target: target with words which we are not predicting replaced + by pad_idx + """ + masked_sent = np.copy(sentence) + sent_length = len(sentence) + mask_num = math.ceil(sent_length * self.masking_ratio) + mask = np.random.choice(sent_length, mask_num, replace=False) + target = np.copy(sentence) + + for i in range(sent_length): + if i in mask: + rand = np.random.random() + + # replace with mask if probability is less than masking_prob + # (Eg: 0.8) + if rand < self.masking_prob: + masked_sent[i] = mask_idx + + # replace with random token if probability is less than + # masking_prob + random_token_prob (Eg: 0.9) + elif rand < (self.masking_prob + self.random_token_prob): + # sample random token from dictionary + masked_sent[i] = np.random.randint( + dictionary_token_range[0], dictionary_token_range[1] + ) + else: + target[i] = pad_idx + + return masked_sent, target + + def _collate(self, samples: List[Dict], pad_idx: int, eos_idx: int): + """ + Does the heavy lifting for creating a batch from the input list of + examples. The logic is as follows: + 1. Mask the input blocks. In case has_pair is True then we have 2 + blocks to mask. + 2. Prepend the first masked block tensor with the special token + used as sentence embedding. Eg: CLS in BERT. This happens + irrespective of the value of has_pair. + 3. If has_pair is True, then append the first masked block with the + special separator token (eg: SEP for BERT) and compute segment + label accordingly. In this case, also append the second masked + block with this special separator token and compute its segment + label. + 4. For the targets tensor, prepend and append with padding index + accordingly. + 5. Concatenate all tensors. + """ + if len(samples) == 0: + return {} + # To ensure determinism, we reset the state of the PRNG after every + # batch based on the seed and the first id of the batch. This ensures + # that across epochs we get the same mask for the same example. This + # is needed for reproducibility and is how BERT does masking + # TODO: Can we add deteminism without this constraint? + with data_utils.numpy_seed(self.seed + samples[0]["id"]): + for s in samples: + + # token range is needed for replacing with random token during + # masking + token_range = (self.vocab.nspecial, len(self.vocab)) + + # mask according to specified probabilities. + masked_blk_one, masked_tgt_one = self._mask_block( + s["block_one"], + self.mask_idx, + self.pad_idx, + token_range, + ) + + tokens = np.concatenate([[self.classif_token_idx], masked_blk_one]) + targets = np.concatenate([[self.pad_idx], masked_tgt_one]) + segments = np.ones(len(tokens)) * self.segment_id + + # if has_pairs is True then we need to add the SEP token to both + # the blocks after masking and re-compute segments based on the new + # lengths. + if self.has_pairs: + tokens_one = np.concatenate([tokens, [self.sep_token_idx]]) + targets_one = np.concatenate([targets, [self.pad_idx]]) + + masked_blk_two, masked_tgt_two = self._mask_block( + s["block_two"], self.mask_idx, self.pad_idx, token_range + ) + tokens_two = np.concatenate([masked_blk_two, [self.sep_token_idx]]) + targets_two = np.concatenate([masked_tgt_two, [self.pad_idx]]) + + # block + 1 sep + 1 special (CLS) + segments_one = np.zeros(len(tokens_one)) + # block + 1 sep + segments_two = np.ones(len(tokens_two)) + + tokens = np.concatenate([tokens_one, tokens_two]) + targets = np.concatenate([targets_one, targets_two]) + segments = np.concatenate([segments_one, segments_two]) + + s["source"] = torch.LongTensor(tokens) + s["segment_labels"] = torch.LongTensor(segments) + s["lm_target"] = torch.LongTensor(targets) + + def merge(key): + return data_utils.collate_tokens( + [s[key] for s in samples], pad_idx, eos_idx, left_pad=False + ) + + return { + "id": torch.LongTensor([s["id"] for s in samples]), + "ntokens": sum(len(s["source"]) for s in samples), + "net_input": { + "src_tokens": merge("source"), + "segment_labels": merge("segment_labels"), + }, + "lm_target": merge("lm_target"), + "sentence_target": torch.LongTensor([s["sentence_target"] for s in samples]) + if self.has_pairs + else None, + "nsentences": len(samples), + } + + def collater(self, samples: List[Dict]): + """Merge a list of samples to form a mini-batch. + + Args: + samples (List[dict]): samples to collate + + Returns: + dict: a mini-batch of data + """ + return self._collate(samples, self.vocab.pad(), self.vocab.eos()) + + def num_tokens(self, index: int): + """ + Return the number of tokens in a sample. This value is used to + enforce max-tokens during batching. + """ + return self.sizes[index] + + def size(self, index: int): + """ + Return an example's size as a float or tuple. This value is used when + filtering a dataset with max-positions. + """ + return self.sizes[index] + + def ordered_indices(self): + """ + Return an ordered list of indices. Batches will be constructed based + on this order. + """ + if self.shuffle: + return np.random.permutation(len(self)) + else: + order = [np.arange(len(self))] + order.append(self.sizes) + return np.lexsort(order) + + @property + def supports_prefetch(self): + return getattr(self.dataset, "supports_prefetch", False) + + def prefetch(self, indices): + self.dataset.prefetch(indices) diff --git a/fairseq/data/legacy/masked_lm_dictionary.py b/fairseq/data/legacy/masked_lm_dictionary.py new file mode 100644 index 0000000000000000000000000000000000000000..dee88f7a3ed72ea465ea4e8ffe7b1c01ff6f57f1 --- /dev/null +++ b/fairseq/data/legacy/masked_lm_dictionary.py @@ -0,0 +1,60 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq.data import Dictionary + + +class MaskedLMDictionary(Dictionary): + """ + Dictionary for Masked Language Modelling tasks. This extends Dictionary by + adding the mask symbol. + """ + + def __init__( + self, + pad="<pad>", + eos="</s>", + unk="<unk>", + mask="<mask>", + ): + super().__init__(pad=pad, eos=eos, unk=unk) + self.mask_word = mask + self.mask_index = self.add_symbol(mask) + self.nspecial = len(self.symbols) + + def mask(self): + """Helper to get index of mask symbol""" + return self.mask_index + + +class BertDictionary(MaskedLMDictionary): + """ + Dictionary for BERT task. This extends MaskedLMDictionary by adding support + for cls and sep symbols. + """ + + def __init__( + self, + pad="<pad>", + eos="</s>", + unk="<unk>", + mask="<mask>", + cls="<cls>", + sep="<sep>", + ): + super().__init__(pad=pad, eos=eos, unk=unk, mask=mask) + self.cls_word = cls + self.sep_word = sep + self.cls_index = self.add_symbol(cls) + self.sep_index = self.add_symbol(sep) + self.nspecial = len(self.symbols) + + def cls(self): + """Helper to get index of cls symbol""" + return self.cls_index + + def sep(self): + """Helper to get index of sep symbol""" + return self.sep_index diff --git a/fairseq/data/list_dataset.py b/fairseq/data/list_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..12f00aa43661d6bad701c9e72653ba8779136906 --- /dev/null +++ b/fairseq/data/list_dataset.py @@ -0,0 +1,32 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from . import BaseWrapperDataset + + +class ListDataset(BaseWrapperDataset): + def __init__(self, dataset, sizes=None): + super().__init__(dataset) + self._sizes = sizes + + def __iter__(self): + for x in self.dataset: + yield x + + def collater(self, samples): + return samples + + @property + def sizes(self): + return self._sizes + + def num_tokens(self, index): + return self.sizes[index] + + def size(self, index): + return self.sizes[index] + + def set_epoch(self, epoch): + pass diff --git a/fairseq/data/lm_context_window_dataset.py b/fairseq/data/lm_context_window_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..1a945927cf0d96719003685676a990737a3762b2 --- /dev/null +++ b/fairseq/data/lm_context_window_dataset.py @@ -0,0 +1,97 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import numpy as np +import torch +from typing import Dict + +from fairseq.data.monolingual_dataset import MonolingualDataset + +from . import FairseqDataset + + +class LMContextWindowDataset(FairseqDataset): + """ + Wraps a MonolingualDataset and provides more context for evaluation. + + Each item in the new dataset will have a maximum size of + ``tokens_per_sample + context_window``. + + Args: + dataset: dataset to wrap + tokens_per_sample (int): the max number of tokens in each dataset item + context_window (int): the number of accumulated tokens to add to each + dataset item + pad_idx (int): padding symbol + """ + + def __init__( + self, + dataset: MonolingualDataset, + tokens_per_sample: int, + context_window: int, + pad_idx: int, + ): + assert context_window > 0 + self.dataset = dataset + self.tokens_per_sample = tokens_per_sample + self.context_window = context_window + self.pad_idx = pad_idx + self.prev_tokens = np.empty([0]) + + def __getitem__(self, index): + return self.dataset[index] + + def __len__(self): + return len(self.dataset) + + def collater(self, samples) -> Dict: + sample = self.dataset.collater(samples) + + pad = self.pad_idx + max_sample_len = self.tokens_per_sample + self.context_window + + bsz, tsz = sample["net_input"]["src_tokens"].shape + start_idxs = [0] * bsz + toks = sample["net_input"]["src_tokens"] + lengths = sample["net_input"]["src_lengths"] + tgt = sample["target"] + new_toks = np.empty([bsz, tsz + self.context_window], dtype=np.int64) + new_tgt = np.full([bsz, tsz + self.context_window], pad, dtype=np.int64) + sample_lens = toks.ne(pad).long().sum(dim=1).cpu() + for i in range(bsz): + sample_len = sample_lens[i] + extra = len(self.prev_tokens) + sample_len - max_sample_len + if extra > 0: + self.prev_tokens = self.prev_tokens[extra:] + pads = np.full(self.context_window - len(self.prev_tokens), pad) + new_toks[i] = np.concatenate([self.prev_tokens, toks[i].numpy(), pads]) + new_tgt[ + i, len(self.prev_tokens) : len(self.prev_tokens) + len(tgt[i]) + ] = tgt[i] + start_idxs[i] = len(self.prev_tokens) + lengths[i] += len(self.prev_tokens) + self.prev_tokens = new_toks[i][new_toks[i] != pad][-self.context_window :] + sample["net_input"]["src_tokens"] = torch.from_numpy(new_toks) + sample["target"] = torch.from_numpy(new_tgt) + sample["start_indices"] = start_idxs + return sample + + def num_tokens(self, index): + return self.dataset.num_tokens(index) + + def size(self, index): + return self.dataset.size(index) + + def ordered_indices(self): + # NOTE we don't shuffle the data to retain access to the previous dataset elements + return np.arange(len(self.dataset)) + + @property + def supports_prefetch(self): + return getattr(self.dataset, "supports_prefetch", False) + + def prefetch(self, indices): + return self.dataset.prefetch(indices) diff --git a/fairseq/data/lru_cache_dataset.py b/fairseq/data/lru_cache_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..a7854ac1701392754ce5795cafe9c634671aebdf --- /dev/null +++ b/fairseq/data/lru_cache_dataset.py @@ -0,0 +1,21 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from functools import lru_cache + +from . import BaseWrapperDataset + + +class LRUCacheDataset(BaseWrapperDataset): + def __init__(self, dataset, token=None): + super().__init__(dataset) + + @lru_cache(maxsize=8) + def __getitem__(self, index): + return self.dataset[index] + + @lru_cache(maxsize=8) + def collater(self, samples): + return self.dataset.collater(samples) diff --git a/fairseq/data/mask_tokens_dataset.py b/fairseq/data/mask_tokens_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..9123235594c3977994a3ae8a03ab4c9e395cc5de --- /dev/null +++ b/fairseq/data/mask_tokens_dataset.py @@ -0,0 +1,220 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from functools import lru_cache + +import numpy as np +import torch +from fairseq.data import Dictionary, data_utils + +from . import BaseWrapperDataset, LRUCacheDataset + + +class MaskTokensDataset(BaseWrapperDataset): + """ + A wrapper Dataset for masked language modeling. + + Input items are masked according to the specified masking probability. + + Args: + dataset: Dataset to wrap. + sizes: Sentence lengths + vocab: Dictionary with the vocabulary and special tokens. + pad_idx: Id of pad token in vocab + mask_idx: Id of mask token in vocab + return_masked_tokens: controls whether to return the non-masked tokens + (the default) or to return a tensor with the original masked token + IDs (and *pad_idx* elsewhere). The latter is useful as targets for + masked LM training. + seed: Seed for random number generator for reproducibility. + mask_prob: probability of replacing a token with *mask_idx*. + leave_unmasked_prob: probability that a masked token is unmasked. + random_token_prob: probability of replacing a masked token with a + random token from the vocabulary. + freq_weighted_replacement: sample random replacement words based on + word frequencies in the vocab. + mask_whole_words: only mask whole words. This should be a byte mask + over vocab indices, indicating whether it is the beginning of a + word. We will extend any mask to encompass the whole word. + bpe: BPE to use for whole-word masking. + mask_multiple_length : repeat each mask index multiple times. Default + value is 1. + mask_stdev : standard deviation of masks distribution in case of + multiple masking. Default value is 0. + """ + + @classmethod + def apply_mask(cls, dataset: torch.utils.data.Dataset, *args, **kwargs): + """Return the source and target datasets for masked LM training.""" + dataset = LRUCacheDataset(dataset) + return ( + LRUCacheDataset(cls(dataset, *args, **kwargs, return_masked_tokens=False)), + LRUCacheDataset(cls(dataset, *args, **kwargs, return_masked_tokens=True)), + ) + + def __init__( + self, + dataset: torch.utils.data.Dataset, + vocab: Dictionary, + pad_idx: int, + mask_idx: int, + return_masked_tokens: bool = False, + seed: int = 1, + mask_prob: float = 0.15, + leave_unmasked_prob: float = 0.1, + random_token_prob: float = 0.1, + freq_weighted_replacement: bool = False, + mask_whole_words: torch.Tensor = None, + mask_multiple_length: int = 1, + mask_stdev: float = 0.0, + ): + assert 0.0 < mask_prob < 1.0 + assert 0.0 <= random_token_prob <= 1.0 + assert 0.0 <= leave_unmasked_prob <= 1.0 + assert random_token_prob + leave_unmasked_prob <= 1.0 + assert mask_multiple_length >= 1 + assert mask_stdev >= 0.0 + + self.dataset = dataset + self.vocab = vocab + self.pad_idx = pad_idx + self.mask_idx = mask_idx + self.return_masked_tokens = return_masked_tokens + self.seed = seed + self.mask_prob = mask_prob + self.leave_unmasked_prob = leave_unmasked_prob + self.random_token_prob = random_token_prob + self.mask_whole_words = mask_whole_words + self.mask_multiple_length = mask_multiple_length + self.mask_stdev = mask_stdev + + if random_token_prob > 0.0: + if freq_weighted_replacement: + weights = np.array(self.vocab.count) + else: + weights = np.ones(len(self.vocab)) + weights[: self.vocab.nspecial] = 0 + self.weights = weights / weights.sum() + + self.epoch = 0 + + @property + def can_reuse_epoch_itr_across_epochs(self): + return True # only the noise changes, not item sizes + + def set_epoch(self, epoch, **unused): + super().set_epoch(epoch) + self.epoch = epoch + + def __getitem__(self, index: int): + return self.__getitem_cached__(self.seed, self.epoch, index) + + @lru_cache(maxsize=8) + def __getitem_cached__(self, seed: int, epoch: int, index: int): + with data_utils.numpy_seed(self.seed, self.epoch, index): + item = self.dataset[index] + sz = len(item) + + assert ( + self.mask_idx not in item + ), "Dataset contains mask_idx (={}), this is not expected!".format( + self.mask_idx, + ) + + if self.mask_whole_words is not None: + word_begins_mask = self.mask_whole_words.gather(0, item) + word_begins_idx = word_begins_mask.nonzero().view(-1) + sz = len(word_begins_idx) + words = np.split(word_begins_mask, word_begins_idx)[1:] + assert len(words) == sz + word_lens = list(map(len, words)) + + # decide elements to mask + mask = np.full(sz, False) + num_mask = int( + # add a random number for probabilistic rounding + self.mask_prob * sz / float(self.mask_multiple_length) + + np.random.rand() + ) + + # multiple masking as described in the vq-wav2vec paper (https://arxiv.org/abs/1910.05453) + mask_idc = np.random.choice(sz, num_mask, replace=False) + if self.mask_stdev > 0.0: + lengths = np.random.normal( + self.mask_multiple_length, self.mask_stdev, size=num_mask + ) + lengths = [max(0, int(round(x))) for x in lengths] + mask_idc = np.asarray( + [ + mask_idc[j] + offset + for j in range(len(mask_idc)) + for offset in range(lengths[j]) + ], + dtype=np.int64, + ) + else: + mask_idc = np.concatenate( + [mask_idc + i for i in range(self.mask_multiple_length)] + ) + mask_idc = mask_idc[mask_idc < len(mask)] + try: + mask[mask_idc] = True + except: # something wrong + print( + "Assigning mask indexes {} to mask {} failed!".format( + mask_idc, mask + ) + ) + raise + + if self.return_masked_tokens: + # exit early if we're just returning the masked tokens + # (i.e., the targets for masked LM training) + if self.mask_whole_words is not None: + mask = np.repeat(mask, word_lens) + new_item = np.full(len(mask), self.pad_idx) + new_item[mask] = item[torch.from_numpy(mask.astype(np.uint8)) == 1] + return torch.from_numpy(new_item) + + # decide unmasking and random replacement + rand_or_unmask_prob = self.random_token_prob + self.leave_unmasked_prob + if rand_or_unmask_prob > 0.0: + rand_or_unmask = mask & (np.random.rand(sz) < rand_or_unmask_prob) + if self.random_token_prob == 0.0: + unmask = rand_or_unmask + rand_mask = None + elif self.leave_unmasked_prob == 0.0: + unmask = None + rand_mask = rand_or_unmask + else: + unmask_prob = self.leave_unmasked_prob / rand_or_unmask_prob + decision = np.random.rand(sz) < unmask_prob + unmask = rand_or_unmask & decision + rand_mask = rand_or_unmask & (~decision) + else: + unmask = rand_mask = None + + if unmask is not None: + mask = mask ^ unmask + + if self.mask_whole_words is not None: + mask = np.repeat(mask, word_lens) + + new_item = np.copy(item) + new_item[mask] = self.mask_idx + if rand_mask is not None: + num_rand = rand_mask.sum() + if num_rand > 0: + if self.mask_whole_words is not None: + rand_mask = np.repeat(rand_mask, word_lens) + num_rand = rand_mask.sum() + + new_item[rand_mask] = np.random.choice( + len(self.vocab), + num_rand, + p=self.weights, + ) + + return torch.from_numpy(new_item) diff --git a/fairseq/data/monolingual_dataset.py b/fairseq/data/monolingual_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..54fd583b64a3a475324ade6eaaeccf593d747fdc --- /dev/null +++ b/fairseq/data/monolingual_dataset.py @@ -0,0 +1,253 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import numpy as np +import torch + +from . import FairseqDataset, data_utils + + +def collate(samples, pad_idx, eos_idx, fixed_pad_length=None, pad_to_bsz=None): + if len(samples) == 0: + return {} + + def merge(key, is_list=False): + if is_list: + res = [] + for i in range(len(samples[0][key])): + res.append( + data_utils.collate_tokens( + [s[key][i] for s in samples], + pad_idx, + eos_idx, + left_pad=False, + pad_to_length=fixed_pad_length, + pad_to_bsz=pad_to_bsz, + ) + ) + return res + else: + return data_utils.collate_tokens( + [s[key] for s in samples], + pad_idx, + eos_idx, + left_pad=False, + pad_to_length=fixed_pad_length, + pad_to_bsz=pad_to_bsz, + ) + + src_tokens = merge("source") + if samples[0]["target"] is not None: + is_target_list = isinstance(samples[0]["target"], list) + target = merge("target", is_target_list) + else: + target = src_tokens + + return { + "id": torch.LongTensor([s["id"] for s in samples]), + "nsentences": len(samples), + "ntokens": sum(len(s["source"]) for s in samples), + "net_input": { + "src_tokens": src_tokens, + "src_lengths": torch.LongTensor([s["source"].numel() for s in samples]), + }, + "target": target, + } + + +class MonolingualDataset(FairseqDataset): + """ + A wrapper around torch.utils.data.Dataset for monolingual data. + + Args: + dataset (torch.utils.data.Dataset): dataset to wrap + sizes (List[int]): sentence lengths + vocab (~fairseq.data.Dictionary): vocabulary + shuffle (bool, optional): shuffle the elements before batching + (default: True). + """ + + def __init__( + self, + dataset, + sizes, + src_vocab, + tgt_vocab=None, + add_eos_for_other_targets=False, + shuffle=False, + targets=None, + add_bos_token=False, + fixed_pad_length=None, + pad_to_bsz=None, + src_lang_idx=None, + tgt_lang_idx=None, + ): + self.dataset = dataset + self.sizes = np.array(sizes) + self.vocab = src_vocab + self.tgt_vocab = tgt_vocab or src_vocab + self.add_eos_for_other_targets = add_eos_for_other_targets + self.shuffle = shuffle + self.add_bos_token = add_bos_token + self.fixed_pad_length = fixed_pad_length + self.pad_to_bsz = pad_to_bsz + self.src_lang_idx = src_lang_idx + self.tgt_lang_idx = tgt_lang_idx + + assert targets is None or all( + t in {"self", "future", "past"} for t in targets + ), "targets must be none or one of 'self', 'future', 'past'" + if targets is not None and len(targets) == 0: + targets = None + self.targets = targets + + def __getitem__(self, index): + if self.targets is not None: + # *future_target* is the original sentence + # *source* is shifted right by 1 (maybe left-padded with eos) + # *past_target* is shifted right by 2 (left-padded as needed) + # + # Left-to-right language models should condition on *source* and + # predict *future_target*. + # Right-to-left language models should condition on *source* and + # predict *past_target*. + source, future_target, past_target = self.dataset[index] + source, target = self._make_source_target( + source, future_target, past_target + ) + else: + source = self.dataset[index] + target = None + source, target = self._maybe_add_bos(source, target) + return {"id": index, "source": source, "target": target} + + def __len__(self): + return len(self.dataset) + + def _make_source_target(self, source, future_target, past_target): + if self.targets is not None: + target = [] + + if ( + self.add_eos_for_other_targets + and (("self" in self.targets) or ("past" in self.targets)) + and source[-1] != self.vocab.eos() + ): + # append eos at the end of source + source = torch.cat([source, source.new([self.vocab.eos()])]) + + if "future" in self.targets: + future_target = torch.cat( + [future_target, future_target.new([self.vocab.pad()])] + ) + if "past" in self.targets: + # first token is before the start of sentence which is only used in "none" break mode when + # add_eos_for_other_targets is False + past_target = torch.cat( + [ + past_target.new([self.vocab.pad()]), + past_target[1:], + source[-2, None], + ] + ) + + for t in self.targets: + if t == "self": + target.append(source) + elif t == "future": + target.append(future_target) + elif t == "past": + target.append(past_target) + else: + raise Exception("invalid target " + t) + + if len(target) == 1: + target = target[0] + else: + target = future_target + + return source, self._filter_vocab(target) + + def _maybe_add_bos(self, source, target): + if self.add_bos_token: + source = torch.cat([source.new([self.vocab.bos()]), source]) + if target is not None: + target = torch.cat([target.new([self.tgt_vocab.bos()]), target]) + return source, target + + def num_tokens_vec(self, indices): + """Return the number of tokens for a set of positions defined by indices. + This value is used to enforce ``--max-tokens`` during batching.""" + return self.sizes[indices] + + def _filter_vocab(self, target): + if len(self.tgt_vocab) != len(self.vocab): + + def _filter(target): + mask = target.ge(len(self.tgt_vocab)) + if mask.any(): + target[mask] = self.tgt_vocab.unk() + return target + + if isinstance(target, list): + return [_filter(t) for t in target] + return _filter(target) + return target + + def collater(self, samples): + """Merge a list of samples to form a mini-batch. + + Args: + samples (List[dict]): samples to collate + + Returns: + dict: a mini-batch with the following keys: + + - `id` (LongTensor): example IDs in the original input order + - `ntokens` (int): total number of tokens in the batch + - `net_input` (dict): the input to the Model, containing keys: + + - `src_tokens` (LongTensor): a padded 2D Tensor of tokens in + the source sentence of shape `(bsz, src_len)`. Padding will + appear on the right. + + - `target` (LongTensor): a padded 2D Tensor of tokens in the + target sentence of shape `(bsz, tgt_len)`. Padding will appear + on the right. + """ + return collate( + samples, + self.vocab.pad(), + self.vocab.eos(), + self.fixed_pad_length, + self.pad_to_bsz, + ) + + def num_tokens(self, index): + """Return the number of tokens in a sample. This value is used to + enforce ``--max-tokens`` during batching.""" + return self.sizes[index] + + def size(self, index): + """Return an example's size as a float or tuple. This value is used when + filtering a dataset with ``--max-positions``.""" + return self.sizes[index] + + def ordered_indices(self): + """Return an ordered list of indices. Batches will be constructed based + on this order.""" + if self.shuffle: + order = [np.random.permutation(len(self))] + else: + order = [np.arange(len(self))] + order.append(self.sizes) + return np.lexsort(order) + + @property + def supports_prefetch(self): + return getattr(self.dataset, "supports_prefetch", False) + + def prefetch(self, indices): + self.dataset.prefetch(indices) diff --git a/fairseq/data/multi_corpus_dataset.py b/fairseq/data/multi_corpus_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..1bd61c32ebcc57759c210f320dd8ac7386c6193d --- /dev/null +++ b/fairseq/data/multi_corpus_dataset.py @@ -0,0 +1,240 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import time +from collections import OrderedDict +from typing import Dict, List + +import numpy as np +from fairseq.data import data_utils + +from . import FairseqDataset + +logger = logging.getLogger(__name__) + + +class MultiCorpusDataset(FairseqDataset): + """ + Stores multiple instances of FairseqDataset together. Requires each instance + to be the same dataset, as the collate method needs to work on batches with + samples from each dataset. + + Allows specifying a distribution over the datasets to use. Note that unlike + MultiCorpusSampledDataset, this distribution allows sampling for each item, + rather than on a batch level. + + Each time ordered_indices() is called, a new sample is generated with + the specified distribution. + + Args: + datasets: a OrderedDict of FairseqDataset instances. + distribution: a List containing the probability of getting an utterance from + corresponding dataset + seed: random seed for sampling the datsets + sort_indices: if true, will sort the ordered indices by size + batch_sample: if true, will ensure each batch is from a single dataset + """ + + def __init__( + self, + datasets: Dict[str, FairseqDataset], + distribution: List[float], + seed: int, + sort_indices: bool = False, + batch_sample: bool = False, + distributed_rank=None, + ): + super().__init__() + assert isinstance(datasets, OrderedDict) + assert len(datasets) == len(distribution) + assert sum(distribution) == 1 + self.datasets = datasets + self.distribution = distribution + self.seed = seed + self.sort_indices = sort_indices + self.batch_sample = batch_sample + self.distributed_rank = distributed_rank + + # Avoid repeated conversions to list later + self.dataset_list = list(datasets.values()) + self.total_num_instances = 0 + + first_dataset = list(self.datasets.values())[0] + + self.dataset_offsets = [] + for dataset in datasets.values(): + assert isinstance(dataset, FairseqDataset) + assert type(dataset) is type(first_dataset) + self.dataset_offsets.append(self.total_num_instances) + self.total_num_instances += len(dataset) + + def ordered_indices(self): + start = time.time() + with data_utils.numpy_seed(self.seed, self.epoch): + logger.info(f"sampling new dataset with seed {self.seed} epoch {self.epoch}") + sampled_indices = [] + num_selected_instances = 0 + + # For each dataset i, sample self.distribution[i] * self.total_num_instances + for i, key in enumerate(self.datasets): + + if i < len(self.datasets) - 1: + num_instances = int(self.distribution[i] * self.total_num_instances) + high = self.dataset_offsets[i + 1] + else: + num_instances = self.total_num_instances - num_selected_instances + high = self.total_num_instances + + logger.info(f"sampling {num_instances} from {key} dataset") + num_selected_instances += num_instances + + # First, add k copies of the dataset where k = num_instances // len(dataset). + # This ensures an equal distribution of the data points as much as possible. + # For the remaining entries randomly sample them + dataset_size = len(self.datasets[key]) + num_copies = num_instances // dataset_size + dataset_indices = ( + np.random.permutation(high - self.dataset_offsets[i]) + + self.dataset_offsets[i] + )[: num_instances - num_copies * dataset_size] + if num_copies > 0: + sampled_indices += list( + np.concatenate( + ( + np.repeat( + np.arange(self.dataset_offsets[i], high), num_copies + ), + dataset_indices, + ) + ) + ) + else: + sampled_indices += list(dataset_indices) + + assert ( + len(sampled_indices) == self.total_num_instances + ), f"{len(sampled_indices)} vs {self.total_num_instances}" + + np.random.shuffle(sampled_indices) + if self.sort_indices: + sampled_indices.sort(key=lambda i: self.num_tokens(i)) + + logger.info( + "multi_corpus_dataset ordered_indices took {}s".format( + time.time() - start + ) + ) + return np.array(sampled_indices, dtype=np.int64) + + def _map_index(self, index: int): + """ + If dataset A has length N and dataset B has length M + then index 1 maps to index 1 of dataset A, and index N + 1 + maps to index 1 of B. + """ + counter = 0 + for key, dataset in self.datasets.items(): + if index < counter + len(dataset): + return index - counter, key + counter += len(dataset) + raise ValueError( + "Invalid index: {}, max: {}".format(index, self.total_num_instances) + ) + + def __len__(self): + """ + Length of this dataset is the sum of individual datasets + """ + return self.total_num_instances + + def __getitem__(self, index): + new_index, key = self._map_index(index) + try: + item = self.datasets[key][new_index] + item["full_id"] = index + return item + except Exception as e: + e.args = (f"Error from {key} dataset", *e.args) + raise + + def collater(self, samples): + """ + If we are doing batch sampling, then pick the right collater to use. + + Otherwise we assume all collaters are the same. + """ + if len(samples) == 0: + return None + if "full_id" in samples[0]: + _, key = self._map_index(samples[0]["full_id"]) + return self.datasets[key].collater(samples) + else: + # Subclasses may override __getitem__ to not specify full_id + return list(self.datasets.values())[0].collater(samples) + + def num_tokens(self, index: int): + index, key = self._map_index(index) + return self.datasets[key].num_tokens(index) + + def size(self, index: int): + index, key = self._map_index(index) + return self.datasets[key].size(index) + + @property + def can_reuse_epoch_itr_across_epochs(self): + return False + + def set_epoch(self, epoch, **unused): + super().set_epoch(epoch) + logger.info(f"setting epoch of multi_corpus_dataset to {epoch}") + self.epoch = epoch + + @property + def supports_prefetch(self): + return False + + @property + def supports_fetch_outside_dataloader(self): + return all( + self.datasets[key].supports_fetch_outside_dataloader + for key in self.datasets + ) + + def batch_by_size( + self, + indices, + max_tokens=None, + max_sentences=None, + required_batch_size_multiple=1, + ): + if not self.batch_sample: + return super().batch_by_size( + indices, max_tokens, max_sentences, required_batch_size_multiple + ) + + dataset_indices = {key: [] for key in self.datasets} + for i in indices: + _, key = self._map_index(i) + dataset_indices[key].append(i) + + batches = [] + for key in dataset_indices: + cur_batches = super().batch_by_size( + np.array(dataset_indices[key], dtype=np.int64), + max_tokens, + max_sentences, + required_batch_size_multiple, + ) + logger.info(f"Created {len(cur_batches)} batches for dataset {key}") + batches += cur_batches + + # If this dataset is used in a distributed training setup, + # then shuffle such that the order is seeded by the distributed rank + # as well + if self.distributed_rank is not None: + with data_utils.numpy_seed(self.seed, self.epoch, self.distributed_rank): + np.random.shuffle(batches) + return batches diff --git a/fairseq/data/multi_corpus_sampled_dataset.py b/fairseq/data/multi_corpus_sampled_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..e2e9fdf004dd1da519a170a5e8bc225775776f72 --- /dev/null +++ b/fairseq/data/multi_corpus_sampled_dataset.py @@ -0,0 +1,152 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from collections import OrderedDict +from typing import Callable, Dict, List + +import numpy as np + +from . import FairseqDataset + + +def uniform_sampler(x): + # Sample from uniform distribution + return np.random.choice(x, 1).item() + + +class MultiCorpusSampledDataset(FairseqDataset): + """ + Stores multiple instances of FairseqDataset together and in every iteration + creates a batch by first sampling a dataset according to a specified + probability distribution and then getting instances from that dataset. + + Args: + datasets: an OrderedDict of FairseqDataset instances. + sampling_func: A function for sampling over list of dataset keys. + The default strategy is to sample uniformly. + """ + + def __init__( + self, + datasets: Dict[str, FairseqDataset], + sampling_func: Callable[[List], int] = None, + ): + super().__init__() + assert isinstance(datasets, OrderedDict) + self.datasets = datasets + if sampling_func is None: + sampling_func = uniform_sampler + self.sampling_func = sampling_func + + self.total_num_instances = 0 + for _, dataset in datasets.items(): + assert isinstance(dataset, FairseqDataset) + self.total_num_instances += len(dataset) + + self._ordered_indices = None + + def __len__(self): + """ + Length of this dataset is the sum of individual datasets + """ + return self.total_num_instances + + def ordered_indices(self): + """ + Ordered indices for batching. Here we call the underlying + dataset's ordered_indices() so that we get the same random ordering + as we would have from using the underlying dataset directly. + """ + if self._ordered_indices is None: + self._ordered_indices = OrderedDict( + [ + (key, dataset.ordered_indices()) + for key, dataset in self.datasets.items() + ] + ) + return np.arange(len(self)) + + def _map_index_to_dataset(self, key: int, index: int): + """ + Different underlying datasets have different lengths. In order to ensure + we are not accessing an index outside the range of the current dataset + size, we wrap around. This function should be called after we have + created an ordering for this and all underlying datasets. + """ + assert ( + self._ordered_indices is not None + ), "Must call MultiCorpusSampledDataset.ordered_indices() first" + mapped_index = index % len(self.datasets[key]) + return self._ordered_indices[key][mapped_index] + + def __getitem__(self, index: int): + """ + Get the item associated with index from each underlying dataset. + Since index is in the range of [0, TotalNumInstances], we need to + map the index to the dataset before retrieving the item. + """ + return OrderedDict( + [ + (key, dataset[self._map_index_to_dataset(key, index)]) + for key, dataset in self.datasets.items() + ] + ) + + def collater(self, samples: List[Dict]): + """ + Generate a mini-batch for this dataset. + To convert this into a regular mini-batch we use the following + logic: + 1. Select a dataset using the specified probability distribution. + 2. Call the collater function of the selected dataset. + """ + if len(samples) == 0: + return None + + selected_key = self.sampling_func(list(self.datasets.keys())) + selected_samples = [sample[selected_key] for sample in samples] + return self.datasets[selected_key].collater(selected_samples) + + def num_tokens(self, index: int): + """ + Return an example's length (number of tokens), used for batching. Here + we return the max across all examples at index across all underlying + datasets. + """ + return max( + dataset.num_tokens(self._map_index_to_dataset(key, index)) + for key, dataset in self.datasets.items() + ) + + def size(self, index: int): + """ + Return an example's size as a float or tuple. Here we return the max + across all underlying datasets. This value is used when filtering a + dataset with max-positions. + """ + return max( + dataset.size(self._map_index_to_dataset(key, index)) + for key, dataset in self.datasets.items() + ) + + @property + def supports_prefetch(self): + return all( + getattr(dataset, "supports_prefetch", False) + for dataset in self.datasets.values() + ) + + def prefetch(self, indices): + for key, dataset in self.datasets.items(): + dataset.prefetch( + [self._map_index_to_dataset(key, index) for index in indices] + ) + + @property + def supports_fetch_outside_dataloader(self): + return all( + self.datasets[key].supports_fetch_outside_dataloader + for key in self.datasets + ) diff --git a/fairseq/data/multilingual/__init__.py b/fairseq/data/multilingual/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..6264236915a7269a4d920ee8213004374dd86a9a --- /dev/null +++ b/fairseq/data/multilingual/__init__.py @@ -0,0 +1,4 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. diff --git a/fairseq/data/multilingual/multilingual_data_manager.py b/fairseq/data/multilingual/multilingual_data_manager.py new file mode 100644 index 0000000000000000000000000000000000000000..a2fae5bf520153a6f41c4bd3410c691c239f9521 --- /dev/null +++ b/fairseq/data/multilingual/multilingual_data_manager.py @@ -0,0 +1,1131 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import itertools +import json +import logging +import math +import os +from collections import OrderedDict, defaultdict + +from fairseq import utils +from fairseq.data import ( + AppendTokenDataset, + ConcatDataset, + Dictionary, + LanguagePairDataset, + PrependTokenDataset, + SampledMultiDataset, + SampledMultiEpochDataset, + StripTokenDataset, + TransformEosLangPairDataset, + TruncateDataset, + data_utils, + indexed_dataset, +) +from fairseq.data.multilingual.multilingual_utils import ( + EncoderLangtok, + LangTokSpec, + LangTokStyle, + augment_dictionary, + get_lang_tok, +) +from fairseq.data.multilingual.sampled_multi_dataset import CollateFormat +from fairseq.file_io import PathManager +from fairseq.utils import FileContentsAction, csv_str_list, eval_str_dict + + +logger = logging.getLogger(__name__) + +SRC_DICT_NAME = 'src' +TGT_DICT_NAME = 'tgt' + + +def _lang_id(dic: Dictionary, lang: str): + """Return language ID index.""" + idx = dic.index(lang) + assert idx != dic.unk_index, "cannot find language ID for lang {}".format(lang) + return idx + + +def load_sampling_weights(from_file): + with open(from_file) as f: + weights = json.load(f) + return weights + + +class MultilingualDatasetManager(object): + def __init__(self, args, lang_pairs, langs, dicts, sampling_method): + super().__init__() + self.args = args + self.seed = args.seed + self.lang_pairs = lang_pairs + self.extra_lang_pairs = ( + list( + {p for _, v in args.extra_lang_pairs.items() for p in v.split(",")} + ) + if args.extra_lang_pairs + else [] + ) + self.src_langs = {p.split("-")[0] for p in args.lang_pairs + self.extra_lang_pairs} + self.tgt_langs = {p.split("-")[1] for p in args.lang_pairs + self.extra_lang_pairs} + self.langs = langs + self.dicts = dicts + self.lang_dict = self.create_lang_dictionary(self.langs) + self.sampling_method = sampling_method + self.sampling_scheduler = None + self._has_sharded_data = False + self._num_shards_dict = {} + self._training_data_sizes = defaultdict(lambda: {}) + + @classmethod + def setup_data_manager(cls, args, lang_pairs, langs, dicts, sampling_method): + return MultilingualDatasetManager( + args, lang_pairs, langs, dicts, sampling_method + ) + + @staticmethod + def add_args(parser): + parser.add_argument( + "data", + help="colon separated path to data directories list, \ + will be iterated upon during epochs in round-robin manner", + action=FileContentsAction, + ) + parser.add_argument( + "--langs", + default=None, + type=csv_str_list, + help="a list of languages comma sperated languages which can appear in lang-pairs; " + "note that the ordering determines language token IDs", + ) + parser.add_argument( + "--lang-dict", + default=None, + type=str, + help="an external file which contains a list of " + "languages which can appear in lang-pairs; " + "note that the ordering determines language token IDs; " + "--langs and --lang-dict are two exclusive options", + ) + parser.add_argument('--source-dict', default=None, type=str, + help='path to source dictionary; if specified it will override per language dictionary loading') + parser.add_argument('--target-dict', default=None, type=str, + help='path to target dictionary; if specified it will override per language dictionary loading') + parser.add_argument( + "--lang-tok-style", + default=LangTokStyle.multilingual.value, + type=str, + choices=[LangTokStyle.multilingual.value, LangTokStyle.mbart.value], + help="language token styles", + ) + + parser.add_argument( + "--load-alignments", + action="store_true", + help="load the binarized alignments", + ) + parser.add_argument( + "--left-pad-source", + default="True", + type=str, + metavar="BOOL", + help="pad the source on the left", + ) + parser.add_argument( + "--left-pad-target", + default="False", + type=str, + metavar="BOOL", + help="pad the target on the left", + ) + parser.add_argument( + "--max-source-positions", + default=1024, + type=int, + metavar="N", + help="max number of tokens in the source sequence", + ) + parser.add_argument( + "--max-target-positions", + default=1024, + type=int, + metavar="N", + help="max number of tokens in the target sequence", + ) + parser.add_argument( + "--upsample-primary", + default=1, + type=int, + help="amount to upsample primary dataset", + ) + parser.add_argument( + "--truncate-source", + action="store_true", + default=False, + help="truncate source to max-source-positions", + ) + parser.add_argument( + "--encoder-langtok", + default=None, + type=str, + choices=[EncoderLangtok.src.value, EncoderLangtok.tgt.value], + metavar="SRCTGT", + help="prepend to the beginning of source sentence the source or target " + "language token. (src/tgt)", + ) + parser.add_argument( + "--decoder-langtok", + action="store_true", + help="prepend to the beginning of target sentence the target language token", + ) + parser.add_argument( + "--lang-tok-replacing-bos-eos", action="store_true", default=False + ) + parser.add_argument( + "--enable-lang-ids", + default=False, + action="store_true", + help="whether to include language IDs in samples", + ) + parser.add_argument( + "--enable-reservsed-directions-shared-datasets", + default=False, + action="store_true", + help="whether to allow datasets be used in reversed directions", + ) + + parser.add_argument( + "--extra-data", + help='a dictionary of data name to this path, \ + e.g. {"mined", path_to_mined_data, "denoised": path_to_denoised_data}', + type=lambda uf: eval_str_dict(uf, type=str), + default=None, + ) + parser.add_argument( + "--extra-lang-pairs", + help='a dictionary of data name to the language pairs they serve, \ + e.g. {"mined": comma-separated-lang-pairs, "denoised": comma-separated-lang-pairs}', + type=lambda uf: eval_str_dict(uf, type=str), + default=None, + ) + parser.add_argument( + "--fixed-dictionary", + help="Fixed dictionary to use with model path", + default=None, + type=str, + ) + parser.add_argument( + "--langtoks-specs", + help='a list of comma separated data types that a set of language tokens to be specialized for, \ + e.g. "main,dae,mined". There will be a set of language tokens added to the vocab to \ + distinguish languages in different training data types. If not specified, default language \ + tokens per languages will be added', + default=LangTokSpec.main.value, + type=csv_str_list, + ) + parser.add_argument( + "--langtoks", + help='a dictionary of how to add language tokens, \ + e.g. {"mined": (None, "tgt"), "mono_dae": ("src.dae", "tgt"), "main": \ + ("src", "tgt")}, or {"mined": ("src.mined", "tgt")}', + default=None, + type=lambda uf: eval_str_dict(uf, type=str), + ) + parser.add_argument( + "--sampling-weights-from-file", + help='a file contain a python dictionary of how to sample data sets, \ + e.g. { "main:en_XX-es_XX": 0.2, "mined:en_XX-pt_XX": 0.5, \ + "mono_dae:es_XX-es_XX: 0.3, "main:en_xx-fr_XX": 0.8 }', + default=None, + type=str, + ) + parser.add_argument( + "--sampling-weights", + help='a dictionary of how to sample data sets, \ + e.g. { "main:en_XX-es_XX": 0.2, "mined:en_XX-pt_XX": 0.5, \ + "mono_dae:es_XX-es_XX: 0.3, "main:en_xx-fr_XX": 0.8 }', + default=None, + type=lambda uf: eval_str_dict(uf, type=str), + ) + parser.add_argument( + "--virtual-epoch-size", + default=None, + type=int, + help="virtual epoch size to speed up data loading", + ) + parser.add_argument( + "--virtual-data-size", + default=None, + type=int, + help="virtual data size of the whole joint dataset to speed" + "up data loading and have specific dynamic sampling strategy interval", + ) + + @classmethod + def load_langs(cls, args, **kwargs): + if args.lang_dict and args.langs: + raise ValueError("--langs and --lang-dict can not both be specified") + if args.lang_dict is None and args.langs is None: + logger.warning( + "External language dictionary is not provided; " + "use lang-pairs to infer the set of supported languages. " + "The language ordering is not stable which might cause " + "misalignment in pretraining and finetuning." + ) + # infer from lang_pairs as it is + langs = list( + {x for lang_pair in args.lang_pairs for x in lang_pair.split("-")} + ) + langs = sorted(langs) + logger.info(f"inferred language list: {langs}") + elif args.lang_dict: + with open( + PathManager.get_local_path(args.lang_dict), "r", encoding="utf-8" + ) as f: + langs = [lang.strip() for lang in f.readlines() if lang.strip()] + logger.info( + f"loaded language list from {args.lang_dict} as they are ordered in file" + ) + elif args.langs: + langs = args.langs + logger.info( + f"parsed the language list as they are ordered in the option: {langs}" + ) + return langs + + def has_sharded_data(self, split): + return self._has_sharded_data and split == getattr( + self.args, "train_subset", None + ) + + def _shared_collater(self): + return not (self.args.extra_data and "mono_dae" in self.args.extra_data) and ( + not self.args.lang_tok_replacing_bos_eos + ) + + def estimate_global_pass_epoch(self, epoch): + if self.args.virtual_epoch_size is None or self.args.virtual_data_size is None: + return None + # one epoch more for remaining data in each shard + virtual_epochs_per_shard = math.ceil( + self.args.virtual_data_size / self.args.virtual_epoch_size + ) + # note that fairseq epoch / shard_epoch starts from 1 + shard_epoch = (epoch - 1) // virtual_epochs_per_shard + 1 + return shard_epoch + + @classmethod + def prepare(cls, load_dictionary, args, **kargs): + args.left_pad_source = utils.eval_bool(args.left_pad_source) + args.left_pad_target = utils.eval_bool(args.left_pad_target) + + if not hasattr(args, "shuffle_instance"): + args.shuffle_instance = False + if args.langtoks is None: + args.langtoks = {} + if "main" not in args.langtoks: + src_langtok_spec = args.encoder_langtok if args.encoder_langtok else None + tgt_langtok_spec = "tgt" if args.decoder_langtok else None + args.langtoks["main"] = (src_langtok_spec, tgt_langtok_spec) + + def check_langs(langs, pairs): + messages = [] + for src, tgt in pairs: + if src not in langs or tgt not in langs: + messages.append( + f"language pair {src}-{tgt} contains languages " + "that are not in the language dictionary" + ) + if len(messages) > 0: + raise ValueError(" ".join(messages) + f"; langs: {langs}") + + if args.lang_pairs is None: + raise ValueError( + "--lang-pairs is required. List all the language pairs in the training objective." + ) + if isinstance(args.lang_pairs, str): + args.lang_pairs = args.lang_pairs.split(",") + if args.source_lang is not None or args.target_lang is not None: + training = False + else: + training = True + language_list = cls.load_langs(args, **kargs) + check_langs( + language_list, + ( + [p.split("-") for p in args.lang_pairs] + if training + else [(args.source_lang, args.target_lang)] + ), + ) + + def load_dictionary_and_postproc(path): + d = load_dictionary(path) + augment_dictionary( + dictionary=d, + language_list=language_list, + lang_tok_style=args.lang_tok_style, + langtoks_specs=args.langtoks_specs, + extra_data=args.extra_data, + ) + return d + + dicts = cls.load_all_dictionaries(args, language_list, load_dictionary_and_postproc, training) + return language_list, dicts, training + + @classmethod + def load_all_dictionaries(cls, args, language_list, load_dictionary, training): + dicts = OrderedDict() + if args.source_dict is not None: + dicts[SRC_DICT_NAME] = load_dictionary(args.source_dict) + if args.target_dict is not None: + dicts[TGT_DICT_NAME] = load_dictionary(args.target_dict) + + if training: + extra_lang_pairs = ( + list( + {p for _, v in args.extra_lang_pairs.items() for p in v.split(",")} + ) + if args.extra_lang_pairs + else [] + ) + src_langs_to_load_dicts = sorted( + {p.split("-")[0] for p in (args.lang_pairs + extra_lang_pairs)} + ) + tgt_langs_to_load_dicts = sorted( + {p.split("-")[1] for p in (args.lang_pairs + extra_lang_pairs)} + ) + else: + src_langs_to_load_dicts = [args.source_lang] + tgt_langs_to_load_dicts = [args.target_lang] + + paths = utils.split_paths(args.data) + assert len(paths) > 0 + + def load_dicts(langs_to_load_dicts): + for lang in langs_to_load_dicts: + dicts[lang] = load_dictionary( + os.path.join(paths[0], "dict.{}.txt".format(lang)) + ) + if len(dicts) > 0: + dict0 = next(iter(dicts.values())) + assert dicts[lang].pad() == dict0.pad() + assert dicts[lang].eos() == dict0.eos() + assert dicts[lang].unk() == dict0.unk() + logger.info("[{}] dictionary: {} types".format(lang, len(dicts[lang]))) + + if args.fixed_dictionary is not None: + fixed_dict = load_dictionary(args.fixed_dictionary) + dicts = {lang: fixed_dict for lang in src_langs_to_load_dicts + tgt_langs_to_load_dicts} + else: + if args.source_dict is None: + load_dicts(src_langs_to_load_dicts) + if args.target_dict is None: + load_dicts(tgt_langs_to_load_dicts) + return dicts + + def get_source_dictionary(self, lang): + if self.args.source_dict is not None: + return self.dicts[SRC_DICT_NAME] + else: + return self.dicts[lang] + + def get_target_dictionary(self, lang): + if self.args.target_dict is not None: + return self.dicts[TGT_DICT_NAME] + else: + return self.dicts[lang] + + @classmethod + def create_lang_dictionary(cls, langs): + unk = "<unk>" + # hack to remove symbols other than unk as they are not needed by lang dict + lang_dict = Dictionary(pad=unk, eos=unk, unk=unk, bos=unk) + for lang in langs: + lang_dict.add_symbol(lang) + return lang_dict + + @classmethod + def get_langtok_index(cls, lang_tok, dic): + idx = dic.index(lang_tok) + assert ( + idx != dic.unk_index + ), "cannot find language token {} in the dictionary".format(lang_tok) + return idx + + def get_encoder_langtok(self, src_lang, tgt_lang, spec=None): + if spec is None: + return None + if spec and spec.startswith("src"): + if src_lang is None: + return None + langtok = get_lang_tok( + lang=src_lang, lang_tok_style=self.args.lang_tok_style, spec=spec + ) + else: + if tgt_lang is None: + return None + langtok = get_lang_tok( + lang=tgt_lang, lang_tok_style=self.args.lang_tok_style, spec=spec + ) + return self.get_langtok_index( + langtok, self.get_source_dictionary(src_lang) if src_lang else self.get_target_dictionary(tgt_lang) + ) + + def get_decoder_langtok(self, tgt_lang, spec=None): + if spec is None: + return None + langtok = get_lang_tok( + lang=tgt_lang, lang_tok_style=self.args.lang_tok_style, spec=spec + ) + return self.get_langtok_index(langtok, self.get_target_dictionary(tgt_lang)) + + @classmethod + def load_data(cls, path, vdict, impl): + dataset = data_utils.load_indexed_dataset(path, vdict, impl) + return dataset + + @classmethod + def split_exists(cls, split, src, tgt, lang, data_path, dataset_impl): + filename = os.path.join(data_path, "{}.{}-{}.{}".format(split, src, tgt, lang)) + return indexed_dataset.dataset_exists(filename, impl=dataset_impl) + + def load_lang_dataset( + self, + data_path, + split, + src, + src_dict, + tgt, + tgt_dict, + combine, + dataset_impl, + upsample_primary, + max_source_positions, + prepend_bos=False, + load_alignments=False, + truncate_source=False, + ): + + src_datasets = [] + tgt_datasets = [] + + for k in itertools.count(): + split_k = split + (str(k) if k > 0 else "") + + # infer langcode + if self.split_exists(split_k, src, tgt, src, data_path, dataset_impl): + prefix = os.path.join(data_path, "{}.{}-{}.".format(split_k, src, tgt)) + elif self.split_exists(split_k, tgt, src, src, data_path, dataset_impl): + prefix = os.path.join(data_path, "{}.{}-{}.".format(split_k, tgt, src)) + else: + if k > 0: + break + else: + logger.error( + f"Dataset not found: {data_path}, {split_k}, {src}, {tgt}" + ) + raise FileNotFoundError( + "Dataset not found: {} ({})".format(split, data_path) + ) + + src_dataset = self.load_data(prefix + src, src_dict, dataset_impl) + if truncate_source: + src_dataset = AppendTokenDataset( + TruncateDataset( + StripTokenDataset(src_dataset, src_dict.eos()), + max_source_positions - 1, + ), + src_dict.eos(), + ) + src_datasets.append(src_dataset) + tgt_datasets.append(self.load_data(prefix + tgt, tgt_dict, dataset_impl)) + + logger.info( + "{} {} {}-{} {} examples".format( + data_path, split_k, src, tgt, len(src_datasets[-1]) + ) + ) + + if not combine: + break + + assert len(src_datasets) == len(tgt_datasets) + + if len(src_datasets) == 1: + src_dataset, tgt_dataset = src_datasets[0], tgt_datasets[0] + else: + sample_ratios = [1] * len(src_datasets) + sample_ratios[0] = upsample_primary + src_dataset = ConcatDataset(src_datasets, sample_ratios) + tgt_dataset = ConcatDataset(tgt_datasets, sample_ratios) + + if prepend_bos: + assert hasattr(src_dict, "bos_index") and hasattr(tgt_dict, "bos_index") + src_dataset = PrependTokenDataset(src_dataset, src_dict.bos()) + tgt_dataset = PrependTokenDataset(tgt_dataset, tgt_dict.bos()) + + align_dataset = None + if load_alignments: + align_path = os.path.join( + data_path, "{}.align.{}-{}".format(split, src, tgt) + ) + if indexed_dataset.dataset_exists(align_path, impl=dataset_impl): + align_dataset = data_utils.load_indexed_dataset( + align_path, None, dataset_impl + ) + + return src_dataset, tgt_dataset, align_dataset + + def load_langpair_dataset( + self, + data_path, + split, + src, + src_dict, + tgt, + tgt_dict, + combine, + dataset_impl, + upsample_primary, + left_pad_source, + left_pad_target, + max_source_positions, + max_target_positions, + prepend_bos=False, + load_alignments=False, + truncate_source=False, + src_dataset_transform_func=lambda dataset: dataset, + tgt_dataset_transform_func=lambda dataset: dataset, + src_lang_id=None, + tgt_lang_id=None, + langpairs_sharing_datasets=None, + ): + norm_direction = "-".join(sorted([src, tgt])) + if langpairs_sharing_datasets is not None: + src_dataset = langpairs_sharing_datasets.get( + (data_path, split, norm_direction, src), "NotInCache" + ) + tgt_dataset = langpairs_sharing_datasets.get( + (data_path, split, norm_direction, tgt), "NotInCache" + ) + align_dataset = langpairs_sharing_datasets.get( + (data_path, split, norm_direction, src, tgt), "NotInCache" + ) + + # a hack: any one is not in cache, we need to reload them + if ( + langpairs_sharing_datasets is None + or src_dataset == "NotInCache" + or tgt_dataset == "NotInCache" + or align_dataset == "NotInCache" + or split != getattr(self.args, "train_subset", None) + ): + # source and target datasets can be reused in reversed directions to save memory + # reversed directions of valid and test data will not share source and target datasets + src_dataset, tgt_dataset, align_dataset = self.load_lang_dataset( + data_path, + split, + src, + src_dict, + tgt, + tgt_dict, + combine, + dataset_impl, + upsample_primary, + max_source_positions=max_source_positions, + prepend_bos=prepend_bos, + load_alignments=load_alignments, + truncate_source=truncate_source, + ) + src_dataset = src_dataset_transform_func(src_dataset) + tgt_dataset = tgt_dataset_transform_func(tgt_dataset) + if langpairs_sharing_datasets is not None: + langpairs_sharing_datasets[ + (data_path, split, norm_direction, src) + ] = src_dataset + langpairs_sharing_datasets[ + (data_path, split, norm_direction, tgt) + ] = tgt_dataset + langpairs_sharing_datasets[ + (data_path, split, norm_direction, src, tgt) + ] = align_dataset + if align_dataset is None: + # no align data so flag the reverse direction as well in sharing + langpairs_sharing_datasets[ + (data_path, split, norm_direction, tgt, src) + ] = align_dataset + else: + logger.info( + f"Reusing source and target datasets of [{split}] {tgt}-{src} for reversed direction: " + f"[{split}] {src}-{tgt}: src length={len(src_dataset)}; tgt length={len(tgt_dataset)}" + ) + + return LanguagePairDataset( + src_dataset, + src_dataset.sizes, + src_dict, + tgt_dataset, + tgt_dataset.sizes if tgt_dataset is not None else None, + tgt_dict, + left_pad_source=left_pad_source, + left_pad_target=left_pad_target, + align_dataset=align_dataset, + src_lang_id=src_lang_id, + tgt_lang_id=tgt_lang_id, + ) + + def src_dataset_tranform_func(self, src_lang, tgt_lang, dataset, spec=None): + if self.args.lang_tok_replacing_bos_eos: + # it is handled by self.alter_dataset_langtok + # TODO: Unifiy with alter_dataset_langtok + return dataset + if spec is None: + return dataset + tok = self.get_encoder_langtok(src_lang, tgt_lang, spec) + if tok: + return PrependTokenDataset(dataset, tok) + return dataset + + def tgt_dataset_tranform_func(self, source_lang, target_lang, dataset, spec=None): + if dataset is None: + # note that target dataset can be None during inference time + return None + if self.args.lang_tok_replacing_bos_eos: + # TODO: Unifiy with alter_dataset_langtok + # It is handled by self.alter_dataset_langtok. + # The complication in self.alter_dataset_langtok + # makes a unified framework difficult. + return dataset + # if not self.args.decoder_langtok: + if not spec: + return dataset + tok = self.get_decoder_langtok(target_lang, spec) + if tok: + return PrependTokenDataset(dataset, tok) + return dataset + + def alter_dataset_langtok( + self, + lang_pair_dataset, + src_eos=None, + src_lang=None, + tgt_eos=None, + tgt_lang=None, + src_langtok_spec=None, + tgt_langtok_spec=None, + ): + if src_langtok_spec is None and tgt_langtok_spec is None: + return lang_pair_dataset + + new_src_eos = None + if ( + src_langtok_spec is not None + and src_eos is not None + and (src_lang is not None or tgt_lang is not None) + ): + new_src_eos = self.get_encoder_langtok(src_lang, tgt_lang, src_langtok_spec) + else: + src_eos = None + + new_tgt_bos = None + if tgt_langtok_spec and tgt_eos is not None and tgt_lang is not None: + new_tgt_bos = self.get_decoder_langtok(tgt_lang, tgt_langtok_spec) + else: + tgt_eos = None + + return TransformEosLangPairDataset( + lang_pair_dataset, + src_eos=src_eos, + new_src_eos=new_src_eos, + tgt_bos=tgt_eos, + new_tgt_bos=new_tgt_bos, + ) + + def load_a_dataset( + self, + split, + data_path, + src, + src_dict, + tgt, + tgt_dict, + combine, + prepend_bos=False, + langpairs_sharing_datasets=None, + data_category=None, + **extra_kwargs, + ): + dataset_impl = self.args.dataset_impl + upsample_primary = self.args.upsample_primary + left_pad_source = self.args.left_pad_source + left_pad_target = self.args.left_pad_target + max_source_positions = self.args.max_source_positions + max_target_positions = self.args.max_target_positions + load_alignments = self.args.load_alignments + truncate_source = self.args.truncate_source + src_dataset_transform_func = self.src_dataset_tranform_func + tgt_dataset_transform_func = self.tgt_dataset_tranform_func + enable_lang_ids = self.args.enable_lang_ids + lang_dictionary = self.lang_dict + src_langtok_spec, tgt_langtok_spec = extra_kwargs["langtok_spec"] + + src_langtok = self.get_encoder_langtok(src, tgt, src_langtok_spec) + tgt_langtok = self.get_decoder_langtok(tgt, tgt_langtok_spec) + logger.info( + f"{data_category}:{src}-{tgt} src_langtok: {src_langtok}; tgt_langtok: {tgt_langtok}" + ) + + langpair_ds = self.load_langpair_dataset( + data_path, + split, + src, + src_dict, + tgt, + tgt_dict, + combine, + dataset_impl, + upsample_primary, + left_pad_source, + left_pad_target, + max_source_positions, + max_target_positions, + prepend_bos, + load_alignments, + truncate_source, + src_dataset_transform_func=lambda dataset: src_dataset_transform_func( + src, tgt, dataset, src_langtok_spec + ), + tgt_dataset_transform_func=lambda dataset: tgt_dataset_transform_func( + src, tgt, dataset, tgt_langtok_spec + ), + src_lang_id=_lang_id(lang_dictionary, src) + if enable_lang_ids and lang_dictionary is not None + else None, + tgt_lang_id=_lang_id(lang_dictionary, tgt) + if enable_lang_ids and lang_dictionary is not None + else None, + langpairs_sharing_datasets=langpairs_sharing_datasets, + ) + # TODO: handle modified lang toks for mined data and dae data + if self.args.lang_tok_replacing_bos_eos: + ds = self.alter_dataset_langtok( + langpair_ds, + src_eos=self.get_source_dictionary(src).eos() if src else self.get_target_dictionary(tgt).eos(), + src_lang=src, + tgt_eos=self.get_target_dictionary(tgt).eos(), + tgt_lang=tgt, + src_langtok_spec=src_langtok_spec, + tgt_langtok_spec=tgt_langtok_spec, + ) + else: + ds = langpair_ds + return ds + + def load_split_langpair_datasets(self, split, data_param_list): + datasets = [] + langpairs_sharing_datasets = ( + {} if self.args.enable_reservsed_directions_shared_datasets else None + ) + for param in data_param_list: + ds = self.load_a_dataset( + split=split, + langpairs_sharing_datasets=langpairs_sharing_datasets, + **param, + ) + datasets.append(ds) + return datasets + + def get_data_paths_and_lang_pairs(self, split): + datapaths = {"main": self.args.data} + lang_pairs = {"main": self.lang_pairs} + if split == getattr(self.args, "train_subset", None): + # only training data can have extra data and extra language pairs + if self.args.extra_data: + extra_datapaths = self.args.extra_data + datapaths.update(extra_datapaths) + if self.args.extra_lang_pairs: + extra_lang_pairs = { + k: v.split(",") for k, v in self.args.extra_lang_pairs.items() + } + lang_pairs.update(extra_lang_pairs) + return datapaths, lang_pairs + + @classmethod + def get_dataset_key(cls, data_category, src, tgt): + return f"{data_category}:{src}-{tgt}" + + @classmethod + def _get_shard_num_dict(cls, split, paths): + shards = defaultdict(int) + for path in paths: + files = PathManager.ls(path) + directions = set() + for f in files: + if f.startswith(split) and f.endswith(".idx"): + # idx files of the form "{split}.{src}-{tgt}.{lang}.idx" + direction = f.split(".")[-3] + directions.add(direction) + for direction in directions: + shards[direction] += 1 + return shards + + def get_split_num_data_shards(self, split): + if split in self._num_shards_dict: + return self._num_shards_dict[split] + num_shards_dict = {} + data_paths, lang_pairs = self.get_data_paths_and_lang_pairs(split) + + for data_category, paths in data_paths.items(): + if data_category not in lang_pairs: + continue + paths = utils.split_paths(paths) + shards_dict = self._get_shard_num_dict(split, paths) + lang_dirs = [ + lang_pair.split("-") for lang_pair in lang_pairs[data_category] + ] + lang_dirs = [x if len(x) > 1 else (x[0], x[0]) for x in lang_dirs] + for src, tgt in lang_dirs: + key = self.get_dataset_key(data_category, src, tgt) + if "mono_" in data_category: + # monolingual data requires tgt only + assert src is None or src == tgt, ( + f"error: src={src}, " + "tgt={tgt} for data_category={data_category}" + ) + num_shards_dict[key] = shards_dict[tgt] + else: + if f"{src}-{tgt}" in shards_dict: + num_shards_dict[key] = shards_dict[f"{src}-{tgt}"] + elif f"{tgt}-{src}" in shards_dict: + # follow the fairseq tradition to use reversed direction data if it is not available + num_shards_dict[key] = shards_dict[f"{tgt}-{src}"] + self._num_shards_dict[split] = num_shards_dict + logger.info(f"[{split}] num of shards: {num_shards_dict}") + return num_shards_dict + + @classmethod + def get_shard_id(cls, num_shards, epoch, shard_epoch=None): + shard = epoch if shard_epoch is None else shard_epoch + shard = (shard - 1) % num_shards + return shard + + def get_split_data_path(self, paths, epoch, shard_epoch, num_shards): + path = paths[self.get_shard_id(num_shards, epoch, shard_epoch)] + return path + + def get_split_data_param_list(self, split, epoch, shard_epoch=None): + # TODO: to extend with extra datasets and keys and loop over different shard data paths + param_list = [] + data_paths, lang_pairs = self.get_data_paths_and_lang_pairs(split) + logger.info(f"langtoks settings: {self.args.langtoks}") + split_num_shards_dict = self.get_split_num_data_shards(split) + for data_category, paths in data_paths.items(): + if data_category not in lang_pairs: + continue + paths = utils.split_paths(paths) + assert len(paths) > 0 + if len(paths) > 1: + self._has_sharded_data = True + if split != getattr(self.args, "train_subset", None): + # if not training data set, use the first shard for valid and test + paths = paths[:1] + + if data_category in self.args.langtoks: + lang_tok_spec = self.args.langtoks[data_category] + else: + # default to None + lang_tok_spec = (None, None) + + # infer langcode + lang_dirs = [ + lang_pair.split("-") for lang_pair in lang_pairs[data_category] + ] + lang_dirs = [x if len(x) > 1 else (x[0], x[0]) for x in lang_dirs] + for src, tgt in lang_dirs: + assert src is not None or data_category == "mono_dae", ( + f"error: src={src}, " "tgt={tgt} for data_category={data_category}" + ) + # logger.info(f"preparing param for {data_category}: {src} - {tgt}") + key = self.get_dataset_key(data_category, src, tgt) + data_path = self.get_split_data_path( + paths, epoch, shard_epoch, split_num_shards_dict[key] + ) + param_list.append( + { + "key": key, + "data_path": data_path, + "split": split, + "src": src, + "src_dict": self.get_source_dictionary(src) + if src and data_category != "mono_dae" + else None, + "tgt": tgt, + "tgt_dict": self.get_target_dictionary(tgt), + "data_category": data_category, + "langtok_spec": lang_tok_spec, + } + ) + return param_list + + def get_train_dataset_sizes( + self, data_param_list, datasets, epoch, shard_epoch=None + ): + num_shards = [ + self.get_split_num_data_shards(param["split"])[param["key"]] + for param in data_param_list + ] + data_sizes = [] + for (key, d), num_shard in zip(datasets, num_shards): + my_data_sizes = self._training_data_sizes[key] + shard_ind = self.get_shard_id(num_shard, epoch, shard_epoch) + if shard_ind not in my_data_sizes: + my_data_sizes[shard_ind] = len(d) + known_size = max(my_data_sizes.values()) + data_sizes.append( + # If we don't know the data size of the shard yet, + # use the the max known data size to approximate. + # Note that we preprocess shards by a designated shard size + # and put any remaining data at the end into the last shard so + # the max shard size approximation is almost correct before loading + # the last shard; after loading the last shard, it will have the + # exact data sizes of the whole data size. + (key, sum(my_data_sizes.get(i, known_size) for i in range(num_shard))) + ) + logger.info( + f"estimated total data sizes of all shards used in sampling ratios: {data_sizes}. " + "Note that if the data a shard has not been loaded yet, use the max known data size to approximate" + ) + return [s for _, s in data_sizes] + + def get_train_sampling_ratios( + self, data_param_list, datasets, epoch=1, shard_epoch=None + ): + data_sizes = self.get_train_dataset_sizes( + data_param_list, datasets, epoch, shard_epoch + ) + sampling_func = self.sampling_method.sampling_method_selector() + sample_ratios = sampling_func(data_sizes) if sampling_func is not None else None + return sample_ratios + + def get_sampling_ratios(self, data_param_list, datasets, epoch, shard_epoch=None): + if self.args.sampling_weights_from_file: + weights = load_sampling_weights(self.args.sampling_weights_from_file) + sample_ratios = [weights[k] for k, _ in datasets] + logger.info( + "| ignoring --sampling-weights when loadding sampling weights " + f"from file {self.args.sampling_weights_from_file}" + ) + elif self.args.sampling_weights: + sample_ratios = [self.args.sampling_weights[k] for k, _ in datasets] + else: + sample_ratios = self.get_train_sampling_ratios( + data_param_list, datasets, epoch, shard_epoch + ) + + if sample_ratios is not None: + logger.info( + "| Upsample ratios: {}".format( + list(zip(map(lambda x: x["key"], data_param_list), sample_ratios)) + ) + ) + assert len(sample_ratios) == len(datasets) + return sample_ratios + + def load_split_datasets( + self, split, training, epoch=1, combine=False, shard_epoch=None, **kwargs + ): + data_param_list = self.get_split_data_param_list( + split, epoch, shard_epoch=shard_epoch + ) + langpairs_sharing_datasets = ( + {} if self.args.enable_reservsed_directions_shared_datasets else None + ) + datasets = [ + ( + param["key"], + self.load_a_dataset( + combine=combine, + langpairs_sharing_datasets=langpairs_sharing_datasets, + **param, + ), + ) + for param in data_param_list + ] + return datasets, data_param_list + + def load_into_concat_dataset(self, split, datasets, data_param_list): + if self.args.lang_tok_replacing_bos_eos: + # TODO: to investigate why TransformEosLangPairDataset doesn't work with ConcatDataset + return SampledMultiDataset( + OrderedDict(datasets), + sampling_ratios=None, + eval_key=None, + collate_format=CollateFormat.single, + virtual_size=None, + split=split, + ) + return ConcatDataset([d for _, d in datasets]) + + def load_sampled_multi_epoch_dataset( + self, split, training, epoch=0, combine=False, shard_epoch=None, **kwargs + ): + datasets, data_param_list = self.load_split_datasets( + split, training, epoch, combine, shard_epoch=shard_epoch, **kwargs + ) + if training and split == getattr(self.args, "train_subset", None): + sample_ratios = self.get_sampling_ratios(data_param_list, datasets, epoch) + return SampledMultiEpochDataset( + OrderedDict(datasets), + epoch=epoch, + shard_epoch=shard_epoch, + # valid and test datasets will be degenerate to concating datasets: + sampling_ratios=sample_ratios, + eval_key=None, + collate_format=CollateFormat.single, + virtual_size=self.args.virtual_data_size, + split=split, + virtual_epoch_size=self.args.virtual_epoch_size, + # if not using lang_tok altering, simplified to use the same collater + shared_collater=self._shared_collater(), + ) + else: + return self.load_into_concat_dataset(split, datasets, data_param_list) + + def load_sampled_multi_dataset( + self, split, training, epoch=0, combine=False, shard_epoch=None, **kwargs + ): + datasets, data_param_list = self.load_split_datasets( + split, training, epoch, combine, shard_epoch=shard_epoch, **kwargs + ) + if training and split == getattr(self.args, "train_subset", None): + sample_ratios = self.get_sampling_ratios(data_param_list, datasets, epoch) + return SampledMultiDataset( + OrderedDict(datasets), + epoch=epoch, + # valid and test datasets will be degerate to concating datasets: + sampling_ratios=sample_ratios, + eval_key=None, + collate_format=CollateFormat.single, + virtual_size=self.args.virtual_data_size, + split=split, + # if not using lang_tok altering, simplified to use the same collater + shared_collater=self._shared_collater(), + ) + else: + return self.load_into_concat_dataset(split, datasets, data_param_list) + + def load_dataset( + self, split, training, epoch=0, combine=False, shard_epoch=None, **kwargs + ): + if self.args.virtual_epoch_size is None: + return self.load_sampled_multi_dataset( + split, training, epoch, combine, shard_epoch, **kwargs + ) + else: + return self.load_sampled_multi_epoch_dataset( + split, training, epoch, combine, shard_epoch, **kwargs + ) diff --git a/fairseq/data/multilingual/multilingual_utils.py b/fairseq/data/multilingual/multilingual_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..b4e0f9828cabfdbe375d05d9152b58bdbd6de7dc --- /dev/null +++ b/fairseq/data/multilingual/multilingual_utils.py @@ -0,0 +1,63 @@ +from enum import Enum +from typing import Dict, List, Optional, Sequence + +import torch +from fairseq.data import Dictionary + + +class EncoderLangtok(Enum): + """ + Prepend to the beginning of source sentence either the + source or target language token. (src/tgt). + """ + + src = "src" + tgt = "tgt" + + +class LangTokSpec(Enum): + main = "main" + mono_dae = "mono_dae" + + +class LangTokStyle(Enum): + multilingual = "multilingual" + mbart = "mbart" + + +@torch.jit.export +def get_lang_tok( + lang: str, lang_tok_style: str, spec: str = LangTokSpec.main.value +) -> str: + # TOKEN_STYLES can't be defined outside this fn since it needs to be + # TorchScriptable. + TOKEN_STYLES: Dict[str, str] = { + LangTokStyle.mbart.value: "[{}]", + LangTokStyle.multilingual.value: "__{}__", + } + + if spec.endswith("dae"): + lang = f"{lang}_dae" + elif spec.endswith("mined"): + lang = f"{lang}_mined" + style = TOKEN_STYLES[lang_tok_style] + return style.format(lang) + + +def augment_dictionary( + dictionary: Dictionary, + language_list: List[str], + lang_tok_style: str, + langtoks_specs: Sequence[str] = (LangTokSpec.main.value,), + extra_data: Optional[Dict[str, str]] = None, +) -> None: + for spec in langtoks_specs: + for language in language_list: + dictionary.add_symbol( + get_lang_tok(lang=language, lang_tok_style=lang_tok_style, spec=spec) + ) + + if lang_tok_style == LangTokStyle.mbart.value or ( + extra_data is not None and LangTokSpec.mono_dae.value in extra_data + ): + dictionary.add_symbol("<mask>") diff --git a/fairseq/data/multilingual/sampled_multi_dataset.py b/fairseq/data/multilingual/sampled_multi_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..b0a617424ee3c5923b37796773da4c97851a16c5 --- /dev/null +++ b/fairseq/data/multilingual/sampled_multi_dataset.py @@ -0,0 +1,467 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import datetime +import hashlib +import logging +import time +from bisect import bisect_right +from collections import OrderedDict, defaultdict +from enum import Enum +from typing import List + +import numpy as np +import torch +from fairseq.data import FairseqDataset, data_utils +from fairseq.distributed import utils as distributed_utils + + +def get_time_gap(s, e): + return ( + datetime.datetime.fromtimestamp(e) - datetime.datetime.fromtimestamp(s) + ).__str__() + + +logger = logging.getLogger(__name__) + + +def default_virtual_size_func(datasets, ratios, max_scale_up=1.5): + sizes = [len(d) for d in datasets] + if ratios is None: + return sum(sizes) + largest_idx = np.argmax(sizes) + largest_r = ratios[largest_idx] + largest_s = sizes[largest_idx] + # set virtual sizes relative to the largest dataset + virtual_sizes = [(r / largest_r) * largest_s for r in ratios] + vsize = sum(virtual_sizes) + max_size = sum(sizes) * max_scale_up + return int(vsize if vsize < max_size else max_size) + + +class CollateFormat(Enum): + single = 1 + ordered_dict = 2 + + +class SampledMultiDataset(FairseqDataset): + """Samples from multiple sub-datasets according to given sampling ratios. + Args: + datasets ( + List[~torch.utils.data.Dataset] + or OrderedDict[str, ~torch.utils.data.Dataset] + ): datasets + sampling_ratios (List[float]): list of probability of each dataset to be sampled + (default: None, which corresponds to concatenating all dataset together). + seed (int): RNG seed to use (default: 2). + epoch (int): starting epoch number (default: 1). + eval_key (str, optional): a key used at evaluation time that causes + this instance to pass-through batches from *datasets[eval_key]*. + collate_format (CollateFormat): collater output format, either CollateFormat.ordered_dict or + CollateFormat.single (default: CollateFormat.single) where CollateFormat.single configures + the collater to output batches of data mixed from all sub-datasets, + and CollateFormat.ordered_dict configures the collater to output a dictionary of batches indexed by keys + of sub-datasets. + Note that not all sub-datasets will present in a single batch in both formats. + virtual_size (int, or callable): the expected virtual size of the dataset (default: default_virtual_size_func). + split (str): the split of the data, e.g. 'train', 'valid' or 'test'. + shared_collater (bool): whether or not to all sub-datasets have the same collater. + shuffle (bool): whether or not to shuffle data (default: True). + """ + + def __init__( + self, + datasets, + sampling_ratios=None, + seed=2, + epoch=1, + eval_key=None, + collate_format=CollateFormat.single, + virtual_size=default_virtual_size_func, + split="", + shared_collater=False, + shuffle=True, + ): + super().__init__() + self.shared_collater = shared_collater + self.shuffle = shuffle + + if isinstance(datasets, OrderedDict): + self.keys = list(datasets.keys()) + datasets = list(datasets.values()) + elif isinstance(datasets, List): + self.keys = list(range(len(datasets))) + else: + raise AssertionError() + self.datasets = datasets + self.split = split + + self.eval_key = eval_key + if self.eval_key is not None: + self.collate_format = CollateFormat.single + else: + self.collate_format = collate_format + + self.seed = seed + self._cur_epoch = None + + self.cumulated_sizes = None + # self.datasets[k][self._cur_indices[i]] is the data item i in this sampled dataset + # namely, data item i is sampled from the kth sub-dataset self.datasets[k] + # where self.cumulated_sizes[k-1] <= i < self.cumulated_sizes[k] + self._cur_indices = None + + self._sizes = None + self.virtual_size_per_dataset = None + # caching properties + self._reset_cached_properties() + self.setup_sampling(sampling_ratios, virtual_size) + self.set_epoch(epoch) + + def _clean_if_not_none(self, var_list): + for v in var_list: + if v is not None: + del v + + def _reset_cached_properties(self): + self._clean_if_not_none([self._sizes, self._cur_indices]) + self._sizes = None + self._cur_indices = None + + def setup_sampling(self, sample_ratios, virtual_size): + sizes = [len(d) for d in self.datasets] + if sample_ratios is None: + # default back to concating datasets + self.sample_ratios = None + self.virtual_size = sum(sizes) + else: + if not isinstance(sample_ratios, np.ndarray): + sample_ratios = np.array(sample_ratios) + self.sample_ratios = sample_ratios + virtual_size = ( + default_virtual_size_func if virtual_size is None else virtual_size + ) + self.virtual_size = ( + virtual_size(self.datasets, self.sample_ratios) + if callable(virtual_size) + else virtual_size + ) + + def adjust_sampling(self, epoch, sampling_ratios, virtual_size): + if sampling_ratios is not None: + sampling_ratios = self._sync_sample_ratios(sampling_ratios) + self.setup_sampling(sampling_ratios, virtual_size) + + def _sync_sample_ratios(self, ratios): + # in case the ratios are not precisely the same across processes + # also to ensure every procresses update the ratios in the same pace + ratios = torch.DoubleTensor(ratios) + if torch.distributed.is_initialized(): + if torch.cuda.is_available(): + distributed_utils.all_reduce( + ratios.cuda(), group=distributed_utils.get_data_parallel_group() + ) + else: + distributed_utils.all_reduce( + ratios, group=distributed_utils.get_data_parallel_group() + ) + ret = ratios.cpu() + ret = ret.numpy() + return ret + + def random_choice_in_dataset(self, rng, dataset, choice_size): + if hasattr(dataset, "random_choice_in_dataset"): + return dataset.random_choice_in_dataset(rng, choice_size) + dataset_size = len(dataset) + return rng.choice( + dataset_size, choice_size, replace=(choice_size > dataset_size) + ) + + def get_virtual_indices(self, rng, datasets, sample_ratios, virtual_size): + def get_counts(sample_ratios): + counts = np.array([virtual_size * r for r in sample_ratios], dtype=np.int64) + diff = virtual_size - counts.sum() + assert diff >= 0 + # due to round-offs, the size might not match the desired sizes + if diff > 0: + dataset_indices = rng.choice( + len(sample_ratios), size=diff, p=sample_ratios + ) + for i in dataset_indices: + counts[i] += 1 + return counts + + def get_in_dataset_indices(datasets, sizes, sample_ratios): + counts = get_counts(sample_ratios) + # uniformally sample desired counts for each dataset + # if the desired counts are large, sample with replacement: + indices = [ + self.random_choice_in_dataset(rng, d, c) + for c, d in zip(counts, datasets) + ] + return indices + + sizes = [len(d) for d in datasets] + if sample_ratios is None: + # default back to concating datasets + in_dataset_indices = [list(range(s)) for s in sizes] + virtual_sizes_per_dataset = sizes + else: + ratios = sample_ratios / sample_ratios.sum() + in_dataset_indices = get_in_dataset_indices(datasets, sizes, ratios) + virtual_sizes_per_dataset = [len(d) for d in in_dataset_indices] + virtual_sizes_per_dataset = np.array(virtual_sizes_per_dataset, np.int64) + cumulative_sizes = np.cumsum(virtual_sizes_per_dataset) + assert sum(virtual_sizes_per_dataset) == virtual_size + assert cumulative_sizes[-1] == virtual_size + if virtual_size < sum(sizes): + logger.warning( + f"virtual data size ({virtual_size}) is less than real data size ({sum(sizes)})." + " If virtual size << real data size, there could be data coverage issue." + ) + in_dataset_indices = np.hstack(in_dataset_indices) + return in_dataset_indices, cumulative_sizes, virtual_sizes_per_dataset + + def _get_dataset_and_index(self, index): + i = bisect_right(self.cumulated_sizes, index) + return i, self._cur_indices[index] + + def __getitem__(self, index): + # self.__getitem__(index) returns self.datasets[k][self._cur_indices[index]] + # where k satisfies self.cumulated_sizes[k - 1] <= k < self.cumulated_sizes[k] + ds_idx, ds_sample_idx = self._get_dataset_and_index(index) + ret = (ds_idx, self.datasets[ds_idx][ds_sample_idx]) + return ret + + def num_tokens(self, index): + return self.sizes[index].max() + + def num_tokens_vec(self, indices): + sizes_vec = self.sizes[np.array(indices)] + # max across all dimensions but first one + return np.amax(sizes_vec, axis=tuple(range(1, len(sizes_vec.shape)))) + + def size(self, index): + return self.sizes[index] + + def __len__(self): + return self.virtual_size + + def collater(self, samples, **extra_args): + """Merge a list of samples to form a mini-batch.""" + if len(samples) == 0: + return None + if self.collate_format == "ordered_dict": + collect_samples = [[] for _ in range(len(self.datasets))] + for (i, sample) in samples: + collect_samples[i].append(sample) + batch = OrderedDict( + [ + (self.keys[i], dataset.collater(collect_samples[i])) + for i, (key, dataset) in enumerate(zip(self.keys, self.datasets)) + if len(collect_samples[i]) > 0 + ] + ) + elif self.shared_collater: + batch = self.datasets[0].collater([s for _, s in samples]) + else: + samples_dict = defaultdict(list) + pad_to_length = ( + defaultdict(int) + if "pad_to_length" not in extra_args + else extra_args["pad_to_length"] + ) + for ds_idx, s in samples: + pad_to_length["source"] = max( + pad_to_length["source"], s["source"].size(0) + ) + if s["target"] is not None: + pad_to_length["target"] = max( + pad_to_length["target"], s["target"].size(0) + ) + samples_dict[ds_idx].append(s) + batches = [ + self.datasets[i].collater(samples_dict[i], pad_to_length=pad_to_length) + for i in range(len(self.datasets)) + if len(samples_dict[i]) > 0 + ] + + def straight_data(tensors): + batch = torch.cat(tensors, dim=0) + return batch + + src_lengths = straight_data( + [b["net_input"]["src_lengths"] for b in batches] + ) + src_lengths, sort_order = src_lengths.sort(descending=True) + + def straight_order(tensors): + batch = straight_data(tensors) + return batch.index_select(0, sort_order) + + batch = { + "id": straight_order([b["id"] for b in batches]), + "nsentences": sum(b["nsentences"] for b in batches), + "ntokens": sum(b["ntokens"] for b in batches), + "net_input": { + "src_tokens": straight_order( + [b["net_input"]["src_tokens"] for b in batches] + ), + "src_lengths": src_lengths, + }, + "target": straight_order([b["target"] for b in batches]) + if batches[0]["target"] is not None + else None, + } + if "prev_output_tokens" in batches[0]["net_input"]: + batch["net_input"]["prev_output_tokens"] = straight_order( + [b["net_input"]["prev_output_tokens"] for b in batches] + ) + if "src_lang_id" in batches[0]["net_input"]: + batch["net_input"]["src_lang_id"] = straight_order( + [b["net_input"]["src_lang_id"] for b in batches] + ) + if "tgt_lang_id" in batches[0]: + batch["tgt_lang_id"] = straight_order( + [b["tgt_lang_id"] for b in batches] + ) + return batch + + @property + def sizes(self): + if self._sizes is not None: + return self._sizes + start_time = time.time() + in_sub_dataset_indices = [ + self._cur_indices[ + 0 if i == 0 else self.cumulated_sizes[i - 1] : self.cumulated_sizes[i] + ] + for i in range(len(self.datasets)) + ] + sub_dataset_sizes = [ + d.sizes[indices] + for d, indices in zip(self.datasets, in_sub_dataset_indices) + ] + self._sizes = np.vstack(sub_dataset_sizes) + logger.info(f"sizes() calling time: {get_time_gap(start_time, time.time())}") + return self._sizes + + def ordered_indices(self): + if self.shuffle: + indices = np.random.permutation(len(self)) + else: + indices = np.arange(len(self)) + + sizes = self.sizes + tgt_sizes = sizes[:, 1] if len(sizes.shape) > 0 and sizes.shape[1] > 1 else None + src_sizes = ( + sizes[:, 0] if len(sizes.shape) > 0 and sizes.shape[1] > 1 else sizes + ) + + # sort by target length, then source length + if tgt_sizes is not None: + indices = indices[np.argsort(tgt_sizes[indices], kind="mergesort")] + sort_indices = indices[np.argsort(src_sizes[indices], kind="mergesort")] + return sort_indices + + def prefetch(self, indices): + prefetch_indices = [[] for _ in range(len(self.datasets))] + for i in indices: + ds_idx, ds_sample_idx = self._get_dataset_and_index(i) + prefetch_indices[ds_idx].append(ds_sample_idx) + for i in range(len(prefetch_indices)): + self.datasets[i].prefetch(prefetch_indices[i]) + + @property + def can_reuse_epoch_itr_across_epochs(self): + return False + + def set_epoch(self, epoch): + super().set_epoch(epoch) + if epoch == self._cur_epoch: + # re-enter so return + return + for d in self.datasets: + if hasattr(d, "set_epoch"): + d.set_epoch(epoch) + self._cur_epoch = epoch + self._establish_virtual_datasets() + + def _establish_virtual_datasets(self): + if self.sample_ratios is None and self._cur_indices is not None: + # not a samping dataset, no need to resample if indices are already established + return + self._reset_cached_properties() + + start_time = time.time() + # Generate a weighted sample of indices as a function of the + # random seed and the current epoch. + rng = np.random.RandomState( + [ + int( + hashlib.sha1( + str(self.__class__.__name__).encode("utf-8") + ).hexdigest(), + 16, + ) + % (2 ** 32), + self.seed % (2 ** 32), # global seed + self._cur_epoch, # epoch index, + ] + ) + self._clean_if_not_none( + [self.cumulated_sizes, self.virtual_size_per_dataset, self._sizes] + ) + self._sizes = None + + indices, cumulated_sizes, virtual_size_per_dataset = self.get_virtual_indices( + rng, self.datasets, self.sample_ratios, self.virtual_size + ) + self._cur_indices = indices + self.cumulated_sizes = cumulated_sizes + self.virtual_size_per_dataset = virtual_size_per_dataset + + raw_sizes = [len(d) for d in self.datasets] + sampled_sizes = self.virtual_size_per_dataset + logger.info( + f"[{self.split}] Raw sizes: {str(dict(zip(self.keys, raw_sizes)))}; " + f"raw total size: {sum(raw_sizes)}" + ) + logger.info( + f"[{self.split}] Resampled sizes: {str(dict(zip(self.keys, sampled_sizes)))}; " + f"resampled total size: {sum(sampled_sizes)}" + ) + if self.sample_ratios is not None: + logger.info( + f"[{self.split}] Upsampling ratios: {str(dict(zip(self.keys, self.sample_ratios)))}" + ) + else: + logger.info(f"[{self.split}] A concat dataset") + logger.info( + f"[{self.split}] virtual dataset established time: {get_time_gap(start_time, time.time())}" + ) + + def filter_indices_by_size(self, indices, max_sizes): + """Filter a list of sample indices. Remove those that are longer + than specified in max_sizes. + + Args: + indices (np.array): original array of sample indices + max_sizes (int or list[int] or tuple[int]): max sample size, + can be defined separately for src and tgt (then list or tuple) + + Returns: + np.array: filtered sample array + list: list of removed indices + """ + sizes = self.sizes + tgt_sizes = sizes[:, 1] if len(sizes.shape) > 0 and sizes.shape[1] > 1 else None + src_sizes = ( + sizes[:, 0] if len(sizes.shape) > 0 and sizes.shape[1] > 1 else sizes + ) + + return data_utils.filter_paired_dataset_indices_by_size( + src_sizes, tgt_sizes, indices, max_sizes + ) diff --git a/fairseq/data/multilingual/sampled_multi_epoch_dataset.py b/fairseq/data/multilingual/sampled_multi_epoch_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..17387b2f85c0ee76db1a003091331b46de8d8def --- /dev/null +++ b/fairseq/data/multilingual/sampled_multi_epoch_dataset.py @@ -0,0 +1,199 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import hashlib +import logging +import math + +import numpy as np +from fairseq.data import SampledMultiDataset + +from .sampled_multi_dataset import CollateFormat, default_virtual_size_func + + +logger = logging.getLogger(__name__) + + +class SampledMultiEpochDataset(SampledMultiDataset): + """Samples from multiple sub-datasets according to sampling ratios + using virtual epoch sizes to speed up dataloading. + Args: + datasets ( + List[~torch.utils.data.Dataset] + or OrderedDict[str, ~torch.utils.data.Dataset] + ): datasets + sampling_ratios (List[float]): list of probability of each dataset to be sampled + (default: None, which corresponds to concating all dataset together). + seed (int): RNG seed to use (default: 2). + epoch (int): starting epoch number (default: 1). + eval_key (str, optional): a key used at evaluation time that causes + this instance to pass-through batches from *datasets[eval_key]*. + collate_format (CollateFormat): collater output format, either CollateFormat.ordered_dict or + CollateFormat.single (default: CollateFormat.single) where CollateFormat.single configures + the collater to output batches of data mixed from all sub-datasets, + and CollateFormat.ordered_dict configures the collater to output a dictionary of batches indexed by keys + of sub-datasets. + Note that not all sub-datasets will present in a single batch in both formats. + virtual_size (int, or callable): the expected virtual size of the dataset (default: default_virtual_size_func). + split (str): the split of the data, e.g. 'train', 'valid' or 'test'. + virtual_epoch_size (int): virtual epoch size, the dataset will go through the data by + this virtual epoch size one by one to speed up data loading, e.g. indicing and filtering + can be performed whenever a virtual epoch is loaded without waiting for the whole dataset to be loaded. + shared_collater (bool): whether or not to all sub-datasets have the same collater. + shard_epoch (int): the real epoch number for shard selection. + shuffle (bool): whether or not to shuffle data (default: True). + """ + + def __init__( + self, + datasets, + sampling_ratios=None, + seed=2, + epoch=1, + eval_key=None, + collate_format=CollateFormat.single, + virtual_size=default_virtual_size_func, + split="", + virtual_epoch_size=None, + shared_collater=False, + shard_epoch=1, + shuffle=True, + ): + self.virtual_epoch_size = virtual_epoch_size + self._current_epoch_start_index = None + self._random_global_indices = None + self.shard_epoch = shard_epoch if shard_epoch is not None else 1 + self.load_next_shard = None + self._epoch_sizes = None + super().__init__( + datasets=datasets, + sampling_ratios=sampling_ratios, + seed=seed, + epoch=epoch, + eval_key=eval_key, + collate_format=collate_format, + virtual_size=virtual_size, + split=split, + shared_collater=shared_collater, + shuffle=shuffle, + ) + + def _setup(self, epoch): + self.virtual_epoch_size = ( + self.virtual_epoch_size + if self.virtual_epoch_size is not None + else self.virtual_size + ) + if self.virtual_epoch_size > self.virtual_size: + logger.warning( + f"virtual epoch size {self.virtual_epoch_size} " + f"is greater than virtual dataset size {self.virtual_size}" + ) + self.virtual_epoch_size = self.virtual_size + self.num_virtual_epochs = math.ceil(self.virtual_size / self.virtual_epoch_size) + self._current_epoch_start_index = self._get_epoch_start_index(epoch) + logger.info( + f"virtual epoch size {self.virtual_epoch_size}; virtual dataset size {self.virtual_size}" + ) + + def _map_epoch_index_to_global(self, index): + index = self._current_epoch_start_index + index + # add randomness + return self._random_global_indices[index] + + @property + def sizes(self): + if self._epoch_sizes is not None: + return self._epoch_sizes + _sizes = super().sizes + indices = self._random_global_indices[ + self._current_epoch_start_index : self._current_epoch_start_index + + len(self) + ] + self._epoch_sizes = _sizes[indices] + # del super()._sizes to save memory + del self._sizes + self._sizes = None + return self._epoch_sizes + + def _get_dataset_and_index(self, index): + i = self._map_epoch_index_to_global(index) + return super()._get_dataset_and_index(i) + + def __len__(self): + return ( + self.virtual_epoch_size + if self._current_epoch_start_index + self.virtual_epoch_size + < self.virtual_size + else self.virtual_size - self._current_epoch_start_index + ) + + def set_epoch(self, epoch): + if self._current_epoch_start_index is None: + # initializing epoch idnices of a virtual dataset + self._setup(epoch) + self._next_virtual_epoch(epoch) + else: + # working on already intialized epoch indices + if epoch == self._cur_epoch: + # re-enter so return + return + self._next_virtual_epoch(epoch) + + def _get_epoch_start_index(self, epoch): + assert epoch >= 1 # fairseq is using 1-based epoch everywhere + return ((epoch - 1) % self.num_virtual_epochs) * self.virtual_epoch_size + + def _next_global_indices(self, epoch): + rng = np.random.RandomState( + [ + int( + hashlib.sha1( + str(self.__class__.__name__).encode("utf-8") + ).hexdigest(), + 16, + ) + % (2 ** 32), + self.seed % (2 ** 32), # global seed + epoch, # epoch index, + ] + ) + del self._random_global_indices + self._random_global_indices = rng.choice( + self.virtual_size, self.virtual_size, replace=False + ) + if self.load_next_shard is None: + self.load_next_shard = False + else: + # increase shard epoch for next loading + self.shard_epoch += 1 + self.load_next_shard = True + logger.info( + "to load next epoch/shard in next load_dataset: " + f"epoch={epoch}/shard_epoch={self.shard_epoch}" + ) + + def _next_virtual_epoch(self, epoch): + index = self._get_epoch_start_index(epoch) + if index == 0 or self._random_global_indices is None: + # need to start from the beginning, + # so call super().set_epoch(epoch) to establish the global virtual indices + logger.info( + "establishing a new set of global virtual indices for " + f"epoch={epoch}/shard_epoch={self.shard_epoch}" + ) + super().set_epoch(epoch) + self._next_global_indices(epoch) + else: + self._cur_epoch = epoch + + # reset cache sizes and ordered_indices for the epoch after moving to a new epoch + self._clean_if_not_none( + [ + self._epoch_sizes, + ] + ) + self._epoch_sizes = None + self._current_epoch_start_index = index diff --git a/fairseq/data/multilingual/sampling_method.py b/fairseq/data/multilingual/sampling_method.py new file mode 100644 index 0000000000000000000000000000000000000000..140c68f01d60e902ef88f11f30f8813dc15fc681 --- /dev/null +++ b/fairseq/data/multilingual/sampling_method.py @@ -0,0 +1,78 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +from typing import List + + +logger = logging.getLogger(__name__) + + +def uniform(dataset_sizes: List[int]): + return [1.0] * len(dataset_sizes) + + +def temperature_sampling(dataset_sizes, temp): + total_size = sum(dataset_sizes) + return [(size / total_size) ** (1.0 / temp) for size in dataset_sizes] + + +def make_temperature_sampling(temp=1.0): + def sampling_func(dataset_sizes): + return temperature_sampling(dataset_sizes, temp) + + return sampling_func + + +def make_ratio_sampling(ratios): + def sampling_func(dataset_sizes): + return ratios + + return sampling_func + + +class SamplingMethod: + @staticmethod + def add_arguments(parser): + parser.add_argument( + "--sampling-method", + choices=[ + "uniform", + "temperature", + "concat", + "RoundRobin", + ], + type=str, + default="concat", + help="The method to sample data per language pairs", + ) + parser.add_argument( + "--sampling-temperature", + default=1.5, + type=float, + help="only work with --sampling-method temperature", + ) + + @staticmethod + def build_sampler(args, task): + return SamplingMethod(args, task) + + def __init__(self, args, task): + self.args = args + self.task = task + + def is_adaptive(self): + return False + + def sampling_method_selector(self): + args = self.args + logger.info(f"selected sampler: {args.sampling_method}") + if args.sampling_method == "uniform": + return uniform + elif args.sampling_method == "temperature" or self.is_adaptive(): + return make_temperature_sampling(float(args.sampling_temperature)) + else: + # default to concating all data set together + return None diff --git a/fairseq/data/nested_dictionary_dataset.py b/fairseq/data/nested_dictionary_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..52e74abddacc923c5e29b0a0c41d7efc85482d3b --- /dev/null +++ b/fairseq/data/nested_dictionary_dataset.py @@ -0,0 +1,125 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from collections import OrderedDict + +import torch +from torch.utils.data.dataloader import default_collate + +from . import FairseqDataset + + +def _flatten(dico, prefix=None): + """Flatten a nested dictionary.""" + new_dico = OrderedDict() + if isinstance(dico, dict): + prefix = prefix + "." if prefix is not None else "" + for k, v in dico.items(): + if v is None: + continue + new_dico.update(_flatten(v, prefix + k)) + elif isinstance(dico, list): + for i, v in enumerate(dico): + new_dico.update(_flatten(v, prefix + ".[" + str(i) + "]")) + else: + new_dico = OrderedDict({prefix: dico}) + return new_dico + + +def _unflatten(dico): + """Unflatten a flattened dictionary into a nested dictionary.""" + new_dico = OrderedDict() + for full_k, v in dico.items(): + full_k = full_k.split(".") + node = new_dico + for k in full_k[:-1]: + if k.startswith("[") and k.endswith("]"): + k = int(k[1:-1]) + if k not in node: + node[k] = OrderedDict() + node = node[k] + node[full_k[-1]] = v + return new_dico + + +class NestedDictionaryDataset(FairseqDataset): + def __init__(self, defn, sizes=None): + super().__init__() + self.defn = _flatten(defn) + self.sizes = [sizes] if not isinstance(sizes, (list, tuple)) else sizes + + first = None + for v in self.defn.values(): + if not isinstance( + v, + ( + FairseqDataset, + torch.utils.data.Dataset, + ), + ): + raise ValueError("Expected Dataset but found: {}".format(v.__class__)) + first = first or v + if len(v) > 0: + assert len(v) == len(first), "dataset lengths must match" + + self._len = len(first) + + def __getitem__(self, index): + return OrderedDict((k, ds[index]) for k, ds in self.defn.items()) + + def __len__(self): + return self._len + + def collater(self, samples): + """Merge a list of samples to form a mini-batch. + + Args: + samples (List[dict]): samples to collate + + Returns: + dict: a mini-batch suitable for forwarding with a Model + """ + if len(samples) == 0: + return {} + sample = OrderedDict() + for k, ds in self.defn.items(): + try: + sample[k] = ds.collater([s[k] for s in samples]) + except NotImplementedError: + sample[k] = default_collate([s[k] for s in samples]) + return _unflatten(sample) + + def num_tokens(self, index): + """Return the number of tokens in a sample. This value is used to + enforce ``--max-tokens`` during batching.""" + return max(s[index] for s in self.sizes) + + def size(self, index): + """Return an example's size as a float or tuple. This value is used when + filtering a dataset with ``--max-positions``.""" + if len(self.sizes) == 1: + return self.sizes[0][index] + else: + return (s[index] for s in self.sizes) + + @property + def supports_prefetch(self): + """Whether this dataset supports prefetching.""" + return any(ds.supports_prefetch for ds in self.defn.values()) + + def prefetch(self, indices): + """Prefetch the data required for this epoch.""" + for ds in self.defn.values(): + if getattr(ds, "supports_prefetch", False): + ds.prefetch(indices) + + @property + def can_reuse_epoch_itr_across_epochs(self): + return all(ds.can_reuse_epoch_itr_across_epochs for ds in self.defn.values()) + + def set_epoch(self, epoch): + super().set_epoch(epoch) + for ds in self.defn.values(): + ds.set_epoch(epoch) diff --git a/fairseq/data/noising.py b/fairseq/data/noising.py new file mode 100644 index 0000000000000000000000000000000000000000..2b1cc347203bfbdc9f1cba29e2e36427b7b5be57 --- /dev/null +++ b/fairseq/data/noising.py @@ -0,0 +1,335 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import numpy as np +import torch +from fairseq.data import data_utils + + +class WordNoising(object): + """Generate a noisy version of a sentence, without changing words themselves.""" + + def __init__(self, dictionary, bpe_cont_marker="@@", bpe_end_marker=None): + self.dictionary = dictionary + self.bpe_end = None + if bpe_cont_marker: + self.bpe_end = np.array( + [ + not self.dictionary[i].endswith(bpe_cont_marker) + for i in range(len(self.dictionary)) + ] + ) + elif bpe_end_marker: + self.bpe_end = np.array( + [ + self.dictionary[i].endswith(bpe_end_marker) + for i in range(len(self.dictionary)) + ] + ) + + self.get_word_idx = ( + self._get_bpe_word_idx if self.bpe_end is not None else self._get_token_idx + ) + + def noising(self, x, lengths, noising_prob=0.0): + raise NotImplementedError() + + def _get_bpe_word_idx(self, x): + """ + Given a list of BPE tokens, for every index in the tokens list, + return the index of the word grouping that it belongs to. + For example, for input x corresponding to ["how", "are", "y@@", "ou"], + return [[0], [1], [2], [2]]. + """ + # x: (T x B) + bpe_end = self.bpe_end[x] + + if x.size(0) == 1 and x.size(1) == 1: + # Special case when we only have one word in x. If x = [[N]], + # bpe_end is a scalar (bool) instead of a 2-dim array of bools, + # which makes the sum operation below fail. + return np.array([[0]]) + + # do a reduce front sum to generate word ids + word_idx = bpe_end[::-1].cumsum(0)[::-1] + word_idx = word_idx.max(0)[None, :] - word_idx + return word_idx + + def _get_token_idx(self, x): + """ + This is to extend noising functions to be able to apply to non-bpe + tokens, e.g. word or characters. + """ + x = torch.t(x) + word_idx = np.array([range(len(x_i)) for x_i in x]) + return np.transpose(word_idx) + + +class WordDropout(WordNoising): + """Randomly drop input words. If not passing blank_idx (default is None), + then dropped words will be removed. Otherwise, it will be replaced by the + blank_idx.""" + + def __init__( + self, + dictionary, + default_dropout_prob=0.1, + bpe_cont_marker="@@", + bpe_end_marker=None, + ): + super().__init__(dictionary, bpe_cont_marker, bpe_end_marker) + self.default_dropout_prob = default_dropout_prob + + def noising(self, x, lengths, dropout_prob=None, blank_idx=None): + if dropout_prob is None: + dropout_prob = self.default_dropout_prob + # x: (T x B), lengths: B + if dropout_prob == 0: + return x, lengths + + assert 0 < dropout_prob < 1 + + # be sure to drop entire words + word_idx = self.get_word_idx(x) + sentences = [] + modified_lengths = [] + for i in range(lengths.size(0)): + # Since dropout probabilities need to apply over non-pad tokens, + # it is not trivial to generate the keep mask without consider + # input lengths; otherwise, this could be done outside the loop + + # We want to drop whole words based on word_idx grouping + num_words = max(word_idx[:, i]) + 1 + + # ith example: [x0, x1, ..., eos, pad, ..., pad] + # We should only generate keep probs for non-EOS tokens. Thus if the + # input sentence ends in EOS, the last word idx is not included in + # the dropout mask generation and we append True to always keep EOS. + # Otherwise, just generate the dropout mask for all word idx + # positions. + has_eos = x[lengths[i] - 1, i] == self.dictionary.eos() + if has_eos: # has eos? + keep = np.random.rand(num_words - 1) >= dropout_prob + keep = np.append(keep, [True]) # keep EOS symbol + else: + keep = np.random.rand(num_words) >= dropout_prob + + words = x[: lengths[i], i].tolist() + + # TODO: speed up the following loop + # drop words from the input according to keep + new_s = [ + w if keep[word_idx[j, i]] else blank_idx for j, w in enumerate(words) + ] + new_s = [w for w in new_s if w is not None] + # we need to have at least one word in the sentence (more than the + # start / end sentence symbols) + if len(new_s) <= 1: + # insert at beginning in case the only token left is EOS + # EOS should be at end of list. + new_s.insert(0, words[np.random.randint(0, len(words))]) + assert len(new_s) >= 1 and ( + not has_eos # Either don't have EOS at end or last token is EOS + or (len(new_s) >= 2 and new_s[-1] == self.dictionary.eos()) + ), "New sentence is invalid." + sentences.append(new_s) + modified_lengths.append(len(new_s)) + # re-construct input + modified_lengths = torch.LongTensor(modified_lengths) + modified_x = torch.LongTensor( + modified_lengths.max(), modified_lengths.size(0) + ).fill_(self.dictionary.pad()) + for i in range(modified_lengths.size(0)): + modified_x[: modified_lengths[i], i].copy_(torch.LongTensor(sentences[i])) + + return modified_x, modified_lengths + + +class WordShuffle(WordNoising): + """Shuffle words by no more than k positions.""" + + def __init__( + self, + dictionary, + default_max_shuffle_distance=3, + bpe_cont_marker="@@", + bpe_end_marker=None, + ): + super().__init__(dictionary, bpe_cont_marker, bpe_end_marker) + self.default_max_shuffle_distance = 3 + + def noising(self, x, lengths, max_shuffle_distance=None): + if max_shuffle_distance is None: + max_shuffle_distance = self.default_max_shuffle_distance + # x: (T x B), lengths: B + if max_shuffle_distance == 0: + return x, lengths + + # max_shuffle_distance < 1 will return the same sequence + assert max_shuffle_distance > 1 + + # define noise word scores + noise = np.random.uniform( + 0, + max_shuffle_distance, + size=(x.size(0), x.size(1)), + ) + noise[0] = -1 # do not move start sentence symbol + # be sure to shuffle entire words + word_idx = self.get_word_idx(x) + x2 = x.clone() + for i in range(lengths.size(0)): + length_no_eos = lengths[i] + if x[lengths[i] - 1, i] == self.dictionary.eos(): + length_no_eos = lengths[i] - 1 + # generate a random permutation + scores = word_idx[:length_no_eos, i] + noise[word_idx[:length_no_eos, i], i] + # ensure no reordering inside a word + scores += 1e-6 * np.arange(length_no_eos.item()) + permutation = scores.argsort() + # shuffle words + x2[:length_no_eos, i].copy_( + x2[:length_no_eos, i][torch.from_numpy(permutation)] + ) + return x2, lengths + + +class UnsupervisedMTNoising(WordNoising): + """ + Implements the default configuration for noising in UnsupervisedMT + (github.com/facebookresearch/UnsupervisedMT) + """ + + def __init__( + self, + dictionary, + max_word_shuffle_distance, + word_dropout_prob, + word_blanking_prob, + bpe_cont_marker="@@", + bpe_end_marker=None, + ): + super().__init__(dictionary) + self.max_word_shuffle_distance = max_word_shuffle_distance + self.word_dropout_prob = word_dropout_prob + self.word_blanking_prob = word_blanking_prob + + self.word_dropout = WordDropout( + dictionary=dictionary, + bpe_cont_marker=bpe_cont_marker, + bpe_end_marker=bpe_end_marker, + ) + self.word_shuffle = WordShuffle( + dictionary=dictionary, + bpe_cont_marker=bpe_cont_marker, + bpe_end_marker=bpe_end_marker, + ) + + def noising(self, x, lengths): + # 1. Word Shuffle + noisy_src_tokens, noisy_src_lengths = self.word_shuffle.noising( + x=x, + lengths=lengths, + max_shuffle_distance=self.max_word_shuffle_distance, + ) + # 2. Word Dropout + noisy_src_tokens, noisy_src_lengths = self.word_dropout.noising( + x=noisy_src_tokens, + lengths=noisy_src_lengths, + dropout_prob=self.word_dropout_prob, + ) + # 3. Word Blanking + noisy_src_tokens, noisy_src_lengths = self.word_dropout.noising( + x=noisy_src_tokens, + lengths=noisy_src_lengths, + dropout_prob=self.word_blanking_prob, + blank_idx=self.dictionary.unk(), + ) + + return noisy_src_tokens + + +class NoisingDataset(torch.utils.data.Dataset): + def __init__( + self, + src_dataset, + src_dict, + seed, + noiser=None, + noising_class=UnsupervisedMTNoising, + **kwargs + ): + """ + Wrap a :class:`~torch.utils.data.Dataset` and apply noise to the + samples based on the supplied noising configuration. + + Args: + src_dataset (~torch.utils.data.Dataset): dataset to wrap. + to build self.src_dataset -- + a LanguagePairDataset with src dataset as the source dataset and + None as the target dataset. Should NOT have padding so that + src_lengths are accurately calculated by language_pair_dataset + collate function. + We use language_pair_dataset here to encapsulate the tgt_dataset + so we can re-use the LanguagePairDataset collater to format the + batches in the structure that SequenceGenerator expects. + src_dict (~fairseq.data.Dictionary): source dictionary + seed (int): seed to use when generating random noise + noiser (WordNoising): a pre-initialized :class:`WordNoising` + instance. If this is None, a new instance will be created using + *noising_class* and *kwargs*. + noising_class (class, optional): class to use to initialize a + default :class:`WordNoising` instance. + kwargs (dict, optional): arguments to initialize the default + :class:`WordNoising` instance given by *noiser*. + """ + self.src_dataset = src_dataset + self.src_dict = src_dict + self.seed = seed + self.noiser = ( + noiser + if noiser is not None + else noising_class( + dictionary=src_dict, + **kwargs, + ) + ) + self.sizes = src_dataset.sizes + + + def __getitem__(self, index): + """ + Returns a single noisy sample. Multiple samples are fed to the collater + create a noising dataset batch. + """ + src_tokens = self.src_dataset[index] + src_lengths = torch.LongTensor([len(src_tokens)]) + src_tokens = src_tokens.unsqueeze(0) + + # Transpose src tokens to fit expected shape of x in noising function + # (batch size, sequence length) -> (sequence length, batch size) + src_tokens_t = torch.t(src_tokens) + + with data_utils.numpy_seed(self.seed + index): + noisy_src_tokens = self.noiser.noising(src_tokens_t, src_lengths) + + # Transpose back to expected src_tokens format + # (sequence length, 1) -> (1, sequence length) + noisy_src_tokens = torch.t(noisy_src_tokens) + return noisy_src_tokens[0] + + def __len__(self): + """ + The length of the noising dataset is the length of src. + """ + return len(self.src_dataset) + + @property + def supports_prefetch(self): + return self.src_dataset.supports_prefetch + + def prefetch(self, indices): + if self.src_dataset.supports_prefetch: + self.src_dataset.prefetch(indices) diff --git a/fairseq/data/num_samples_dataset.py b/fairseq/data/num_samples_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..99a17495c701d8a05e0268f98bf453905e11d078 --- /dev/null +++ b/fairseq/data/num_samples_dataset.py @@ -0,0 +1,17 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from . import FairseqDataset + + +class NumSamplesDataset(FairseqDataset): + def __getitem__(self, index): + return 1 + + def __len__(self): + return 0 + + def collater(self, samples): + return sum(samples) diff --git a/fairseq/data/numel_dataset.py b/fairseq/data/numel_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..ac86dfd2f1d89055de909656d61d6aca85523f00 --- /dev/null +++ b/fairseq/data/numel_dataset.py @@ -0,0 +1,31 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import numpy as np +import torch + +from . import BaseWrapperDataset + + +class NumelDataset(BaseWrapperDataset): + def __init__(self, dataset, reduce=False): + super().__init__(dataset) + self.reduce = reduce + + def __getitem__(self, index): + item = self.dataset[index] + if torch.is_tensor(item): + return torch.numel(item) + else: + return np.size(item) + + def __len__(self): + return len(self.dataset) + + def collater(self, samples): + if self.reduce: + return sum(samples) + else: + return torch.tensor(samples) diff --git a/fairseq/data/offset_tokens_dataset.py b/fairseq/data/offset_tokens_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..6fabbdcdaa1a8f70d8d8c07db4cd53754503c194 --- /dev/null +++ b/fairseq/data/offset_tokens_dataset.py @@ -0,0 +1,15 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from . import BaseWrapperDataset + + +class OffsetTokensDataset(BaseWrapperDataset): + def __init__(self, dataset, offset): + super().__init__(dataset) + self.offset = offset + + def __getitem__(self, idx): + return self.dataset[idx] + self.offset diff --git a/fairseq/data/pad_dataset.py b/fairseq/data/pad_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..8075bba6a9efc5f8421368ee0b2ae66afe3f5009 --- /dev/null +++ b/fairseq/data/pad_dataset.py @@ -0,0 +1,28 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq.data import data_utils + +from . import BaseWrapperDataset + + +class PadDataset(BaseWrapperDataset): + def __init__(self, dataset, pad_idx, left_pad): + super().__init__(dataset) + self.pad_idx = pad_idx + self.left_pad = left_pad + + def collater(self, samples): + return data_utils.collate_tokens(samples, self.pad_idx, left_pad=self.left_pad) + + +class LeftPadDataset(PadDataset): + def __init__(self, dataset, pad_idx): + super().__init__(dataset, pad_idx, left_pad=True) + + +class RightPadDataset(PadDataset): + def __init__(self, dataset, pad_idx): + super().__init__(dataset, pad_idx, left_pad=False) diff --git a/fairseq/data/plasma_utils.py b/fairseq/data/plasma_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..b9fab3b739db46b685fa6859a2f851a14eef8407 --- /dev/null +++ b/fairseq/data/plasma_utils.py @@ -0,0 +1,197 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +import subprocess +import json +import tempfile +import hashlib +from typing import Hashable + +try: + import pyarrow.plasma as plasma + + PYARROW_AVAILABLE = True +except ImportError: + plasma = None + PYARROW_AVAILABLE = False + + +class PlasmaArray: + """ + Wrapper around numpy arrays that automatically moves the data to shared + memory upon serialization. This is particularly helpful when passing numpy + arrays through multiprocessing, so that data is not unnecessarily + duplicated or pickled. + """ + + def __init__(self, array): + super().__init__() + self.array = array + self.disable = array.nbytes < 134217728 # disable for arrays <128MB + self.object_id = None + self.path = None + + # variables with underscores shouldn't be pickled + self._client = None + self._server = None + self._server_tmp = None + self._plasma = None + + @property + def plasma(self): + if self._plasma is None and not self.disable: + self._plasma = plasma + return self._plasma + + def start_server(self): + if self.plasma is None or self._server is not None: + return + assert self.object_id is None + assert self.path is None + self._server_tmp = tempfile.NamedTemporaryFile() + self.path = self._server_tmp.name + self._server = subprocess.Popen( + ["plasma_store", "-m", str(int(1.05 * self.array.nbytes)), "-s", self.path] + ) + + @property + def client(self): + if self._client is None: + assert self.path is not None + self._client = self.plasma.connect(self.path, num_retries=200) + return self._client + + def __getstate__(self): + """Called on pickle load""" + if self.plasma is None: + return self.__dict__ + if self.object_id is None: + self.start_server() + self.object_id = self.client.put(self.array) + state = self.__dict__.copy() + del state["array"] + state["_client"] = None + state["_server"] = None + state["_server_tmp"] = None + state["_plasma"] = None + return state + + def __setstate__(self, state): + """Called on pickle save""" + self.__dict__.update(state) + if self.plasma is None: + return + self.array = self.client.get(self.object_id) + + def __del__(self): + if self._server is not None: + self._server.kill() + self._server = None + self._server_tmp.close() + self._server_tmp = None + + +DEFAULT_PLASMA_PATH = "/tmp/plasma" + + +class PlasmaView: + """Interface to write and read from shared memory. Whereas PlasmaArray writes to plasma on serialization, + PlasmaView writes to shared memory on instantiation.""" + + def __init__(self, array, split_path: str, hash_data: Hashable, plasma_path=None): + """ + Args: + array: numpy array to store. This can be read with ``PlasmaView().array`` + split_path: the path whence the data was read, used for hashing + hash_data: other metadata about the array that can be used to create a unique key. + as of writing, the 3 callers in ``TokenBlockDataset`` use:: + + hash_data = ((block_size, document_sep_len, str(break_mode), len(dataset)), 0|1|2) + + + """ + assert PYARROW_AVAILABLE + assert split_path is not None + if plasma_path is None: + plasma_path = DEFAULT_PLASMA_PATH + + self.path = plasma_path + self.split_path = split_path + self._client = None # Initialize lazily for pickle. plasma clients should not be deep copied or serialized. + self._n = None + + self.object_id = self.get_object_id(self.split_path, hash_data) + try: + self.client.put(array, object_id=self.object_id) + except plasma.PlasmaObjectExists: + pass + + @property + def client(self): + if self._client is None: + self._client = plasma.connect(self.path, num_retries=200) + return self._client + + @property + def array(self): + """Fetch a read only view of an np.array, stored in plasma.""" + ret = self.client.get(self.object_id) + return ret + + @staticmethod + def get_object_id(split_path: str, hash_data: Hashable): + """Returns plasma.ObjectID from hashing split_path and object_num.""" + hash = hashlib.blake2b(bytes(split_path, "utf-8"), digest_size=20) + harg = json.dumps(hash_data).encode("utf-8") + hash.update(harg) + return plasma.ObjectID(hash.digest()) + + def __getstate__(self): + """Called on pickle save""" + self.disconnect() + state = self.__dict__.copy() + assert state["_client"] is None + assert "object_id" in state + return state + + def __setstate__(self, state): + """Called on pickle load""" + self.__dict__.update(state) + + def __del__(self): + self.disconnect() + + def disconnect(self): + if self._client is not None: + self._client.disconnect() + self._client = None + + def __len__(self): + """Save reads by caching len""" + if self._n is None: + self._n = len(self.array) + return self._n + + +GB100 = (1024 ** 3) * 100 + + +class PlasmaStore: + def __init__(self, path=DEFAULT_PLASMA_PATH, nbytes: int = GB100): + + self.server = self.start(path, nbytes) + + def __del__(self): + self.server.kill() + + @staticmethod + def start(path=DEFAULT_PLASMA_PATH, nbytes: int = GB100) -> subprocess.Popen: + if not PYARROW_AVAILABLE: + raise ImportError("please run pip install pyarrow to use --use_plasma_view") + # best practice is to allocate more space than we need. The limitation seems to be the size of /dev/shm + _server = subprocess.Popen(["plasma_store", "-m", str(nbytes), "-s", path]) + plasma.connect(path, num_retries=200) # If we can't connect we fail immediately + return _server diff --git a/fairseq/data/prepend_dataset.py b/fairseq/data/prepend_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..ad74784d2d7920e4a6225282d95543ce16ea50d9 --- /dev/null +++ b/fairseq/data/prepend_dataset.py @@ -0,0 +1,28 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import numpy as np +import torch + +from . import BaseWrapperDataset + + +class PrependDataset(BaseWrapperDataset): + def __init__(self, dataset, prepend_getter, ensure_first_token_is=None): + super().__init__(dataset) + self.prepend_getter = prepend_getter + self.ensure_first_token = ensure_first_token_is + + def __getitem__(self, idx): + item = self.dataset[idx] + is_tuple = isinstance(item, tuple) + src = item[0] if is_tuple else item + + assert self.ensure_first_token is None or src[0] == self.ensure_first_token + prepend_idx = self.prepend_getter(self.dataset, idx) + assert isinstance(prepend_idx, int) + src[0] = prepend_idx + item = tuple((src,) + item[1:]) if is_tuple else src + return item diff --git a/fairseq/data/prepend_token_dataset.py b/fairseq/data/prepend_token_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..fd1331f4c44c1595eb9bb78baa0cf5cf3bcce9ad --- /dev/null +++ b/fairseq/data/prepend_token_dataset.py @@ -0,0 +1,41 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import numpy as np +import torch + +from . import BaseWrapperDataset + + +class PrependTokenDataset(BaseWrapperDataset): + def __init__(self, dataset, token=None): + super().__init__(dataset) + self.token = token + if token is not None: + self._sizes = np.array(dataset.sizes) + 1 + else: + self._sizes = dataset.sizes + + def __getitem__(self, idx): + item = self.dataset[idx] + if self.token is not None: + item = torch.cat([item.new([self.token]), item]) + return item + + @property + def sizes(self): + return self._sizes + + def num_tokens(self, index): + n = self.dataset.num_tokens(index) + if self.token is not None: + n += 1 + return n + + def size(self, index): + n = self.dataset.size(index) + if self.token is not None: + n += 1 + return n diff --git a/fairseq/data/raw_label_dataset.py b/fairseq/data/raw_label_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d054904f419bd64855d33a2a770b43f671c7c8d8 --- /dev/null +++ b/fairseq/data/raw_label_dataset.py @@ -0,0 +1,23 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch + +from . import FairseqDataset + + +class RawLabelDataset(FairseqDataset): + def __init__(self, labels): + super().__init__() + self.labels = labels + + def __getitem__(self, index): + return self.labels[index] + + def __len__(self): + return len(self.labels) + + def collater(self, samples): + return torch.tensor(samples) diff --git a/fairseq/data/replace_dataset.py b/fairseq/data/replace_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..5aac2ba96bee0a8bb65f4c9e56fa0b17248ee1d9 --- /dev/null +++ b/fairseq/data/replace_dataset.py @@ -0,0 +1,36 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from . import BaseWrapperDataset + + +class ReplaceDataset(BaseWrapperDataset): + """Replaces tokens found in the dataset by a specified replacement token + + Args: + dataset (~torch.utils.data.Dataset): dataset to replace tokens in + replace_map(Dictionary[int,int]): map of token to replace -> replacement token + offsets (List[int]): do not replace tokens before (from left if pos, right if neg) this offset. should be + as many as the number of objects returned by the underlying dataset __getitem__ method. + """ + + def __init__(self, dataset, replace_map, offsets): + super().__init__(dataset) + assert len(replace_map) > 0 + self.replace_map = replace_map + self.offsets = offsets + + def __getitem__(self, index): + item = self.dataset[index] + is_tuple = isinstance(item, tuple) + srcs = item if is_tuple else [item] + + for offset, src in zip(self.offsets, srcs): + for k, v in self.replace_map.items(): + src_off = src[offset:] if offset >= 0 else src[:offset] + src_off.masked_fill_(src_off == k, v) + + item = srcs if is_tuple else srcs[0] + return item diff --git a/fairseq/data/resampling_dataset.py b/fairseq/data/resampling_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..3d3b993164dc3962df48bacff26714328e843e80 --- /dev/null +++ b/fairseq/data/resampling_dataset.py @@ -0,0 +1,139 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging + +import numpy as np +from fairseq.data import BaseWrapperDataset, plasma_utils + + +logger = logging.getLogger(__name__) + + +class ResamplingDataset(BaseWrapperDataset): + """Randomly samples from a given dataset at each epoch. + + Sampling is done with or without replacement, depending on the "replace" + parameter. + + Optionally, the epoch size can be rescaled. This is potentially desirable + to increase per-epoch coverage of the base dataset (since sampling with + replacement means that many items in the dataset will be left out). In the + case of sampling without replacement, size_ratio should be strictly less + than 1. + + Args: + dataset (~torch.utils.data.Dataset): dataset on which to sample. + weights (List[float]): list of probability weights + (default: None, which corresponds to uniform sampling). + replace (bool): sampling mode; True for "with replacement", or False + for "without replacement" (default: True) + size_ratio (float): the ratio to subsample to; must be positive + (default: 1.0). + batch_by_size (bool): whether or not to batch by sequence length + (default: True). + seed (int): RNG seed to use (default: 0). + epoch (int): starting epoch number (default: 1). + """ + + def __init__( + self, + dataset, + weights=None, + replace=True, + size_ratio=1.0, + batch_by_size=True, + seed=0, + epoch=1, + ): + super().__init__(dataset) + + if weights is None: + self.weights = None + + else: + assert len(weights) == len(dataset) + weights_arr = np.array(weights, dtype=np.float64) + weights_arr /= weights_arr.sum() + self.weights = plasma_utils.PlasmaArray(weights_arr) + + self.replace = replace + + assert size_ratio > 0.0 + if not self.replace: + assert size_ratio < 1.0 + self.size_ratio = float(size_ratio) + self.actual_size = np.ceil(len(dataset) * self.size_ratio).astype(int) + + self.batch_by_size = batch_by_size + self.seed = seed + + self._cur_epoch = None + self._cur_indices = None + + self.set_epoch(epoch) + + def __getitem__(self, index): + return self.dataset[self._cur_indices.array[index]] + + def __len__(self): + return self.actual_size + + @property + def sizes(self): + if isinstance(self.dataset.sizes, list): + return [s[self._cur_indices.array] for s in self.dataset.sizes] + return self.dataset.sizes[self._cur_indices.array] + + def num_tokens(self, index): + return self.dataset.num_tokens(self._cur_indices.array[index]) + + def size(self, index): + return self.dataset.size(self._cur_indices.array[index]) + + def ordered_indices(self): + if self.batch_by_size: + order = [ + np.arange(len(self)), + self.sizes, + ] # No need to handle `self.shuffle == True` + return np.lexsort(order) + else: + return np.arange(len(self)) + + def prefetch(self, indices): + self.dataset.prefetch(self._cur_indices.array[indices]) + + @property + def can_reuse_epoch_itr_across_epochs(self): + return False + + def set_epoch(self, epoch): + logger.debug("ResamplingDataset.set_epoch: {}".format(epoch)) + super().set_epoch(epoch) + + if epoch == self._cur_epoch: + return + + self._cur_epoch = epoch + + # Generate a weighted sample of indices as a function of the + # random seed and the current epoch. + + rng = np.random.RandomState( + [ + 42, # magic number + self.seed % (2 ** 32), # global seed + self._cur_epoch, # epoch index + ] + ) + self._cur_indices = plasma_utils.PlasmaArray( + rng.choice( + len(self.dataset), + self.actual_size, + replace=self.replace, + p=(None if self.weights is None else self.weights.array), + ) + ) diff --git a/fairseq/data/roll_dataset.py b/fairseq/data/roll_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..a2915eeb3e8fb4dfb4b2bb33e0464ad0783d854c --- /dev/null +++ b/fairseq/data/roll_dataset.py @@ -0,0 +1,18 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch + +from . import BaseWrapperDataset + + +class RollDataset(BaseWrapperDataset): + def __init__(self, dataset, shifts): + super().__init__(dataset) + self.shifts = shifts + + def __getitem__(self, index): + item = self.dataset[index] + return torch.roll(item, self.shifts) diff --git a/fairseq/data/round_robin_zip_datasets.py b/fairseq/data/round_robin_zip_datasets.py new file mode 100644 index 0000000000000000000000000000000000000000..2cb7447ea955a7c3ae7372f09ee426c08acd430e --- /dev/null +++ b/fairseq/data/round_robin_zip_datasets.py @@ -0,0 +1,160 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +from collections import OrderedDict +from typing import Dict, Sequence + +import numpy as np + +from . import FairseqDataset, LanguagePairDataset + +logger = logging.getLogger(__name__) + + +class RoundRobinZipDatasets(FairseqDataset): + """Zip multiple :class:`~fairseq.data.FairseqDataset` instances together. + + Shorter datasets are repeated in a round-robin fashion to match the length + of the longest one. + + Args: + datasets (Dict[~fairseq.data.FairseqDataset]): a dictionary of + :class:`~fairseq.data.FairseqDataset` instances. + eval_key (str, optional): a key used at evaluation time that causes + this instance to pass-through batches from *datasets[eval_key]*. + """ + + def __init__(self, datasets, eval_key=None): + super().__init__() + if isinstance(datasets, dict): + datasets = OrderedDict(datasets) + assert isinstance(datasets, OrderedDict) + assert datasets, "Can't make a RoundRobinZipDatasets out of nothing" + for dataset in datasets.values(): + assert isinstance(dataset, FairseqDataset) + + self.datasets = datasets + self.eval_key = eval_key + + self.longest_dataset_key = max(datasets, key=lambda k: len(datasets[k])) + self.longest_dataset = datasets[self.longest_dataset_key] + self._ordered_indices: Dict[str, Sequence[int]] = None + + def _map_index(self, key, index): + assert ( + self._ordered_indices is not None + ), "Must call RoundRobinZipDatasets.ordered_indices() first" + o = self._ordered_indices[key] + return o[index % len(o)] + + def __getitem__(self, index): + if self.eval_key is None: + return OrderedDict( + [ + (key, dataset[self._map_index(key, index)]) + for key, dataset in self.datasets.items() + ] + ) + else: + # at evaluation time it's useful to pass-through batches from a single key + return self.datasets[self.eval_key][self._map_index(self.eval_key, index)] + + def __len__(self): + if self._ordered_indices is not None: + return len(self._ordered_indices[self.longest_dataset_key]) + return len(self.longest_dataset) + + def collater(self, samples): + """Merge a list of samples to form a mini-batch.""" + if len(samples) == 0: + return None + if self.eval_key is None: + return OrderedDict( + [ + (key, dataset.collater([sample[key] for sample in samples])) + for key, dataset in self.datasets.items() + ] + ) + else: + # at evaluation time it's useful to pass-through batches from a single key + return self.datasets[self.eval_key].collater(samples) + + def num_tokens(self, index): + """Return an example's length (number of tokens), used for batching.""" + # TODO make it configurable whether to use max() or sum() here + return max( + dataset.num_tokens(self._map_index(key, index)) + for key, dataset in self.datasets.items() + ) + + def size(self, index): + """Return an example's size as a float or tuple. This value is used when + filtering a dataset with ``--max-positions``.""" + return { + key: dataset.size(self._map_index(key, index)) + for key, dataset in self.datasets.items() + } + + def ordered_indices(self): + """Ordered indices for batching.""" + if self._ordered_indices is None: + # Call the underlying dataset's ordered_indices() here, so that we + # get the same random ordering as we would have from using the + # underlying sub-datasets directly. + self._ordered_indices = OrderedDict( + [ + (key, dataset.ordered_indices()) + for key, dataset in self.datasets.items() + ] + ) + return np.arange(len(self)) + + def filter_indices_by_size(self, indices, max_positions=None): + """ + Filter each sub-dataset independently, then update the round robin to work + on the filtered sub-datasets. + """ + + def _deep_until_language_pair(dataset): + if isinstance(dataset, LanguagePairDataset): + return dataset + if hasattr(dataset, "tgt_dataset"): + return _deep_until_language_pair(dataset.tgt_dataset) + if hasattr(dataset, "dataset"): + return _deep_until_language_pair(dataset.dataset) + raise Exception(f"Don't know how to unwrap this dataset: {dataset}") + + if not isinstance(max_positions, dict): + max_positions = {k: max_positions for k in self.datasets.keys()} + ignored_some = False + for key, dataset in self.datasets.items(): + dataset = _deep_until_language_pair(dataset) + self._ordered_indices[key], ignored = dataset.filter_indices_by_size( + self._ordered_indices[key], max_positions[key] + ) + if len(ignored) > 0: + ignored_some = True + logger.warning( + f"{len(ignored)} samples from {key} have invalid sizes and will be skipped, " + f"max_positions={max_positions[key]}, first few sample ids={ignored[:10]}" + ) + # Since we are modifying in place the _ordered_indices, + # it's not possible anymore to return valid ignored indices. + # Hopefully the extra debug information print above should be enough to debug. + # Ideally we would receive ignore_invalid_inputs so that we could have + # a proper error message. + return (np.arange(len(self)), [0] if ignored_some else []) + + @property + def supports_prefetch(self): + return all( + getattr(dataset, "supports_prefetch", False) + for dataset in self.datasets.values() + ) + + def prefetch(self, indices): + for key, dataset in self.datasets.items(): + dataset.prefetch([self._map_index(key, index) for index in indices]) diff --git a/fairseq/data/shorten_dataset.py b/fairseq/data/shorten_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..6ebb5d88feb3f29d1512a0873df304915d051209 --- /dev/null +++ b/fairseq/data/shorten_dataset.py @@ -0,0 +1,78 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import numpy as np +from fairseq.data import data_utils + +from . import BaseWrapperDataset + + +class TruncateDataset(BaseWrapperDataset): + """Truncate a sequence by returning the first truncation_length tokens""" + + def __init__(self, dataset, truncation_length): + super().__init__(dataset) + assert truncation_length is not None + self.truncation_length = truncation_length + self.dataset = dataset + + def __getitem__(self, index): + item = self.dataset[index] + item_len = item.size(0) + if item_len > self.truncation_length: + item = item[: self.truncation_length] + return item + + @property + def sizes(self): + return np.minimum(self.dataset.sizes, self.truncation_length) + + def __len__(self): + return len(self.dataset) + + +class RandomCropDataset(TruncateDataset): + """Truncate a sequence by returning a random crop of truncation_length tokens""" + + def __init__(self, dataset, truncation_length, seed=1): + super().__init__(dataset, truncation_length) + self.seed = seed + self.epoch = 0 + + @property + def can_reuse_epoch_itr_across_epochs(self): + return True # only the crop changes, not item sizes + + def set_epoch(self, epoch, **unused): + super().set_epoch(epoch) + self.epoch = epoch + + def __getitem__(self, index): + with data_utils.numpy_seed(self.seed, self.epoch, index): + item = self.dataset[index] + item_len = item.size(0) + excess = item_len - self.truncation_length + if excess > 0: + start_idx = np.random.randint(0, excess) + item = item[start_idx : start_idx + self.truncation_length] + return item + + +def maybe_shorten_dataset( + dataset, + split, + shorten_data_split_list, + shorten_method, + tokens_per_sample, + seed, +): + truncate_split = ( + split in shorten_data_split_list.split(",") or len(shorten_data_split_list) == 0 + ) + if shorten_method == "truncate" and truncate_split: + dataset = TruncateDataset(dataset, tokens_per_sample) + elif shorten_method == "random_crop" and truncate_split: + dataset = RandomCropDataset(dataset, tokens_per_sample, seed) + return dataset diff --git a/fairseq/data/sort_dataset.py b/fairseq/data/sort_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..b3890e7279e1f26db2e48ec0a91c639e9299d60f --- /dev/null +++ b/fairseq/data/sort_dataset.py @@ -0,0 +1,21 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import numpy as np + +from . import BaseWrapperDataset + + +class SortDataset(BaseWrapperDataset): + def __init__(self, dataset, sort_order): + super().__init__(dataset) + if not isinstance(sort_order, (list, tuple)): + sort_order = [sort_order] + self.sort_order = sort_order + + assert all(len(so) == len(dataset) for so in sort_order) + + def ordered_indices(self): + return np.lexsort(self.sort_order) diff --git a/fairseq/data/strip_token_dataset.py b/fairseq/data/strip_token_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..cae39ba4d2f8106398eccd7eb0cf5c2194ec0db5 --- /dev/null +++ b/fairseq/data/strip_token_dataset.py @@ -0,0 +1,20 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from . import BaseWrapperDataset + + +class StripTokenDataset(BaseWrapperDataset): + def __init__(self, dataset, id_to_strip): + super().__init__(dataset) + self.id_to_strip = id_to_strip + + def __getitem__(self, index): + item = self.dataset[index] + while len(item) > 0 and item[-1] == self.id_to_strip: + item = item[:-1] + while len(item) > 0 and item[0] == self.id_to_strip: + item = item[1:] + return item diff --git a/fairseq/data/subsample_dataset.py b/fairseq/data/subsample_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..48feaf883f87dc95f8637c24d3c96f3f9fd8bd1d --- /dev/null +++ b/fairseq/data/subsample_dataset.py @@ -0,0 +1,72 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging + +import numpy as np + +from . import BaseWrapperDataset + + +logger = logging.getLogger(__name__) + + +class SubsampleDataset(BaseWrapperDataset): + """Subsamples a given dataset by a specified ratio. Subsampling is done on the number of examples + + Args: + dataset (~torch.utils.data.Dataset): dataset to subsample + size_ratio(float): the ratio to subsample to. must be between 0 and 1 (exclusive) + """ + + def __init__(self, dataset, size_ratio, shuffle=False): + super().__init__(dataset) + assert size_ratio < 1 + self.actual_size = np.ceil(len(dataset) * size_ratio).astype(int) + self.indices = np.random.choice( + list(range(len(self.dataset))), self.actual_size, replace=False + ) + self.shuffle = shuffle + logger.info( + "subsampled dataset from {} to {} (ratio={})".format( + len(self.dataset), self.actual_size, size_ratio + ) + ) + + def __getitem__(self, index): + return self.dataset[self.indices[index]] + + def __len__(self): + return self.actual_size + + def collater(self, samples): + return self.dataset.collater(samples) + + @property + def sizes(self): + return self.dataset.sizes[self.indices] + + @property + def name(self): + return self.dataset.name + + def num_tokens(self, index): + return self.dataset.num_tokens(self.indices[index]) + + def size(self, index): + return self.dataset.size(self.indices[index]) + + def ordered_indices(self): + """Return an ordered list of indices. Batches will be constructed based + on this order.""" + if self.shuffle: + order = [np.random.permutation(len(self))] + else: + order = [np.arange(len(self))] + order.append(self.sizes) + return np.lexsort(order) + + def prefetch(self, indices): + self.dataset.prefetch(self.indices[indices]) diff --git a/fairseq/data/token_block_dataset.py b/fairseq/data/token_block_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d2c65fd7e058072911c3aa60bfc760288a0f83e5 --- /dev/null +++ b/fairseq/data/token_block_dataset.py @@ -0,0 +1,202 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import numpy as np +import torch +from fairseq.data import FairseqDataset, plasma_utils +from fairseq.data.indexed_dataset import best_fitting_int_dtype +from typing import Tuple + + +class TokenBlockDataset(FairseqDataset): + """Break a Dataset of tokens into blocks. + + Args: + dataset (~torch.utils.data.Dataset): dataset to break into blocks + sizes (List[int]): sentence lengths (required for 'complete' and 'eos') + block_size (int): maximum block size (ignored in 'eos' break mode) + break_mode (str, optional): Mode used for breaking tokens. Values can + be one of: + - 'none': break tokens into equally sized blocks (up to block_size) + - 'complete': break tokens into blocks (up to block_size) such that + blocks contains complete sentences, although block_size may be + exceeded if some sentences exceed block_size + - 'complete_doc': similar to 'complete' mode, but do not + cross document boundaries + - 'eos': each block contains one sentence (block_size is ignored) + include_targets (bool, optional): return next tokens as targets + (default: False). + document_sep_len (int, optional): document separator size (required for + 'complete_doc' break mode). Typically 1 if the sentences have eos + and 0 otherwise. + """ + + def __init__( + self, + dataset, + sizes, + block_size, + pad, + eos, + break_mode=None, + include_targets=False, + document_sep_len=1, + use_plasma_view=False, + split_path=None, + plasma_path=None, + ): + + super().__init__() + self.dataset = dataset + self.pad = pad + self.eos = eos + self.include_targets = include_targets + + assert len(dataset) > 0 + + assert len(dataset) == len(sizes) + _sizes, block_to_dataset_index, slice_indices = self._build_slice_indices( + sizes, break_mode, document_sep_len, block_size + ) + if use_plasma_view: + plasma_id = (block_size, document_sep_len, str(break_mode), len(dataset)) + self._slice_indices = plasma_utils.PlasmaView( + slice_indices, split_path, (plasma_id, 0), plasma_path=plasma_path + ) + self._sizes = plasma_utils.PlasmaView( + _sizes, split_path, (plasma_id, 1), plasma_path=plasma_path + ) + self._block_to_dataset_index = plasma_utils.PlasmaView( + block_to_dataset_index, split_path, (plasma_id, 2), plasma_path=plasma_path, + ) + else: + self._slice_indices = plasma_utils.PlasmaArray(slice_indices) + self._sizes = plasma_utils.PlasmaArray(_sizes) + self._block_to_dataset_index = plasma_utils.PlasmaArray( + block_to_dataset_index + ) + + @staticmethod + def _build_slice_indices( + sizes, break_mode, document_sep_len, block_size + ) -> Tuple[np.ndarray]: + """Use token_block_utils_fast to build arrays for indexing into self.dataset""" + try: + from fairseq.data.token_block_utils_fast import ( + _get_slice_indices_fast, + _get_block_to_dataset_index_fast, + ) + except ImportError: + raise ImportError( + "Please build Cython components with: `pip install --editable .` " + "or `python setup.py build_ext --inplace`" + ) + + if isinstance(sizes, list): + sizes = np.array(sizes, dtype=np.int64) + else: + if torch.is_tensor(sizes): + sizes = sizes.numpy() + sizes = sizes.astype(np.int64) + + break_mode = break_mode if break_mode is not None else "none" + + # For "eos" break-mode, block_size is not required parameters. + if break_mode == "eos" and block_size is None: + block_size = 0 + + slice_indices = _get_slice_indices_fast( + sizes, str(break_mode), block_size, document_sep_len + ) + _sizes = slice_indices[:, 1] - slice_indices[:, 0] + + # build index mapping block indices to the underlying dataset indices + if break_mode == "eos": + # much faster version for eos break mode + block_to_dataset_index = np.stack( + [ + np.arange(len(sizes)), # starting index in dataset + np.zeros( + len(sizes), dtype=np.compat.long + ), # starting offset within starting index + np.arange(len(sizes)), # ending index in dataset + ], + 1, + ) + else: + block_to_dataset_index = _get_block_to_dataset_index_fast( + sizes, slice_indices, + ) + size_dtype = np.uint16 if block_size < 65535 else np.uint32 + num_tokens = slice_indices[-1].max() + slice_indices_dtype = best_fitting_int_dtype(num_tokens) + slice_indices = slice_indices.astype(slice_indices_dtype) + _sizes = _sizes.astype(size_dtype) + block_to_dataset_index = block_to_dataset_index.astype(slice_indices_dtype) + return _sizes, block_to_dataset_index, slice_indices + + @property + def slice_indices(self): + return self._slice_indices.array + + @property + def sizes(self): + return self._sizes.array + + @property + def block_to_dataset_index(self): + return self._block_to_dataset_index.array + + def attr(self, attr: str, index: int): + start_ds_idx, _, _ = self.block_to_dataset_index[index] + return self.dataset.attr(attr, start_ds_idx) + + def __getitem__(self, index): + start_ds_idx, start_offset, end_ds_idx = self.block_to_dataset_index[index] + + buffer = torch.cat( + [self.dataset[idx] for idx in range(start_ds_idx, end_ds_idx + 1)] + ) + slice_s, slice_e = self.slice_indices[index] + length = slice_e - slice_s + s, e = start_offset, start_offset + length + item = buffer[s:e] + + if self.include_targets: + # *target* is the original sentence (=item) + # *source* is shifted right by 1 (maybe left-padded with eos) + # *past_target* is shifted right by 2 (left-padded as needed) + if s == 0: + source = torch.cat([item.new([self.eos]), buffer[0 : e - 1]]) + past_target = torch.cat( + [item.new([self.pad, self.eos]), buffer[0 : e - 2]] + ) + else: + source = buffer[s - 1 : e - 1] + if s == 1: + past_target = torch.cat([item.new([self.eos]), buffer[0 : e - 2]]) + else: + past_target = buffer[s - 2 : e - 2] + + return source, item, past_target + + return item + + def __len__(self): + return len(self.slice_indices) + + @property + def supports_prefetch(self): + return getattr(self.dataset, "supports_prefetch", False) + + def prefetch(self, indices): + self.dataset.prefetch( + { + ds_idx + for index in indices + for start_ds_idx, _, end_ds_idx in [self.block_to_dataset_index[index]] + for ds_idx in range(start_ds_idx, end_ds_idx + 1) + } + ) diff --git a/fairseq/data/token_block_utils_fast.pyx b/fairseq/data/token_block_utils_fast.pyx new file mode 100644 index 0000000000000000000000000000000000000000..08af4f30613a7b6ffa965a7c7084acabec8f8749 --- /dev/null +++ b/fairseq/data/token_block_utils_fast.pyx @@ -0,0 +1,187 @@ +# cython: language_level=3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import numpy as np +import torch +from itertools import chain +from libc.math cimport ceil + +cimport cython +cimport numpy as np + +from libc.stdint cimport int32_t, int64_t + +DTYPE = np.int64 +ctypedef int64_t DTYPE_t + + +@cython.boundscheck(False) +@cython.wraparound(False) +@cython.nonecheck(False) +cdef np.ndarray[DTYPE_t, ndim=2] _get_slice_indices_none_mode(np.ndarray[DTYPE_t, ndim=1] sizes, int block_size): + cdef DTYPE_t total_size = sizes.sum() + cdef DTYPE_t length = <DTYPE_t> ceil(total_size / <double> block_size) + cdef np.ndarray[DTYPE_t, ndim=2] slice_indices = np.zeros([length, 2], dtype=DTYPE) + cdef DTYPE_t[:, :] slice_indices_view = slice_indices + cdef DTYPE_t i + cdef DTYPE_t start + cdef DTYPE_t end + for i in range(length): + start = i * block_size + end = min(start + block_size, total_size) + slice_indices_view[i][0] = start + slice_indices_view[i][1] = end + return slice_indices + + +cdef np.ndarray[DTYPE_t, ndim=2] _fast_convert_to_np_array(list list_of_list): + """ + Faster function to convert DTYPE_t list of list. + Only fast when there are huge number of rows and low number of columns. + """ + cdef np.ndarray[DTYPE_t, ndim=1] flat = np.fromiter(chain.from_iterable(list_of_list), DTYPE, -1) + return flat.reshape((len(list_of_list), -1)) + + +@cython.boundscheck(False) +@cython.wraparound(False) +@cython.nonecheck(False) +cpdef np.ndarray[DTYPE_t, ndim=2] _get_slice_indices_fast(np.ndarray[DTYPE_t, ndim=1] sizes, str break_mode, int block_size, int document_sep_len): + cdef DTYPE_t tok_idx = 0 + cdef DTYPE_t sz_idx = 0 + cdef DTYPE_t curr_size = 0 + cdef DTYPE_t i = 0 + cdef DTYPE_t length + cdef DTYPE_t total_size + cdef DTYPE_t[:] sizes_view = sizes + cdef np.ndarray[DTYPE_t, ndim=2] slice_indices + cdef list slice_indices_list = [] + + if break_mode is None or break_mode == 'none': + slice_indices = _get_slice_indices_none_mode(sizes, block_size) + elif break_mode == 'complete': + while sz_idx < len(sizes_view): + if curr_size + sizes_view[sz_idx] <= block_size or curr_size == 0: + curr_size += sizes_view[sz_idx] + sz_idx += 1 + else: + slice_indices_list.append((tok_idx, tok_idx + curr_size)) + tok_idx += curr_size + curr_size = 0 + if curr_size > 0: + slice_indices_list.append((tok_idx, tok_idx + curr_size)) + slice_indices = _fast_convert_to_np_array(slice_indices_list) + elif break_mode == 'complete_doc': + while sz_idx < len(sizes_view): + if ( + (curr_size + sizes_view[sz_idx] <= block_size or curr_size == 0) + # an empty sentence indicates end-of-document: + and sizes_view[sz_idx] != document_sep_len + ): + curr_size += sizes_view[sz_idx] + sz_idx += 1 + else: + # Only keep non-empty documents. + if curr_size > 1: + slice_indices_list.append((tok_idx, tok_idx + curr_size)) + tok_idx += curr_size + curr_size = 0 + if sizes_view[sz_idx] == document_sep_len: + tok_idx += sizes_view[sz_idx] + sz_idx += 1 + if curr_size > 1: + slice_indices_list.append((tok_idx, tok_idx + curr_size)) + slice_indices = _fast_convert_to_np_array(slice_indices_list) + elif break_mode == 'eos': + slice_indices = np.zeros((len(sizes), 2), dtype=DTYPE) + cumsum = sizes.cumsum(axis=0) + slice_indices[1:, 0] = cumsum[:cumsum.shape[0] - 1] + slice_indices[:, 1] = cumsum + else: + raise ValueError('Invalid break_mode: ' + break_mode) + return slice_indices + + +@cython.boundscheck(False) +@cython.wraparound(False) +@cython.nonecheck(False) +cpdef np.ndarray[DTYPE_t, ndim=2] _get_block_to_dataset_index_fast(np.ndarray[DTYPE_t, ndim=1] sizes, np.ndarray[DTYPE_t, ndim=2] slice_indices): + cdef DTYPE_t start_ds_idx + cdef DTYPE_t start_offset + cdef DTYPE_t end_ds_idx + cdef DTYPE_t i + cdef DTYPE_t s + cdef DTYPE_t e + cdef DatasetSearcher ds = DatasetSearcher(sizes) + cdef np.ndarray[DTYPE_t, ndim=2] block_to_dataset_index = np.zeros([len(slice_indices), 3], dtype=DTYPE) + cdef DTYPE_t[:, :] block_to_dataset_index_view = block_to_dataset_index + cdef DTYPE_t[:, :] slice_indices_view = slice_indices + cdef Py_ssize_t x_max = slice_indices.shape[0] + + for i in range(x_max): + s = slice_indices_view[i][0] + e = slice_indices_view[i][1] + ds.seek(s) + start_ds_idx = ds.current_index + start_offset = ds.current_offset + if e <= s: + end_ds_idx = start_ds_idx + else: + ds.seek(e - 1) + end_ds_idx = ds.current_index + block_to_dataset_index_view[i][0] = start_ds_idx # starting index in dataset + block_to_dataset_index_view[i][1] = start_offset # starting offset within starting index + block_to_dataset_index_view[i][2] = end_ds_idx # ending index in dataset + return block_to_dataset_index + + +cdef class DatasetSearcher(object): + """Helper for mapping "flat" indices to indices and offsets in an + underlying dataset.""" + cdef DTYPE_t current_i + cdef DTYPE_t current_offset + cdef DTYPE_t current_index + cdef DTYPE_t[:] sizes + + def __init__(self, DTYPE_t[:] sizes): + self.sizes = sizes + self.reset() + + cdef reset(self): + self.current_offset = 0 # offset within current index in underlying dataset + self.current_i = 0 # "flat" index + self.current_index = 0 # index in underlying dataset + + @cython.boundscheck(False) + @cython.wraparound(False) + @cython.nonecheck(False) + cdef int step(self, DTYPE_t i): + cdef DTYPE_t to_consume + cdef DTYPE_t remaining + if i < self.current_i: + self.reset() + if i > self.current_i: + to_consume = i - self.current_i + remaining = self.sizes[self.current_index] - self.current_offset + if remaining > to_consume: + self.current_offset += to_consume + self.current_i += to_consume + else: + assert remaining >= 0 + self.current_i += remaining + self.current_index += 1 + self.current_offset = 0 + return 1 + return 0 + + @cython.boundscheck(False) + @cython.wraparound(False) + @cython.nonecheck(False) + cdef seek(self, DTYPE_t i): + cdef int not_done = 1 + while not_done == 1: + not_done = self.step(i) + assert self.current_i == i diff --git a/fairseq/data/transform_eos_dataset.py b/fairseq/data/transform_eos_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..fb14ff018edf13b20f5d0e486692dfb0a37ec6d1 --- /dev/null +++ b/fairseq/data/transform_eos_dataset.py @@ -0,0 +1,120 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch + +from . import FairseqDataset + + +class TransformEosDataset(FairseqDataset): + """A :class:`~fairseq.data.FairseqDataset` wrapper that appends/prepends/strips EOS. + + Note that the transformation is applied in :func:`collater`. + + Args: + dataset (~fairseq.data.FairseqDataset): dataset to wrap + eos (int): index of the end-of-sentence symbol + append_eos_to_src (bool, optional): append EOS to the end of src + remove_eos_from_src (bool, optional): remove EOS from the end of src + append_eos_to_tgt (bool, optional): append EOS to the end of tgt + remove_eos_from_tgt (bool, optional): remove EOS from the end of tgt + """ + + def __init__( + self, + dataset, + eos, + append_eos_to_src=False, + remove_eos_from_src=False, + append_eos_to_tgt=False, + remove_eos_from_tgt=False, + has_target=True, + ): + if not isinstance(dataset, FairseqDataset): + raise ValueError("dataset must be an instance of FairseqDataset") + if append_eos_to_src and remove_eos_from_src: + raise ValueError("cannot combine append_eos_to_src and remove_eos_from_src") + if append_eos_to_tgt and remove_eos_from_tgt: + raise ValueError("cannot combine append_eos_to_tgt and remove_eos_from_tgt") + + self.dataset = dataset + self.eos = torch.LongTensor([eos]) + self.append_eos_to_src = append_eos_to_src + self.remove_eos_from_src = remove_eos_from_src + self.append_eos_to_tgt = append_eos_to_tgt + self.remove_eos_from_tgt = remove_eos_from_tgt + self.has_target = has_target + + # precompute how we should adjust the reported sizes + self._src_delta = 0 + self._src_delta += 1 if append_eos_to_src else 0 + self._src_delta -= 1 if remove_eos_from_src else 0 + self._tgt_delta = 0 + self._tgt_delta += 1 if append_eos_to_tgt else 0 + self._tgt_delta -= 1 if remove_eos_from_tgt else 0 + + self._checked_src = False + self._checked_tgt = False + + def _check_src(self, src, expect_eos): + if not self._checked_src: + assert (src[-1] == self.eos[0]) == expect_eos + self._checked_src = True + + def _check_tgt(self, tgt, expect_eos): + if self.has_target and not self._checked_tgt: + assert (tgt[-1] == self.eos[0]) == expect_eos + self._checked_tgt = True + + def __getitem__(self, index): + return self.dataset[index] + + def __len__(self): + return len(self.dataset) + + def collater(self, samples): + def transform(item): + if self.append_eos_to_src: + self.eos = self.eos.to(device=item["source"].device) + self._check_src(item["source"], expect_eos=False) + item["source"] = torch.cat([item["source"], self.eos]) + if self.remove_eos_from_src: + self.eos = self.eos.to(device=item["source"].device) + self._check_src(item["source"], expect_eos=True) + item["source"] = item["source"][:-1] + if self.append_eos_to_tgt: + self.eos = self.eos.to(device=item["target"].device) + self._check_tgt(item["target"], expect_eos=False) + item["target"] = torch.cat([item["target"], self.eos]) + if self.remove_eos_from_tgt: + self.eos = self.eos.to(device=item["target"].device) + self._check_tgt(item["target"], expect_eos=True) + item["target"] = item["target"][:-1] + return item + + samples = list(map(transform, samples)) + return self.dataset.collater(samples) + + def num_tokens(self, index): + return self.dataset.num_tokens(index) + + def size(self, index): + if self.has_target: + src_len, tgt_len = self.dataset.size(index) + return (src_len + self._src_delta, tgt_len + self._tgt_delta) + else: + return self.dataset.size(index) + + def ordered_indices(self): + # NOTE: we assume that the ordering does not change based on the + # addition or removal of eos + return self.dataset.ordered_indices() + + @property + def supports_prefetch(self): + return getattr(self.dataset, "supports_prefetch", False) + + def prefetch(self, indices): + return self.dataset.prefetch(indices) diff --git a/fairseq/data/transform_eos_lang_pair_dataset.py b/fairseq/data/transform_eos_lang_pair_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..07ebdd5f3882b50cca39665715fd2b2af45f0825 --- /dev/null +++ b/fairseq/data/transform_eos_lang_pair_dataset.py @@ -0,0 +1,111 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +from typing import Optional + +import torch + +from . import FairseqDataset + + +class TransformEosLangPairDataset(FairseqDataset): + """A :class:`~fairseq.data.FairseqDataset` wrapper that transform bos on + collated samples of language pair dataset. + + Note that the transformation is applied in :func:`collater`. + + Args: + dataset (~fairseq.data.FairseqDataset): dataset that collates sample into + LanguagePairDataset schema + src_eos (int): original source end-of-sentence symbol index to be replaced + new_src_eos (int, optional): new end-of-sentence symbol index to replace source eos symbol + tgt_bos (int, optional): original target beginning-of-sentence symbol index to be replaced + new_tgt_bos (int, optional): new beginning-of-sentence symbol index to replace at the + beginning of 'prev_output_tokens' + """ + + def __init__( + self, + dataset: FairseqDataset, + src_eos: int, + new_src_eos: Optional[int] = None, + tgt_bos: Optional[int] = None, + new_tgt_bos: Optional[int] = None, + ): + self.dataset = dataset + self.src_eos = src_eos + self.new_src_eos = new_src_eos + self.tgt_bos = tgt_bos + self.new_tgt_bos = new_tgt_bos + + def __getitem__(self, index): + return self.dataset[index] + + def __len__(self): + return len(self.dataset) + + def collater(self, samples, **extra_args): + samples = self.dataset.collater(samples, **extra_args) + + if 'net_input' not in samples: + return samples + + if self.new_src_eos is not None: + if self.dataset.left_pad_source: + assert ( + samples["net_input"]["src_tokens"][:, -1] != self.src_eos + ).sum() == 0 + samples["net_input"]["src_tokens"][:, -1] = self.new_src_eos + else: + eos_idx = samples["net_input"]["src_lengths"] - 1 + assert ( + samples["net_input"]["src_tokens"][ + torch.arange(eos_idx.size(0)), eos_idx + ] + != self.src_eos + ).sum() == 0 + eos_idx = eos_idx.resize_(len(samples["net_input"]["src_lengths"]), 1) + samples["net_input"]["src_tokens"].scatter_( + 1, eos_idx, self.new_src_eos + ) + + if ( + self.new_tgt_bos is not None + and "prev_output_tokens" in samples["net_input"] + ): + if self.dataset.left_pad_target: + # TODO: support different padding direction on target side + raise NotImplementedError( + "TransformEosLangPairDataset does not implement --left-pad-target True option" + ) + else: + assert ( + samples["net_input"]["prev_output_tokens"][:, 0] != self.tgt_bos + ).sum() == 0 + samples["net_input"]["prev_output_tokens"][:, 0] = self.new_tgt_bos + + return samples + + def num_tokens(self, index): + return self.dataset.num_tokens(index) + + def size(self, index): + return self.dataset.size(index) + + @property + def sizes(self): + # dataset.sizes can be a dynamically computed sizes: + return self.dataset.sizes + + def ordered_indices(self): + return self.dataset.ordered_indices() + + @property + def supports_prefetch(self): + return getattr(self.dataset, "supports_prefetch", False) + + def prefetch(self, indices): + return self.dataset.prefetch(indices) diff --git a/fairseq/dataclass/__init__.py b/fairseq/dataclass/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..25408d28ec44cee56eb5fb3ab0c817dc04159e95 --- /dev/null +++ b/fairseq/dataclass/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .configs import FairseqDataclass +from .constants import ChoiceEnum + + +__all__ = [ + "FairseqDataclass", + "ChoiceEnum", +] diff --git a/fairseq/dataclass/configs.py b/fairseq/dataclass/configs.py new file mode 100644 index 0000000000000000000000000000000000000000..b0146fa4c7332c9f8b1f6bcff7977399dfc46f08 --- /dev/null +++ b/fairseq/dataclass/configs.py @@ -0,0 +1,990 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import sys +from dataclasses import _MISSING_TYPE, dataclass, field +from typing import Any, List, Optional + +import torch + +from fairseq.dataclass.constants import ( + DATASET_IMPL_CHOICES, + DDP_BACKEND_CHOICES, + DDP_COMM_HOOK_CHOICES, + GENERATION_CONSTRAINTS_CHOICES, + GENERATION_DECODING_FORMAT_CHOICES, + LOG_FORMAT_CHOICES, + PIPELINE_CHECKPOINT_CHOICES, + PRINT_ALIGNMENT_CHOICES, + ZERO_SHARDING_CHOICES, +) + +from omegaconf import II, MISSING + + +@dataclass +class FairseqDataclass: + """fairseq base dataclass that supported fetching attributes and metas""" + + _name: Optional[str] = None + + @staticmethod + def name(): + return None + + def _get_all_attributes(self) -> List[str]: + return [k for k in self.__dataclass_fields__.keys()] + + def _get_meta( + self, attribute_name: str, meta: str, default: Optional[Any] = None + ) -> Any: + return self.__dataclass_fields__[attribute_name].metadata.get(meta, default) + + def _get_name(self, attribute_name: str) -> str: + return self.__dataclass_fields__[attribute_name].name + + def _get_default(self, attribute_name: str) -> Any: + if hasattr(self, attribute_name): + if str(getattr(self, attribute_name)).startswith("${"): + return str(getattr(self, attribute_name)) + elif str(self.__dataclass_fields__[attribute_name].default).startswith( + "${" + ): + return str(self.__dataclass_fields__[attribute_name].default) + elif ( + getattr(self, attribute_name) + != self.__dataclass_fields__[attribute_name].default + ): + return getattr(self, attribute_name) + + f = self.__dataclass_fields__[attribute_name] + if not isinstance(f.default_factory, _MISSING_TYPE): + return f.default_factory() + return f.default + + def _get_type(self, attribute_name: str) -> Any: + return self.__dataclass_fields__[attribute_name].type + + def _get_help(self, attribute_name: str) -> Any: + return self._get_meta(attribute_name, "help") + + def _get_argparse_const(self, attribute_name: str) -> Any: + return self._get_meta(attribute_name, "argparse_const") + + def _get_argparse_alias(self, attribute_name: str) -> Any: + return self._get_meta(attribute_name, "argparse_alias") + + def _get_choices(self, attribute_name: str) -> Any: + return self._get_meta(attribute_name, "choices") + + +@dataclass +class CommonConfig(FairseqDataclass): + # This is the core dataclass including common parameters shared by all different jobs. Please append your params to other dataclasses if they were + # used for a particular purpose or task, such as those dedicated for `distributed training`, `optimization`, etc. + no_progress_bar: bool = field( + default=False, metadata={"help": "disable progress bar"} + ) + log_interval: int = field( + default=100, + metadata={ + "help": "log progress every N batches (when progress bar is disabled)" + }, + ) + log_format: Optional[LOG_FORMAT_CHOICES] = field( + default=None, metadata={"help": "log format to use"} + ) + log_file: Optional[str] = field( + default=None, metadata={"help": "log file to copy metrics to."} + ) + tensorboard_logdir: Optional[str] = field( + default=None, + metadata={ + "help": "path to save logs for tensorboard, should match --logdir " + "of running tensorboard (default: no tensorboard logging)" + }, + ) + wandb_project: Optional[str] = field( + default=None, + metadata={"help": "Weights and Biases project name to use for logging"}, + ) + azureml_logging: Optional[bool] = field( + default=False, metadata={"help": "Log scalars to AzureML context"}, + ) + seed: int = field( + default=1, metadata={"help": "pseudo random number generator seed"} + ) + cpu: bool = field(default=False, metadata={"help": "use CPU instead of CUDA"}) + tpu: bool = field(default=False, metadata={"help": "use TPU instead of CUDA"}) + bf16: bool = field(default=False, metadata={"help": "use bfloat16; implies --tpu"}) + memory_efficient_bf16: bool = field( + default=False, + metadata={ + "help": "use a memory-efficient version of BF16 training; implies --bf16" + }, + ) + fp16: bool = field(default=False, metadata={"help": "use FP16"}) + memory_efficient_fp16: bool = field( + default=False, + metadata={ + "help": "use a memory-efficient version of FP16 training; implies --fp16" + }, + ) + fp16_no_flatten_grads: bool = field( + default=False, metadata={"help": "don't flatten FP16 grads tensor"} + ) + fp16_init_scale: int = field( + default=2 ** 7, metadata={"help": "default FP16 loss scale"} + ) + fp16_scale_window: Optional[int] = field( + default=None, + metadata={"help": "number of updates before increasing loss scale"}, + ) + fp16_scale_tolerance: float = field( + default=0.0, + metadata={ + "help": "pct of updates that can overflow before decreasing the loss scale" + }, + ) + on_cpu_convert_precision: bool = field( + default=False, + metadata={ + "help": "if set, the floating point conversion to fp16/bf16 runs on CPU. " + "This reduces bus transfer time and GPU memory usage." + } + ) + min_loss_scale: float = field( + default=1e-4, + metadata={"help": "minimum FP16/AMP loss scale, after which training is stopped"}, + ) + threshold_loss_scale: Optional[float] = field( + default=None, metadata={"help": "threshold FP16 loss scale from below"} + ) + amp: bool = field(default=False, metadata={"help": "use automatic mixed precision"}) + amp_batch_retries: int = field( + default=2, + metadata={"help": "number of retries of same batch after reducing loss scale with AMP"}, + ) + amp_init_scale: int = field( + default=2 ** 7, metadata={"help": "default AMP loss scale"} + ) + amp_scale_window: Optional[int] = field( + default=None, + metadata={"help": "number of updates before increasing AMP loss scale"}, + ) + user_dir: Optional[str] = field( + default=None, + metadata={ + "help": "path to a python module containing custom extensions (tasks and/or architectures)" + }, + ) + empty_cache_freq: int = field( + default=0, + metadata={"help": "how often to clear the PyTorch CUDA cache (0 to disable)"}, + ) + all_gather_list_size: int = field( + default=16384, + metadata={"help": "number of bytes reserved for gathering stats from workers"}, + ) + model_parallel_size: int = field( + default=1, metadata={"help": "total number of GPUs to parallelize model over"} + ) + quantization_config_path: Optional[str] = field( + default=None, metadata={"help": "path to quantization config file"} + ) + profile: bool = field( + default=False, metadata={"help": "enable autograd profiler emit_nvtx"} + ) + reset_logging: bool = field( + default=False, + metadata={ + "help": "when using Hydra, reset the logging at the beginning of training" + }, + ) + suppress_crashes: bool = field( + default=False, + metadata={ + "help": "suppress crashes when training with the hydra_train entry point so that the " + "main method can return a value (useful for sweeps)" + }, + ) + use_plasma_view: bool = field( + default=False, metadata={"help": "Store indices and sizes in shared memory"} + ) + plasma_path: Optional[str] = field( + default="/tmp/plasma", + metadata={ + "help": "path to run plasma_store, defaults to /tmp/plasma. Paths outside /tmp tend to fail." + }, + ) + + +@dataclass +class DistributedTrainingConfig(FairseqDataclass): + distributed_world_size: int = field( + default=max(1, torch.cuda.device_count()), + metadata={ + "help": "total number of GPUs across all nodes (default: all visible GPUs)" + }, + ) + distributed_num_procs: Optional[int] = field( + default=max(1, torch.cuda.device_count()), + metadata={ + "help": "total number of processes to fork (default: all visible GPUs)" + }, + ) + distributed_rank: Optional[int] = field( + default=0, metadata={"help": "rank of the current worker"} + ) + distributed_backend: str = field( + default="nccl", metadata={"help": "distributed backend"} + ) + distributed_init_method: Optional[str] = field( + default=None, + metadata={ + "help": "typically tcp://hostname:port that will be used to " + "establish initial connetion" + }, + ) + distributed_port: int = field( + default=-1, + metadata={ + "help": "port number (not required if using --distributed-init-method)" + }, + ) + device_id: int = field( + default=0, + metadata={ + "help": "which GPU to use (usually configured automatically)", + "argparse_alias": "--local_rank", + }, + ) + distributed_no_spawn: bool = field( + default=False, + metadata={ + "help": "do not spawn multiple processes even if multiple GPUs are visible" + }, + ) + ddp_backend: DDP_BACKEND_CHOICES = field( + default="pytorch_ddp", metadata={"help": "DistributedDataParallel backend"} + ) + ddp_comm_hook: DDP_COMM_HOOK_CHOICES = field( + default="none", metadata={"help": "communication hook"} + ) + bucket_cap_mb: int = field( + default=25, metadata={"help": "bucket size for reduction"} + ) + fix_batches_to_gpus: bool = field( + default=False, + metadata={ + "help": "don't shuffle batches between GPUs; this reduces overall " + "randomness and may affect precision but avoids the cost of re-reading the data" + }, + ) + find_unused_parameters: bool = field( + default=False, + metadata={ + "help": "disable unused parameter detection (not applicable to " + "--ddp-backend=legacy_ddp)" + }, + ) + fast_stat_sync: bool = field( + default=False, + metadata={"help": "[deprecated] this is now defined per Criterion"}, + ) + heartbeat_timeout: int = field( + default=-1, + metadata={ + "help": "kill the job if no progress is made in N seconds; " + "set to -1 to disable" + }, + ) + broadcast_buffers: bool = field( + default=False, + metadata={ + "help": "Copy non-trainable parameters between GPUs, such as " + "batchnorm population statistics" + }, + ) + slowmo_momentum: Optional[float] = field( + default=None, + metadata={ + "help": "SlowMo momentum term; by default use 0.0 for 16 GPUs, " + "0.2 for 32 GPUs; 0.5 for 64 GPUs, 0.6 for > 64 GPUs" + }, + ) + slowmo_algorithm: str = field( + default="LocalSGD", metadata={"help": "whether to use LocalSGD or SGP"} + ) + localsgd_frequency: int = field( + default=3, metadata={"help": "Local SGD allreduce frequency"} + ) + nprocs_per_node: int = field( + default=max(1, torch.cuda.device_count()), + metadata={ + "help": "number of GPUs in each node. An allreduce operation across GPUs in " + "a node is very fast. Hence, we do allreduce across GPUs in a node, " + "and gossip across different nodes" + }, + ) + pipeline_model_parallel: bool = field( + default=False, + metadata={"help": "if set, use pipeline model parallelism across GPUs"}, + ) + pipeline_balance: Optional[str] = field( + default=None, + metadata={ + "help": "partition the model into N_K pieces, where each piece " + "contains N_i layers. The sum(args.pipeline_balance) " + "should equal the total number of layers in the model" + }, + ) + pipeline_devices: Optional[str] = field( + default=None, + metadata={ + "help": "a list of device indices indicating which device to place " + "each of the N_K partitions. The length of this list should " + "equal the length of the --pipeline-balance argument" + }, + ) + pipeline_chunks: Optional[int] = field( + default=0, metadata={"help": "microbatch count for pipeline model parallelism"} + ) + pipeline_encoder_balance: Optional[str] = field( + default=None, + metadata={ + "help": "partition the pipeline parallel encoder into N_K pieces, where each piece " + "contains N_i layers. The sum(args.pipeline_encoder_balance) " + "should equal the total number of encoder layers in the model" + }, + ) + pipeline_encoder_devices: Optional[str] = field( + default=None, + metadata={ + "help": "a list of device indices indicating which device to place " + "each of the N_K partitions. The length of this list should " + "equal the length of the --pipeline-encoder-balance argument" + }, + ) + pipeline_decoder_balance: Optional[str] = field( + default=None, + metadata={ + "help": "partition the pipeline parallel decoder into N_K pieces, where each piece " + "contains N_i layers. The sum(args.pipeline_decoder_balance) " + "should equal the total number of decoder layers in the model" + }, + ) + pipeline_decoder_devices: Optional[str] = field( + default=None, + metadata={ + "help": "a list of device indices indicating which device to place " + "each of the N_K partitions. The length of this list should " + "equal the length of the --pipeline-decoder-balance argument" + }, + ) + pipeline_checkpoint: PIPELINE_CHECKPOINT_CHOICES = field( + default="never", + metadata={"help": "checkpointing mode for pipeline model parallelism"}, + ) + zero_sharding: ZERO_SHARDING_CHOICES = field( + default="none", metadata={"help": "ZeRO sharding"} + ) + fp16: bool = II("common.fp16") + memory_efficient_fp16: bool = II("common.memory_efficient_fp16") + tpu: bool = II("common.tpu") + # configuration for --ddp-backend=fully_sharded + no_reshard_after_forward: bool = field( + default=False, metadata={"help": "don't reshard parameters after forward pass"}, + ) + fp32_reduce_scatter: bool = field( + default=False, metadata={"help": "reduce-scatter grads in FP32"}, + ) + cpu_offload: bool = field( + default=False, metadata={"help": "offload FP32 params to CPU"} + ) + use_sharded_state: bool = field( + default=False, metadata={"help": "use sharded checkpoint files"}, + ) + + +@dataclass +class DatasetConfig(FairseqDataclass): + num_workers: int = field( + default=1, metadata={"help": "how many subprocesses to use for data loading"} + ) + skip_invalid_size_inputs_valid_test: bool = field( + default=False, + metadata={"help": "ignore too long or too short lines in valid and test set"}, + ) + max_tokens: Optional[int] = field( + default=None, metadata={"help": "maximum number of tokens in a batch"} + ) + batch_size: Optional[int] = field( + default=None, + metadata={ + "help": "number of examples in a batch", + "argparse_alias": "--max-sentences", + }, + ) + required_batch_size_multiple: int = field( + default=8, metadata={"help": "batch size will be a multiplier of this value"} + ) + required_seq_len_multiple: int = field( + default=1, + metadata={ + "help": "maximum sequence length in batch will be a multiplier of this value" + }, + ) + dataset_impl: Optional[DATASET_IMPL_CHOICES] = field( + default=None, metadata={"help": "output dataset implementation"} + ) + data_buffer_size: int = field( + default=10, metadata={"help": "Number of batches to preload"} + ) + train_subset: str = field( + default="train", + metadata={"help": "data subset to use for training (e.g. train, valid, test)"}, + ) + valid_subset: str = field( + default="valid", + metadata={ + "help": "comma separated list of data subsets to use for validation" + " (e.g. train, valid, test)" + }, + ) + combine_valid_subsets: Optional[bool] = field( + default=None, + metadata={ + "help": "comma separated list of data subsets to use for validation" + " (e.g. train, valid, test)", + "argparse_alias": "--combine-val", + }, + ) + ignore_unused_valid_subsets: Optional[bool] = field( + default=False, + metadata={"help": "do not raise error if valid subsets are ignored"}, + ) + + validate_interval: int = field( + default=1, metadata={"help": "validate every N epochs"} + ) + validate_interval_updates: int = field( + default=0, metadata={"help": "validate every N updates"} + ) + validate_after_updates: int = field( + default=0, metadata={"help": "dont validate until reaching this many updates"} + ) + fixed_validation_seed: Optional[int] = field( + default=None, metadata={"help": "specified random seed for validation"} + ) + disable_validation: bool = field( + default=False, metadata={"help": "disable validation"} + ) + max_tokens_valid: Optional[int] = field( + default=II("dataset.max_tokens"), + metadata={ + "help": "maximum number of tokens in a validation batch" + " (defaults to --max-tokens)" + }, + ) + batch_size_valid: Optional[int] = field( + default=II("dataset.batch_size"), + metadata={ + "help": "batch size of the validation batch (defaults to --batch-size)", + "argparse_alias": "--max-sentences-valid", + }, + ) + max_valid_steps: Optional[int] = field(default=None, metadata={'help': 'How many batches to evaluate', + "argparse_alias": "--nval"}) + curriculum: int = field( + default=0, metadata={"help": "don't shuffle batches for first N epochs"} + ) + gen_subset: str = field( + default="test", + metadata={"help": "data subset to generate (train, valid, test)"}, + ) + num_shards: int = field( + default=1, metadata={"help": "shard generation over N shards"} + ) + shard_id: int = field( + default=0, metadata={"help": "id of the shard to generate (id < num_shards)"} + ) + + +@dataclass +class OptimizationConfig(FairseqDataclass): + max_epoch: int = field( + default=0, metadata={"help": "force stop training at specified epoch"} + ) + max_update: int = field( + default=0, metadata={"help": "force stop training at specified update"} + ) + stop_time_hours: float = field( + default=0, + metadata={ + "help": "force stop training after specified cumulative time (if >0)" + }, + ) + clip_norm: float = field( + default=0.0, metadata={"help": "clip threshold of gradients"} + ) + sentence_avg: bool = field( + default=False, + metadata={ + "help": "normalize gradients by the number of sentences in a batch" + " (default is to normalize by number of tokens)" + }, + ) + update_freq: List[int] = field( + default_factory=lambda: [1], + metadata={"help": "update parameters every N_i batches, when in epoch i"}, + ) + lr: List[float] = field( + default_factory=lambda: [0.25], + metadata={ + "help": "learning rate for the first N epochs; all epochs >N using LR_N" + " (note: this may be interpreted differently depending on --lr-scheduler)" + }, + ) + stop_min_lr: float = field( + default=-1.0, + metadata={"help": "stop training when the learning rate reaches this minimum"}, + ) + use_bmuf: bool = field( + default=False, + metadata={ + "help": "specify global optimizer for syncing models on different GPUs/shards" + }, + ) + + +@dataclass +class CheckpointConfig(FairseqDataclass): + save_dir: str = field( + default="checkpoints", metadata={"help": "path to save checkpoints"} + ) + restore_file: str = field( + default="checkpoint_last.pt", + metadata={ + "help": "filename from which to load checkpoint " + "(default: <save-dir>/checkpoint_last.pt" + }, + ) + finetune_from_model: Optional[str] = field( + default=None, + metadata={ + "help": "finetune from a pretrained model; note that meters and lr scheduler will be reset" + }, + ) + reset_dataloader: bool = field( + default=False, + metadata={ + "help": "if set, does not reload dataloader state from the checkpoint" + }, + ) + reset_lr_scheduler: bool = field( + default=False, + metadata={ + "help": "if set, does not load lr scheduler state from the checkpoint" + }, + ) + reset_meters: bool = field( + default=False, + metadata={"help": "if set, does not load meters from the checkpoint"}, + ) + reset_optimizer: bool = field( + default=False, + metadata={"help": "if set, does not load optimizer state from the checkpoint"}, + ) + optimizer_overrides: str = field( + default="{}", + metadata={ + "help": "a dictionary used to override optimizer args when loading a checkpoint" + }, + ) + save_interval: int = field( + default=1, metadata={"help": "save a checkpoint every N epochs"} + ) + save_interval_updates: int = field( + default=0, metadata={"help": "save a checkpoint (and validate) every N updates"} + ) + keep_interval_updates: int = field( + default=-1, + metadata={ + "help": "keep the last N checkpoints saved with --save-interval-updates" + }, + ) + keep_interval_updates_pattern: int = field( + default=-1, + metadata={ + "help": "when used with --keep-interval-updates, skips deleting " + "any checkpoints with update X where " + "X %% keep_interval_updates_pattern == 0" + }, + ) + keep_last_epochs: int = field( + default=-1, metadata={"help": "keep last N epoch checkpoints"} + ) + keep_best_checkpoints: int = field( + default=-1, metadata={"help": "keep best N checkpoints based on scores"} + ) + no_save: bool = field( + default=False, metadata={"help": "don't save models or checkpoints"} + ) + no_epoch_checkpoints: bool = field( + default=False, metadata={"help": "only store last and best checkpoints"} + ) + no_last_checkpoints: bool = field( + default=False, metadata={"help": "don't store last checkpoints"} + ) + no_save_optimizer_state: bool = field( + default=False, + metadata={"help": "don't save optimizer-state as part of checkpoint"}, + ) + best_checkpoint_metric: str = field( + default="loss", metadata={"help": 'metric to use for saving "best" checkpoints'} + ) + maximize_best_checkpoint_metric: bool = field( + default=False, + metadata={ + "help": 'select the largest metric value for saving "best" checkpoints' + }, + ) + patience: int = field( + default=-1, + metadata={ + "help": ( + "early stop training if valid performance doesn't " + "improve for N consecutive validation runs; note " + "that this is influenced by --validate-interval" + ) + }, + ) + checkpoint_suffix: str = field( + default="", metadata={"help": "suffix to add to the checkpoint file name"} + ) + checkpoint_shard_count: int = field( + default=1, + metadata={ + "help": "Number of shards containing the checkpoint - " + "if the checkpoint is over 300GB, it is preferable " + "to split it into shards to prevent OOM on CPU while loading " + "the checkpoint" + }, + ) + load_checkpoint_on_all_dp_ranks: bool = field( + default=False, + metadata={ + "help": "load checkpoints on all data parallel devices " + "(default: only load on rank 0 and broadcast to other devices)" + }, + ) + write_checkpoints_asynchronously: bool = field( + default=False, + metadata={ + "help": ( + "Write checkpoints asynchronously in a separate " + "thread. NOTE: This feature is currently being tested." + ), + "argparse_alias": "--save-async", + }, + ) + model_parallel_size: int = II("common.model_parallel_size") + + +@dataclass +class FairseqBMUFConfig(FairseqDataclass): + block_lr: float = field( + default=1, metadata={"help": "block learning rate for bmuf"} + ) + block_momentum: float = field( + default=0.875, metadata={"help": "block momentum for bmuf"} + ) + global_sync_iter: int = field( + default=50, metadata={"help": "Iteration for syncing global model"} + ) + warmup_iterations: int = field( + default=500, metadata={"help": "warmup iterations for model to broadcast"} + ) + use_nbm: bool = field( + default=False, + metadata={"help": "Specify whether you want to use classical BM / Nesterov BM"}, + ) + average_sync: bool = field( + default=False, + metadata={ + "help": "Specify whether you want to average the local momentum after each sync" + }, + ) + distributed_world_size: int = II("distributed_training.distributed_world_size") + + +@dataclass +class GenerationConfig(FairseqDataclass): + beam: int = field( + default=5, metadata={"help": "beam size"}, + ) + nbest: int = field( + default=1, metadata={"help": "number of hypotheses to output"}, + ) + max_len_a: float = field( + default=0, + metadata={ + "help": "generate sequences of maximum length ax + b, where x is the source length" + }, + ) + max_len_b: int = field( + default=200, + metadata={ + "help": "generate sequences of maximum length ax + b, where x is the source length" + }, + ) + min_len: int = field( + default=1, metadata={"help": "minimum generation length"}, + ) + match_source_len: bool = field( + default=False, metadata={"help": "generations should match the source length"}, + ) + unnormalized: bool = field( + default=False, metadata={"help": "compare unnormalized hypothesis scores"}, + ) + no_early_stop: bool = field( + default=False, metadata={"help": "deprecated"}, + ) + no_beamable_mm: bool = field( + default=False, metadata={"help": "don't use BeamableMM in attention layers"}, + ) + lenpen: float = field( + default=1, + metadata={ + "help": "length penalty: <1.0 favors shorter, >1.0 favors longer sentences" + }, + ) + unkpen: float = field( + default=0, + metadata={ + "help": "unknown word penalty: <0 produces more unks, >0 produces fewer" + }, + ) + replace_unk: Optional[str] = field( + default=None, + metadata={ + "help": "perform unknown replacement (optionally with alignment dictionary)", + "argparse_const": "@@ ", + }, + ) + sacrebleu: bool = field( + default=False, metadata={"help": "score with sacrebleu"}, + ) + score_reference: bool = field( + default=False, metadata={"help": "just score the reference translation"}, + ) + prefix_size: int = field( + default=0, + metadata={"help": "initialize generation by target prefix of given length"}, + ) + no_repeat_ngram_size: int = field( + default=0, + metadata={ + "help": "ngram blocking such that this size ngram cannot be repeated in the generation" + }, + ) + sampling: bool = field( + default=False, + metadata={"help": "sample hypotheses instead of using beam search"}, + ) + sampling_topk: int = field( + default=-1, + metadata={"help": "sample from top K likely next words instead of all words"}, + ) + sampling_topp: float = field( + default=-1.0, + metadata={ + "help": "sample from the smallest set whose cumulative probability mass exceeds p for next words" + }, + ) + constraints: Optional[GENERATION_CONSTRAINTS_CHOICES] = field( + default=None, + metadata={ + "help": "enables lexically constrained decoding", + "argparse_const": "ordered", + }, + ) + temperature: float = field( + default=1.0, metadata={"help": "temperature for generation"}, + ) + diverse_beam_groups: int = field( + default=-1, metadata={"help": "number of groups for Diverse Beam Search"}, + ) + diverse_beam_strength: float = field( + default=0.5, + metadata={"help": "strength of diversity penalty for Diverse Beam Search"}, + ) + diversity_rate: float = field( + default=-1.0, + metadata={"help": "strength of diversity penalty for Diverse Siblings Search"}, + ) + print_alignment: Optional[PRINT_ALIGNMENT_CHOICES] = field( + default=None, + metadata={ + "help": "if set, uses attention feedback to compute and print alignment to source tokens " + "(valid options are: hard, soft, otherwise treated as hard alignment)", + "argparse_const": "hard", + }, + ) + print_step: bool = field( + default=False, metadata={"help": "print steps"}, + ) + lm_path: Optional[str] = field( + default=None, metadata={"help": "path to lm checkpoint for lm fusion"}, + ) + lm_weight: float = field( + default=0.0, metadata={"help": "weight for lm probs for lm fusion"}, + ) + + # arguments for iterative refinement generator + iter_decode_eos_penalty: float = field( + default=0.0, + metadata={"help": "if > 0.0, it penalized early-stopping in decoding."}, + ) + iter_decode_max_iter: int = field( + default=10, metadata={"help": "maximum iterations for iterative refinement."}, + ) + iter_decode_force_max_iter: bool = field( + default=False, + metadata={ + "help": "if set, run exact the maximum number of iterations without early stop" + }, + ) + iter_decode_with_beam: int = field( + default=1, + metadata={ + "help": "if > 1, model will generate translations varying by the lengths." + }, + ) + iter_decode_with_external_reranker: bool = field( + default=False, + metadata={ + "help": "if set, the last checkpoint are assumed to be a reranker to rescore the translations" + }, + ) + retain_iter_history: bool = field( + default=False, + metadata={ + "help": "if set, decoding returns the whole history of iterative refinement" + }, + ) + retain_dropout: bool = field( + default=False, metadata={"help": "Use dropout at inference time"}, + ) + # temporarily set to Any until https://github.com/facebookresearch/hydra/issues/1117 is fixed + # retain_dropout_modules: Optional[List[str]] = field( + retain_dropout_modules: Any = field( + default=None, + metadata={ + "help": "if set, only retain dropout for the specified modules; " + "if not set, then dropout will be retained for all modules" + }, + ) + # special decoding format for advanced decoding. + decoding_format: Optional[GENERATION_DECODING_FORMAT_CHOICES] = field( + default=None, + metadata={"help": "special decoding format for advanced decoding."}, + ) + no_seed_provided: bool = field( + default=False, + metadata={"help": "if set, dont use seed for initializing random generators"}, + ) + + +@dataclass +class CommonEvalConfig(FairseqDataclass): + path: Optional[str] = field( + default=None, metadata={"help": "path(s) to model file(s), colon separated"}, + ) + post_process: Optional[str] = field( + default=None, + metadata={ + "help": ( + "post-process text by removing BPE, letter segmentation, etc. " + "Valid options can be found in fairseq.data.utils.post_process." + ), + "argparse_const": "subword_nmt", + "argparse_alias": "--remove-bpe", + }, + ) + quiet: bool = field(default=False, metadata={"help": "only print final scores"}) + model_overrides: str = field( + default="{}", + metadata={ + "help": "a dictionary used to override model args at generation that were used during model training" + }, + ) + results_path: Optional[str] = field( + default=None, metadata={"help": "path to save eval results (optional)"} + ) + + +@dataclass +class EvalLMConfig(FairseqDataclass): + output_word_probs: bool = field( + default=False, + metadata={ + "help": "if set, outputs words and their predicted log probabilities to standard output" + }, + ) + output_word_stats: bool = field( + default=False, + metadata={ + "help": "if set, outputs word statistics such as word count, average probability, etc" + }, + ) + context_window: int = field( + default=0, + metadata={ + "help": "ensures that every evaluated token has access to a context of at least this size, if possible" + }, + ) + softmax_batch: int = field( + default=sys.maxsize, + metadata={ + "help": "if BxT is more than this, will batch the softmax over vocab to this amount of tokens, in order to fit into GPU memory" + }, + ) + + +@dataclass +class InteractiveConfig(FairseqDataclass): + buffer_size: int = field( + default=0, + metadata={ + "help": "read this many sentences into a buffer before processing them" + }, + ) + input: str = field( + default="-", metadata={"help": "file to read from; use - for stdin"}, + ) + + +@dataclass +class FairseqConfig(FairseqDataclass): + common: CommonConfig = CommonConfig() + common_eval: CommonEvalConfig = CommonEvalConfig() + distributed_training: DistributedTrainingConfig = DistributedTrainingConfig() + dataset: DatasetConfig = DatasetConfig() + optimization: OptimizationConfig = OptimizationConfig() + checkpoint: CheckpointConfig = CheckpointConfig() + bmuf: FairseqBMUFConfig = FairseqBMUFConfig() + generation: GenerationConfig = GenerationConfig() + eval_lm: EvalLMConfig = EvalLMConfig() + interactive: InteractiveConfig = InteractiveConfig() + model: Any = MISSING + task: Any = None + criterion: Any = None + optimizer: Any = None + lr_scheduler: Any = None + scoring: Any = None + bpe: Any = None + tokenizer: Any = None diff --git a/fairseq/dataclass/constants.py b/fairseq/dataclass/constants.py new file mode 100644 index 0000000000000000000000000000000000000000..442c25982b55f680880147feb64b9c2e6756142c --- /dev/null +++ b/fairseq/dataclass/constants.py @@ -0,0 +1,54 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from enum import Enum, EnumMeta +from typing import List + + +class StrEnumMeta(EnumMeta): + # this is workaround for submitit pickling leading to instance checks failing in hydra for StrEnum, see + # https://github.com/facebookresearch/hydra/issues/1156 + @classmethod + def __instancecheck__(cls, other): + return "enum" in str(type(other)) + + +class StrEnum(Enum, metaclass=StrEnumMeta): + def __str__(self): + return self.value + + def __eq__(self, other: str): + return self.value == other + + def __repr__(self): + return self.value + + def __hash__(self): + return hash(str(self)) + + +def ChoiceEnum(choices: List[str]): + """return the Enum class used to enforce list of choices""" + return StrEnum("Choices", {k: k for k in choices}) + + +LOG_FORMAT_CHOICES = ChoiceEnum(["json", "none", "simple", "tqdm"]) +DDP_BACKEND_CHOICES = ChoiceEnum([ + "c10d", # alias for pytorch_ddp + "fully_sharded", # FullyShardedDataParallel from fairscale + "legacy_ddp", + "no_c10d", # alias for legacy_ddp + "pytorch_ddp", + "slow_mo", +]) +DDP_COMM_HOOK_CHOICES = ChoiceEnum(["none", "fp16"]) +DATASET_IMPL_CHOICES = ChoiceEnum(["raw", "lazy", "cached", "mmap", "fasta"]) +GENERATION_CONSTRAINTS_CHOICES = ChoiceEnum(["ordered", "unordered"]) +GENERATION_DECODING_FORMAT_CHOICES = ChoiceEnum( + ["unigram", "ensemble", "vote", "dp", "bs"] +) +ZERO_SHARDING_CHOICES = ChoiceEnum(["none", "os"]) +PIPELINE_CHECKPOINT_CHOICES = ChoiceEnum(["always", "never", "except_last"]) +PRINT_ALIGNMENT_CHOICES = ChoiceEnum(["hard", "soft"]) diff --git a/fairseq/dataclass/initialize.py b/fairseq/dataclass/initialize.py new file mode 100644 index 0000000000000000000000000000000000000000..479aeb8b16ad230c424da353a689fe3505b449e5 --- /dev/null +++ b/fairseq/dataclass/initialize.py @@ -0,0 +1,61 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +"""isort:skip_file""" + +import logging +from hydra.core.config_store import ConfigStore +from fairseq.dataclass.configs import FairseqConfig +from omegaconf import DictConfig, OmegaConf + + +logger = logging.getLogger(__name__) + + +def hydra_init(cfg_name="config") -> None: + + cs = ConfigStore.instance() + cs.store(name=cfg_name, node=FairseqConfig) + + for k in FairseqConfig.__dataclass_fields__: + v = FairseqConfig.__dataclass_fields__[k].default + try: + cs.store(name=k, node=v) + except BaseException: + logger.error(f"{k} - {v}") + raise + + +def add_defaults(cfg: DictConfig) -> None: + """This function adds default values that are stored in dataclasses that hydra doesn't know about """ + + from fairseq.registry import REGISTRIES + from fairseq.tasks import TASK_DATACLASS_REGISTRY + from fairseq.models import ARCH_MODEL_NAME_REGISTRY, MODEL_DATACLASS_REGISTRY + from fairseq.dataclass.utils import merge_with_parent + from typing import Any + + OmegaConf.set_struct(cfg, False) + + for k, v in FairseqConfig.__dataclass_fields__.items(): + field_cfg = cfg.get(k) + if field_cfg is not None and v.type == Any: + dc = None + + if isinstance(field_cfg, str): + field_cfg = DictConfig({"_name": field_cfg}) + field_cfg.__dict__["_parent"] = field_cfg.__dict__["_parent"] + + name = field_cfg.get("_name") + + if k == "task": + dc = TASK_DATACLASS_REGISTRY.get(name) + elif k == "model": + name = ARCH_MODEL_NAME_REGISTRY.get(name, name) + dc = MODEL_DATACLASS_REGISTRY.get(name) + elif k in REGISTRIES: + dc = REGISTRIES[k]["dataclass_registry"].get(name) + + if dc is not None: + cfg[k] = merge_with_parent(dc, field_cfg) diff --git a/fairseq/dataclass/utils.py b/fairseq/dataclass/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..89206125d1d50ccc4b4d56394a76bc07bb32927a --- /dev/null +++ b/fairseq/dataclass/utils.py @@ -0,0 +1,476 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import ast +import inspect +import logging +import os +import re +from argparse import ArgumentError, ArgumentParser, Namespace +from dataclasses import _MISSING_TYPE, MISSING, is_dataclass +from enum import Enum +from typing import Any, Dict, List, Optional, Tuple, Type + +from fairseq.dataclass import FairseqDataclass +from fairseq.dataclass.configs import FairseqConfig +from hydra.core.global_hydra import GlobalHydra +from hydra.experimental import compose, initialize +from omegaconf import DictConfig, OmegaConf, open_dict + +logger = logging.getLogger(__name__) + + +def eval_str_list(x, x_type=float): + if x is None: + return None + if isinstance(x, str): + if len(x) == 0: + return [] + x = ast.literal_eval(x) + try: + return list(map(x_type, x)) + except TypeError: + return [x_type(x)] + + +def interpret_dc_type(field_type): + if isinstance(field_type, str): + raise RuntimeError("field should be a type") + + if field_type == Any: + return str + + typestring = str(field_type) + if re.match( + r"(typing.|^)Union\[(.*), NoneType\]$", typestring + ) or typestring.startswith("typing.Optional"): + return field_type.__args__[0] + return field_type + + +def gen_parser_from_dataclass( + parser: ArgumentParser, + dataclass_instance: FairseqDataclass, + delete_default: bool = False, +) -> None: + """convert a dataclass instance to tailing parser arguments""" + + def argparse_name(name: str): + if name == "data": + # normally data is positional args + return name + if name == "_name": + # private member, skip + return None + return "--" + name.replace("_", "-") + + def get_kwargs_from_dc( + dataclass_instance: FairseqDataclass, k: str + ) -> Dict[str, Any]: + """k: dataclass attributes""" + + kwargs = {} + + field_type = dataclass_instance._get_type(k) + inter_type = interpret_dc_type(field_type) + + field_default = dataclass_instance._get_default(k) + + if isinstance(inter_type, type) and issubclass(inter_type, Enum): + field_choices = [t.value for t in list(inter_type)] + else: + field_choices = None + + field_help = dataclass_instance._get_help(k) + field_const = dataclass_instance._get_argparse_const(k) + + if isinstance(field_default, str) and field_default.startswith("${"): + kwargs["default"] = field_default + else: + if field_default is MISSING: + kwargs["required"] = True + if field_choices is not None: + kwargs["choices"] = field_choices + if ( + isinstance(inter_type, type) + and (issubclass(inter_type, List) or issubclass(inter_type, Tuple)) + ) or ("List" in str(inter_type) or "Tuple" in str(inter_type)): + if "int" in str(inter_type): + kwargs["type"] = lambda x: eval_str_list(x, int) + elif "float" in str(inter_type): + kwargs["type"] = lambda x: eval_str_list(x, float) + elif "str" in str(inter_type): + kwargs["type"] = lambda x: eval_str_list(x, str) + else: + raise NotImplementedError( + "parsing of type " + str(inter_type) + " is not implemented" + ) + if field_default is not MISSING: + kwargs["default"] = ( + ",".join(map(str, field_default)) + if field_default is not None + else None + ) + elif ( + isinstance(inter_type, type) and issubclass(inter_type, Enum) + ) or "Enum" in str(inter_type): + kwargs["type"] = str + if field_default is not MISSING: + if isinstance(field_default, Enum): + kwargs["default"] = field_default.value + else: + kwargs["default"] = field_default + elif inter_type is bool: + kwargs["action"] = ( + "store_false" if field_default is True else "store_true" + ) + kwargs["default"] = field_default + else: + kwargs["type"] = inter_type + if field_default is not MISSING: + kwargs["default"] = field_default + + kwargs["help"] = field_help + if field_const is not None: + kwargs["const"] = field_const + kwargs["nargs"] = "?" + + return kwargs + + for k in dataclass_instance._get_all_attributes(): + field_name = argparse_name(dataclass_instance._get_name(k)) + field_type = dataclass_instance._get_type(k) + if field_name is None: + continue + elif inspect.isclass(field_type) and issubclass(field_type, FairseqDataclass): + gen_parser_from_dataclass(parser, field_type(), delete_default) + continue + + kwargs = get_kwargs_from_dc(dataclass_instance, k) + + field_args = [field_name] + alias = dataclass_instance._get_argparse_alias(k) + if alias is not None: + field_args.append(alias) + + if "default" in kwargs: + if isinstance(kwargs["default"], str) and kwargs["default"].startswith( + "${" + ): + if kwargs["help"] is None: + # this is a field with a name that will be added elsewhere + continue + else: + del kwargs["default"] + if delete_default and "default" in kwargs: + del kwargs["default"] + try: + parser.add_argument(*field_args, **kwargs) + except ArgumentError: + pass + + +def _set_legacy_defaults(args, cls): + """Helper to set default arguments based on *add_args*.""" + if not hasattr(cls, "add_args"): + return + + import argparse + + parser = argparse.ArgumentParser( + argument_default=argparse.SUPPRESS, allow_abbrev=False + ) + cls.add_args(parser) + # copied from argparse.py: + defaults = argparse.Namespace() + for action in parser._actions: + if action.dest is not argparse.SUPPRESS: + if not hasattr(defaults, action.dest): + if action.default is not argparse.SUPPRESS: + setattr(defaults, action.dest, action.default) + for key, default_value in vars(defaults).items(): + if not hasattr(args, key): + setattr(args, key, default_value) + + +def _override_attr( + sub_node: str, data_class: Type[FairseqDataclass], args: Namespace +) -> List[str]: + overrides = [] + + if not inspect.isclass(data_class) or not issubclass(data_class, FairseqDataclass): + return overrides + + def get_default(f): + if not isinstance(f.default_factory, _MISSING_TYPE): + return f.default_factory() + return f.default + + for k, v in data_class.__dataclass_fields__.items(): + if k.startswith("_"): + # private member, skip + continue + + val = get_default(v) if not hasattr(args, k) else getattr(args, k) + + field_type = interpret_dc_type(v.type) + if ( + isinstance(val, str) + and not val.startswith("${") # not interpolation + and field_type != str + and ( + not inspect.isclass(field_type) or not issubclass(field_type, Enum) + ) # not choices enum + ): + # upgrade old models that stored complex parameters as string + val = ast.literal_eval(val) + + if isinstance(val, tuple): + val = list(val) + + v_type = getattr(v.type, "__origin__", None) + if ( + (v_type is List or v_type is list or v_type is Optional) + # skip interpolation + and not (isinstance(val, str) and val.startswith("${")) + ): + # if type is int but val is float, then we will crash later - try to convert here + if hasattr(v.type, "__args__"): + t_args = v.type.__args__ + if len(t_args) == 1 and (t_args[0] is float or t_args[0] is int): + val = list(map(t_args[0], val)) + elif val is not None and ( + field_type is int or field_type is bool or field_type is float + ): + try: + val = field_type(val) + except: + pass # ignore errors here, they are often from interpolation args + + if val is None: + overrides.append("{}.{}=null".format(sub_node, k)) + elif val == "": + overrides.append("{}.{}=''".format(sub_node, k)) + elif isinstance(val, str): + val = val.replace("'", r"\'") + overrides.append("{}.{}='{}'".format(sub_node, k, val)) + elif isinstance(val, FairseqDataclass): + overrides += _override_attr(f"{sub_node}.{k}", type(val), args) + elif isinstance(val, Namespace): + sub_overrides, _ = override_module_args(val) + for so in sub_overrides: + overrides.append(f"{sub_node}.{k}.{so}") + else: + overrides.append("{}.{}={}".format(sub_node, k, val)) + + return overrides + + +def migrate_registry( + name, value, registry, args, overrides, deletes, use_name_as_val=False +): + if value in registry: + overrides.append("{}={}".format(name, value)) + overrides.append("{}._name={}".format(name, value)) + overrides.extend(_override_attr(name, registry[value], args)) + elif use_name_as_val and value is not None: + overrides.append("{}={}".format(name, value)) + else: + deletes.append(name) + + +def override_module_args(args: Namespace) -> Tuple[List[str], List[str]]: + """use the field in args to overrides those in cfg""" + overrides = [] + deletes = [] + + for k in FairseqConfig.__dataclass_fields__.keys(): + overrides.extend( + _override_attr(k, FairseqConfig.__dataclass_fields__[k].type, args) + ) + + if args is not None: + if hasattr(args, "task"): + from fairseq.tasks import TASK_DATACLASS_REGISTRY + + migrate_registry( + "task", args.task, TASK_DATACLASS_REGISTRY, args, overrides, deletes + ) + else: + deletes.append("task") + + # these options will be set to "None" if they have not yet been migrated + # so we can populate them with the entire flat args + CORE_REGISTRIES = {"criterion", "optimizer", "lr_scheduler"} + + from fairseq.registry import REGISTRIES + + for k, v in REGISTRIES.items(): + if hasattr(args, k): + migrate_registry( + k, + getattr(args, k), + v["dataclass_registry"], + args, + overrides, + deletes, + use_name_as_val=k not in CORE_REGISTRIES, + ) + else: + deletes.append(k) + + no_dc = True + if hasattr(args, "arch"): + from fairseq.models import ARCH_MODEL_REGISTRY, ARCH_MODEL_NAME_REGISTRY + + if args.arch in ARCH_MODEL_REGISTRY: + m_cls = ARCH_MODEL_REGISTRY[args.arch] + dc = getattr(m_cls, "__dataclass", None) + if dc is not None: + m_name = ARCH_MODEL_NAME_REGISTRY[args.arch] + overrides.append("model={}".format(m_name)) + overrides.append("model._name={}".format(args.arch)) + # override model params with those exist in args + overrides.extend(_override_attr("model", dc, args)) + no_dc = False + if no_dc: + deletes.append("model") + + return overrides, deletes + + +def convert_namespace_to_omegaconf(args: Namespace) -> DictConfig: + """Convert a flat argparse.Namespace to a structured DictConfig.""" + + # Here we are using field values provided in args to override counterparts inside config object + overrides, deletes = override_module_args(args) + + # configs will be in fairseq/config after installation + config_path = os.path.join("..", "config") + + GlobalHydra.instance().clear() + + with initialize(config_path=config_path): + try: + composed_cfg = compose("config", overrides=overrides, strict=False) + except: + logger.error("Error when composing. Overrides: " + str(overrides)) + raise + + for k in deletes: + composed_cfg[k] = None + + cfg = OmegaConf.create( + OmegaConf.to_container(composed_cfg, resolve=True, enum_to_str=True) + ) + + # hack to be able to set Namespace in dict config. this should be removed when we update to newer + # omegaconf version that supports object flags, or when we migrate all existing models + from omegaconf import _utils + + old_primitive = _utils.is_primitive_type + _utils.is_primitive_type = lambda _: True + + if cfg.task is None and getattr(args, "task", None): + cfg.task = Namespace(**vars(args)) + from fairseq.tasks import TASK_REGISTRY + + _set_legacy_defaults(cfg.task, TASK_REGISTRY[args.task]) + cfg.task._name = args.task + if cfg.model is None and getattr(args, "arch", None): + cfg.model = Namespace(**vars(args)) + from fairseq.models import ARCH_MODEL_REGISTRY + + _set_legacy_defaults(cfg.model, ARCH_MODEL_REGISTRY[args.arch]) + cfg.model._name = args.arch + if cfg.optimizer is None and getattr(args, "optimizer", None): + cfg.optimizer = Namespace(**vars(args)) + from fairseq.optim import OPTIMIZER_REGISTRY + + _set_legacy_defaults(cfg.optimizer, OPTIMIZER_REGISTRY[args.optimizer]) + cfg.optimizer._name = args.optimizer + if cfg.lr_scheduler is None and getattr(args, "lr_scheduler", None): + cfg.lr_scheduler = Namespace(**vars(args)) + from fairseq.optim.lr_scheduler import LR_SCHEDULER_REGISTRY + + _set_legacy_defaults(cfg.lr_scheduler, LR_SCHEDULER_REGISTRY[args.lr_scheduler]) + cfg.lr_scheduler._name = args.lr_scheduler + if cfg.criterion is None and getattr(args, "criterion", None): + cfg.criterion = Namespace(**vars(args)) + from fairseq.criterions import CRITERION_REGISTRY + + _set_legacy_defaults(cfg.criterion, CRITERION_REGISTRY[args.criterion]) + cfg.criterion._name = args.criterion + + _utils.is_primitive_type = old_primitive + OmegaConf.set_struct(cfg, True) + return cfg + + +def populate_dataclass( + dataclass: FairseqDataclass, + args: Namespace, +) -> FairseqDataclass: + for k in dataclass.__dataclass_fields__.keys(): + if k.startswith("_"): + # private member, skip + continue + if hasattr(args, k): + setattr(dataclass, k, getattr(args, k)) + + return dataclass + + +def overwrite_args_by_name(cfg: DictConfig, overrides: Dict[str, any]): + # this will be deprecated when we get rid of argparse and model_overrides logic + + from fairseq.registry import REGISTRIES + + with open_dict(cfg): + for k in cfg.keys(): + # "k in cfg" will return false if its a "mandatory value (e.g. ???)" + if k in cfg and isinstance(cfg[k], DictConfig): + if k in overrides and isinstance(overrides[k], dict): + for ok, ov in overrides[k].items(): + if isinstance(ov, dict) and cfg[k][ok] is not None: + overwrite_args_by_name(cfg[k][ok], ov) + else: + cfg[k][ok] = ov + else: + overwrite_args_by_name(cfg[k], overrides) + elif k in cfg and isinstance(cfg[k], Namespace): + for override_key, val in overrides.items(): + setattr(cfg[k], override_key, val) + elif k in overrides: + if ( + k in REGISTRIES + and overrides[k] in REGISTRIES[k]["dataclass_registry"] + ): + cfg[k] = DictConfig( + REGISTRIES[k]["dataclass_registry"][overrides[k]] + ) + overwrite_args_by_name(cfg[k], overrides) + cfg[k]._name = overrides[k] + else: + cfg[k] = overrides[k] + + +def merge_with_parent(dc: FairseqDataclass, cfg: DictConfig, remove_missing=True): + if remove_missing: + + if is_dataclass(dc): + target_keys = set(dc.__dataclass_fields__.keys()) + else: + target_keys = set(dc.keys()) + + with open_dict(cfg): + for k in list(cfg.keys()): + if k not in target_keys: + del cfg[k] + + merged_cfg = OmegaConf.merge(dc, cfg) + merged_cfg.__dict__["_parent"] = cfg.__dict__["_parent"] + OmegaConf.set_struct(merged_cfg, True) + return merged_cfg diff --git a/fairseq/distributed/__init__.py b/fairseq/distributed/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..d0b96b734c4b5e7cd5d295238d0764c05093dc27 --- /dev/null +++ b/fairseq/distributed/__init__.py @@ -0,0 +1,21 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .distributed_timeout_wrapper import DistributedTimeoutWrapper +from .fully_sharded_data_parallel import fsdp_enable_wrap, fsdp_wrap, FullyShardedDataParallel +from .legacy_distributed_data_parallel import LegacyDistributedDataParallel +from .module_proxy_wrapper import ModuleProxyWrapper +from .tpu_distributed_data_parallel import TPUDistributedDataParallel + + +__all__ = [ + "DistributedTimeoutWrapper", + "fsdp_enable_wrap", + "fsdp_wrap", + "FullyShardedDataParallel", + "LegacyDistributedDataParallel", + "ModuleProxyWrapper", + "TPUDistributedDataParallel", +] diff --git a/fairseq/distributed/distributed_timeout_wrapper.py b/fairseq/distributed/distributed_timeout_wrapper.py new file mode 100644 index 0000000000000000000000000000000000000000..18107ef27ea837b8c72dcaa49db18fd8e64267b1 --- /dev/null +++ b/fairseq/distributed/distributed_timeout_wrapper.py @@ -0,0 +1,94 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import os +import signal +import threading + +from torch import nn + + +logger = logging.getLogger(__name__) + + +class DistributedTimeoutWrapper(nn.Module): + """ + A wrapper that kills the process if no progress is made within a given + *timeout*. The timer is reset every time :func:`forward` is called. + + Usage:: + + module = DistributedTimeoutWrapper(module, timeout=30) + x = module(input) + time.sleep(20) # safe + x = module(input) + time.sleep(45) # job will be killed before this returns + + Args: + module (nn.Module): module to wrap + timeout (int): number of seconds before killing the process + (set to a value <= 0 to disable the timeout) + signal (Optional): signal to send once timeout is triggered + """ + def __init__(self, module: nn.Module, timeout: int, signal=signal.SIGINT): + super().__init__() + self.module = module + self.timeout = timeout + self.signal = signal + + if timeout > 0: + self._heartbeat = threading.Event() + self._heartbeat_thread = threading.Thread( + target=self._check_heartbeat, + args=(os.getpid(),), + daemon=True, + ) + self._heartbeat_thread.start() + self._terminated = False + else: + self._heartbeat = None + self._heartbeat_thread = None + + def __del__(self): + self.stop_timeout() + + def __getattr__(self, name): + """Forward missing attributes to wrapped module.""" + try: + return super().__getattr__(name) # defer to nn.Module's logic + except AttributeError: + return getattr(self.module, name) + + def stop_timeout(self): + if self._heartbeat_thread is not None: + self._terminated = True + self._heartbeat_thread.join() + + def state_dict(self, *args, **kwargs): + return self.module.state_dict(*args, **kwargs) + + def load_state_dict(self, *args, **kwargs): + return self.module.load_state_dict(*args, **kwargs) + + def forward(self, *args, **kwargs): + if self._heartbeat is not None: + self._heartbeat.set() + return self.module(*args, **kwargs) + + def _check_heartbeat(self, parent_pid): + self._heartbeat.wait() # wait for the first forward pass + while True: + self._heartbeat.clear() + success = self._heartbeat.wait(timeout=self.timeout) + if self._terminated: + break + elif not success: + logger.error(( + "Killing job for not making progress in {} seconds. " + "Set --heartbeat-timeout=-1 to disable this timeout." + ).format(int(self.timeout))) + os.kill(parent_pid, self.signal) + return diff --git a/fairseq/distributed/fully_sharded_data_parallel.py b/fairseq/distributed/fully_sharded_data_parallel.py new file mode 100644 index 0000000000000000000000000000000000000000..8a96bfc76516682ac8e2b7e2c3bc2e6aa3d8ef0c --- /dev/null +++ b/fairseq/distributed/fully_sharded_data_parallel.py @@ -0,0 +1,135 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import contextlib +from typing import Optional + +import torch +from fairseq.dataclass.configs import DistributedTrainingConfig +from fairseq.distributed import utils as dist_utils + + +try: + from fairscale.nn.data_parallel import FullyShardedDataParallel as FSDP + + has_FSDP = True +except ImportError: + FSDP = torch.nn.Module + has_FSDP = False + + +class FullyShardedDataParallel(FSDP): + """ + A small wrapper around fairscale's FullyShardedDataParallel (FSDP) with some + fairseq-specific checkpoint saving/loading logic. + + Args: + use_sharded_state (bool): if True, then ``state_dict`` will return + ``FSDP.local_state_dict`` and ``load_state_dict`` will call + ``FSDP.load_local_state_dict``. Otherwise, ``state_dict`` will + return the full model weights on data parallel rank 0 (empty on + other ranks) and ``load_state_dict`` will broadcast model weights + from rank 0 to other ranks. + """ + + def __init__(self, *args, use_sharded_state: bool = False, **kwargs): + if not has_FSDP: + raise ImportError( + "Cannot find FullyShardedDataParallel. " + "Please install fairscale with: pip install fairscale" + ) + super().__init__(*args, **kwargs) + self.use_sharded_state = use_sharded_state + + @property + def unwrapped_module(self) -> torch.nn.Module: + if self.flatten_parameters: + return self.module.module + else: + return self.module + + def state_dict(self, destination=None, prefix="", keep_vars=False): + if self.use_sharded_state: + return super().local_state_dict( + destination=destination, prefix=prefix, keep_vars=keep_vars + ) + else: + if self.rank == 0: + return super().state_dict( + destination=destination, prefix=prefix, keep_vars=keep_vars + ) + else: + # We must call state_dict() due to use of communication + # primitives. But we don't use the result. + super().state_dict() + return destination or {} + + def load_state_dict(self, state_dict, strict=True, model_cfg=None): + if self.use_sharded_state: + return super().load_local_state_dict(state_dict, strict=strict) + else: + state_dict = dist_utils.broadcast_object( + state_dict, src_rank=0, group=self.process_group + ) + return super().load_state_dict(state_dict, strict=strict) + + +@contextlib.contextmanager +def fsdp_enable_wrap(cfg: DistributedTrainingConfig): + try: + from fairscale.nn import enable_wrap + except ImportError: + raise ImportError( + "Cannot find FullyShardedDataParallel. " + "Please install fairscale with: pip install fairscale" + ) + if cfg.memory_efficient_fp16: + assert cfg.fp16 # memory_efficient_fp16 should imply fp16 + group = dist_utils.get_data_parallel_group() + if group is None and cfg.distributed_world_size == 1: + from fairscale.utils.testing import DummyProcessGroup + + group = DummyProcessGroup(rank=0, size=1) + fsdp_config = { + "process_group": group, + "reshard_after_forward": not cfg.no_reshard_after_forward, + "mixed_precision": cfg.fp16 and not cfg.memory_efficient_fp16, + "fp32_reduce_scatter": cfg.fp32_reduce_scatter, + "flatten_parameters": True, + "cpu_offload": cfg.cpu_offload, + "compute_dtype": torch.float16 if cfg.fp16 else torch.float32, + "bucket_cap_mb": cfg.bucket_cap_mb, + "state_dict_device": torch.device("cpu"), # reduce GPU mem usage + } + with enable_wrap( + wrapper_cls=FullyShardedDataParallel, + use_sharded_state=cfg.use_sharded_state, + **fsdp_config, + ): + yield + + +def fsdp_wrap(module, min_num_params: Optional[int] = None, **kwargs): + """ + Helper to wrap layers/modules in FSDP. This falls back to a no-op if + fairscale is not available. + + Args: + module (nn.Module): module to (maybe) wrap + min_num_params (int, Optional): minimum number of layer params to wrap + """ + try: + from fairscale.nn import wrap + + if min_num_params is not None: + num_params = sum(p.numel() for p in module.parameters()) + if num_params >= min_num_params: + return wrap(module, **kwargs) + else: + return module + else: + return wrap(module, **kwargs) + except ImportError: + return module diff --git a/fairseq/distributed/legacy_distributed_data_parallel.py b/fairseq/distributed/legacy_distributed_data_parallel.py new file mode 100644 index 0000000000000000000000000000000000000000..f2308f87c5233625a3fe1b27104f5ead003ae3cb --- /dev/null +++ b/fairseq/distributed/legacy_distributed_data_parallel.py @@ -0,0 +1,165 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +""" +A modified version of the legacy DistributedDataParallel module that uses c10d +communication primitives. This version is simpler than the latest PyTorch +version and is useful for debugging. Notably it does not overlap gradient +communication with the backward pass, which makes it slower but more robust +than the PyTorch version. + +This version also supports the *no_sync* context manager, which allows faster +training with `--update-freq`. +""" + +from collections import OrderedDict +from contextlib import contextmanager + +import torch +from torch import nn + +from fairseq.distributed import utils + + +class LegacyDistributedDataParallel(nn.Module): + """Implements distributed data parallelism at the module level. + + A simplified version of :class:`torch.nn.parallel.DistributedDataParallel`. + This version uses a c10d process group for communication and does not + broadcast buffers. + + Args: + module (~torch.nn.Module): module to be parallelized + process_group: the c10d process group to be used for distributed data + parallel all-reduction. + buffer_size (int, optional): number of elements to buffer before + performing all-reduce (default: 256M). + """ + + def __init__(self, module, process_group, buffer_size=2 ** 28): + super().__init__() + + self.module = module + self.process_group = process_group + self.world_size = utils.get_world_size(self.process_group) + + # Never use a bigger buffer than the number of model params + self.buffer_size = min(buffer_size, sum(p.numel() for p in module.parameters())) + self.buffer = None + + # We can also forcibly accumulate grads locally and only do the + # all-reduce at some later time + self.accumulate_grads = False + + # make per-device lists of parameters + paramlists = OrderedDict() + for param in self.module.parameters(): + device = param.device + if paramlists.get(device) is None: + paramlists[device] = [] + paramlists[device] += [param] + self.per_device_params = list(paramlists.values()) + + @contextmanager + def no_sync(self): + """A context manager to disable gradient synchronization.""" + old_accumulate_grads = self.accumulate_grads + self.accumulate_grads = True + yield + self.accumulate_grads = old_accumulate_grads + + def forward(self, *inputs, **kwargs): + return self.module(*inputs, **kwargs) + + def all_reduce_grads(self): + """ + This function must be called explicitly after backward to reduce + gradients. There is no automatic hook like c10d. + """ + + def all_reduce_params(params): + buffer = self.buffer + nonzero_buffer = False + if len(params) > 1: + offset = 0 + for p in params: + sz = p.numel() + if p.grad is not None: + buffer[offset : offset + sz].copy_(p.grad.data.view(-1)) + nonzero_buffer = True + else: + buffer[offset : offset + sz].zero_() + offset += sz + else: + # we only have a single grad to all-reduce + p = params[0] + if p.grad is not None: + buffer = p.grad.data + nonzero_buffer = True + elif p.numel() <= self.buffer.numel(): + buffer = buffer[: p.numel()] + buffer.zero_() + else: + buffer = torch.zeros_like(p) + + if nonzero_buffer: + buffer.div_(self.world_size) + + utils.all_reduce(buffer, self.process_group) + + # copy all-reduced grads back into their original place + offset = 0 + for p in params: + sz = p.numel() + if p.grad is not None: + p.grad.data.copy_(buffer[offset : offset + sz].view_as(p)) + else: + p.grad = buffer[offset : offset + sz].view_as(p).clone() + offset += sz + + def reduction_fn(): + # This function only needs to be called once + if self.accumulate_grads: + return + + if self.buffer is None: + self.buffer = next(self.module.parameters()).new(self.buffer_size) + + for params in self.per_device_params: + # All-reduce the gradients in buckets + offset = 0 + buffered_params = [] + for param in params: + if not param.requires_grad: + continue + if param.grad is None: + param.grad = torch.zeros_like(param) + + if hasattr(param, 'expert'): + # Skip gradient sync for unshared parameters + continue + + if param.grad.requires_grad: + raise RuntimeError( + "DistributedDataParallel only works " + "with gradients that don't require " + "grad" + ) + sz = param.numel() + if sz > self.buffer.numel(): + # all-reduce big params directly + all_reduce_params([param]) + else: + if offset + sz > self.buffer.numel(): + all_reduce_params(buffered_params) + offset = 0 + buffered_params.clear() + buffered_params.append(param) + offset += sz + + if len(buffered_params) > 0: + all_reduce_params(buffered_params) + + reduction_fn() diff --git a/fairseq/distributed/module_proxy_wrapper.py b/fairseq/distributed/module_proxy_wrapper.py new file mode 100644 index 0000000000000000000000000000000000000000..fc2c6f8c718f2ac8ece308e50f7ba74a05474f4a --- /dev/null +++ b/fairseq/distributed/module_proxy_wrapper.py @@ -0,0 +1,55 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from torch import nn + + +class ModuleProxyWrapper(nn.Module): + """ + Wrap a DistributedDataParallel module and forward requests for missing + attributes to the module wrapped by DDP (the twice-wrapped module). + Also forward calls to :func:`state_dict` and :func:`load_state_dict`. + + Usage:: + + module.xyz = "hello world" + wrapped_module = DistributedDataParallel(module, **ddp_args) + wrapped_module = ModuleProxyWrapper(wrapped_module) + assert wrapped_module.xyz == "hello world" + assert wrapped_module.state_dict().keys() == module.state_dict().keys() + + Args: + module (nn.Module): module to wrap + """ + + def __init__(self, module: nn.Module): + super().__init__() + assert hasattr(module, "module"), \ + "ModuleProxyWrapper expects input to wrap another module" + self.module = module + + def __getattr__(self, name): + """Forward missing attributes to twice-wrapped module.""" + try: + # defer to nn.Module's logic + return super().__getattr__(name) + except AttributeError: + try: + # forward to the once-wrapped module + return getattr(self.module, name) + except AttributeError: + # forward to the twice-wrapped module + return getattr(self.module.module, name) + + def state_dict(self, *args, **kwargs): + """Forward to the twice-wrapped module.""" + return self.module.module.state_dict(*args, **kwargs) + + def load_state_dict(self, *args, **kwargs): + """Forward to the twice-wrapped module.""" + return self.module.module.load_state_dict(*args, **kwargs) + + def forward(self, *args, **kwargs): + return self.module(*args, **kwargs) diff --git a/fairseq/distributed/tpu_distributed_data_parallel.py b/fairseq/distributed/tpu_distributed_data_parallel.py new file mode 100644 index 0000000000000000000000000000000000000000..e971cf07c57c4e864726781092a690dd4d7d3e46 --- /dev/null +++ b/fairseq/distributed/tpu_distributed_data_parallel.py @@ -0,0 +1,43 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +from torch import nn + +from fairseq.distributed import utils + + +class TPUDistributedDataParallel(nn.Module): + + def __init__(self, module, process_group): + super().__init__() + self.module = module + self.process_group = process_group + self.world_size = utils.get_world_size(self.process_group) + + def forward(self, *inputs, **kwargs): + return self.module(*inputs, **kwargs) + + def all_reduce_grads(self): + gradients = [] + for p in self.parameters(): + if not p.requires_grad: + continue + if p.grad is None: + p.grad = torch.zeros_like(p) + if p.grad.requires_grad: + raise RuntimeError( + "TPUDistributedDataParallel only works with gradients that don't " + "require grad" + ) + gradients.append(p.grad) + + import torch_xla.core.xla_model as xm + xm.all_reduce( + 'sum', + gradients, + scale=1. / self.world_size, + groups=self.process_group[1], + ) diff --git a/fairseq/distributed/utils.py b/fairseq/distributed/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..b7736116f97bd2b9f3a72339e179f06be5c33cfd --- /dev/null +++ b/fairseq/distributed/utils.py @@ -0,0 +1,805 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import io +import logging +import os +import pickle +import random +import socket +import struct +import subprocess +import warnings +from argparse import Namespace +from collections import OrderedDict +from dataclasses import dataclass +from typing import Any, Dict, List, Mapping, Optional + +import torch +import torch.distributed as dist +from fairseq.dataclass.configs import DistributedTrainingConfig, FairseqConfig +from omegaconf import open_dict + +try: + import torch_xla.core.xla_model as xm +except ImportError: + xm = None + + +# Flag to indicate if we're using Megatron +# NOTE: this is a temporary hack until we move away from Megatron's model parallel init +_USE_MEGATRON = False + +# Whether to use XLA ops (e.g., on TPUs) instead of CUDA ops. +_USE_XLA = False + + +logger = logging.getLogger(__name__) + + +def is_master(cfg: DistributedTrainingConfig): + return cfg.distributed_rank == 0 + + +def infer_init_method(cfg: DistributedTrainingConfig, force_distributed=False): + if cfg.distributed_init_method is not None or cfg.tpu: + return + + num_pipelines_per_node = None + if cfg.pipeline_model_parallel: + num_pipeline_devices, num_pipelines_per_node = _pipeline_parallel_pre_init(cfg) + + if all( + key in os.environ + for key in ["MASTER_ADDR", "MASTER_PORT", "WORLD_SIZE", "RANK"] + ): + # support torch.distributed.launch + _infer_torch_distributed_launch_init(cfg) + elif cfg.distributed_port > 0: + # we can determine the init method automatically for Slurm + _infer_slurm_init(cfg, num_pipelines_per_node) + elif cfg.distributed_world_size > 1 or force_distributed: + # fallback for single node with multiple GPUs + _infer_single_node_init(cfg) + + if cfg.pipeline_model_parallel: + _pipeline_parallel_post_init(cfg, num_pipeline_devices, num_pipelines_per_node) + elif not cfg.distributed_no_spawn: + with open_dict(cfg): + cfg.distributed_num_procs = min( + torch.cuda.device_count(), cfg.distributed_world_size + ) + + +def _infer_torch_distributed_launch_init(cfg: DistributedTrainingConfig): + cfg.distributed_init_method = "env://" + cfg.distributed_world_size = int(os.environ["WORLD_SIZE"]) + cfg.distributed_rank = int(os.environ["RANK"]) + # processes are created by torch.distributed.launch + cfg.distributed_no_spawn = True + + +def _infer_slurm_init(cfg: DistributedTrainingConfig, num_pipelines_per_node): + node_list = os.environ.get("SLURM_STEP_NODELIST") + if node_list is None: + node_list = os.environ.get("SLURM_JOB_NODELIST") + if node_list is not None: + try: + hostnames = subprocess.check_output( + ["scontrol", "show", "hostnames", node_list] + ) + cfg.distributed_init_method = "tcp://{host}:{port}".format( + host=hostnames.split()[0].decode("utf-8"), + port=cfg.distributed_port, + ) + nnodes = int(os.environ.get("SLURM_NNODES")) + ntasks_per_node = os.environ.get("SLURM_NTASKS_PER_NODE") + if ntasks_per_node is not None: + ntasks_per_node = int(ntasks_per_node) + else: + ntasks = int(os.environ.get("SLURM_NTASKS")) + nnodes = int(os.environ.get("SLURM_NNODES")) + assert ntasks % nnodes == 0 + ntasks_per_node = int(ntasks / nnodes) + if ntasks_per_node == 1: + gpus_per_node = torch.cuda.device_count() + node_id = int(os.environ.get("SLURM_NODEID")) + cfg.distributed_rank = node_id * gpus_per_node + cfg.distributed_world_size = nnodes * gpus_per_node + elif cfg.pipeline_model_parallel: + assert ntasks_per_node == num_pipelines_per_node, ( + "SLURM --ntasks-per-node must match number of pipelines per " + "node (={})".format(num_pipelines_per_node) + ) + cfg.distributed_no_spawn = True + # For 4-way MP on nodes with 8 GPUs, ranks will be [0, 1] on + # the first node, [1, 2] on the second node, etc. This + # matches torch.distributed.launch. + node_id = int(os.environ.get("SLURM_NODEID")) + local_id = int(os.environ.get("SLURM_LOCALID")) + cfg.distributed_rank = node_id * num_pipelines_per_node + local_id + # In the above example, device_id will always be in [0, 1], + # which also matches torch.distributed.launch. + cfg.device_id = local_id + # We also want to set distributed_world_size to be the total + # number of pipelines across all nodes. + cfg.distributed_world_size = nnodes * num_pipelines_per_node + else: + assert ntasks_per_node == cfg.distributed_world_size // nnodes + cfg.distributed_no_spawn = True + cfg.distributed_rank = int(os.environ.get("SLURM_PROCID")) + cfg.device_id = int(os.environ.get("SLURM_LOCALID")) + except subprocess.CalledProcessError as e: # scontrol failed + raise e + except FileNotFoundError: # Slurm is not installed + pass + + +def _infer_single_node_init(cfg: DistributedTrainingConfig): + assert ( + cfg.distributed_world_size <= torch.cuda.device_count() + ), f"world size is {cfg.distributed_world_size} but have {torch.cuda.device_count()} available devices" + port = random.randint(10000, 20000) + cfg.distributed_init_method = "tcp://localhost:{port}".format(port=port) + + +def _pipeline_parallel_pre_init(cfg: DistributedTrainingConfig): + from fairseq import utils + + balance_exists = ( + cfg.pipeline_balance is not None + or cfg.pipeline_encoder_balance is not None + or cfg.pipeline_decoder_balance is not None + ) + devices_exist = ( + cfg.pipeline_devices is not None + or cfg.pipeline_encoder_devices is not None + or cfg.pipeline_decoder_devices is not None + ) + if not balance_exists: + raise ValueError( + "--pipeline-balance is currently required for pipeline model parallelism" + ) + if not devices_exist: + raise ValueError( + "--pipeline-devices is currently required for pipeline model parallelism" + ) + + cfg.pipeline_balance = utils.eval_str_list(cfg.pipeline_balance, type=int) + if cfg.pipeline_devices is not None: + cfg.pipeline_devices = utils.eval_str_list(cfg.pipeline_devices, type=int) + num_pipeline_devices = len(set(cfg.pipeline_devices)) + else: + cfg.pipeline_encoder_devices = utils.eval_str_list( + cfg.pipeline_encoder_devices, type=int + ) + cfg.pipeline_decoder_devices = utils.eval_str_list( + cfg.pipeline_decoder_devices, type=int + ) + num_pipeline_devices = len( + set(cfg.pipeline_encoder_devices + cfg.pipeline_decoder_devices) + ) + gpus_per_node = torch.cuda.device_count() + assert ( + gpus_per_node >= num_pipeline_devices + and gpus_per_node % num_pipeline_devices == 0 + ), ( + "the number of unique device IDs in --pipeline-devices must evenly divide " + "the number of GPUs per node (multi-node pipelining is not yet supported)" + ) + num_pipelines_per_node = gpus_per_node // num_pipeline_devices + return num_pipeline_devices, num_pipelines_per_node + + +def _pipeline_parallel_post_init( + cfg: DistributedTrainingConfig, num_pipeline_devices, num_pipelines_per_node +): + if not cfg.distributed_no_spawn: + # When distributed_no_spawn is False, we expect distributed_rank and + # distributed_world_size to be based on the total number of GPUs, so + # we need to correct them to be based on the number of pipelines. + assert cfg.distributed_world_size % num_pipeline_devices == 0 + cfg.distributed_world_size = ( + cfg.distributed_world_size // num_pipeline_devices + ) + # In the case of 4-way MP on nodes with 8 GPUs, we want + # distributed_rank to be the starting GPU index for each pipeline + # i.e., 0, 2, ... + gpus_per_node = torch.cuda.device_count() + assert cfg.distributed_rank % gpus_per_node == 0 + assert cfg.distributed_rank % num_pipeline_devices == 0 + + with open_dict(cfg): + cfg.distributed_rank = cfg.distributed_rank // num_pipeline_devices + # launch one process per pipeline + cfg.distributed_num_procs = num_pipelines_per_node + + # if we have 4-way MP on a node with 8 GPUs, we want device_ids to be 0 + # and 4, indicating the starting device IDs for each pipeline + cfg.device_id *= num_pipeline_devices + + if cfg.device_id > 0: + # if there's multiple pipelines on a node (e.g., 4-way MP on an 8 + # GPU node), we need to adjust pipeline_devices accordingly + logger.debug( + "setting CUDA device={} on rank {}".format( + cfg.device_id, cfg.distributed_rank + ) + ) + torch.cuda.set_device(cfg.device_id) + with open_dict(cfg): + cfg.pipeline_devices = [cfg.device_id + d for d in cfg.pipeline_devices] + logger.info( + "setting pipeline_devices={} on rank {}".format( + cfg.pipeline_devices, cfg.distributed_rank + ) + ) + + +def distributed_init(cfg: FairseqConfig): + if isinstance(cfg, Namespace): + from fairseq.dataclass.utils import convert_namespace_to_omegaconf + + cfg = convert_namespace_to_omegaconf(cfg) + + if not cfg.common.tpu: + if torch.distributed.is_available() and torch.distributed.is_initialized(): + warnings.warn( + "Distributed is already initialized, cannot initialize twice!" + ) + else: + logger.info( + "distributed init (rank {}): {}".format( + cfg.distributed_training.distributed_rank, + cfg.distributed_training.distributed_init_method, + ) + ) + dist.init_process_group( + backend=cfg.distributed_training.distributed_backend, + init_method=cfg.distributed_training.distributed_init_method, + world_size=cfg.distributed_training.distributed_world_size, + rank=cfg.distributed_training.distributed_rank, + ) + logger.info( + "initialized host {} as rank {}".format( + socket.gethostname(), + cfg.distributed_training.distributed_rank, + ) + ) + + # perform a dummy all-reduce to initialize the NCCL communicator + if torch.cuda.is_available(): + dist.all_reduce(torch.zeros(1).cuda()) + + cfg.distributed_training.distributed_rank = torch.distributed.get_rank() + else: + assert xm.xrt_world_size() == cfg.distributed_training.distributed_world_size + global _USE_XLA + _USE_XLA = True + cfg.distributed_training.device_id = xm.get_local_ordinal() + cfg.distributed_training.distributed_rank = xm.get_ordinal() + xm.rendezvous("distributed_init") # wait for all workers + + if is_master(cfg.distributed_training): + logging.getLogger().setLevel(logging.INFO) + else: + logging.getLogger().setLevel(logging.WARNING) + + if cfg.common.model_parallel_size > 1: + try: + from fairseq.model_parallel.megatron.mpu import ( + initialize_model_parallel, + model_parallel_cuda_manual_seed, + ) + except ImportError: + raise ImportError( + "\n\nPlease install the megatron submodule:" + "\n\n git submodule update --init " + "fairseq/model_parallel/megatron" + ) + global _USE_MEGATRON + _USE_MEGATRON = True + initialize_model_parallel(cfg.common.model_parallel_size) + model_parallel_cuda_manual_seed(cfg.common.seed) + model_part_number = get_model_parallel_rank() + cfg.checkpoint.checkpoint_suffix += "-model_part-{0}".format(model_part_number) + + if hasattr(cfg, "model") and getattr(cfg.model, "base_layers", 0) > 0: + cfg.checkpoint.checkpoint_suffix = f"-rank-{cfg.distributed_training.distributed_rank}" + + return cfg.distributed_training.distributed_rank + + +def distributed_main(i, main, cfg: FairseqConfig, kwargs): + cfg.distributed_training.device_id = i + if torch.cuda.is_available() and not cfg.common.cpu and not cfg.common.tpu: + torch.cuda.set_device(cfg.distributed_training.device_id) + if cfg.distributed_training.distributed_rank is None: # torch.multiprocessing.spawn + cfg.distributed_training.distributed_rank = kwargs.pop("start_rank", 0) + i + + cfg.distributed_training.distributed_rank = distributed_init(cfg) + + after_distributed_init_fn = kwargs.pop("after_distributed_init_fn", None) + if after_distributed_init_fn: + cfg = after_distributed_init_fn(cfg) + + main(cfg, **kwargs) + + if torch.distributed.is_initialized(): + torch.distributed.barrier(get_global_group()) + + +def call_main(cfg: FairseqConfig, main, **kwargs): + if cfg.distributed_training.distributed_init_method is None: + infer_init_method(cfg.distributed_training) + + if cfg.distributed_training.distributed_init_method is not None: + # distributed training + if not cfg.distributed_training.distributed_no_spawn: + start_rank = cfg.distributed_training.distributed_rank + cfg.distributed_training.distributed_rank = None # assign automatically + kwargs["start_rank"] = start_rank + torch.multiprocessing.spawn( + fn=distributed_main, + args=(main, cfg, kwargs), + nprocs=min( + torch.cuda.device_count(), + cfg.distributed_training.distributed_world_size, + ), + join=True, + ) + else: + distributed_main(cfg.distributed_training.device_id, main, cfg, kwargs) + elif cfg.common.tpu and cfg.distributed_training.distributed_world_size > 1: + import torch_xla.distributed.xla_multiprocessing as xmp + + torch.multiprocessing.set_sharing_strategy("file_system") + xmp.spawn( + fn=distributed_main, + args=(main, cfg, kwargs), + # tpu-comment: + # 8 devices in one TPU VM, is the max processes to be spawned. + # The rest is driven by xm.distributed.xla_dist + nprocs=min(cfg.distributed_training.distributed_world_size, 8), + ) + else: + # single GPU main + main(cfg, **kwargs) + + +def use_xla(): + global _USE_XLA + return _USE_XLA + + +def new_groups(grouped_ranks: List[List[int]]): + if use_xla(): + return ("tpu", grouped_ranks) + else: + groups = [dist.new_group(g) for g in grouped_ranks] + my_group_idx = _find_my_group_index(grouped_ranks) + return groups[my_group_idx] + + +def _find_my_group_index(grouped_ranks): + my_rank = get_global_rank() + for i, group in enumerate(grouped_ranks): + if my_rank in group: + return i + raise RuntimeError + + +def _find_my_group(grouped_ranks): + index = _find_my_group_index(grouped_ranks) + return grouped_ranks[index] + + +def get_rank(group): + if use_xla(): + assert group[0] == "tpu" + my_group = _find_my_group(group[1]) + return my_group.index(get_global_rank()) + else: + return dist.get_rank(group=group) + + +def get_world_size(group): + if use_xla(): + assert group[0] == "tpu" + my_group = _find_my_group(group[1]) + return len(my_group) + elif torch.distributed.is_initialized(): + return dist.get_world_size(group=group) + else: + return 1 + + +def get_global_group(): + if use_xla(): + return new_groups([list(range(get_global_world_size()))]) + elif torch.distributed.is_initialized(): + if not hasattr(get_global_group, "_global_group"): + # ideally we could use torch.distributed.group.WORLD, but it seems + # to cause random NCCL hangs in some cases + get_global_group._global_group = dist.new_group() + return get_global_group._global_group + else: + return None + + +def get_global_rank(): + if use_xla(): + return xm.get_ordinal() + elif torch.distributed.is_initialized(): + return torch.distributed.get_rank() + else: + return 0 + + +def get_global_world_size(): + if use_xla(): + return xm.xrt_world_size() + elif torch.distributed.is_initialized(): + return torch.distributed.get_world_size() + else: + return 1 + + +def get_data_parallel_group(): + """Get the data parallel group the caller rank belongs to.""" + global _USE_MEGATRON + if _USE_MEGATRON: + from fairseq.model_parallel.megatron import mpu + + return mpu.get_data_parallel_group() + else: + return get_global_group() + + +def get_data_parallel_rank(): + """Return my rank for the data parallel group.""" + return get_rank(get_data_parallel_group()) + + +def get_data_parallel_world_size(): + """Return world size for the data parallel group.""" + return get_world_size(get_data_parallel_group()) + + +def get_model_parallel_group(): + global _USE_MEGATRON + if _USE_MEGATRON: + from fairseq.model_parallel.megatron import mpu + + return mpu.get_model_parallel_group() + else: + return None + + +def get_model_parallel_rank(): + """Return my rank for the model parallel group.""" + return get_rank(get_model_parallel_group()) + + +def get_model_parallel_world_size(): + """Return world size for the model parallel group.""" + return get_world_size(get_model_parallel_group()) + + +def all_reduce(tensor, group, op="sum"): + if use_xla(): + assert isinstance(group, tuple) and group[0] == "tpu" + tensor = [tensor] # wrap in a list to make xm.all_reduce in-place + return xm.all_reduce(op, tensor, groups=group[1])[0] + else: + if op == "sum": + op = dist.ReduceOp.SUM + elif op == "max": + op = dist.ReduceOp.MAX + else: + raise NotImplementedError + dist.all_reduce(tensor, op=op, group=group) + return tensor + + +def broadcast(tensor, src, group): + if use_xla(): + # XLA doesn't support broadcast, hack it with all_reduce + if get_rank(group) != src: + tensor.zero_() + all_reduce(tensor, group) + else: + dist.broadcast(tensor, src=src, group=group) + + +def all_to_all(tensor, group): + """Perform an all-to-all operation on a 1D Tensor.""" + assert tensor.dim() == 1 + split_count = get_world_size(group=group) + assert tensor.numel() % split_count == 0 + if use_xla(): + assert isinstance(group, tuple) and group[0] == "tpu" + return xm.all_to_all( + tensor, + split_dimension=0, + concat_dimension=0, + split_count=split_count, + groups=group[1], + ) + else: + output = torch.zeros_like(tensor) + dist.all_to_all_single(output, tensor, group=group) + return output + + +def all_gather(tensor, group, return_tensor=False): + """Perform an all-gather operation.""" + if use_xla(): + result = xm.all_gather(tensor, groups=group[1]) + world_size = get_world_size(group=group) + result = result.view(world_size, *tensor.size()) + if return_tensor: + return result + else: + return [result[i] for i in range(world_size)] + else: + world_size = get_world_size(group=group) + rank = get_rank(group=group) + tensor_list = [ + tensor if i == rank else torch.empty_like(tensor) for i in range(world_size) + ] + dist.all_gather(tensor_list, tensor, group=group) + if return_tensor: + return torch.stack(tensor_list, dim=0) + else: + return tensor_list + + +def all_gather_list(data, group=None, max_size=16384): + """Gathers arbitrary data from all nodes into a list. + + Similar to :func:`~torch.distributed.all_gather` but for arbitrary Python + data. Note that *data* must be picklable and any CUDA tensors will be moved + to CPU and returned on CPU as well. + + Args: + data (Any): data from the local worker to be gathered on other workers + group: group of the collective + max_size (int, optional): maximum size of the data to be gathered + across workers + """ + from fairseq import utils + + if group is None: + group = get_global_group() + rank = get_rank(group=group) + world_size = get_world_size(group=group) + + buffer_size = max_size * world_size + if ( + not hasattr(all_gather_list, "_buffer") + or all_gather_list._buffer.numel() < buffer_size + ): + all_gather_list._buffer = torch.cuda.ByteTensor(buffer_size) + all_gather_list._cpu_buffer = torch.ByteTensor(max_size).pin_memory() + buffer = all_gather_list._buffer + buffer.zero_() + cpu_buffer = all_gather_list._cpu_buffer + + data = utils.move_to_cpu(data) + enc = pickle.dumps(data) + enc_size = len(enc) + header_size = 4 # size of header that contains the length of the encoded data + size = header_size + enc_size + if size > max_size: + raise ValueError( + "encoded data size ({}) exceeds max_size ({})".format(size, max_size) + ) + + header = struct.pack(">I", enc_size) + cpu_buffer[:size] = torch.ByteTensor(list(header + enc)) + start = rank * max_size + buffer[start : start + size].copy_(cpu_buffer[:size]) + + all_reduce(buffer, group=group) + + buffer = buffer.cpu() + try: + result = [] + for i in range(world_size): + out_buffer = buffer[i * max_size : (i + 1) * max_size] + (enc_size,) = struct.unpack(">I", bytes(out_buffer[:header_size].tolist())) + if enc_size > 0: + result.append( + pickle.loads( + bytes(out_buffer[header_size : header_size + enc_size].tolist()) + ) + ) + return result + except pickle.UnpicklingError: + raise Exception( + "Unable to unpickle data from other workers. all_gather_list requires all " + "workers to enter the function together, so this error usually indicates " + "that the workers have fallen out of sync somehow. Workers can fall out of " + "sync if one of them runs out of memory, or if there are other conditions " + "in your training script that can cause one worker to finish an epoch " + "while other workers are still iterating over their portions of the data. " + "Try rerunning with --ddp-backend=legacy_ddp and see if that helps." + ) + + +def all_reduce_dict(data: Mapping[str, Any], device, group) -> Dict[str, Any]: + """ + AllReduce a dictionary of values across workers. We separately + reduce items that are already on the device and items on CPU for + better performance. + + Args: + data (Mapping[str, Any]): dictionary of data to all-reduce, but + cannot be a nested dictionary + device (torch.device): device for the reduction + group: group of the collective + """ + data_keys = list(data.keys()) + + # We want to separately reduce items that are already on the + # device and items on CPU for performance reasons. + cpu_data = OrderedDict() + device_data = OrderedDict() + for k in data_keys: + t = data[k] + if not torch.is_tensor(t): + cpu_data[k] = torch.tensor(t, dtype=torch.double) + elif t.device.type != device.type: + cpu_data[k] = t.to(dtype=torch.double) + else: + device_data[k] = t.to(dtype=torch.double) + + def _all_reduce_dict(data: OrderedDict): + if len(data) == 0: + return data + buf = torch.cat([t.view(-1) for t in data.values()]).to(device=device) + all_reduce(buf, group=group) + split_buf = torch.split(buf, [t.numel() for t in data.values()]) + reduced_data = [t.view_as(orig) for t, orig in zip(split_buf, data.values())] + return OrderedDict(zip(data.keys(), reduced_data)) + + cpu_data = _all_reduce_dict(cpu_data) + device_data = _all_reduce_dict(device_data) + + def get_from_stack(key): + if key in cpu_data: + return cpu_data[key] + elif key in device_data: + return device_data[key] + raise KeyError + + return OrderedDict([(key, get_from_stack(key)) for key in data_keys]) + + +def broadcast_tensors( + tensors: Optional[List[torch.Tensor]], + src_rank: int, + group: object, + dist_device: Optional[torch.device] = None, +) -> List[torch.Tensor]: + """ + Broadcasts a list of tensors without other (non-src) ranks needing to know + the dtypes/shapes of the tensors. + """ + if dist_device is None: + if torch.distributed.get_backend(group) == "nccl": + dist_device = torch.device("cuda") + else: + dist_device = torch.device("cpu") + + # share metadata first to simplify transfer + is_src_rank = (get_rank(group) == src_rank) + if is_src_rank: + metadata = [ + {"size": t.size(), "dtype": t.dtype, "device": t.device} for t in tensors + ] + metadata = _broadcast_object_slow(metadata, src_rank, group, dist_device) + else: + metadata = _broadcast_object_slow(None, src_rank, group, dist_device) + + out_tensors = [] + for i, meta in enumerate(metadata): + if is_src_rank: + tensor = tensors[i] + broadcast(tensors[i].to(dist_device), src=src_rank, group=group) + else: + tensor = torch.zeros( + [meta["size"].numel()], dtype=meta["dtype"], device=dist_device + ) + broadcast(tensor, src=src_rank, group=group) + tensor = tensor.view(meta["size"]).to(meta["device"]) + out_tensors.append(tensor) + return out_tensors + + +def broadcast_object( + obj: Any, + src_rank: int, + group: object, + dist_device: Optional[torch.device] = None, +) -> Any: + """Broadcast an arbitrary Python object to other workers.""" + if dist_device is None: + if torch.distributed.get_backend(group) == "nccl": + dist_device = torch.device("cuda") + else: + dist_device = torch.device("cpu") + + if get_rank(group) == src_rank: + # split the tensors from the non-tensors so we can broadcast them + # directly, avoiding unnecessary serialization/deserialization + tensors = [] + obj = _split_tensors_from_obj(obj, tensors) + obj = _broadcast_object_slow(obj, src_rank, group, dist_device) + tensors = broadcast_tensors(tensors, src_rank, group, dist_device) + else: + obj = _broadcast_object_slow(None, src_rank, group, dist_device) + tensors = broadcast_tensors(None, src_rank, group, dist_device) + return _put_tensors_in_obj(obj, tensors) + + +def _broadcast_object_slow( + obj: Any, src_rank: int, group: object, dist_device: torch.device, +) -> Any: + if get_rank(group) == src_rank: + # Emit data + buffer = io.BytesIO() + torch.save(obj, buffer) + buffer = torch.ByteTensor(buffer.getbuffer()).to(dist_device) + length = torch.LongTensor([len(buffer)]).to(dist_device) + broadcast(length, src=src_rank, group=group) + broadcast(buffer, src=src_rank, group=group) + else: + # Fetch from the source + length = torch.LongTensor([0]).to(dist_device) + broadcast(length, src=src_rank, group=group) + buffer = torch.ByteTensor(int(length.item())).to(dist_device) + broadcast(buffer, src=src_rank, group=group) + buffer = io.BytesIO(buffer.cpu().numpy()) + obj = torch.load(buffer, map_location="cpu") + return obj + + +@dataclass(frozen=True) +class _TensorPlaceholder: + index: int + + +def _split_tensors_from_obj(obj: Any, tensors: List[torch.Tensor]) -> Any: + if torch.is_tensor(obj): + placeholder = _TensorPlaceholder(index=len(tensors)) + tensors.append(obj) + return placeholder + elif isinstance(obj, dict): + return {k: _split_tensors_from_obj(v, tensors) for k, v in obj.items()} + elif isinstance(obj, list): + return [_split_tensors_from_obj(v, tensors) for v in obj] + elif isinstance(obj, tuple): + return tuple(_split_tensors_from_obj(v, tensors) for v in obj) + elif isinstance(obj, set): + return {_split_tensors_from_obj(v, tensors) for v in obj} + else: + return obj + + +def _put_tensors_in_obj(obj: Any, tensors: List[torch.Tensor]) -> Any: + if isinstance(obj, _TensorPlaceholder): + return tensors[obj.index] + elif isinstance(obj, dict): + return {k: _put_tensors_in_obj(v, tensors) for k, v in obj.items()} + elif isinstance(obj, list): + return [_put_tensors_in_obj(v, tensors) for v in obj] + elif isinstance(obj, tuple): + return tuple(_put_tensors_in_obj(v, tensors) for v in obj) + elif isinstance(obj, set): + return {_put_tensors_in_obj(v, tensors) for v in obj} + else: + return obj diff --git a/fairseq/file_io.py b/fairseq/file_io.py new file mode 100644 index 0000000000000000000000000000000000000000..dba663d4aafeb925ddffa50f5055933d6531a069 --- /dev/null +++ b/fairseq/file_io.py @@ -0,0 +1,194 @@ +#!/usr/bin/env python3 + +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import os +import shutil +from typing import List, Optional + + +logger = logging.getLogger(__file__) + + +try: + from iopath.common.file_io import g_pathmgr as IOPathManager + + try: + # [FB only - for now] AWS PathHandler for PathManager + from .fb_pathhandlers import S3PathHandler + + IOPathManager.register_handler(S3PathHandler()) + except KeyError: + logging.warning("S3PathHandler already registered.") + except ImportError: + logging.debug( + "S3PathHandler couldn't be imported. Either missing fb-only files, or boto3 module." + ) + +except ImportError: + IOPathManager = None + + +class PathManager: + """ + Wrapper for insulating OSS I/O (using Python builtin operations) from + iopath's PathManager abstraction (for transparently handling various + internal backends). + """ + + @staticmethod + def open( + path: str, + mode: str = "r", + buffering: int = -1, + encoding: Optional[str] = None, + errors: Optional[str] = None, + newline: Optional[str] = None, + ): + if IOPathManager: + return IOPathManager.open( + path=path, + mode=mode, + buffering=buffering, + encoding=encoding, + errors=errors, + newline=newline, + ) + return open( + path, + mode=mode, + buffering=buffering, + encoding=encoding, + errors=errors, + newline=newline, + ) + + @staticmethod + def copy(src_path: str, dst_path: str, overwrite: bool = False) -> bool: + if IOPathManager: + return IOPathManager.copy( + src_path=src_path, dst_path=dst_path, overwrite=overwrite + ) + return shutil.copyfile(src_path, dst_path) + + @staticmethod + def get_local_path(path: str, **kwargs) -> str: + if IOPathManager: + return IOPathManager.get_local_path(path, **kwargs) + return path + + @staticmethod + def exists(path: str) -> bool: + if IOPathManager: + return IOPathManager.exists(path) + return os.path.exists(path) + + @staticmethod + def isfile(path: str) -> bool: + if IOPathManager: + return IOPathManager.isfile(path) + return os.path.isfile(path) + + @staticmethod + def ls(path: str) -> List[str]: + if IOPathManager: + return IOPathManager.ls(path) + return os.listdir(path) + + @staticmethod + def mkdirs(path: str) -> None: + if IOPathManager: + return IOPathManager.mkdirs(path) + os.makedirs(path, exist_ok=True) + + @staticmethod + def rm(path: str) -> None: + if IOPathManager: + return IOPathManager.rm(path) + os.remove(path) + + @staticmethod + def chmod(path: str, mode: int) -> None: + if not PathManager.path_requires_pathmanager(path): + os.chmod(path, mode) + + @staticmethod + def register_handler(handler) -> None: + if IOPathManager: + return IOPathManager.register_handler(handler=handler) + + @staticmethod + def copy_from_local( + local_path: str, dst_path: str, overwrite: bool = False, **kwargs + ) -> None: + if IOPathManager: + return IOPathManager.copy_from_local( + local_path=local_path, dst_path=dst_path, overwrite=overwrite, **kwargs + ) + return shutil.copyfile(local_path, dst_path) + + @staticmethod + def path_requires_pathmanager(path: str) -> bool: + """Do we require PathManager to access given path?""" + if IOPathManager: + for p in IOPathManager._path_handlers.keys(): + if path.startswith(p): + return True + return False + + @staticmethod + def supports_rename(path: str) -> bool: + # PathManager doesn't yet support renames + return not PathManager.path_requires_pathmanager(path) + + @staticmethod + def rename(src: str, dst: str): + os.rename(src, dst) + + """ + ioPath async PathManager methods: + """ + @staticmethod + def opena( + path: str, + mode: str = "r", + buffering: int = -1, + encoding: Optional[str] = None, + errors: Optional[str] = None, + newline: Optional[str] = None, + ): + """ + Return file descriptor with asynchronous write operations. + """ + global IOPathManager + if not IOPathManager: + logging.info("ioPath is initializing PathManager.") + try: + from iopath.common.file_io import PathManager + IOPathManager = PathManager() + except Exception: + logging.exception("Failed to initialize ioPath PathManager object.") + return IOPathManager.opena( + path=path, + mode=mode, + buffering=buffering, + encoding=encoding, + errors=errors, + newline=newline, + ) + + @staticmethod + def async_close() -> bool: + """ + Wait for files to be written and clean up asynchronous PathManager. + NOTE: `PathManager.async_close()` must be called at the end of any + script that uses `PathManager.opena(...)`. + """ + global IOPathManager + if IOPathManager: + return IOPathManager.async_close() + return False diff --git a/fairseq/file_utils.py b/fairseq/file_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..d1d5ea65746682881264e4a9c462854dcfb3413f --- /dev/null +++ b/fairseq/file_utils.py @@ -0,0 +1,369 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +""" +Utilities for working with the local dataset cache. +This file is adapted from `AllenNLP <https://github.com/allenai/allennlp>`_. +and `huggingface <https://github.com/huggingface>`_. +""" + +import fnmatch +import json +import logging +import os +import shutil +import tarfile +import tempfile +from functools import partial, wraps +from hashlib import sha256 +from io import open + + +try: + from torch.hub import _get_torch_home + + torch_cache_home = _get_torch_home() +except ImportError: + torch_cache_home = os.path.expanduser( + os.getenv( + "TORCH_HOME", os.path.join(os.getenv("XDG_CACHE_HOME", "~/.cache"), "torch") + ) + ) +default_cache_path = os.path.join(torch_cache_home, "pytorch_fairseq") + +try: + from urllib.parse import urlparse +except ImportError: + from urlparse import urlparse + +try: + from pathlib import Path + + PYTORCH_FAIRSEQ_CACHE = Path(os.getenv("PYTORCH_FAIRSEQ_CACHE", default_cache_path)) +except (AttributeError, ImportError): + PYTORCH_FAIRSEQ_CACHE = os.getenv("PYTORCH_FAIRSEQ_CACHE", default_cache_path) + +CONFIG_NAME = "config.json" +WEIGHTS_NAME = "pytorch_model.bin" + +logger = logging.getLogger(__name__) # pylint: disable=invalid-name + + +def load_archive_file(archive_file): + # redirect to the cache, if necessary + try: + resolved_archive_file = cached_path(archive_file, cache_dir=None) + except EnvironmentError: + logger.info( + "Archive name '{}' was not found in archive name list. " + "We assumed '{}' was a path or URL but couldn't find any file " + "associated to this path or URL.".format( + archive_file, + archive_file, + ) + ) + return None + + if resolved_archive_file == archive_file: + logger.info("loading archive file {}".format(archive_file)) + else: + logger.info( + "loading archive file {} from cache at {}".format( + archive_file, resolved_archive_file + ) + ) + + # Extract archive to temp dir and replace .tar.bz2 if necessary + tempdir = None + if not os.path.isdir(resolved_archive_file): + tempdir = tempfile.mkdtemp() + logger.info( + "extracting archive file {} to temp dir {}".format( + resolved_archive_file, tempdir + ) + ) + ext = os.path.splitext(archive_file)[1][1:] + with tarfile.open(resolved_archive_file, "r:" + ext) as archive: + top_dir = os.path.commonprefix(archive.getnames()) + archive.extractall(tempdir) + os.remove(resolved_archive_file) + shutil.move(os.path.join(tempdir, top_dir), resolved_archive_file) + shutil.rmtree(tempdir) + + return resolved_archive_file + + +def url_to_filename(url, etag=None): + """ + Convert `url` into a hashed filename in a repeatable way. + If `etag` is specified, append its hash to the URL's, delimited + by a period. + """ + url_bytes = url.encode("utf-8") + url_hash = sha256(url_bytes) + filename = url_hash.hexdigest() + + if etag: + etag_bytes = etag.encode("utf-8") + etag_hash = sha256(etag_bytes) + filename += "." + etag_hash.hexdigest() + + return filename + + +def filename_to_url(filename, cache_dir=None): + """ + Return the url and etag (which may be ``None``) stored for `filename`. + Raise ``EnvironmentError`` if `filename` or its stored metadata do not exist. + """ + if cache_dir is None: + cache_dir = PYTORCH_FAIRSEQ_CACHE + if isinstance(cache_dir, Path): + cache_dir = str(cache_dir) + + cache_path = os.path.join(cache_dir, filename) + if not os.path.exists(cache_path): + raise EnvironmentError("file {} not found".format(cache_path)) + + meta_path = cache_path + ".json" + if not os.path.exists(meta_path): + raise EnvironmentError("file {} not found".format(meta_path)) + + with open(meta_path, encoding="utf-8") as meta_file: + metadata = json.load(meta_file) + url = metadata["url"] + etag = metadata["etag"] + + return url, etag + + +def cached_path_from_pm(url_or_filename): + """ + Tries to cache the specified URL using PathManager class. + Returns the cached path if success otherwise failure. + """ + try: + from fairseq.file_io import PathManager + local_path = PathManager.get_local_path(url_or_filename) + return local_path + except Exception: + return None + + +def cached_path(url_or_filename, cache_dir=None): + """ + Given something that might be a URL (or might be a local path), + determine which. If it's a URL, download the file and cache it, and + return the path to the cached file. If it's already a local path, + make sure the file exists and then return the path. + """ + if cache_dir is None: + cache_dir = PYTORCH_FAIRSEQ_CACHE + if isinstance(url_or_filename, Path): + url_or_filename = str(url_or_filename) + if isinstance(cache_dir, Path): + cache_dir = str(cache_dir) + + parsed = urlparse(url_or_filename) + + if parsed.scheme in ("http", "https", "s3"): + # URL, so get it from the cache (downloading if necessary) + return get_from_cache(url_or_filename, cache_dir) + elif os.path.exists(url_or_filename): + # File, and it exists. + return url_or_filename + elif parsed.scheme == "": + # File, but it doesn't exist. + raise EnvironmentError("file {} not found".format(url_or_filename)) + else: + cached_path = cached_path_from_pm(url_or_filename) + if cached_path: + return cached_path + # Something unknown + raise ValueError( + "unable to parse {} as a URL or as a local path".format(url_or_filename) + ) + + +def split_s3_path(url): + """Split a full s3 path into the bucket name and path.""" + parsed = urlparse(url) + if not parsed.netloc or not parsed.path: + raise ValueError("bad s3 path {}".format(url)) + bucket_name = parsed.netloc + s3_path = parsed.path + # Remove '/' at beginning of path. + if s3_path.startswith("/"): + s3_path = s3_path[1:] + return bucket_name, s3_path + + +def s3_request(func): + """ + Wrapper function for s3 requests in order to create more helpful error + messages. + """ + + @wraps(func) + def wrapper(url, *args, **kwargs): + from botocore.exceptions import ClientError + + try: + return func(url, *args, **kwargs) + except ClientError as exc: + if int(exc.response["Error"]["Code"]) == 404: + raise EnvironmentError("file {} not found".format(url)) + else: + raise + + return wrapper + + +@s3_request +def s3_etag(url): + """Check ETag on S3 object.""" + import boto3 + + s3_resource = boto3.resource("s3") + bucket_name, s3_path = split_s3_path(url) + s3_object = s3_resource.Object(bucket_name, s3_path) + return s3_object.e_tag + + +@s3_request +def s3_get(url, temp_file): + """Pull a file directly from S3.""" + import boto3 + + s3_resource = boto3.resource("s3") + bucket_name, s3_path = split_s3_path(url) + s3_resource.Bucket(bucket_name).download_fileobj(s3_path, temp_file) + + +def request_wrap_timeout(func, url): + import requests + + for attempt, timeout in enumerate([10, 20, 40, 60, 60]): + try: + return func(timeout=timeout) + except requests.exceptions.Timeout as e: + logger.warning( + "Request for %s timed-out (attempt %d). Retrying with a timeout of %d secs", + url, + attempt, + timeout, + exc_info=e, + ) + continue + raise RuntimeError(f"Unable to fetch file {url}") + + +def http_get(url, temp_file): + import requests + from tqdm import tqdm + + req = request_wrap_timeout(partial(requests.get, url, stream=True), url) + content_length = req.headers.get("Content-Length") + total = int(content_length) if content_length is not None else None + progress = tqdm(unit="B", total=total) + for chunk in req.iter_content(chunk_size=1024): + if chunk: # filter out keep-alive new chunks + progress.update(len(chunk)) + temp_file.write(chunk) + progress.close() + + +def get_from_cache(url, cache_dir=None): + """ + Given a URL, look for the corresponding dataset in the local cache. + If it's not there, download it. Then return the path to the cached file. + """ + if cache_dir is None: + cache_dir = PYTORCH_FAIRSEQ_CACHE + if isinstance(cache_dir, Path): + cache_dir = str(cache_dir) + + if not os.path.exists(cache_dir): + os.makedirs(cache_dir) + + # Get eTag to add to filename, if it exists. + if url.startswith("s3://"): + etag = s3_etag(url) + else: + try: + import requests + + response = request_wrap_timeout( + partial(requests.head, url, allow_redirects=True), url + ) + if response.status_code != 200: + etag = None + else: + etag = response.headers.get("ETag") + except RuntimeError: + etag = None + + filename = url_to_filename(url, etag) + + # get cache path to put the file + cache_path = os.path.join(cache_dir, filename) + + # If we don't have a connection (etag is None) and can't identify the file + # try to get the last downloaded one + if not os.path.exists(cache_path) and etag is None: + matching_files = fnmatch.filter(os.listdir(cache_dir), filename + ".*") + matching_files = list(filter(lambda s: not s.endswith(".json"), matching_files)) + if matching_files: + cache_path = os.path.join(cache_dir, matching_files[-1]) + + if not os.path.exists(cache_path): + # Download to temporary file, then copy to cache dir once finished. + # Otherwise you get corrupt cache entries if the download gets interrupted. + with tempfile.NamedTemporaryFile() as temp_file: + logger.info("%s not found in cache, downloading to %s", url, temp_file.name) + + # GET file object + if url.startswith("s3://"): + s3_get(url, temp_file) + else: + http_get(url, temp_file) + + # we are copying the file before closing it, so flush to avoid truncation + temp_file.flush() + # shutil.copyfileobj() starts at the current position, so go to the start + temp_file.seek(0) + + logger.info("copying %s to cache at %s", temp_file.name, cache_path) + with open(cache_path, "wb") as cache_file: + shutil.copyfileobj(temp_file, cache_file) + + logger.info("creating metadata file for %s", cache_path) + meta = {"url": url, "etag": etag} + meta_path = cache_path + ".json" + with open(meta_path, "w") as meta_file: + output_string = json.dumps(meta) + meta_file.write(output_string) + + logger.info("removing temp file %s", temp_file.name) + + return cache_path + + +def read_set_from_file(filename): + """ + Extract a de-duped collection (set) of text from a file. + Expected file format is one item per line. + """ + collection = set() + with open(filename, "r", encoding="utf-8") as file_: + for line in file_: + collection.add(line.rstrip()) + return collection + + +def get_file_extension(path, dot=True, lower=True): + ext = os.path.splitext(path)[1] + ext = ext if dot else ext[1:] + return ext.lower() if lower else ext diff --git a/fairseq/hub_utils.py b/fairseq/hub_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..d74470d2ecba2825221a2efa2ce21a9b698340df --- /dev/null +++ b/fairseq/hub_utils.py @@ -0,0 +1,303 @@ +#!/usr/bin/env python3 -u +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import copy +import logging +import os +from typing import Any, Dict, Iterator, List + +import torch +from fairseq import utils +from fairseq.data import encoders +from omegaconf import open_dict +from torch import nn + + +logger = logging.getLogger(__name__) + + +def from_pretrained( + model_name_or_path, + checkpoint_file="model.pt", + data_name_or_path=".", + archive_map=None, + **kwargs +): + from fairseq import checkpoint_utils, file_utils + + if archive_map is not None: + if model_name_or_path in archive_map: + model_name_or_path = archive_map[model_name_or_path] + if data_name_or_path is not None and data_name_or_path in archive_map: + data_name_or_path = archive_map[data_name_or_path] + + # allow archive_map to set default arg_overrides (e.g., tokenizer, bpe) + # for each model + if isinstance(model_name_or_path, dict): + for k, v in model_name_or_path.items(): + if k == "checkpoint_file": + checkpoint_file = v + elif ( + k != "path" + # only set kwargs that don't already have overrides + and k not in kwargs + ): + kwargs[k] = v + model_name_or_path = model_name_or_path["path"] + + model_path = file_utils.load_archive_file(model_name_or_path) + + # convenience hack for loading data and BPE codes from model archive + if data_name_or_path.startswith("."): + kwargs["data"] = os.path.abspath(os.path.join(model_path, data_name_or_path)) + else: + kwargs["data"] = file_utils.load_archive_file(data_name_or_path) + for file, arg in { + "code": "bpe_codes", + "bpecodes": "bpe_codes", + "sentencepiece.bpe.model": "sentencepiece_model", + "merges.txt": "bpe_merges", + "vocab.json": "bpe_vocab", + }.items(): + path = os.path.join(model_path, file) + if os.path.exists(path): + kwargs[arg] = path + + if "user_dir" in kwargs: + utils.import_user_module(argparse.Namespace(user_dir=kwargs["user_dir"])) + + models, args, task = checkpoint_utils.load_model_ensemble_and_task( + [os.path.join(model_path, cpt) for cpt in checkpoint_file.split(os.pathsep)], + arg_overrides=kwargs, + ) + + return { + "args": args, + "task": task, + "models": models, + } + + +class GeneratorHubInterface(nn.Module): + """ + PyTorch Hub interface for generating sequences from a pre-trained + translation or language model. + """ + + def __init__(self, cfg, task, models): + super().__init__() + self.cfg = cfg + self.task = task + self.models = nn.ModuleList(models) + self.src_dict = task.source_dictionary + self.tgt_dict = task.target_dictionary + + # optimize model for generation + for model in self.models: + model.prepare_for_inference_(cfg) + + # Load alignment dictionary for unknown word replacement + # (None if no unknown word replacement, empty if no path to align dictionary) + self.align_dict = utils.load_align_dict(cfg.generation.replace_unk) + + self.tokenizer = encoders.build_tokenizer(cfg.tokenizer) + self.bpe = encoders.build_bpe(cfg.bpe) + + self.max_positions = utils.resolve_max_positions( + self.task.max_positions(), *[model.max_positions() for model in models] + ) + + # this is useful for determining the device + self.register_buffer("_float_tensor", torch.tensor([0], dtype=torch.float)) + + @property + def device(self): + return self._float_tensor.device + + def translate( + self, sentences: List[str], beam: int = 5, verbose: bool = False, **kwargs + ) -> List[str]: + return self.sample(sentences, beam, verbose, **kwargs) + + def sample( + self, sentences: List[str], beam: int = 1, verbose: bool = False, **kwargs + ) -> List[str]: + if isinstance(sentences, str): + return self.sample([sentences], beam=beam, verbose=verbose, **kwargs)[0] + tokenized_sentences = [self.encode(sentence) for sentence in sentences] + batched_hypos = self.generate(tokenized_sentences, beam, verbose, **kwargs) + return [self.decode(hypos[0]["tokens"]) for hypos in batched_hypos] + + def score(self, sentences: List[str], **kwargs): + if isinstance(sentences, str): + return self.score([sentences], **kwargs)[0] + # NOTE: this doesn't support translation tasks currently + tokenized_sentences = [self.encode(sentence) for sentence in sentences] + return [ + hypos[0] + for hypos in self.generate( + tokenized_sentences, score_reference=True, **kwargs + ) + ] + + def generate( + self, + tokenized_sentences: List[torch.LongTensor], + beam: int = 5, + verbose: bool = False, + skip_invalid_size_inputs=False, + inference_step_args=None, + prefix_allowed_tokens_fn=None, + **kwargs + ) -> List[List[Dict[str, torch.Tensor]]]: + if torch.is_tensor(tokenized_sentences) and tokenized_sentences.dim() == 1: + return self.generate( + tokenized_sentences.unsqueeze(0), beam=beam, verbose=verbose, **kwargs + )[0] + + # build generator using current args as well as any kwargs + gen_args = copy.deepcopy(self.cfg.generation) + with open_dict(gen_args): + gen_args.beam = beam + for k, v in kwargs.items(): + setattr(gen_args, k, v) + generator = self.task.build_generator( + self.models, + gen_args, + prefix_allowed_tokens_fn=prefix_allowed_tokens_fn, + ) + + inference_step_args = inference_step_args or {} + results = [] + for batch in self._build_batches(tokenized_sentences, skip_invalid_size_inputs): + batch = utils.apply_to_sample(lambda t: t.to(self.device), batch) + translations = self.task.inference_step( + generator, self.models, batch, **inference_step_args + ) + for id, hypos in zip(batch["id"].tolist(), translations): + results.append((id, hypos)) + + # sort output to match input order + outputs = [hypos for _, hypos in sorted(results, key=lambda x: x[0])] + + if verbose: + + def getarg(name, default): + return getattr(gen_args, name, getattr(self.cfg, name, default)) + + for source_tokens, target_hypotheses in zip(tokenized_sentences, outputs): + src_str_with_unk = self.string(source_tokens) + logger.info("S\t{}".format(src_str_with_unk)) + for hypo in target_hypotheses: + hypo_str = self.decode(hypo["tokens"]) + logger.info("H\t{}\t{}".format(hypo["score"], hypo_str)) + logger.info( + "P\t{}".format( + " ".join( + map( + lambda x: "{:.4f}".format(x), + hypo["positional_scores"].tolist(), + ) + ) + ) + ) + if hypo["alignment"] is not None and getarg( + "print_alignment", False + ): + logger.info( + "A\t{}".format( + " ".join( + [ + "{}-{}".format(src_idx, tgt_idx) + for src_idx, tgt_idx in hypo["alignment"] + ] + ) + ) + ) + return outputs + + def encode(self, sentence: str) -> torch.LongTensor: + sentence = self.tokenize(sentence) + sentence = self.apply_bpe(sentence) + return self.binarize(sentence) + + def decode(self, tokens: torch.LongTensor) -> str: + sentence = self.string(tokens) + sentence = self.remove_bpe(sentence) + return self.detokenize(sentence) + + def tokenize(self, sentence: str) -> str: + if self.tokenizer is not None: + sentence = self.tokenizer.encode(sentence) + return sentence + + def detokenize(self, sentence: str) -> str: + if self.tokenizer is not None: + sentence = self.tokenizer.decode(sentence) + return sentence + + def apply_bpe(self, sentence: str) -> str: + if self.bpe is not None: + sentence = self.bpe.encode(sentence) + return sentence + + def remove_bpe(self, sentence: str) -> str: + if self.bpe is not None: + sentence = self.bpe.decode(sentence) + return sentence + + def binarize(self, sentence: str) -> torch.LongTensor: + return self.src_dict.encode_line(sentence, add_if_not_exist=False).long() + + def string(self, tokens: torch.LongTensor) -> str: + return self.tgt_dict.string(tokens) + + def _build_batches( + self, tokens: List[List[int]], skip_invalid_size_inputs: bool + ) -> Iterator[Dict[str, Any]]: + lengths = torch.LongTensor([t.numel() for t in tokens]) + batch_iterator = self.task.get_batch_iterator( + dataset=self.task.build_dataset_for_inference(tokens, lengths), + max_tokens=self.cfg.dataset.max_tokens, + max_sentences=self.cfg.dataset.batch_size, + max_positions=self.max_positions, + ignore_invalid_inputs=skip_invalid_size_inputs, + disable_iterator_cache=True, + ).next_epoch_itr(shuffle=False) + return batch_iterator + + +class BPEHubInterface(object): + """PyTorch Hub interface for Byte-Pair Encoding (BPE).""" + + def __init__(self, bpe, **kwargs): + super().__init__() + args = argparse.Namespace(bpe=bpe, **kwargs) + self.bpe = encoders.build_bpe(args) + assert self.bpe is not None + + def encode(self, sentence: str) -> str: + return self.bpe.encode(sentence) + + def decode(self, sentence: str) -> str: + return self.bpe.decode(sentence) + + +class TokenizerHubInterface(object): + """PyTorch Hub interface for tokenization.""" + + def __init__(self, tokenizer, **kwargs): + super().__init__() + args = argparse.Namespace(tokenizer=tokenizer, **kwargs) + self.tokenizer = encoders.build_tokenizer(args) + assert self.tokenizer is not None + + def encode(self, sentence: str) -> str: + return self.tokenizer.encode(sentence) + + def decode(self, sentence: str) -> str: + return self.tokenizer.decode(sentence) diff --git a/fairseq/incremental_decoding_utils.py b/fairseq/incremental_decoding_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..b26e6cd01cd4cbdffa23d88b354eb4a55a94189b --- /dev/null +++ b/fairseq/incremental_decoding_utils.py @@ -0,0 +1,51 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import uuid +from typing import Dict, Optional + +from torch import Tensor + + +class FairseqIncrementalState(object): + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + self.init_incremental_state() + + def init_incremental_state(self): + self._incremental_state_id = str(uuid.uuid4()) + + def _get_full_incremental_state_key(self, key: str) -> str: + return "{}.{}".format(self._incremental_state_id, key) + + def get_incremental_state( + self, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]], + key: str, + ) -> Optional[Dict[str, Optional[Tensor]]]: + """Helper for getting incremental state for an nn.Module.""" + full_key = self._get_full_incremental_state_key(key) + if incremental_state is None or full_key not in incremental_state: + return None + return incremental_state[full_key] + + def set_incremental_state( + self, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]], + key: str, + value: Dict[str, Optional[Tensor]], + ) -> Optional[Dict[str, Dict[str, Optional[Tensor]]]]: + """Helper for setting incremental state for an nn.Module.""" + if incremental_state is not None: + full_key = self._get_full_incremental_state_key(key) + incremental_state[full_key] = value + return incremental_state + + +def with_incremental_state(cls): + cls.__bases__ = (FairseqIncrementalState,) + tuple( + b for b in cls.__bases__ if b != FairseqIncrementalState + ) + return cls diff --git a/fairseq/iterative_refinement_generator.py b/fairseq/iterative_refinement_generator.py new file mode 100644 index 0000000000000000000000000000000000000000..4fb0946f499329ceb130761b59675d761df1c158 --- /dev/null +++ b/fairseq/iterative_refinement_generator.py @@ -0,0 +1,359 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from collections import namedtuple + +import numpy as np +import torch +from fairseq import utils + + +DecoderOut = namedtuple( + "IterativeRefinementDecoderOut", + ["output_tokens", "output_scores", "attn", "step", "max_step", "history"], +) + + +class IterativeRefinementGenerator(object): + def __init__( + self, + tgt_dict, + models=None, + eos_penalty=0.0, + max_iter=10, + max_ratio=2, + beam_size=1, + decoding_format=None, + retain_dropout=False, + adaptive=True, + retain_history=False, + reranking=False, + ): + """ + Generates translations based on iterative refinement. + + Args: + tgt_dict: target dictionary + eos_penalty: if > 0.0, it penalized early-stopping in decoding + max_iter: maximum number of refinement iterations + max_ratio: generate sequences of maximum length ax, where x is the source length + decoding_format: decoding mode in {'unigram', 'ensemble', 'vote', 'dp', 'bs'} + retain_dropout: retaining dropout in the inference + adaptive: decoding with early stop + """ + self.bos = tgt_dict.bos() + self.pad = tgt_dict.pad() + self.unk = tgt_dict.unk() + self.eos = tgt_dict.eos() + self.vocab_size = len(tgt_dict) + self.eos_penalty = eos_penalty + self.max_iter = max_iter + self.max_ratio = max_ratio + self.beam_size = beam_size + self.reranking = reranking + self.decoding_format = decoding_format + self.retain_dropout = retain_dropout + self.retain_history = retain_history + self.adaptive = adaptive + self.models = models + + def generate_batched_itr( + self, + data_itr, + maxlen_a=None, + maxlen_b=None, + cuda=False, + timer=None, + prefix_size=0, + ): + """Iterate over a batched dataset and yield individual translations. + + Args: + maxlen_a/b: generate sequences of maximum length ax + b, + where x is the source sentence length. + cuda: use GPU for generation + timer: StopwatchMeter for timing generations. + """ + + for sample in data_itr: + if "net_input" not in sample: + continue + if timer is not None: + timer.start() + with torch.no_grad(): + hypos = self.generate( + self.models, + sample, + prefix_tokens=sample["target"][:, :prefix_size] + if prefix_size > 0 + else None, + ) + if timer is not None: + timer.stop(sample["ntokens"]) + for i, id in enumerate(sample["id"]): + # remove padding + src = utils.strip_pad(sample["net_input"]["src_tokens"][i, :], self.pad) + ref = utils.strip_pad(sample["target"][i, :], self.pad) + yield id, src, ref, hypos[i] + + @torch.no_grad() + def generate(self, models, sample, prefix_tokens=None, constraints=None): + if constraints is not None: + raise NotImplementedError( + "Constrained decoding with the IterativeRefinementGenerator is not supported" + ) + + # TODO: iterative refinement generator does not support ensemble for now. + if not self.retain_dropout: + for model in models: + model.eval() + + model, reranker = models[0], None + if self.reranking: + assert len(models) > 1, "Assuming the last checkpoint is the reranker" + assert ( + self.beam_size > 1 + ), "Reranking requires multiple translation for each example" + + reranker = models[-1] + models = models[:-1] + + if len(models) > 1 and hasattr(model, "enable_ensemble"): + assert model.allow_ensemble, "{} does not support ensembling".format( + model.__class__.__name__ + ) + model.enable_ensemble(models) + + # TODO: better encoder inputs? + src_tokens = sample["net_input"]["src_tokens"] + src_lengths = sample["net_input"]["src_lengths"] + bsz, src_len = src_tokens.size() + + # initialize + encoder_out = model.forward_encoder([src_tokens, src_lengths]) + prev_decoder_out = model.initialize_output_tokens(encoder_out, src_tokens) + + if self.beam_size > 1: + assert ( + model.allow_length_beam + ), "{} does not support decoding with length beam.".format( + model.__class__.__name__ + ) + + # regenerate data based on length-beam + length_beam_order = ( + utils.new_arange(src_tokens, self.beam_size, bsz).t().reshape(-1) + ) + encoder_out = model.encoder.reorder_encoder_out( + encoder_out, length_beam_order + ) + prev_decoder_out = model.regenerate_length_beam( + prev_decoder_out, self.beam_size + ) + bsz = bsz * self.beam_size + + sent_idxs = torch.arange(bsz) + prev_output_tokens = prev_decoder_out.output_tokens.clone() + + if self.retain_history: + prev_decoder_out = prev_decoder_out._replace(history=[prev_output_tokens]) + + finalized = [[] for _ in range(bsz)] + + def is_a_loop(x, y, s, a): + b, l_x, l_y = x.size(0), x.size(1), y.size(1) + if l_x > l_y: + y = torch.cat([y, x.new_zeros(b, l_x - l_y).fill_(self.pad)], 1) + s = torch.cat([s, s.new_zeros(b, l_x - l_y)], 1) + if a is not None: + a = torch.cat([a, a.new_zeros(b, l_x - l_y, a.size(2))], 1) + elif l_x < l_y: + x = torch.cat([x, y.new_zeros(b, l_y - l_x).fill_(self.pad)], 1) + return (x == y).all(1), y, s, a + + def finalized_hypos(step, prev_out_token, prev_out_score, prev_out_attn): + cutoff = prev_out_token.ne(self.pad) + tokens = prev_out_token[cutoff] + if prev_out_score is None: + scores, score = None, None + else: + scores = prev_out_score[cutoff] + score = scores.mean() + + if prev_out_attn is None: + hypo_attn, alignment = None, None + else: + hypo_attn = prev_out_attn[cutoff] + alignment = hypo_attn.max(dim=1)[1] + return { + "steps": step, + "tokens": tokens, + "positional_scores": scores, + "score": score, + "hypo_attn": hypo_attn, + "alignment": alignment, + } + + for step in range(self.max_iter + 1): + + decoder_options = { + "eos_penalty": self.eos_penalty, + "max_ratio": self.max_ratio, + "decoding_format": self.decoding_format, + } + prev_decoder_out = prev_decoder_out._replace( + step=step, + max_step=self.max_iter + 1, + ) + + decoder_out = model.forward_decoder( + prev_decoder_out, encoder_out, **decoder_options + ) + + if self.adaptive: + # terminate if there is a loop + terminated, out_tokens, out_scores, out_attn = is_a_loop( + prev_output_tokens, + decoder_out.output_tokens, + decoder_out.output_scores, + decoder_out.attn, + ) + decoder_out = decoder_out._replace( + output_tokens=out_tokens, + output_scores=out_scores, + attn=out_attn, + ) + + else: + terminated = decoder_out.output_tokens.new_zeros( + decoder_out.output_tokens.size(0) + ).bool() + + if step == self.max_iter: # reach last iteration, terminate + terminated.fill_(1) + + # collect finalized sentences + finalized_idxs = sent_idxs[terminated] + finalized_tokens = decoder_out.output_tokens[terminated] + finalized_scores = decoder_out.output_scores[terminated] + finalized_attn = ( + None + if (decoder_out.attn is None or decoder_out.attn.size(0) == 0) + else decoder_out.attn[terminated] + ) + + if self.retain_history: + finalized_history_tokens = [h[terminated] for h in decoder_out.history] + + for i in range(finalized_idxs.size(0)): + finalized[finalized_idxs[i]] = [ + finalized_hypos( + step, + finalized_tokens[i], + finalized_scores[i], + None if finalized_attn is None else finalized_attn[i], + ) + ] + + if self.retain_history: + finalized[finalized_idxs[i]][0]["history"] = [] + for j in range(len(finalized_history_tokens)): + finalized[finalized_idxs[i]][0]["history"].append( + finalized_hypos( + step, finalized_history_tokens[j][i], None, None + ) + ) + + # check if all terminated + if terminated.sum() == terminated.size(0): + break + + # for next step + not_terminated = ~terminated + prev_decoder_out = decoder_out._replace( + output_tokens=decoder_out.output_tokens[not_terminated], + output_scores=decoder_out.output_scores[not_terminated], + attn=decoder_out.attn[not_terminated] + if (decoder_out.attn is not None and decoder_out.attn.size(0) > 0) + else None, + history=[h[not_terminated] for h in decoder_out.history] + if decoder_out.history is not None + else None, + ) + encoder_out = model.encoder.reorder_encoder_out( + encoder_out, not_terminated.nonzero(as_tuple=False).squeeze() + ) + sent_idxs = sent_idxs[not_terminated] + prev_output_tokens = prev_decoder_out.output_tokens.clone() + + if self.beam_size > 1: + if reranker is not None: + finalized = self.rerank( + reranker, finalized, [src_tokens, src_lengths], self.beam_size + ) + + # aggregate information from length beam + finalized = [ + finalized[ + np.argmax( + [ + finalized[self.beam_size * i + j][0]["score"] + for j in range(self.beam_size) + ] + ) + + self.beam_size * i + ] + for i in range(len(finalized) // self.beam_size) + ] + + return finalized + + def rerank(self, reranker, finalized, encoder_input, beam_size): + def rebuild_batch(finalized): + finalized_tokens = [f[0]["tokens"] for f in finalized] + finalized_maxlen = max(f.size(0) for f in finalized_tokens) + final_output_tokens = ( + finalized_tokens[0] + .new_zeros(len(finalized_tokens), finalized_maxlen) + .fill_(self.pad) + ) + for i, f in enumerate(finalized_tokens): + final_output_tokens[i, : f.size(0)] = f + return final_output_tokens + + final_output_tokens = rebuild_batch(finalized) + final_output_tokens[ + :, 0 + ] = self.eos # autoregressive model assumes starting with EOS + + reranker_encoder_out = reranker.encoder(*encoder_input) + length_beam_order = ( + utils.new_arange( + final_output_tokens, beam_size, reranker_encoder_out.encoder_out.size(1) + ) + .t() + .reshape(-1) + ) + reranker_encoder_out = reranker.encoder.reorder_encoder_out( + reranker_encoder_out, length_beam_order + ) + reranking_scores = reranker.get_normalized_probs( + reranker.decoder(final_output_tokens[:, :-1], reranker_encoder_out), + True, + None, + ) + reranking_scores = reranking_scores.gather(2, final_output_tokens[:, 1:, None]) + reranking_masks = final_output_tokens[:, 1:].ne(self.pad) + reranking_scores = ( + reranking_scores[:, :, 0].masked_fill_(~reranking_masks, 0).sum(1) + ) + reranking_scores = reranking_scores / reranking_masks.sum(1).type_as( + reranking_scores + ) + + for i in range(len(finalized)): + finalized[i][0]["score"] = reranking_scores[i] + + return finalized diff --git a/fairseq/logging/__init__.py b/fairseq/logging/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/fairseq/logging/meters.py b/fairseq/logging/meters.py new file mode 100644 index 0000000000000000000000000000000000000000..2100b1fa0b2704b1c585f59e9349655bba0cc9e6 --- /dev/null +++ b/fairseq/logging/meters.py @@ -0,0 +1,323 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import bisect +import time +from collections import OrderedDict +from typing import Dict, Optional + + +try: + import torch + + def type_as(a, b): + if torch.is_tensor(a) and torch.is_tensor(b): + return a.to(b) + else: + return a + + +except ImportError: + torch = None + + def type_as(a, b): + return a + + +try: + import numpy as np +except ImportError: + np = None + + +class Meter(object): + """Base class for Meters.""" + + def __init__(self): + pass + + def state_dict(self): + return {} + + def load_state_dict(self, state_dict): + pass + + def reset(self): + raise NotImplementedError + + @property + def smoothed_value(self) -> float: + """Smoothed value used for logging.""" + raise NotImplementedError + + +def safe_round(number, ndigits): + if hasattr(number, "__round__"): + return round(number, ndigits) + elif torch is not None and torch.is_tensor(number) and number.numel() == 1: + return safe_round(number.item(), ndigits) + elif np is not None and np.ndim(number) == 0 and hasattr(number, "item"): + return safe_round(number.item(), ndigits) + else: + return number + + +class AverageMeter(Meter): + """Computes and stores the average and current value""" + + def __init__(self, round: Optional[int] = None): + self.round = round + self.reset() + + def reset(self): + self.val = None # most recent update + self.sum = 0 # sum from all updates + self.count = 0 # total n from all updates + + def update(self, val, n=1): + if val is not None: + self.val = val + if n > 0: + self.sum = type_as(self.sum, val) + (val * n) + self.count = type_as(self.count, n) + n + + def state_dict(self): + return { + "val": self.val, + "sum": self.sum, + "count": self.count, + "round": self.round, + } + + def load_state_dict(self, state_dict): + self.val = state_dict["val"] + self.sum = state_dict["sum"] + self.count = state_dict["count"] + self.round = state_dict.get("round", None) + + @property + def avg(self): + return self.sum / self.count if self.count > 0 else self.val + + @property + def smoothed_value(self) -> float: + val = self.avg + if self.round is not None and val is not None: + val = safe_round(val, self.round) + return val + + +class SumMeter(Meter): + """Computes and stores the sum""" + + def __init__(self, round: Optional[int] = None): + self.round = round + self.reset() + + def reset(self): + self.sum = 0 # sum from all updates + + def update(self, val): + if val is not None: + self.sum = type_as(self.sum, val) + val + + def state_dict(self): + return { + "sum": self.sum, + "round": self.round, + } + + def load_state_dict(self, state_dict): + self.sum = state_dict["sum"] + self.round = state_dict.get("round", None) + + @property + def smoothed_value(self) -> float: + val = self.sum + if self.round is not None and val is not None: + val = safe_round(val, self.round) + return val + + +class TimeMeter(Meter): + """Computes the average occurrence of some event per second""" + + def __init__( + self, + init: int = 0, + n: int = 0, + round: Optional[int] = None, + ): + self.round = round + self.reset(init, n) + + def reset(self, init=0, n=0): + self.init = init + self.start = time.perf_counter() + self.n = n + self.i = 0 + + def update(self, val=1): + self.n = type_as(self.n, val) + val + self.i += 1 + + def state_dict(self): + return { + "init": self.elapsed_time, + "n": self.n, + "round": self.round, + } + + def load_state_dict(self, state_dict): + if "start" in state_dict: + # backwards compatibility for old state_dicts + self.reset(init=state_dict["init"]) + else: + self.reset(init=state_dict["init"], n=state_dict["n"]) + self.round = state_dict.get("round", None) + + @property + def avg(self): + return self.n / self.elapsed_time + + @property + def elapsed_time(self): + return self.init + (time.perf_counter() - self.start) + + @property + def smoothed_value(self) -> float: + val = self.avg + if self.round is not None and val is not None: + val = safe_round(val, self.round) + return val + + +class StopwatchMeter(Meter): + """Computes the sum/avg duration of some event in seconds""" + + def __init__(self, round: Optional[int] = None): + self.round = round + self.sum = 0 + self.n = 0 + self.start_time = None + + def start(self): + self.start_time = time.perf_counter() + + def stop(self, n=1, prehook=None): + if self.start_time is not None: + if prehook is not None: + prehook() + delta = time.perf_counter() - self.start_time + self.sum = self.sum + delta + self.n = type_as(self.n, n) + n + + def reset(self): + self.sum = 0 # cumulative time during which stopwatch was active + self.n = 0 # total n across all start/stop + self.start() + + def state_dict(self): + return { + "sum": self.sum, + "n": self.n, + "round": self.round, + } + + def load_state_dict(self, state_dict): + self.sum = state_dict["sum"] + self.n = state_dict["n"] + self.start_time = None + self.round = state_dict.get("round", None) + + @property + def avg(self): + return self.sum / self.n if self.n > 0 else self.sum + + @property + def elapsed_time(self): + if self.start_time is None: + return 0.0 + return time.perf_counter() - self.start_time + + @property + def smoothed_value(self) -> float: + val = self.avg if self.sum > 0 else self.elapsed_time + if self.round is not None and val is not None: + val = safe_round(val, self.round) + return val + + +class MetersDict(OrderedDict): + """A sorted dictionary of :class:`Meters`. + + Meters are sorted according to a priority that is given when the + meter is first added to the dictionary. + """ + + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + self.priorities = [] + + def __setitem__(self, key, value): + assert key not in self, "MetersDict doesn't support reassignment" + priority, value = value + bisect.insort(self.priorities, (priority, len(self.priorities), key)) + super().__setitem__(key, value) + for _, _, key in self.priorities: # reorder dict to match priorities + self.move_to_end(key) + + def add_meter(self, key, meter, priority): + self.__setitem__(key, (priority, meter)) + + def state_dict(self): + return [ + (pri, key, self[key].__class__.__name__, self[key].state_dict()) + for pri, _, key in self.priorities + # can't serialize DerivedMeter instances + if not isinstance(self[key], MetersDict._DerivedMeter) + ] + + def load_state_dict(self, state_dict): + self.clear() + self.priorities.clear() + for pri, key, meter_cls, meter_state in state_dict: + meter = globals()[meter_cls]() + meter.load_state_dict(meter_state) + self.add_meter(key, meter, pri) + + def get_smoothed_value(self, key: str) -> float: + """Get a single smoothed value.""" + meter = self[key] + if isinstance(meter, MetersDict._DerivedMeter): + return meter.fn(self) + else: + return meter.smoothed_value + + def get_smoothed_values(self) -> Dict[str, float]: + """Get all smoothed values.""" + return OrderedDict( + [ + (key, self.get_smoothed_value(key)) + for key in self.keys() + if not key.startswith("_") + ] + ) + + def reset(self): + """Reset Meter instances.""" + for meter in self.values(): + if isinstance(meter, MetersDict._DerivedMeter): + continue + meter.reset() + + class _DerivedMeter(Meter): + """A Meter whose values are derived from other Meters.""" + + def __init__(self, fn): + self.fn = fn + + def reset(self): + pass diff --git a/fairseq/logging/metrics.py b/fairseq/logging/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..58c2fb64e186ed9d5e9a06c73194d98a21bb7560 --- /dev/null +++ b/fairseq/logging/metrics.py @@ -0,0 +1,314 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +A standalone module for aggregating metrics. + +Metrics can be logged from anywhere using the `log_*` functions defined +in this module. The logged values will be aggregated dynamically based +on the aggregation context in which the logging occurs. See the +:func:`aggregate` context manager for more details. +""" + +import contextlib +import uuid +from collections import defaultdict +from typing import Callable, List, Optional + +from .meters import * + + +# Aggregation contexts are considered "active" when inside the scope +# created by the :func:`aggregate` context manager. +_aggregators = OrderedDict() +_active_aggregators = OrderedDict() +_active_aggregators_cnt = defaultdict(lambda: 0) + + +def reset() -> None: + """Reset all metrics aggregators.""" + _aggregators.clear() + _active_aggregators.clear() + _active_aggregators_cnt.clear() + + # The "default" aggregator observes all logged values. + _aggregators["default"] = MetersDict() + _active_aggregators["default"] = _aggregators["default"] + _active_aggregators_cnt["default"] = 1 + + +reset() + + +@contextlib.contextmanager +def aggregate(name: Optional[str] = None, new_root: bool = False): + """Context manager to aggregate metrics under a given name. + + Aggregations can be nested. If *new_root* is ``False``, then logged + metrics will be recorded along the entire stack of nested + aggregators, including a global "default" aggregator. If *new_root* + is ``True``, then this aggregator will be the root of a new + aggregation stack, thus bypassing any parent aggregators. + + Note that aggregation contexts are uniquely identified by their + *name* (e.g., train, valid). Creating a context with an existing + name will reuse the corresponding :class:`MetersDict` instance. + If no name is given, then a temporary aggregator will be created. + + Usage:: + + with metrics.aggregate("train"): + for step, batch in enumerate(epoch): + with metrics.aggregate("train_inner") as agg: + metrics.log_scalar("loss", get_loss(batch)) + if step % log_interval == 0: + print(agg.get_smoothed_value("loss")) + agg.reset() + print(metrics.get_smoothed_values("train")["loss"]) + + Args: + name (str): name of the aggregation. Defaults to a + random/temporary name if not given explicitly. + new_root (bool): make this aggregation the root of a new + aggregation stack. + """ + if name is None: + # generate a temporary name + name = str(uuid.uuid4()) + assert name not in _aggregators + agg = MetersDict() + else: + assert name != "default" + agg = _aggregators.setdefault(name, MetersDict()) + + if new_root: + backup_aggregators = _active_aggregators.copy() + _active_aggregators.clear() + backup_aggregators_cnt = _active_aggregators_cnt.copy() + _active_aggregators_cnt.clear() + + _active_aggregators[name] = agg + _active_aggregators_cnt[name] += 1 + + yield agg + + _active_aggregators_cnt[name] -= 1 + if _active_aggregators_cnt[name] == 0 and name in _active_aggregators: + del _active_aggregators[name] + + if new_root: + _active_aggregators.clear() + _active_aggregators.update(backup_aggregators) + _active_aggregators_cnt.clear() + _active_aggregators_cnt.update(backup_aggregators_cnt) + + +def get_active_aggregators() -> List[MetersDict]: + return list(_active_aggregators.values()) + + +def log_scalar( + key: str, + value: float, + weight: float = 1, + priority: int = 10, + round: Optional[int] = None, +): + """Log a scalar value. + + Args: + key (str): name of the field to log + value (float): value to log + weight (float): weight that this value contributes to the average. + A weight of 0 will always log the latest value. + priority (int): smaller values are logged earlier in the output + round (Optional[int]): number of digits to round to when displaying + """ + for agg in get_active_aggregators(): + if key not in agg: + agg.add_meter(key, AverageMeter(round=round), priority) + agg[key].update(value, weight) + +def log_scalar_sum( + key: str, + value: float, + priority: int = 10, + round: Optional[int] = None, +): + """Log a scalar value that is summed for reporting. + + Args: + key (str): name of the field to log + value (float): value to log + priority (int): smaller values are logged earlier in the output + round (Optional[int]): number of digits to round to when displaying + """ + for agg in get_active_aggregators(): + if key not in agg: + agg.add_meter(key, SumMeter(round=round), priority) + agg[key].update(value) + + +def log_derived(key: str, fn: Callable[[MetersDict], float], priority: int = 20): + """Log a scalar value derived from other meters. + + Args: + key (str): name of the field to log + fn (Callable[[MetersDict], float]): function that takes a single + argument *meters* and returns the derived value + priority (int): smaller values are logged earlier in the output + """ + for agg in get_active_aggregators(): + if key not in agg: + agg.add_meter(key, MetersDict._DerivedMeter(fn), priority) + + +def log_speed( + key: str, + value: float, + priority: int = 30, + round: Optional[int] = None, +): + """Log the rate of some quantity per second. + + Args: + key (str): name of the field to log + value (float): value to log + priority (int): smaller values are logged earlier in the output + round (Optional[int]): number of digits to round to when displaying + """ + for agg in get_active_aggregators(): + if key not in agg: + agg.add_meter(key, TimeMeter(round=round), priority) + agg[key].reset() # reset meter on the first call + else: + agg[key].update(value) + + +def log_start_time(key: str, priority: int = 40, round: Optional[int] = None): + """Log the duration of some event in seconds. + + The duration will be computed once :func:`log_stop_time` is called. + + Args: + key (str): name of the field to log + priority (int): smaller values are logged earlier in the output + round (Optional[int]): number of digits to round to when displaying + """ + for agg in get_active_aggregators(): + if key not in agg: + agg.add_meter(key, StopwatchMeter(round=round), priority) + agg[key].start() + + +def log_stop_time(key: str, weight: float = 0.0, prehook=None): + """Log the duration of some event in seconds. + + The duration will be computed since :func:`log_start_time` was called. + Set weight > 0 to report the average time instead of the sum. + + Args: + key (str): name of the field to log + weight (float): weight that this time contributes to the average + prehook (function, no arguments): will be called before the timer + is stopped. For example, use prehook=torch.cuda.synchronize to + make sure all gpu operations are done before timer is stopped. + """ + for agg in get_active_aggregators(): + if key in agg: + agg[key].stop(weight, prehook) + + +def log_custom( + new_meter_fn: Callable[[], Meter], + key: str, + *args, + priority: int = 50, + **kwargs, +): + """Log using a custom Meter. + + Any extra *args* or *kwargs* will be passed through to the Meter's + *update* method. + + Args: + new_meter_fn (Callable[[], Meter]): function that returns a new + Meter instance + key (str): name of the field to log + priority (int): smaller values are logged earlier in the output + """ + for agg in get_active_aggregators(): + if key not in agg: + agg.add_meter(key, new_meter_fn(), priority) + agg[key].update(*args, **kwargs) + + +def reset_meter(name: str, key: str) -> None: + """Reset Meter instance aggregated under a given *name* and *key*.""" + meter = get_meter(name, key) + if meter is not None: + meter.reset() + + +def reset_meters(name: str) -> None: + """Reset Meter instances aggregated under a given *name*.""" + meters = get_meters(name) + if meters is not None: + meters.reset() + + +def get_meter(name: str, key: str) -> Meter: + """Get a single Meter instance aggregated under *name* and *key*. + + Returns: + Meter or None if no metrics have been logged under *name* and *key*. + """ + if name not in _aggregators: + return None + return _aggregators[name].get(key, None) + + +def get_meters(name: str) -> MetersDict: + """Get Meter instances aggregated under a given *name*. + + Returns: + MetersDict or None if no metrics have been logged under *name*. + """ + return _aggregators.get(name, None) + + +def get_smoothed_value(name: str, key: str) -> float: + """Get a single smoothed value. + + Raises: + KeyError: if no metrics have been logged under *name* and *key*. + """ + return _aggregators[name].get_smoothed_value(key) + + +def get_smoothed_values(name: str) -> Dict[str, float]: + """Get smoothed values aggregated under a given *name*. + + Raises: + KeyError: if no metrics have been logged under *name*. + """ + return _aggregators[name].get_smoothed_values() + + +def state_dict(): + return OrderedDict([(name, agg.state_dict()) for name, agg in _aggregators.items()]) + + +def load_state_dict(state_dict): + for name, agg_state in state_dict.items(): + _aggregators[name] = MetersDict() + _aggregators[name].load_state_dict(agg_state) + + +def xla_metrics_report(): + try: + import torch_xla.debug.metrics as met + print(met.metrics_report()) + except ImportError: + return diff --git a/fairseq/logging/progress_bar.py b/fairseq/logging/progress_bar.py new file mode 100644 index 0000000000000000000000000000000000000000..061082caefe542c5f0f87e04d9472583874126a3 --- /dev/null +++ b/fairseq/logging/progress_bar.py @@ -0,0 +1,490 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +""" +Wrapper around various loggers and progress bars (e.g., tqdm). +""" + +import atexit +import json +import logging +import os +import sys +from collections import OrderedDict +from contextlib import contextmanager +from numbers import Number +from typing import Optional + +import torch + +from .meters import AverageMeter, StopwatchMeter, TimeMeter + + +logger = logging.getLogger(__name__) + + +def progress_bar( + iterator, + log_format: Optional[str] = None, + log_interval: int = 100, + log_file: Optional[str] = None, + epoch: Optional[int] = None, + prefix: Optional[str] = None, + tensorboard_logdir: Optional[str] = None, + default_log_format: str = "tqdm", + wandb_project: Optional[str] = None, + wandb_run_name: Optional[str] = None, + azureml_logging: Optional[bool] = False, +): + if log_format is None: + log_format = default_log_format + if log_file is not None: + handler = logging.FileHandler(filename=log_file) + logger.addHandler(handler) + + if log_format == "tqdm" and not sys.stderr.isatty(): + log_format = "simple" + + if log_format == "json": + bar = JsonProgressBar(iterator, epoch, prefix, log_interval) + elif log_format == "none": + bar = NoopProgressBar(iterator, epoch, prefix) + elif log_format == "simple": + bar = SimpleProgressBar(iterator, epoch, prefix, log_interval) + elif log_format == "tqdm": + bar = TqdmProgressBar(iterator, epoch, prefix) + else: + raise ValueError("Unknown log format: {}".format(log_format)) + + if tensorboard_logdir: + try: + # [FB only] custom wrapper for TensorBoard + import palaas # noqa + from .fb_tbmf_wrapper import FbTbmfWrapper + + bar = FbTbmfWrapper(bar, log_interval) + except ImportError: + bar = TensorboardProgressBarWrapper(bar, tensorboard_logdir) + + if wandb_project: + bar = WandBProgressBarWrapper(bar, wandb_project, run_name=wandb_run_name) + + if azureml_logging: + bar = AzureMLProgressBarWrapper(bar) + + return bar + + +def build_progress_bar( + args, + iterator, + epoch: Optional[int] = None, + prefix: Optional[str] = None, + default: str = "tqdm", + no_progress_bar: str = "none", +): + """Legacy wrapper that takes an argparse.Namespace.""" + if getattr(args, "no_progress_bar", False): + default = no_progress_bar + if getattr(args, "distributed_rank", 0) == 0: + tensorboard_logdir = getattr(args, "tensorboard_logdir", None) + else: + tensorboard_logdir = None + return progress_bar( + iterator, + log_format=args.log_format, + log_interval=args.log_interval, + epoch=epoch, + prefix=prefix, + tensorboard_logdir=tensorboard_logdir, + default_log_format=default, + ) + + +def format_stat(stat): + if isinstance(stat, Number): + stat = "{:g}".format(stat) + elif isinstance(stat, AverageMeter): + stat = "{:.3f}".format(stat.avg) + elif isinstance(stat, TimeMeter): + stat = "{:g}".format(round(stat.avg)) + elif isinstance(stat, StopwatchMeter): + stat = "{:g}".format(round(stat.sum)) + elif torch.is_tensor(stat): + stat = stat.tolist() + return stat + + +class BaseProgressBar(object): + """Abstract class for progress bars.""" + + def __init__(self, iterable, epoch=None, prefix=None): + self.iterable = iterable + self.n = getattr(iterable, "n", 0) + self.epoch = epoch + self.prefix = "" + if epoch is not None: + self.prefix += "epoch {:03d}".format(epoch) + if prefix is not None: + self.prefix += (" | " if self.prefix != "" else "") + prefix + + def __len__(self): + return len(self.iterable) + + def __enter__(self): + return self + + def __exit__(self, *exc): + return False + + def __iter__(self): + raise NotImplementedError + + def log(self, stats, tag=None, step=None): + """Log intermediate stats according to log_interval.""" + raise NotImplementedError + + def print(self, stats, tag=None, step=None): + """Print end-of-epoch stats.""" + raise NotImplementedError + + def update_config(self, config): + """Log latest configuration.""" + pass + + def _str_commas(self, stats): + return ", ".join(key + "=" + stats[key].strip() for key in stats.keys()) + + def _str_pipes(self, stats): + return " | ".join(key + " " + stats[key].strip() for key in stats.keys()) + + def _format_stats(self, stats): + postfix = OrderedDict(stats) + # Preprocess stats according to datatype + for key in postfix.keys(): + postfix[key] = str(format_stat(postfix[key])) + return postfix + + +@contextmanager +def rename_logger(logger, new_name): + old_name = logger.name + if new_name is not None: + logger.name = new_name + yield logger + logger.name = old_name + + +class JsonProgressBar(BaseProgressBar): + """Log output in JSON format.""" + + def __init__(self, iterable, epoch=None, prefix=None, log_interval=1000): + super().__init__(iterable, epoch, prefix) + self.log_interval = log_interval + self.i = None + self.size = None + + def __iter__(self): + self.size = len(self.iterable) + for i, obj in enumerate(self.iterable, start=self.n): + self.i = i + yield obj + + def log(self, stats, tag=None, step=None): + """Log intermediate stats according to log_interval.""" + step = step or self.i or 0 + if step > 0 and self.log_interval is not None and step % self.log_interval == 0: + update = ( + self.epoch - 1 + (self.i + 1) / float(self.size) + if self.epoch is not None + else None + ) + stats = self._format_stats(stats, epoch=self.epoch, update=update) + with rename_logger(logger, tag): + logger.info(json.dumps(stats)) + + def print(self, stats, tag=None, step=None): + """Print end-of-epoch stats.""" + self.stats = stats + if tag is not None: + self.stats = OrderedDict( + [(tag + "_" + k, v) for k, v in self.stats.items()] + ) + stats = self._format_stats(self.stats, epoch=self.epoch) + with rename_logger(logger, tag): + logger.info(json.dumps(stats)) + + def _format_stats(self, stats, epoch=None, update=None): + postfix = OrderedDict() + if epoch is not None: + postfix["epoch"] = epoch + if update is not None: + postfix["update"] = round(update, 3) + # Preprocess stats according to datatype + for key in stats.keys(): + postfix[key] = format_stat(stats[key]) + return postfix + + +class NoopProgressBar(BaseProgressBar): + """No logging.""" + + def __init__(self, iterable, epoch=None, prefix=None): + super().__init__(iterable, epoch, prefix) + + def __iter__(self): + for obj in self.iterable: + yield obj + + def log(self, stats, tag=None, step=None): + """Log intermediate stats according to log_interval.""" + pass + + def print(self, stats, tag=None, step=None): + """Print end-of-epoch stats.""" + pass + + +class SimpleProgressBar(BaseProgressBar): + """A minimal logger for non-TTY environments.""" + + def __init__(self, iterable, epoch=None, prefix=None, log_interval=1000): + super().__init__(iterable, epoch, prefix) + self.log_interval = log_interval + self.i = None + self.size = None + + def __iter__(self): + self.size = len(self.iterable) + for i, obj in enumerate(self.iterable, start=self.n): + self.i = i + yield obj + + def log(self, stats, tag=None, step=None): + """Log intermediate stats according to log_interval.""" + step = step or self.i or 0 + if step > 0 and self.log_interval is not None and step % self.log_interval == 0: + stats = self._format_stats(stats) + postfix = self._str_commas(stats) + with rename_logger(logger, tag): + logger.info( + "{}: {:5d} / {:d} {}".format( + self.prefix, self.i + 1, self.size, postfix + ) + ) + + def print(self, stats, tag=None, step=None): + """Print end-of-epoch stats.""" + postfix = self._str_pipes(self._format_stats(stats)) + with rename_logger(logger, tag): + logger.info("{} | {}".format(self.prefix, postfix)) + + +class TqdmProgressBar(BaseProgressBar): + """Log to tqdm.""" + + def __init__(self, iterable, epoch=None, prefix=None): + super().__init__(iterable, epoch, prefix) + from tqdm import tqdm + + self.tqdm = tqdm( + iterable, + self.prefix, + leave=False, + disable=(logger.getEffectiveLevel() > logging.INFO), + ) + + def __iter__(self): + return iter(self.tqdm) + + def log(self, stats, tag=None, step=None): + """Log intermediate stats according to log_interval.""" + self.tqdm.set_postfix(self._format_stats(stats), refresh=False) + + def print(self, stats, tag=None, step=None): + """Print end-of-epoch stats.""" + postfix = self._str_pipes(self._format_stats(stats)) + with rename_logger(logger, tag): + logger.info("{} | {}".format(self.prefix, postfix)) + + +try: + _tensorboard_writers = {} + from torch.utils.tensorboard import SummaryWriter +except ImportError: + try: + from tensorboardX import SummaryWriter + except ImportError: + SummaryWriter = None + + +def _close_writers(): + for w in _tensorboard_writers.values(): + w.close() + + +atexit.register(_close_writers) + + +class TensorboardProgressBarWrapper(BaseProgressBar): + """Log to tensorboard.""" + + def __init__(self, wrapped_bar, tensorboard_logdir): + self.wrapped_bar = wrapped_bar + self.tensorboard_logdir = tensorboard_logdir + + if SummaryWriter is None: + logger.warning( + "tensorboard not found, please install with: pip install tensorboard" + ) + + def _writer(self, key): + if SummaryWriter is None: + return None + _writers = _tensorboard_writers + if key not in _writers: + _writers[key] = SummaryWriter(os.path.join(self.tensorboard_logdir, key)) + _writers[key].add_text("sys.argv", " ".join(sys.argv)) + return _writers[key] + + def __iter__(self): + return iter(self.wrapped_bar) + + def log(self, stats, tag=None, step=None): + """Log intermediate stats to tensorboard.""" + self._log_to_tensorboard(stats, tag, step) + self.wrapped_bar.log(stats, tag=tag, step=step) + + def print(self, stats, tag=None, step=None): + """Print end-of-epoch stats.""" + self._log_to_tensorboard(stats, tag, step) + self.wrapped_bar.print(stats, tag=tag, step=step) + + def update_config(self, config): + """Log latest configuration.""" + # TODO add hparams to Tensorboard + self.wrapped_bar.update_config(config) + + def _log_to_tensorboard(self, stats, tag=None, step=None): + writer = self._writer(tag or "") + if writer is None: + return + if step is None: + step = stats["num_updates"] + for key in stats.keys() - {"num_updates"}: + if isinstance(stats[key], AverageMeter): + writer.add_scalar(key, stats[key].val, step) + elif isinstance(stats[key], Number): + writer.add_scalar(key, stats[key], step) + elif torch.is_tensor(stats[key]) and stats[key].numel() == 1: + writer.add_scalar(key, stats[key].item(), step) + writer.flush() + + +try: + import wandb +except ImportError: + wandb = None + + +class WandBProgressBarWrapper(BaseProgressBar): + """Log to Weights & Biases.""" + + def __init__(self, wrapped_bar, wandb_project, run_name=None): + self.wrapped_bar = wrapped_bar + if wandb is None: + logger.warning("wandb not found, pip install wandb") + return + + # reinit=False to ensure if wandb.init() is called multiple times + # within one process it still references the same run + wandb.init(project=wandb_project, reinit=False, name=run_name) + + def __iter__(self): + return iter(self.wrapped_bar) + + def log(self, stats, tag=None, step=None): + """Log intermediate stats to tensorboard.""" + self._log_to_wandb(stats, tag, step) + self.wrapped_bar.log(stats, tag=tag, step=step) + + def print(self, stats, tag=None, step=None): + """Print end-of-epoch stats.""" + self._log_to_wandb(stats, tag, step) + self.wrapped_bar.print(stats, tag=tag, step=step) + + def update_config(self, config): + """Log latest configuration.""" + if wandb is not None: + wandb.config.update(config) + self.wrapped_bar.update_config(config) + + def _log_to_wandb(self, stats, tag=None, step=None): + if wandb is None: + return + if step is None: + step = stats["num_updates"] + + prefix = "" if tag is None else tag + "/" + + for key in stats.keys() - {"num_updates"}: + if isinstance(stats[key], AverageMeter): + wandb.log({prefix + key: stats[key].val}, step=step) + elif isinstance(stats[key], Number): + wandb.log({prefix + key: stats[key]}, step=step) + + +try: + from azureml.core import Run +except ImportError: + Run = None + + +class AzureMLProgressBarWrapper(BaseProgressBar): + """Log to Azure ML""" + + def __init__(self, wrapped_bar): + self.wrapped_bar = wrapped_bar + if Run is None: + logger.warning("azureml.core not found, pip install azureml-core") + return + self.run = Run.get_context() + + def __exit__(self, *exc): + if Run is not None: + self.run.complete() + return False + + def __iter__(self): + return iter(self.wrapped_bar) + + def log(self, stats, tag=None, step=None): + """Log intermediate stats to AzureML""" + self._log_to_azureml(stats, tag, step) + self.wrapped_bar.log(stats, tag=tag, step=step) + + def print(self, stats, tag=None, step=None): + """Print end-of-epoch stats""" + self._log_to_azureml(stats, tag, step) + self.wrapped_bar.print(stats, tag=tag, step=step) + + def update_config(self, config): + """Log latest configuration.""" + self.wrapped_bar.update_config(config) + + def _log_to_azureml(self, stats, tag=None, step=None): + if Run is None: + return + if step is None: + step = stats["num_updates"] + + prefix = "" if tag is None else tag + "/" + + for key in stats.keys() - {"num_updates"}: + name = prefix + key + if isinstance(stats[key], AverageMeter): + self.run.log_row(name=name, **{"step": step, key: stats[key].val}) + elif isinstance(stats[key], Number): + self.run.log_row(name=name, **{"step": step, key: stats[key]}) diff --git a/fairseq/model_parallel/__init__.py b/fairseq/model_parallel/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..69f21684872f72ae8ee26d9ff7d2d2b6e6d526c3 --- /dev/null +++ b/fairseq/model_parallel/__init__.py @@ -0,0 +1,6 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from . import criterions, models, modules # noqa diff --git a/fairseq/model_parallel/criterions/__init__.py b/fairseq/model_parallel/criterions/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..5fae7bd4c2cfa7b4f64ad62dd9b9082f59f0e50d --- /dev/null +++ b/fairseq/model_parallel/criterions/__init__.py @@ -0,0 +1,14 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import importlib +import os + + +# automatically import any Python files in the criterions/ directory +for file in sorted(os.listdir(os.path.dirname(__file__))): + if file.endswith(".py") and not file.startswith("_"): + module = file[: file.find(".py")] + importlib.import_module("fairseq.model_parallel.criterions." + module) diff --git a/fairseq/model_parallel/criterions/vocab_parallel_cross_entropy.py b/fairseq/model_parallel/criterions/vocab_parallel_cross_entropy.py new file mode 100644 index 0000000000000000000000000000000000000000..35c50ee1521963c5cb6dfb7036ccf43401c6c6ac --- /dev/null +++ b/fairseq/model_parallel/criterions/vocab_parallel_cross_entropy.py @@ -0,0 +1,87 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +from fairseq import metrics, utils +from fairseq.criterions import FairseqCriterion, register_criterion + + +try: + from fairseq.model_parallel.megatron.mpu.cross_entropy import ( + vocab_parallel_cross_entropy, + ) + + has_megatron_submodule = True +except (ImportError, ModuleNotFoundError): + has_megatron_submodule = False + + +@register_criterion("vocab_parallel_cross_entropy") +class VocabParallelCrossEntropyCriterion(FairseqCriterion): + def __init__(self, task, sentence_avg): + super().__init__(task) + self.sentence_avg = sentence_avg + if not has_megatron_submodule: + raise ImportError( + "\n\nPlease install the megatron submodule:" + "\n\n git submodule update --init " + "fairseq/model_parallel/megatron" + ) + + def forward(self, model, sample, reduce=True): + """Compute the loss for the given sample. + + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + net_output = model(**sample["net_input"]) + target = sample["target"] + + loss = vocab_parallel_cross_entropy(net_output[0].float(), target) + loss = (loss * (target != self.padding_idx)).sum() + sample_size = ( + sample["target"].size(0) if self.sentence_avg else sample["ntokens"] + ) + logging_output = { + "loss": utils.item(loss.data) if reduce else loss.data, + "ntokens": sample["ntokens"], + "nsentences": sample["target"].size(0), + "sample_size": sample_size, + } + return loss, sample_size, logging_output + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + """Aggregate logging outputs from data parallel training.""" + loss_sum = sum(log.get("loss", 0) for log in logging_outputs) + ntokens = sum(log.get("ntokens", 0) for log in logging_outputs) + sample_size = sum(log.get("sample_size", 0) for log in logging_outputs) + + metrics.log_scalar( + "loss", loss_sum / sample_size / math.log(2), sample_size, round=3 + ) + if sample_size != ntokens: + metrics.log_scalar( + "nll_loss", loss_sum / ntokens / math.log(2), ntokens, round=3 + ) + metrics.log_derived( + "ppl", lambda meters: utils.get_perplexity(meters["nll_loss"].avg) + ) + else: + metrics.log_derived( + "ppl", lambda meters: utils.get_perplexity(meters["loss"].avg) + ) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return True diff --git a/fairseq/model_parallel/megatron_trainer.py b/fairseq/model_parallel/megatron_trainer.py new file mode 100644 index 0000000000000000000000000000000000000000..8ab4657f73c6cda91e95637921edb84ccb76b3d0 --- /dev/null +++ b/fairseq/model_parallel/megatron_trainer.py @@ -0,0 +1,71 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +""" +Train a network across multiple GPUs. +""" + +from fairseq.dataclass.configs import FairseqConfig +from fairseq.distributed import utils as distributed_utils +from fairseq.trainer import Trainer + +try: + from fairseq.model_parallel.megatron.mpu import ( + get_data_parallel_rank, + get_data_parallel_world_size, + get_model_parallel_src_rank, + get_cuda_rng_tracker, + ) + + has_megatron_submodule = True +except (ImportError, ModuleNotFoundError): + has_megatron_submodule = False + + +class MegatronTrainer(Trainer): + """Main class for model parallel with data parallel training.""" + + def __init__(self, cfg: FairseqConfig, task, model, criterion, **kwargs): + if not has_megatron_submodule: + raise ImportError( + "\n\nPlease install the megatron submodule:" + "\n\n git submodule update --init " + "fairseq/model_parallel/megatron" + ) + super().__init__(cfg, task, model, criterion, **kwargs) + + def clip_grad_norm(self, clip_norm): + def _aggregate_model_parallel_grad_norm(total_norm): + total_norm = total_norm ** 2 + distributed_utils.all_reduce( + total_norm, group=distributed_utils.get_model_parallel_group() + ) + total_norm = total_norm ** 0.5 + return total_norm + + return self.optimizer.clip_grad_norm( + clip_norm, + aggregate_norm_fn=_aggregate_model_parallel_grad_norm, + ) + + def save_checkpoint(self, filename, extra_state): + """Save all training state in a checkpoint file.""" + extra_state['rng_tracker_states'] \ + = get_cuda_rng_tracker().get_states() + super().save_checkpoint(filename, extra_state) + + def load_checkpoint( + self, + filename, + reset_optimizer=False, + reset_lr_scheduler=False, + optimizer_overrides=None, + reset_meters=False, + ): + extra_state = super().load_checkpoint(filename, reset_optimizer=reset_optimizer, reset_lr_scheduler=reset_lr_scheduler, optimizer_overrides=optimizer_overrides, reset_meters=reset_meters) + if extra_state is not None and 'rng_tracker_states' in extra_state: + get_cuda_rng_tracker().set_states( + extra_state['rng_tracker_states']) + return extra_state diff --git a/fairseq/model_parallel/models/__init__.py b/fairseq/model_parallel/models/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..3532479e52a0e1f1ba204c6f5d51c71c98ee5df0 --- /dev/null +++ b/fairseq/model_parallel/models/__init__.py @@ -0,0 +1,20 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import importlib +import os + + +# automatically import any Python files in the models/ directory +models_dir = os.path.dirname(__file__) +for file in os.listdir(models_dir): + path = os.path.join(models_dir, file) + if ( + not file.startswith("_") + and not file.startswith(".") + and (file.endswith(".py") or os.path.isdir(path)) + ): + model_name = file[: file.find(".py")] if file.endswith(".py") else file + module = importlib.import_module("fairseq.model_parallel.models." + model_name) diff --git a/fairseq/model_parallel/models/pipeline_parallel_transformer/__init__.py b/fairseq/model_parallel/models/pipeline_parallel_transformer/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..117827c3e9c176477f33e3a6fd7fe19a922411a2 --- /dev/null +++ b/fairseq/model_parallel/models/pipeline_parallel_transformer/__init__.py @@ -0,0 +1,6 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .model import * # noqa diff --git a/fairseq/model_parallel/models/pipeline_parallel_transformer/layers.py b/fairseq/model_parallel/models/pipeline_parallel_transformer/layers.py new file mode 100644 index 0000000000000000000000000000000000000000..eb81ded341257ba0a43c4d0867e8f3c83f276bc7 --- /dev/null +++ b/fairseq/model_parallel/models/pipeline_parallel_transformer/layers.py @@ -0,0 +1,600 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +from collections import namedtuple + +import torch +import torch.nn as nn +import torch.nn.functional as F +from fairseq import options, utils +from fairseq.modules import ( + AdaptiveSoftmax, + LayerNorm, + MultiheadAttention, + PositionalEmbedding, +) + + +EncoderOut = namedtuple( + "TransformerEncoderOut", + [ + "encoder_out", # T x B x C + "encoder_padding_mask", # B x T + "encoder_embedding", # B x T x C + "encoder_states", # List[T x B x C] + ], +) + + +class TransformerEncoderEmbedding(nn.Module): + """ Encoder Embedding + Positional Embedding """ + + def __init__(self, args, embed_tokens): + super().__init__() + self.dropout = args.dropout + self.max_source_positions = args.max_source_positions + self.embed_tokens = embed_tokens + if isinstance(embed_tokens, nn.ModuleList): + self.padding_idx = embed_tokens[0].padding_idx + embed_dim = sum(e.embedding_dim for e in embed_tokens) + else: + self.padding_idx = embed_tokens.padding_idx + embed_dim = embed_tokens.embedding_dim + self.embed_scale = math.sqrt(embed_dim) + self.embed_positions = ( + PositionalEmbedding( + args.max_source_positions, + embed_dim, + self.padding_idx, + learned=args.encoder_learned_pos, + ) + if not args.no_token_positional_embeddings + else None + ) + if getattr(args, "layernorm_embedding", False): + self.layernorm_embedding = LayerNorm(embed_dim) + else: + self.layernorm_embedding = None + + def forward(self, input): + # embed tokens and positions + src_tokens = input[0] + prev_output_tokens = input[2] + if isinstance(self.embed_tokens, nn.ModuleList): + x_embed_list = [] + for embed_tokens_part in self.embed_tokens: + x_embed_list.append(embed_tokens_part(src_tokens)) + + embedded = torch.cat(x_embed_list, dim=-1) + else: + embedded = self.embed_tokens(src_tokens) + x = embed = self.embed_scale * embedded + if self.embed_positions is not None: + x = embed + self.embed_positions(src_tokens) + if self.layernorm_embedding: + x = self.layernorm_embedding(x) + x = F.dropout(x, p=self.dropout, training=self.training) + # B x T x C -> T x B x C + x = x.transpose(0, 1) + + # compute padding mask + encoder_padding_mask = src_tokens.eq(self.padding_idx) + return (x, encoder_padding_mask, prev_output_tokens) + + +class TransformerEncoderLayerNorm(nn.Module): + """ + Layer norm at the the end of all encoder layers if + args.encoder_enormalize_before = True + """ + + def __init__(self, args, embed_dim): + super().__init__() + if args.encoder_normalize_before: + self.layer_norm = LayerNorm(embed_dim) + else: + self.layer_norm = None + + def forward(self, input): + x = input[0] + encoder_padding_mask = input[1] + prev_output_tokens = input[2] + if self.layer_norm: + x = self.layer_norm(x) + # keeping track of the incremental_state is not supported yet + return (x, encoder_padding_mask, prev_output_tokens) + + +class TransformerDecoderEmbedding(nn.Module): + """ Decoder Embedding + Positional Embedding """ + + def __init__(self, args, embed_tokens): + super().__init__() + self.dropout = args.dropout + self.share_input_output_embed = args.share_decoder_input_output_embed + input_embed_dim = ( + sum(e.embedding_dim for e in embed_tokens) + if isinstance(embed_tokens, nn.ModuleList) + else embed_tokens.embedding_dim + ) + embed_dim = args.decoder_embed_dim + self.output_embed_dim = args.decoder_output_dim + + padding_idx = ( + embed_tokens[0].padding_idx + if isinstance(embed_tokens, nn.ModuleList) + else embed_tokens.padding_idx + ) + self.max_target_positions = args.max_target_positions + + self.embed_tokens = embed_tokens + self.embed_scale = math.sqrt(embed_dim) # todo: try with input_embed_dim + + self.project_in_dim = ( + Linear(input_embed_dim, embed_dim, bias=False) + if embed_dim != input_embed_dim + else None + ) + + self.embed_positions = ( + PositionalEmbedding( + args.max_target_positions, + embed_dim, + padding_idx, + learned=args.decoder_learned_pos, + ) + if not args.no_token_positional_embeddings + else None + ) + + def forward(self, input): + mt_task = False + if isinstance(input, tuple): + if len(input) == 3: + encoder_out = input[0] + encoder_padding_mask = input[1] + prev_output_tokens = input[2] + incremental_state = None # Hardcoding to avoid passing of None objects + mt_task = True + else: + # HACK for now, need to fix (TODO sidgoyal) + prev_output_tokens = input[0] + # discard "src_lengths" + encoder_out = None + encoder_padding_mask = None + incremental_state = None + + else: + prev_output_tokens = input + encoder_out = None + encoder_padding_mask = None + incremental_state = None + + positions = ( + self.embed_positions( + prev_output_tokens, + incremental_state=incremental_state, + ) + if self.embed_positions is not None + else None + ) + + if incremental_state is not None: + prev_output_tokens = prev_output_tokens[:, -1:] + if positions is not None: + positions = positions[:, -1:] + + # embed tokens and positions + + if isinstance(self.embed_tokens, nn.ModuleList): + x_embed_list = [] + for embed_tokens_part in self.embed_tokens: + x_embed_list.append(embed_tokens_part(prev_output_tokens)) + + x = self.embed_scale * torch.cat(x_embed_list, dim=-1) + else: + x = self.embed_scale * self.embed_tokens(prev_output_tokens) + + if self.project_in_dim is not None: + x = self.project_in_dim(x) + + if positions is not None: + x += positions + x = F.dropout(x, p=self.dropout, training=self.training) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + if mt_task: + return (x, encoder_out, encoder_padding_mask) + return x + + +class TransformerDecoderOutputLayer(nn.Module): + def __init__(self, args, embed_tokens, dictionary): + super().__init__() + self.share_input_output_embed = args.share_decoder_input_output_embed + self.embed_tokens = embed_tokens + self.output_embed_dim = args.decoder_output_dim + embed_dim = args.decoder_embed_dim + + self.project_out_dim = ( + Linear(embed_dim, self.output_embed_dim, bias=False) + if embed_dim != self.output_embed_dim and not args.tie_adaptive_weights + else None + ) + self.adaptive_softmax = None + if args.adaptive_softmax_cutoff is not None: + assert not isinstance(embed_tokens, nn.ModuleList) + self.adaptive_softmax = AdaptiveSoftmax( + len(dictionary), + self.output_embed_dim, + options.eval_str_list(args.adaptive_softmax_cutoff, type=int), + dropout=args.adaptive_softmax_dropout, + adaptive_inputs=embed_tokens if args.tie_adaptive_weights else None, + factor=args.adaptive_softmax_factor, + tie_proj=args.tie_adaptive_proj, + ) + elif not self.share_input_output_embed: + self.embed_tokens = nn.Parameter( + torch.Tensor(len(dictionary), self.output_embed_dim) + ) + nn.init.normal_( + self.embed_tokens, mean=0, std=self.output_embed_dim ** -0.5 + ) + + if args.decoder_normalize_before and not getattr( + args, "no_decoder_final_norm", False + ): + self.layer_norm = LayerNorm(embed_dim) + else: + self.layer_norm = None + + def forward(self, input, apply_final_proj=True): + if isinstance(input, tuple): + x = input[0] + else: + x = input + + if self.layer_norm: + x = self.layer_norm(x) + + # T x B x C -> B x T x C + x = x.transpose(0, 1) + + if self.project_out_dim is not None: + x = self.project_out_dim(x) + if apply_final_proj: + x = self.output_layer(x) + return x + + def output_layer(self, features, **kwargs): + """Project features to the vocabulary size.""" + if self.adaptive_softmax is None: + # project back to size of vocabulary + if self.share_input_output_embed: + if isinstance(self.embed_tokens, nn.ModuleList): + output = None + for i, emb in enumerate(self.embed_tokens): + sidx = i * emb.embedding_dim + eidx = (i + 1) * emb.embedding_dim + if output is None: + output = F.linear(features[:, :, sidx:eidx], emb.weight) + else: + output += F.linear(features[:, :, sidx:eidx], emb.weight) + + return output + else: + return F.linear(features, self.embed_tokens.weight) + else: + return F.linear(features, self.embed_tokens) + else: + return features + + +class TransformerEncoderLayer(nn.Module): + """Encoder layer block. + In the original paper each operation (multi-head attention or FFN) is + postprocessed with: `dropout -> add residual -> layernorm`. In the + tensor2tensor code they suggest that learning is more robust when + preprocessing each layer with layernorm and postprocessing with: + `dropout -> add residual`. We default to the approach in the paper, but the + tensor2tensor approach can be enabled by setting + *args.encoder_normalize_before* to ``True``. + + Args: + args (argparse.Namespace): parsed command-line arguments + """ + + def __init__(self, args): + super().__init__() + self.embed_dim = args.encoder_embed_dim + self.self_attn = MultiheadAttention( + self.embed_dim, + args.encoder_attention_heads, + dropout=args.attention_dropout, + self_attention=True, + ) + self.self_attn_layer_norm = LayerNorm(self.embed_dim) + self.dropout = args.dropout + self.activation_fn = utils.get_activation_fn( + activation=getattr(args, "activation_fn", "relu") + ) + self.activation_dropout = getattr(args, "activation_dropout", 0) + if self.activation_dropout == 0: + # for backwards compatibility with models that use args.relu_dropout + self.activation_dropout = getattr(args, "relu_dropout", 0) + self.normalize_before = args.encoder_normalize_before + self.fc1 = Linear(self.embed_dim, args.encoder_ffn_embed_dim) + self.fc2 = Linear(args.encoder_ffn_embed_dim, self.embed_dim) + self.final_layer_norm = LayerNorm(self.embed_dim) + + def upgrade_state_dict_named(self, state_dict, name): + """ + Rename layer norm states from `...layer_norms.0.weight` to + `...self_attn_layer_norm.weight` and `...layer_norms.1.weight` to + `...final_layer_norm.weight` + """ + layer_norm_map = {"0": "self_attn_layer_norm", "1": "final_layer_norm"} + for old, new in layer_norm_map.items(): + for m in ("weight", "bias"): + k = "{}.layer_norms.{}.{}".format(name, old, m) + if k in state_dict: + state_dict["{}.{}.{}".format(name, new, m)] = state_dict[k] + del state_dict[k] + + def forward(self, input): + """ + Args: + input (Tuple): + input[0] (Tensor): input to the layer of shape `(seq_len, batch, embed_dim)` + input[1] (ByteTensor/FloatTensor): encoder padding mask - + binary ByteTensor of shape `(batch, src_len)` where padding elements + are indicated by ``1``. + input[2] (LongTensor): previous decoder outputs of shape + `(batch, tgt_len)`, for teacher forcing) + Returns: + output (Tuple): + output[0] (Tensor): encoded output of shape `(batch, src_len, embed_dim)` + output[1] (ByteTensor/FloatTensor): encoder padding mask + output[2] (LongTensor): previous decoder outputs + """ + x = input[0] + encoder_padding_mask = input[1] + prev_output_tokens = input[2] + residual = x + x = self.maybe_layer_norm(self.self_attn_layer_norm, x, before=True) + x, _ = self.self_attn( + query=x, key=x, value=x, key_padding_mask=encoder_padding_mask + ) + x = F.dropout(x, p=self.dropout, training=self.training) + x = residual + x + x = self.maybe_layer_norm(self.self_attn_layer_norm, x, after=True) + + residual = x + x = self.maybe_layer_norm(self.final_layer_norm, x, before=True) + x = self.activation_fn(self.fc1(x)) + x = F.dropout(x, p=self.activation_dropout, training=self.training) + x = self.fc2(x) + x = F.dropout(x, p=self.dropout, training=self.training) + x = residual + x + x = self.maybe_layer_norm(self.final_layer_norm, x, after=True) + return (x, encoder_padding_mask, prev_output_tokens) + + def maybe_layer_norm(self, layer_norm, x, before=False, after=False): + assert before ^ after + if after ^ self.normalize_before: + return layer_norm(x) + else: + return x + + +class TransformerDecoderLayer(nn.Module): + """Decoder layer block. + + In the original paper each operation (multi-head attention, encoder + attention or FFN) is postprocessed with: `dropout -> add residual -> + layernorm`. In the tensor2tensor code they suggest that learning is more + robust when preprocessing each layer with layernorm and postprocessing with: + `dropout -> add residual`. We default to the approach in the paper, but the + tensor2tensor approach can be enabled by setting + *args.decoder_normalize_before* to ``True``. + + Args: + args (argparse.Namespace): parsed command-line arguments + no_encoder_attn (bool, optional): whether to attend to encoder outputs + (default: False). + """ + + def __init__( + self, args, no_encoder_attn=False, add_bias_kv=False, add_zero_attn=False + ): + super().__init__() + self.embed_dim = args.decoder_embed_dim + self.self_attn = MultiheadAttention( + embed_dim=self.embed_dim, + num_heads=args.decoder_attention_heads, + dropout=args.attention_dropout, + add_bias_kv=add_bias_kv, + add_zero_attn=add_zero_attn, + self_attention=True, + ) + self.dropout = args.dropout + self.activation_fn = utils.get_activation_fn( + activation=getattr(args, "activation_fn", "relu") + ) + self.activation_dropout = getattr(args, "activation_dropout", 0) + if self.activation_dropout == 0: + # for backwards compatibility with models that use args.relu_dropout + self.activation_dropout = getattr(args, "relu_dropout", 0) + self.normalize_before = args.decoder_normalize_before + + # use layerNorm rather than FusedLayerNorm for exporting. + # char_inputs can be used to determint this. + # TODO remove this once we update apex with the fix + export = getattr(args, "char_inputs", False) + self.self_attn_layer_norm = LayerNorm(self.embed_dim, export=export) + + if no_encoder_attn: + self.encoder_attn = None + self.encoder_attn_layer_norm = None + else: + self.encoder_attn = MultiheadAttention( + self.embed_dim, + args.decoder_attention_heads, + kdim=getattr(args, "encoder_embed_dim", None), + vdim=getattr(args, "encoder_embed_dim", None), + dropout=args.attention_dropout, + encoder_decoder_attention=True, + ) + self.encoder_attn_layer_norm = LayerNorm(self.embed_dim, export=export) + + self.fc1 = Linear(self.embed_dim, args.decoder_ffn_embed_dim) + self.fc2 = Linear(args.decoder_ffn_embed_dim, self.embed_dim) + + self.final_layer_norm = LayerNorm(self.embed_dim, export=export) + self.need_attn = True + + self.onnx_trace = False + + def prepare_for_onnx_export_(self): + self.onnx_trace = True + + def forward(self, input): + """ + Args: + input (Tuple): + input[0] (Tensor): input to the layer of shape `(seq_len, batch, embed_dim)` + input[1] (Tensor): encoder output of shape `(batch, src_len, embed_dim)` + input[2] (ByteTensor/FloatTensor): encoder padding mask - + binary ByteTensor of shape `(batch, src_len)` where padding elements + are indicated by ``1``. + Returns: + output (Tuple): + output[0] (Tensor): encoded output of shape `(batch, src_len, embed_dim)` + output[1] (ByteTensor/FloatTensor): encoder padding mask + output[2] (LongTensor): previous decoder outputs + """ + # Note: incremental state is not yet supported + mt_task = False + if isinstance(input, tuple): + x = input[0] + encoder_out = input[1] + encoder_padding_mask = input[2] + incremental_state = None + mt_task = True + else: + x = input + encoder_out = None + encoder_padding_mask = None + incremental_state = None + + if incremental_state is None: + self_attn_mask = self.buffered_future_mask(x) + else: + self_attn_mask = None + + # TODO: add back prev_self_attn_state, prev_attn_state, + # self_attn_padding_mask + prev_self_attn_state = None + prev_attn_state = None + self_attn_padding_mask = None + + residual = x + x = self.maybe_layer_norm(self.self_attn_layer_norm, x, before=True) + if prev_self_attn_state is not None: + if incremental_state is None: + incremental_state = {} + prev_key, prev_value = prev_self_attn_state + saved_state = {"prev_key": prev_key, "prev_value": prev_value} + self.self_attn._set_input_buffer(incremental_state, saved_state) + x, attn = self.self_attn( + query=x, + key=x, + value=x, + key_padding_mask=self_attn_padding_mask, + incremental_state=incremental_state, + need_weights=False, + attn_mask=self_attn_mask, + ) + x = F.dropout(x, p=self.dropout, training=self.training) + x = residual + x + x = self.maybe_layer_norm(self.self_attn_layer_norm, x, after=True) + + if self.encoder_attn is not None: + residual = x + x = self.maybe_layer_norm(self.encoder_attn_layer_norm, x, before=True) + if prev_attn_state is not None: + if incremental_state is None: + incremental_state = {} + prev_key, prev_value = prev_attn_state + saved_state = {"prev_key": prev_key, "prev_value": prev_value} + self.encoder_attn._set_input_buffer(incremental_state, saved_state) + x, attn = self.encoder_attn( + query=x, + key=encoder_out, + value=encoder_out, + key_padding_mask=encoder_padding_mask, + incremental_state=incremental_state, + static_kv=True, + need_weights=(not self.training and self.need_attn), + ) + x = F.dropout(x, p=self.dropout, training=self.training) + x = residual + x + x = self.maybe_layer_norm(self.encoder_attn_layer_norm, x, after=True) + + residual = x + x = self.maybe_layer_norm(self.final_layer_norm, x, before=True) + x = self.activation_fn(self.fc1(x)) + x = F.dropout(x, p=self.activation_dropout, training=self.training) + x = self.fc2(x) + x = F.dropout(x, p=self.dropout, training=self.training) + x = residual + x + x = self.maybe_layer_norm(self.final_layer_norm, x, after=True) + + if mt_task: + return (x, encoder_out, encoder_padding_mask) + return x + + def buffered_future_mask(self, tensor): + dim = tensor.size(0) + if ( + not hasattr(self, "_future_mask") + or self._future_mask is None + or self._future_mask.device != tensor.device + ): + self._future_mask = torch.triu( + utils.fill_with_neg_inf(tensor.new(dim, dim)), 1 + ) + if self._future_mask.size(0) < dim: + self._future_mask = torch.triu( + utils.fill_with_neg_inf(self._future_mask.resize_(dim, dim)), 1 + ) + return self._future_mask[:dim, :dim] + + def maybe_layer_norm(self, layer_norm, x, before=False, after=False): + assert before ^ after + if after ^ self.normalize_before: + return layer_norm(x) + else: + return x + + def make_generation_fast_(self, need_attn=False, **kwargs): + self.need_attn = need_attn + + +def Embedding(num_embeddings, embedding_dim, padding_idx): + m = nn.Embedding(num_embeddings, embedding_dim, padding_idx=padding_idx) + nn.init.normal_(m.weight, mean=0, std=embedding_dim ** -0.5) + nn.init.constant_(m.weight[padding_idx], 0) + return m + + +def Linear(in_features, out_features, bias=True): + m = nn.Linear(in_features, out_features, bias) + nn.init.xavier_uniform_(m.weight) + if bias: + nn.init.constant_(m.bias, 0.0) + return m diff --git a/fairseq/model_parallel/models/pipeline_parallel_transformer/model.py b/fairseq/model_parallel/models/pipeline_parallel_transformer/model.py new file mode 100644 index 0000000000000000000000000000000000000000..7f30dd98bb19b7bc414790787053efb231855129 --- /dev/null +++ b/fairseq/model_parallel/models/pipeline_parallel_transformer/model.py @@ -0,0 +1,767 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging + +import torch +import torch.nn as nn +import torch.nn.functional as F +from fairseq import utils +from fairseq.model_parallel.models.pipeline_parallel_transformer.layers import ( + Embedding, + TransformerDecoderEmbedding, + TransformerDecoderLayer, + TransformerDecoderOutputLayer, + TransformerEncoderEmbedding, + TransformerEncoderLayer, + TransformerEncoderLayerNorm, +) +from fairseq.models import ( + BaseFairseqModel, + FairseqDecoder, + FairseqEncoder, + register_model, + register_model_architecture, +) +from fairseq.models.fairseq_encoder import EncoderOut +from fairseq.models.transformer import ( + base_architecture, + transformer_iwslt_de_en, + transformer_wmt_en_de_big, +) +from fairseq.modules import SinusoidalPositionalEmbedding + + +logger = logging.getLogger(__name__) + + +DEFAULT_MAX_SOURCE_POSITIONS = 1024 +DEFAULT_MAX_TARGET_POSITIONS = 1024 +TORCH_PIPE = False +RPC_INIT = False + +def import_pipe(): + global TORCH_PIPE + global RPC_INIT + try: + from torch.distributed.pipeline.sync import Pipe # noqa + global Pipe + from torch.distributed.pipeline.sync.utils import partition_model + global partition_model + from torch.distributed import rpc + import tempfile + TORCH_PIPE = True + # Initialize single process RPC agent since TORCH_PIPE requires + # RRef. RRef depends on RPC being initialized and as a result we initialize + # RPC with a single node. + tmpfile = tempfile.NamedTemporaryFile() + if not RPC_INIT: + rpc.init_rpc( + name="worker", + rank=0, + world_size=1, + rpc_backend_options=rpc.TensorPipeRpcBackendOptions( + init_method="file://{}".format(tmpfile.name), + ) + ) + RPC_INIT = True + logger.info('Using torch pipe') + except ImportError: + try: + from fairscale.nn import Pipe # noqa + logger.info('Using fairscale pipe') + except ImportError: + raise ImportError("Please install fairscale with: pip install fairscale") + + +@register_model("pipeline_parallel_transformer") +class PipelineParallelTransformerModel(BaseFairseqModel): + def __init__(self, encoder, decoder, balance, devices, chunks, checkpoint): + import_pipe() + super().__init__() + assert isinstance(encoder, FairseqEncoder) + assert isinstance(decoder, FairseqDecoder) + encoder_module_list = ( + [encoder.embedding_layer] + + list(encoder.encoder_layers) + + [encoder.final_layer_norm] + ) + self.num_encoder_modules = len(encoder_module_list) + decoder_module_list = ( + [decoder.embedding_layer] + + list(decoder.decoder_layers) + + [decoder.decoder_output_layer] + ) + self.num_decoder_modules = len(decoder_module_list) + module_list = encoder_module_list + decoder_module_list + self.devices = devices + if TORCH_PIPE: + self.model = Pipe( + partition_model(nn.Sequential(*module_list), balance, devices), + chunks=chunks, + checkpoint=checkpoint, + ) + else: + self.model = Pipe( + nn.Sequential(*module_list), + balance=balance, + devices=devices, + chunks=chunks, + checkpoint=checkpoint, + ) + self.encoder_max_positions = self.max_positions_helper( + encoder.embedding_layer, "max_source_positions" + ) + self.decoder_max_positions = self.max_positions_helper( + decoder.embedding_layer, "max_target_positions" + ) + self.adaptive_softmax = getattr(decoder, "adaptive_softmax", None) + # Note: To be populated during inference + self.encoder = None + self.decoder = None + + def forward(self, src_tokens, src_lengths, prev_output_tokens): + if self.training: + input_lst = [src_tokens, src_lengths, prev_output_tokens] + input = tuple(i.to(self.devices[0], non_blocking=True) for i in input_lst) + if TORCH_PIPE: + return self.model(input).local_value() + else: + return self.model(input) + else: + assert self.encoder is not None and self.decoder is not None, ( + "encoder and decoder need to be initialized by " + + "calling the `prepare_for_inference_()` method" + ) + encoder_output_tuple = self.encoder(input) + return self.decoder(encoder_output_tuple) + + def prepare_for_inference_(self, cfg): + if self.encoder is not None and self.decoder is not None: + logger.info("Encoder and Decoder already initialized") + return + encoder_module_list = [] + decoder_module_list = [] + module_count = 0 + for partition in self.model.partitions: + for module in partition: + if module_count < self.num_encoder_modules: + encoder_module_list.append(module) + else: + decoder_module_list.append(module) + module_count += 1 + self.model = None + self.encoder = TransformerEncoder(cfg.distributed_training, None, None, encoder_module_list) + self.decoder = TransformerDecoder( + cfg.distributed_training, None, None, decoder_module_list=decoder_module_list + ) + + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + # fmt: off + parser.add_argument('--activation-fn', + choices=utils.get_available_activation_fns(), + help='activation function to use') + parser.add_argument('--dropout', type=float, metavar='D', + help='dropout probability') + parser.add_argument('--attention-dropout', type=float, metavar='D', + help='dropout probability for attention weights') + parser.add_argument('--activation-dropout', '--relu-dropout', type=float, metavar='D', + help='dropout probability after activation in FFN.') + parser.add_argument('--encoder-embed-path', type=str, metavar='STR', + help='path to pre-trained encoder embedding') + parser.add_argument('--encoder-embed-dim', type=int, metavar='N', + help='encoder embedding dimension') + parser.add_argument('--encoder-ffn-embed-dim', type=int, metavar='N', + help='encoder embedding dimension for FFN') + parser.add_argument('--encoder-layers', type=int, metavar='N', + help='num encoder layers') + parser.add_argument('--encoder-attention-heads', type=int, metavar='N', + help='num encoder attention heads') + parser.add_argument('--encoder-normalize-before', action='store_true', + help='apply layernorm before each encoder block') + parser.add_argument('--encoder-learned-pos', action='store_true', + help='use learned positional embeddings in the encoder') + parser.add_argument('--decoder-embed-path', type=str, metavar='STR', + help='path to pre-trained decoder embedding') + parser.add_argument('--decoder-embed-dim', type=int, metavar='N', + help='decoder embedding dimension') + parser.add_argument('--decoder-ffn-embed-dim', type=int, metavar='N', + help='decoder embedding dimension for FFN') + parser.add_argument('--decoder-layers', type=int, metavar='N', + help='num decoder layers') + parser.add_argument('--decoder-attention-heads', type=int, metavar='N', + help='num decoder attention heads') + parser.add_argument('--decoder-learned-pos', action='store_true', + help='use learned positional embeddings in the decoder') + parser.add_argument('--decoder-normalize-before', action='store_true', + help='apply layernorm before each decoder block') + parser.add_argument('--share-decoder-input-output-embed', action='store_true', + help='share decoder input and output embeddings') + parser.add_argument('--share-all-embeddings', action='store_true', + help='share encoder, decoder and output embeddings' + ' (requires shared dictionary and embed dim)') + parser.add_argument('--no-token-positional-embeddings', default=False, action='store_true', + help='if set, disables positional embeddings (outside self attention)') + parser.add_argument('--adaptive-softmax-cutoff', metavar='EXPR', + help='comma separated list of adaptive softmax cutoff points. ' + 'Must be used with adaptive_loss criterion'), + parser.add_argument('--adaptive-softmax-dropout', type=float, metavar='D', + help='sets adaptive softmax dropout for the tail projections') + parser.add_argument('--num-embedding-chunks', type=int, metavar='N', default=1, + help='Number of embedding layer chunks (enables more even distribution' + 'of optimizer states across data parallel nodes' + 'when using optimizer state sharding and' + 'a big embedding vocabulary)') + # fmt: on + + @classmethod + def build_model_base(cls, args, task): + """Build a new model instance.""" + + # make sure all arguments are present in older models + base_architecture(args) + + if not hasattr(args, "max_source_positions"): + args.max_source_positions = DEFAULT_MAX_SOURCE_POSITIONS + if not hasattr(args, "max_target_positions"): + args.max_target_positions = DEFAULT_MAX_TARGET_POSITIONS + + src_dict, tgt_dict = task.source_dictionary, task.target_dictionary + + def build_embedding(dictionary, embed_dim, path=None, num_embed_chunks=1): + assert embed_dim % num_embed_chunks == 0, ( + f"Number of embedding chunks = {num_embed_chunks} should be " + + f"divisible by the embedding dimension = {embed_dim}" + ) + assert path is None or num_embed_chunks == 1, ( + "Loading embedding from a path with number of embedding chunks > 1" + + " is not yet supported" + ) + num_embeddings = len(dictionary) + padding_idx = dictionary.pad() + # if provided, load from preloaded dictionaries + if path: + emb = Embedding(num_embeddings, embed_dim, padding_idx) + embed_dict = utils.parse_embedding(path) + utils.load_embedding(embed_dict, dictionary, emb) + else: + embed_chunk_dim = embed_dim // num_embed_chunks + emb = nn.ModuleList() + for i in range(num_embed_chunks): + emb.append(Embedding(num_embeddings, embed_chunk_dim, padding_idx)) + return emb + + num_embed_chunks = args.num_embedding_chunks + if args.share_all_embeddings: + if src_dict != tgt_dict: + raise ValueError("--share-all-embeddings requires a joined dictionary") + if args.encoder_embed_dim != args.decoder_embed_dim: + raise ValueError( + "--share-all-embeddings requires --encoder-embed-dim to match --decoder-embed-dim" + ) + if args.decoder_embed_path and ( + args.decoder_embed_path != args.encoder_embed_path + ): + raise ValueError( + "--share-all-embeddings not compatible with --decoder-embed-path" + ) + encoder_embed_tokens = build_embedding( + src_dict, + args.encoder_embed_dim, + args.encoder_embed_path, + num_embed_chunks, + ) + decoder_embed_tokens = encoder_embed_tokens + args.share_decoder_input_output_embed = True + else: + assert args.share_decoder_input_output_embed or num_embed_chunks == 1, ( + "Not sharing decoder I/O embeddings is not yet supported with number of " + + "embedding chunks > 1" + ) + encoder_embed_tokens = build_embedding( + src_dict, + args.encoder_embed_dim, + args.encoder_embed_path, + num_embed_chunks, + ) + decoder_embed_tokens = build_embedding( + tgt_dict, + args.decoder_embed_dim, + args.decoder_embed_path, + num_embed_chunks, + ) + + encoder = cls.build_encoder(args, src_dict, encoder_embed_tokens) + decoder = cls.build_decoder(args, tgt_dict, decoder_embed_tokens) + return (encoder, decoder) + + @classmethod + def build_encoder(cls, args, src_dict, embed_tokens): + return TransformerEncoder(args, src_dict, embed_tokens) + + @classmethod + def build_decoder(cls, args, tgt_dict, embed_tokens): + return TransformerDecoder(args, tgt_dict, embed_tokens) + + @classmethod + def build_model(cls, args, task): + encoder, decoder = cls.build_model_base(args, task) + return PipelineParallelTransformerModel( + encoder=encoder, + decoder=decoder, + balance=utils.eval_str_list(args.pipeline_balance, type=int), + devices=utils.eval_str_list(args.pipeline_devices, type=int), + chunks=args.pipeline_chunks, + checkpoint=args.pipeline_checkpoint, + ) + + def output_layer(self, features, **kwargs): + """Project features to the default output size (typically vocabulary size).""" + return self.decoder.output_layer(features, **kwargs) + + def max_positions(self): + """Maximum length supported by the model.""" + return (self.encoder_max_positions, self.decoder_max_positions) + + def max_positions_helper( + self, embedding_layer, max_positions_field="max_source_positions" + ): + """Maximum input length supported by the encoder or decoder.""" + if embedding_layer.embed_positions is None: + return getattr(embedding_layer, max_positions_field) + return min( + getattr(embedding_layer, max_positions_field), + embedding_layer.embed_positions.max_positions, + ) + + def get_normalized_probs(self, net_output, log_probs, sample=None): + """Get normalized probabilities (or log probs) from a net's output.""" + + if hasattr(self, "adaptive_softmax") and self.adaptive_softmax is not None: + if sample is not None: + assert "target" in sample + target = sample["target"] + else: + target = None + out = self.adaptive_softmax.get_log_prob(net_output, target=target) + return out.exp_() if not log_probs else out + + # A Pipe() module returns a tuple of tensors as the output. + # In this case, the tuple has one element - the output tensor of logits + logits = net_output if isinstance(net_output, torch.Tensor) else net_output[0] + if log_probs: + return utils.log_softmax(logits, dim=-1, onnx_trace=False) + else: + return utils.softmax(logits, dim=-1, onnx_trace=False) + + def max_decoder_positions(self): + """Maximum length supported by the decoder.""" + return self.decoder_max_positions + + def load_state_dict(self, state_dict, strict=True, model_cfg=None): + """Copies parameters and buffers from *state_dict* into this module and + its descendants. + + Overrides the method in :class:`nn.Module`. Compared with that method + this additionally "upgrades" *state_dicts* from old checkpoints. + """ + self.upgrade_state_dict(state_dict) + is_regular_transformer = not any("model.partitions" in k for k in state_dict) + if is_regular_transformer: + state_dict = self.convert_to_pipeline_parallel_state_dict(state_dict) + return super().load_state_dict(state_dict, strict) + + def convert_to_pipeline_parallel_state_dict(self, state_dict): + new_state_dict = self.state_dict() + encoder_layer_idx = 0 + decoder_layer_idx = 0 + encoder_key_suffixes = [ + "self_attn.k_proj.weight", + "self_attn.k_proj.bias", + "self_attn.v_proj.weight", + "self_attn.v_proj.bias", + "self_attn.q_proj.weight", + "self_attn.q_proj.bias", + "self_attn.out_proj.weight", + "self_attn.out_proj.bias", + "self_attn_layer_norm.weight", + "self_attn_layer_norm.bias", + "fc1.weight", + "fc1.bias", + "fc2.weight", + "fc2.bias", + "final_layer_norm.weight", + "final_layer_norm.bias", + ] + decoder_key_suffixes = [ + "self_attn.k_proj.weight", + "self_attn.k_proj.bias", + "self_attn.v_proj.weight", + "self_attn.v_proj.bias", + "self_attn.q_proj.weight", + "self_attn.q_proj.bias", + "self_attn.out_proj.weight", + "self_attn.out_proj.bias", + "self_attn_layer_norm.weight", + "self_attn_layer_norm.bias", + "encoder_attn.k_proj.weight", + "encoder_attn.k_proj.bias", + "encoder_attn.v_proj.weight", + "encoder_attn.v_proj.bias", + "encoder_attn.q_proj.weight", + "encoder_attn.q_proj.bias", + "encoder_attn.out_proj.weight", + "encoder_attn.out_proj.bias", + "encoder_attn_layer_norm.weight", + "encoder_attn_layer_norm.bias", + "fc1.weight", + "fc1.bias", + "fc2.weight", + "fc2.bias", + "final_layer_norm.weight", + "final_layer_norm.bias", + ] + for pid, partition in enumerate(self.model.partitions): + logger.info(f"Begin Partition {pid}") + for mid, module in enumerate(partition): + # fmt: off + if isinstance(module, TransformerEncoderEmbedding): + new_state_dict[f'model.partitions.{pid}.{mid}.embed_tokens.weight'] = state_dict['encoder.embed_tokens.weight'] + new_state_dict[f'model.partitions.{pid}.{mid}.embed_positions._float_tensor'] = state_dict['encoder.embed_positions._float_tensor'] + if isinstance(module, TransformerEncoderLayer): + for suffix in encoder_key_suffixes: + new_state_dict[f'model.partitions.{pid}.{mid}.{suffix}'] = state_dict[f'encoder.layers.{encoder_layer_idx}.{suffix}'] + encoder_layer_idx += 1 + if isinstance(module, TransformerDecoderLayer): + for suffix in decoder_key_suffixes: + new_state_dict[f'model.partitions.{pid}.{mid}.{suffix}'] = state_dict[f'decoder.layers.{decoder_layer_idx}.{suffix}'] + decoder_layer_idx += 1 + if isinstance(module, TransformerEncoderLayerNorm): + if 'encoder.layer_norm.weight' in state_dict: + new_state_dict[f'model.partitions.{pid}.{mid}.layer_norm.weight'] = state_dict['encoder.layer_norm.weight'] + new_state_dict[f'model.partitions.{pid}.{mid}.layer_norm.bias'] = state_dict['encoder.layer_norm.bias'] + if isinstance(module, TransformerDecoderEmbedding): + new_state_dict[f'model.partitions.{pid}.{mid}.embed_tokens.weight'] = state_dict['decoder.embed_tokens.weight'] + new_state_dict[f'model.partitions.{pid}.{mid}.embed_positions._float_tensor'] = state_dict['decoder.embed_positions._float_tensor'] + if isinstance(module, TransformerDecoderOutputLayer): + new_state_dict[f'model.partitions.{pid}.{mid}.output_projection.weight'] = state_dict['decoder.output_projection.weight'] + # fmt: on + return new_state_dict + + +class TransformerEncoder(FairseqEncoder): + """ + Transformer encoder consisting of *args.encoder_layers* layers. Each layer + is a :class:`TransformerEncoderLayer`. + + Args: + args (argparse.Namespace): parsed command-line arguments + dictionary (~fairseq.data.Dictionary): encoding dictionary + embed_tokens (torch.nn.Embedding): input embedding + """ + + def __init__(self, args, dictionary, embed_tokens, encoder_module_list=None): + super().__init__(dictionary) + self.register_buffer("version", torch.Tensor([3])) + import_pipe() + self.use_pipeline = encoder_module_list is not None + if not self.use_pipeline: + self.embedding_layer = TransformerEncoderEmbedding(args, embed_tokens) + self.encoder_layers = nn.Sequential(*[TransformerEncoderLayer(args) for i in range(args.encoder_layers)]) + if isinstance(embed_tokens, nn.ModuleList): + emb_dim = sum(e.embedding_dim for e in embed_tokens) + else: + emb_dim = embed_tokens.embedding_dim + self.final_layer_norm = TransformerEncoderLayerNorm(args, emb_dim) + else: + encoder_balance = utils.eval_str_list( + args.pipeline_encoder_balance, type=int + ) + encoder_devices = utils.eval_str_list( + args.pipeline_encoder_devices, type=int + ) + assert sum(encoder_balance) == len(encoder_module_list), ( + f"Sum of encoder_balance={encoder_balance} is not equal " + + f"to num_encoder_modules={len(encoder_module_list)}" + ) + if TORCH_PIPE: + self.model = Pipe( + module=partition_model(nn.Sequential(*encoder_module_list), encoder_balance, encoder_devices), + chunks=args.pipeline_chunks, + checkpoint=args.pipeline_checkpoint, + ) + else: + self.model = Pipe( + module=nn.Sequential(*encoder_module_list), + balance=encoder_balance, + devices=encoder_devices, + chunks=args.pipeline_chunks, + checkpoint=args.pipeline_checkpoint, + ) + + def forward(self, src_tokens, src_lengths): + """ + Args: + input_tuple( + src_tokens (LongTensor): tokens in the source language of shape + `(batch, src_len)` + src_lengths (torch.LongTensor): lengths of each source sentence of + shape `(batch)` + ) + + Returns: + output_tuple( + - **encoder_out** (Tensor): the last encoder layer's output of + shape `(src_len, batch, embed_dim)` + - **encoder_padding_mask** (ByteTensor): the positions of + padding elements of shape `(batch, src_len)` + - prev_output_tokens + - **encoder_states** (List[Tensor]): all intermediate + hidden states of shape `(src_len, batch, embed_dim)`. + Only populated if *return_all_hiddens* is True. + ) + """ + dummy_prev_output_tokens = torch.zeros( + 1, dtype=src_tokens.dtype, device=src_tokens.device + ) + input_tuple = (src_tokens, src_lengths, dummy_prev_output_tokens) + if self.use_pipeline: + input_tuple = tuple(i.to(self.model.devices[0]) for i in input_tuple) + if TORCH_PIPE: + encoder_out = self.model(input_tuple).local_value() + else: + encoder_out = self.model(input_tuple) + else: + encoder_embed_output_tuple = self.embedding_layer(input_tuple) + encoder_layers_output = self.encoder_layers(encoder_embed_output_tuple) + encoder_out = self.final_layer_norm(encoder_layers_output) + # first element is the encoder output + # second element is the encoder padding mask + # the remaining elements of EncoderOut are not computed by + # the PipelineParallelTransformer + return EncoderOut(encoder_out[0], encoder_out[1], None, None, None, None) + + def reorder_encoder_out(self, encoder_out, new_order): + """ + Reorder encoder output according to *new_order*. + + Args: + encoder_out: output from the ``forward()`` method + new_order (LongTensor): desired order + + Returns: + *encoder_out* rearranged according to *new_order* + """ + if encoder_out.encoder_out is not None: + encoder_out = encoder_out._replace( + encoder_out=encoder_out.encoder_out.index_select(1, new_order) + ) + if encoder_out.encoder_padding_mask is not None: + encoder_out = encoder_out._replace( + encoder_padding_mask=encoder_out.encoder_padding_mask.index_select( + 0, new_order + ) + ) + if encoder_out.encoder_embedding is not None: + encoder_out = encoder_out._replace( + encoder_embedding=encoder_out.encoder_embedding.index_select( + 0, new_order + ) + ) + if encoder_out.encoder_states is not None: + for idx, state in enumerate(encoder_out.encoder_states): + encoder_out.encoder_states[idx] = state.index_select(1, new_order) + return encoder_out + + def max_positions(self): + """Maximum input length supported by the encoder.""" + if self.embedding_layer.embed_positions is None: + return self.embedding_layer.max_source_positions + return min( + self.embedding_layer.max_source_positions, + self.embedding_layer.embed_positions.max_positions, + ) + + +class TransformerDecoder(FairseqDecoder): + """ + Transformer decoder consisting of *args.decoder_layers* layers. Each layer + is a :class:`TransformerDecoderLayer`. + + Args: + args (argparse.Namespace): parsed command-line arguments + dictionary (~fairseq.data.Dictionary): decoding dictionary + embed_tokens (torch.nn.Embedding): output embedding + no_encoder_attn (bool, optional): whether to attend to encoder outputs + (default: False). + """ + + def __init__( + self, + args, + dictionary, + embed_tokens, + no_encoder_attn=False, + decoder_module_list=None, + ): + super().__init__(dictionary) + self.register_buffer("version", torch.Tensor([3])) + import_pipe() + self.use_pipeline = decoder_module_list is not None + if not self.use_pipeline: + self.embedding_layer = TransformerDecoderEmbedding(args, embed_tokens) + self.decoder_layers = nn.Sequential(*[ + TransformerDecoderLayer(args, no_encoder_attn) + for _ in range(args.decoder_layers) + ]) + self.decoder_output_layer = TransformerDecoderOutputLayer( + args, embed_tokens, dictionary + ) + else: + decoder_balance = utils.eval_str_list( + args.pipeline_decoder_balance, type=int + ) + decoder_devices = utils.eval_str_list( + args.pipeline_decoder_devices, type=int + ) + assert sum(decoder_balance) == len(decoder_module_list), ( + f"Sum of decoder_balance={decoder_balance} is not equal " + + f"to num_decoder_modules={len(decoder_module_list)}" + ) + if TORCH_PIPE: + self.model = Pipe( + module=partition_model(nn.Sequential(*decoder_module_list), decoder_balance, decoder_devices), + chunks=args.pipeline_chunks, + checkpoint=args.pipeline_checkpoint, + ) + else: + self.model = Pipe( + module=nn.Sequential(*decoder_module_list), + balance=decoder_balance, + devices=decoder_devices, + chunks=args.pipeline_chunks, + checkpoint=args.pipeline_checkpoint, + ) + + def forward( + self, + prev_output_tokens, + encoder_out=None, + ): + """ + Args: + prev_output_tokens (LongTensor): previous decoder outputs of shape + `(batch, tgt_len)`, for teacher forcing + encoder_out (optional): output from the encoder, used for + encoder-side attention + incremental_state (dict): dictionary used for storing state during + :ref:`Incremental decoding` + features_only (bool, optional): only return features without + applying output layer (default: False). + + Returns: + tuple: + - the decoder's output of shape `(batch, tgt_len, vocab)` + - a dictionary with any model-specific outputs + """ + input_tuple = ( + encoder_out.encoder_out, + encoder_out.encoder_padding_mask, + prev_output_tokens, + ) + if self.use_pipeline: + input_tuple = tuple(i.to(self.model.devices[0]) for i in input_tuple) + if TORCH_PIPE: + return (self.model(input_tuple).local_value(),) + else: + return (self.model(input_tuple),) + else: + embed_layer_output = self.embedding_layer(input_tuple) + state = self.decoder_layers(embed_layer_output) + return (self.decoder_output_layer(state),) + + def output_layer(self, features, **kwargs): + """Project features to the vocabulary size.""" + if self.adaptive_softmax is None: + # project back to size of vocabulary + if self.share_input_output_embed: + return F.linear(features, self.embed_tokens.weight) + else: + return F.linear(features, self.embed_out) + else: + return features + + def max_positions(self): + """Maximum output length supported by the decoder.""" + if self.embedding_layer.embed_positions is None: + return self.embedding_layer.max_target_positions + return min( + self.embedding_layer.max_target_positions, + self.embedding_layer.embed_positions.max_positions, + ) + + def buffered_future_mask(self, tensor): + dim = tensor.size(0) + if ( + not hasattr(self, "_future_mask") + or self._future_mask is None + or self._future_mask.device != tensor.device + or self._future_mask.size(0) < dim + ): + self._future_mask = torch.triu( + utils.fill_with_neg_inf(tensor.new(dim, dim)), 1 + ) + return self._future_mask[:dim, :dim] + + def upgrade_state_dict_named(self, state_dict, name): + """Upgrade a (possibly old) state dict for new versions of fairseq.""" + if isinstance(self.embed_positions, SinusoidalPositionalEmbedding): + weights_key = "{}.embed_positions.weights".format(name) + if weights_key in state_dict: + del state_dict[weights_key] + state_dict[ + "{}.embed_positions._float_tensor".format(name) + ] = torch.FloatTensor(1) + + for i in range(len(self.layers)): + # update layer norms + layer_norm_map = { + "0": "self_attn_layer_norm", + "1": "encoder_attn_layer_norm", + "2": "final_layer_norm", + } + for old, new in layer_norm_map.items(): + for m in ("weight", "bias"): + k = "{}.layers.{}.layer_norms.{}.{}".format(name, i, old, m) + if k in state_dict: + state_dict[ + "{}.layers.{}.{}.{}".format(name, i, new, m) + ] = state_dict[k] + del state_dict[k] + + version_key = "{}.version".format(name) + if utils.item(state_dict.get(version_key, torch.Tensor([1]))[0]) <= 2: + # earlier checkpoints did not normalize after the stack of layers + self.layer_norm = None + self.normalize = False + state_dict[version_key] = torch.Tensor([1]) + + return state_dict + + +@register_model_architecture( + "pipeline_parallel_transformer", "transformer_iwslt_de_en_pipeline_parallel" +) +def transformer_iwslt_de_en_dist(args): + transformer_iwslt_de_en(args) + + +@register_model_architecture( + "pipeline_parallel_transformer", "transformer_wmt_en_de_big_pipeline_parallel" +) +def transformer_wmt_en_de_big_dist(args): + transformer_wmt_en_de_big(args) diff --git a/fairseq/model_parallel/models/roberta/__init__.py b/fairseq/model_parallel/models/roberta/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..117827c3e9c176477f33e3a6fd7fe19a922411a2 --- /dev/null +++ b/fairseq/model_parallel/models/roberta/__init__.py @@ -0,0 +1,6 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .model import * # noqa diff --git a/fairseq/model_parallel/models/roberta/model.py b/fairseq/model_parallel/models/roberta/model.py new file mode 100644 index 0000000000000000000000000000000000000000..77a80ef72057219110b34678a38705549910edd3 --- /dev/null +++ b/fairseq/model_parallel/models/roberta/model.py @@ -0,0 +1,225 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +RoBERTa: A Robustly Optimized BERT Pretraining Approach. +""" + +import logging + +import torch +import torch.nn as nn +import torch.nn.functional as F +from fairseq import utils +from fairseq.model_parallel.models.transformer import ModelParallelTransformerEncoder +from fairseq.models import register_model, register_model_architecture +from fairseq.models.roberta import ( + roberta_base_architecture, + roberta_prenorm_architecture, + RobertaEncoder, + RobertaModel, +) +from fairseq.modules import LayerNorm + + +try: + from fairseq.model_parallel.megatron.mpu import ( + copy_to_model_parallel_region, + gather_from_model_parallel_region, + ColumnParallelLinear, + VocabParallelEmbedding, + ) + + has_megatron_submodule = True +except (ImportError, ModuleNotFoundError): + has_megatron_submodule = False + +logger = logging.getLogger(__name__) + + +@register_model("model_parallel_roberta") +class ModelParallelRobertaModel(RobertaModel): + def __init__(self, args, encoder): + super().__init__(args, encoder) + + self.classification_heads = nn.ModuleDict() + + @staticmethod + def add_args(parser): + RobertaModel.add_args(parser) + parser.add_argument( + "--no-final-layer-norm", + action="store_true", + help=( + "don't add final layernorm (only applicable when " + "--encoder-normalize-before=True" + ), + ) + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + + # make sure all arguments are present + base_architecture(args) + + task.source_dictionary.pad_to_multiple_(args.model_parallel_size * 8) + task.target_dictionary.pad_to_multiple_(args.model_parallel_size * 8) + + if not hasattr(args, "max_positions"): + args.max_positions = args.tokens_per_sample + + if getattr(args, "untie_weights_roberta", False): + raise NotImplementedError( + "--untie-weights-roberta is not supported in model parallel mode" + ) + + encoder = ModelParallelRobertaEncoder(args, task.source_dictionary) + return cls(args, encoder) + + def forward( + self, + src_tokens, + features_only=False, + return_all_hiddens=False, + classification_head_name=None, + **kwargs + ): + if classification_head_name is not None: + features_only = True + + x, extra = self.encoder(src_tokens, features_only, return_all_hiddens, **kwargs) + + if classification_head_name is not None: + x = self.classification_heads[classification_head_name](x) + return x, extra + + def register_classification_head( + self, name, num_classes=None, inner_dim=None, **kwargs + ): + """Register a classification head.""" + if name in self.classification_heads: + prev_num_classes = self.classification_heads[name].out_proj.out_features + prev_inner_dim = self.classification_heads[name].dense.out_features + if num_classes != prev_num_classes or inner_dim != prev_inner_dim: + logger.warning( + 're-registering head "{}" with num_classes {} (prev: {}) ' + "and inner_dim {} (prev: {})".format( + name, num_classes, prev_num_classes, inner_dim, prev_inner_dim + ) + ) + self.classification_heads[name] = ModelParallelRobertaClassificationHead( + self.args.encoder_embed_dim, + inner_dim or self.args.encoder_embed_dim, + num_classes, + self.args.pooler_activation_fn, + self.args.pooler_dropout, + ) + + +class ModelParallelRobertaLMHead(nn.Module): + """Head for masked language modeling.""" + + def __init__(self, embed_dim, output_dim, activation_fn, weight=None): + super().__init__() + self.dense = ColumnParallelLinear(embed_dim, embed_dim, gather_output=True) + self.activation_fn = utils.get_activation_fn(activation_fn) + self.layer_norm = LayerNorm(embed_dim) + + if weight is None: + weight = nn.Linear(embed_dim, output_dim, bias=False).weight + self.weight = weight + self.bias = nn.Parameter(torch.zeros(output_dim)) + + def forward(self, features, masked_tokens=None, **kwargs): + # Only project the unmasked tokens while training, + # saves both memory and computation + if masked_tokens is not None: + features = features[masked_tokens, :] + + x = self.dense(features) + x = self.activation_fn(x) + x = self.layer_norm(x) + + x = copy_to_model_parallel_region(x) + # project back to size of vocabulary with bias + x = F.linear(x, self.weight) + x = gather_from_model_parallel_region(x).contiguous() + x = x + self.bias + return x + + +class ModelParallelRobertaClassificationHead(nn.Module): + """Head for sentence-level classification tasks.""" + + def __init__( + self, input_dim, inner_dim, num_classes, activation_fn, pooler_dropout + ): + super().__init__() + self.dense = ColumnParallelLinear(input_dim, inner_dim, gather_output=True) + self.activation_fn = utils.get_activation_fn(activation_fn) + self.dropout = nn.Dropout(p=pooler_dropout) + self.out_proj = nn.Linear(inner_dim, num_classes) + + def forward(self, features, **kwargs): + x = features[:, 0, :] # take <s> token (equiv. to [CLS]) + x = self.dropout(x) + x = self.dense(x) + x = self.activation_fn(x) + x = self.dropout(x) + x = self.out_proj(x) + return x + + +class ModelParallelRobertaEncoder(RobertaEncoder): + """RoBERTa encoder.""" + + def __init__(self, args, dictionary): + super().__init__(args, dictionary) + assert not self.args.untie_weights_roberta + + def build_embedding(self, vocab_size, embedding_dim, padding_idx): + return VocabParallelEmbedding(vocab_size, embedding_dim, padding_idx) + + def build_encoder(self, args, dictionary, embed_tokens): + return ModelParallelTransformerEncoder(args, dictionary, embed_tokens) + + def build_lm_head(self, embed_dim, output_dim, activation_fn, weight): + return ModelParallelRobertaLMHead(embed_dim, output_dim, activation_fn, weight) + + +@register_model_architecture("model_parallel_roberta", "model_parallel_roberta") +def base_architecture(args): + args.no_final_layer_norm = getattr(args, "no_final_layer_norm", False) + # model parallel RoBERTa defaults to "Pre-LN" formulation + roberta_prenorm_architecture(args) + + +# earlier versions of model parallel RoBERTa removed the final layer norm +@register_model_architecture("model_parallel_roberta", "model_parallel_roberta_v1") +def model_parallel_roberta_v1_architecture(args): + args.no_final_layer_norm = getattr(args, "no_final_layer_norm", True) + base_architecture(args) + + +@register_model_architecture( + "model_parallel_roberta", "model_parallel_roberta_postnorm" +) +def model_parallel_roberta_postnorm_architecture(args): + # the original BERT/RoBERTa uses the "Post-LN" formulation + roberta_base_architecture(args) + + +@register_model_architecture("model_parallel_roberta", "model_parallel_roberta_base") +def model_parallel_roberta_base_architecture(args): + base_architecture(args) + + +@register_model_architecture("model_parallel_roberta", "model_parallel_roberta_large") +def model_parallel_roberta_large_architecture(args): + args.encoder_layers = getattr(args, "encoder_layers", 24) + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 1024) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 4096) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 16) + base_architecture(args) diff --git a/fairseq/model_parallel/models/transformer.py b/fairseq/model_parallel/models/transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..6b330ef1b7f7a506e7e8176f20a0e722b5fd5149 --- /dev/null +++ b/fairseq/model_parallel/models/transformer.py @@ -0,0 +1,121 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging + +import torch.nn as nn +from fairseq.model_parallel.modules import ( + ModelParallelTransformerDecoderLayer, + ModelParallelTransformerEncoderLayer, +) +from fairseq.models import register_model +from fairseq.models.transformer import ( + TransformerDecoder, + TransformerEncoder, + TransformerModel, +) + + +try: + from fairseq.model_parallel.megatron.mpu import ( + copy_to_model_parallel_region, + gather_from_model_parallel_region, + VocabParallelEmbedding, + ) + + has_megatron_submodule = True +except (ImportError, ModuleNotFoundError): + has_megatron_submodule = False + + +logger = logging.getLogger(__name__) + + +@register_model("model_parallel_transformer") +class ModelParallelTransformerModel(TransformerModel): + """ + Model parallel Transformer model. + """ + + @classmethod + def build_embedding(cls, args, dictionary, embed_dim, path=None): + if not has_megatron_submodule: + raise ImportError( + "\n\nPlease install the megatron submodule:" + "\n\n git submodule update --init " + "fairseq/model_parallel/megatron" + ) + dictionary.pad_to_multiple_(args.model_parallel_size * 8) + num_embeddings = len(dictionary) + padding_idx = dictionary.pad() + + def _vocab_init(tensor, **kwargs): + nn.init.normal_(tensor, mean=0, std=num_embeddings ** -0.5) + nn.init.constant_(tensor[1], 0) + + emb = VocabParallelEmbedding( + num_embeddings, embed_dim, padding_idx, init_method=_vocab_init + ) + # if provided, load from preloaded dictionaries + if path: + raise NotImplementedError( + "Loading of embedding from path is not supported for model parallel" + ) + return emb + + @classmethod + def build_encoder(cls, args, src_dict, embed_tokens): + return ModelParallelTransformerEncoder(args, src_dict, embed_tokens) + + @classmethod + def build_decoder(cls, args, tgt_dict, embed_tokens): + return ModelParallelTransformerDecoder( + args, + tgt_dict, + embed_tokens, + no_encoder_attn=getattr(args, "no_cross_attention", False), + ) + + +class ModelParallelTransformerEncoder(TransformerEncoder): + """ + Model parallel Transformer encoder consisting of *args.encoder_layers* layers. Each layer + is a :class:`ModelParallelTransformerEncoderLayer`. + """ + + def __init__(self, args, dictionary, embed_tokens): + super().__init__(args, dictionary, embed_tokens) + + if args.no_final_layer_norm: + self.layer_norm = None + + def build_encoder_layer(self, args): + return ModelParallelTransformerEncoderLayer(args) + + +class ModelParallelTransformerDecoder(TransformerDecoder): + """ + Model Parallel Transformer decoder consisting of *args.decoder_layers* layers. Each layer + is a :class:`ModelParallelTransformerDecoderLayer`. + """ + + def build_decoder_layer(self, args, no_encoder_attn=False): + return ModelParallelTransformerDecoderLayer(args, no_encoder_attn) + + def output_layer(self, features, **kwargs): + """Project features to the vocabulary size.""" + if not self.share_input_output_embed: + raise NotImplementedError( + "Model parallel training currently requires --share-decoder-input-output-embed" + ) + + features = copy_to_model_parallel_region(features) + + # project back to size of vocabulary + x = self.output_projection(features) + + if getattr(self.args, "criterion") != "vocab_parallel_cross_entropy": + x = gather_from_model_parallel_region(x).contiguous() + return x diff --git a/fairseq/model_parallel/models/transformer_lm.py b/fairseq/model_parallel/models/transformer_lm.py new file mode 100644 index 0000000000000000000000000000000000000000..dc52f6e8dd3899b6bf9bebae7415cee20baf9884 --- /dev/null +++ b/fairseq/model_parallel/models/transformer_lm.py @@ -0,0 +1,174 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch.nn as nn +from fairseq.model_parallel.models.transformer import ModelParallelTransformerDecoder +from fairseq.models import register_model, register_model_architecture +from fairseq.models.transformer_lm import TransformerLanguageModel + + +try: + from fairseq.model_parallel.megatron.mpu import VocabParallelEmbedding + + has_megatron_submodule = True +except (ImportError, ModuleNotFoundError): + has_megatron_submodule = False + + +DEFAULT_MAX_TARGET_POSITIONS = 1024 + + +@register_model("model_parallel_transformer_lm") +class ModelParallelTransformerLanguageModel(TransformerLanguageModel): + + @staticmethod + def add_args(parser): + TransformerLanguageModel.add_args(parser) + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + if not has_megatron_submodule: + raise ImportError( + "\n\nPlease install the megatron submodule:" + "\n\n git submodule update --init " + "fairseq/model_parallel/megatron" + ) + + # make sure all arguments are present in older models + base_lm_architecture(args) + + task.source_dictionary.pad_to_multiple_(args.model_parallel_size * 8) + task.target_dictionary.pad_to_multiple_(args.model_parallel_size * 8) + + if args.decoder_layers_to_keep: + args.decoder_layers = len(args.decoder_layers_to_keep.split(",")) + + if getattr(args, "max_target_positions", None) is None: + args.max_target_positions = getattr( + args, "tokens_per_sample", DEFAULT_MAX_TARGET_POSITIONS + ) + + if args.character_embeddings: + raise NotImplementedError( + "Character embeddings is not supported for model parallel" + ) + elif args.adaptive_input: + raise NotImplementedError( + "Adaptive input is not supported for model parallel" + ) + else: + embed_tokens = cls.build_embedding( + args, task.source_dictionary, args.decoder_input_dim + ) + + decoder = ModelParallelTransformerDecoder( + args, + task.target_dictionary, + embed_tokens, + no_encoder_attn=True, + ) + return cls(decoder) + + @staticmethod + def add_args(parser): + TransformerLanguageModel.add_args(parser) + + @classmethod + def build_embedding(cls, args, dictionary, embed_dim, path=None): + def _vocab_init(tensor, **kwargs): + nn.init.normal_(tensor, mean=0, std=embed_dim ** -0.5) + nn.init.constant_(tensor[1], 0) + + embed_tokens = VocabParallelEmbedding( + len(dictionary), embed_dim, dictionary.pad(), init_method=_vocab_init + ) + return embed_tokens + + +def base_lm_architecture(args): + # backward compatibility for older model checkpoints + if hasattr(args, "no_tie_adaptive_proj"): + # previous models defined --no-tie-adaptive-proj, so use the existence of + # that option to determine if this is an "old" model checkpoint + args.no_decoder_final_norm = True # old models always set this to True + if args.no_tie_adaptive_proj is False: + args.tie_adaptive_proj = True + if hasattr(args, "decoder_final_norm"): + args.no_decoder_final_norm = not args.decoder_final_norm + + args.activation_fn = getattr(args, "activation_fn", "relu") + args.dropout = getattr(args, "dropout", 0.1) + args.attention_dropout = getattr(args, "attention_dropout", 0.0) + args.activation_dropout = getattr(args, "activation_dropout", 0.0) + args.relu_dropout = getattr(args, "relu_dropout", 0.0) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 512) + args.decoder_output_dim = getattr( + args, "decoder_output_dim", args.decoder_embed_dim + ) + args.decoder_input_dim = getattr(args, "decoder_input_dim", args.decoder_embed_dim) + args.decoder_ffn_embed_dim = getattr(args, "decoder_ffn_embed_dim", 2048) + args.decoder_layers = getattr(args, "decoder_layers", 6) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 8) + # Model training is not stable without this + args.decoder_normalize_before = True + args.no_decoder_final_norm = getattr(args, "no_decoder_final_norm", False) + args.adaptive_softmax_cutoff = getattr(args, "adaptive_softmax_cutoff", None) + args.adaptive_softmax_dropout = getattr(args, "adaptive_softmax_dropout", 0) + args.adaptive_softmax_factor = getattr(args, "adaptive_softmax_factor", 4) + args.no_token_positional_embeddings = getattr( + args, "no_token_positional_embeddings", False + ) + args.share_decoder_input_output_embed = getattr( + args, "share_decoder_input_output_embed", False + ) + args.character_embeddings = getattr(args, "character_embeddings", False) + args.character_filters = getattr( + args, + "character_filters", + "[(1, 64), (2, 128), (3, 192), (4, 256), (5, 256), (6, 256), (7, 256)]", + ) + args.character_embedding_dim = getattr(args, "character_embedding_dim", 4) + args.char_embedder_highway_layers = getattr(args, "char_embedder_highway_layers", 2) + args.adaptive_input = getattr(args, "adaptive_input", False) + args.adaptive_input_factor = getattr(args, "adaptive_input_factor", 4) + args.adaptive_input_cutoff = getattr(args, "adaptive_input_cutoff", None) + args.tie_adaptive_weights = getattr(args, "tie_adaptive_weights", False) + args.tie_adaptive_proj = getattr(args, "tie_adaptive_proj", False) + args.decoder_learned_pos = getattr(args, "decoder_learned_pos", False) + args.decoder_layerdrop = getattr(args, "decoder_layerdrop", 0.0) + args.decoder_layers_to_keep = getattr(args, "decoder_layers_to_keep", None) + args.layernorm_embedding = getattr(args, "layernorm_embedding", False) + args.no_scale_embedding = getattr(args, "no_scale_embedding", False) + args.quant_noise_pq = getattr(args, "quant_noise_pq", 0.0) + args.quant_noise_pq_block_size = getattr(args, "quant_noise_pq_block_size", 8) + args.quant_noise_scalar = getattr(args, "quant_noise_scalar", 0.0) + args.add_bos_token = getattr(args, "add_bos_token", False) + + +@register_model_architecture("model_parallel_transformer_lm", "transformer_lm_megatron") +def transformer_lm_megatron(args): + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 3072) + args.decoder_ffn_embed_dim = getattr(args, "decoder_ffn_embed_dim", 3072 * 4) + args.decoder_layers = getattr(args, "decoder_layers", 72) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 32) + args.dropout = getattr(args, "dropout", 0.1) + args.attention_dropout = getattr(args, "attention_dropout", 0.1) + args.activation_fn = getattr(args, "activation_fn", "gelu") + base_lm_architecture(args) + + +@register_model_architecture( + "model_parallel_transformer_lm", "transformer_lm_megatron_11b" +) +def transformer_lm_megatron_11b(args): + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 3072) + args.decoder_ffn_embed_dim = getattr(args, "decoder_ffn_embed_dim", 3072 * 6) + args.decoder_layers = getattr(args, "decoder_layers", 72) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 32) + args.dropout = getattr(args, "dropout", 0.1) + args.attention_dropout = getattr(args, "attention_dropout", 0.1) + args.activation_fn = getattr(args, "activation_fn", "gelu") + base_lm_architecture(args) diff --git a/fairseq/model_parallel/modules/__init__.py b/fairseq/model_parallel/modules/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..11603217a188f420ea849ae0fde19979736ba208 --- /dev/null +++ b/fairseq/model_parallel/modules/__init__.py @@ -0,0 +1,17 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +"""isort:skip_file""" + +from .multihead_attention import ModelParallelMultiheadAttention +from .transformer_layer import ( + ModelParallelTransformerEncoderLayer, + ModelParallelTransformerDecoderLayer, +) + +__all__ = [ + "ModelParallelMultiheadAttention", + "ModelParallelTransformerEncoderLayer", + "ModelParallelTransformerDecoderLayer", +] diff --git a/fairseq/model_parallel/modules/multihead_attention.py b/fairseq/model_parallel/modules/multihead_attention.py new file mode 100644 index 0000000000000000000000000000000000000000..8eb9d09dad37ab132295166d691873beec63eaf1 --- /dev/null +++ b/fairseq/model_parallel/modules/multihead_attention.py @@ -0,0 +1,349 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from typing import Dict, Optional, Tuple + +import torch +import torch.nn.functional as F +from fairseq import utils +from fairseq.incremental_decoding_utils import with_incremental_state +from fairseq.modules.fairseq_dropout import FairseqDropout +from torch import Tensor, nn + + +try: + from fairseq.model_parallel.megatron.mpu import ( + get_cuda_rng_tracker, + get_model_parallel_world_size, + ColumnParallelLinear, + RowParallelLinear, + ) + + has_megatron_submodule = True +except (ImportError, ModuleNotFoundError): + has_megatron_submodule = False + + +@with_incremental_state +class ModelParallelMultiheadAttention(nn.Module): + """Model parallel Multi-headed attention. + This performs the Multi-headed attention over multiple gpus. + + See "Megatron-LM: https://arxiv.org/pdf/1909.08053.pdf" for more details. + """ + + def __init__( + self, + embed_dim, + num_heads, + kdim=None, + vdim=None, + dropout=0.0, + bias=True, + self_attention=False, + encoder_decoder_attention=False, + ): + super().__init__() + if not has_megatron_submodule: + raise ImportError( + "\n\nPlease install the megatron submodule:" + "\n\n git submodule update --init " + "fairseq/model_parallel/megatron" + ) + self.embed_dim = embed_dim + self.kdim = kdim if kdim is not None else embed_dim + self.vdim = vdim if vdim is not None else embed_dim + self.qkv_same_dim = self.kdim == embed_dim and self.vdim == embed_dim + + self.model_parallel_size = get_model_parallel_world_size() + + self.num_heads_partition = num_heads // self.model_parallel_size + assert ( + self.num_heads_partition * self.model_parallel_size == num_heads + ), "Number of heads must be divisible by model parallel size" + + self.dropout_module = FairseqDropout( + dropout, module_name=self.__class__.__name__ + ) + self.head_dim = embed_dim // num_heads + assert ( + self.head_dim * num_heads == self.embed_dim + ), "embed_dim must be divisible by num_heads" + self.scaling = self.head_dim ** -0.5 + + self.self_attention = self_attention + self.encoder_decoder_attention = encoder_decoder_attention + + assert ( + not self.self_attention or self.qkv_same_dim + ), "Self-attention requires query, key and value to be of the same size" + + self.k_proj = ColumnParallelLinear( + self.kdim, embed_dim, bias=bias, gather_output=False + ) + self.v_proj = ColumnParallelLinear( + self.vdim, embed_dim, bias=bias, gather_output=False + ) + self.q_proj = ColumnParallelLinear( + embed_dim, embed_dim, bias=bias, gather_output=False + ) + self.out_proj = RowParallelLinear( + embed_dim, embed_dim, bias=bias, input_is_parallel=True + ) + + def forward( + self, + query, + key: Optional[Tensor], + value: Optional[Tensor], + key_padding_mask: Optional[Tensor] = None, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + static_kv: bool = False, + attn_mask: Optional[Tensor] = None, + **unused_kwargs, + ) -> Tuple[Tensor, Optional[Tensor]]: + """Input shape: Time x Batch x Channel + + Args: + key_padding_mask (ByteTensor, optional): mask to exclude + keys that are pads, of shape `(batch, src_len)`, where + padding elements are indicated by 1s. + attn_mask (ByteTensor, optional): typically used to + implement causal attention, where the mask prevents the + attention from looking forward in time (default: None). + """ + tgt_len, bsz, embed_dim = query.size() + assert embed_dim == self.embed_dim + assert list(query.size()) == [tgt_len, bsz, embed_dim] + + is_tpu = query.device.type == "xla" + + if incremental_state is not None: + saved_state = self._get_input_buffer(incremental_state) + if saved_state is not None and "prev_key" in saved_state: + # previous time steps are cached - no need to recompute + # key and value if they are static + if static_kv: + assert self.encoder_decoder_attention and not self.self_attention + key = value = None + else: + saved_state = None + + if self.self_attention: + q = self.q_proj(query) + k = self.k_proj(query) + v = self.v_proj(query) + elif self.encoder_decoder_attention: + # encoder-decoder attention + q = self.q_proj(query) + if key is None: + assert value is None + k = v = None + else: + k = self.k_proj(key) + v = self.v_proj(key) + + else: + assert key is not None and value is not None + q = self.q_proj(query) + k = self.k_proj(key) + v = self.v_proj(value) + q *= self.scaling + + q = ( + q.contiguous() + .view(tgt_len, bsz * self.num_heads_partition, self.head_dim) + .transpose(0, 1) + ) + if k is not None: + k = ( + k.contiguous() + .view(-1, bsz * self.num_heads_partition, self.head_dim) + .transpose(0, 1) + ) + if v is not None: + v = ( + v.contiguous() + .view(-1, bsz * self.num_heads_partition, self.head_dim) + .transpose(0, 1) + ) + + if saved_state is not None: + # saved states are stored with shape (bsz, num_heads_partition, seq_len, head_dim) + if "prev_key" in saved_state: + _prev_key = saved_state["prev_key"] + assert _prev_key is not None + prev_key = _prev_key.view( + bsz * self.num_heads_partition, -1, self.head_dim + ) + if static_kv: + k = prev_key + else: + assert k is not None + k = torch.cat([prev_key, k], dim=1) + if "prev_value" in saved_state: + _prev_value = saved_state["prev_value"] + assert _prev_value is not None + prev_value = _prev_value.view( + bsz * self.num_heads_partition, -1, self.head_dim + ) + if static_kv: + v = prev_value + else: + assert v is not None + v = torch.cat([prev_value, v], dim=1) + prev_key_padding_mask: Optional[Tensor] = None + if "prev_key_padding_mask" in saved_state: + prev_key_padding_mask = saved_state["prev_key_padding_mask"] + assert k is not None and v is not None + key_padding_mask = ( + ModelParallelMultiheadAttention._append_prev_key_padding_mask( + key_padding_mask=key_padding_mask, + prev_key_padding_mask=prev_key_padding_mask, + batch_size=bsz, + src_len=k.size(1), + static_kv=static_kv, + ) + ) + + saved_state["prev_key"] = k.view( + bsz, self.num_heads_partition, -1, self.head_dim + ) + saved_state["prev_value"] = v.view( + bsz, self.num_heads_partition, -1, self.head_dim + ) + saved_state["prev_key_padding_mask"] = key_padding_mask + # In this branch incremental_state is never None + assert incremental_state is not None + incremental_state = self._set_input_buffer(incremental_state, saved_state) + assert k is not None + src_len = k.size(1) + + # This is part of a workaround to get around fork/join parallelism + # not supporting Optional types. + if key_padding_mask is not None and key_padding_mask.dim() == 0: + key_padding_mask = None + + if key_padding_mask is not None: + assert key_padding_mask.size(0) == bsz + assert key_padding_mask.size(1) == src_len + + attn_weights = torch.bmm(q, k.transpose(1, 2)) + + assert list(attn_weights.size()) == [ + bsz * self.num_heads_partition, + tgt_len, + src_len, + ] + + if attn_mask is not None: + attn_mask = attn_mask.unsqueeze(0) + attn_weights += attn_mask + + if key_padding_mask is not None: + # don't attend to padding symbols + attn_weights = attn_weights.view( + bsz, self.num_heads_partition, tgt_len, src_len + ) + if not is_tpu: + attn_weights = attn_weights.masked_fill( + key_padding_mask.unsqueeze(1).unsqueeze(2).to(torch.bool), + float("-inf"), + ) + else: + attn_weights = attn_weights.transpose(0, 2) + attn_weights = attn_weights.masked_fill(key_padding_mask, float("-inf")) + attn_weights = attn_weights.transpose(0, 2) + attn_weights = attn_weights.view( + bsz * self.num_heads_partition, tgt_len, src_len + ) + + attn_weights_float = utils.softmax(attn_weights, dim=-1) + attn_weights = attn_weights_float.type_as(attn_weights) + + with get_cuda_rng_tracker().fork(): + attn_probs = self.dropout_module(attn_weights) + + assert v is not None + attn = torch.bmm(attn_probs, v) + assert list(attn.size()) == [ + bsz * self.num_heads_partition, + tgt_len, + self.head_dim, + ] + embed_dim_partition = embed_dim // self.model_parallel_size + attn = attn.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim_partition) + attn = self.out_proj(attn) + # return attn_weights None to keep the return type same as single gpu multihead attention + # This will be deprecated. + attn_weights: Optional[Tensor] = None + + return attn, attn_weights + + @staticmethod + def _append_prev_key_padding_mask( + key_padding_mask: Optional[Tensor], + prev_key_padding_mask: Optional[Tensor], + batch_size: int, + src_len: int, + static_kv: bool, + ) -> Optional[Tensor]: + # saved key padding masks have shape (bsz, seq_len) + if prev_key_padding_mask is not None and static_kv: + new_key_padding_mask = prev_key_padding_mask + elif prev_key_padding_mask is not None and key_padding_mask is not None: + new_key_padding_mask = torch.cat( + [prev_key_padding_mask.float(), key_padding_mask.float()], dim=1 + ) + # During incremental decoding, as the padding token enters and + # leaves the frame, there will be a time when prev or current + # is None + elif prev_key_padding_mask is not None: + + filler = torch.zeros(batch_size, src_len - prev_key_padding_mask.size(1)) + if prev_key_padding_mask.is_cuda: + filler = filler.cuda() + new_key_padding_mask = torch.cat( + [prev_key_padding_mask.float(), filler.float()], dim=1 + ) + elif key_padding_mask is not None: + filler = torch.zeros(batch_size, src_len - key_padding_mask.size(1)) + if key_padding_mask.is_cuda: + filler = filler.cuda() + new_key_padding_mask = torch.cat( + [filler.float(), key_padding_mask.float()], dim=1 + ) + else: + new_key_padding_mask = prev_key_padding_mask + return new_key_padding_mask + + def reorder_incremental_state( + self, incremental_state: Dict[str, Dict[str, Optional[Tensor]]], new_order + ): + """Reorder buffered internal state (for incremental generation).""" + input_buffer = self._get_input_buffer(incremental_state) + if input_buffer is not None: + for k in input_buffer.keys(): + if input_buffer[k] is not None: + input_buffer[k] = input_buffer[k].index_select(0, new_order) + incremental_state = self._set_input_buffer(incremental_state, input_buffer) + return incremental_state + + def _get_input_buffer( + self, incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] + ) -> Dict[str, Optional[Tensor]]: + result = self.get_incremental_state(incremental_state, "attn_state") + if result is not None: + return result + else: + empty_result: Dict[str, Optional[Tensor]] = {} + return empty_result + + def _set_input_buffer( + self, + incremental_state: Dict[str, Dict[str, Optional[Tensor]]], + buffer: Dict[str, Optional[Tensor]], + ): + return self.set_incremental_state(incremental_state, "attn_state", buffer) diff --git a/fairseq/model_parallel/modules/transformer_layer.py b/fairseq/model_parallel/modules/transformer_layer.py new file mode 100644 index 0000000000000000000000000000000000000000..7ab53c6e5f12f15562717effb86ab8cb8d6b4fa3 --- /dev/null +++ b/fairseq/model_parallel/modules/transformer_layer.py @@ -0,0 +1,78 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq.model_parallel.modules import ModelParallelMultiheadAttention +from fairseq.modules import TransformerDecoderLayer, TransformerEncoderLayer + + +try: + from fairseq.model_parallel.megatron.mpu import ( + ColumnParallelLinear, + RowParallelLinear, + ) + + has_megatron_submodule = True +except (ImportError, ModuleNotFoundError): + has_megatron_submodule = False + + +class ModelParallelTransformerEncoderLayer(TransformerEncoderLayer): + """Encoder layer block over multiple gpus. + + See "Megatron-LM: https://arxiv.org/pdf/1909.08053.pdf" for more details. + """ + + def build_fc1(self, input_dim, output_dim, q_noise, qn_block_size): + if q_noise > 0: + raise NotImplementedError + return ColumnParallelLinear(input_dim, output_dim, gather_output=False) + + def build_fc2(self, input_dim, output_dim, q_noise, qn_block_size): + if q_noise > 0: + raise NotImplementedError + return RowParallelLinear(input_dim, output_dim, input_is_parallel=True) + + def build_self_attention(self, embed_dim, args, **unused_kwargs): + return ModelParallelMultiheadAttention( + embed_dim, + args.encoder_attention_heads, + dropout=args.attention_dropout, + self_attention=True, + ) + + +class ModelParallelTransformerDecoderLayer(TransformerDecoderLayer): + """Decoder layer block. + + See "Megatron-LM: https://arxiv.org/pdf/1909.08053.pdf" for more details. + """ + + def build_fc1(self, input_dim, output_dim, q_noise, qn_block_size): + if q_noise > 0: + raise NotImplementedError + return ColumnParallelLinear(input_dim, output_dim, gather_output=False) + + def build_fc2(self, input_dim, output_dim, q_noise, qn_block_size): + if q_noise > 0: + raise NotImplementedError + return RowParallelLinear(input_dim, output_dim, input_is_parallel=True) + + def build_self_attention(self, embed_dim, args, **unused_kwargs): + return ModelParallelMultiheadAttention( + embed_dim=embed_dim, + num_heads=args.decoder_attention_heads, + dropout=args.attention_dropout, + self_attention=not getattr(args, "cross_self_attention", False), + ) + + def build_encoder_attention(self, embed_dim, args, **unused_kwargs): + return ModelParallelMultiheadAttention( + embed_dim=embed_dim, + num_heads=args.decoder_attention_heads, + kdim=getattr(args, "encoder_embed_dim", None), + vdim=getattr(args, "encoder_embed_dim", None), + dropout=args.attention_dropout, + encoder_decoder_attention=True, + ) diff --git a/fairseq/models/__init__.py b/fairseq/models/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..61425c8ef5e386c035d97a7ddaf773ff39dde61c --- /dev/null +++ b/fairseq/models/__init__.py @@ -0,0 +1,225 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +"""isort:skip_file""" + +import argparse +import importlib +import os + +from fairseq.dataclass import FairseqDataclass +from fairseq.dataclass.utils import merge_with_parent, populate_dataclass +from hydra.core.config_store import ConfigStore + +from .composite_encoder import CompositeEncoder +from .distributed_fairseq_model import DistributedFairseqModel +from .fairseq_decoder import FairseqDecoder +from .fairseq_encoder import FairseqEncoder +from .fairseq_incremental_decoder import FairseqIncrementalDecoder +from .fairseq_model import ( + BaseFairseqModel, + FairseqEncoderDecoderModel, + FairseqEncoderModel, + FairseqLanguageModel, + FairseqModel, + FairseqMultiModel, +) + + +MODEL_REGISTRY = {} +MODEL_DATACLASS_REGISTRY = {} +ARCH_MODEL_REGISTRY = {} +ARCH_MODEL_NAME_REGISTRY = {} +ARCH_MODEL_INV_REGISTRY = {} +ARCH_CONFIG_REGISTRY = {} + + +__all__ = [ + "BaseFairseqModel", + "CompositeEncoder", + "DistributedFairseqModel", + "FairseqDecoder", + "FairseqEncoder", + "FairseqEncoderDecoderModel", + "FairseqEncoderModel", + "FairseqIncrementalDecoder", + "FairseqLanguageModel", + "FairseqModel", + "FairseqMultiModel", +] + + +def build_model(cfg: FairseqDataclass, task): + + model = None + model_type = getattr(cfg, "_name", None) or getattr(cfg, "arch", None) + + if not model_type and len(cfg) == 1: + # this is hit if config object is nested in directory that is named after model type + + model_type = next(iter(cfg)) + if model_type in MODEL_DATACLASS_REGISTRY: + cfg = cfg[model_type] + else: + raise Exception( + "Could not infer model type from directory. Please add _name field to indicate model type. " + "Available models: " + + str(MODEL_DATACLASS_REGISTRY.keys()) + + " Requested model type: " + + model_type + ) + + if model_type in ARCH_MODEL_REGISTRY: + # case 1: legacy models + model = ARCH_MODEL_REGISTRY[model_type] + elif model_type in MODEL_DATACLASS_REGISTRY: + # case 2: config-driven models + model = MODEL_REGISTRY[model_type] + + if model_type in MODEL_DATACLASS_REGISTRY: + # set defaults from dataclass. note that arch name and model name can be the same + dc = MODEL_DATACLASS_REGISTRY[model_type] + if isinstance(cfg, argparse.Namespace): + cfg = populate_dataclass(dc(), cfg) + else: + cfg = merge_with_parent(dc(), cfg) + + assert model is not None, ( + f"Could not infer model type from {cfg}. " + f"Available models: " + + str(MODEL_DATACLASS_REGISTRY.keys()) + + " Requested model type: " + + model_type + ) + + return model.build_model(cfg, task) + + +def register_model(name, dataclass=None): + """ + New model types can be added to fairseq with the :func:`register_model` + function decorator. + + For example:: + + @register_model('lstm') + class LSTM(FairseqEncoderDecoderModel): + (...) + + .. note:: All models must implement the :class:`BaseFairseqModel` interface. + Typically you will extend :class:`FairseqEncoderDecoderModel` for + sequence-to-sequence tasks or :class:`FairseqLanguageModel` for + language modeling tasks. + + Args: + name (str): the name of the model + """ + + def register_model_cls(cls): + if name in MODEL_REGISTRY: + raise ValueError("Cannot register duplicate model ({})".format(name)) + if not issubclass(cls, BaseFairseqModel): + raise ValueError( + "Model ({}: {}) must extend BaseFairseqModel".format(name, cls.__name__) + ) + MODEL_REGISTRY[name] = cls + if dataclass is not None and not issubclass(dataclass, FairseqDataclass): + raise ValueError( + "Dataclass {} must extend FairseqDataclass".format(dataclass) + ) + + cls.__dataclass = dataclass + if dataclass is not None: + MODEL_DATACLASS_REGISTRY[name] = dataclass + + cs = ConfigStore.instance() + node = dataclass() + node._name = name + cs.store(name=name, group="model", node=node, provider="fairseq") + + @register_model_architecture(name, name) + def noop(_): + pass + + return cls + + return register_model_cls + + +def register_model_architecture(model_name, arch_name): + """ + New model architectures can be added to fairseq with the + :func:`register_model_architecture` function decorator. After registration, + model architectures can be selected with the ``--arch`` command-line + argument. + + For example:: + + @register_model_architecture('lstm', 'lstm_luong_wmt_en_de') + def lstm_luong_wmt_en_de(cfg): + args.encoder_embed_dim = getattr(cfg.model, 'encoder_embed_dim', 1000) + (...) + + The decorated function should take a single argument *cfg*, which is a + :class:`omegaconf.DictConfig`. The decorated function should modify these + arguments in-place to match the desired architecture. + + Args: + model_name (str): the name of the Model (Model must already be + registered) + arch_name (str): the name of the model architecture (``--arch``) + """ + + def register_model_arch_fn(fn): + if model_name not in MODEL_REGISTRY: + raise ValueError( + "Cannot register model architecture for unknown model type ({})".format( + model_name + ) + ) + if arch_name in ARCH_MODEL_REGISTRY: + raise ValueError( + "Cannot register duplicate model architecture ({})".format(arch_name) + ) + if not callable(fn): + raise ValueError( + "Model architecture must be callable ({})".format(arch_name) + ) + ARCH_MODEL_REGISTRY[arch_name] = MODEL_REGISTRY[model_name] + ARCH_MODEL_NAME_REGISTRY[arch_name] = model_name + ARCH_MODEL_INV_REGISTRY.setdefault(model_name, []).append(arch_name) + ARCH_CONFIG_REGISTRY[arch_name] = fn + return fn + + return register_model_arch_fn + + +def import_models(models_dir, namespace): + for file in os.listdir(models_dir): + path = os.path.join(models_dir, file) + if ( + not file.startswith("_") + and not file.startswith(".") + and (file.endswith(".py") or os.path.isdir(path)) + ): + model_name = file[: file.find(".py")] if file.endswith(".py") else file + importlib.import_module(namespace + "." + model_name) + + # extra `model_parser` for sphinx + if model_name in MODEL_REGISTRY: + parser = argparse.ArgumentParser(add_help=False) + group_archs = parser.add_argument_group("Named architectures") + group_archs.add_argument( + "--arch", choices=ARCH_MODEL_INV_REGISTRY[model_name] + ) + group_args = parser.add_argument_group( + "Additional command-line arguments" + ) + MODEL_REGISTRY[model_name].add_args(group_args) + globals()[model_name + "_parser"] = parser + + +# automatically import any Python files in the models/ directory +models_dir = os.path.dirname(__file__) +import_models(models_dir, "fairseq.models") diff --git a/fairseq/models/bart/__init__.py b/fairseq/models/bart/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..a701923f7e5a2a8aa9b75e5580ddea22907f53ee --- /dev/null +++ b/fairseq/models/bart/__init__.py @@ -0,0 +1,7 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .hub_interface import * # noqa +from .model import * # noqa diff --git a/fairseq/models/bart/hub_interface.py b/fairseq/models/bart/hub_interface.py new file mode 100644 index 0000000000000000000000000000000000000000..9afe385b9d93e29f81709b088c945b73639bf583 --- /dev/null +++ b/fairseq/models/bart/hub_interface.py @@ -0,0 +1,208 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import copy +import logging +from typing import Dict, List + +import numpy as np +import torch +import torch.nn as nn +import torch.nn.functional as F +from fairseq import utils +from fairseq.data import encoders +from fairseq.hub_utils import GeneratorHubInterface +from omegaconf import open_dict + + +logger = logging.getLogger(__name__) + + +class BARTHubInterface(GeneratorHubInterface): + """A simple PyTorch Hub interface to BART. + + Usage: https://github.com/pytorch/fairseq/tree/master/examples/bart + """ + + def __init__(self, cfg, task, model): + super().__init__(cfg, task, [model]) + self.model = self.models[0] + + def encode( + self, sentence: str, *addl_sentences, no_separator=True + ) -> torch.LongTensor: + """ + BPE-encode a sentence (or multiple sentences). + + Every sequence begins with a beginning-of-sentence (`<s>`) symbol. + Every sentence ends with an end-of-sentence (`</s>`). + + Example (single sentence): `<s> a b c </s>` + Example (sentence pair): `<s> d e f </s> 1 2 3 </s>` + + The BPE encoding follows GPT-2. One subtle detail is that the GPT-2 BPE + requires leading spaces. For example:: + + >>> bart.encode('Hello world').tolist() + [0, 31414, 232, 2] + >>> bart.encode(' world').tolist() + [0, 232, 2] + >>> bart.encode('world').tolist() + [0, 8331, 2] + """ + tokens = self.bpe.encode(sentence) + if len(tokens.split(" ")) > min(self.max_positions) - 2: + tokens = " ".join(tokens.split(" ")[: min(self.max_positions) - 2]) + bpe_sentence = "<s> " + tokens + " </s>" + for s in addl_sentences: + bpe_sentence += " </s>" if not no_separator else "" + bpe_sentence += " " + self.bpe.encode(s) + " </s>" + tokens = self.task.source_dictionary.encode_line(bpe_sentence, append_eos=False) + return tokens.long() + + def decode(self, tokens: torch.LongTensor): + assert tokens.dim() == 1 + tokens = tokens.cpu().numpy() + if tokens[0] == self.task.source_dictionary.bos(): + tokens = tokens[1:] # remove <s> + eos_mask = tokens == self.task.source_dictionary.eos() + doc_mask = eos_mask[1:] & eos_mask[:-1] + sentences = np.split(tokens, doc_mask.nonzero()[0] + 1) + sentences = [ + self.bpe.decode(self.task.source_dictionary.string(s)) for s in sentences + ] + if len(sentences) == 1: + return sentences[0] + return sentences + + def _build_sample(self, src_tokens: List[torch.LongTensor]): + # assert torch.is_tensor(src_tokens) + dataset = self.task.build_dataset_for_inference( + src_tokens, + [x.numel() for x in src_tokens], + ) + sample = dataset.collater(dataset) + sample = utils.apply_to_sample(lambda tensor: tensor.to(self.device), sample) + return sample + + def generate( + self, + tokenized_sentences: List[torch.LongTensor], + *args, + inference_step_args=None, + skip_invalid_size_inputs=False, + **kwargs + ) -> List[List[Dict[str, torch.Tensor]]]: + inference_step_args = inference_step_args or {} + if "prefix_tokens" in inference_step_args: + raise NotImplementedError("prefix generation not implemented for BART") + res = [] + for batch in self._build_batches(tokenized_sentences, skip_invalid_size_inputs): + src_tokens = batch['net_input']['src_tokens'] + inference_step_args["prefix_tokens"] =src_tokens.new_full( + (src_tokens.size(0), 1), fill_value=self.task.source_dictionary.bos() + ).to(device=self.device) + results = super().generate( + src_tokens, + *args, + inference_step_args=inference_step_args, + skip_invalid_size_inputs=skip_invalid_size_inputs, + **kwargs + ) + for id, hypos in zip(batch['id'].tolist(), results): + res.append((id, hypos)) + res = [hypos for _, hypos in sorted(res, key=lambda x: x[0])] + return res + + def extract_features( + self, tokens: torch.LongTensor, return_all_hiddens: bool = False + ) -> torch.Tensor: + if tokens.dim() == 1: + tokens = tokens.unsqueeze(0) + if tokens.size(-1) > min(self.model.max_positions()): + raise ValueError( + "tokens exceeds maximum length: {} > {}".format( + tokens.size(-1), self.model.max_positions() + ) + ) + tokens.to(device=self.device), + prev_output_tokens = tokens.clone() + + prev_output_tokens[:, 0] = tokens.gather( + 1, + (tokens.ne(self.task.source_dictionary.pad()).sum(dim=1) - 1).unsqueeze(-1), + ).squeeze() + + prev_output_tokens[:, 1:] = tokens[:, :-1] + features, extra = self.model( + src_tokens=tokens, + src_lengths=None, + prev_output_tokens=prev_output_tokens, + features_only=True, + return_all_hiddens=return_all_hiddens, + ) + if return_all_hiddens: + # convert from T x B x C -> B x T x C + inner_states = extra["inner_states"] + return [inner_state.transpose(0, 1) for inner_state in inner_states] + else: + return features # just the last layer's features + + def register_classification_head( + self, name: str, num_classes: int = None, embedding_size: int = None, **kwargs + ): + self.model.register_classification_head( + name, num_classes=num_classes, embedding_size=embedding_size, **kwargs + ) + + def predict(self, head: str, tokens: torch.LongTensor, return_logits: bool = False): + if tokens.dim() == 1: + tokens = tokens.unsqueeze(0) + features = self.extract_features(tokens.to(device=self.device)) + sentence_representation = features[ + tokens.eq(self.task.source_dictionary.eos()), : + ].view(features.size(0), -1, features.size(-1))[:, -1, :] + + logits = self.model.classification_heads[head](sentence_representation) + if return_logits: + return logits + return F.log_softmax(logits, dim=-1) + + def fill_mask( + self, + masked_inputs: List[str], + topk: int = 5, + match_source_len: bool = True, + **generate_kwargs + ): + masked_token = '<mask>' + batch_tokens = [] + for masked_input in masked_inputs: + assert masked_token in masked_input, \ + "please add one {} token for the input".format(masked_token) + + text_spans = masked_input.split(masked_token) + text_spans_bpe = (' {0} '.format(masked_token)).join( + [self.bpe.encode(text_span.rstrip()) for text_span in text_spans] + ).strip() + tokens = self.task.source_dictionary.encode_line( + '<s> ' + text_spans_bpe + ' </s>', + append_eos=False, + add_if_not_exist=False, + ).long() + batch_tokens.append(tokens) + + # ensure beam size is at least as big as topk + generate_kwargs['beam'] = max( + topk, + generate_kwargs.get('beam', -1), + ) + generate_kwargs['match_source_len'] = match_source_len + batch_hypos = self.generate(batch_tokens, **generate_kwargs) + + return [ + [(self.decode(hypo['tokens']), hypo['score']) for hypo in hypos[:topk]] + for hypos in batch_hypos + ] diff --git a/fairseq/models/bart/model.py b/fairseq/models/bart/model.py new file mode 100644 index 0000000000000000000000000000000000000000..71d0b27cd2c0655fe3b00479b672d6d042a4d5ed --- /dev/null +++ b/fairseq/models/bart/model.py @@ -0,0 +1,384 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +BART: Denoising Sequence-to-Sequence Pre-training for +Natural Language Generation, Translation, and Comprehension +""" +from typing import Optional + +import logging + +import torch +import torch.nn as nn +from fairseq import utils +from fairseq.models import register_model, register_model_architecture +from fairseq.models.transformer import TransformerModel +from fairseq.modules.transformer_sentence_encoder import init_bert_params + +from .hub_interface import BARTHubInterface + + +logger = logging.getLogger(__name__) + + +@register_model("bart") +class BARTModel(TransformerModel): + __jit_unused_properties__ = ["supported_targets"] + + @classmethod + def hub_models(cls): + return { + "bart.base": "http://dl.fbaipublicfiles.com/fairseq/models/bart.base.tar.gz", + "bart.large": "http://dl.fbaipublicfiles.com/fairseq/models/bart.large.tar.gz", + "bart.large.mnli": "http://dl.fbaipublicfiles.com/fairseq/models/bart.large.mnli.tar.gz", + "bart.large.cnn": "http://dl.fbaipublicfiles.com/fairseq/models/bart.large.cnn.tar.gz", + "bart.large.xsum": "http://dl.fbaipublicfiles.com/fairseq/models/bart.large.xsum.tar.gz", + } + + def __init__(self, args, encoder, decoder): + super().__init__(args, encoder, decoder) + + # We follow BERT's random weight initialization + self.apply(init_bert_params) + + self.classification_heads = nn.ModuleDict() + if hasattr(self.encoder, "dictionary"): + self.eos: int = self.encoder.dictionary.eos() + + @staticmethod + def add_args(parser): + super(BARTModel, BARTModel).add_args(parser) + parser.add_argument( + "--pooler-dropout", + type=float, + metavar="D", + help="dropout probability in the masked_lm pooler layers", + ) + parser.add_argument( + "--pooler-activation-fn", + choices=utils.get_available_activation_fns(), + help="activation function to use for pooler layer", + ) + parser.add_argument( + "--spectral-norm-classification-head", + action="store_true", + help="Apply spectral normalization on the classification head", + ) + + @property + def supported_targets(self): + return {"self"} + + def forward( + self, + src_tokens, + src_lengths, + prev_output_tokens, + features_only: bool = False, + classification_head_name: Optional[str] = None, + token_embeddings: Optional[torch.Tensor] = None, + return_all_hiddens: bool = True, + alignment_layer: Optional[int] = None, + alignment_heads: Optional[int] = None, + ): + if classification_head_name is not None: + features_only = True + + encoder_out = self.encoder( + src_tokens, + src_lengths=src_lengths, + token_embeddings=token_embeddings, + return_all_hiddens=return_all_hiddens + ) + x, extra = self.decoder( + prev_output_tokens, + encoder_out=encoder_out, + features_only=features_only, + alignment_layer=alignment_layer, + alignment_heads=alignment_heads, + src_lengths=src_lengths, + return_all_hiddens=return_all_hiddens, + ) + eos: int = self.eos + if classification_head_name is not None: + sentence_representation = x[ + src_tokens.eq(eos), : + ].view(x.size(0), -1, x.size(-1))[:, -1, :] + for k, head in self.classification_heads.items(): + # for torch script only supports iteration + if k == classification_head_name: + x = head(sentence_representation) + break + return x, extra + + @classmethod + def from_pretrained( + cls, + model_name_or_path, + checkpoint_file="model.pt", + data_name_or_path=".", + bpe="gpt2", + sample_break_mode="eos", + **kwargs, + ): + from fairseq import hub_utils + + x = hub_utils.from_pretrained( + model_name_or_path, + checkpoint_file, + data_name_or_path, + archive_map=cls.hub_models(), + bpe=bpe, + load_checkpoint_heads=True, + sample_break_mode=sample_break_mode, + **kwargs, + ) + return BARTHubInterface(x["args"], x["task"], x["models"][0]) + + def register_classification_head( + self, name, num_classes=None, inner_dim=None, **kwargs + ): + """Register a classification head.""" + logger.info("Registering classification head: {0}".format(name)) + if name in self.classification_heads: + prev_num_classes = self.classification_heads[name].out_proj.out_features + prev_inner_dim = self.classification_heads[name].dense.out_features + if num_classes != prev_num_classes or inner_dim != prev_inner_dim: + logger.warning( + 're-registering head "{}" with num_classes {} (prev: {}) ' + "and inner_dim {} (prev: {})".format( + name, num_classes, prev_num_classes, inner_dim, prev_inner_dim + ) + ) + self.classification_heads[name] = BARTClassificationHead( + input_dim=self.args.encoder_embed_dim, + inner_dim=inner_dim or self.args.encoder_embed_dim, + num_classes=num_classes, + activation_fn=self.args.pooler_activation_fn, + pooler_dropout=self.args.pooler_dropout, + do_spectral_norm=getattr( + self.args, "spectral_norm_classification_head", False + ), + ) + + def upgrade_state_dict_named(self, state_dict, name): + super().upgrade_state_dict_named(state_dict, name) + + prefix = name + "." if name != "" else "" + current_head_names = ( + [] + if not hasattr(self, "classification_heads") + else self.classification_heads.keys() + ) + + # Handle new classification heads present in the state dict. + keys_to_delete = [] + for k in state_dict.keys(): + if not k.startswith(prefix + "classification_heads."): + continue + + head_name = k[len(prefix + "classification_heads.") :].split(".")[0] + num_classes = state_dict[ + prefix + "classification_heads." + head_name + ".out_proj.weight" + ].size(0) + inner_dim = state_dict[ + prefix + "classification_heads." + head_name + ".dense.weight" + ].size(0) + + if getattr(self.args, "load_checkpoint_heads", False): + if head_name not in current_head_names: + self.register_classification_head(head_name, num_classes, inner_dim) + else: + if head_name not in current_head_names: + logger.warning( + "deleting classification head ({}) from checkpoint " + "not present in current model: {}".format(head_name, k) + ) + keys_to_delete.append(k) + elif ( + num_classes + != self.classification_heads[head_name].out_proj.out_features + or inner_dim + != self.classification_heads[head_name].dense.out_features + ): + logger.warning( + "deleting classification head ({}) from checkpoint " + "with different dimensions than current model: {}".format( + head_name, k + ) + ) + keys_to_delete.append(k) + for k in keys_to_delete: + del state_dict[k] + + def truncate_emb(key): + if key in state_dict: + state_dict[key] = state_dict[key][:-1, :] + + # When finetuning on translation task, remove last row of + # embedding matrix that corresponds to mask_idx token. + loaded_dict_size = state_dict["encoder.embed_tokens.weight"].size(0) + if ( + loaded_dict_size == len(self.encoder.dictionary) + 1 + and "<mask>" not in self.encoder.dictionary + ): + truncate_emb("encoder.embed_tokens.weight") + truncate_emb("decoder.embed_tokens.weight") + truncate_emb("encoder.output_projection.weight") + truncate_emb("decoder.output_projection.weight") + + # When continued pretraining on new set of languages for mbart, + # add extra lang embeddings at the end of embed_tokens. + # Note: newly added languages are assumed to have been added at the end. + if self.args.task == "multilingual_denoising" and loaded_dict_size < len( + self.encoder.dictionary + ): + logger.info( + "Adding extra language embeddings not found in pretrained model for " + "continued pretraining of MBART on new set of languages." + ) + loaded_mask_token_embedding = state_dict["encoder.embed_tokens.weight"][ + -1, : + ] + + num_langids_to_add = len(self.encoder.dictionary) - loaded_dict_size + embed_dim = state_dict["encoder.embed_tokens.weight"].size(1) + + new_lang_embed_to_add = torch.zeros(num_langids_to_add, embed_dim) + nn.init.normal_(new_lang_embed_to_add, mean=0, std=embed_dim ** -0.5) + new_lang_embed_to_add = new_lang_embed_to_add.to( + dtype=state_dict["encoder.embed_tokens.weight"].dtype, + ) + + state_dict["encoder.embed_tokens.weight"] = torch.cat( + [ + state_dict["encoder.embed_tokens.weight"][ + : loaded_dict_size - 1, : + ], + new_lang_embed_to_add, + loaded_mask_token_embedding.unsqueeze(0), + ] + ) + state_dict["decoder.embed_tokens.weight"] = torch.cat( + [ + state_dict["decoder.embed_tokens.weight"][ + : loaded_dict_size - 1, : + ], + new_lang_embed_to_add, + loaded_mask_token_embedding.unsqueeze(0), + ] + ) + + # Copy any newly-added classification heads into the state dict + # with their current weights. + if hasattr(self, "classification_heads"): + cur_state = self.classification_heads.state_dict() + for k, v in cur_state.items(): + if prefix + "classification_heads." + k not in state_dict: + logger.info("Overwriting " + prefix + "classification_heads." + k) + state_dict[prefix + "classification_heads." + k] = v + + +class BARTClassificationHead(nn.Module): + """Head for sentence-level classification tasks.""" + + def __init__( + self, + input_dim, + inner_dim, + num_classes, + activation_fn, + pooler_dropout, + do_spectral_norm=False, + ): + super().__init__() + self.dense = nn.Linear(input_dim, inner_dim) + self.activation_fn = utils.get_activation_fn(activation_fn) + self.dropout = nn.Dropout(p=pooler_dropout) + self.out_proj = nn.Linear(inner_dim, num_classes) + + if do_spectral_norm: + self.out_proj = torch.nn.utils.spectral_norm(self.out_proj) + + def forward(self, features, **kwargs): + x = features + x = self.dropout(x) + x = self.dense(x) + x = self.activation_fn(x) + x = self.dropout(x) + x = self.out_proj(x) + return x + + +@register_model_architecture("bart", "bart_large") +def bart_large_architecture(args): + args.encoder_embed_path = getattr(args, "encoder_embed_path", None) + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 1024) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 4 * 1024) + args.encoder_layers = getattr(args, "encoder_layers", 12) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 16) + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False) + args.encoder_learned_pos = getattr(args, "encoder_learned_pos", True) + args.decoder_embed_path = getattr(args, "decoder_embed_path", None) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", args.encoder_embed_dim) + args.decoder_ffn_embed_dim = getattr( + args, "decoder_ffn_embed_dim", args.encoder_ffn_embed_dim + ) + args.decoder_layers = getattr(args, "decoder_layers", 12) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 16) + args.decoder_normalize_before = getattr(args, "decoder_normalize_before", False) + args.decoder_learned_pos = getattr(args, "decoder_learned_pos", True) + args.attention_dropout = getattr(args, "attention_dropout", 0.0) + args.relu_dropout = getattr(args, "relu_dropout", 0.0) + args.dropout = getattr(args, "dropout", 0.1) + args.max_target_positions = getattr(args, "max_target_positions", 1024) + args.max_source_positions = getattr(args, "max_source_positions", 1024) + args.adaptive_softmax_cutoff = getattr(args, "adaptive_softmax_cutoff", None) + args.adaptive_softmax_dropout = getattr(args, "adaptive_softmax_dropout", 0) + args.share_decoder_input_output_embed = getattr( + args, "share_decoder_input_output_embed", True + ) + args.share_all_embeddings = getattr(args, "share_all_embeddings", True) + + args.decoder_output_dim = getattr( + args, "decoder_output_dim", args.decoder_embed_dim + ) + args.decoder_input_dim = getattr(args, "decoder_input_dim", args.decoder_embed_dim) + + args.no_scale_embedding = getattr(args, "no_scale_embedding", True) + args.layernorm_embedding = getattr(args, "layernorm_embedding", True) + + args.activation_fn = getattr(args, "activation_fn", "gelu") + args.pooler_activation_fn = getattr(args, "pooler_activation_fn", "tanh") + args.pooler_dropout = getattr(args, "pooler_dropout", 0.0) + + +@register_model_architecture("bart", "bart_base") +def bart_base_architecture(args): + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 768) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 4 * 768) + args.encoder_layers = getattr(args, "encoder_layers", 6) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 12) + args.decoder_layers = getattr(args, "decoder_layers", 6) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 12) + bart_large_architecture(args) + + +@register_model_architecture("bart", "mbart_large") +def mbart_large_architecture(args): + args.no_scale_embedding = getattr(args, "no_scale_embedding", False) + bart_large_architecture(args) + + +@register_model_architecture("bart", "mbart_base") +def mbart_base_architecture(args): + args.no_scale_embedding = getattr(args, "no_scale_embedding", False) + bart_base_architecture(args) + + +@register_model_architecture("bart", "mbart_base_wmt20") +def mbart_base_wmt20_architecture(args): + args.layernorm_embedding = getattr(args, "layernorm_embedding", False) + mbart_base_architecture(args) diff --git a/fairseq/models/composite_encoder.py b/fairseq/models/composite_encoder.py new file mode 100644 index 0000000000000000000000000000000000000000..4e20fe3a833a2d87876cbec294ad2bebfba7f591 --- /dev/null +++ b/fairseq/models/composite_encoder.py @@ -0,0 +1,57 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .fairseq_encoder import FairseqEncoder + + +class CompositeEncoder(FairseqEncoder): + """ + A wrapper around a dictionary of :class:`FairseqEncoder` objects. + + We run forward on each encoder and return a dictionary of outputs. The first + encoder's dictionary is used for initialization. + + Args: + encoders (dict): a dictionary of :class:`FairseqEncoder` objects. + """ + + def __init__(self, encoders): + super().__init__(next(iter(encoders.values())).dictionary) + self.encoders = encoders + for key in self.encoders: + self.add_module(key, self.encoders[key]) + + def forward(self, src_tokens, src_lengths): + """ + Args: + src_tokens (LongTensor): tokens in the source language of shape + `(batch, src_len)` + src_lengths (LongTensor): lengths of each source sentence of shape + `(batch)` + + Returns: + dict: + the outputs from each Encoder + """ + encoder_out = {} + for key in self.encoders: + encoder_out[key] = self.encoders[key](src_tokens, src_lengths) + return encoder_out + + def reorder_encoder_out(self, encoder_out, new_order): + """Reorder encoder output according to new_order.""" + for key in self.encoders: + encoder_out[key] = self.encoders[key].reorder_encoder_out( + encoder_out[key], new_order + ) + return encoder_out + + def max_positions(self): + return min(self.encoders[key].max_positions() for key in self.encoders) + + def upgrade_state_dict(self, state_dict): + for key in self.encoders: + self.encoders[key].upgrade_state_dict(state_dict) + return state_dict diff --git a/fairseq/models/distributed_fairseq_model.py b/fairseq/models/distributed_fairseq_model.py new file mode 100644 index 0000000000000000000000000000000000000000..06905455fd615ea962d8478c6093e7b4bbcc83c4 --- /dev/null +++ b/fairseq/models/distributed_fairseq_model.py @@ -0,0 +1,145 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import os +import signal +import threading + +import torch +import torch.nn as nn +from torch.nn.parallel import DistributedDataParallel + +from fairseq.distributed import ( + DistributedTimeoutWrapper, + LegacyDistributedDataParallel, + ModuleProxyWrapper, + TPUDistributedDataParallel, +) + + +logger = logging.getLogger(__name__) + + +_GOSSIP_DISABLED = False +try: + import gossip +except ImportError: + _GOSSIP_DISABLED = True + + +def DistributedFairseqModel(args, model, process_group, device): + """ + Wrap a *model* to support distributed data parallel training. + + This is similar to the built-in DistributedDataParallel, but allows + additional configuration of the DistributedDataParallel class to + use, and also provides easier access to the wrapped model by + forwarding requests for missing attributes to the wrapped model. + + Args: + args (argparse.Namespace): fairseq args + model (BaseFairseqModel): model to wrap + process_group: the c10d process group to be used for distributed data + parallel all-reduction. + device: device to move model to + """ + assert isinstance(model, nn.Module) + if args.tpu: + wrapped_model = TPUDistributedDataParallel( + module=model.to(device), + process_group=process_group, + ) + # forward missing getattr and state_dict/load_state_dict to orig model + wrapped_model = ModuleProxyWrapper(wrapped_model) + elif args.ddp_backend in {"c10d", "pytorch_ddp"}: + wrapped_model = DistributedDataParallel( + module=model.to(device), + device_ids=[args.device_id], + output_device=args.device_id, + broadcast_buffers=args.broadcast_buffers, + bucket_cap_mb=args.bucket_cap_mb, + process_group=process_group, + find_unused_parameters=args.find_unused_parameters, + ) + if args.ddp_comm_hook == "fp16": + logger.info("enable fp16 communication hook in DDP") + try: + from torch.distributed.algorithms.ddp_comm_hooks import ( + register_ddp_comm_hook, + DDPCommHookType, + ) + except: + logger.error( + "Could not import from torch.distributed.algorithms.ddp_comm_hooks; you may need to update your pytorch version" + ) + raise + + register_ddp_comm_hook(DDPCommHookType.FP16_COMPRESS, wrapped_model) + # forward missing getattr and state_dict/load_state_dict to orig model + wrapped_model = ModuleProxyWrapper(wrapped_model) + elif args.ddp_backend in {"no_c10d", "legacy_ddp"}: + wrapped_model = LegacyDistributedDataParallel( + module=model.to(device), + buffer_size=2 ** 28, + process_group=process_group, + ) + # forward missing getattr and state_dict/load_state_dict to orig model + wrapped_model = ModuleProxyWrapper(wrapped_model) + elif args.ddp_backend == "slow_mo": + if _GOSSIP_DISABLED: + raise ImportError( + "Cannot find gossip library. Please install from: " + "github.com/facebookresearch/stochastic_gradient_push" + ) + + # The values of slowmo_momentum below were obtained by tuning on the + # En-De 16 dataset by training the transformer_wmt_en_de_large model + if args.slowmo_momentum is None: + if args.distributed_world_size <= 16: + args.slowmo_momentum = 0.0 + elif args.distributed_world_size <= 32: + args.slowmo_momentum = 0.2 + elif args.distributed_world_size <= 64: + args.slowmo_momentum = 0.5 + else: + args.slowmo_momentum = 0.6 + + wrapped_model = gossip.GossipDataParallel( + module=model.to(device), + device_ids=[args.device_id], + output_device=args.device_id, + broadcast_buffers=args.broadcast_buffers, + nprocs_per_node=args.nprocs_per_node, + slowmo_momentum=args.slowmo_momentum, + localsgd=(args.slowmo_algorithm == "LocalSGD"), + localsgd_frequency=args.localsgd_frequency, + ) + # forward missing getattr and state_dict/load_state_dict to orig model + wrapped_model = ModuleProxyWrapper(wrapped_model) + elif args.ddp_backend == "fully_sharded": + try: + from fairscale.nn.data_parallel import FullyShardedDataParallel as FSDP + except ImportError: + raise ImportError( + "Cannot find FullyShardedDataParallel. " + "Please install fairscale with: pip install fairscale" + ) + assert isinstance(model, FSDP), "expected model to already be wrapped in FSDP" + wrapped_model = model + if args.memory_efficient_fp16: + wrapped_model = wrapped_model.half() + if not args.cpu_offload: + wrapped_model = wrapped_model.to(device=device) + else: + raise ValueError("Unknown --ddp-backend: " + args.ddp_backend) + + # kill hung distributed jobs after a timeout + if getattr(args, "heartbeat_timeout", -1) > 0: + wrapped_model = DistributedTimeoutWrapper( + wrapped_model, timeout=getattr(args, "heartbeat_timeout", -1) + ) + + return wrapped_model diff --git a/fairseq/models/fairseq_decoder.py b/fairseq/models/fairseq_decoder.py new file mode 100644 index 0000000000000000000000000000000000000000..4f1e8b52a2e0a50199050f11cc613ab02ca9febe --- /dev/null +++ b/fairseq/models/fairseq_decoder.py @@ -0,0 +1,105 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from typing import Dict, List, Optional, Tuple + +import torch.nn as nn +from fairseq import utils +from torch import Tensor + + +class FairseqDecoder(nn.Module): + """Base class for decoders.""" + + def __init__(self, dictionary): + super().__init__() + self.dictionary = dictionary + self.onnx_trace = False + self.adaptive_softmax = None + + + def forward(self, prev_output_tokens, encoder_out=None, **kwargs): + """ + Args: + prev_output_tokens (LongTensor): shifted output tokens of shape + `(batch, tgt_len)`, for teacher forcing + encoder_out (dict, optional): output from the encoder, used for + encoder-side attention + + Returns: + tuple: + - the decoder's output of shape `(batch, tgt_len, vocab)` + - a dictionary with any model-specific outputs + """ + x, extra = self.extract_features( + prev_output_tokens, encoder_out=encoder_out, **kwargs + ) + x = self.output_layer(x) + return x, extra + + def extract_features(self, prev_output_tokens, encoder_out=None, **kwargs): + """ + Returns: + tuple: + - the decoder's features of shape `(batch, tgt_len, embed_dim)` + - a dictionary with any model-specific outputs + """ + raise NotImplementedError + + def output_layer(self, features, **kwargs): + """ + Project features to the default output size, e.g., vocabulary size. + + Args: + features (Tensor): features returned by *extract_features*. + """ + raise NotImplementedError + + def get_normalized_probs( + self, + net_output: Tuple[Tensor, Optional[Dict[str, List[Optional[Tensor]]]]], + log_probs: bool, + sample: Optional[Dict[str, Tensor]] = None, + ): + """Get normalized probabilities (or log probs) from a net's output.""" + return self.get_normalized_probs_scriptable(net_output, log_probs, sample) + + # TorchScript doesn't support super() method so that the scriptable Subclass + # can't access the base class model in Torchscript. + # Current workaround is to add a helper function with different name and + # call the helper function from scriptable Subclass. + def get_normalized_probs_scriptable( + self, + net_output: Tuple[Tensor, Optional[Dict[str, List[Optional[Tensor]]]]], + log_probs: bool, + sample: Optional[Dict[str, Tensor]] = None, + ): + """Get normalized probabilities (or log probs) from a net's output.""" + + if hasattr(self, "adaptive_softmax") and self.adaptive_softmax is not None: + if sample is not None: + assert "target" in sample + target = sample["target"] + else: + target = None + out = self.adaptive_softmax.get_log_prob(net_output[0], target=target) + return out.exp_() if not log_probs else out + + logits = net_output[0] + if log_probs: + return utils.log_softmax(logits, dim=-1, onnx_trace=self.onnx_trace) + else: + return utils.softmax(logits, dim=-1, onnx_trace=self.onnx_trace) + + def max_positions(self): + """Maximum input length supported by the decoder.""" + return 1e6 # an arbitrary large number + + def upgrade_state_dict_named(self, state_dict, name): + """Upgrade old state dicts to work with newer code.""" + return state_dict + + def prepare_for_onnx_export_(self): + self.onnx_trace = True diff --git a/fairseq/models/fairseq_encoder.py b/fairseq/models/fairseq_encoder.py new file mode 100644 index 0000000000000000000000000000000000000000..08cbde15a46e9b6d58e11c2f6052e7cf2d0cc8b2 --- /dev/null +++ b/fairseq/models/fairseq_encoder.py @@ -0,0 +1,92 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from typing import Dict, List, NamedTuple, Optional + +import torch +import torch.nn as nn +from torch import Tensor + + +EncoderOut = NamedTuple( + "EncoderOut", + [ + ("encoder_out", Tensor), # T x B x C + ("encoder_padding_mask", Optional[Tensor]), # B x T + ("encoder_embedding", Optional[Tensor]), # B x T x C + ("encoder_states", Optional[List[Tensor]]), # List[T x B x C] + ("src_tokens", Optional[Tensor]), # B x T + ("src_lengths", Optional[Tensor]), # B x 1 + ], +) + + +class FairseqEncoder(nn.Module): + """Base class for encoders.""" + + def __init__(self, dictionary): + super().__init__() + self.dictionary = dictionary + + def forward(self, src_tokens, src_lengths=None, **kwargs): + """ + Args: + src_tokens (LongTensor): tokens in the source language of shape + `(batch, src_len)` + src_lengths (LongTensor): lengths of each source sentence of shape + `(batch)` + """ + raise NotImplementedError + + def forward_torchscript(self, net_input: Dict[str, Tensor]): + """A TorchScript-compatible version of forward. + + Encoders which use additional arguments may want to override + this method for TorchScript compatibility. + """ + if torch.jit.is_scripting(): + return self.forward( + src_tokens=net_input["src_tokens"], + src_lengths=net_input["src_lengths"], + ) + else: + return self.forward_non_torchscript(net_input) + + @torch.jit.unused + def forward_non_torchscript(self, net_input: Dict[str, Tensor]): + encoder_input = { + k: v for k, v in net_input.items() if k != "prev_output_tokens" + } + return self.forward(**encoder_input) + + def reorder_encoder_out(self, encoder_out, new_order): + """ + Reorder encoder output according to `new_order`. + + Args: + encoder_out: output from the ``forward()`` method + new_order (LongTensor): desired order + + Returns: + `encoder_out` rearranged according to `new_order` + """ + raise NotImplementedError + + def max_positions(self): + """Maximum input length supported by the encoder.""" + return 1e6 # an arbitrary large number + + def upgrade_state_dict_named(self, state_dict, name): + """Upgrade old state dicts to work with newer code.""" + return state_dict + + def set_num_updates(self, num_updates): + """State from trainer to pass along to model at every update.""" + + def _apply(m): + if hasattr(m, "set_num_updates") and m != self: + m.set_num_updates(num_updates) + + self.apply(_apply) diff --git a/fairseq/models/fairseq_incremental_decoder.py b/fairseq/models/fairseq_incremental_decoder.py new file mode 100644 index 0000000000000000000000000000000000000000..cc72a0f8f3da238a8ce846240e5008d91ce1bc1a --- /dev/null +++ b/fairseq/models/fairseq_incremental_decoder.py @@ -0,0 +1,118 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +from typing import Dict, Optional + +from fairseq.incremental_decoding_utils import with_incremental_state +from fairseq.models import FairseqDecoder +from torch import Tensor + + +logger = logging.getLogger(__name__) + + +@with_incremental_state +class FairseqIncrementalDecoder(FairseqDecoder): + """Base class for incremental decoders. + + Incremental decoding is a special mode at inference time where the Model + only receives a single timestep of input corresponding to the previous + output token (for teacher forcing) and must produce the next output + *incrementally*. Thus the model must cache any long-term state that is + needed about the sequence, e.g., hidden states, convolutional states, etc. + + Compared to the standard :class:`FairseqDecoder` interface, the incremental + decoder interface allows :func:`forward` functions to take an extra keyword + argument (*incremental_state*) that can be used to cache state across + time-steps. + + The :class:`FairseqIncrementalDecoder` interface also defines the + :func:`reorder_incremental_state` method, which is used during beam search + to select and reorder the incremental state based on the selection of beams. + + To learn more about how incremental decoding works, refer to `this blog + <http://www.telesens.co/2019/04/21/understanding-incremental-decoding-in-fairseq/>`_. + """ + + def __init__(self, dictionary): + super().__init__(dictionary) + + def forward( + self, prev_output_tokens, encoder_out=None, incremental_state=None, **kwargs + ): + """ + Args: + prev_output_tokens (LongTensor): shifted output tokens of shape + `(batch, tgt_len)`, for teacher forcing + encoder_out (dict, optional): output from the encoder, used for + encoder-side attention + incremental_state (dict, optional): dictionary used for storing + state during :ref:`Incremental decoding` + + Returns: + tuple: + - the decoder's output of shape `(batch, tgt_len, vocab)` + - a dictionary with any model-specific outputs + """ + raise NotImplementedError + + def extract_features( + self, prev_output_tokens, encoder_out=None, incremental_state=None, **kwargs + ): + """ + Returns: + tuple: + - the decoder's features of shape `(batch, tgt_len, embed_dim)` + - a dictionary with any model-specific outputs + """ + raise NotImplementedError + + def reorder_incremental_state( + self, + incremental_state: Dict[str, Dict[str, Optional[Tensor]]], + new_order: Tensor, + ): + """Reorder incremental state. + + This will be called when the order of the input has changed from the + previous time step. A typical use case is beam search, where the input + order changes between time steps based on the selection of beams. + """ + pass + + def reorder_incremental_state_scripting( + self, + incremental_state: Dict[str, Dict[str, Optional[Tensor]]], + new_order: Tensor, + ): + """Main entry point for reordering the incremental state. + + Due to limitations in TorchScript, we call this function in + :class:`fairseq.sequence_generator.SequenceGenerator` instead of + calling :func:`reorder_incremental_state` directly. + """ + for module in self.modules(): + if hasattr(module, "reorder_incremental_state"): + result = module.reorder_incremental_state(incremental_state, new_order) + if result is not None: + incremental_state = result + + def set_beam_size(self, beam_size): + """Sets the beam size in the decoder and all children.""" + if getattr(self, "_beam_size", -1) != beam_size: + seen = set() + + def apply_set_beam_size(module): + if ( + module != self + and hasattr(module, "set_beam_size") + and module not in seen + ): + seen.add(module) + module.set_beam_size(beam_size) + + self.apply(apply_set_beam_size) + self._beam_size = beam_size diff --git a/fairseq/models/fairseq_model.py b/fairseq/models/fairseq_model.py new file mode 100644 index 0000000000000000000000000000000000000000..e55c7ba1ad90f4e2f12db6c814d04a90c4e3b77c --- /dev/null +++ b/fairseq/models/fairseq_model.py @@ -0,0 +1,569 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +Base classes for various fairseq models. +""" + +import logging +from argparse import Namespace +from typing import Dict, List, Optional, Tuple + +import torch +import torch.nn as nn +import torch.nn.functional as F +from fairseq import utils +from fairseq.data import Dictionary +from fairseq.dataclass.utils import ( + convert_namespace_to_omegaconf, + gen_parser_from_dataclass, +) +from fairseq.models import FairseqDecoder, FairseqEncoder +from omegaconf import DictConfig +from torch import Tensor + + +logger = logging.getLogger(__name__) + + +def check_type(module, expected_type): + if hasattr(module, "unwrapped_module"): + assert isinstance(module.unwrapped_module, expected_type), \ + f"{type(module.unwrapped_module)} != {expected_type}" + else: + assert isinstance(module, expected_type), f"{type(module)} != {expected_type}" + + +class BaseFairseqModel(nn.Module): + """Base class for fairseq models.""" + + def __init__(self): + super().__init__() + self._is_generation_fast = False + + @classmethod + def add_args(cls, parser): + """Add model-specific arguments to the parser.""" + dc = getattr(cls, "__dataclass", None) + if dc is not None: + # do not set defaults so that settings defaults from various architectures still works + gen_parser_from_dataclass(parser, dc(), delete_default=True) + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + raise NotImplementedError("Model must implement the build_model method") + + def get_targets(self, sample, net_output): + """Get targets from either the sample or the net's output.""" + return sample["target"] + + def get_normalized_probs( + self, + net_output: Tuple[Tensor, Optional[Dict[str, List[Optional[Tensor]]]]], + log_probs: bool, + sample: Optional[Dict[str, Tensor]] = None, + ): + """Get normalized probabilities (or log probs) from a net's output.""" + return self.get_normalized_probs_scriptable(net_output, log_probs, sample) + + # TorchScript doesn't support super() method so that the scriptable Subclass + # can't access the base class model in Torchscript. + # Current workaround is to add a helper function with different name and + # call the helper function from scriptable Subclass. + def get_normalized_probs_scriptable( + self, + net_output: Tuple[Tensor, Optional[Dict[str, List[Optional[Tensor]]]]], + log_probs: bool, + sample: Optional[Dict[str, Tensor]] = None, + ): + """Scriptable helper function for get_normalized_probs in ~BaseFairseqModel""" + if hasattr(self, "decoder"): + return self.decoder.get_normalized_probs(net_output, log_probs, sample) + elif torch.is_tensor(net_output): + # syntactic sugar for simple models which don't have a decoder + # (e.g., the classification tutorial) + logits = net_output.float() + if log_probs: + return F.log_softmax(logits, dim=-1) + else: + return F.softmax(logits, dim=-1) + raise NotImplementedError + + def extract_features(self, *args, **kwargs): + """Similar to *forward* but only return features.""" + return self(*args, **kwargs) + + def max_positions(self): + """Maximum length supported by the model.""" + return None + + def load_state_dict( + self, + state_dict, + strict=True, + model_cfg: Optional[DictConfig] = None, + args: Optional[Namespace] = None, + ): + """Copies parameters and buffers from *state_dict* into this module and + its descendants. + + Overrides the method in :class:`nn.Module`. Compared with that method + this additionally "upgrades" *state_dicts* from old checkpoints. + """ + + if model_cfg is None and args is not None: + logger.warn("using 'args' is deprecated, please update your code to use dataclass config") + model_cfg = convert_namespace_to_omegaconf(args).model + + self.upgrade_state_dict(state_dict) + + from fairseq.checkpoint_utils import prune_state_dict + + new_state_dict = prune_state_dict(state_dict, model_cfg) + return super().load_state_dict(new_state_dict, strict) + + def upgrade_state_dict(self, state_dict): + """Upgrade old state dicts to work with newer code.""" + self.upgrade_state_dict_named(state_dict, "") + + def upgrade_state_dict_named(self, state_dict, name): + """Upgrade old state dicts to work with newer code. + + Args: + state_dict (dict): state dictionary to upgrade, in place + name (str): the state dict key corresponding to the current module + """ + assert state_dict is not None + + def do_upgrade(m, prefix): + if len(prefix) > 0: + prefix += "." + + for n, c in m.named_children(): + name = prefix + n + if hasattr(c, "upgrade_state_dict_named"): + c.upgrade_state_dict_named(state_dict, name) + elif hasattr(c, "upgrade_state_dict"): + c.upgrade_state_dict(state_dict) + do_upgrade(c, name) + + do_upgrade(self, name) + + def set_num_updates(self, num_updates): + """State from trainer to pass along to model at every update.""" + for m in self.modules(): + if hasattr(m, "set_num_updates") and m != self: + m.set_num_updates(num_updates) + + def prepare_for_inference_(self, cfg: DictConfig): + """Prepare model for inference.""" + kwargs = {} + kwargs["beamable_mm_beam_size"] = ( + None + if getattr(cfg.generation, "no_beamable_mm", False) + else getattr(cfg.generation, "beam", 5) + ) + kwargs["need_attn"] = getattr(cfg.generation, "print_alignment", False) + if getattr(cfg.generation, "retain_dropout", False): + kwargs["retain_dropout"] = cfg.generation.retain_dropout + kwargs["retain_dropout_modules"] = cfg.generation.retain_dropout_modules + self.make_generation_fast_(**kwargs) + + def make_generation_fast_(self, **kwargs): + """ + Legacy entry point to optimize model for faster generation. + Prefer prepare_for_inference_. + """ + if self._is_generation_fast: + return # only apply once + self._is_generation_fast = True + + # remove weight norm from all modules in the network + def apply_remove_weight_norm(module): + try: + nn.utils.remove_weight_norm(module) + except (AttributeError, ValueError): # this module didn't have weight norm + return + + self.apply(apply_remove_weight_norm) + + def apply_make_generation_fast_(module, prefix): + if len(prefix) > 0: + prefix += "." + + base_func = BaseFairseqModel.make_generation_fast_ + for n, m in module.named_modules(): + if ( + m != self + and hasattr(m, "make_generation_fast_") + # don't call this implementation again, e.g., if + # children modules also inherit from BaseFairseqModel + and m.make_generation_fast_.__func__ is not base_func + ): + name = prefix + n + m.make_generation_fast_(name=name, **kwargs) + + apply_make_generation_fast_(self, "") + + def train(mode=True): + if mode: + raise RuntimeError("cannot train after make_generation_fast") + + # this model should no longer be used for training + self.eval() + self.train = train + + def prepare_for_onnx_export_(self, **kwargs): + """Make model exportable via ONNX trace.""" + seen = set() + + def apply_prepare_for_onnx_export_(module): + if ( + module != self + and hasattr(module, "prepare_for_onnx_export_") + and module not in seen + ): + seen.add(module) + module.prepare_for_onnx_export_(**kwargs) + + self.apply(apply_prepare_for_onnx_export_) + + @classmethod + def from_pretrained( + cls, + model_name_or_path, + checkpoint_file="model.pt", + data_name_or_path=".", + **kwargs, + ): + """ + Load a :class:`~fairseq.models.FairseqModel` from a pre-trained model + file. Downloads and caches the pre-trained model file if needed. + + The base implementation returns a + :class:`~fairseq.hub_utils.GeneratorHubInterface`, which can be used to + generate translations or sample from language models. The underlying + :class:`~fairseq.models.FairseqModel` can be accessed via the + *generator.models* attribute. + + Other models may override this to implement custom hub interfaces. + + Args: + model_name_or_path (str): either the name of a pre-trained model to + load or a path/URL to a pre-trained model state dict + checkpoint_file (str, optional): colon-separated list of checkpoint + files in the model archive to ensemble (default: 'model.pt') + data_name_or_path (str, optional): point args.data to the archive + at the given path/URL. Can start with '.' or './' to reuse the + model archive path. + """ + from fairseq import hub_utils + + x = hub_utils.from_pretrained( + model_name_or_path, + checkpoint_file, + data_name_or_path, + archive_map=cls.hub_models(), + **kwargs, + ) + logger.info(x["args"]) + return hub_utils.GeneratorHubInterface(x["args"], x["task"], x["models"]) + + @classmethod + def hub_models(cls): + return {} + + +class FairseqEncoderDecoderModel(BaseFairseqModel): + """Base class for encoder-decoder models. + + Args: + encoder (FairseqEncoder): the encoder + decoder (FairseqDecoder): the decoder + """ + + def __init__(self, encoder, decoder): + super().__init__() + + self.encoder = encoder + self.decoder = decoder + + check_type(self.encoder, FairseqEncoder) + check_type(self.decoder, FairseqDecoder) + + def forward(self, src_tokens, src_lengths, prev_output_tokens, **kwargs): + """ + Run the forward pass for an encoder-decoder model. + + First feed a batch of source tokens through the encoder. Then, feed the + encoder output and previous decoder outputs (i.e., teacher forcing) to + the decoder to produce the next outputs:: + + encoder_out = self.encoder(src_tokens, src_lengths) + return self.decoder(prev_output_tokens, encoder_out) + + Args: + src_tokens (LongTensor): tokens in the source language of shape + `(batch, src_len)` + src_lengths (LongTensor): source sentence lengths of shape `(batch)` + prev_output_tokens (LongTensor): previous decoder outputs of shape + `(batch, tgt_len)`, for teacher forcing + + Returns: + tuple: + - the decoder's output of shape `(batch, tgt_len, vocab)` + - a dictionary with any model-specific outputs + """ + encoder_out = self.encoder(src_tokens, src_lengths=src_lengths, **kwargs) + decoder_out = self.decoder( + prev_output_tokens, encoder_out=encoder_out, **kwargs + ) + return decoder_out + + def forward_decoder(self, prev_output_tokens, **kwargs): + return self.decoder(prev_output_tokens, **kwargs) + + def extract_features(self, src_tokens, src_lengths, prev_output_tokens, **kwargs): + """ + Similar to *forward* but only return features. + + Returns: + tuple: + - the decoder's features of shape `(batch, tgt_len, embed_dim)` + - a dictionary with any model-specific outputs + """ + encoder_out = self.encoder(src_tokens, src_lengths=src_lengths, **kwargs) + features = self.decoder.extract_features( + prev_output_tokens, encoder_out=encoder_out, **kwargs + ) + return features + + def output_layer(self, features, **kwargs): + """Project features to the default output size (typically vocabulary size).""" + return self.decoder.output_layer(features, **kwargs) + + def max_positions(self): + """Maximum length supported by the model.""" + return (self.encoder.max_positions(), self.decoder.max_positions()) + + def max_decoder_positions(self): + """Maximum length supported by the decoder.""" + return self.decoder.max_positions() + + +class FairseqModel(FairseqEncoderDecoderModel): + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + utils.deprecation_warning( + "FairseqModel is deprecated, please use FairseqEncoderDecoderModel " + "or BaseFairseqModel instead", + stacklevel=4, + ) + + +class FairseqMultiModel(BaseFairseqModel): + """Base class for combining multiple encoder-decoder models.""" + + def __init__(self, encoders, decoders): + super().__init__() + assert encoders.keys() == decoders.keys() + self.keys = list(encoders.keys()) + for key in self.keys: + check_type(encoders[key], FairseqEncoder) + check_type(decoders[key], FairseqDecoder) + + self.models = nn.ModuleDict( + { + key: FairseqEncoderDecoderModel(encoders[key], decoders[key]) + for key in self.keys + } + ) + + @staticmethod + def build_shared_embeddings( + dicts: Dict[str, Dictionary], + langs: List[str], + embed_dim: int, + build_embedding: callable, + pretrained_embed_path: Optional[str] = None, + ): + """ + Helper function to build shared embeddings for a set of languages after + checking that all dicts corresponding to those languages are equivalent. + + Args: + dicts: Dict of lang_id to its corresponding Dictionary + langs: languages that we want to share embeddings for + embed_dim: embedding dimension + build_embedding: callable function to actually build the embedding + pretrained_embed_path: Optional path to load pretrained embeddings + """ + shared_dict = dicts[langs[0]] + if any(dicts[lang] != shared_dict for lang in langs): + raise ValueError( + "--share-*-embeddings requires a joined dictionary: " + "--share-encoder-embeddings requires a joined source " + "dictionary, --share-decoder-embeddings requires a joined " + "target dictionary, and --share-all-embeddings requires a " + "joint source + target dictionary." + ) + return build_embedding(shared_dict, embed_dim, pretrained_embed_path) + + def forward(self, src_tokens, src_lengths, prev_output_tokens, **kwargs): + raise NotImplementedError + + def max_positions(self): + """Maximum length supported by the model.""" + return { + key: ( + self.models[key].encoder.max_positions(), + self.models[key].decoder.max_positions(), + ) + for key in self.keys + } + + def max_decoder_positions(self): + """Maximum length supported by the decoder.""" + return min(model.decoder.max_positions() for model in self.models.values()) + + @property + def encoder(self): + return self.models[self.keys[0]].encoder + + @property + def decoder(self): + return self.models[self.keys[0]].decoder + + def forward_decoder(self, prev_output_tokens, **kwargs): + return self.decoder(prev_output_tokens, **kwargs) + + def load_state_dict( + self, + state_dict, + strict=True, + model_cfg=None, + args: Optional[Namespace] = None, + ): + """Copies parameters and buffers from *state_dict* into this module and + its descendants. + + Overrides the method in :class:`nn.Module`. Compared with that method + this additionally "upgrades" *state_dicts* from old checkpoints. + """ + + if model_cfg is None and args is not None: + logger.warn("using 'args' is deprecated, please update your code to use dataclass config") + model_cfg = convert_namespace_to_omegaconf(args).model + + self.upgrade_state_dict(state_dict) + + from fairseq.checkpoint_utils import prune_state_dict + + new_state_dict = prune_state_dict(state_dict, model_cfg) + return super().load_state_dict(new_state_dict, strict) + + +class FairseqLanguageModel(BaseFairseqModel): + """Base class for decoder-only models. + + Args: + decoder (FairseqDecoder): the decoder + """ + + def __init__(self, decoder): + super().__init__() + self.decoder = decoder + check_type(self.decoder, FairseqDecoder) + + def forward(self, src_tokens, **kwargs): + """ + Run the forward pass for a decoder-only model. + + Feeds a batch of tokens through the decoder to predict the next tokens. + + Args: + src_tokens (LongTensor): tokens on which to condition the decoder, + of shape `(batch, tgt_len)` + src_lengths (LongTensor): source sentence lengths of shape `(batch)` + + Returns: + tuple: + - the decoder's output of shape `(batch, seq_len, vocab)` + - a dictionary with any model-specific outputs + """ + return self.decoder(src_tokens, **kwargs) + + def forward_decoder(self, prev_output_tokens, **kwargs): + return self.decoder(prev_output_tokens, **kwargs) + + def extract_features(self, src_tokens, **kwargs): + """ + Similar to *forward* but only return features. + + Returns: + tuple: + - the decoder's features of shape `(batch, seq_len, embed_dim)` + - a dictionary with any model-specific outputs + """ + return self.decoder.extract_features(src_tokens, **kwargs) + + def output_layer(self, features, **kwargs): + """Project features to the default output size (typically vocabulary size).""" + return self.decoder.output_layer(features, **kwargs) + + def max_positions(self): + """Maximum length supported by the model.""" + return self.decoder.max_positions() + + def max_decoder_positions(self): + """Maximum length supported by the decoder.""" + return self.decoder.max_positions() + + @property + def supported_targets(self): + return {"future"} + + +class FairseqEncoderModel(BaseFairseqModel): + """Base class for encoder-only models. + + Args: + encoder (FairseqEncoder): the encoder + """ + + def __init__(self, encoder): + super().__init__() + self.encoder = encoder + check_type(self.encoder, FairseqEncoder) + + def forward(self, src_tokens, src_lengths, **kwargs): + """ + Run the forward pass for a encoder-only model. + + Feeds a batch of tokens through the encoder to generate features. + + Args: + src_tokens (LongTensor): input tokens of shape `(batch, src_len)` + src_lengths (LongTensor): source sentence lengths of shape `(batch)` + + Returns: + the encoder's output, typically of shape `(batch, src_len, features)` + """ + return self.encoder(src_tokens, src_lengths, **kwargs) + + def get_normalized_probs(self, net_output, log_probs, sample=None): + """Get normalized probabilities (or log probs) from a net's output.""" + encoder_out = net_output["encoder_out"] + if torch.is_tensor(encoder_out): + logits = encoder_out.float() + if log_probs: + return F.log_softmax(logits, dim=-1) + else: + return F.softmax(logits, dim=-1) + raise NotImplementedError + + def max_positions(self): + """Maximum length supported by the model.""" + return self.encoder.max_positions() diff --git a/fairseq/models/fconv.py b/fairseq/models/fconv.py new file mode 100644 index 0000000000000000000000000000000000000000..c99a2151014d816ec9aff6f4b27d71224dd7b4cf --- /dev/null +++ b/fairseq/models/fconv.py @@ -0,0 +1,756 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +import torch +import torch.nn as nn +import torch.nn.functional as F +from fairseq import utils +from fairseq.models import ( + FairseqEncoder, + FairseqEncoderDecoderModel, + FairseqIncrementalDecoder, + register_model, + register_model_architecture, +) +from fairseq.modules import ( + AdaptiveSoftmax, + BeamableMM, + FairseqDropout, + GradMultiply, + LearnedPositionalEmbedding, + LinearizedConvolution, +) + + +@register_model("fconv") +class FConvModel(FairseqEncoderDecoderModel): + """ + A fully convolutional model, i.e. a convolutional encoder and a + convolutional decoder, as described in `"Convolutional Sequence to Sequence + Learning" (Gehring et al., 2017) <https://arxiv.org/abs/1705.03122>`_. + + Args: + encoder (FConvEncoder): the encoder + decoder (FConvDecoder): the decoder + + The Convolutional model provides the following named architectures and + command-line arguments: + + .. argparse:: + :ref: fairseq.models.fconv_parser + :prog: + """ + + @classmethod + def hub_models(cls): + def moses_subword(path): + return { + "path": path, + "tokenizer": "moses", + "bpe": "subword_nmt", + } + + return { + "conv.wmt14.en-fr": moses_subword( + "https://dl.fbaipublicfiles.com/fairseq/models/wmt14.v2.en-fr.fconv-py.tar.bz2" + ), + "conv.wmt14.en-de": moses_subword( + "https://dl.fbaipublicfiles.com/fairseq/models/wmt14.en-de.fconv-py.tar.bz2" + ), + "conv.wmt17.en-de": moses_subword( + "https://dl.fbaipublicfiles.com/fairseq/models/wmt17.v2.en-de.fconv-py.tar.bz2" + ), + } + + def __init__(self, encoder, decoder): + super().__init__(encoder, decoder) + self.encoder.num_attention_layers = sum( + layer is not None for layer in decoder.attention + ) + + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + # fmt: off + parser.add_argument('--dropout', type=float, metavar='D', + help='dropout probability') + parser.add_argument('--encoder-embed-dim', type=int, metavar='N', + help='encoder embedding dimension') + parser.add_argument('--encoder-embed-path', type=str, metavar='STR', + help='path to pre-trained encoder embedding') + parser.add_argument('--encoder-layers', type=str, metavar='EXPR', + help='encoder layers [(dim, kernel_size), ...]') + parser.add_argument('--decoder-embed-dim', type=int, metavar='N', + help='decoder embedding dimension') + parser.add_argument('--decoder-embed-path', type=str, metavar='STR', + help='path to pre-trained decoder embedding') + parser.add_argument('--decoder-layers', type=str, metavar='EXPR', + help='decoder layers [(dim, kernel_size), ...]') + parser.add_argument('--decoder-out-embed-dim', type=int, metavar='N', + help='decoder output embedding dimension') + parser.add_argument('--decoder-attention', type=str, metavar='EXPR', + help='decoder attention [True, ...]') + parser.add_argument('--share-input-output-embed', action='store_true', + help='share input and output embeddings (requires' + ' --decoder-out-embed-dim and --decoder-embed-dim' + ' to be equal)') + # fmt: on + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + # make sure that all args are properly defaulted (in case there are any new ones) + base_architecture(args) + + encoder_embed_dict = None + if args.encoder_embed_path: + encoder_embed_dict = utils.parse_embedding(args.encoder_embed_path) + utils.print_embed_overlap(encoder_embed_dict, task.source_dictionary) + + decoder_embed_dict = None + if args.decoder_embed_path: + decoder_embed_dict = utils.parse_embedding(args.decoder_embed_path) + utils.print_embed_overlap(decoder_embed_dict, task.target_dictionary) + + encoder = FConvEncoder( + dictionary=task.source_dictionary, + embed_dim=args.encoder_embed_dim, + embed_dict=encoder_embed_dict, + convolutions=eval(args.encoder_layers), + dropout=args.dropout, + max_positions=args.max_source_positions, + ) + decoder = FConvDecoder( + dictionary=task.target_dictionary, + embed_dim=args.decoder_embed_dim, + embed_dict=decoder_embed_dict, + convolutions=eval(args.decoder_layers), + out_embed_dim=args.decoder_out_embed_dim, + attention=eval(args.decoder_attention), + dropout=args.dropout, + max_positions=args.max_target_positions, + share_embed=args.share_input_output_embed, + ) + return FConvModel(encoder, decoder) + + +class FConvEncoder(FairseqEncoder): + """ + Convolutional encoder consisting of `len(convolutions)` layers. + + Args: + dictionary (~fairseq.data.Dictionary): encoding dictionary + embed_dim (int, optional): embedding dimension + embed_dict (str, optional): filename from which to load pre-trained + embeddings + max_positions (int, optional): maximum supported input sequence length + convolutions (list, optional): the convolutional layer structure. Each + list item `i` corresponds to convolutional layer `i`. Layers are + given as ``(out_channels, kernel_width, [residual])``. Residual + connections are added between layers when ``residual=1`` (which is + the default behavior). + dropout (float, optional): dropout to be applied before each conv layer + """ + + def __init__( + self, + dictionary, + embed_dim=512, + embed_dict=None, + max_positions=1024, + convolutions=((512, 3),) * 20, + dropout=0.1, + ): + super().__init__(dictionary) + self.dropout_module = FairseqDropout( + dropout, module_name=self.__class__.__name__ + ) + self.num_attention_layers = None + + num_embeddings = len(dictionary) + self.padding_idx = dictionary.pad() + self.embed_tokens = Embedding(num_embeddings, embed_dim, self.padding_idx) + if embed_dict: + self.embed_tokens = utils.load_embedding( + embed_dict, self.dictionary, self.embed_tokens + ) + + self.embed_positions = PositionalEmbedding( + max_positions, + embed_dim, + self.padding_idx, + ) + + convolutions = extend_conv_spec(convolutions) + in_channels = convolutions[0][0] + self.fc1 = Linear(embed_dim, in_channels, dropout=dropout) + self.projections = nn.ModuleList() + self.convolutions = nn.ModuleList() + self.residuals = [] + + layer_in_channels = [in_channels] + for _, (out_channels, kernel_size, residual) in enumerate(convolutions): + if residual == 0: + residual_dim = out_channels + else: + residual_dim = layer_in_channels[-residual] + self.projections.append( + Linear(residual_dim, out_channels) + if residual_dim != out_channels + else None + ) + if kernel_size % 2 == 1: + padding = kernel_size // 2 + else: + padding = 0 + self.convolutions.append( + ConvTBC( + in_channels, + out_channels * 2, + kernel_size, + dropout=dropout, + padding=padding, + ) + ) + self.residuals.append(residual) + in_channels = out_channels + layer_in_channels.append(out_channels) + self.fc2 = Linear(in_channels, embed_dim) + + def forward(self, src_tokens, src_lengths): + """ + Args: + src_tokens (LongTensor): tokens in the source language of shape + `(batch, src_len)` + src_lengths (LongTensor): lengths of each source sentence of shape + `(batch)` + + Returns: + dict: + - **encoder_out** (tuple): a tuple with two elements, where the + first element is the last encoder layer's output and the + second element is the same quantity summed with the input + embedding (used for attention). The shape of both tensors is + `(batch, src_len, embed_dim)`. + - **encoder_padding_mask** (ByteTensor): the positions of + padding elements of shape `(batch, src_len)` + """ + # embed tokens and positions + x = self.embed_tokens(src_tokens) + self.embed_positions(src_tokens) + x = self.dropout_module(x) + input_embedding = x + + # project to size of convolution + x = self.fc1(x) + + # used to mask padding in input + encoder_padding_mask = src_tokens.eq(self.padding_idx).t() # -> T x B + if not encoder_padding_mask.any(): + encoder_padding_mask = None + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + + residuals = [x] + # temporal convolutions + for proj, conv, res_layer in zip( + self.projections, self.convolutions, self.residuals + ): + if res_layer > 0: + residual = residuals[-res_layer] + residual = residual if proj is None else proj(residual) + else: + residual = None + + if encoder_padding_mask is not None: + x = x.masked_fill(encoder_padding_mask.unsqueeze(-1), 0) + + x = self.dropout_module(x) + if conv.kernel_size[0] % 2 == 1: + # padding is implicit in the conv + x = conv(x) + else: + padding_l = (conv.kernel_size[0] - 1) // 2 + padding_r = conv.kernel_size[0] // 2 + x = F.pad(x, (0, 0, 0, 0, padding_l, padding_r)) + x = conv(x) + x = F.glu(x, dim=2) + + if residual is not None: + x = (x + residual) * math.sqrt(0.5) + residuals.append(x) + + # T x B x C -> B x T x C + x = x.transpose(1, 0) + + # project back to size of embedding + x = self.fc2(x) + + if encoder_padding_mask is not None: + encoder_padding_mask = encoder_padding_mask.t() # -> B x T + x = x.masked_fill(encoder_padding_mask.unsqueeze(-1), 0) + + # scale gradients (this only affects backward, not forward) + x = GradMultiply.apply(x, 1.0 / (2.0 * self.num_attention_layers)) + + # add output to input embedding for attention + y = (x + input_embedding) * math.sqrt(0.5) + + return { + "encoder_out": (x, y), + "encoder_padding_mask": encoder_padding_mask, # B x T + } + + def reorder_encoder_out(self, encoder_out, new_order): + if encoder_out["encoder_out"] is not None: + encoder_out["encoder_out"] = ( + encoder_out["encoder_out"][0].index_select(0, new_order), + encoder_out["encoder_out"][1].index_select(0, new_order), + ) + if encoder_out["encoder_padding_mask"] is not None: + encoder_out["encoder_padding_mask"] = encoder_out[ + "encoder_padding_mask" + ].index_select(0, new_order) + return encoder_out + + def max_positions(self): + """Maximum input length supported by the encoder.""" + return self.embed_positions.max_positions + + +class AttentionLayer(nn.Module): + def __init__(self, conv_channels, embed_dim, bmm=None): + super().__init__() + # projects from output of convolution to embedding dimension + self.in_projection = Linear(conv_channels, embed_dim) + # projects from embedding dimension to convolution size + self.out_projection = Linear(embed_dim, conv_channels) + + self.bmm = bmm if bmm is not None else torch.bmm + + def forward(self, x, target_embedding, encoder_out, encoder_padding_mask): + residual = x + + # attention + x = (self.in_projection(x) + target_embedding) * math.sqrt(0.5) + x = self.bmm(x, encoder_out[0]) + + # don't attend over padding + if encoder_padding_mask is not None: + x = ( + x.float() + .masked_fill(encoder_padding_mask.unsqueeze(1), float("-inf")) + .type_as(x) + ) # FP16 support: cast to float and back + + # softmax over last dim + sz = x.size() + x = F.softmax(x.view(sz[0] * sz[1], sz[2]), dim=1) + x = x.view(sz) + attn_scores = x + + x = self.bmm(x, encoder_out[1]) + + # scale attention output (respecting potentially different lengths) + s = encoder_out[1].size(1) + if encoder_padding_mask is None: + x = x * (s * math.sqrt(1.0 / s)) + else: + s = s - encoder_padding_mask.type_as(x).sum( + dim=1, keepdim=True + ) # exclude padding + s = s.unsqueeze(-1) + x = x * (s * s.rsqrt()) + + # project back + x = (self.out_projection(x) + residual) * math.sqrt(0.5) + return x, attn_scores + + def make_generation_fast_(self, beamable_mm_beam_size=None, **kwargs): + """Replace torch.bmm with BeamableMM.""" + if beamable_mm_beam_size is not None: + del self.bmm + self.add_module("bmm", BeamableMM(beamable_mm_beam_size)) + + +class FConvDecoder(FairseqIncrementalDecoder): + """Convolutional decoder""" + + def __init__( + self, + dictionary, + embed_dim=512, + embed_dict=None, + out_embed_dim=256, + max_positions=1024, + convolutions=((512, 3),) * 20, + attention=True, + dropout=0.1, + share_embed=False, + positional_embeddings=True, + adaptive_softmax_cutoff=None, + adaptive_softmax_dropout=0.0, + ): + super().__init__(dictionary) + self.register_buffer("version", torch.Tensor([2])) + self.dropout_module = FairseqDropout( + dropout, module_name=self.__class__.__name__ + ) + self.need_attn = True + + convolutions = extend_conv_spec(convolutions) + in_channels = convolutions[0][0] + if isinstance(attention, bool): + # expand True into [True, True, ...] and do the same with False + attention = [attention] * len(convolutions) + if not isinstance(attention, list) or len(attention) != len(convolutions): + raise ValueError( + "Attention is expected to be a list of booleans of " + "length equal to the number of layers." + ) + + num_embeddings = len(dictionary) + padding_idx = dictionary.pad() + self.embed_tokens = Embedding(num_embeddings, embed_dim, padding_idx) + if embed_dict: + self.embed_tokens = utils.load_embedding( + embed_dict, self.dictionary, self.embed_tokens + ) + + self.embed_positions = ( + PositionalEmbedding( + max_positions, + embed_dim, + padding_idx, + ) + if positional_embeddings + else None + ) + + self.fc1 = Linear(embed_dim, in_channels, dropout=dropout) + self.projections = nn.ModuleList() + self.convolutions = nn.ModuleList() + self.attention = nn.ModuleList() + self.residuals = [] + + layer_in_channels = [in_channels] + for i, (out_channels, kernel_size, residual) in enumerate(convolutions): + if residual == 0: + residual_dim = out_channels + else: + residual_dim = layer_in_channels[-residual] + self.projections.append( + Linear(residual_dim, out_channels) + if residual_dim != out_channels + else None + ) + self.convolutions.append( + LinearizedConv1d( + in_channels, + out_channels * 2, + kernel_size, + padding=(kernel_size - 1), + dropout=dropout, + ) + ) + self.attention.append( + AttentionLayer(out_channels, embed_dim) if attention[i] else None + ) + self.residuals.append(residual) + in_channels = out_channels + layer_in_channels.append(out_channels) + + self.adaptive_softmax = None + self.fc2 = self.fc3 = None + + if adaptive_softmax_cutoff is not None: + assert not share_embed + self.adaptive_softmax = AdaptiveSoftmax( + num_embeddings, + in_channels, + adaptive_softmax_cutoff, + dropout=adaptive_softmax_dropout, + ) + else: + self.fc2 = Linear(in_channels, out_embed_dim) + if share_embed: + assert out_embed_dim == embed_dim, ( + "Shared embed weights implies same dimensions " + " out_embed_dim={} vs embed_dim={}".format(out_embed_dim, embed_dim) + ) + self.fc3 = nn.Linear(out_embed_dim, num_embeddings) + self.fc3.weight = self.embed_tokens.weight + else: + self.fc3 = Linear(out_embed_dim, num_embeddings, dropout=dropout) + + def forward( + self, prev_output_tokens, encoder_out=None, incremental_state=None, **unused + ): + if encoder_out is not None: + encoder_padding_mask = encoder_out["encoder_padding_mask"] + encoder_out = encoder_out["encoder_out"] + + # split and transpose encoder outputs + encoder_a, encoder_b = self._split_encoder_out( + encoder_out, incremental_state + ) + + if self.embed_positions is not None: + pos_embed = self.embed_positions(prev_output_tokens, incremental_state) + else: + pos_embed = 0 + + if incremental_state is not None: + prev_output_tokens = prev_output_tokens[:, -1:] + x = self._embed_tokens(prev_output_tokens, incremental_state) + + # embed tokens and combine with positional embeddings + x += pos_embed + x = self.dropout_module(x) + target_embedding = x + + # project to size of convolution + x = self.fc1(x) + + # B x T x C -> T x B x C + x = self._transpose_if_training(x, incremental_state) + + # temporal convolutions + avg_attn_scores = None + num_attn_layers = len(self.attention) + residuals = [x] + for proj, conv, attention, res_layer in zip( + self.projections, self.convolutions, self.attention, self.residuals + ): + if res_layer > 0: + residual = residuals[-res_layer] + residual = residual if proj is None else proj(residual) + else: + residual = None + + x = self.dropout_module(x) + x = conv(x, incremental_state) + x = F.glu(x, dim=2) + + # attention + if attention is not None: + x = self._transpose_if_training(x, incremental_state) + + x, attn_scores = attention( + x, target_embedding, (encoder_a, encoder_b), encoder_padding_mask + ) + + if not self.training and self.need_attn: + attn_scores = attn_scores / num_attn_layers + if avg_attn_scores is None: + avg_attn_scores = attn_scores + else: + avg_attn_scores.add_(attn_scores) + + x = self._transpose_if_training(x, incremental_state) + + # residual + if residual is not None: + x = (x + residual) * math.sqrt(0.5) + residuals.append(x) + + # T x B x C -> B x T x C + x = self._transpose_if_training(x, incremental_state) + + # project back to size of vocabulary if not using adaptive softmax + if self.fc2 is not None and self.fc3 is not None: + x = self.fc2(x) + x = self.dropout_module(x) + x = self.fc3(x) + + return x, avg_attn_scores + + def reorder_incremental_state(self, incremental_state, new_order): + super().reorder_incremental_state(incremental_state, new_order) + encoder_out = utils.get_incremental_state( + self, incremental_state, "encoder_out" + ) + if encoder_out is not None: + encoder_out = tuple(eo.index_select(0, new_order) for eo in encoder_out) + utils.set_incremental_state( + self, incremental_state, "encoder_out", encoder_out + ) + + def max_positions(self): + """Maximum output length supported by the decoder.""" + return ( + self.embed_positions.max_positions + if self.embed_positions is not None + else float("inf") + ) + + def upgrade_state_dict(self, state_dict): + if utils.item(state_dict.get("decoder.version", torch.Tensor([1]))[0]) < 2: + # old models use incorrect weight norm dimension + for i, conv in enumerate(self.convolutions): + # reconfigure weight norm + nn.utils.remove_weight_norm(conv) + self.convolutions[i] = nn.utils.weight_norm(conv, dim=0) + state_dict["decoder.version"] = torch.Tensor([1]) + return state_dict + + def make_generation_fast_(self, need_attn=False, **kwargs): + self.need_attn = need_attn + + def _embed_tokens(self, tokens, incremental_state): + if incremental_state is not None: + # keep only the last token for incremental forward pass + tokens = tokens[:, -1:] + return self.embed_tokens(tokens) + + def _split_encoder_out(self, encoder_out, incremental_state): + """Split and transpose encoder outputs. + + This is cached when doing incremental inference. + """ + cached_result = utils.get_incremental_state( + self, incremental_state, "encoder_out" + ) + if cached_result is not None: + return cached_result + + # transpose only once to speed up attention layers + encoder_a, encoder_b = encoder_out + encoder_a = encoder_a.transpose(1, 2).contiguous() + result = (encoder_a, encoder_b) + + if incremental_state is not None: + utils.set_incremental_state(self, incremental_state, "encoder_out", result) + return result + + def _transpose_if_training(self, x, incremental_state): + if incremental_state is None: + x = x.transpose(0, 1) + return x + + +def extend_conv_spec(convolutions): + """ + Extends convolutional spec that is a list of tuples of 2 or 3 parameters + (kernel size, dim size and optionally how many layers behind to look for residual) + to default the residual propagation param if it is not specified + """ + extended = [] + for spec in convolutions: + if len(spec) == 3: + extended.append(spec) + elif len(spec) == 2: + extended.append(spec + (1,)) + else: + raise Exception( + "invalid number of parameters in convolution spec " + + str(spec) + + ". expected 2 or 3" + ) + return tuple(extended) + + +def Embedding(num_embeddings, embedding_dim, padding_idx): + m = nn.Embedding(num_embeddings, embedding_dim, padding_idx=padding_idx) + nn.init.normal_(m.weight, 0, 0.1) + nn.init.constant_(m.weight[padding_idx], 0) + return m + + +def PositionalEmbedding(num_embeddings, embedding_dim, padding_idx): + m = LearnedPositionalEmbedding(num_embeddings, embedding_dim, padding_idx) + nn.init.normal_(m.weight, 0, 0.1) + nn.init.constant_(m.weight[padding_idx], 0) + return m + + +def Linear(in_features, out_features, dropout=0.0): + """Weight-normalized Linear layer (input: N x T x C)""" + m = nn.Linear(in_features, out_features) + nn.init.normal_(m.weight, mean=0, std=math.sqrt((1 - dropout) / in_features)) + nn.init.constant_(m.bias, 0) + return nn.utils.weight_norm(m) + + +def LinearizedConv1d(in_channels, out_channels, kernel_size, dropout=0.0, **kwargs): + """Weight-normalized Conv1d layer optimized for decoding""" + m = LinearizedConvolution(in_channels, out_channels, kernel_size, **kwargs) + std = math.sqrt((4 * (1.0 - dropout)) / (m.kernel_size[0] * in_channels)) + nn.init.normal_(m.weight, mean=0, std=std) + nn.init.constant_(m.bias, 0) + return nn.utils.weight_norm(m, dim=2) + + +def ConvTBC(in_channels, out_channels, kernel_size, dropout=0.0, **kwargs): + """Weight-normalized Conv1d layer""" + from fairseq.modules import ConvTBC + + m = ConvTBC(in_channels, out_channels, kernel_size, **kwargs) + std = math.sqrt((4 * (1.0 - dropout)) / (m.kernel_size[0] * in_channels)) + nn.init.normal_(m.weight, mean=0, std=std) + nn.init.constant_(m.bias, 0) + return nn.utils.weight_norm(m, dim=2) + + +@register_model_architecture("fconv", "fconv") +def base_architecture(args): + args.dropout = getattr(args, "dropout", 0.1) + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512) + args.encoder_embed_path = getattr(args, "encoder_embed_path", None) + args.encoder_layers = getattr(args, "encoder_layers", "[(512, 3)] * 20") + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 512) + args.decoder_embed_path = getattr(args, "decoder_embed_path", None) + args.decoder_layers = getattr(args, "decoder_layers", "[(512, 3)] * 20") + args.decoder_out_embed_dim = getattr(args, "decoder_out_embed_dim", 256) + args.decoder_attention = getattr(args, "decoder_attention", "True") + args.share_input_output_embed = getattr(args, "share_input_output_embed", False) + + +@register_model_architecture("fconv", "fconv_iwslt_de_en") +def fconv_iwslt_de_en(args): + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 256) + args.encoder_layers = getattr(args, "encoder_layers", "[(256, 3)] * 4") + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 256) + args.decoder_layers = getattr(args, "decoder_layers", "[(256, 3)] * 3") + args.decoder_out_embed_dim = getattr(args, "decoder_out_embed_dim", 256) + base_architecture(args) + + +@register_model_architecture("fconv", "fconv_wmt_en_ro") +def fconv_wmt_en_ro(args): + args.decoder_out_embed_dim = getattr(args, "decoder_out_embed_dim", 512) + base_architecture(args) + + +@register_model_architecture("fconv", "fconv_wmt_en_de") +def fconv_wmt_en_de(args): + convs = "[(512, 3)] * 9" # first 9 layers have 512 units + convs += " + [(1024, 3)] * 4" # next 4 layers have 1024 units + convs += " + [(2048, 1)] * 2" # final 2 layers use 1x1 convolutions + + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 768) + args.encoder_layers = getattr(args, "encoder_layers", convs) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 768) + args.decoder_layers = getattr(args, "decoder_layers", convs) + args.decoder_out_embed_dim = getattr(args, "decoder_out_embed_dim", 512) + base_architecture(args) + + +@register_model_architecture("fconv", "fconv_wmt_en_fr") +def fconv_wmt_en_fr(args): + convs = "[(512, 3)] * 6" # first 6 layers have 512 units + convs += " + [(768, 3)] * 4" # next 4 layers have 768 units + convs += " + [(1024, 3)] * 3" # next 3 layers have 1024 units + convs += " + [(2048, 1)] * 1" # next 1 layer uses 1x1 convolutions + convs += " + [(4096, 1)] * 1" # final 1 layer uses 1x1 convolutions + + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 768) + args.encoder_layers = getattr(args, "encoder_layers", convs) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 768) + args.decoder_layers = getattr(args, "decoder_layers", convs) + args.decoder_out_embed_dim = getattr(args, "decoder_out_embed_dim", 512) + base_architecture(args) diff --git a/fairseq/models/fconv_lm.py b/fairseq/models/fconv_lm.py new file mode 100644 index 0000000000000000000000000000000000000000..07391eaa2908eacd2709176942d920c483c4f066 --- /dev/null +++ b/fairseq/models/fconv_lm.py @@ -0,0 +1,135 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq import utils +from fairseq.models import ( + FairseqLanguageModel, + register_model, + register_model_architecture, +) +from fairseq.models.fconv import FConvDecoder + + +@register_model("fconv_lm") +class FConvLanguageModel(FairseqLanguageModel): + def __init__(self, decoder): + super().__init__(decoder) + + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + parser.add_argument( + "--dropout", type=float, metavar="D", help="dropout probability" + ) + parser.add_argument( + "--decoder-embed-dim", + type=int, + metavar="N", + help="decoder embedding dimension", + ) + parser.add_argument( + "--decoder-layers", + type=str, + metavar="EXPR", + help="decoder layers [(dim, kernel_size), ...]", + ) + parser.add_argument( + "--decoder-out-embed-dim", + type=int, + metavar="N", + help="decoder output embedding dimension", + ) + parser.add_argument( + "--adaptive-softmax-cutoff", + metavar="EXPR", + help="comma separated list of adaptive softmax cutoff points. " + "Must be used with adaptive_loss criterion", + ) + parser.add_argument( + "--adaptive-softmax-dropout", + type=float, + metavar="D", + help="sets adaptive softmax dropout for the tail projections", + ) + parser.add_argument( + "--decoder-attention", + type=str, + metavar="EXPR", + help="decoder attention [True, ...]", + ) + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + # make sure all arguments are present in older models + base_lm_architecture(args) + + if hasattr(args, "max_target_positions") and not hasattr( + args, "tokens_per_sample" + ): + args.tokens_per_sample = args.max_target_positions + + decoder = FConvDecoder( + dictionary=task.target_dictionary, + embed_dim=args.decoder_embed_dim, + convolutions=eval(args.decoder_layers), + out_embed_dim=args.decoder_embed_dim, + attention=eval(args.decoder_attention), + dropout=args.dropout, + max_positions=args.tokens_per_sample, + share_embed=False, + positional_embeddings=False, + adaptive_softmax_cutoff=( + utils.eval_str_list(args.adaptive_softmax_cutoff, type=int) + if args.criterion == "adaptive_loss" + else None + ), + adaptive_softmax_dropout=args.adaptive_softmax_dropout, + ) + return FConvLanguageModel(decoder) + + +@register_model_architecture("fconv_lm", "fconv_lm") +def base_lm_architecture(args): + args.dropout = getattr(args, "dropout", 0.1) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 128) + args.decoder_layers = getattr(args, "decoder_layers", "[(1268, 4)] * 13") + args.decoder_attention = getattr(args, "decoder_attention", "False") + args.adaptive_softmax_cutoff = getattr(args, "adaptive_softmax_cutoff", None) + args.adaptive_softmax_dropout = getattr(args, "adaptive_softmax_dropout", 0) + + +@register_model_architecture("fconv_lm", "fconv_lm_dauphin_wikitext103") +def fconv_lm_dauphin_wikitext103(args): + layers = "[(850, 6)] * 3" + layers += " + [(850, 1)] * 1" + layers += " + [(850, 5)] * 4" + layers += " + [(850, 1)] * 1" + layers += " + [(850, 4)] * 3" + layers += " + [(1024, 4)] * 1" + layers += " + [(2048, 4)] * 1" + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 280) + args.decoder_layers = getattr(args, "decoder_layers", layers) + args.decoder_attention = getattr(args, "decoder_attention", "False") + args.adaptive_softmax_cutoff = getattr( + args, "adaptive_softmax_cutoff", "10000,20000,200000" + ) + base_lm_architecture(args) + + +@register_model_architecture("fconv_lm", "fconv_lm_dauphin_gbw") +def fconv_lm_dauphin_gbw(args): + layers = "[(512, 5)]" + layers += " + [(128, 1, 0), (128, 5, 0), (512, 1, 3)] * 3" + layers += " + [(512, 1, 0), (512, 5, 0), (1024, 1, 3)] * 3" + layers += " + [(1024, 1, 0), (1024, 5, 0), (2048, 1, 3)] * 6" + layers += " + [(1024, 1, 0), (1024, 5, 0), (4096, 1, 3)]" + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 128) + args.decoder_layers = getattr(args, "decoder_layers", layers) + args.decoder_attention = getattr(args, "decoder_attention", "False") + args.adaptive_softmax_cutoff = getattr( + args, "adaptive_softmax_cutoff", "10000,50000,200000" + ) + base_lm_architecture(args) diff --git a/fairseq/models/fconv_self_att.py b/fairseq/models/fconv_self_att.py new file mode 100644 index 0000000000000000000000000000000000000000..8357ef7847ed25a62345e219c41906156828c233 --- /dev/null +++ b/fairseq/models/fconv_self_att.py @@ -0,0 +1,674 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import math +import os + +import torch +import torch.nn as nn +import torch.nn.functional as F +from fairseq import checkpoint_utils +from fairseq.incremental_decoding_utils import with_incremental_state +from fairseq.models import ( + CompositeEncoder, + FairseqDecoder, + FairseqEncoder, + FairseqEncoderDecoderModel, + register_model, + register_model_architecture, +) +from fairseq.modules import ( + DownsampledMultiHeadAttention, + FairseqDropout, + GradMultiply, + LayerNorm, + LearnedPositionalEmbedding, + LinearizedConvolution, +) + + +logger = logging.getLogger(__name__) + + +@register_model("fconv_self_att") +class FConvModelSelfAtt(FairseqEncoderDecoderModel): + @classmethod + def hub_models(cls): + return { + "conv.stories.pretrained": { + "path": "https://dl.fbaipublicfiles.com/fairseq/models/stories_checkpoint.tar.gz", + "checkpoint_file": "pretrained_checkpoint.pt", + "tokenizer": "nltk", + }, + "conv.stories": { + "path": "https://dl.fbaipublicfiles.com/fairseq/models/stories_checkpoint.tar.gz", + "checkpoint_file": "fusion_checkpoint.pt", + "tokenizer": "nltk", + "pretrained": "True", + "pretrained_checkpoint": "./pretrained_checkpoint.pt", + }, + # Test set containing dictionaries + "data.stories": "https://dl.fbaipublicfiles.com/fairseq/data/stories_test.tar.bz2", + } + + def __init__(self, encoder, decoder, pretrained_encoder=None): + super().__init__(encoder, decoder) + self.encoder.num_attention_layers = sum( + layer is not None for layer in decoder.attention + ) + self.pretrained_encoder = pretrained_encoder + if self.pretrained_encoder is None: + encoders = {"encoder": encoder} + else: + encoders = {"encoder": encoder, "pretrained": self.pretrained_encoder} + # for fusion model, CompositeEncoder contains both pretrained and training encoders + # these are forwarded and then combined in the decoder + self.encoder = CompositeEncoder(encoders) + + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + # fmt: off + parser.add_argument('--dropout', type=float, metavar='D', + help='dropout probability') + parser.add_argument('--encoder-embed-dim', type=int, metavar='N', + help='encoder embedding dimension') + parser.add_argument('--encoder-layers', type=str, metavar='EXPR', + help='encoder layers [(dim, kernel_size), ...]') + parser.add_argument('--decoder-embed-dim', type=int, metavar='N', + help='decoder embedding dimension') + parser.add_argument('--decoder-layers', type=str, metavar='EXPR', + help='decoder layers [(dim, kernel_size), ...]') + parser.add_argument('--decoder-out-embed-dim', type=int, metavar='N', + help='decoder output embedding dimension') + parser.add_argument('--decoder-attention', type=str, metavar='EXPR', + help='decoder attention [True, ...]') + parser.add_argument('--self-attention', type=str, metavar='EXPR', + help='decoder self-attention layers, ex: [True] + [False]*5') + parser.add_argument('--multihead-attention-nheads', type=int, + help='Number of heads to use in attention') + parser.add_argument('--multihead-self-attention-nheads', type=int, + help='Number of heads to use in self-attention') + parser.add_argument('--encoder-attention', type=str, metavar='EXPR', + help='encoder attention [True, ...]') + parser.add_argument('--encoder-attention-nheads', type=int, + help='Number of heads to use in encoder attention') + parser.add_argument('--project-input', type=str, metavar='EXPR', + help='Use projections in self-attention [True, ...]') + parser.add_argument('--gated-attention', type=str, metavar='EXPR', + help='Use GLU layers in self-attention projections [True, ...]') + parser.add_argument('--downsample', type=str, metavar='EXPR', + help='Use downsampling in self-attention [True, ...]') + parser.add_argument('--pretrained-checkpoint', metavar='DIR', + help='path to load checkpoint from pretrained model') + parser.add_argument('--pretrained', type=str, metavar='EXPR', + help='use pretrained model when training [True, ...]') + # fmt: on + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + trained_encoder, trained_decoder = None, None + pretrained = eval(args.pretrained) + if pretrained: + logger.info("loading pretrained model") + if not os.path.exists(args.pretrained_checkpoint): + new_pretrained_checkpoint = os.path.join( + args.data, args.pretrained_checkpoint + ) + if os.path.exists(new_pretrained_checkpoint): + args.pretrained_checkpoint = new_pretrained_checkpoint + trained_model = checkpoint_utils.load_model_ensemble( + filenames=[args.pretrained_checkpoint], + task=task, + )[0][0] + trained_decoder = list(trained_model.children())[1] + trained_encoder = list(trained_model.children())[0] + + # freeze pretrained model + for param in trained_decoder.parameters(): + param.requires_grad = False + for param in trained_encoder.parameters(): + param.requires_grad = False + + encoder = FConvEncoder( + task.source_dictionary, + embed_dim=args.encoder_embed_dim, + convolutions=eval(args.encoder_layers), + dropout=args.dropout, + max_positions=args.max_source_positions, + attention=eval(args.encoder_attention), + attention_nheads=args.encoder_attention_nheads, + ) + + decoder = FConvDecoder( + task.target_dictionary, + embed_dim=args.decoder_embed_dim, + convolutions=eval(args.decoder_layers), + out_embed_dim=args.decoder_out_embed_dim, + attention=eval(args.decoder_attention), + dropout=args.dropout, + max_positions=args.max_target_positions, + selfattention=eval(args.self_attention), + attention_nheads=args.multihead_attention_nheads, + selfattention_nheads=args.multihead_self_attention_nheads, + project_input=eval(args.project_input), + gated_attention=eval(args.gated_attention), + downsample=eval(args.downsample), + pretrained=pretrained, + trained_decoder=trained_decoder, + ) + model = FConvModelSelfAtt(encoder, decoder, trained_encoder) + + return model + + @property + def pretrained(self): + return self.pretrained_encoder is not None + + +class FConvEncoder(FairseqEncoder): + """Convolutional encoder""" + + def __init__( + self, + dictionary, + embed_dim=512, + max_positions=1024, + convolutions=((512, 3),) * 20, + dropout=0.1, + attention=False, + attention_nheads=1, + ): + super().__init__(dictionary) + self.dropout_module = FairseqDropout( + dropout, module_name=self.__class__.__name__ + ) + self.num_attention_layers = None + + num_embeddings = len(dictionary) + self.padding_idx = dictionary.pad() + self.embed_tokens = Embedding(num_embeddings, embed_dim, self.padding_idx) + self.embed_positions = PositionalEmbedding( + max_positions, + embed_dim, + self.padding_idx, + ) + + def expand_bool_array(val): + if isinstance(val, bool): + # expand True into [True, True, ...] and do the same with False + return [val] * len(convolutions) + return val + + attention = expand_bool_array(attention) + + in_channels = convolutions[0][0] + self.fc1 = Linear(embed_dim, in_channels, dropout=dropout) + self.projections = nn.ModuleList() + self.convolutions = nn.ModuleList() + self.attention = nn.ModuleList() + self.attproj = nn.ModuleList() + for i, (out_channels, kernel_size) in enumerate(convolutions): + self.projections.append( + Linear(in_channels, out_channels) + if in_channels != out_channels + else None + ) + self.convolutions.append( + ConvTBC(in_channels, out_channels * 2, kernel_size, dropout=dropout) + ) + + self.attention.append( + SelfAttention(out_channels, embed_dim, attention_nheads) + if attention[i] + else None + ) + in_channels = out_channels + + self.fc2 = Linear(in_channels, embed_dim) + + def forward(self, src_tokens, src_lengths): + # embed tokens and positions + x = self.embed_tokens(src_tokens) + self.embed_positions(src_tokens) + x = self.dropout_module(x) + input_embedding = x.transpose(0, 1) + + # project to size of convolution + x = self.fc1(x) + + encoder_padding_mask = src_tokens.eq(self.padding_idx).t() # -> T x B + if not encoder_padding_mask.any(): + encoder_padding_mask = None + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + + # temporal convolutions + for proj, conv, attention in zip( + self.projections, self.convolutions, self.attention + ): + residual = x if proj is None else proj(x) + + if encoder_padding_mask is not None: + x = x.masked_fill(encoder_padding_mask.unsqueeze(-1), 0) + + x = self.dropout_module(x) + padding_l = (conv.kernel_size[0] - 1) // 2 + padding_r = conv.kernel_size[0] // 2 + x = F.pad(x, (0, 0, 0, 0, padding_l, padding_r)) + x = conv(x) + x = F.glu(x, dim=2) + if attention is not None: + x = attention(x) + x = (x + residual) * math.sqrt(0.5) + + # T x B x C -> B x T x C + x = x.transpose(1, 0) + + # project back to size of embedding + x = self.fc2(x) + + if encoder_padding_mask is not None: + encoder_padding_mask = encoder_padding_mask.t() # -> B x T + x = x.masked_fill(encoder_padding_mask.unsqueeze(-1), 0) + + # scale gradients (this only affects backward, not forward) + x = GradMultiply.apply(x, 1.0 / (2.0 * self.num_attention_layers)) + + # add output to input embedding for attention + y = (x + input_embedding.transpose(0, 1)) * math.sqrt(0.5) + + return { + "encoder_out": (x, y), + "encoder_padding_mask": encoder_padding_mask, # B x T + } + + def reorder_encoder_out(self, encoder_out, new_order): + encoder_out["encoder_out"] = tuple( + eo.index_select(0, new_order) for eo in encoder_out["encoder_out"] + ) + + if encoder_out["encoder_padding_mask"] is not None: + encoder_out["encoder_padding_mask"] = encoder_out[ + "encoder_padding_mask" + ].index_select(0, new_order) + + if "pretrained" in encoder_out: + encoder_out["pretrained"]["encoder_out"] = tuple( + eo.index_select(0, new_order) + for eo in encoder_out["pretrained"]["encoder_out"] + ) + + return encoder_out + + def max_positions(self): + """Maximum input length supported by the encoder.""" + return self.embed_positions.max_positions + + +@with_incremental_state +class FConvDecoder(FairseqDecoder): + """Convolutional decoder""" + + def __init__( + self, + dictionary, + embed_dim=512, + out_embed_dim=256, + max_positions=1024, + convolutions=((512, 3),) * 8, + attention=True, + dropout=0.1, + selfattention=False, + attention_nheads=1, + selfattention_nheads=1, + project_input=False, + gated_attention=False, + downsample=False, + pretrained=False, + trained_decoder=None, + ): + super().__init__(dictionary) + self.register_buffer("version", torch.Tensor([2])) + self.pretrained = pretrained + self.pretrained_decoder = trained_decoder + self.dropout_module = FairseqDropout( + dropout, module_name=self.__class__.__name__ + ) + self.need_attn = True + in_channels = convolutions[0][0] + + def expand_bool_array(val): + if isinstance(val, bool): + # expand True into [True, True, ...] and do the same with False + return [val] * len(convolutions) + return val + + attention = expand_bool_array(attention) + selfattention = expand_bool_array(selfattention) + + if not isinstance(attention, list) or len(attention) != len(convolutions): + raise ValueError( + "Attention is expected to be a list of booleans of " + "length equal to the number of layers." + ) + + num_embeddings = len(dictionary) + padding_idx = dictionary.pad() + self.embed_tokens = Embedding(num_embeddings, embed_dim, padding_idx) + + self.embed_positions = PositionalEmbedding( + max_positions, + embed_dim, + padding_idx, + ) + + self.fc1 = Linear(embed_dim, in_channels, dropout=dropout) + self.projections = nn.ModuleList() + self.convolutions = nn.ModuleList() + self.attention = nn.ModuleList() + self.selfattention = nn.ModuleList() + self.attproj = nn.ModuleList() + for i, (out_channels, kernel_size) in enumerate(convolutions): + self.projections.append( + Linear(in_channels, out_channels) + if in_channels != out_channels + else None + ) + self.convolutions.append( + LinearizedConv1d( + in_channels, + out_channels * 2, + kernel_size, + padding=(kernel_size - 1), + dropout=dropout, + ) + ) + + self.attention.append( + DownsampledMultiHeadAttention( + out_channels, + embed_dim, + attention_nheads, + project_input=project_input, + gated=False, + downsample=False, + ) + if attention[i] + else None + ) + + self.attproj.append( + Linear(out_channels, embed_dim, dropout=dropout) + if attention[i] + else None + ) + self.selfattention.append( + SelfAttention( + out_channels, + embed_dim, + selfattention_nheads, + project_input=project_input, + gated=gated_attention, + downsample=downsample, + ) + if selfattention[i] + else None + ) + in_channels = out_channels + + self.fc2 = Linear(in_channels, out_embed_dim) + self.fc3 = Linear(out_embed_dim, num_embeddings, dropout=dropout) + + # model fusion + if self.pretrained: + # independent gates are learned from the concatenated input + self.gate1 = nn.Sequential( + Linear(out_embed_dim * 2, out_embed_dim), nn.Sigmoid() + ) + self.gate2 = nn.Sequential( + Linear(out_embed_dim * 2, out_embed_dim), nn.Sigmoid() + ) + # pretrained and trained models are joined + self.joining = nn.Sequential( + Linear(out_embed_dim * 2, out_embed_dim * 2), + LayerNorm(out_embed_dim * 2), + nn.GLU(), + Linear(out_embed_dim, out_embed_dim * 2), + LayerNorm(out_embed_dim * 2), + nn.GLU(), + Linear(out_embed_dim, out_embed_dim), + LayerNorm(out_embed_dim), + ) + # pretrained model contains an output layer that is nhid -> vocab size + # but the models are combined in their hidden state + # the hook stores the output of the pretrained model forward + self.pretrained_outputs = {} + + def save_output(): + def hook(a, b, output): + self.pretrained_outputs["out"] = output + + return hook + + self.pretrained_decoder.fc2.register_forward_hook(save_output()) + + def forward(self, prev_output_tokens, encoder_out): + trained_encoder_out = encoder_out["pretrained"] if self.pretrained else None + encoder_out = encoder_out["encoder"]["encoder_out"] + + encoder_a, encoder_b = self._split_encoder_out(encoder_out) + + # embed positions + positions = self.embed_positions(prev_output_tokens) + + # embed tokens and positions + x = self.embed_tokens(prev_output_tokens) + positions + x = self.dropout_module(x) + target_embedding = x.transpose(0, 1) + + # project to size of convolution + x = self.fc1(x) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + + # temporal convolutions + avg_attn_scores = None + for proj, conv, attention, selfattention, attproj in zip( + self.projections, + self.convolutions, + self.attention, + self.selfattention, + self.attproj, + ): + residual = x if proj is None else proj(x) + + x = self.dropout_module(x) + x = conv(x) + x = F.glu(x, dim=2) + + # attention + if attention is not None: + r = x + x, attn_scores = attention( + attproj(x) + target_embedding, encoder_a, encoder_b + ) + x = x + r + if not self.training and self.need_attn: + if avg_attn_scores is None: + avg_attn_scores = attn_scores + else: + avg_attn_scores.add_(attn_scores) + + if selfattention is not None: + x = selfattention(x) + + x = (x + residual) * math.sqrt(0.5) + + # T x B x C -> B x T x C + x = x.transpose(0, 1) + + # project back to size of vocabulary + x = self.fc2(x) + x = self.dropout_module(x) + if not self.pretrained: + x = self.fc3(x) + + # fusion gating + if self.pretrained: + trained_x, _ = self.pretrained_decoder.forward( + prev_output_tokens, trained_encoder_out + ) + y = torch.cat([x, self.pretrained_outputs["out"]], dim=-1) + gate1 = self.gate1(y) + gate2 = self.gate2(y) + gated_x1 = gate1 * x + gated_x2 = gate2 * self.pretrained_outputs["out"] + fusion = torch.cat([gated_x1, gated_x2], dim=-1) + fusion = self.joining(fusion) + fusion_output = self.fc3(fusion) + return fusion_output, avg_attn_scores + else: + return x, avg_attn_scores + + def max_positions(self): + """Maximum output length supported by the decoder.""" + return self.embed_positions.max_positions + + def make_generation_fast_(self, need_attn=False, **kwargs): + self.need_attn = need_attn + + def _split_encoder_out(self, encoder_out): + """Split and transpose encoder outputs.""" + # transpose only once to speed up attention layers + encoder_a, encoder_b = encoder_out + encoder_a = encoder_a.transpose(0, 1).contiguous() + encoder_b = encoder_b.transpose(0, 1).contiguous() + result = (encoder_a, encoder_b) + return result + + +class SelfAttention(nn.Module): + def __init__( + self, + out_channels, + embed_dim, + num_heads, + project_input=False, + gated=False, + downsample=False, + ): + super().__init__() + self.attention = DownsampledMultiHeadAttention( + out_channels, + embed_dim, + num_heads, + dropout=0, + bias=True, + project_input=project_input, + gated=gated, + downsample=downsample, + ) + self.in_proj_q = Linear(out_channels, embed_dim) + self.in_proj_k = Linear(out_channels, embed_dim) + self.in_proj_v = Linear(out_channels, embed_dim) + self.ln = LayerNorm(out_channels) + + def forward(self, x): + residual = x + query = self.in_proj_q(x) + key = self.in_proj_k(x) + value = self.in_proj_v(x) + x, _ = self.attention( + query, key, value, mask_future_timesteps=True, use_scalar_bias=True + ) + return self.ln(x + residual) + + +def Embedding(num_embeddings, embedding_dim, padding_idx): + m = nn.Embedding(num_embeddings, embedding_dim, padding_idx=padding_idx) + m.weight.data.normal_(0, 0.1) + return m + + +def PositionalEmbedding(num_embeddings, embedding_dim, padding_idx): + m = LearnedPositionalEmbedding(num_embeddings, embedding_dim, padding_idx) + m.weight.data.normal_(0, 0.1) + return m + + +def Linear(in_features, out_features, dropout=0.0): + """Weight-normalized Linear layer (input: N x T x C)""" + m = nn.Linear(in_features, out_features) + m.weight.data.normal_(mean=0, std=math.sqrt((1 - dropout) / in_features)) + m.bias.data.zero_() + return m + + +def LinearizedConv1d(in_channels, out_channels, kernel_size, dropout=0.0, **kwargs): + """Weight-normalized Conv1d layer optimized for decoding""" + m = LinearizedConvolution(in_channels, out_channels, kernel_size, **kwargs) + std = math.sqrt((4 * (1.0 - dropout)) / (m.kernel_size[0] * in_channels)) + m.weight.data.normal_(mean=0, std=std) + m.bias.data.zero_() + return m + + +def ConvTBC(in_channels, out_channels, kernel_size, dropout=0.0, **kwargs): + """Weight-normalized Conv1d layer""" + from fairseq.modules import ConvTBC + + m = ConvTBC(in_channels, out_channels, kernel_size, **kwargs) + std = math.sqrt((4 * (1.0 - dropout)) / (m.kernel_size[0] * in_channels)) + m.weight.data.normal_(mean=0, std=std) + m.bias.data.zero_() + return m + + +@register_model_architecture("fconv_self_att", "fconv_self_att") +def base_architecture(args): + args.dropout = getattr(args, "dropout", 0.1) + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512) + args.encoder_layers = getattr(args, "encoder_layers", "[(512, 3)] * 3") + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 512) + args.decoder_layers = getattr(args, "decoder_layers", "[(512, 3)] * 8") + args.decoder_out_embed_dim = getattr(args, "decoder_out_embed_dim", 256) + args.decoder_attention = getattr(args, "decoder_attention", "True") + args.self_attention = getattr(args, "self_attention", "False") + args.encoder_attention = getattr(args, "encoder_attention", "False") + args.multihead_attention_nheads = getattr(args, "multihead_attention_nheads", 1) + args.multihead_self_attention_nheads = getattr( + args, "multihead_self_attention_nheads", 1 + ) + args.encoder_attention_nheads = getattr(args, "encoder_attention_nheads", 1) + args.project_input = getattr(args, "project_input", "False") + args.gated_attention = getattr(args, "gated_attention", "False") + args.downsample = getattr(args, "downsample", "False") + args.pretrained_checkpoint = getattr(args, "pretrained_checkpoint", "") + args.pretrained = getattr(args, "pretrained", "False") + + +@register_model_architecture("fconv_self_att", "fconv_self_att_wp") +def fconv_self_att_wp(args): + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 256) + args.encoder_layers = getattr( + args, "encoder_layers", "[(128, 3)] * 2 + [(512,3)] * 1" + ) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 256) + args.decoder_layers = getattr( + args, "decoder_layers", "[(512, 4)] * 4 + [(768, 4)] * 2 + [(1024, 4)] * 1" + ) + args.decoder_out_embed_dim = getattr(args, "decoder_out_embed_dim", 256) + args.self_attention = getattr(args, "self_attention", "True") + args.multihead_self_attention_nheads = getattr( + args, "multihead_self_attention_nheads", 4 + ) + args.project_input = getattr(args, "project_input", "True") + args.gated_attention = getattr(args, "gated_attention", "True") + args.downsample = getattr(args, "downsample", "True") + base_architecture(args) diff --git a/fairseq/models/hubert/__init__.py b/fairseq/models/hubert/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..a1b0eabbdbcaf12b15bb96b329ab1e276256f79a --- /dev/null +++ b/fairseq/models/hubert/__init__.py @@ -0,0 +1,7 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .hubert import * # noqa +from .hubert_asr import * # noqa diff --git a/fairseq/models/hubert/hubert.py b/fairseq/models/hubert/hubert.py new file mode 100644 index 0000000000000000000000000000000000000000..232a5e402a146023e5c93f3c2574ecec98faf9d5 --- /dev/null +++ b/fairseq/models/hubert/hubert.py @@ -0,0 +1,563 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +from typing import Dict, List, Optional, Tuple + +import numpy as np + +import torch +import torch.nn as nn +from dataclasses import dataclass, field +from fairseq import utils +from fairseq.data.data_utils import compute_mask_indices +from fairseq.data.dictionary import Dictionary +from fairseq.dataclass import ChoiceEnum, FairseqDataclass +from fairseq.models import BaseFairseqModel, register_model +from fairseq.models.wav2vec.wav2vec2 import ( + ConvFeatureExtractionModel, + TransformerEncoder, +) +from fairseq.modules import GradMultiply, LayerNorm +from fairseq.tasks.hubert_pretraining import ( + HubertPretrainingConfig, + HubertPretrainingTask, +) +from omegaconf import II + +logger = logging.getLogger(__name__) + +EXTRACTOR_MODE_CHOICES = ChoiceEnum(["default", "layer_norm"]) +MASKING_DISTRIBUTION_CHOICES = ChoiceEnum( + ["static", "uniform", "normal", "poisson"] +) + + +@dataclass +class HubertConfig(FairseqDataclass): + label_rate: int = II("task.label_rate") + + extractor_mode: EXTRACTOR_MODE_CHOICES = field( + default="default", + metadata={ + "help": "mode for feature extractor. default has a single group " + "norm with d groups in the first conv block, whereas layer_norm " + "has layer norms in every block (meant to use with normalize=True)" + }, + ) + encoder_layers: int = field( + default=12, metadata={"help": "num encoder layers in the transformer"} + ) + encoder_embed_dim: int = field( + default=768, metadata={"help": "encoder embedding dimension"} + ) + encoder_ffn_embed_dim: int = field( + default=3072, metadata={"help": "encoder embedding dimension for FFN"} + ) + encoder_attention_heads: int = field( + default=12, metadata={"help": "num encoder attention heads"} + ) + activation_fn: ChoiceEnum(utils.get_available_activation_fns()) = field( + default="gelu", metadata={"help": "activation function to use"} + ) + + # dropouts + dropout: float = field( + default=0.1, + metadata={"help": "dropout probability for the transformer"}, + ) + attention_dropout: float = field( + default=0.1, + metadata={"help": "dropout probability for attention weights"}, + ) + activation_dropout: float = field( + default=0.0, + metadata={"help": "dropout probability after activation in FFN"}, + ) + encoder_layerdrop: float = field( + default=0.0, + metadata={"help": "probability of dropping a tarnsformer layer"}, + ) + dropout_input: float = field( + default=0.0, + metadata={"help": "dropout to apply to the input (after feat extr)"}, + ) + dropout_features: float = field( + default=0.0, + metadata={ + "help": "dropout to apply to the features (after feat extr)" + }, + ) + + final_dim: int = field( + default=0, + metadata={ + "help": "project final representations and targets to this many " + "dimensions. set to encoder_embed_dim is <= 0" + }, + ) + untie_final_proj: bool = field( + default=False, + metadata={"help": "use separate projection for each target"}, + ) + layer_norm_first: bool = field( + default=False, + metadata={"help": "apply layernorm first in the transformer"}, + ) + conv_feature_layers: str = field( + default="[(512,10,5)] + [(512,3,2)] * 4 + [(512,2,2)] * 2", + metadata={ + "help": "string describing convolutional feature extraction " + "layers in form of a python list that contains " + "[(dim, kernel_size, stride), ...]" + }, + ) + conv_bias: bool = field( + default=False, metadata={"help": "include bias in conv encoder"} + ) + logit_temp: float = field( + default=0.1, metadata={"help": "temperature to divide logits by"} + ) + target_glu: bool = field( + default=False, metadata={"help": "adds projection + glu to targets"} + ) + feature_grad_mult: float = field( + default=1.0, + metadata={"help": "multiply feature extractor var grads by this"}, + ) + + # masking + mask_length: int = field(default=10, metadata={"help": "mask length"}) + mask_prob: float = field( + default=0.65, + metadata={"help": "probability of replacing a token with mask"}, + ) + mask_selection: MASKING_DISTRIBUTION_CHOICES = field( + default="static", metadata={"help": "how to choose mask length"} + ) + mask_other: float = field( + default=0, + metadata={ + "help": "secondary mask argument " + "(used for more complex distributions), " + "see help in compute_mask_indicesh" + }, + ) + no_mask_overlap: bool = field( + default=False, metadata={"help": "whether to allow masks to overlap"} + ) + mask_min_space: int = field( + default=1, + metadata={ + "help": "min space between spans (if no overlap is enabled)" + }, + ) + + # channel masking + mask_channel_length: int = field( + default=10, + metadata={"help": "length of the mask for features (channels)"}, + ) + mask_channel_prob: float = field( + default=0.0, + metadata={"help": "probability of replacing a feature with 0"}, + ) + mask_channel_selection: MASKING_DISTRIBUTION_CHOICES = field( + default="static", + metadata={"help": "how to choose mask length for channel masking"}, + ) + mask_channel_other: float = field( + default=0, + metadata={ + "help": "secondary mask argument " + "(used for more complex distributions), " + "see help in compute_mask_indicesh" + }, + ) + no_mask_channel_overlap: bool = field( + default=False, + metadata={"help": "whether to allow channel masks to overlap"}, + ) + mask_channel_min_space: int = field( + default=1, + metadata={ + "help": "min space between spans (if no overlap is enabled)" + }, + ) + + # positional embeddings + conv_pos: int = field( + default=128, + metadata={ + "help": "number of filters for convolutional positional embeddings" + }, + ) + conv_pos_groups: int = field( + default=16, + metadata={ + "help": "number of groups for convolutional positional embedding" + }, + ) + + latent_temp: Tuple[float, float, float] = field( + default=(2, 0.5, 0.999995), + metadata={"help": "legacy (to be removed)"}, + ) + + # loss computation + skip_masked: bool = field( + default=False, + metadata={"help": "skip computing losses over masked frames"}, + ) + skip_nomask: bool = field( + default=False, + metadata={"help": "skip computing losses over unmasked frames"}, + ) + + +@register_model("hubert", dataclass=HubertConfig) +class HubertModel(BaseFairseqModel): + def __init__( + self, + cfg: HubertConfig, + task_cfg: HubertPretrainingConfig, + dictionaries: List[Dictionary], + ) -> None: + super().__init__() + logger.info(f"HubertModel Config: {cfg}") + + feature_enc_layers = eval(cfg.conv_feature_layers) # noqa + self.embed = feature_enc_layers[-1][0] + + self.feature_extractor = ConvFeatureExtractionModel( + conv_layers=feature_enc_layers, + dropout=0.0, + mode=cfg.extractor_mode, + conv_bias=cfg.conv_bias, + ) + feature_ds_rate = np.prod([s for _, _, s in feature_enc_layers]) + self.feat2tar_ratio = ( + cfg.label_rate * feature_ds_rate / task_cfg.sample_rate + ) + + self.post_extract_proj = ( + nn.Linear(self.embed, cfg.encoder_embed_dim) + if self.embed != cfg.encoder_embed_dim + else None + ) + + self.mask_prob = cfg.mask_prob + self.mask_selection = cfg.mask_selection + self.mask_other = cfg.mask_other + self.mask_length = cfg.mask_length + self.no_mask_overlap = cfg.no_mask_overlap + self.mask_min_space = cfg.mask_min_space + + self.mask_channel_prob = cfg.mask_channel_prob + self.mask_channel_selection = cfg.mask_channel_selection + self.mask_channel_other = cfg.mask_channel_other + self.mask_channel_length = cfg.mask_channel_length + self.no_mask_channel_overlap = cfg.no_mask_channel_overlap + self.mask_channel_min_space = cfg.mask_channel_min_space + + self.dropout_input = nn.Dropout(cfg.dropout_input) + self.dropout_features = nn.Dropout(cfg.dropout_features) + + self.feature_grad_mult = cfg.feature_grad_mult + self.logit_temp = cfg.logit_temp + self.skip_masked = cfg.skip_masked + self.skip_nomask = cfg.skip_nomask + + final_dim = ( + cfg.final_dim if cfg.final_dim > 0 else cfg.encoder_embed_dim + ) + + self.mask_emb = nn.Parameter( + torch.FloatTensor(cfg.encoder_embed_dim).uniform_() + ) + + self.encoder = TransformerEncoder(cfg) + self.layer_norm = LayerNorm(self.embed) + + self.target_glu = None + if cfg.target_glu: + self.target_glu = nn.Sequential( + nn.Linear(final_dim, final_dim * 2), nn.GLU() + ) + + self.untie_final_proj = cfg.untie_final_proj + if self.untie_final_proj: + self.final_proj = nn.Linear( + cfg.encoder_embed_dim, final_dim * len(dictionaries) + ) + else: + self.final_proj = nn.Linear(cfg.encoder_embed_dim, final_dim) + + # modules below are not needed during fine-tuning + if any([d is None for d in dictionaries]): + logger.info( + "cannot find dictionary. assume will be used for fine-tuning" + ) + else: + self.num_classes = [len(d) for d in dictionaries] + self.label_embs_concat = nn.Parameter( + torch.FloatTensor(sum(self.num_classes), final_dim) + ) + nn.init.uniform_(self.label_embs_concat) + + def upgrade_state_dict_named(self, state_dict, name): + """Upgrade a (possibly old) state dict for new versions of fairseq.""" + + super().upgrade_state_dict_named(state_dict, name) + return state_dict + + @classmethod + def build_model(cls, cfg: HubertConfig, task: HubertPretrainingTask): + """Build a new model instance.""" + + model = HubertModel(cfg, task.cfg, task.dictionaries) + return model + + def apply_mask(self, x, padding_mask, target_list): + B, T, C = x.shape + if self.mask_prob > 0: + mask_indices = compute_mask_indices( + (B, T), + padding_mask, + self.mask_prob, + self.mask_length, + self.mask_selection, + self.mask_other, + min_masks=2, + no_overlap=self.no_mask_overlap, + min_space=self.mask_min_space, + ) + mask_indices = torch.from_numpy(mask_indices).to(x.device) + x[mask_indices] = self.mask_emb + else: + mask_indices = None + + if self.mask_channel_prob > 0: + mask_channel_indices = compute_mask_indices( + (B, C), + None, + self.mask_channel_prob, + self.mask_channel_length, + self.mask_channel_selection, + self.mask_channel_other, + no_overlap=self.no_mask_channel_overlap, + min_space=self.mask_channel_min_space, + ) + mask_channel_indices = ( + torch.from_numpy(mask_channel_indices) + .to(x.device) + .unsqueeze(1) + .expand(-1, T, -1) + ) + x[mask_channel_indices] = 0 + + return x, mask_indices + + def compute_nce(self, x, pos, negs): + neg_is_pos = (pos == negs).all(-1) + pos = pos.unsqueeze(0) + targets = torch.cat([pos, negs], dim=0) + + logits = torch.cosine_similarity( + x.float(), targets.float(), dim=-1 + ).type_as(x) + logits /= self.logit_temp + if neg_is_pos.any(): + logits[1:][neg_is_pos] = float("-inf") + logits = logits.transpose(0, 1) # (num_x, num_cls+1) + return logits + + def forward_features(self, source: torch.Tensor) -> torch.Tensor: + if self.feature_grad_mult > 0: + features = self.feature_extractor(source) + if self.feature_grad_mult != 1.0: + features = GradMultiply.apply(features, self.feature_grad_mult) + else: + with torch.no_grad(): + features = self.feature_extractor(source) + return features + + def forward_targets( + self, features: torch.Tensor, target_list: List[torch.Tensor], + ) -> Tuple[torch.Tensor, torch.Tensor]: + # Trim features to ensure labels exist and then get aligned labels + feat_tsz = features.size(2) + targ_tsz = min([t.size(1) for t in target_list]) + if self.feat2tar_ratio * feat_tsz > targ_tsz: + feat_tsz = int(targ_tsz / self.feat2tar_ratio) + features = features[..., :feat_tsz] + target_inds = torch.arange(feat_tsz).float() * self.feat2tar_ratio + target_list = [t[:, target_inds.long()] for t in target_list] + return features, target_list + + def forward_padding_mask( + self, features: torch.Tensor, padding_mask: torch.Tensor, + ) -> torch.Tensor: + extra = padding_mask.size(1) % features.size(1) + if extra > 0: + padding_mask = padding_mask[:, :-extra] + padding_mask = padding_mask.view( + padding_mask.size(0), features.size(1), -1 + ) + padding_mask = padding_mask.all(-1) + return padding_mask + + def forward( + self, + source: torch.Tensor, + target_list: Optional[List[torch.Tensor]] = None, + padding_mask: Optional[torch.Tensor] = None, + mask: bool = True, + features_only: bool = False, + output_layer: Optional[int] = None, + ) -> Dict[str, torch.Tensor]: + """output layer is 1-based""" + features = self.forward_features(source) + if target_list is not None: + features, target_list = self.forward_targets(features, target_list) + + features_pen = features.float().pow(2).mean() + + features = features.transpose(1, 2) + features = self.layer_norm(features) + unmasked_features = features.clone() + + if padding_mask is not None: + padding_mask = self.forward_padding_mask(features, padding_mask) + + if self.post_extract_proj is not None: + features = self.post_extract_proj(features) + + features = self.dropout_input(features) + unmasked_features = self.dropout_features(unmasked_features) + + if mask: + x, mask_indices = self.apply_mask( + features, padding_mask, target_list + ) + else: + x = features + mask_indices = None + + # feature: (B, T, D), float + # target: (B, T), long + # x: (B, T, D), float + # padding_mask: (B, T), bool + # mask_indices: (B, T), bool + x, _ = self.encoder( + x, + padding_mask=padding_mask, + layer=None if output_layer is None else output_layer - 1 + ) + + if features_only: + return {"x": x, "padding_mask": padding_mask, "features": features} + + def compute_pred(proj_x, target, label_embs): + # compute logits for the i-th label set + y = torch.index_select(label_embs, 0, target.long()) + negs = label_embs.unsqueeze(1).expand(-1, proj_x.size(0), -1) + if self.target_glu: + y = self.target_glu(y) + negs = self.target_glu(negs) + # proj_x: (S, D) + # y: (S, D) + # negs: (Neg, S, D) + return self.compute_nce(proj_x, y, negs) + + label_embs_list = self.label_embs_concat.split(self.num_classes, 0) + + if not self.skip_masked: + masked_indices = torch.logical_and(~padding_mask, mask_indices) + proj_x_m = self.final_proj(x[masked_indices]) + if self.untie_final_proj: + proj_x_m_list = proj_x_m.chunk(len(target_list), dim=-1) + else: + proj_x_m_list = [proj_x_m for _ in range(len(target_list))] + logit_m_list = [ + compute_pred(proj_x_m, t[masked_indices], label_embs_list[i]) + for i, (proj_x_m, t) in enumerate( + zip(proj_x_m_list, target_list) + ) + ] + else: + logit_m_list = [None for _ in target_list] + + if not self.skip_nomask: + nomask_indices = torch.logical_and(~padding_mask, ~mask_indices) + proj_x_u = self.final_proj(x[nomask_indices]) + if self.untie_final_proj: + proj_x_u_list = proj_x_u.chunk(len(target_list), dim=-1) + else: + proj_x_u_list = [proj_x_u for _ in range(len(target_list))] + + logit_u_list = [ + compute_pred(proj_x_u, t[nomask_indices], label_embs_list[i]) + for i, (proj_x_u, t) in enumerate( + zip(proj_x_u_list, target_list) + ) + ] + else: + logit_u_list = [None for _ in target_list] + + result = { + "logit_m_list": logit_m_list, + "logit_u_list": logit_u_list, + "padding_mask": padding_mask, + "features_pen": features_pen, + } + return result + + def extract_features( + self, + source: torch.Tensor, + padding_mask: Optional[torch.Tensor] = None, + mask: bool = False, + ret_conv: bool = False, + output_layer: Optional[int] = None, + ) -> Tuple[torch.Tensor, torch.Tensor]: + res = self.forward( + source, + padding_mask=padding_mask, + mask=mask, + features_only=True, + output_layer=output_layer, + ) + feature = res["features"] if ret_conv else res["x"] + return feature, res["padding_mask"] + + def get_logits(self, net_output, is_masked=True): + if is_masked: + logits_list = net_output["logit_m_list"] + else: + logits_list = net_output["logit_u_list"] + logits_list = [x.float() for x in logits_list if x is not None] + return logits_list + + def get_targets(self, net_output, is_masked=True): + logits_list = self.get_logits(net_output, is_masked) + targets_list = [ + x.new_zeros(x.size(0), dtype=torch.long) for x in logits_list + ] + return targets_list + + def get_extra_losses(self, net_output): + extra_losses = [] + names = [] + + if "features_pen" in net_output: + extra_losses.append(net_output["features_pen"]) + names.append("features_pen") + + return extra_losses, names + + def remove_pretraining_modules(self): + self.target_glu = None + self.final_proj = None diff --git a/fairseq/models/hubert/hubert_asr.py b/fairseq/models/hubert/hubert_asr.py new file mode 100644 index 0000000000000000000000000000000000000000..4cb3fb71537643b560b493ff1c7fc17843e1e49e --- /dev/null +++ b/fairseq/models/hubert/hubert_asr.py @@ -0,0 +1,373 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import contextlib +from argparse import Namespace +from typing import Any + +import torch +import torch.nn as nn +from dataclasses import dataclass, field +from fairseq import checkpoint_utils, tasks, utils +from fairseq.dataclass import FairseqDataclass +from fairseq.dataclass.utils import convert_namespace_to_omegaconf +from fairseq.models import BaseFairseqModel, FairseqEncoder, register_model +from fairseq.models.hubert.hubert import MASKING_DISTRIBUTION_CHOICES +from fairseq.tasks import FairseqTask +from omegaconf import II, MISSING + + +@dataclass +class HubertAsrConfig(FairseqDataclass): + w2v_path: str = field( + default=MISSING, metadata={"help": "path to hubert model"} + ) + no_pretrained_weights: bool = field( + default=False, + metadata={"help": "if true, does not load pretrained weights"}, + ) + dropout_input: float = field( + default=0.0, + metadata={"help": "dropout to apply to the input (after feat extr)"}, + ) + final_dropout: float = field( + default=0.0, + metadata={ + "help": "dropout after transformer and before final projection" + }, + ) + dropout: float = field( + default=0.0, + metadata={"help": "dropout probability inside hubert model"}, + ) + attention_dropout: float = field( + default=0.0, + metadata={ + "help": "dropout probability for attention weights " + "inside hubert model" + }, + ) + activation_dropout: float = field( + default=0.0, + metadata={ + "help": "dropout probability after activation in FFN " + "inside hubert model" + }, + ) + + # masking + apply_mask: bool = field( + default=False, metadata={"help": "apply masking during fine-tuning"} + ) + mask_length: int = field( + default=10, metadata={"help": "repeat the mask indices multiple times"} + ) + mask_prob: float = field( + default=0.5, + metadata={ + "help": "probability of replacing a token with mask " + "(normalized by length)" + }, + ) + mask_selection: MASKING_DISTRIBUTION_CHOICES = field( + default="static", metadata={"help": "how to choose masks"} + ) + mask_other: float = field( + default=0, + metadata={ + "help": "secondary mask argument " + "(used for more complex distributions), " + "see help in compute_mask_indices" + }, + ) + no_mask_overlap: bool = field( + default=False, metadata={"help": "whether to allow masks to overlap"} + ) + + # channel masking + mask_channel_length: int = field( + default=10, + metadata={"help": "length of the mask for features (channels)"}, + ) + mask_channel_prob: float = field( + default=0.0, + metadata={"help": "probability of replacing a feature with 0"}, + ) + mask_channel_selection: MASKING_DISTRIBUTION_CHOICES = field( + default="static", + metadata={"help": "how to choose mask length for channel masking"}, + ) + mask_channel_other: float = field( + default=0, + metadata={ + "help": "secondary mask argument " + "(used for more complex distributions), " + "see help in compute_mask_indices" + }, + ) + no_mask_channel_overlap: bool = field( + default=False, + metadata={"help": "whether to allow channel masks to overlap"}, + ) + freeze_finetune_updates: int = field( + default=0, + metadata={"help": "dont finetune hubert for this many updates"}, + ) + feature_grad_mult: float = field( + default=0.0, + metadata={"help": "reset feature grad mult in hubert to this"}, + ) + layerdrop: float = field( + default=0.0, + metadata={"help": "probability of dropping a layer in hubert"}, + ) + normalize: bool = II("task.normalize") + data: str = II("task.data") + + # this holds the loaded hubert args + w2v_args: Any = None + + +@dataclass +class HubertCtcConfig(HubertAsrConfig): + pass + + +@register_model("hubert_ctc", dataclass=HubertCtcConfig) +class HubertCtc(BaseFairseqModel): + def __init__(self, cfg: HubertCtcConfig, w2v_encoder: BaseFairseqModel): + super().__init__() + self.cfg = cfg + self.w2v_encoder = w2v_encoder + + def upgrade_state_dict_named(self, state_dict, name): + super().upgrade_state_dict_named(state_dict, name) + return state_dict + + @classmethod + def build_model(cls, cfg: HubertCtcConfig, task: FairseqTask): + """Build a new model instance.""" + w2v_encoder = HubertEncoder(cfg, task.target_dictionary) + return cls(cfg, w2v_encoder) + + def get_normalized_probs(self, net_output, log_probs): + """Get normalized probabilities (or log probs) from a net's output.""" + + logits = net_output["encoder_out"] + if log_probs: + return utils.log_softmax(logits.float(), dim=-1) + else: + return utils.softmax(logits.float(), dim=-1) + + def get_logits(self, net_output): + logits = net_output["encoder_out"] + padding = net_output["encoder_padding_mask"] + if padding is not None and padding.any(): + padding = padding.T + logits[padding][..., 0] = 0 + logits[padding][..., 1:] = float("-inf") + + return logits + + def forward(self, **kwargs): + x = self.w2v_encoder(**kwargs) + return x + + +@dataclass +class HubertSeq2SeqConfig(HubertAsrConfig): + decoder_embed_dim: int = field( + default=768, metadata={"help": "decoder embedding dimension"} + ) + decoder_ffn_embed_dim: int = field( + default=3072, metadata={"help": "decoder embedding dimension for FFN"} + ) + decoder_layers: int = field( + default=6, metadata={"help": "num of decoder layers"} + ) + decoder_layerdrop: float = field( + default=0.0, metadata={"help": "decoder layerdrop chance"} + ) + decoder_attention_heads: int = field( + default=4, metadata={"help": "num decoder attention heads"} + ) + decoder_learned_pos: bool = field( + default=False, + metadata={"help": "use learned positional embeddings in the decoder"}, + ) + decoder_normalize_before: bool = field( + default=False, + metadata={"help": "apply layernorm before each decoder block"}, + ) + no_token_positional_embeddings: bool = field( + default=False, + metadata={ + "help": "if set, disables positional embeddings " + "(outside self attention)" + }, + ) + decoder_dropout: float = field( + default=0.0, metadata={"help": "dropout probability in the decoder"} + ) + decoder_attention_dropout: float = field( + default=0.0, + metadata={ + "help": "dropout probability for attention weights " + "inside the decoder" + }, + ) + decoder_activation_dropout: float = field( + default=0.0, + metadata={ + "help": "dropout probability after activation in FFN " + "inside the decoder" + }, + ) + max_target_positions: int = field( + default=2048, metadata={"help": "max target positions"} + ) + share_decoder_input_output_embed: bool = field( + default=False, + metadata={"help": "share decoder input and output embeddings"}, + ) + + +class HubertEncoder(FairseqEncoder): + def __init__(self, cfg: HubertAsrConfig, tgt_dict=None): + self.apply_mask = cfg.apply_mask + + arg_overrides = { + "dropout": cfg.dropout, + "activation_dropout": cfg.activation_dropout, + "dropout_input": cfg.dropout_input, + "attention_dropout": cfg.attention_dropout, + "mask_length": cfg.mask_length, + "mask_prob": cfg.mask_prob, + "mask_selection": cfg.mask_selection, + "mask_other": cfg.mask_other, + "no_mask_overlap": cfg.no_mask_overlap, + "mask_channel_length": cfg.mask_channel_length, + "mask_channel_prob": cfg.mask_channel_prob, + "mask_channel_selection": cfg.mask_channel_selection, + "mask_channel_other": cfg.mask_channel_other, + "no_mask_channel_overlap": cfg.no_mask_channel_overlap, + "encoder_layerdrop": cfg.layerdrop, + "feature_grad_mult": cfg.feature_grad_mult, + } + + if cfg.w2v_args is None: + state = checkpoint_utils.load_checkpoint_to_cpu( + cfg.w2v_path, arg_overrides + ) + w2v_args = state.get("cfg", None) + if w2v_args is None: + w2v_args = convert_namespace_to_omegaconf(state["args"]) + cfg.w2v_args = w2v_args + else: + state = None + w2v_args = cfg.w2v_args + if isinstance(w2v_args, Namespace): + cfg.w2v_args = w2v_args = convert_namespace_to_omegaconf( + w2v_args + ) + + assert cfg.normalize == w2v_args.task.normalize, ( + "Fine-tuning works best when data normalization is the same. " + "Please check that --normalize is set or unset for " + "both pre-training and here" + ) + + w2v_args.task.data = cfg.data + task = tasks.setup_task(w2v_args.task) + model = task.build_model(w2v_args.model) + + if state is not None and not cfg.no_pretrained_weights: + # set strict=False because we omit some modules + model.load_state_dict(state["model"], strict=False) + + model.remove_pretraining_modules() + + super().__init__(task.source_dictionary) + + d = w2v_args.model.encoder_embed_dim + + self.w2v_model = model + + self.final_dropout = nn.Dropout(cfg.final_dropout) + self.freeze_finetune_updates = cfg.freeze_finetune_updates + self.num_updates = 0 + + if tgt_dict is not None: + self.proj = Linear(d, len(tgt_dict)) + elif getattr(cfg, "decoder_embed_dim", d) != d: + self.proj = Linear(d, cfg.decoder_embed_dim) + else: + self.proj = None + + def set_num_updates(self, num_updates): + """Set the number of parameters updates.""" + super().set_num_updates(num_updates) + self.num_updates = num_updates + + def forward(self, source, padding_mask, tbc=True, **kwargs): + + w2v_args = { + "source": source, + "padding_mask": padding_mask, + "mask": self.apply_mask and self.training, + } + + ft = self.freeze_finetune_updates <= self.num_updates + + with torch.no_grad() if not ft else contextlib.ExitStack(): + x, padding_mask = self.w2v_model.extract_features(**w2v_args) + + if tbc: + # B x T x C -> T x B x C + x = x.transpose(0, 1) + + x = self.final_dropout(x) + + if self.proj: + x = self.proj(x) + + return { + "encoder_out": x, # T x B x C + "encoder_padding_mask": padding_mask, # B x T + "padding_mask": padding_mask, + } + + def reorder_encoder_out(self, encoder_out, new_order): + if encoder_out["encoder_out"] is not None: + encoder_out["encoder_out"] = encoder_out[ + "encoder_out" + ].index_select(1, new_order) + if encoder_out["encoder_padding_mask"] is not None: + encoder_out["encoder_padding_mask"] = encoder_out[ + "encoder_padding_mask" + ].index_select(0, new_order) + return encoder_out + + def max_positions(self): + """Maximum input length supported by the encoder.""" + return None + + def upgrade_state_dict_named(self, state_dict, name): + return state_dict + + +def Embedding(num_embeddings, embedding_dim, padding_idx): + m = nn.Embedding(num_embeddings, embedding_dim, padding_idx=padding_idx) + nn.init.normal_(m.weight, mean=0, std=embedding_dim ** -0.5) + nn.init.constant_(m.weight[padding_idx], 0) + return m + + +def Linear(in_features, out_features, bias=True): + m = nn.Linear(in_features, out_features, bias) + nn.init.xavier_uniform_(m.weight) + if bias: + nn.init.constant_(m.bias, 0.0) + return m diff --git a/fairseq/models/huggingface/__init__.py b/fairseq/models/huggingface/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..f7911c2c8edf516855023a285b18935e5389ec02 --- /dev/null +++ b/fairseq/models/huggingface/__init__.py @@ -0,0 +1,20 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import importlib +import os + + +# automatically import any Python files in the models/huggingface/ directory +models_dir = os.path.dirname(__file__) +for file in os.listdir(models_dir): + path = os.path.join(models_dir, file) + if ( + not file.startswith("_") + and not file.startswith(".") + and (file.endswith(".py") or os.path.isdir(path)) + ): + model_name = file[: file.find(".py")] if file.endswith(".py") else file + module = importlib.import_module("fairseq.models.huggingface." + model_name) diff --git a/fairseq/models/huggingface/hf_gpt2.py b/fairseq/models/huggingface/hf_gpt2.py new file mode 100644 index 0000000000000000000000000000000000000000..3a8eb78198f5808557092f814e92f1c9d72933ec --- /dev/null +++ b/fairseq/models/huggingface/hf_gpt2.py @@ -0,0 +1,168 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import os +import sys +from typing import Dict, List, Optional + +import torch +from fairseq.models import ( + FairseqIncrementalDecoder, + FairseqLanguageModel, + register_model, + register_model_architecture, +) + + +logger = logging.getLogger(__name__) + + +DEFAULT_MAX_TARGET_POSITIONS = 1024 + + +@register_model("hf_gpt2") +class HuggingFaceGPT2LanguageModel(FairseqLanguageModel): + def __init__(self, decoder): + super().__init__(decoder) + + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + # fmt: off + parser.add_argument('--embed-dim', type=int, metavar='N', + help='embedding dimension') + parser.add_argument('--num-attention-heads', type=int, metavar='N', + help='num attention heads') + parser.add_argument('--num-layers', type=int, metavar='N', + help='num layers') + parser.add_argument('--dropout', type=float, metavar='D', + help='dropout probability for all fully connected layers ' + 'in the embeddings, encoder, and pooler') + parser.add_argument('--attention-dropout', type=float, metavar='D', + help='dropout probability for attention weights') + # fmt: on + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + default_architecture(args) + return cls(HuggingFaceGPT2Decoder(args, task)) + + +class HuggingFaceGPT2Decoder(FairseqIncrementalDecoder): + def __init__(self, args, task): + try: + from transformers import GPT2Config, GPT2LMHeadModel + except ImportError: + raise ImportError( + "\n\nPlease install huggingface/transformers with:" + "\n\n pip install transformers" + ) + + super().__init__(task.target_dictionary) + + config = GPT2Config( + vocab_size=len(task.target_dictionary), + n_positions=args.max_target_positions + 1, + n_ctx=args.max_target_positions, + n_embd=args.embed_dim, + n_layer=args.num_layers, + n_head=args.num_attention_heads, + resid_pdrop=args.dropout, + embd_pdrop=args.dropout, + attn_pdrop=args.attention_dropout, + layer_norm_epsilon=1e-6, + ) + self.model = GPT2LMHeadModel(config) + + # set zero embedding for padding symbol + self.pad_idx = task.target_dictionary.pad() + self.model.transformer.wte.weight.data[self.pad_idx].zero_() + self.model.transformer.wpe.weight.data[0].zero_() + + def forward( + self, + prev_output_tokens, + src_lengths=None, + incremental_state: Optional[Dict[str, List[torch.Tensor]]] = None, + encoder_out=None, + ): + features = self.extract_features(prev_output_tokens, incremental_state) + lm_logits = self.model.lm_head(features) + return (lm_logits,) + + def extract_features( + self, + prev_output_tokens, + incremental_state: Optional[Dict[str, List[torch.Tensor]]] = None, + ): + if incremental_state: + past = self.get_incremental_state("past") + else: + past = None + + # don't attend to padding symbols + attention_mask = prev_output_tokens.ne(self.pad_idx).int() + + # set position ids to exclude padding symbols + position_ids = attention_mask * ( + torch.arange(1, 1 + prev_output_tokens.size(1)) + .to(prev_output_tokens) + .repeat(prev_output_tokens.size(0), 1) + ) + + outputs = self.model.transformer( + input_ids=prev_output_tokens, + past=past, + attention_mask=attention_mask, + position_ids=position_ids, + ) + last_hidden_states = outputs[0] + + if incremental_state: + self.set_incremental_state(incremental_state, "past", outputs[1]) + + return last_hidden_states + + def max_positions(self): + return self.model.config.n_positions - 1 + + +@register_model_architecture("hf_gpt2", "hf_gpt2") +def default_architecture(args): + if getattr(args, "max_target_positions", None) is None: + args.max_target_positions = getattr( + args, "tokens_per_sample", DEFAULT_MAX_TARGET_POSITIONS + ) + args.embed_dim = getattr(args, "embed_dim", 768) + args.num_attention_heads = getattr(args, "num_attention_heads", 12) + args.num_layers = getattr(args, "num_layers", 12) + args.dropout = getattr(args, "dropout", 0.1) + args.attention_dropout = getattr(args, "attention_dropout", 0.1) + + +@register_model_architecture("hf_gpt2", "hf_gpt2_medium") +def hf_gpt2_medium(args): + args.embed_dim = getattr(args, "embed_dim", 1024) + args.num_attention_heads = getattr(args, "num_attention_heads", 16) + args.num_layers = getattr(args, "num_layers", 24) + default_architecture(args) + + +@register_model_architecture("hf_gpt2", "hf_gpt2_large") +def hf_gpt2_large(args): + args.embed_dim = getattr(args, "embed_dim", 1280) + args.num_attention_heads = getattr(args, "num_attention_heads", 20) + args.num_layers = getattr(args, "num_layers", 36) + default_architecture(args) + + +@register_model_architecture("hf_gpt2", "hf_gpt2_xl") +def hf_gpt2_xl(args): + args.embed_dim = getattr(args, "embed_dim", 1600) + args.num_attention_heads = getattr(args, "num_attention_heads", 25) + args.num_layers = getattr(args, "num_layers", 48) + default_architecture(args) diff --git a/fairseq/models/lightconv.py b/fairseq/models/lightconv.py new file mode 100644 index 0000000000000000000000000000000000000000..b614da366513091132c8b6bd8b8e170cce33a1c4 --- /dev/null +++ b/fairseq/models/lightconv.py @@ -0,0 +1,1018 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +import torch +import torch.nn as nn +import torch.nn.functional as F +from fairseq import utils +from fairseq.models import ( + FairseqEncoder, + FairseqEncoderDecoderModel, + FairseqIncrementalDecoder, + register_model, + register_model_architecture, +) +from fairseq.modules import ( + AdaptiveSoftmax, + DynamicConv, + FairseqDropout, + LayerNorm, + LightweightConv, + MultiheadAttention, + PositionalEmbedding, +) + + +@register_model("lightconv") +class LightConvModel(FairseqEncoderDecoderModel): + """ + LightConv and DynamicConv model from `"Pay Less Attention with Lightweight and Dynamic Convolutions" (Wu, et al, 2019) + <https://openreview.net/pdf?id=SkVhlh09tX>`_. + To use LightConv please set ``--encoder-conv-type lightweight --decoder-conv-type lightweight`` + To use DynamicConv please set ``--encoder-conv-type dynamic --decoder-conv-type dynamic`` + + Args: + encoder (LightConvEncoder): the encoder + decoder (LightConvDecoder): the decoder + + The LightConv model provides the following named architectures and + command-line arguments: + + .. argparse:: + :ref: fairseq.models.lightconv_parser + :prog: + """ + + @classmethod + def hub_models(cls): + # fmt: off + + def moses_subword(path): + return { + 'path': path, + 'tokenizer': 'moses', + 'bpe': 'subword_nmt', + } + + return { + 'lightconv.no_glu.iwslt14.de-en': moses_subword('https://dl.fbaipublicfiles.com/fairseq/models/dynamicconv/iwslt14.de-en.lightconv.tar.gz'), + 'dynamicconv.no_glu.iwslt14.de-en': moses_subword('https://dl.fbaipublicfiles.com/fairseq/models/dynamicconv/iwslt14.de-en.dynamicconv.tar.gz'), + 'lightconv.no_glu.wmt16.en-de': moses_subword('https://dl.fbaipublicfiles.com/fairseq/models/dynamicconv/wmt16.en-de.joined-dict.lightconv.tar.gz'), + 'dynamicconv.no_glu.wmt16.en-de': moses_subword('https://dl.fbaipublicfiles.com/fairseq/models/dynamicconv/wmt16.en-de.joined-dict.dynamicconv.tar.gz'), + 'lightconv.glu.wmt16.en-de': moses_subword('https://dl.fbaipublicfiles.com/fairseq/models/dynamicconv/wmt16.en-de.joined-dict.lightconv-glu.tar.gz'), + 'dynamicconv.glu.wmt16.en-de': moses_subword('https://dl.fbaipublicfiles.com/fairseq/models/dynamicconv/wmt16.en-de.joined-dict.dynamicconv-glu.tar.gz'), + 'lightconv.glu.wmt17.en-de': moses_subword('https://dl.fbaipublicfiles.com/fairseq/models/dynamicconv/wmt16.en-de.joined-dict.lightconv-glu.tar.gz'), + 'dynamicconv.glu.wmt17.en-de': moses_subword('https://dl.fbaipublicfiles.com/fairseq/models/dynamicconv/wmt16.en-de.joined-dict.dynamicconv-glu.tar.gz'), + 'lightconv.glu.wmt14.en-fr': moses_subword('https://dl.fbaipublicfiles.com/fairseq/models/dynamicconv/wmt14.en-fr.joined-dict.lightconv-glu.tar.gz'), + 'dynamicconv.glu.wmt14.en-fr': moses_subword('https://dl.fbaipublicfiles.com/fairseq/models/dynamicconv/wmt14.en-fr.joined-dict.dynamicconv-glu.tar.gz'), + 'lightconv.glu.wmt17.zh-en': moses_subword('https://dl.fbaipublicfiles.com/fairseq/models/dynamicconv/wmt17.zh-en.lightconv-glu.tar.gz'), + 'dynamicconv.glu.wmt17.zh-en': moses_subword('https://dl.fbaipublicfiles.com/fairseq/models/dynamicconv/wmt17.zh-en.dynamicconv-glu.tar.gz'), + } + # fmt: on + + def __init__(self, encoder, decoder): + super().__init__(encoder, decoder) + + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + parser.add_argument( + "--dropout", type=float, metavar="D", help="dropout probability" + ) + parser.add_argument( + "--attention-dropout", + type=float, + metavar="D", + help="dropout probability for attention weights", + ) + parser.add_argument( + "--relu-dropout", + type=float, + metavar="D", + help="dropout probability after ReLU in FFN", + ) + parser.add_argument( + "--input-dropout", + type=float, + metavar="D", + help="dropout probability of the inputs", + ) + parser.add_argument( + "--encoder-embed-path", + type=str, + metavar="STR", + help="path to pre-trained encoder embedding", + ) + parser.add_argument( + "--encoder-embed-dim", + type=int, + metavar="N", + help="encoder embedding dimension", + ) + parser.add_argument( + "--encoder-conv-dim", + type=int, + metavar="N", + help="encoder embedding dimension", + ) + parser.add_argument( + "--encoder-ffn-embed-dim", + type=int, + metavar="N", + help="encoder embedding dimension for FFN", + ) + parser.add_argument( + "--encoder-layers", type=int, metavar="N", help="num encoder layers" + ) + parser.add_argument( + "--encoder-attention-heads", + type=int, + metavar="N", + help="num encoder attention heads or LightConv/DynamicConv heads", + ) + parser.add_argument( + "--encoder-normalize-before", + action="store_true", + help="apply layernorm before each encoder block", + ) + parser.add_argument( + "--encoder-learned-pos", + action="store_true", + help="use learned positional embeddings in the encoder", + ) + parser.add_argument( + "--decoder-embed-path", + type=str, + metavar="STR", + help="path to pre-trained decoder embedding", + ) + parser.add_argument( + "--decoder-embed-dim", + type=int, + metavar="N", + help="decoder embedding dimension", + ) + parser.add_argument( + "--decoder-conv-dim", + type=int, + metavar="N", + help="decoder embedding dimension", + ) + parser.add_argument( + "--decoder-ffn-embed-dim", + type=int, + metavar="N", + help="decoder embedding dimension for FFN", + ) + parser.add_argument( + "--decoder-layers", type=int, metavar="N", help="num decoder layers" + ) + parser.add_argument( + "--decoder-attention-heads", + type=int, + metavar="N", + help="num decoder attention heads or LightConv/DynamicConv heads", + ) + parser.add_argument( + "--decoder-learned-pos", + action="store_true", + help="use learned positional embeddings in the decoder", + ) + parser.add_argument( + "--decoder-normalize-before", + action="store_true", + help="apply layernorm before each decoder block", + ) + parser.add_argument( + "--share-decoder-input-output-embed", + action="store_true", + help="share decoder input and output embeddings", + ) + parser.add_argument( + "--share-all-embeddings", + action="store_true", + help="share encoder, decoder and output embeddings" + " (requires shared dictionary and embed dim)", + ) + parser.add_argument( + "--adaptive-softmax-cutoff", + metavar="EXPR", + help="comma separated list of adaptive softmax cutoff points. " + "Must be used with adaptive_loss criterion", + ), + parser.add_argument( + "--adaptive-softmax-dropout", + type=float, + metavar="D", + help="sets adaptive softmax dropout for the tail projections", + ) + + """LightConv and DynamicConv arguments""" + parser.add_argument( + "--encoder-kernel-size-list", + type=lambda x: utils.eval_str_list(x, int), + help='list of kernel size (default: "[3,7,15,31,31,31,31]")', + ) + parser.add_argument( + "--decoder-kernel-size-list", + type=lambda x: utils.eval_str_list(x, int), + help='list of kernel size (default: "[3,7,15,31,31,31]")', + ) + parser.add_argument( + "--encoder-glu", type=utils.eval_bool, help="glu after in proj" + ) + parser.add_argument( + "--decoder-glu", type=utils.eval_bool, help="glu after in proj" + ) + parser.add_argument( + "--encoder-conv-type", + default="dynamic", + type=str, + choices=["dynamic", "lightweight"], + help="type of convolution", + ) + parser.add_argument( + "--decoder-conv-type", + default="dynamic", + type=str, + choices=["dynamic", "lightweight"], + help="type of convolution", + ) + parser.add_argument("--weight-softmax", default=True, type=utils.eval_bool) + parser.add_argument( + "--weight-dropout", + type=float, + metavar="D", + help="dropout probability for conv weights", + ) + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + + # make sure all arguments are present in older models + base_architecture(args) + + if not hasattr(args, "max_source_positions"): + args.max_source_positions = 1024 + if not hasattr(args, "max_target_positions"): + args.max_target_positions = 1024 + + src_dict, tgt_dict = task.source_dictionary, task.target_dictionary + + def build_embedding(dictionary, embed_dim, path=None): + num_embeddings = len(dictionary) + padding_idx = dictionary.pad() + emb = Embedding(num_embeddings, embed_dim, padding_idx) + # if provided, load from preloaded dictionaries + if path: + embed_dict = utils.parse_embedding(path) + utils.load_embedding(embed_dict, dictionary, emb) + return emb + + if args.share_all_embeddings: + if src_dict != tgt_dict: + raise RuntimeError( + "--share-all-embeddings requires a joined dictionary" + ) + if args.encoder_embed_dim != args.decoder_embed_dim: + raise RuntimeError( + "--share-all-embeddings requires --encoder-embed-dim to match --decoder-embed-dim" + ) + if args.decoder_embed_path and ( + args.decoder_embed_path != args.encoder_embed_path + ): + raise RuntimeError( + "--share-all-embeddings not compatible with --decoder-embed-path" + ) + encoder_embed_tokens = build_embedding( + src_dict, args.encoder_embed_dim, args.encoder_embed_path + ) + decoder_embed_tokens = encoder_embed_tokens + args.share_decoder_input_output_embed = True + else: + encoder_embed_tokens = build_embedding( + src_dict, args.encoder_embed_dim, args.encoder_embed_path + ) + decoder_embed_tokens = build_embedding( + tgt_dict, args.decoder_embed_dim, args.decoder_embed_path + ) + + encoder = LightConvEncoder(args, src_dict, encoder_embed_tokens) + decoder = LightConvDecoder(args, tgt_dict, decoder_embed_tokens) + return LightConvModel(encoder, decoder) + + +class LightConvEncoder(FairseqEncoder): + """ + LightConv encoder consisting of *args.encoder_layers* layers. Each layer + is a :class:`LightConvEncoderLayer`. + + Args: + args (argparse.Namespace): parsed command-line arguments + dictionary (~fairseq.data.Dictionary): encoding dictionary + embed_tokens (torch.nn.Embedding): input embedding + """ + + def __init__(self, args, dictionary, embed_tokens): + super().__init__(dictionary) + self.dropout_module = FairseqDropout( + args.dropout, module_name=self.__class__.__name__ + ) + + embed_dim = embed_tokens.embedding_dim + self.padding_idx = embed_tokens.padding_idx + self.max_source_positions = args.max_source_positions + + self.embed_tokens = embed_tokens + self.embed_scale = math.sqrt(embed_dim) + self.embed_positions = ( + PositionalEmbedding( + args.max_source_positions, + embed_dim, + self.padding_idx, + learned=args.encoder_learned_pos, + ) + if not args.no_token_positional_embeddings + else None + ) + + self.layers = nn.ModuleList([]) + self.layers.extend( + [ + LightConvEncoderLayer( + args, kernel_size=args.encoder_kernel_size_list[i] + ) + for i in range(args.encoder_layers) + ] + ) + self.register_buffer("version", torch.Tensor([2])) + self.normalize = args.encoder_normalize_before + if self.normalize: + self.layer_norm = LayerNorm(embed_dim) + + def forward(self, src_tokens, **unused): + """ + Args: + src_tokens (LongTensor): tokens in the source language of shape + `(batch, src_len)` + + Returns: + dict: + - **encoder_out** (Tensor): the last encoder layer's output of + shape `(src_len, batch, embed_dim)` + - **encoder_padding_mask** (ByteTensor): the positions of + padding elements of shape `(batch, src_len)` + """ + # embed tokens and positions + x = self.embed_scale * self.embed_tokens(src_tokens) + if self.embed_positions is not None: + x += self.embed_positions(src_tokens) + x = self.dropout_module(x) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + + # compute padding mask + encoder_padding_mask = src_tokens.eq(self.padding_idx) + if not encoder_padding_mask.any(): + encoder_padding_mask = None + + # encoder layers + for layer in self.layers: + x = layer(x, encoder_padding_mask) + + if self.normalize: + x = self.layer_norm(x) + + return { + "encoder_out": x, # T x B x C + "encoder_padding_mask": encoder_padding_mask, # B x T + } + + def reorder_encoder_out(self, encoder_out, new_order): + """ + Reorder encoder output according to *new_order*. + + Args: + encoder_out: output from the ``forward()`` method + new_order (LongTensor): desired order + + Returns: + *encoder_out* rearranged according to *new_order* + """ + if encoder_out["encoder_out"] is not None: + encoder_out["encoder_out"] = encoder_out["encoder_out"].index_select( + 1, new_order + ) + if encoder_out["encoder_padding_mask"] is not None: + encoder_out["encoder_padding_mask"] = encoder_out[ + "encoder_padding_mask" + ].index_select(0, new_order) + return encoder_out + + def max_positions(self): + """Maximum input length supported by the encoder.""" + if self.embed_positions is None: + return self.max_source_positions + return min(self.max_source_positions, self.embed_positions.max_positions) + + +class LightConvDecoder(FairseqIncrementalDecoder): + """ + LightConv decoder consisting of *args.decoder_layers* layers. Each layer + is a :class:`LightConvDecoderLayer`. + + Args: + args (argparse.Namespace): parsed command-line arguments + dictionary (~fairseq.data.Dictionary): decoding dictionary + embed_tokens (torch.nn.Embedding): output embedding + no_encoder_attn (bool, optional): whether to attend to encoder outputs. + Default: ``False`` + """ + + def __init__( + self, args, dictionary, embed_tokens, no_encoder_attn=False, final_norm=True + ): + super().__init__(dictionary) + self.dropout_module = FairseqDropout( + args.dropout, module_name=self.__class__.__name__ + ) + self.share_input_output_embed = args.share_decoder_input_output_embed + + input_embed_dim = embed_tokens.embedding_dim + embed_dim = args.decoder_embed_dim + output_embed_dim = args.decoder_output_dim + + padding_idx = embed_tokens.padding_idx + self.max_target_positions = args.max_target_positions + + self.embed_tokens = embed_tokens + self.embed_scale = math.sqrt(embed_dim) # todo: try with input_embed_dim + + self.project_in_dim = ( + Linear(input_embed_dim, embed_dim, bias=False) + if embed_dim != input_embed_dim + else None + ) + + self.embed_positions = ( + PositionalEmbedding( + args.max_target_positions, + embed_dim, + padding_idx, + learned=args.decoder_learned_pos, + ) + if not args.no_token_positional_embeddings + else None + ) + + self.layers = nn.ModuleList([]) + self.layers.extend( + [ + LightConvDecoderLayer( + args, no_encoder_attn, kernel_size=args.decoder_kernel_size_list[i] + ) + for i in range(args.decoder_layers) + ] + ) + + self.adaptive_softmax = None + + self.project_out_dim = ( + Linear(embed_dim, output_embed_dim, bias=False) + if embed_dim != output_embed_dim and not args.tie_adaptive_weights + else None + ) + + if args.adaptive_softmax_cutoff is not None: + self.adaptive_softmax = AdaptiveSoftmax( + len(dictionary), + output_embed_dim, + utils.eval_str_list(args.adaptive_softmax_cutoff, type=int), + dropout=args.adaptive_softmax_dropout, + adaptive_inputs=embed_tokens if args.tie_adaptive_weights else None, + factor=args.adaptive_softmax_factor, + tie_proj=args.tie_adaptive_proj, + ) + elif not self.share_input_output_embed: + self.embed_out = nn.Parameter( + torch.Tensor(len(dictionary), output_embed_dim) + ) + nn.init.normal_(self.embed_out, mean=0, std=output_embed_dim ** -0.5) + self.register_buffer("version", torch.Tensor([2])) + self.normalize = args.decoder_normalize_before and final_norm + if self.normalize: + self.layer_norm = LayerNorm(embed_dim) + + def forward( + self, prev_output_tokens, encoder_out=None, incremental_state=None, **kwargs + ): + """ + Args: + prev_output_tokens (LongTensor): previous decoder outputs of shape + `(batch, tgt_len)`, for teacher forcing + encoder_out (Tensor, optional): output from the encoder, used for + encoder-side attention + incremental_state (dict): dictionary used for storing state during + :ref:`Incremental decoding` + + Returns: + tuple: + - the last decoder layer's output of shape `(batch, tgt_len, + vocab)` + - the last decoder layer's attention weights of shape `(batch, + tgt_len, src_len)` + """ + # embed positions + positions = ( + self.embed_positions( + prev_output_tokens, + incremental_state=incremental_state, + ) + if self.embed_positions is not None + else None + ) + + if incremental_state is not None: + prev_output_tokens = prev_output_tokens[:, -1:] + if positions is not None: + positions = positions[:, -1:] + + # embed tokens and positions + x = self.embed_scale * self.embed_tokens(prev_output_tokens) + + if self.project_in_dim is not None: + x = self.project_in_dim(x) + + if positions is not None: + x += positions + x = self.dropout_module(x) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + attn = None + + inner_states = [x] + + # decoder layers + for layer in self.layers: + x, attn = layer( + x, + encoder_out["encoder_out"] if encoder_out is not None else None, + encoder_out["encoder_padding_mask"] + if encoder_out is not None + else None, + incremental_state, + ) + inner_states.append(x) + + if self.normalize: + x = self.layer_norm(x) + + # T x B x C -> B x T x C + x = x.transpose(0, 1) + + if self.project_out_dim is not None: + x = self.project_out_dim(x) + + if self.adaptive_softmax is None: + # project back to size of vocabulary + if self.share_input_output_embed: + x = F.linear(x, self.embed_tokens.weight) + else: + x = F.linear(x, self.embed_out) + + return x, {"attn": attn, "inner_states": inner_states} + + def max_positions(self): + """Maximum output length supported by the decoder.""" + if self.embed_positions is None: + return self.max_target_positions + return min(self.max_target_positions, self.embed_positions.max_positions) + + def buffered_future_mask(self, tensor): + dim = tensor.size(0) + if ( + not hasattr(self, "_future_mask") + or self._future_mask is None + or self._future_mask.device != tensor.device + ): + self._future_mask = torch.triu( + utils.fill_with_neg_inf(tensor.new(dim, dim)), 1 + ) + if self._future_mask.size(0) < dim: + self._future_mask = torch.triu( + utils.fill_with_neg_inf(self._future_mask.resize_(dim, dim)), 1 + ) + return self._future_mask[:dim, :dim] + + +class LightConvEncoderLayer(nn.Module): + """Encoder layer block. + + Args: + args (argparse.Namespace): parsed command-line arguments + kernel_size: kernel size of the convolution + """ + + def __init__(self, args, kernel_size=0): + super().__init__() + self.embed_dim = args.encoder_embed_dim + self.conv_dim = args.encoder_conv_dim + padding_l = ( + kernel_size // 2 + if kernel_size % 2 == 1 + else ((kernel_size - 1) // 2, kernel_size // 2) + ) + + if args.encoder_glu: + self.linear1 = Linear(self.embed_dim, 2 * self.conv_dim) + self.act = nn.GLU() + else: + self.linear1 = Linear(self.embed_dim, self.conv_dim) + self.act = None + if args.encoder_conv_type == "lightweight": + self.conv = LightweightConv( + self.conv_dim, + kernel_size, + padding_l=padding_l, + weight_softmax=args.weight_softmax, + num_heads=args.encoder_attention_heads, + weight_dropout=args.weight_dropout, + ) + elif args.encoder_conv_type == "dynamic": + self.conv = DynamicConv( + self.conv_dim, + kernel_size, + padding_l=padding_l, + weight_softmax=args.weight_softmax, + num_heads=args.encoder_attention_heads, + weight_dropout=args.weight_dropout, + ) + else: + raise NotImplementedError + self.linear2 = Linear(self.conv_dim, self.embed_dim) + + self.dropout_module = FairseqDropout( + args.dropout, module_name=self.__class__.__name__ + ) + self.relu_dropout_module = FairseqDropout( + args.relu_dropout, module_name=self.__class__.__name__ + ) + self.input_dropout_module = FairseqDropout( + args.input_dropout, module_name=self.__class__.__name__ + ) + self.normalize_before = args.encoder_normalize_before + self.fc1 = Linear(self.embed_dim, args.encoder_ffn_embed_dim) + self.fc2 = Linear(args.encoder_ffn_embed_dim, self.embed_dim) + self.layer_norms = nn.ModuleList([LayerNorm(self.embed_dim) for _ in range(2)]) + + def forward(self, x, encoder_padding_mask): + """ + Args: + x (Tensor): input to the layer of shape `(seq_len, batch, embed_dim)` + encoder_padding_mask (ByteTensor): binary ByteTensor of shape + `(batch, src_len)` where padding elements are indicated by ``1``. + + Returns: + encoded output of shape `(batch, src_len, embed_dim)` + """ + residual = x + x = self.maybe_layer_norm(0, x, before=True) + x = self.input_dropout_module(x) + x = self.linear1(x) + if self.act is not None: + x = self.act(x) + if encoder_padding_mask is not None: + x = x.masked_fill(encoder_padding_mask.transpose(0, 1).unsqueeze(2), 0) + x = self.conv(x) + x = self.linear2(x) + x = self.dropout_module(x) + x = residual + x + x = self.maybe_layer_norm(0, x, after=True) + + residual = x + x = self.maybe_layer_norm(1, x, before=True) + x = F.relu(self.fc1(x)) + x = self.relu_dropout_module(x) + x = self.fc2(x) + x = self.dropout_module(x) + x = residual + x + x = self.maybe_layer_norm(1, x, after=True) + return x + + def maybe_layer_norm(self, i, x, before=False, after=False): + assert before ^ after + if after ^ self.normalize_before: + return self.layer_norms[i](x) + else: + return x + + def extra_repr(self): + return ( + "dropout={}, relu_dropout={}, input_dropout={}, normalize_before={}".format( + self.dropout_module.p, + self.relu_dropout_module.p, + self.input_dropout_module.p, + self.normalize_before, + ) + ) + + +class LightConvDecoderLayer(nn.Module): + """Decoder layer block. + + Args: + args (argparse.Namespace): parsed command-line arguments + no_encoder_attn (bool, optional): whether to attend to encoder outputs. + Default: ``False`` + kernel_size: kernel size of the convolution + """ + + def __init__(self, args, no_encoder_attn=False, kernel_size=0): + super().__init__() + self.embed_dim = args.decoder_embed_dim + self.conv_dim = args.decoder_conv_dim + if args.decoder_glu: + self.linear1 = Linear(self.embed_dim, 2 * self.conv_dim) + self.act = nn.GLU() + else: + self.linear1 = Linear(self.embed_dim, self.conv_dim) + self.act = None + if args.decoder_conv_type == "lightweight": + self.conv = LightweightConv( + self.conv_dim, + kernel_size, + padding_l=kernel_size - 1, + weight_softmax=args.weight_softmax, + num_heads=args.decoder_attention_heads, + weight_dropout=args.weight_dropout, + ) + elif args.decoder_conv_type == "dynamic": + self.conv = DynamicConv( + self.conv_dim, + kernel_size, + padding_l=kernel_size - 1, + weight_softmax=args.weight_softmax, + num_heads=args.decoder_attention_heads, + weight_dropout=args.weight_dropout, + ) + else: + raise NotImplementedError + self.linear2 = Linear(self.conv_dim, self.embed_dim) + + self.dropout_module = FairseqDropout( + args.dropout, module_name=self.__class__.__name__ + ) + self.relu_dropout_module = FairseqDropout( + args.relu_dropout, module_name=self.__class__.__name__ + ) + self.input_dropout_module = FairseqDropout( + args.input_dropout, module_name=self.__class__.__name__ + ) + self.normalize_before = args.decoder_normalize_before + + self.conv_layer_norm = LayerNorm(self.embed_dim) + + if no_encoder_attn: + self.encoder_attn = None + self.encoder_attn_layer_norm = None + else: + self.encoder_attn = MultiheadAttention( + self.embed_dim, + args.decoder_attention_heads, + dropout=args.attention_dropout, + encoder_decoder_attention=True, + ) + self.encoder_attn_layer_norm = LayerNorm(self.embed_dim) + + self.fc1 = Linear(self.embed_dim, args.decoder_ffn_embed_dim) + self.fc2 = Linear(args.decoder_ffn_embed_dim, self.embed_dim) + + self.final_layer_norm = LayerNorm(self.embed_dim) + self.need_attn = True + + def forward( + self, + x, + encoder_out, + encoder_padding_mask, + incremental_state, + prev_conv_state=None, + prev_attn_state=None, + conv_mask=None, + conv_padding_mask=None, + ): + """ + Args: + x (Tensor): input to the layer of shape `(seq_len, batch, embed_dim)` + encoder_padding_mask (ByteTensor): binary ByteTensor of shape + `(batch, src_len)` where padding elements are indicated by ``1``. + + Returns: + encoded output of shape `(batch, src_len, embed_dim)` + """ + residual = x + x = self.maybe_layer_norm(self.conv_layer_norm, x, before=True) + if prev_conv_state is not None: + if incremental_state is None: + incremental_state = {} + self.conv._set_input_buffer(incremental_state, prev_conv_state) + x = self.input_dropout_module(x) + x = self.linear1(x) + if self.act is not None: + x = self.act(x) + x = self.conv(x, incremental_state=incremental_state) + x = self.linear2(x) + x = self.dropout_module(x) + x = residual + x + x = self.maybe_layer_norm(self.conv_layer_norm, x, after=True) + + attn = None + if self.encoder_attn is not None: + residual = x + x = self.maybe_layer_norm(self.encoder_attn_layer_norm, x, before=True) + if prev_attn_state is not None: + if incremental_state is None: + incremental_state = {} + prev_key, prev_value = prev_attn_state + saved_state = {"prev_key": prev_key, "prev_value": prev_value} + self.encoder_attn._set_input_buffer(incremental_state, saved_state) + x, attn = self.encoder_attn( + query=x, + key=encoder_out, + value=encoder_out, + key_padding_mask=encoder_padding_mask, + incremental_state=incremental_state, + static_kv=True, + need_weights=(not self.training and self.need_attn), + ) + x = self.dropout_module(x) + x = residual + x + x = self.maybe_layer_norm(self.encoder_attn_layer_norm, x, after=True) + + residual = x + x = self.maybe_layer_norm(self.final_layer_norm, x, before=True) + x = F.relu(self.fc1(x)) + x = self.relu_dropout_module(x) + x = self.fc2(x) + x = self.dropout_module(x) + x = residual + x + x = self.maybe_layer_norm(self.final_layer_norm, x, after=True) + return x, attn + + def maybe_layer_norm(self, layer_norm, x, before=False, after=False): + assert before ^ after + if after ^ self.normalize_before: + return layer_norm(x) + else: + return x + + def make_generation_fast_(self, need_attn=False, **kwargs): + self.need_attn = need_attn + + def extra_repr(self): + return ( + "dropout={}, relu_dropout={}, input_dropout={}, normalize_before={}".format( + self.dropout_module.p, + self.relu_dropout_module.p, + self.input_dropout_module.p, + self.normalize_before, + ) + ) + + +def Embedding(num_embeddings, embedding_dim, padding_idx): + m = nn.Embedding(num_embeddings, embedding_dim, padding_idx=padding_idx) + nn.init.normal_(m.weight, mean=0, std=embedding_dim ** -0.5) + nn.init.constant_(m.weight[padding_idx], 0) + return m + + +def Linear(in_features, out_features, bias=True): + m = nn.Linear(in_features, out_features, bias) + nn.init.xavier_uniform_(m.weight) + if bias: + nn.init.constant_(m.bias, 0.0) + return m + + +@register_model_architecture("lightconv", "lightconv") +def base_architecture(args): + args.encoder_embed_path = getattr(args, "encoder_embed_path", None) + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 2048) + args.encoder_layers = getattr(args, "encoder_layers", 7) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 8) + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False) + args.encoder_learned_pos = getattr(args, "encoder_learned_pos", False) + args.decoder_embed_path = getattr(args, "decoder_embed_path", None) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", args.encoder_embed_dim) + args.decoder_ffn_embed_dim = getattr( + args, "decoder_ffn_embed_dim", args.encoder_ffn_embed_dim + ) + args.decoder_layers = getattr(args, "decoder_layers", 6) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 8) + args.decoder_normalize_before = getattr(args, "decoder_normalize_before", False) + args.decoder_learned_pos = getattr(args, "decoder_learned_pos", False) + args.attention_dropout = getattr(args, "attention_dropout", 0.0) + args.relu_dropout = getattr(args, "relu_dropout", 0.0) + args.dropout = getattr(args, "dropout", 0.1) + args.adaptive_softmax_cutoff = getattr(args, "adaptive_softmax_cutoff", None) + args.adaptive_softmax_dropout = getattr(args, "adaptive_softmax_dropout", 0) + args.share_decoder_input_output_embed = getattr( + args, "share_decoder_input_output_embed", False + ) + args.share_all_embeddings = getattr(args, "share_all_embeddings", False) + args.no_token_positional_embeddings = getattr( + args, "no_token_positional_embeddings", False + ) + + args.decoder_output_dim = getattr( + args, "decoder_output_dim", args.decoder_embed_dim + ) + args.decoder_input_dim = getattr(args, "decoder_input_dim", args.decoder_embed_dim) + + args.encoder_conv_dim = getattr(args, "encoder_conv_dim", args.encoder_embed_dim) + args.decoder_conv_dim = getattr(args, "decoder_conv_dim", args.decoder_embed_dim) + + args.encoder_kernel_size_list = getattr( + args, "encoder_kernel_size_list", [3, 7, 15, 31, 31, 31, 31] + ) + args.decoder_kernel_size_list = getattr( + args, "decoder_kernel_size_list", [3, 7, 15, 31, 31, 31] + ) + if len(args.encoder_kernel_size_list) == 1: + args.encoder_kernel_size_list = ( + args.encoder_kernel_size_list * args.encoder_layers + ) + if len(args.decoder_kernel_size_list) == 1: + args.decoder_kernel_size_list = ( + args.decoder_kernel_size_list * args.decoder_layers + ) + assert ( + len(args.encoder_kernel_size_list) == args.encoder_layers + ), "encoder_kernel_size_list doesn't match encoder_layers" + assert ( + len(args.decoder_kernel_size_list) == args.decoder_layers + ), "decoder_kernel_size_list doesn't match decoder_layers" + args.encoder_glu = getattr(args, "encoder_glu", True) + args.decoder_glu = getattr(args, "decoder_glu", True) + args.input_dropout = getattr(args, "input_dropout", 0.1) + args.weight_dropout = getattr(args, "weight_dropout", args.attention_dropout) + + +@register_model_architecture("lightconv", "lightconv_iwslt_de_en") +def lightconv_iwslt_de_en(args): + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 1024) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 4) + args.encoder_layers = getattr(args, "encoder_layers", 7) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 512) + args.decoder_ffn_embed_dim = getattr(args, "decoder_ffn_embed_dim", 1024) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 4) + args.decoder_layers = getattr(args, "decoder_layers", 6) + args.attention_dropout = getattr(args, "attention_dropout", 0.1) + args.weight_dropout = getattr(args, "weight_dropout", 0.1) + args.encoder_glu = getattr(args, "encoder_glu", False) + args.decoder_glu = getattr(args, "decoder_glu", False) + args.input_dropout = getattr(args, "input_dropout", 0.0) + base_architecture(args) + + +@register_model_architecture("lightconv", "lightconv_wmt_en_de") +def lightconv_wmt_en_de(args): + base_architecture(args) + + +@register_model_architecture("lightconv", "lightconv_wmt_en_de_big") +def lightconv_wmt_en_de_big(args): + args.attention_dropout = getattr(args, "attention_dropout", 0.1) + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 1024) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 4096) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 16) + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 1024) + args.decoder_ffn_embed_dim = getattr(args, "decoder_ffn_embed_dim", 4096) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 16) + args.dropout = getattr(args, "dropout", 0.3) + base_architecture(args) + + +@register_model_architecture("lightconv", "lightconv_wmt_en_fr_big") +def lightconv_wmt_en_fr_big(args): + args.dropout = getattr(args, "dropout", 0.1) + lightconv_wmt_en_de_big(args) + + +@register_model_architecture("lightconv", "lightconv_wmt_zh_en_big") +def lightconv_wmt_zh_en_big(args): + args.dropout = getattr(args, "dropout", 0.2) + args.attention_dropout = getattr(args, "attention_dropout", 0.2) + args.weight_dropout = getattr(args, "weight_dropout", 0.2) + lightconv_wmt_en_de_big(args) diff --git a/fairseq/models/lightconv_lm.py b/fairseq/models/lightconv_lm.py new file mode 100644 index 0000000000000000000000000000000000000000..1d9efc4e42a5ecc1b83338055f18ade5a83ea666 --- /dev/null +++ b/fairseq/models/lightconv_lm.py @@ -0,0 +1,306 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq import utils +from fairseq.models import ( + FairseqLanguageModel, + register_model, + register_model_architecture, +) +from fairseq.models.lightconv import Embedding, LightConvDecoder +from fairseq.modules import AdaptiveInput, CharacterTokenEmbedder + + +@register_model("lightconv_lm") +class LightConvLanguageModel(FairseqLanguageModel): + def __init__(self, decoder): + super().__init__(decoder) + + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + parser.add_argument( + "--dropout", + default=0.1, + type=float, + metavar="D", + help="dropout probability", + ) + parser.add_argument( + "--attention-dropout", + default=0.0, + type=float, + metavar="D", + help="dropout probability for attention weights", + ) + parser.add_argument( + "--relu-dropout", + default=0.0, + type=float, + metavar="D", + help="dropout probability after ReLU in FFN", + ) + parser.add_argument( + "--input-dropout", + type=float, + metavar="D", + help="dropout probability of the inputs", + ) + parser.add_argument( + "--decoder-embed-dim", + type=int, + metavar="N", + help="decoder embedding dimension", + ) + parser.add_argument( + "--decoder-output-dim", + type=int, + metavar="N", + help="decoder output dimension", + ) + parser.add_argument( + "--decoder-input-dim", type=int, metavar="N", help="decoder input dimension" + ) + parser.add_argument( + "--decoder-ffn-embed-dim", + type=int, + metavar="N", + help="decoder embedding dimension for FFN", + ) + parser.add_argument( + "--decoder-layers", type=int, metavar="N", help="num decoder layers" + ) + parser.add_argument( + "--decoder-attention-heads", + type=int, + metavar="N", + help="num decoder attention heads or LightConv/DynamicConv heads", + ) + parser.add_argument( + "--decoder-normalize-before", + default=False, + action="store_true", + help="apply layernorm before each decoder block", + ) + parser.add_argument( + "--adaptive-softmax-cutoff", + metavar="EXPR", + help="comma separated list of adaptive softmax cutoff points. " + "Must be used with adaptive_loss criterion", + ) + parser.add_argument( + "--adaptive-softmax-dropout", + type=float, + metavar="D", + help="sets adaptive softmax dropout for the tail projections", + ) + parser.add_argument( + "--adaptive-softmax-factor", + type=float, + metavar="N", + help="adaptive input factor", + ) + parser.add_argument( + "--no-token-positional-embeddings", + default=False, + action="store_true", + help="if set, disables positional embeddings (outside self attention)", + ) + parser.add_argument( + "--share-decoder-input-output-embed", + default=False, + action="store_true", + help="share decoder input and output embeddings", + ) + parser.add_argument( + "--character-embeddings", + default=False, + action="store_true", + help="if set, uses character embedding convolutions to produce token embeddings", + ) + parser.add_argument( + "--character-filters", + type=str, + metavar="LIST", + default="[(1, 64), (2, 128), (3, 192), (4, 256), (5, 256), (6, 256), (7, 256)]", + help="size of character embeddings", + ) + parser.add_argument( + "--character-embedding-dim", + type=int, + metavar="N", + default=4, + help="size of character embeddings", + ) + parser.add_argument( + "--char-embedder-highway-layers", + type=int, + metavar="N", + default=2, + help="number of highway layers for character token embeddder", + ) + parser.add_argument( + "--adaptive-input", + default=False, + action="store_true", + help="if set, uses adaptive input", + ) + parser.add_argument( + "--adaptive-input-factor", + type=float, + metavar="N", + help="adaptive input factor", + ) + parser.add_argument( + "--adaptive-input-cutoff", + metavar="EXPR", + help="comma separated list of adaptive input cutoff points.", + ) + parser.add_argument( + "--tie-adaptive-weights", + action="store_true", + help="if set, ties the weights of adaptive softmax and adaptive input", + ) + parser.add_argument( + "--tie-adaptive-proj", + action="store_true", + help="if set, ties the projection weights of adaptive softmax and adaptive input", + ) + parser.add_argument( + "--decoder-learned-pos", + action="store_true", + help="use learned positional embeddings in the decoder", + ) + + """LightConv and DynamicConv arguments""" + parser.add_argument( + "--decoder-kernel-size-list", + type=lambda x: utils.eval_str_list(x, int), + help='list of kernel size (default: "[3,7,15,31,31,31]")', + ) + parser.add_argument( + "--decoder-glu", type=utils.eval_bool, help="glu after in proj" + ) + parser.add_argument( + "--decoder-conv-type", + default="dynamic", + type=str, + choices=["dynamic", "lightweight"], + help="type of convolution", + ) + parser.add_argument("--weight-softmax", default=True, type=utils.eval_bool) + parser.add_argument( + "--weight-dropout", + type=float, + metavar="D", + help="dropout probability for conv weights", + ) + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + + # make sure all arguments are present in older models + base_lm_architecture(args) + + if getattr(args, "max_source_positions", None) is None: + args.max_source_positions = args.tokens_per_sample + if getattr(args, "max_target_positions", None) is None: + args.max_target_positions = args.tokens_per_sample + + if args.character_embeddings: + embed_tokens = CharacterTokenEmbedder( + task.dictionary, + eval(args.character_filters), + args.character_embedding_dim, + args.decoder_embed_dim, + args.char_embedder_highway_layers, + ) + elif args.adaptive_input: + embed_tokens = AdaptiveInput( + len(task.dictionary), + task.dictionary.pad(), + args.decoder_input_dim, + args.adaptive_input_factor, + args.decoder_embed_dim, + utils.eval_str_list(args.adaptive_input_cutoff, type=int), + ) + else: + embed_tokens = Embedding( + len(task.dictionary), args.decoder_input_dim, task.dictionary.pad() + ) + + if args.tie_adaptive_weights: + assert args.adaptive_input + assert args.adaptive_input_factor == args.adaptive_softmax_factor + assert ( + args.adaptive_softmax_cutoff == args.adaptive_input_cutoff + ), "{} != {}".format( + args.adaptive_softmax_cutoff, args.adaptive_input_cutoff + ) + assert args.decoder_input_dim == args.decoder_output_dim + + decoder = LightConvDecoder( + args, + task.output_dictionary, + embed_tokens, + no_encoder_attn=True, + final_norm=False, + ) + return LightConvLanguageModel(decoder) + + +@register_model_architecture("lightconv_lm", "lightconv_lm") +def base_lm_architecture(args): + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 512) + args.decoder_ffn_embed_dim = getattr(args, "decoder_ffn_embed_dim", 2048) + args.decoder_layers = getattr(args, "decoder_layers", 6) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 8) + args.adaptive_softmax_cutoff = getattr(args, "adaptive_softmax_cutoff", None) + args.adaptive_softmax_dropout = getattr(args, "adaptive_softmax_dropout", 0) + args.adaptive_softmax_factor = getattr(args, "adaptive_softmax_factor", 4) + args.decoder_learned_pos = getattr(args, "decoder_learned_pos", False) + + args.character_embeddings = getattr(args, "character_embeddings", False) + + args.decoder_output_dim = getattr( + args, "decoder_output_dim", args.decoder_embed_dim + ) + args.decoder_input_dim = getattr(args, "decoder_input_dim", args.decoder_embed_dim) + args.decoder_conv_dim = getattr(args, "decoder_conv_dim", args.decoder_embed_dim) + + # The model training is not stable without this + args.decoder_normalize_before = True + + args.adaptive_input = getattr(args, "adaptive_input", False) + args.adaptive_input_factor = getattr(args, "adaptive_input_factor", 4) + args.adaptive_input_cutoff = getattr(args, "adaptive_input_cutoff", None) + + args.tie_adaptive_weights = getattr(args, "tie_adaptive_weights", False) + args.tie_adaptive_proj = getattr(args, "tie_adaptive_proj", False) + + args.decoder_kernel_size_list = getattr( + args, "decoder_kernel_size_list", [3, 7, 15, 31, 31, 31] + ) + if len(args.decoder_kernel_size_list) == 1: + args.decoder_kernel_size_list = ( + args.decoder_kernel_size_list * args.decoder_layers + ) + assert ( + len(args.decoder_kernel_size_list) == args.decoder_layers + ), "decoder_kernel_size_list doesn't match decoder_layers" + args.decoder_glu = getattr(args, "decoder_glu", True) + args.input_dropout = getattr(args, "input_dropout", 0.1) + args.weight_dropout = getattr(args, "weight_dropout", args.attention_dropout) + + +@register_model_architecture("lightconv_lm", "lightconv_lm_gbw") +def lightconv_lm_gbw(args): + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 512) + args.dropout = getattr(args, "dropout", 0.1) + args.attention_dropout = getattr(args, "attention_dropout", 0.1) + args.decoder_ffn_embed_dim = getattr(args, "decoder_ffn_embed_dim", 4096) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 16) + base_lm_architecture(args) diff --git a/fairseq/models/lstm.py b/fairseq/models/lstm.py new file mode 100644 index 0000000000000000000000000000000000000000..12e3aff85dc02604a380546cd654719e4ab445f7 --- /dev/null +++ b/fairseq/models/lstm.py @@ -0,0 +1,753 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from typing import Dict, List, Optional, Tuple + +import torch +import torch.nn as nn +import torch.nn.functional as F +from fairseq import utils +from fairseq.models import ( + FairseqEncoder, + FairseqEncoderDecoderModel, + FairseqIncrementalDecoder, + register_model, + register_model_architecture, +) +from fairseq.modules import AdaptiveSoftmax, FairseqDropout +from torch import Tensor + + +DEFAULT_MAX_SOURCE_POSITIONS = 1e5 +DEFAULT_MAX_TARGET_POSITIONS = 1e5 + + +@register_model("lstm") +class LSTMModel(FairseqEncoderDecoderModel): + def __init__(self, encoder, decoder): + super().__init__(encoder, decoder) + + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + # fmt: off + parser.add_argument('--dropout', type=float, metavar='D', + help='dropout probability') + parser.add_argument('--encoder-embed-dim', type=int, metavar='N', + help='encoder embedding dimension') + parser.add_argument('--encoder-embed-path', type=str, metavar='STR', + help='path to pre-trained encoder embedding') + parser.add_argument('--encoder-freeze-embed', action='store_true', + help='freeze encoder embeddings') + parser.add_argument('--encoder-hidden-size', type=int, metavar='N', + help='encoder hidden size') + parser.add_argument('--encoder-layers', type=int, metavar='N', + help='number of encoder layers') + parser.add_argument('--encoder-bidirectional', action='store_true', + help='make all layers of encoder bidirectional') + parser.add_argument('--decoder-embed-dim', type=int, metavar='N', + help='decoder embedding dimension') + parser.add_argument('--decoder-embed-path', type=str, metavar='STR', + help='path to pre-trained decoder embedding') + parser.add_argument('--decoder-freeze-embed', action='store_true', + help='freeze decoder embeddings') + parser.add_argument('--decoder-hidden-size', type=int, metavar='N', + help='decoder hidden size') + parser.add_argument('--decoder-layers', type=int, metavar='N', + help='number of decoder layers') + parser.add_argument('--decoder-out-embed-dim', type=int, metavar='N', + help='decoder output embedding dimension') + parser.add_argument('--decoder-attention', type=str, metavar='BOOL', + help='decoder attention') + parser.add_argument('--adaptive-softmax-cutoff', metavar='EXPR', + help='comma separated list of adaptive softmax cutoff points. ' + 'Must be used with adaptive_loss criterion') + parser.add_argument('--share-decoder-input-output-embed', default=False, + action='store_true', + help='share decoder input and output embeddings') + parser.add_argument('--share-all-embeddings', default=False, action='store_true', + help='share encoder, decoder and output embeddings' + ' (requires shared dictionary and embed dim)') + + # Granular dropout settings (if not specified these default to --dropout) + parser.add_argument('--encoder-dropout-in', type=float, metavar='D', + help='dropout probability for encoder input embedding') + parser.add_argument('--encoder-dropout-out', type=float, metavar='D', + help='dropout probability for encoder output') + parser.add_argument('--decoder-dropout-in', type=float, metavar='D', + help='dropout probability for decoder input embedding') + parser.add_argument('--decoder-dropout-out', type=float, metavar='D', + help='dropout probability for decoder output') + # fmt: on + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + # make sure that all args are properly defaulted (in case there are any new ones) + base_architecture(args) + + if args.encoder_layers != args.decoder_layers: + raise ValueError("--encoder-layers must match --decoder-layers") + + max_source_positions = getattr( + args, "max_source_positions", DEFAULT_MAX_SOURCE_POSITIONS + ) + max_target_positions = getattr( + args, "max_target_positions", DEFAULT_MAX_TARGET_POSITIONS + ) + + def load_pretrained_embedding_from_file(embed_path, dictionary, embed_dim): + num_embeddings = len(dictionary) + padding_idx = dictionary.pad() + embed_tokens = Embedding(num_embeddings, embed_dim, padding_idx) + embed_dict = utils.parse_embedding(embed_path) + utils.print_embed_overlap(embed_dict, dictionary) + return utils.load_embedding(embed_dict, dictionary, embed_tokens) + + if args.encoder_embed_path: + pretrained_encoder_embed = load_pretrained_embedding_from_file( + args.encoder_embed_path, task.source_dictionary, args.encoder_embed_dim + ) + else: + num_embeddings = len(task.source_dictionary) + pretrained_encoder_embed = Embedding( + num_embeddings, args.encoder_embed_dim, task.source_dictionary.pad() + ) + + if args.share_all_embeddings: + # double check all parameters combinations are valid + if task.source_dictionary != task.target_dictionary: + raise ValueError("--share-all-embeddings requires a joint dictionary") + if args.decoder_embed_path and ( + args.decoder_embed_path != args.encoder_embed_path + ): + raise ValueError( + "--share-all-embed not compatible with --decoder-embed-path" + ) + if args.encoder_embed_dim != args.decoder_embed_dim: + raise ValueError( + "--share-all-embeddings requires --encoder-embed-dim to " + "match --decoder-embed-dim" + ) + pretrained_decoder_embed = pretrained_encoder_embed + args.share_decoder_input_output_embed = True + else: + # separate decoder input embeddings + pretrained_decoder_embed = None + if args.decoder_embed_path: + pretrained_decoder_embed = load_pretrained_embedding_from_file( + args.decoder_embed_path, + task.target_dictionary, + args.decoder_embed_dim, + ) + # one last double check of parameter combinations + if args.share_decoder_input_output_embed and ( + args.decoder_embed_dim != args.decoder_out_embed_dim + ): + raise ValueError( + "--share-decoder-input-output-embeddings requires " + "--decoder-embed-dim to match --decoder-out-embed-dim" + ) + + if args.encoder_freeze_embed: + pretrained_encoder_embed.weight.requires_grad = False + if args.decoder_freeze_embed: + pretrained_decoder_embed.weight.requires_grad = False + + encoder = LSTMEncoder( + dictionary=task.source_dictionary, + embed_dim=args.encoder_embed_dim, + hidden_size=args.encoder_hidden_size, + num_layers=args.encoder_layers, + dropout_in=args.encoder_dropout_in, + dropout_out=args.encoder_dropout_out, + bidirectional=args.encoder_bidirectional, + pretrained_embed=pretrained_encoder_embed, + max_source_positions=max_source_positions, + ) + decoder = LSTMDecoder( + dictionary=task.target_dictionary, + embed_dim=args.decoder_embed_dim, + hidden_size=args.decoder_hidden_size, + out_embed_dim=args.decoder_out_embed_dim, + num_layers=args.decoder_layers, + dropout_in=args.decoder_dropout_in, + dropout_out=args.decoder_dropout_out, + attention=utils.eval_bool(args.decoder_attention), + encoder_output_units=encoder.output_units, + pretrained_embed=pretrained_decoder_embed, + share_input_output_embed=args.share_decoder_input_output_embed, + adaptive_softmax_cutoff=( + utils.eval_str_list(args.adaptive_softmax_cutoff, type=int) + if args.criterion == "adaptive_loss" + else None + ), + max_target_positions=max_target_positions, + residuals=False, + ) + return cls(encoder, decoder) + + def forward( + self, + src_tokens, + src_lengths, + prev_output_tokens, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + ): + encoder_out = self.encoder(src_tokens, src_lengths=src_lengths) + decoder_out = self.decoder( + prev_output_tokens, + encoder_out=encoder_out, + incremental_state=incremental_state, + ) + return decoder_out + + +class LSTMEncoder(FairseqEncoder): + """LSTM encoder.""" + + def __init__( + self, + dictionary, + embed_dim=512, + hidden_size=512, + num_layers=1, + dropout_in=0.1, + dropout_out=0.1, + bidirectional=False, + left_pad=True, + pretrained_embed=None, + padding_idx=None, + max_source_positions=DEFAULT_MAX_SOURCE_POSITIONS, + ): + super().__init__(dictionary) + self.num_layers = num_layers + self.dropout_in_module = FairseqDropout( + dropout_in, module_name=self.__class__.__name__ + ) + self.dropout_out_module = FairseqDropout( + dropout_out, module_name=self.__class__.__name__ + ) + self.bidirectional = bidirectional + self.hidden_size = hidden_size + self.max_source_positions = max_source_positions + + num_embeddings = len(dictionary) + self.padding_idx = padding_idx if padding_idx is not None else dictionary.pad() + if pretrained_embed is None: + self.embed_tokens = Embedding(num_embeddings, embed_dim, self.padding_idx) + else: + self.embed_tokens = pretrained_embed + + self.lstm = LSTM( + input_size=embed_dim, + hidden_size=hidden_size, + num_layers=num_layers, + dropout=self.dropout_out_module.p if num_layers > 1 else 0.0, + bidirectional=bidirectional, + ) + self.left_pad = left_pad + + self.output_units = hidden_size + if bidirectional: + self.output_units *= 2 + + def forward( + self, + src_tokens: Tensor, + src_lengths: Tensor, + enforce_sorted: bool = True, + ): + """ + Args: + src_tokens (LongTensor): tokens in the source language of + shape `(batch, src_len)` + src_lengths (LongTensor): lengths of each source sentence of + shape `(batch)` + enforce_sorted (bool, optional): if True, `src_tokens` is + expected to contain sequences sorted by length in a + decreasing order. If False, this condition is not + required. Default: True. + """ + if self.left_pad: + # nn.utils.rnn.pack_padded_sequence requires right-padding; + # convert left-padding to right-padding + src_tokens = utils.convert_padding_direction( + src_tokens, + torch.zeros_like(src_tokens).fill_(self.padding_idx), + left_to_right=True, + ) + + bsz, seqlen = src_tokens.size() + + # embed tokens + x = self.embed_tokens(src_tokens) + x = self.dropout_in_module(x) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + + # pack embedded source tokens into a PackedSequence + packed_x = nn.utils.rnn.pack_padded_sequence( + x, src_lengths.cpu(), enforce_sorted=enforce_sorted + ) + + # apply LSTM + if self.bidirectional: + state_size = 2 * self.num_layers, bsz, self.hidden_size + else: + state_size = self.num_layers, bsz, self.hidden_size + h0 = x.new_zeros(*state_size) + c0 = x.new_zeros(*state_size) + packed_outs, (final_hiddens, final_cells) = self.lstm(packed_x, (h0, c0)) + + # unpack outputs and apply dropout + x, _ = nn.utils.rnn.pad_packed_sequence( + packed_outs, padding_value=self.padding_idx * 1.0 + ) + x = self.dropout_out_module(x) + assert list(x.size()) == [seqlen, bsz, self.output_units] + + if self.bidirectional: + final_hiddens = self.combine_bidir(final_hiddens, bsz) + final_cells = self.combine_bidir(final_cells, bsz) + + encoder_padding_mask = src_tokens.eq(self.padding_idx).t() + + return tuple( + ( + x, # seq_len x batch x hidden + final_hiddens, # num_layers x batch x num_directions*hidden + final_cells, # num_layers x batch x num_directions*hidden + encoder_padding_mask, # seq_len x batch + ) + ) + + def combine_bidir(self, outs, bsz: int): + out = outs.view(self.num_layers, 2, bsz, -1).transpose(1, 2).contiguous() + return out.view(self.num_layers, bsz, -1) + + def reorder_encoder_out(self, encoder_out, new_order): + return tuple( + ( + encoder_out[0].index_select(1, new_order), + encoder_out[1].index_select(1, new_order), + encoder_out[2].index_select(1, new_order), + encoder_out[3].index_select(1, new_order), + ) + ) + + def max_positions(self): + """Maximum input length supported by the encoder.""" + return self.max_source_positions + + +class AttentionLayer(nn.Module): + def __init__(self, input_embed_dim, source_embed_dim, output_embed_dim, bias=False): + super().__init__() + + self.input_proj = Linear(input_embed_dim, source_embed_dim, bias=bias) + self.output_proj = Linear( + input_embed_dim + source_embed_dim, output_embed_dim, bias=bias + ) + + def forward(self, input, source_hids, encoder_padding_mask): + # input: bsz x input_embed_dim + # source_hids: srclen x bsz x source_embed_dim + + # x: bsz x source_embed_dim + x = self.input_proj(input) + + # compute attention + attn_scores = (source_hids * x.unsqueeze(0)).sum(dim=2) + + # don't attend over padding + if encoder_padding_mask is not None: + attn_scores = ( + attn_scores.float() + .masked_fill_(encoder_padding_mask, float("-inf")) + .type_as(attn_scores) + ) # FP16 support: cast to float and back + + attn_scores = F.softmax(attn_scores, dim=0) # srclen x bsz + + # sum weighted sources + x = (attn_scores.unsqueeze(2) * source_hids).sum(dim=0) + + x = torch.tanh(self.output_proj(torch.cat((x, input), dim=1))) + return x, attn_scores + + +class LSTMDecoder(FairseqIncrementalDecoder): + """LSTM decoder.""" + + def __init__( + self, + dictionary, + embed_dim=512, + hidden_size=512, + out_embed_dim=512, + num_layers=1, + dropout_in=0.1, + dropout_out=0.1, + attention=True, + encoder_output_units=512, + pretrained_embed=None, + share_input_output_embed=False, + adaptive_softmax_cutoff=None, + max_target_positions=DEFAULT_MAX_TARGET_POSITIONS, + residuals=False, + ): + super().__init__(dictionary) + self.dropout_in_module = FairseqDropout( + dropout_in, module_name=self.__class__.__name__ + ) + self.dropout_out_module = FairseqDropout( + dropout_out, module_name=self.__class__.__name__ + ) + self.hidden_size = hidden_size + self.share_input_output_embed = share_input_output_embed + self.need_attn = True + self.max_target_positions = max_target_positions + self.residuals = residuals + self.num_layers = num_layers + + self.adaptive_softmax = None + num_embeddings = len(dictionary) + padding_idx = dictionary.pad() + if pretrained_embed is None: + self.embed_tokens = Embedding(num_embeddings, embed_dim, padding_idx) + else: + self.embed_tokens = pretrained_embed + + self.encoder_output_units = encoder_output_units + if encoder_output_units != hidden_size and encoder_output_units != 0: + self.encoder_hidden_proj = Linear(encoder_output_units, hidden_size) + self.encoder_cell_proj = Linear(encoder_output_units, hidden_size) + else: + self.encoder_hidden_proj = self.encoder_cell_proj = None + + # disable input feeding if there is no encoder + # input feeding is described in arxiv.org/abs/1508.04025 + input_feed_size = 0 if encoder_output_units == 0 else hidden_size + self.layers = nn.ModuleList( + [ + LSTMCell( + input_size=input_feed_size + embed_dim + if layer == 0 + else hidden_size, + hidden_size=hidden_size, + ) + for layer in range(num_layers) + ] + ) + + if attention: + # TODO make bias configurable + self.attention = AttentionLayer( + hidden_size, encoder_output_units, hidden_size, bias=False + ) + else: + self.attention = None + + if hidden_size != out_embed_dim: + self.additional_fc = Linear(hidden_size, out_embed_dim) + + if adaptive_softmax_cutoff is not None: + # setting adaptive_softmax dropout to dropout_out for now but can be redefined + self.adaptive_softmax = AdaptiveSoftmax( + num_embeddings, + hidden_size, + adaptive_softmax_cutoff, + dropout=dropout_out, + ) + elif not self.share_input_output_embed: + self.fc_out = Linear(out_embed_dim, num_embeddings, dropout=dropout_out) + + def forward( + self, + prev_output_tokens, + encoder_out: Optional[Tuple[Tensor, Tensor, Tensor, Tensor]] = None, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + src_lengths: Optional[Tensor] = None, + ): + x, attn_scores = self.extract_features( + prev_output_tokens, encoder_out, incremental_state + ) + return self.output_layer(x), attn_scores + + def extract_features( + self, + prev_output_tokens, + encoder_out: Optional[Tuple[Tensor, Tensor, Tensor, Tensor]] = None, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + ): + """ + Similar to *forward* but only return features. + """ + # get outputs from encoder + if encoder_out is not None: + encoder_outs = encoder_out[0] + encoder_hiddens = encoder_out[1] + encoder_cells = encoder_out[2] + encoder_padding_mask = encoder_out[3] + else: + encoder_outs = torch.empty(0) + encoder_hiddens = torch.empty(0) + encoder_cells = torch.empty(0) + encoder_padding_mask = torch.empty(0) + srclen = encoder_outs.size(0) + + if incremental_state is not None and len(incremental_state) > 0: + prev_output_tokens = prev_output_tokens[:, -1:] + + bsz, seqlen = prev_output_tokens.size() + + # embed tokens + x = self.embed_tokens(prev_output_tokens) + x = self.dropout_in_module(x) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + + # initialize previous states (or get from cache during incremental generation) + if incremental_state is not None and len(incremental_state) > 0: + prev_hiddens, prev_cells, input_feed = self.get_cached_state( + incremental_state + ) + elif encoder_out is not None: + # setup recurrent cells + prev_hiddens = [encoder_hiddens[i] for i in range(self.num_layers)] + prev_cells = [encoder_cells[i] for i in range(self.num_layers)] + if self.encoder_hidden_proj is not None: + prev_hiddens = [self.encoder_hidden_proj(y) for y in prev_hiddens] + prev_cells = [self.encoder_cell_proj(y) for y in prev_cells] + input_feed = x.new_zeros(bsz, self.hidden_size) + else: + # setup zero cells, since there is no encoder + zero_state = x.new_zeros(bsz, self.hidden_size) + prev_hiddens = [zero_state for i in range(self.num_layers)] + prev_cells = [zero_state for i in range(self.num_layers)] + input_feed = None + + assert ( + srclen > 0 or self.attention is None + ), "attention is not supported if there are no encoder outputs" + attn_scores: Optional[Tensor] = ( + x.new_zeros(srclen, seqlen, bsz) if self.attention is not None else None + ) + outs = [] + for j in range(seqlen): + # input feeding: concatenate context vector from previous time step + if input_feed is not None: + input = torch.cat((x[j, :, :], input_feed), dim=1) + else: + input = x[j] + + for i, rnn in enumerate(self.layers): + # recurrent cell + hidden, cell = rnn(input, (prev_hiddens[i], prev_cells[i])) + + # hidden state becomes the input to the next layer + input = self.dropout_out_module(hidden) + if self.residuals: + input = input + prev_hiddens[i] + + # save state for next time step + prev_hiddens[i] = hidden + prev_cells[i] = cell + + # apply attention using the last layer's hidden state + if self.attention is not None: + assert attn_scores is not None + out, attn_scores[:, j, :] = self.attention( + hidden, encoder_outs, encoder_padding_mask + ) + else: + out = hidden + out = self.dropout_out_module(out) + + # input feeding + if input_feed is not None: + input_feed = out + + # save final output + outs.append(out) + + # Stack all the necessary tensors together and store + prev_hiddens_tensor = torch.stack(prev_hiddens) + prev_cells_tensor = torch.stack(prev_cells) + cache_state = torch.jit.annotate( + Dict[str, Optional[Tensor]], + { + "prev_hiddens": prev_hiddens_tensor, + "prev_cells": prev_cells_tensor, + "input_feed": input_feed, + }, + ) + self.set_incremental_state(incremental_state, "cached_state", cache_state) + + # collect outputs across time steps + x = torch.cat(outs, dim=0).view(seqlen, bsz, self.hidden_size) + + # T x B x C -> B x T x C + x = x.transpose(1, 0) + + if hasattr(self, "additional_fc") and self.adaptive_softmax is None: + x = self.additional_fc(x) + x = self.dropout_out_module(x) + # srclen x tgtlen x bsz -> bsz x tgtlen x srclen + if not self.training and self.need_attn and self.attention is not None: + assert attn_scores is not None + attn_scores = attn_scores.transpose(0, 2) + else: + attn_scores = None + return x, attn_scores + + def output_layer(self, x): + """Project features to the vocabulary size.""" + if self.adaptive_softmax is None: + if self.share_input_output_embed: + x = F.linear(x, self.embed_tokens.weight) + else: + x = self.fc_out(x) + return x + + def get_cached_state( + self, + incremental_state: Dict[str, Dict[str, Optional[Tensor]]], + ) -> Tuple[List[Tensor], List[Tensor], Optional[Tensor]]: + cached_state = self.get_incremental_state(incremental_state, "cached_state") + assert cached_state is not None + prev_hiddens_ = cached_state["prev_hiddens"] + assert prev_hiddens_ is not None + prev_cells_ = cached_state["prev_cells"] + assert prev_cells_ is not None + prev_hiddens = [prev_hiddens_[i] for i in range(self.num_layers)] + prev_cells = [prev_cells_[j] for j in range(self.num_layers)] + input_feed = cached_state[ + "input_feed" + ] # can be None for decoder-only language models + return prev_hiddens, prev_cells, input_feed + + def reorder_incremental_state( + self, + incremental_state: Dict[str, Dict[str, Optional[Tensor]]], + new_order: Tensor, + ): + if incremental_state is None or len(incremental_state) == 0: + return + prev_hiddens, prev_cells, input_feed = self.get_cached_state(incremental_state) + prev_hiddens = [p.index_select(0, new_order) for p in prev_hiddens] + prev_cells = [p.index_select(0, new_order) for p in prev_cells] + if input_feed is not None: + input_feed = input_feed.index_select(0, new_order) + cached_state_new = torch.jit.annotate( + Dict[str, Optional[Tensor]], + { + "prev_hiddens": torch.stack(prev_hiddens), + "prev_cells": torch.stack(prev_cells), + "input_feed": input_feed, + }, + ) + self.set_incremental_state(incremental_state, "cached_state", cached_state_new), + return + + def max_positions(self): + """Maximum output length supported by the decoder.""" + return self.max_target_positions + + def make_generation_fast_(self, need_attn=False, **kwargs): + self.need_attn = need_attn + + +def Embedding(num_embeddings, embedding_dim, padding_idx): + m = nn.Embedding(num_embeddings, embedding_dim, padding_idx=padding_idx) + nn.init.uniform_(m.weight, -0.1, 0.1) + nn.init.constant_(m.weight[padding_idx], 0) + return m + + +def LSTM(input_size, hidden_size, **kwargs): + m = nn.LSTM(input_size, hidden_size, **kwargs) + for name, param in m.named_parameters(): + if "weight" in name or "bias" in name: + param.data.uniform_(-0.1, 0.1) + return m + + +def LSTMCell(input_size, hidden_size, **kwargs): + m = nn.LSTMCell(input_size, hidden_size, **kwargs) + for name, param in m.named_parameters(): + if "weight" in name or "bias" in name: + param.data.uniform_(-0.1, 0.1) + return m + + +def Linear(in_features, out_features, bias=True, dropout=0.0): + """Linear layer (input: N x T x C)""" + m = nn.Linear(in_features, out_features, bias=bias) + m.weight.data.uniform_(-0.1, 0.1) + if bias: + m.bias.data.uniform_(-0.1, 0.1) + return m + + +@register_model_architecture("lstm", "lstm") +def base_architecture(args): + args.dropout = getattr(args, "dropout", 0.1) + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512) + args.encoder_embed_path = getattr(args, "encoder_embed_path", None) + args.encoder_freeze_embed = getattr(args, "encoder_freeze_embed", False) + args.encoder_hidden_size = getattr( + args, "encoder_hidden_size", args.encoder_embed_dim + ) + args.encoder_layers = getattr(args, "encoder_layers", 1) + args.encoder_bidirectional = getattr(args, "encoder_bidirectional", False) + args.encoder_dropout_in = getattr(args, "encoder_dropout_in", args.dropout) + args.encoder_dropout_out = getattr(args, "encoder_dropout_out", args.dropout) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 512) + args.decoder_embed_path = getattr(args, "decoder_embed_path", None) + args.decoder_freeze_embed = getattr(args, "decoder_freeze_embed", False) + args.decoder_hidden_size = getattr( + args, "decoder_hidden_size", args.decoder_embed_dim + ) + args.decoder_layers = getattr(args, "decoder_layers", 1) + args.decoder_out_embed_dim = getattr(args, "decoder_out_embed_dim", 512) + args.decoder_attention = getattr(args, "decoder_attention", "1") + args.decoder_dropout_in = getattr(args, "decoder_dropout_in", args.dropout) + args.decoder_dropout_out = getattr(args, "decoder_dropout_out", args.dropout) + args.share_decoder_input_output_embed = getattr( + args, "share_decoder_input_output_embed", False + ) + args.share_all_embeddings = getattr(args, "share_all_embeddings", False) + args.adaptive_softmax_cutoff = getattr( + args, "adaptive_softmax_cutoff", "10000,50000,200000" + ) + + +@register_model_architecture("lstm", "lstm_wiseman_iwslt_de_en") +def lstm_wiseman_iwslt_de_en(args): + args.dropout = getattr(args, "dropout", 0.1) + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 256) + args.encoder_dropout_in = getattr(args, "encoder_dropout_in", 0) + args.encoder_dropout_out = getattr(args, "encoder_dropout_out", 0) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 256) + args.decoder_out_embed_dim = getattr(args, "decoder_out_embed_dim", 256) + args.decoder_dropout_in = getattr(args, "decoder_dropout_in", 0) + args.decoder_dropout_out = getattr(args, "decoder_dropout_out", args.dropout) + base_architecture(args) + + +@register_model_architecture("lstm", "lstm_luong_wmt_en_de") +def lstm_luong_wmt_en_de(args): + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 1000) + args.encoder_layers = getattr(args, "encoder_layers", 4) + args.encoder_dropout_out = getattr(args, "encoder_dropout_out", 0) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 1000) + args.decoder_layers = getattr(args, "decoder_layers", 4) + args.decoder_out_embed_dim = getattr(args, "decoder_out_embed_dim", 1000) + args.decoder_dropout_out = getattr(args, "decoder_dropout_out", 0) + base_architecture(args) diff --git a/fairseq/models/lstm_lm.py b/fairseq/models/lstm_lm.py new file mode 100644 index 0000000000000000000000000000000000000000..454f0ac36fab78bf02a8e2f07ed9607d1da87e34 --- /dev/null +++ b/fairseq/models/lstm_lm.py @@ -0,0 +1,142 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq import utils +from fairseq.models import ( + FairseqLanguageModel, + register_model, + register_model_architecture, +) +from fairseq.models.lstm import Embedding, LSTMDecoder + + +DEFAULT_MAX_TARGET_POSITIONS = 1e5 + + +@register_model("lstm_lm") +class LSTMLanguageModel(FairseqLanguageModel): + def __init__(self, decoder): + super().__init__(decoder) + + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + # fmt: off + parser.add_argument('--dropout', type=float, metavar='D', + help='dropout probability') + parser.add_argument('--decoder-embed-dim', type=int, metavar='N', + help='decoder embedding dimension') + parser.add_argument('--decoder-embed-path', type=str, metavar='STR', + help='path to pre-trained decoder embedding') + parser.add_argument('--decoder-hidden-size', type=int, metavar='N', + help='decoder hidden size') + parser.add_argument('--decoder-layers', type=int, metavar='N', + help='number of decoder layers') + parser.add_argument('--decoder-out-embed-dim', type=int, metavar='N', + help='decoder output embedding dimension') + parser.add_argument('--decoder-attention', type=str, metavar='BOOL', + help='decoder attention') + parser.add_argument('--adaptive-softmax-cutoff', metavar='EXPR', + help='comma separated list of adaptive softmax cutoff points. ' + 'Must be used with adaptive_loss criterion') + parser.add_argument('--residuals', default=False, + action='store_true', + help='applying residuals between LSTM layers') + + # Granular dropout settings (if not specified these default to --dropout) + parser.add_argument('--decoder-dropout-in', type=float, metavar='D', + help='dropout probability for decoder input embedding') + parser.add_argument('--decoder-dropout-out', type=float, metavar='D', + help='dropout probability for decoder output') + parser.add_argument('--share-decoder-input-output-embed', default=False, + action='store_true', + help='share decoder input and output embeddings') + # fmt: on + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + + # make sure all arguments are present in older models + base_architecture(args) + + if getattr(args, "max_target_positions", None) is not None: + max_target_positions = args.max_target_positions + else: + max_target_positions = getattr( + args, "tokens_per_sample", DEFAULT_MAX_TARGET_POSITIONS + ) + + def load_pretrained_embedding_from_file(embed_path, dictionary, embed_dim): + num_embeddings = len(dictionary) + padding_idx = dictionary.pad() + embed_tokens = Embedding(num_embeddings, embed_dim, padding_idx) + embed_dict = utils.parse_embedding(embed_path) + utils.print_embed_overlap(embed_dict, dictionary) + return utils.load_embedding(embed_dict, dictionary, embed_tokens) + + pretrained_decoder_embed = None + if args.decoder_embed_path: + pretrained_decoder_embed = load_pretrained_embedding_from_file( + args.decoder_embed_path, task.target_dictionary, args.decoder_embed_dim + ) + + if args.share_decoder_input_output_embed: + # double check all parameters combinations are valid + if task.source_dictionary != task.target_dictionary: + raise ValueError( + "--share-decoder-input-output-embeddings requires a joint dictionary" + ) + + if args.decoder_embed_dim != args.decoder_out_embed_dim: + raise ValueError( + "--share-decoder-input-output-embeddings requires " + "--decoder-embed-dim to match --decoder-out-embed-dim" + ) + + decoder = LSTMDecoder( + dictionary=task.dictionary, + embed_dim=args.decoder_embed_dim, + hidden_size=args.decoder_hidden_size, + out_embed_dim=args.decoder_out_embed_dim, + num_layers=args.decoder_layers, + dropout_in=args.decoder_dropout_in, + dropout_out=args.decoder_dropout_out, + attention=False, # decoder-only language model doesn't support attention + encoder_output_units=0, + pretrained_embed=pretrained_decoder_embed, + share_input_output_embed=args.share_decoder_input_output_embed, + adaptive_softmax_cutoff=( + utils.eval_str_list(args.adaptive_softmax_cutoff, type=int) + if args.criterion == "adaptive_loss" + else None + ), + max_target_positions=max_target_positions, + residuals=args.residuals, + ) + + return cls(decoder) + + +@register_model_architecture("lstm_lm", "lstm_lm") +def base_architecture(args): + args.dropout = getattr(args, "dropout", 0.1) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 512) + args.decoder_embed_path = getattr(args, "decoder_embed_path", None) + args.decoder_hidden_size = getattr( + args, "decoder_hidden_size", args.decoder_embed_dim + ) + args.decoder_layers = getattr(args, "decoder_layers", 1) + args.decoder_out_embed_dim = getattr(args, "decoder_out_embed_dim", 512) + args.decoder_attention = getattr(args, "decoder_attention", "0") + args.decoder_dropout_in = getattr(args, "decoder_dropout_in", args.dropout) + args.decoder_dropout_out = getattr(args, "decoder_dropout_out", args.dropout) + args.share_decoder_input_output_embed = getattr( + args, "share_decoder_input_output_embed", False + ) + args.adaptive_softmax_cutoff = getattr( + args, "adaptive_softmax_cutoff", "10000,50000,200000" + ) + args.residuals = getattr(args, "residuals", False) diff --git a/fairseq/models/masked_lm.py b/fairseq/models/masked_lm.py new file mode 100644 index 0000000000000000000000000000000000000000..c786de9125551f7247618b0a1d0867477894c755 --- /dev/null +++ b/fairseq/models/masked_lm.py @@ -0,0 +1,403 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging + +import torch +import torch.nn as nn +import torch.nn.functional as F +from fairseq import utils +from fairseq.models import ( + FairseqEncoder, + FairseqEncoderModel, + register_model, + register_model_architecture, +) +from fairseq.modules import ( + LayerNorm, + SinusoidalPositionalEmbedding, + TransformerSentenceEncoder, +) +from fairseq.modules.transformer_sentence_encoder import init_bert_params + + +logger = logging.getLogger(__name__) + + +@register_model("masked_lm") +class MaskedLMModel(FairseqEncoderModel): + """ + Class for training a Masked Language Model. It also supports an + additional sentence level prediction if the sent-loss argument is set. + """ + + def __init__(self, args, encoder): + super().__init__(encoder) + self.args = args + + # if specified then apply bert initialization on the model. We need + # to explictly call this to make sure that the output embeddings + # and projection layers are also correctly initialized + if getattr(args, "apply_bert_init", False): + self.apply(init_bert_params) + + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + # Arguments related to dropout + parser.add_argument( + "--dropout", type=float, metavar="D", help="dropout probability" + ) + parser.add_argument( + "--attention-dropout", + type=float, + metavar="D", + help="dropout probability for" " attention weights", + ) + parser.add_argument( + "--act-dropout", + type=float, + metavar="D", + help="dropout probability after" " activation in FFN", + ) + + # Arguments related to hidden states and self-attention + parser.add_argument( + "--encoder-ffn-embed-dim", + type=int, + metavar="N", + help="encoder embedding dimension for FFN", + ) + parser.add_argument( + "--encoder-layers", type=int, metavar="N", help="num encoder layers" + ) + parser.add_argument( + "--encoder-attention-heads", + type=int, + metavar="N", + help="num encoder attention heads", + ) + + # Arguments related to input and output embeddings + parser.add_argument( + "--encoder-embed-dim", + type=int, + metavar="N", + help="encoder embedding dimension", + ) + parser.add_argument( + "--share-encoder-input-output-embed", + action="store_true", + help="share encoder input" " and output embeddings", + ) + parser.add_argument( + "--encoder-learned-pos", + action="store_true", + help="use learned positional embeddings in the encoder", + ) + parser.add_argument( + "--no-token-positional-embeddings", + action="store_true", + help="if set, disables positional embeddings" " (outside self attention)", + ) + parser.add_argument( + "--num-segment", type=int, metavar="N", help="num segment in the input" + ) + parser.add_argument( + "--max-positions", type=int, help="number of positional embeddings to learn" + ) + + # Arguments related to sentence level prediction + parser.add_argument( + "--sentence-class-num", + type=int, + metavar="N", + help="number of classes for sentence task", + ) + parser.add_argument( + "--sent-loss", + action="store_true", + help="if set," " calculate sentence level predictions", + ) + + # Arguments related to parameter initialization + parser.add_argument( + "--apply-bert-init", + action="store_true", + help="use custom param initialization for BERT", + ) + + # misc params + parser.add_argument( + "--activation-fn", + choices=utils.get_available_activation_fns(), + help="activation function to use", + ) + parser.add_argument( + "--pooler-activation-fn", + choices=utils.get_available_activation_fns(), + help="Which activation function to use for pooler layer.", + ) + parser.add_argument( + "--encoder-normalize-before", + action="store_true", + help="apply layernorm before each encoder block", + ) + + def forward(self, src_tokens, segment_labels=None, **kwargs): + return self.encoder(src_tokens, segment_labels=segment_labels, **kwargs) + + def max_positions(self): + return self.encoder.max_positions + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + # make sure all arguments are present in older models + base_architecture(args) + + if not hasattr(args, "max_positions"): + args.max_positions = args.tokens_per_sample + + logger.info(args) + + encoder = MaskedLMEncoder(args, task.dictionary) + return cls(args, encoder) + + +class MaskedLMEncoder(FairseqEncoder): + """ + Encoder for Masked Language Modelling. + """ + + def __init__(self, args, dictionary): + super().__init__(dictionary) + + self.padding_idx = dictionary.pad() + self.vocab_size = dictionary.__len__() + self.max_positions = args.max_positions + + self.sentence_encoder = TransformerSentenceEncoder( + padding_idx=self.padding_idx, + vocab_size=self.vocab_size, + num_encoder_layers=args.encoder_layers, + embedding_dim=args.encoder_embed_dim, + ffn_embedding_dim=args.encoder_ffn_embed_dim, + num_attention_heads=args.encoder_attention_heads, + dropout=args.dropout, + attention_dropout=args.attention_dropout, + activation_dropout=args.act_dropout, + max_seq_len=self.max_positions, + num_segments=args.num_segment, + use_position_embeddings=not args.no_token_positional_embeddings, + encoder_normalize_before=args.encoder_normalize_before, + apply_bert_init=args.apply_bert_init, + activation_fn=args.activation_fn, + learned_pos_embedding=args.encoder_learned_pos, + ) + + self.share_input_output_embed = args.share_encoder_input_output_embed + self.embed_out = None + self.sentence_projection_layer = None + self.sentence_out_dim = args.sentence_class_num + self.lm_output_learned_bias = None + + # Remove head is set to true during fine-tuning + self.load_softmax = not getattr(args, "remove_head", False) + + self.masked_lm_pooler = nn.Linear( + args.encoder_embed_dim, args.encoder_embed_dim + ) + self.pooler_activation = utils.get_activation_fn(args.pooler_activation_fn) + + self.lm_head_transform_weight = nn.Linear( + args.encoder_embed_dim, args.encoder_embed_dim + ) + self.activation_fn = utils.get_activation_fn(args.activation_fn) + self.layer_norm = LayerNorm(args.encoder_embed_dim) + + self.lm_output_learned_bias = None + if self.load_softmax: + self.lm_output_learned_bias = nn.Parameter(torch.zeros(self.vocab_size)) + + if not self.share_input_output_embed: + self.embed_out = nn.Linear( + args.encoder_embed_dim, self.vocab_size, bias=False + ) + + if args.sent_loss: + self.sentence_projection_layer = nn.Linear( + args.encoder_embed_dim, self.sentence_out_dim, bias=False + ) + + def forward(self, src_tokens, segment_labels=None, masked_tokens=None, **unused): + """ + Forward pass for Masked LM encoder. This first computes the token + embedding using the token embedding matrix, position embeddings (if + specified) and segment embeddings (if specified). + + Here we assume that the sentence representation corresponds to the + output of the classification_token (see bert_task or cross_lingual_lm + task for more details). + Args: + - src_tokens: B x T matrix representing sentences + - segment_labels: B x T matrix representing segment label for tokens + Returns: + - a tuple of the following: + - logits for predictions in format B x T x C to be used in + softmax afterwards + - a dictionary of additional data, where 'pooled_output' contains + the representation for classification_token and 'inner_states' + is a list of internal model states used to compute the + predictions (similar in ELMO). 'sentence_logits' + is the prediction logit for NSP task and is only computed if + this is specified in the input arguments. + """ + + inner_states, sentence_rep = self.sentence_encoder( + src_tokens, + segment_labels=segment_labels, + ) + + x = inner_states[-1].transpose(0, 1) + # project masked tokens only + if masked_tokens is not None: + x = x[masked_tokens, :] + x = self.layer_norm(self.activation_fn(self.lm_head_transform_weight(x))) + + pooled_output = self.pooler_activation(self.masked_lm_pooler(sentence_rep)) + + # project back to size of vocabulary + if self.share_input_output_embed and hasattr( + self.sentence_encoder.embed_tokens, "weight" + ): + x = F.linear(x, self.sentence_encoder.embed_tokens.weight) + elif self.embed_out is not None: + x = self.embed_out(x) + if self.lm_output_learned_bias is not None: + x = x + self.lm_output_learned_bias + sentence_logits = None + if self.sentence_projection_layer: + sentence_logits = self.sentence_projection_layer(pooled_output) + + return x, { + "inner_states": inner_states, + "pooled_output": pooled_output, + "sentence_logits": sentence_logits, + } + + def max_positions(self): + """Maximum output length supported by the encoder.""" + return self.max_positions + + def upgrade_state_dict_named(self, state_dict, name): + if isinstance( + self.sentence_encoder.embed_positions, SinusoidalPositionalEmbedding + ): + state_dict[ + name + ".sentence_encoder.embed_positions._float_tensor" + ] = torch.FloatTensor(1) + if not self.load_softmax: + for k in list(state_dict.keys()): + if ( + "embed_out.weight" in k + or "sentence_projection_layer.weight" in k + or "lm_output_learned_bias" in k + ): + del state_dict[k] + return state_dict + + +@register_model_architecture("masked_lm", "masked_lm") +def base_architecture(args): + args.dropout = getattr(args, "dropout", 0.1) + args.attention_dropout = getattr(args, "attention_dropout", 0.1) + args.act_dropout = getattr(args, "act_dropout", 0.0) + + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 4096) + args.encoder_layers = getattr(args, "encoder_layers", 6) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 8) + + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 1024) + args.share_encoder_input_output_embed = getattr( + args, "share_encoder_input_output_embed", False + ) + args.encoder_learned_pos = getattr(args, "encoder_learned_pos", False) + args.no_token_positional_embeddings = getattr( + args, "no_token_positional_embeddings", False + ) + args.num_segment = getattr(args, "num_segment", 2) + + args.sentence_class_num = getattr(args, "sentence_class_num", 2) + args.sent_loss = getattr(args, "sent_loss", False) + + args.apply_bert_init = getattr(args, "apply_bert_init", False) + + args.activation_fn = getattr(args, "activation_fn", "relu") + args.pooler_activation_fn = getattr(args, "pooler_activation_fn", "tanh") + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False) + + +@register_model_architecture("masked_lm", "bert_base") +def bert_base_architecture(args): + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 768) + args.share_encoder_input_output_embed = getattr( + args, "share_encoder_input_output_embed", True + ) + args.no_token_positional_embeddings = getattr( + args, "no_token_positional_embeddings", False + ) + args.encoder_learned_pos = getattr(args, "encoder_learned_pos", True) + args.num_segment = getattr(args, "num_segment", 2) + + args.encoder_layers = getattr(args, "encoder_layers", 12) + + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 12) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 3072) + + args.sentence_class_num = getattr(args, "sentence_class_num", 2) + args.sent_loss = getattr(args, "sent_loss", True) + + args.apply_bert_init = getattr(args, "apply_bert_init", True) + + args.activation_fn = getattr(args, "activation_fn", "gelu") + args.pooler_activation_fn = getattr(args, "pooler_activation_fn", "tanh") + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", True) + base_architecture(args) + + +@register_model_architecture("masked_lm", "bert_large") +def bert_large_architecture(args): + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 1024) + args.encoder_layers = getattr(args, "encoder_layers", 24) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 16) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 4096) + bert_base_architecture(args) + + +@register_model_architecture("masked_lm", "xlm_base") +def xlm_architecture(args): + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 1024) + args.share_encoder_input_output_embed = getattr( + args, "share_encoder_input_output_embed", True + ) + args.no_token_positional_embeddings = getattr( + args, "no_token_positional_embeddings", False + ) + args.encoder_learned_pos = getattr(args, "encoder_learned_pos", True) + args.num_segment = getattr(args, "num_segment", 1) + + args.encoder_layers = getattr(args, "encoder_layers", 6) + + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 8) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 4096) + + args.sent_loss = getattr(args, "sent_loss", False) + + args.activation_fn = getattr(args, "activation_fn", "gelu") + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False) + args.pooler_activation_fn = getattr(args, "pooler_activation_fn", "tanh") + args.apply_bert_init = getattr(args, "apply_bert_init", True) + base_architecture(args) diff --git a/fairseq/models/model_utils.py b/fairseq/models/model_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..732d66b1d5f695151c26d29eb7f6b53179c269f1 --- /dev/null +++ b/fairseq/models/model_utils.py @@ -0,0 +1,92 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from typing import List, Optional + +import torch +from torch import Tensor + + +@torch.jit.script +def script_skip_tensor_list(x: List[Tensor], mask): + res = [xi[mask] if xi.size(0) == mask.size(0) else xi[:, mask] for xi in x] + outputs = [] + for i, t in enumerate(res): + if t.numel() != 0: + outputs.append(t) + else: + outputs.append(x[i]) + return outputs + + +@torch.jit.script +def script_skip_tensor(x: Tensor, mask): + # None case + if x.size(0) == 0: + return x + res = x[mask] if x.size(0) == mask.size(0) else x[:, mask] + if res.numel() == 0: + return x + else: + return res + + +@torch.jit.script +def expand_2d_or_3d_tensor(x, trg_dim: int, padding_idx: int): + """ + Expand 2D/3D tensor on dim=1 + """ + if x is None: + return None + + assert x.dim() == 2 or x.dim() == 3 + assert trg_dim >= x.size(1), (trg_dim, x.size()) + if trg_dim == x.size(1): + return x + + dims = [x.size(0), trg_dim - x.size(1)] + if x.dim() == 3: + dims.append(x.size(2)) + x = torch.cat([x, torch.zeros(dims).to(x).fill_(padding_idx)], 1) + + return x + + +@torch.jit.script +def coalesce(x: Optional[Tensor], y: Tensor) -> Tensor: + return x if x is not None else y + + +@torch.jit.script +def fill_tensors( + x: Optional[Tensor], mask, y: Optional[Tensor], padding_idx: int +) -> Optional[Tensor]: + """ + Filling tensor x with y at masked positions (dim=0). + """ + if x is None or x.size()[0] == 0 or y is None: + return x + assert x.dim() == y.dim() and mask.size(0) == x.size(0) + assert x.dim() == 2 or (x.dim() == 3 and x.size(2) == y.size(2)) + + n_selected = mask.sum() + if n_selected == 0: + return x + assert n_selected == y.size(0) + if n_selected == x.size(0): + return y + + if x.size(1) < y.size(1): + x = expand_2d_or_3d_tensor(x, y.size(1), padding_idx) + x[mask] = y + elif x.size(1) > y.size(1): + x[mask] = torch.tensor(padding_idx).type_as(x) + if x.dim() == 2: + x[mask, : y.size(1)] = y + else: + x[mask, : y.size(1), :] = y + else: + x[mask] = y + return x diff --git a/fairseq/models/multilingual_transformer.py b/fairseq/models/multilingual_transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..2e1f86f36e01a2dd105c13f2e69b0eb25caa9fca --- /dev/null +++ b/fairseq/models/multilingual_transformer.py @@ -0,0 +1,228 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from collections import OrderedDict + +from fairseq import utils +from fairseq.models import ( + FairseqMultiModel, + register_model, + register_model_architecture, +) +from fairseq.models.transformer import ( + Embedding, + TransformerDecoder, + TransformerEncoder, + TransformerModel, + base_architecture, +) + + +@register_model("multilingual_transformer") +class MultilingualTransformerModel(FairseqMultiModel): + """Train Transformer models for multiple language pairs simultaneously. + + Requires `--task multilingual_translation`. + + We inherit all arguments from TransformerModel and assume that all language + pairs use a single Transformer architecture. In addition, we provide several + options that are specific to the multilingual setting. + + Args: + --share-encoder-embeddings: share encoder embeddings across all source languages + --share-decoder-embeddings: share decoder embeddings across all target languages + --share-encoders: share all encoder params (incl. embeddings) across all source languages + --share-decoders: share all decoder params (incl. embeddings) across all target languages + """ + + def __init__(self, encoders, decoders): + super().__init__(encoders, decoders) + + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + TransformerModel.add_args(parser) + parser.add_argument( + "--share-encoder-embeddings", + action="store_true", + help="share encoder embeddings across languages", + ) + parser.add_argument( + "--share-decoder-embeddings", + action="store_true", + help="share decoder embeddings across languages", + ) + parser.add_argument( + "--share-encoders", + action="store_true", + help="share encoders across languages", + ) + parser.add_argument( + "--share-decoders", + action="store_true", + help="share decoders across languages", + ) + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + from fairseq.tasks.multilingual_translation import MultilingualTranslationTask + + assert isinstance(task, MultilingualTranslationTask) + + # make sure all arguments are present in older models + base_multilingual_architecture(args) + + if not hasattr(args, "max_source_positions"): + args.max_source_positions = 1024 + if not hasattr(args, "max_target_positions"): + args.max_target_positions = 1024 + + src_langs = [lang_pair.split("-")[0] for lang_pair in task.model_lang_pairs] + tgt_langs = [lang_pair.split("-")[1] for lang_pair in task.model_lang_pairs] + + if args.share_encoders: + args.share_encoder_embeddings = True + if args.share_decoders: + args.share_decoder_embeddings = True + + def build_embedding(dictionary, embed_dim, path=None): + num_embeddings = len(dictionary) + padding_idx = dictionary.pad() + emb = Embedding(num_embeddings, embed_dim, padding_idx) + # if provided, load from preloaded dictionaries + if path: + embed_dict = utils.parse_embedding(path) + utils.load_embedding(embed_dict, dictionary, emb) + return emb + + # build shared embeddings (if applicable) + shared_encoder_embed_tokens, shared_decoder_embed_tokens = None, None + if args.share_all_embeddings: + if args.encoder_embed_dim != args.decoder_embed_dim: + raise ValueError( + "--share-all-embeddings requires --encoder-embed-dim to match --decoder-embed-dim" + ) + if args.decoder_embed_path and ( + args.decoder_embed_path != args.encoder_embed_path + ): + raise ValueError( + "--share-all-embeddings not compatible with --decoder-embed-path" + ) + shared_encoder_embed_tokens = FairseqMultiModel.build_shared_embeddings( + dicts=task.dicts, + langs=task.langs, + embed_dim=args.encoder_embed_dim, + build_embedding=build_embedding, + pretrained_embed_path=args.encoder_embed_path, + ) + shared_decoder_embed_tokens = shared_encoder_embed_tokens + args.share_decoder_input_output_embed = True + else: + if args.share_encoder_embeddings: + shared_encoder_embed_tokens = FairseqMultiModel.build_shared_embeddings( + dicts=task.dicts, + langs=src_langs, + embed_dim=args.encoder_embed_dim, + build_embedding=build_embedding, + pretrained_embed_path=args.encoder_embed_path, + ) + if args.share_decoder_embeddings: + shared_decoder_embed_tokens = FairseqMultiModel.build_shared_embeddings( + dicts=task.dicts, + langs=tgt_langs, + embed_dim=args.decoder_embed_dim, + build_embedding=build_embedding, + pretrained_embed_path=args.decoder_embed_path, + ) + + # encoders/decoders for each language + lang_encoders, lang_decoders = {}, {} + + def get_encoder(lang): + if lang not in lang_encoders: + if shared_encoder_embed_tokens is not None: + encoder_embed_tokens = shared_encoder_embed_tokens + else: + encoder_embed_tokens = build_embedding( + task.dicts[lang], + args.encoder_embed_dim, + args.encoder_embed_path, + ) + lang_encoders[lang] = cls._get_module_class( + True, args, task.dicts[lang], encoder_embed_tokens, src_langs + ) + return lang_encoders[lang] + + def get_decoder(lang): + if lang not in lang_decoders: + if shared_decoder_embed_tokens is not None: + decoder_embed_tokens = shared_decoder_embed_tokens + else: + decoder_embed_tokens = build_embedding( + task.dicts[lang], + args.decoder_embed_dim, + args.decoder_embed_path, + ) + lang_decoders[lang] = cls._get_module_class( + False, args, task.dicts[lang], decoder_embed_tokens, tgt_langs + ) + return lang_decoders[lang] + + # shared encoders/decoders (if applicable) + shared_encoder, shared_decoder = None, None + if args.share_encoders: + shared_encoder = get_encoder(src_langs[0]) + if args.share_decoders: + shared_decoder = get_decoder(tgt_langs[0]) + + encoders, decoders = OrderedDict(), OrderedDict() + for lang_pair, src, tgt in zip(task.model_lang_pairs, src_langs, tgt_langs): + encoders[lang_pair] = ( + shared_encoder if shared_encoder is not None else get_encoder(src) + ) + decoders[lang_pair] = ( + shared_decoder if shared_decoder is not None else get_decoder(tgt) + ) + + return MultilingualTransformerModel(encoders, decoders) + + @classmethod + def _get_module_class(cls, is_encoder, args, lang_dict, embed_tokens, langs): + module_class = TransformerEncoder if is_encoder else TransformerDecoder + return module_class(args, lang_dict, embed_tokens) + + def load_state_dict(self, state_dict, strict=True, model_cfg=None): + state_dict_subset = state_dict.copy() + for k, _ in state_dict.items(): + assert k.startswith("models.") + lang_pair = k.split(".")[1] + if lang_pair not in self.models: + del state_dict_subset[k] + super().load_state_dict(state_dict_subset, strict=strict, model_cfg=model_cfg) + + +@register_model_architecture("multilingual_transformer", "multilingual_transformer") +def base_multilingual_architecture(args): + base_architecture(args) + args.share_encoder_embeddings = getattr(args, "share_encoder_embeddings", False) + args.share_decoder_embeddings = getattr(args, "share_decoder_embeddings", False) + args.share_encoders = getattr(args, "share_encoders", False) + args.share_decoders = getattr(args, "share_decoders", False) + + +@register_model_architecture( + "multilingual_transformer", "multilingual_transformer_iwslt_de_en" +) +def multilingual_transformer_iwslt_de_en(args): + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 1024) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 4) + args.encoder_layers = getattr(args, "encoder_layers", 6) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 512) + args.decoder_ffn_embed_dim = getattr(args, "decoder_ffn_embed_dim", 1024) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 4) + args.decoder_layers = getattr(args, "decoder_layers", 6) + base_multilingual_architecture(args) diff --git a/fairseq/models/nat/__init__.py b/fairseq/models/nat/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..05fe822487c3bcde8346648d5826f1669c6bc1ca --- /dev/null +++ b/fairseq/models/nat/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +"""isort:skip_file""" + +from .fairseq_nat_model import * +from .nonautoregressive_transformer import * +from .nat_crf_transformer import * +from .iterative_nonautoregressive_transformer import * +from .cmlm_transformer import * +from .levenshtein_transformer import * +from .insertion_transformer import * diff --git a/fairseq/models/nat/cmlm_transformer.py b/fairseq/models/nat/cmlm_transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..c876e9453c101c00bd8e93e6e6f1fb48dc26f993 --- /dev/null +++ b/fairseq/models/nat/cmlm_transformer.py @@ -0,0 +1,162 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +""" +This file implements: +Ghazvininejad, Marjan, et al. +"Constant-time machine translation with conditional masked language models." +arXiv preprint arXiv:1904.09324 (2019). +""" + +from fairseq.models import register_model, register_model_architecture +from fairseq.models.nat import NATransformerModel +from fairseq.utils import new_arange + + +def _skeptical_unmasking(output_scores, output_masks, p): + sorted_index = output_scores.sort(-1)[1] + boundary_len = ( + (output_masks.sum(1, keepdim=True).type_as(output_scores) - 2) * p + ).long() + skeptical_mask = new_arange(output_masks) < boundary_len + return skeptical_mask.scatter(1, sorted_index, skeptical_mask) + + +@register_model("cmlm_transformer") +class CMLMNATransformerModel(NATransformerModel): + @staticmethod + def add_args(parser): + NATransformerModel.add_args(parser) + + def forward( + self, src_tokens, src_lengths, prev_output_tokens, tgt_tokens, **kwargs + ): + assert not self.decoder.src_embedding_copy, "do not support embedding copy." + + # encoding + encoder_out = self.encoder(src_tokens, src_lengths=src_lengths, **kwargs) + # length prediction + length_out = self.decoder.forward_length( + normalize=False, encoder_out=encoder_out + ) + length_tgt = self.decoder.forward_length_prediction( + length_out, encoder_out, tgt_tokens + ) + + # decoding + word_ins_out = self.decoder( + normalize=False, + prev_output_tokens=prev_output_tokens, + encoder_out=encoder_out, + ) + word_ins_mask = prev_output_tokens.eq(self.unk) + + return { + "word_ins": { + "out": word_ins_out, + "tgt": tgt_tokens, + "mask": word_ins_mask, + "ls": self.args.label_smoothing, + "nll_loss": True, + }, + "length": { + "out": length_out, + "tgt": length_tgt, + "factor": self.decoder.length_loss_factor, + }, + } + + def forward_decoder(self, decoder_out, encoder_out, decoding_format=None, **kwargs): + + step = decoder_out.step + max_step = decoder_out.max_step + + output_tokens = decoder_out.output_tokens + output_scores = decoder_out.output_scores + history = decoder_out.history + + # execute the decoder + output_masks = output_tokens.eq(self.unk) + _scores, _tokens = self.decoder( + normalize=True, + prev_output_tokens=output_tokens, + encoder_out=encoder_out, + ).max(-1) + output_tokens.masked_scatter_(output_masks, _tokens[output_masks]) + output_scores.masked_scatter_(output_masks, _scores[output_masks]) + + if history is not None: + history.append(output_tokens.clone()) + + # skeptical decoding (depend on the maximum decoding steps.) + if (step + 1) < max_step: + skeptical_mask = _skeptical_unmasking( + output_scores, output_tokens.ne(self.pad), 1 - (step + 1) / max_step + ) + + output_tokens.masked_fill_(skeptical_mask, self.unk) + output_scores.masked_fill_(skeptical_mask, 0.0) + + if history is not None: + history.append(output_tokens.clone()) + + return decoder_out._replace( + output_tokens=output_tokens, + output_scores=output_scores, + attn=None, + history=history, + ) + + +@register_model_architecture("cmlm_transformer", "cmlm_transformer") +def cmlm_base_architecture(args): + args.encoder_embed_path = getattr(args, "encoder_embed_path", None) + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 2048) + args.encoder_layers = getattr(args, "encoder_layers", 6) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 8) + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False) + args.encoder_learned_pos = getattr(args, "encoder_learned_pos", False) + args.decoder_embed_path = getattr(args, "decoder_embed_path", None) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", args.encoder_embed_dim) + args.decoder_ffn_embed_dim = getattr( + args, "decoder_ffn_embed_dim", args.encoder_ffn_embed_dim + ) + args.decoder_layers = getattr(args, "decoder_layers", 6) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 8) + args.decoder_normalize_before = getattr(args, "decoder_normalize_before", False) + args.decoder_learned_pos = getattr(args, "decoder_learned_pos", False) + args.attention_dropout = getattr(args, "attention_dropout", 0.0) + args.activation_dropout = getattr(args, "activation_dropout", 0.0) + args.activation_fn = getattr(args, "activation_fn", "relu") + args.dropout = getattr(args, "dropout", 0.1) + args.adaptive_softmax_cutoff = getattr(args, "adaptive_softmax_cutoff", None) + args.adaptive_softmax_dropout = getattr(args, "adaptive_softmax_dropout", 0) + args.share_decoder_input_output_embed = getattr( + args, "share_decoder_input_output_embed", False + ) + args.share_all_embeddings = getattr(args, "share_all_embeddings", True) + args.no_token_positional_embeddings = getattr( + args, "no_token_positional_embeddings", False + ) + args.adaptive_input = getattr(args, "adaptive_input", False) + args.apply_bert_init = getattr(args, "apply_bert_init", False) + + args.decoder_output_dim = getattr( + args, "decoder_output_dim", args.decoder_embed_dim + ) + args.decoder_input_dim = getattr(args, "decoder_input_dim", args.decoder_embed_dim) + + # --- special arguments --- + args.sg_length_pred = getattr(args, "sg_length_pred", False) + args.pred_length_offset = getattr(args, "pred_length_offset", False) + args.length_loss_factor = getattr(args, "length_loss_factor", 0.1) + args.ngram_predictor = getattr(args, "ngram_predictor", 1) + args.src_embedding_copy = getattr(args, "src_embedding_copy", False) + + +@register_model_architecture("cmlm_transformer", "cmlm_transformer_wmt_en_de") +def cmlm_wmt_en_de(args): + cmlm_base_architecture(args) diff --git a/fairseq/models/nat/fairseq_nat_model.py b/fairseq/models/nat/fairseq_nat_model.py new file mode 100644 index 0000000000000000000000000000000000000000..b09394112f57d9e82f2a4cbc371af888281b9e8a --- /dev/null +++ b/fairseq/models/nat/fairseq_nat_model.py @@ -0,0 +1,170 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +import torch +from fairseq.models.transformer import ( + TransformerDecoder, + TransformerEncoder, + TransformerModel, +) +from fairseq.modules.transformer_sentence_encoder import init_bert_params + + +def ensemble_encoder(func): + def wrapper(self, *args, **kwargs): + if self.ensemble_models is None or len(self.ensemble_models) == 1: + return func(self, *args, **kwargs) + encoder_outs = [func(model, *args, **kwargs, return_all_hiddens=True) for model in self.ensemble_models] + _encoder_out = encoder_outs[0].copy() + + def stack(key): + outs = [e[key][0] for e in encoder_outs] + return [torch.stack(outs, -1) if outs[0] is not None else None] + + _encoder_out["encoder_out"] = stack("encoder_out") + _encoder_out["encoder_embedding"] = stack("encoder_embedding") + + num_layers = len(_encoder_out["encoder_states"]) + if num_layers > 0: + _encoder_out["encoder_states"] = [ + torch.stack([e["encoder_states"][i] for e in encoder_outs], -1) + for i in range(num_layers) + ] + return _encoder_out + + return wrapper + + +def ensemble_decoder(func): + def wrapper(self, normalize=False, encoder_out=None, *args, **kwargs): + if self.ensemble_models is None or len(self.ensemble_models) == 1: + return func( + self, normalize=normalize, encoder_out=encoder_out, *args, **kwargs + ) + + def _replace(encoder_out, new_val): + new_encoder_out = encoder_out.copy() + new_encoder_out["encoder_out"] = [new_val] + return new_encoder_out + + action_outs = [ + func( + model, + normalize=normalize, + encoder_out=_replace( + encoder_out, + encoder_out["encoder_out"][0][:, :, :, i] + ), + *args, + **kwargs + ) + for i, model in enumerate(self.ensemble_models) + ] + + if not isinstance(action_outs[0], tuple): # return multiple values + action_outs = [[a] for a in action_outs] + else: + action_outs = [list(a) for a in action_outs] + + ensembled_outs = [] + for i in range(len(action_outs[0])): + if i == 0 and normalize: + ensembled_outs += [ + torch.logsumexp( + torch.stack([a[i] for a in action_outs], -1), dim=-1 + ) + - math.log(len(self.ensemble_models)) + ] + elif action_outs[0][i] is not None: + ensembled_outs += [torch.stack([a[i] for a in action_outs], -1)] + else: + ensembled_outs += [None] + + if len(ensembled_outs) == 1: + return ensembled_outs[0] + return tuple(ensembled_outs) + + return wrapper + + +class FairseqNATModel(TransformerModel): + """ + Abstract class for all nonautoregressive-based models + """ + + def __init__(self, args, encoder, decoder): + super().__init__(args, encoder, decoder) + self.tgt_dict = decoder.dictionary + self.bos = decoder.dictionary.bos() + self.eos = decoder.dictionary.eos() + self.pad = decoder.dictionary.pad() + self.unk = decoder.dictionary.unk() + + self.ensemble_models = None + + @property + def allow_length_beam(self): + return False + + @property + def allow_ensemble(self): + return True + + def enable_ensemble(self, models): + self.encoder.ensemble_models = [m.encoder for m in models] + self.decoder.ensemble_models = [m.decoder for m in models] + + @staticmethod + def add_args(parser): + TransformerModel.add_args(parser) + parser.add_argument( + "--apply-bert-init", + action="store_true", + help="use custom param initialization for BERT", + ) + + @classmethod + def build_decoder(cls, args, tgt_dict, embed_tokens): + decoder = FairseqNATDecoder(args, tgt_dict, embed_tokens) + if getattr(args, "apply_bert_init", False): + decoder.apply(init_bert_params) + return decoder + + @classmethod + def build_encoder(cls, args, src_dict, embed_tokens): + encoder = FairseqNATEncoder(args, src_dict, embed_tokens) + if getattr(args, "apply_bert_init", False): + encoder.apply(init_bert_params) + return encoder + + def forward_encoder(self, encoder_inputs): + return self.encoder(*encoder_inputs) + + def forward_decoder(self, *args, **kwargs): + return NotImplementedError + + def initialize_output_tokens(self, *args, **kwargs): + return NotImplementedError + + def forward(self, *args, **kwargs): + return NotImplementedError + + +class FairseqNATEncoder(TransformerEncoder): + def __init__(self, args, dictionary, embed_tokens): + super().__init__(args, dictionary, embed_tokens) + self.ensemble_models = None + + @ensemble_encoder + def forward(self, *args, **kwargs): + return super().forward(*args, **kwargs) + + +class FairseqNATDecoder(TransformerDecoder): + def __init__(self, args, dictionary, embed_tokens, no_encoder_attn=False): + super().__init__(args, dictionary, embed_tokens, no_encoder_attn) + self.ensemble_models = None diff --git a/fairseq/models/nat/insertion_transformer.py b/fairseq/models/nat/insertion_transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..bc28000f59a3b9e8098f9fe710cc8335d39eea3e --- /dev/null +++ b/fairseq/models/nat/insertion_transformer.py @@ -0,0 +1,280 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import numpy as np +import torch +import torch.nn.functional as F +from fairseq.models import register_model, register_model_architecture +from fairseq.models.nat import ( + FairseqNATModel, + LevenshteinTransformerDecoder, + LevenshteinTransformerModel, + ensemble_decoder, +) +from fairseq.models.transformer import Linear +from fairseq.modules.transformer_sentence_encoder import init_bert_params +from fairseq.utils import new_arange + + +class NegativeDistanceScore(object): + def __init__(self): + + # pre-compute some values + self.scores = {} + + self.scores[0.5] = self.compute_score_full(50, 0.5) + self.scores[1.0] = self.compute_score_full(50, 1.0) + self.scores[2.0] = self.compute_score_full(50, 2.0) + + def __call__(self, i, L, tau): + if (tau is None) or (tau > 1000): + return 1 / L + + if tau in self.scores: + if L < self.scores[tau].shape[0]: + return self.scores[tau][L - 1, i] + return self.compute_score(L, tau)[i] + + def compute_score(self, L, tau): + s = np.array([-abs(L / 2 - i) / tau for i in range(L)]) + s = np.exp(s - s.max()) + return s / s.sum() + + def compute_score_full(self, L, tau): + s = -abs(np.arange(0, L - 1)[:, None] / 2 - np.arange(L)[None, :]) / tau + s = np.tril(s, 0) + np.triu(s - float("inf"), 1) + s = np.exp(s - s.max(1, keepdims=True)) + return s / s.sum(1, keepdims=True) + + +neg_scorer = NegativeDistanceScore() + + +def _get_ins_targets(in_tokens, out_tokens, padding_idx, unk_idx, vocab_size, tau=None): + try: + from fairseq import libnat + except ImportError as e: + import sys + + sys.stderr.write("ERROR: missing libnat. run `pip install --editable .`\n") + raise e + + B = in_tokens.size(0) + T = in_tokens.size(1) + V = vocab_size + + with torch.cuda.device_of(in_tokens): + in_tokens_list = [ + [t for t in s if t != padding_idx] for i, s in enumerate(in_tokens.tolist()) + ] + out_tokens_list = [ + [t for t in s if t != padding_idx] + for i, s in enumerate(out_tokens.tolist()) + ] + + full_labels = libnat.suggested_ed2_path( + in_tokens_list, out_tokens_list, padding_idx + ) + insert_labels = [a[:-1] for a in full_labels] + + # numericalize1 + insert_label_tensors = in_tokens.new_zeros(B * (T - 1) * V).float() + insert_index, insert_labels = zip( + *[ + (w + (j + i * (T - 1)) * V, neg_scorer(k, len(label), tau)) + for i, labels in enumerate(insert_labels) + for j, label in enumerate(labels[1:-1]) + for k, w in enumerate(label) + ] + ) # HACK 1:-1 + insert_index, insert_labels = [ + torch.tensor(list(a), device=in_tokens.device) + for a in [insert_index, insert_labels] + ] + insert_label_tensors.scatter_(0, insert_index.long(), insert_labels) + insert_label_tensors = insert_label_tensors.view(B, T - 1, V) + + return insert_label_tensors + + +def _apply_ins_words(in_tokens, in_scores, word_ins_pred, word_ins_scores, padding_idx): + + padding_masks = in_tokens[:, 1:].eq(padding_idx) + word_ins_scores.masked_fill_(padding_masks, 0.0) + word_ins_pred.masked_fill_(padding_masks, padding_idx) + + in_coords = new_arange(in_tokens).type_as(in_scores) + + # shift all padding predictions to infinite + out_coords = (in_coords[:, 1:] - 0.5).masked_fill( + word_ins_pred.eq(padding_idx), float("inf") + ) + out_coords = torch.cat([in_coords, out_coords], 1).sort(-1)[1] + out_tokens = torch.cat([in_tokens, word_ins_pred], 1).gather(1, out_coords) + out_scores = torch.cat([in_scores, word_ins_scores], 1).gather(1, out_coords) + return out_tokens, out_scores + + +@register_model("insertion_transformer") +class InsertionTransformerModel(LevenshteinTransformerModel): + def __init__(self, args, encoder, decoder): + super().__init__(args, encoder, decoder) + + @staticmethod + def add_args(parser): + FairseqNATModel.add_args(parser) + parser.add_argument("--label-tau", default=None, type=float) + + @classmethod + def build_decoder(cls, args, tgt_dict, embed_tokens): + decoder = InsertionTransformerDecoder(args, tgt_dict, embed_tokens) + if getattr(args, "apply_bert_init", False): + decoder.apply(init_bert_params) + return decoder + + def forward( + self, src_tokens, src_lengths, prev_output_tokens, tgt_tokens, **kwargs + ): + + assert tgt_tokens is not None, "forward function only supports training." + + # encoding + encoder_out = self.encoder(src_tokens, src_lengths=src_lengths, **kwargs) + + # generate training labels for insertion + word_ins_out = self.decoder.forward_word_ins( + normalize=False, + prev_output_tokens=prev_output_tokens, + encoder_out=encoder_out, + ) + + word_ins_tgt = _get_ins_targets( + prev_output_tokens, + tgt_tokens, + self.pad, + self.unk, + len(self.tgt_dict), + tau=self.decoder.label_tau, + ).type_as(word_ins_out) + word_ins_masks = prev_output_tokens[:, 1:].ne(self.pad) + + return { + "word_ins": { + "out": word_ins_out, + "tgt": word_ins_tgt, + "mask": word_ins_masks, + "ls": self.args.label_smoothing, + "nll_loss": True, + } + } + + def forward_decoder( + self, decoder_out, encoder_out, eos_penalty=0.0, max_ratio=None, **kwargs + ): + + output_tokens = decoder_out.output_tokens + output_scores = decoder_out.output_scores + history = decoder_out.history + + # TODO: decoding for InsertionTransformer + word_ins_score = self.decoder.forward_word_ins( + normalize=True, prev_output_tokens=output_tokens, encoder_out=encoder_out + ) + + if eos_penalty > 0.0: + word_ins_score[:, :, self.pad] -= eos_penalty + word_ins_score, word_ins_pred = word_ins_score.max(-1) + output_tokens, output_scores = _apply_ins_words( + output_tokens, output_scores, word_ins_pred, word_ins_score, self.pad + ) + + # delete some unnecessary paddings + cut_off = output_tokens.ne(self.pad).sum(1).max() + output_tokens = output_tokens[:, :cut_off] + output_scores = output_scores[:, :cut_off] + + if history is not None: + history.append(output_tokens.clone()) + + return decoder_out._replace( + output_tokens=output_tokens, + output_scores=output_scores, + attn=None, + history=history, + ) + + +class InsertionTransformerDecoder(LevenshteinTransformerDecoder): + def __init__(self, args, dictionary, embed_tokens, no_encoder_attn=False): + # use the TransformerDecoder's __init__ + super(LevenshteinTransformerDecoder, self).__init__( + args, dictionary, embed_tokens, no_encoder_attn=no_encoder_attn + ) + + self.dictionary = dictionary + self.bos = dictionary.bos() + self.unk = dictionary.unk() + self.eos = dictionary.eos() + self.pool_out = Linear(self.output_embed_dim * 2, self.output_embed_dim) + + self.label_tau = getattr(args, "label_tau", None) + + @ensemble_decoder + def forward_word_ins(self, normalize, encoder_out, prev_output_tokens): + features = self.extract_features(prev_output_tokens, encoder_out=encoder_out)[0] + features = self.pool_out( + torch.cat([features[:, :-1, :], features[:, 1:, :]], 2) + ) + decoder_out = self.output_layer(features) + return F.log_softmax(decoder_out, -1) if normalize else decoder_out + + def forward_mask_ins(self, *args, **kwargs): + raise NotImplementedError + + def forward_word_del(self, *args, **kwargs): + raise NotImplementedError + + +@register_model_architecture("insertion_transformer", "insertion_transformer") +def insertion_base_architecture(args): + args.encoder_embed_path = getattr(args, "encoder_embed_path", None) + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 2048) + args.encoder_layers = getattr(args, "encoder_layers", 6) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 8) + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False) + args.encoder_learned_pos = getattr(args, "encoder_learned_pos", False) + args.decoder_embed_path = getattr(args, "decoder_embed_path", None) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", args.encoder_embed_dim) + args.decoder_ffn_embed_dim = getattr( + args, "decoder_ffn_embed_dim", args.encoder_ffn_embed_dim + ) + args.decoder_layers = getattr(args, "decoder_layers", 6) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 8) + args.decoder_normalize_before = getattr(args, "decoder_normalize_before", False) + args.decoder_learned_pos = getattr(args, "decoder_learned_pos", False) + args.attention_dropout = getattr(args, "attention_dropout", 0.0) + args.activation_dropout = getattr(args, "activation_dropout", 0.0) + args.activation_fn = getattr(args, "activation_fn", "relu") + args.dropout = getattr(args, "dropout", 0.1) + args.adaptive_softmax_cutoff = getattr(args, "adaptive_softmax_cutoff", None) + args.adaptive_softmax_dropout = getattr(args, "adaptive_softmax_dropout", 0) + args.share_decoder_input_output_embed = getattr( + args, "share_decoder_input_output_embed", False + ) + args.share_all_embeddings = getattr(args, "share_all_embeddings", False) + args.no_token_positional_embeddings = getattr( + args, "no_token_positional_embeddings", False + ) + args.adaptive_input = getattr(args, "adaptive_input", False) + args.apply_bert_init = getattr(args, "apply_bert_init", False) + + args.decoder_output_dim = getattr( + args, "decoder_output_dim", args.decoder_embed_dim + ) + args.decoder_input_dim = getattr(args, "decoder_input_dim", args.decoder_embed_dim) + + # special for insertion transformer + args.label_tau = getattr(args, "label_tau", None) diff --git a/fairseq/models/nat/iterative_nonautoregressive_transformer.py b/fairseq/models/nat/iterative_nonautoregressive_transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..bc39509980a80eb8c21e0bfdb304649ad3acc4d0 --- /dev/null +++ b/fairseq/models/nat/iterative_nonautoregressive_transformer.py @@ -0,0 +1,228 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +from fairseq.models import register_model, register_model_architecture +from fairseq.models.nat import NATransformerModel + + +def _sequential_poisoning(s, V, beta=0.33, bos=2, eos=3, pad=1): + # s: input batch + # V: vocabulary size + rand_words = torch.randint(low=4, high=V, size=s.size(), device=s.device) + choices = torch.rand(size=s.size(), device=s.device) + choices.masked_fill_((s == pad) | (s == bos) | (s == eos), 1) + + replace = choices < beta / 3 + repeat = (choices >= beta / 3) & (choices < beta * 2 / 3) + swap = (choices >= beta * 2 / 3) & (choices < beta) + safe = choices >= beta + + for i in range(s.size(1) - 1): + rand_word = rand_words[:, i] + next_word = s[:, i + 1] + self_word = s[:, i] + + replace_i = replace[:, i] + swap_i = swap[:, i] & (next_word != 3) + repeat_i = repeat[:, i] & (next_word != 3) + safe_i = safe[:, i] | ((next_word == 3) & (~replace_i)) + + s[:, i] = ( + self_word * (safe_i | repeat_i).long() + + next_word * swap_i.long() + + rand_word * replace_i.long() + ) + s[:, i + 1] = ( + next_word * (safe_i | replace_i).long() + + self_word * (swap_i | repeat_i).long() + ) + return s + + +def gumbel_noise(input, TINY=1e-8): + return ( + input.new_zeros(*input.size()) + .uniform_() + .add_(TINY) + .log_() + .neg_() + .add_(TINY) + .log_() + .neg_() + ) + + +@register_model("iterative_nonautoregressive_transformer") +class IterNATransformerModel(NATransformerModel): + @staticmethod + def add_args(parser): + NATransformerModel.add_args(parser) + parser.add_argument( + "--train-step", + type=int, + help="number of refinement iterations during training", + ) + parser.add_argument( + "--dae-ratio", + type=float, + help="the probability of switching to the denoising auto-encoder loss", + ) + parser.add_argument( + "--stochastic-approx", + action="store_true", + help="sampling from the decoder as the inputs for next iteration", + ) + + @classmethod + def build_model(cls, args, task): + model = super().build_model(args, task) + model.train_step = getattr(args, "train_step", 4) + model.dae_ratio = getattr(args, "dae_ratio", 0.5) + model.stochastic_approx = getattr(args, "stochastic_approx", False) + return model + + def forward( + self, src_tokens, src_lengths, prev_output_tokens, tgt_tokens, **kwargs + ): + + B, T = prev_output_tokens.size() + + # encoding + encoder_out = self.encoder(src_tokens, src_lengths=src_lengths, **kwargs) + + # length prediction + length_out = self.decoder.forward_length( + normalize=False, encoder_out=encoder_out + ) + length_tgt = self.decoder.forward_length_prediction( + length_out, encoder_out, tgt_tokens + ) + + # decoding + word_ins_outs, word_ins_tgts, word_ins_masks = [], [], [] + for t in range(self.train_step): + word_ins_out = self.decoder( + normalize=False, + prev_output_tokens=prev_output_tokens, + encoder_out=encoder_out, + step=t, + ) + word_ins_tgt = tgt_tokens + word_ins_mask = word_ins_tgt.ne(self.pad) + + word_ins_outs.append(word_ins_out) + word_ins_tgts.append(word_ins_tgt) + word_ins_masks.append(word_ins_mask) + + if t < (self.train_step - 1): + # prediction for next iteration + if self.stochastic_approx: + word_ins_prediction = ( + word_ins_out + gumbel_noise(word_ins_out) + ).max(-1)[1] + else: + word_ins_prediction = word_ins_out.max(-1)[1] + + prev_output_tokens = prev_output_tokens.masked_scatter( + word_ins_mask, word_ins_prediction[word_ins_mask] + ) + + if self.dae_ratio > 0: + # we do not perform denoising for the first iteration + corrputed = ( + torch.rand(size=(B,), device=prev_output_tokens.device) + < self.dae_ratio + ) + corrputed_tokens = _sequential_poisoning( + tgt_tokens[corrputed], + len(self.tgt_dict), + 0.33, + self.bos, + self.eos, + self.pad, + ) + prev_output_tokens[corrputed] = corrputed_tokens + + # concat everything + word_ins_out = torch.cat(word_ins_outs, 0) + word_ins_tgt = torch.cat(word_ins_tgts, 0) + word_ins_mask = torch.cat(word_ins_masks, 0) + + return { + "word_ins": { + "out": word_ins_out, + "tgt": word_ins_tgt, + "mask": word_ins_mask, + "ls": self.args.label_smoothing, + "nll_loss": True, + }, + "length": { + "out": length_out, + "tgt": length_tgt, + "factor": self.decoder.length_loss_factor, + }, + } + + +@register_model_architecture( + "iterative_nonautoregressive_transformer", "iterative_nonautoregressive_transformer" +) +def inat_base_architecture(args): + args.encoder_embed_path = getattr(args, "encoder_embed_path", None) + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 2048) + args.encoder_layers = getattr(args, "encoder_layers", 6) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 8) + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False) + args.encoder_learned_pos = getattr(args, "encoder_learned_pos", False) + args.decoder_embed_path = getattr(args, "decoder_embed_path", None) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", args.encoder_embed_dim) + args.decoder_ffn_embed_dim = getattr( + args, "decoder_ffn_embed_dim", args.encoder_ffn_embed_dim + ) + args.decoder_layers = getattr(args, "decoder_layers", 6) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 8) + args.decoder_normalize_before = getattr(args, "decoder_normalize_before", False) + args.decoder_learned_pos = getattr(args, "decoder_learned_pos", False) + args.attention_dropout = getattr(args, "attention_dropout", 0.0) + args.activation_dropout = getattr(args, "activation_dropout", 0.0) + args.activation_fn = getattr(args, "activation_fn", "relu") + args.dropout = getattr(args, "dropout", 0.1) + args.adaptive_softmax_cutoff = getattr(args, "adaptive_softmax_cutoff", None) + args.adaptive_softmax_dropout = getattr(args, "adaptive_softmax_dropout", 0) + args.share_decoder_input_output_embed = getattr( + args, "share_decoder_input_output_embed", False + ) + args.share_all_embeddings = getattr(args, "share_all_embeddings", False) + args.no_token_positional_embeddings = getattr( + args, "no_token_positional_embeddings", False + ) + args.adaptive_input = getattr(args, "adaptive_input", False) + args.apply_bert_init = getattr(args, "apply_bert_init", False) + + args.decoder_output_dim = getattr( + args, "decoder_output_dim", args.decoder_embed_dim + ) + args.decoder_input_dim = getattr(args, "decoder_input_dim", args.decoder_embed_dim) + + # --- special arguments --- + args.sg_length_pred = getattr(args, "sg_length_pred", False) + args.pred_length_offset = getattr(args, "pred_length_offset", False) + args.length_loss_factor = getattr(args, "length_loss_factor", 0.1) + args.ngram_predictor = getattr(args, "ngram_predictor", 1) + args.src_embedding_copy = getattr(args, "src_embedding_copy", False) + + args.train_step = getattr(args, "train_step", 4) + args.dae_ratio = getattr(args, "dae_ratio", 0.5) + args.stochastic_approx = getattr(args, "stochastic_approx", False) + + +@register_model_architecture( + "iterative_nonautoregressive_transformer", + "iterative_nonautoregressive_transformer_wmt_en_de", +) +def iter_nat_wmt_en_de(args): + inat_base_architecture(args) diff --git a/fairseq/models/nat/levenshtein_transformer.py b/fairseq/models/nat/levenshtein_transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..9377c3c7f5ad6b298eedfb2dc11f1a7a52d1cf26 --- /dev/null +++ b/fairseq/models/nat/levenshtein_transformer.py @@ -0,0 +1,509 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn as nn +import torch.nn.functional as F +from fairseq.iterative_refinement_generator import DecoderOut +from fairseq.models import register_model, register_model_architecture +from fairseq.models.nat import FairseqNATDecoder, FairseqNATModel, ensemble_decoder +from fairseq.models.transformer import Embedding, TransformerDecoderLayer +from fairseq.modules.transformer_sentence_encoder import init_bert_params + +from .levenshtein_utils import ( + _apply_del_words, + _apply_ins_masks, + _apply_ins_words, + _fill, + _get_del_targets, + _get_ins_targets, + _skip, + _skip_encoder_out, +) + + +@register_model("levenshtein_transformer") +class LevenshteinTransformerModel(FairseqNATModel): + @property + def allow_length_beam(self): + return False + + @staticmethod + def add_args(parser): + FairseqNATModel.add_args(parser) + parser.add_argument( + "--early-exit", + default="6,6,6", + type=str, + help="number of decoder layers before word_del, mask_ins, word_ins", + ) + parser.add_argument( + "--no-share-discriminator", + action="store_true", + help="separate parameters for discriminator", + ) + parser.add_argument( + "--no-share-maskpredictor", + action="store_true", + help="separate parameters for mask-predictor", + ) + parser.add_argument( + "--share-discriminator-maskpredictor", + action="store_true", + help="share the parameters for both mask-predictor and discriminator", + ) + parser.add_argument( + "--sampling-for-deletion", + action="store_true", + help="instead of argmax, use sampling to predict the tokens", + ) + + @classmethod + def build_decoder(cls, args, tgt_dict, embed_tokens): + decoder = LevenshteinTransformerDecoder(args, tgt_dict, embed_tokens) + if getattr(args, "apply_bert_init", False): + decoder.apply(init_bert_params) + return decoder + + def forward( + self, src_tokens, src_lengths, prev_output_tokens, tgt_tokens, **kwargs + ): + + assert tgt_tokens is not None, "forward function only supports training." + + # encoding + encoder_out = self.encoder(src_tokens, src_lengths=src_lengths, **kwargs) + + # generate training labels for insertion + masked_tgt_masks, masked_tgt_tokens, mask_ins_targets = _get_ins_targets( + prev_output_tokens, tgt_tokens, self.pad, self.unk + ) + mask_ins_targets = mask_ins_targets.clamp(min=0, max=255) # for safe prediction + mask_ins_masks = prev_output_tokens[:, 1:].ne(self.pad) + + mask_ins_out, _ = self.decoder.forward_mask_ins( + normalize=False, + prev_output_tokens=prev_output_tokens, + encoder_out=encoder_out, + ) + word_ins_out, _ = self.decoder.forward_word_ins( + normalize=False, + prev_output_tokens=masked_tgt_tokens, + encoder_out=encoder_out, + ) + + # make online prediction + if self.decoder.sampling_for_deletion: + word_predictions = torch.multinomial( + F.softmax(word_ins_out, -1).view(-1, word_ins_out.size(-1)), 1 + ).view(word_ins_out.size(0), -1) + else: + word_predictions = F.log_softmax(word_ins_out, dim=-1).max(2)[1] + + word_predictions.masked_scatter_( + ~masked_tgt_masks, tgt_tokens[~masked_tgt_masks] + ) + + # generate training labels for deletion + word_del_targets = _get_del_targets(word_predictions, tgt_tokens, self.pad) + word_del_out, _ = self.decoder.forward_word_del( + normalize=False, + prev_output_tokens=word_predictions, + encoder_out=encoder_out, + ) + word_del_masks = word_predictions.ne(self.pad) + + return { + "mask_ins": { + "out": mask_ins_out, + "tgt": mask_ins_targets, + "mask": mask_ins_masks, + "ls": 0.01, + }, + "word_ins": { + "out": word_ins_out, + "tgt": tgt_tokens, + "mask": masked_tgt_masks, + "ls": self.args.label_smoothing, + "nll_loss": True, + }, + "word_del": { + "out": word_del_out, + "tgt": word_del_targets, + "mask": word_del_masks, + }, + } + + def forward_decoder( + self, decoder_out, encoder_out, eos_penalty=0.0, max_ratio=None, **kwargs + ): + + output_tokens = decoder_out.output_tokens + output_scores = decoder_out.output_scores + attn = decoder_out.attn + history = decoder_out.history + + bsz = output_tokens.size(0) + if max_ratio is None: + max_lens = torch.zeros_like(output_tokens).fill_(255) + else: + if not encoder_out["encoder_padding_mask"]: + max_src_len = encoder_out["encoder_out"].size(0) + src_lens = encoder_out["encoder_out"].new(bsz).fill_(max_src_len) + else: + src_lens = (~encoder_out["encoder_padding_mask"][0]).sum(1) + max_lens = (src_lens * max_ratio).clamp(min=10).long() + + # delete words + # do not delete tokens if it is <s> </s> + can_del_word = output_tokens.ne(self.pad).sum(1) > 2 + if can_del_word.sum() != 0: # we cannot delete, skip + word_del_score, word_del_attn = self.decoder.forward_word_del( + normalize=True, + prev_output_tokens=_skip(output_tokens, can_del_word), + encoder_out=_skip_encoder_out(self.encoder, encoder_out, can_del_word), + ) + word_del_pred = word_del_score.max(-1)[1].bool() + + _tokens, _scores, _attn = _apply_del_words( + output_tokens[can_del_word], + output_scores[can_del_word], + word_del_attn, + word_del_pred, + self.pad, + self.bos, + self.eos, + ) + output_tokens = _fill(output_tokens, can_del_word, _tokens, self.pad) + output_scores = _fill(output_scores, can_del_word, _scores, 0) + attn = _fill(attn, can_del_word, _attn, 0.0) + + if history is not None: + history.append(output_tokens.clone()) + + # insert placeholders + can_ins_mask = output_tokens.ne(self.pad).sum(1) < max_lens + if can_ins_mask.sum() != 0: + mask_ins_score, _ = self.decoder.forward_mask_ins( + normalize=True, + prev_output_tokens=_skip(output_tokens, can_ins_mask), + encoder_out=_skip_encoder_out(self.encoder, encoder_out, can_ins_mask), + ) + if eos_penalty > 0.0: + mask_ins_score[:, :, 0] = mask_ins_score[:, :, 0] - eos_penalty + mask_ins_pred = mask_ins_score.max(-1)[1] + mask_ins_pred = torch.min( + mask_ins_pred, max_lens[can_ins_mask, None].expand_as(mask_ins_pred) + ) + + _tokens, _scores = _apply_ins_masks( + output_tokens[can_ins_mask], + output_scores[can_ins_mask], + mask_ins_pred, + self.pad, + self.unk, + self.eos, + ) + output_tokens = _fill(output_tokens, can_ins_mask, _tokens, self.pad) + output_scores = _fill(output_scores, can_ins_mask, _scores, 0) + + if history is not None: + history.append(output_tokens.clone()) + + # insert words + can_ins_word = output_tokens.eq(self.unk).sum(1) > 0 + if can_ins_word.sum() != 0: + word_ins_score, word_ins_attn = self.decoder.forward_word_ins( + normalize=True, + prev_output_tokens=_skip(output_tokens, can_ins_word), + encoder_out=_skip_encoder_out(self.encoder, encoder_out, can_ins_word), + ) + word_ins_score, word_ins_pred = word_ins_score.max(-1) + _tokens, _scores = _apply_ins_words( + output_tokens[can_ins_word], + output_scores[can_ins_word], + word_ins_pred, + word_ins_score, + self.unk, + ) + + output_tokens = _fill(output_tokens, can_ins_word, _tokens, self.pad) + output_scores = _fill(output_scores, can_ins_word, _scores, 0) + attn = _fill(attn, can_ins_word, word_ins_attn, 0.0) + + if history is not None: + history.append(output_tokens.clone()) + + # delete some unnecessary paddings + cut_off = output_tokens.ne(self.pad).sum(1).max() + output_tokens = output_tokens[:, :cut_off] + output_scores = output_scores[:, :cut_off] + attn = None if attn is None else attn[:, :cut_off, :] + + return decoder_out._replace( + output_tokens=output_tokens, + output_scores=output_scores, + attn=attn, + history=history, + ) + + def initialize_output_tokens(self, encoder_out, src_tokens): + initial_output_tokens = src_tokens.new_zeros(src_tokens.size(0), 2) + initial_output_tokens[:, 0] = self.bos + initial_output_tokens[:, 1] = self.eos + + initial_output_scores = initial_output_tokens.new_zeros( + *initial_output_tokens.size() + ).type_as(encoder_out["encoder_out"][0]) + + return DecoderOut( + output_tokens=initial_output_tokens, + output_scores=initial_output_scores, + attn=None, + step=0, + max_step=0, + history=None, + ) + + +class LevenshteinTransformerDecoder(FairseqNATDecoder): + def __init__(self, args, dictionary, embed_tokens, no_encoder_attn=False): + super().__init__( + args, dictionary, embed_tokens, no_encoder_attn=no_encoder_attn + ) + self.dictionary = dictionary + self.bos = dictionary.bos() + self.unk = dictionary.unk() + self.eos = dictionary.eos() + self.sampling_for_deletion = getattr(args, "sampling_for_deletion", False) + self.embed_mask_ins = Embedding(256, self.output_embed_dim * 2, None) + self.embed_word_del = Embedding(2, self.output_embed_dim, None) + + # del_word, ins_mask, ins_word + self.early_exit = [int(i) for i in args.early_exit.split(",")] + assert len(self.early_exit) == 3 + + # copy layers for mask-predict/deletion + self.layers_msk = None + if getattr(args, "no_share_maskpredictor", False): + self.layers_msk = nn.ModuleList( + [ + TransformerDecoderLayer(args, no_encoder_attn) + for _ in range(self.early_exit[1]) + ] + ) + self.layers_del = None + if getattr(args, "no_share_discriminator", False): + self.layers_del = nn.ModuleList( + [ + TransformerDecoderLayer(args, no_encoder_attn) + for _ in range(self.early_exit[0]) + ] + ) + + if getattr(args, "share_discriminator_maskpredictor", False): + assert getattr( + args, "no_share_discriminator", False + ), "must set saperate discriminator" + self.layers_msk = self.layers_del + + def extract_features( + self, + prev_output_tokens, + encoder_out=None, + early_exit=None, + layers=None, + **unused + ): + """ + Similar to *forward* but only return features. + Inputs: + prev_output_tokens: Tensor(B, T) + encoder_out: a dictionary of hidden states and masks + + Returns: + tuple: + - the decoder's features of shape `(batch, tgt_len, embed_dim)` + - a dictionary with any model-specific outputs + the LevenshteinTransformer decoder has full-attention to all generated tokens + """ + # embed positions + positions = ( + self.embed_positions(prev_output_tokens) + if self.embed_positions is not None + else None + ) + + # embed tokens and positions + x = self.embed_scale * self.embed_tokens(prev_output_tokens) + if self.project_in_dim is not None: + x = self.project_in_dim(x) + + if positions is not None: + x += positions + x = self.dropout_module(x) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + attn = None + inner_states = [x] + + # decoder layers + decoder_padding_mask = prev_output_tokens.eq(self.padding_idx) + layers = self.layers if layers is None else layers + early_exit = len(layers) if early_exit is None else early_exit + for _, layer in enumerate(layers[:early_exit]): + x, attn, _ = layer( + x, + encoder_out["encoder_out"][0] + if (encoder_out is not None and len(encoder_out["encoder_out"]) > 0) + else None, + encoder_out["encoder_padding_mask"][0] + if ( + encoder_out is not None + and len(encoder_out["encoder_padding_mask"]) > 0 + ) + else None, + self_attn_mask=None, + self_attn_padding_mask=decoder_padding_mask, + ) + inner_states.append(x) + + if self.layer_norm: + x = self.layer_norm(x) + + # T x B x C -> B x T x C + x = x.transpose(0, 1) + + if self.project_out_dim is not None: + x = self.project_out_dim(x) + + return x, {"attn": attn, "inner_states": inner_states} + + @ensemble_decoder + def forward_mask_ins(self, normalize, encoder_out, prev_output_tokens, **unused): + features, extra = self.extract_features( + prev_output_tokens, + encoder_out=encoder_out, + early_exit=self.early_exit[1], + layers=self.layers_msk, + **unused + ) + features_cat = torch.cat([features[:, :-1, :], features[:, 1:, :]], 2) + decoder_out = F.linear(features_cat, self.embed_mask_ins.weight) + if normalize: + return F.log_softmax(decoder_out, -1), extra["attn"] + return decoder_out, extra["attn"] + + @ensemble_decoder + def forward_word_ins(self, normalize, encoder_out, prev_output_tokens, **unused): + features, extra = self.extract_features( + prev_output_tokens, + encoder_out=encoder_out, + early_exit=self.early_exit[2], + layers=self.layers, + **unused + ) + decoder_out = self.output_layer(features) + if normalize: + return F.log_softmax(decoder_out, -1), extra["attn"] + return decoder_out, extra["attn"] + + @ensemble_decoder + def forward_word_del(self, normalize, encoder_out, prev_output_tokens, **unused): + features, extra = self.extract_features( + prev_output_tokens, + encoder_out=encoder_out, + early_exit=self.early_exit[0], + layers=self.layers_del, + **unused + ) + decoder_out = F.linear(features, self.embed_word_del.weight) + if normalize: + return F.log_softmax(decoder_out, -1), extra["attn"] + return decoder_out, extra["attn"] + + +@register_model_architecture("levenshtein_transformer", "levenshtein_transformer") +def levenshtein_base_architecture(args): + args.encoder_embed_path = getattr(args, "encoder_embed_path", None) + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 2048) + args.encoder_layers = getattr(args, "encoder_layers", 6) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 8) + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False) + args.encoder_learned_pos = getattr(args, "encoder_learned_pos", False) + args.decoder_embed_path = getattr(args, "decoder_embed_path", None) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", args.encoder_embed_dim) + args.decoder_ffn_embed_dim = getattr( + args, "decoder_ffn_embed_dim", args.encoder_ffn_embed_dim + ) + args.decoder_layers = getattr(args, "decoder_layers", 6) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 8) + args.decoder_normalize_before = getattr(args, "decoder_normalize_before", False) + args.decoder_learned_pos = getattr(args, "decoder_learned_pos", False) + args.attention_dropout = getattr(args, "attention_dropout", 0.0) + args.activation_dropout = getattr(args, "activation_dropout", 0.0) + args.activation_fn = getattr(args, "activation_fn", "relu") + args.dropout = getattr(args, "dropout", 0.1) + args.adaptive_softmax_cutoff = getattr(args, "adaptive_softmax_cutoff", None) + args.adaptive_softmax_dropout = getattr(args, "adaptive_softmax_dropout", 0) + args.share_decoder_input_output_embed = getattr( + args, "share_decoder_input_output_embed", False + ) + args.share_all_embeddings = getattr(args, "share_all_embeddings", False) + args.no_token_positional_embeddings = getattr( + args, "no_token_positional_embeddings", False + ) + args.adaptive_input = getattr(args, "adaptive_input", False) + args.apply_bert_init = getattr(args, "apply_bert_init", False) + + args.decoder_output_dim = getattr( + args, "decoder_output_dim", args.decoder_embed_dim + ) + args.sampling_for_deletion = getattr(args, "sampling_for_deletion", False) + args.decoder_input_dim = getattr(args, "decoder_input_dim", args.decoder_embed_dim) + args.early_exit = getattr(args, "early_exit", "6,6,6") + args.no_share_discriminator = getattr(args, "no_share_discriminator", False) + args.no_share_maskpredictor = getattr(args, "no_share_maskpredictor", False) + args.share_discriminator_maskpredictor = getattr( + args, "share_discriminator_maskpredictor", False + ) + args.no_share_last_layer = getattr(args, "no_share_last_layer", False) + + +@register_model_architecture( + "levenshtein_transformer", "levenshtein_transformer_wmt_en_de" +) +def levenshtein_transformer_wmt_en_de(args): + levenshtein_base_architecture(args) + + +# similar parameters used in the "Attention Is All You Need" paper (Vaswani et al., 2017) +@register_model_architecture( + "levenshtein_transformer", "levenshtein_transformer_vaswani_wmt_en_de_big" +) +def levenshtein_transformer_vaswani_wmt_en_de_big(args): + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 1024) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 4096) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 16) + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 1024) + args.decoder_ffn_embed_dim = getattr(args, "decoder_ffn_embed_dim", 4096) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 16) + args.dropout = getattr(args, "dropout", 0.3) + levenshtein_base_architecture(args) + + +# default parameters used in tensor2tensor implementation +@register_model_architecture( + "levenshtein_transformer", "levenshtein_transformer_wmt_en_de_big" +) +def levenshtein_transformer_wmt_en_de_big_t2t(args): + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", True) + args.decoder_normalize_before = getattr(args, "decoder_normalize_before", True) + args.attention_dropout = getattr(args, "attention_dropout", 0.1) + args.activation_dropout = getattr(args, "activation_dropout", 0.1) + levenshtein_transformer_vaswani_wmt_en_de_big(args) diff --git a/fairseq/models/nat/levenshtein_utils.py b/fairseq/models/nat/levenshtein_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..375a98c2e11354de085f0a7926f407bd1a6a2ad4 --- /dev/null +++ b/fairseq/models/nat/levenshtein_utils.py @@ -0,0 +1,293 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +from fairseq.utils import new_arange + + +# -------------- Helper Functions --------------------------------------------------- # + + +def load_libnat(): + try: + from fairseq import libnat_cuda + + return libnat_cuda, True + + except ImportError as e: + print(str(e) + "... fall back to CPU version") + + try: + from fairseq import libnat + + return libnat, False + + except ImportError as e: + import sys + + sys.stderr.write( + "ERROR: missing libnat_cuda. run `python setup.py build_ext --inplace`\n" + ) + raise e + + +def _get_ins_targets(in_tokens, out_tokens, padding_idx, unk_idx): + libnat, use_cuda = load_libnat() + + def _get_ins_targets_cuda(in_tokens, out_tokens, padding_idx, unk_idx): + in_masks = in_tokens.ne(padding_idx) + out_masks = out_tokens.ne(padding_idx) + mask_ins_targets, masked_tgt_masks = libnat.generate_insertion_labels( + out_tokens.int(), + libnat.levenshtein_distance( + in_tokens.int(), + out_tokens.int(), + in_masks.sum(1).int(), + out_masks.sum(1).int(), + ), + ) + masked_tgt_masks = masked_tgt_masks.bool() & out_masks + mask_ins_targets = mask_ins_targets.type_as(in_tokens)[ + :, 1 : in_masks.size(1) + ].masked_fill_(~in_masks[:, 1:], 0) + masked_tgt_tokens = out_tokens.masked_fill(masked_tgt_masks, unk_idx) + return masked_tgt_masks, masked_tgt_tokens, mask_ins_targets + + def _get_ins_targets_cpu(in_tokens, out_tokens, padding_idx, unk_idx): + in_seq_len, out_seq_len = in_tokens.size(1), out_tokens.size(1) + + in_tokens_list = [ + [t for t in s if t != padding_idx] for i, s in enumerate(in_tokens.tolist()) + ] + out_tokens_list = [ + [t for t in s if t != padding_idx] + for i, s in enumerate(out_tokens.tolist()) + ] + + full_labels = libnat.suggested_ed2_path( + in_tokens_list, out_tokens_list, padding_idx + ) + mask_inputs = [ + [len(c) if c[0] != padding_idx else 0 for c in a[:-1]] for a in full_labels + ] + + # generate labels + masked_tgt_masks = [] + for mask_input in mask_inputs: + mask_label = [] + for beam_size in mask_input[1:-1]: # HACK 1:-1 + mask_label += [0] + [1 for _ in range(beam_size)] + masked_tgt_masks.append( + mask_label + [0 for _ in range(out_seq_len - len(mask_label))] + ) + mask_ins_targets = [ + mask_input[1:-1] + + [0 for _ in range(in_seq_len - 1 - len(mask_input[1:-1]))] + for mask_input in mask_inputs + ] + + # transform to tensor + masked_tgt_masks = torch.tensor( + masked_tgt_masks, device=out_tokens.device + ).bool() + mask_ins_targets = torch.tensor(mask_ins_targets, device=in_tokens.device) + masked_tgt_tokens = out_tokens.masked_fill(masked_tgt_masks, unk_idx) + return masked_tgt_masks, masked_tgt_tokens, mask_ins_targets + + if use_cuda: + return _get_ins_targets_cuda(in_tokens, out_tokens, padding_idx, unk_idx) + return _get_ins_targets_cpu(in_tokens, out_tokens, padding_idx, unk_idx) + + +def _get_del_targets(in_tokens, out_tokens, padding_idx): + libnat, use_cuda = load_libnat() + + def _get_del_targets_cuda(in_tokens, out_tokens, padding_idx): + in_masks = in_tokens.ne(padding_idx) + out_masks = out_tokens.ne(padding_idx) + + word_del_targets = libnat.generate_deletion_labels( + in_tokens.int(), + libnat.levenshtein_distance( + in_tokens.int(), + out_tokens.int(), + in_masks.sum(1).int(), + out_masks.sum(1).int(), + ), + ) + word_del_targets = word_del_targets.type_as(in_tokens).masked_fill_( + ~in_masks, 0 + ) + return word_del_targets + + def _get_del_targets_cpu(in_tokens, out_tokens, padding_idx): + out_seq_len = out_tokens.size(1) + with torch.cuda.device_of(in_tokens): + in_tokens_list = [ + [t for t in s if t != padding_idx] + for i, s in enumerate(in_tokens.tolist()) + ] + out_tokens_list = [ + [t for t in s if t != padding_idx] + for i, s in enumerate(out_tokens.tolist()) + ] + + full_labels = libnat.suggested_ed2_path( + in_tokens_list, out_tokens_list, padding_idx + ) + word_del_targets = [b[-1] for b in full_labels] + word_del_targets = [ + labels + [0 for _ in range(out_seq_len - len(labels))] + for labels in word_del_targets + ] + + # transform to tensor + word_del_targets = torch.tensor(word_del_targets, device=out_tokens.device) + return word_del_targets + + if use_cuda: + return _get_del_targets_cuda(in_tokens, out_tokens, padding_idx) + return _get_del_targets_cpu(in_tokens, out_tokens, padding_idx) + + +def _apply_ins_masks( + in_tokens, in_scores, mask_ins_pred, padding_idx, unk_idx, eos_idx +): + + in_masks = in_tokens.ne(padding_idx) + in_lengths = in_masks.sum(1) + + # HACK: hacky way to shift all the paddings to eos first. + in_tokens.masked_fill_(~in_masks, eos_idx) + mask_ins_pred.masked_fill_(~in_masks[:, 1:], 0) + + out_lengths = in_lengths + mask_ins_pred.sum(1) + out_max_len = out_lengths.max() + out_masks = new_arange(out_lengths, out_max_len)[None, :] < out_lengths[:, None] + + reordering = (mask_ins_pred + in_masks[:, 1:].long()).cumsum(1) + out_tokens = ( + in_tokens.new_zeros(in_tokens.size(0), out_max_len) + .fill_(padding_idx) + .masked_fill_(out_masks, unk_idx) + ) + out_tokens[:, 0] = in_tokens[:, 0] + out_tokens.scatter_(1, reordering, in_tokens[:, 1:]) + + out_scores = None + if in_scores is not None: + in_scores.masked_fill_(~in_masks, 0) + out_scores = in_scores.new_zeros(*out_tokens.size()) + out_scores[:, 0] = in_scores[:, 0] + out_scores.scatter_(1, reordering, in_scores[:, 1:]) + + return out_tokens, out_scores + + +def _apply_ins_words(in_tokens, in_scores, word_ins_pred, word_ins_scores, unk_idx): + word_ins_masks = in_tokens.eq(unk_idx) + out_tokens = in_tokens.masked_scatter(word_ins_masks, word_ins_pred[word_ins_masks]) + + if in_scores is not None: + out_scores = in_scores.masked_scatter( + word_ins_masks, word_ins_scores[word_ins_masks] + ) + else: + out_scores = None + + return out_tokens, out_scores + + +def _apply_del_words( + in_tokens, in_scores, in_attn, word_del_pred, padding_idx, bos_idx, eos_idx +): + # apply deletion to a tensor + in_masks = in_tokens.ne(padding_idx) + bos_eos_masks = in_tokens.eq(bos_idx) | in_tokens.eq(eos_idx) + + max_len = in_tokens.size(1) + word_del_pred.masked_fill_(~in_masks, 1) + word_del_pred.masked_fill_(bos_eos_masks, 0) + + reordering = new_arange(in_tokens).masked_fill_(word_del_pred, max_len).sort(1)[1] + + out_tokens = in_tokens.masked_fill(word_del_pred, padding_idx).gather(1, reordering) + + out_scores = None + if in_scores is not None: + out_scores = in_scores.masked_fill(word_del_pred, 0).gather(1, reordering) + + out_attn = None + if in_attn is not None: + _mask = word_del_pred[:, :, None].expand_as(in_attn) + _reordering = reordering[:, :, None].expand_as(in_attn) + out_attn = in_attn.masked_fill(_mask, 0.0).gather(1, _reordering) + + return out_tokens, out_scores, out_attn + + +def _skip(x, mask): + """ + Getting sliced (dim=0) tensor by mask. Supporting tensor and list/dict of tensors. + """ + if isinstance(x, int): + return x + + if x is None: + return None + + if isinstance(x, torch.Tensor): + if x.size(0) == mask.size(0): + return x[mask] + elif x.size(1) == mask.size(0): + return x[:, mask] + + if isinstance(x, list): + return [_skip(x_i, mask) for x_i in x] + + if isinstance(x, dict): + return {k: _skip(v, mask) for k, v in x.items()} + + raise NotImplementedError + + +def _skip_encoder_out(encoder, encoder_out, mask): + if not mask.any(): + return encoder_out + else: + return encoder.reorder_encoder_out( + encoder_out, mask.nonzero(as_tuple=False).squeeze() + ) + + +def _fill(x, mask, y, padding_idx): + """ + Filling tensor x with y at masked positions (dim=0). + """ + if x is None: + return y + assert x.dim() == y.dim() and mask.size(0) == x.size(0) + assert x.dim() == 2 or (x.dim() == 3 and x.size(2) == y.size(2)) + n_selected = mask.sum() + assert n_selected == y.size(0) + + if n_selected == x.size(0): + return y + + if x.size(1) < y.size(1): + dims = [x.size(0), y.size(1) - x.size(1)] + if x.dim() == 3: + dims.append(x.size(2)) + x = torch.cat([x, x.new_zeros(*dims).fill_(padding_idx)], 1) + x[mask] = y + elif x.size(1) > y.size(1): + x[mask] = padding_idx + if x.dim() == 2: + x[mask, : y.size(1)] = y + else: + x[mask, : y.size(1), :] = y + else: + x[mask] = y + return x diff --git a/fairseq/models/nat/nat_crf_transformer.py b/fairseq/models/nat/nat_crf_transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..d4b3cd931ceb077eb30db73df1d5d6cd714a86c2 --- /dev/null +++ b/fairseq/models/nat/nat_crf_transformer.py @@ -0,0 +1,121 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +from fairseq.models import register_model, register_model_architecture +from fairseq.models.nat import NATransformerModel, base_architecture +from fairseq.modules import DynamicCRF + + +@register_model("nacrf_transformer") +class NACRFTransformerModel(NATransformerModel): + def __init__(self, args, encoder, decoder): + super().__init__(args, encoder, decoder) + self.crf_layer = DynamicCRF( + num_embedding=len(self.tgt_dict), + low_rank=args.crf_lowrank_approx, + beam_size=args.crf_beam_approx, + ) + + @property + def allow_ensemble(self): + return False + + @staticmethod + def add_args(parser): + NATransformerModel.add_args(parser) + parser.add_argument( + "--crf-lowrank-approx", + type=int, + help="the dimension of low-rank approximation of transition", + ) + parser.add_argument( + "--crf-beam-approx", + type=int, + help="the beam size for apporixmating the normalizing factor", + ) + parser.add_argument( + "--word-ins-loss-factor", + type=float, + help="weights on NAT loss used to co-training with CRF loss.", + ) + + def forward( + self, src_tokens, src_lengths, prev_output_tokens, tgt_tokens, **kwargs + ): + # encoding + encoder_out = self.encoder(src_tokens, src_lengths=src_lengths, **kwargs) + + # length prediction + length_out = self.decoder.forward_length( + normalize=False, encoder_out=encoder_out + ) + length_tgt = self.decoder.forward_length_prediction( + length_out, encoder_out, tgt_tokens + ) + + # decoding + word_ins_out = self.decoder( + normalize=False, + prev_output_tokens=prev_output_tokens, + encoder_out=encoder_out, + ) + word_ins_tgt, word_ins_mask = tgt_tokens, tgt_tokens.ne(self.pad) + + # compute the log-likelihood of CRF + crf_nll = -self.crf_layer(word_ins_out, word_ins_tgt, word_ins_mask) + crf_nll = (crf_nll / word_ins_mask.type_as(crf_nll).sum(-1)).mean() + + return { + "word_ins": { + "out": word_ins_out, + "tgt": word_ins_tgt, + "mask": word_ins_mask, + "ls": self.args.label_smoothing, + "nll_loss": True, + "factor": self.args.word_ins_loss_factor, + }, + "word_crf": {"loss": crf_nll}, + "length": { + "out": length_out, + "tgt": length_tgt, + "factor": self.decoder.length_loss_factor, + }, + } + + def forward_decoder(self, decoder_out, encoder_out, decoding_format=None, **kwargs): + output_tokens = decoder_out.output_tokens + output_scores = decoder_out.output_scores + history = decoder_out.history + + # execute the decoder and get emission scores + output_masks = output_tokens.ne(self.pad) + word_ins_out = self.decoder( + normalize=False, prev_output_tokens=output_tokens, encoder_out=encoder_out + ) + + # run viterbi decoding through CRF + _scores, _tokens = self.crf_layer.forward_decoder(word_ins_out, output_masks) + output_tokens.masked_scatter_(output_masks, _tokens[output_masks]) + output_scores.masked_scatter_(output_masks, _scores[output_masks]) + if history is not None: + history.append(output_tokens.clone()) + + return decoder_out._replace( + output_tokens=output_tokens, + output_scores=output_scores, + attn=None, + history=history, + ) + + +@register_model_architecture("nacrf_transformer", "nacrf_transformer") +def nacrf_base_architecture(args): + args.crf_lowrank_approx = getattr(args, "crf_lowrank_approx", 32) + args.crf_beam_approx = getattr(args, "crf_beam_approx", 64) + args.word_ins_loss_factor = getattr(args, "word_ins_loss_factor", 0.5) + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", True) + args.decoder_normalize_before = getattr(args, "decoder_normalize_before", True) + base_architecture(args) diff --git a/fairseq/models/nat/nonautoregressive_ensembles.py b/fairseq/models/nat/nonautoregressive_ensembles.py new file mode 100644 index 0000000000000000000000000000000000000000..705a04fb49658c91114a26efd411b4653c65b943 --- /dev/null +++ b/fairseq/models/nat/nonautoregressive_ensembles.py @@ -0,0 +1,253 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +import torch +import torch.nn.functional as F +from fairseq.models.nat import ( + _apply_del_words, + _apply_ins_masks, + _apply_ins_words, + _fill, + _skip, + _skip_encoder_out, +) + + +class _EnsembleModelEncoder(object): + def __init__(self, models): + self.models = models + + def reorder_encoder_out(self, encoder_outs, new_order): + encoder_outs = [ + model.encoder.reorder_encoder_out(encoder_out, new_order) + for model, encoder_out in zip(self.models, encoder_outs) + ] + return encoder_outs + + +class BasicEnsembleModel(torch.nn.Module): + """A wrapper around an ensemble of models.""" + + def __init__(self, models): + super().__init__() + self.models = torch.nn.ModuleList(models) + self.bos = self.models[0].decoder.dictionary.bos() + self.eos = self.models[0].decoder.dictionary.eos() + self.pad = self.models[0].decoder.dictionary.pad() + self.unk = self.models[0].decoder.dictionary.unk() + self.encoder = _EnsembleModelEncoder(self.models) + + def has_encoder(self): + return hasattr(self.models[0], "encoder") + + def max_decoder_positions(self): + return min(m.max_decoder_positions() for m in self.models) + + @torch.no_grad() + def forward_encoder(self, encoder_input): + if not self.has_encoder(): + return None + return [model.forward_encoder(encoder_input) for model in self.models] + + @torch.no_grad() + def forward_decoder(self, *inputs): + raise NotImplementedError + + def initialize_output_tokens(self, *inputs): + raise NotImplementedError + + +class EnsembleLevT(BasicEnsembleModel): + """A wrapper around an ensemble of models.""" + + def __init__(self, models): + super().__init__(models) + + @torch.no_grad() + def forward_decoder( + self, decoder_out, encoder_outs, eos_penalty=0.0, max_ratio=None, **kwargs + ): + # LevT ensembling + # A pipeline of three steps: deletion, placeholder, and word insertion. + # We need to average scores in each step in a pipeline way because of dependence. + # deletion + output_tokens = decoder_out.output_tokens + output_scores = decoder_out.output_scores + attn = decoder_out.attn + + bsz = output_tokens.size(0) + if max_ratio is None: + max_lens = output_tokens.new().fill_(255) + else: + if not encoder_outs[0]["encoder_padding_mask"]: + src_lens = ( + encoder_outs[0]["encoder_out"][0].new(bsz) + .fill_(encoder_outs[0]["encoder_out"][0].size(1)) + ) + else: + src_lens = (~encoder_outs[0]["encoder_padding_mask"][0]).sum(1) + max_lens = (src_lens * max_ratio).clamp(min=10).long() + + # delete words + # do not delete tokens if it is <s> </s> + can_del_word = output_tokens.ne(self.pad).sum(1) > 2 + if can_del_word.sum() != 0: # we cannot delete, skip + output_tokens, output_scores, attn = self.forward_word_del( + encoder_outs, + output_tokens, + output_scores, + attn, + can_del_word, + ) + + # insert placeholders + can_ins_mask = output_tokens.ne(self.pad).sum(1) < max_lens + if can_ins_mask.sum() != 0: + output_tokens, output_scores = self.forward_mask_ins( + encoder_outs, + output_tokens, + output_scores, + can_ins_mask, + eos_penalty, + max_lens, + ) + + # insert words + can_ins_word = output_tokens.eq(self.unk).sum(1) > 0 + if can_ins_word.sum() != 0: + output_tokens, output_scores, attn = self.forward_word_ins( + encoder_outs, + output_tokens, + output_scores, + attn, + can_ins_word, + ) + + # delete some unnecessary paddings + cut_off = output_tokens.ne(self.pad).sum(1).max() + output_tokens = output_tokens[:, :cut_off] + output_scores = output_scores[:, :cut_off] + attn = None if attn is None else attn[:, :cut_off, :] + return decoder_out._replace( + output_tokens=output_tokens, + output_scores=output_scores, + attn=attn, + history=None, + ) + + def forward_word_del( + self, encoder_outs, output_tokens, output_scores, attn, can_del_word + ): + word_del_score_avg = [] + word_del_attn_avg = [] + for model, encoder_out in zip(self.models, encoder_outs): + word_del_out, word_del_attn = model.decoder.forward_word_del( + _skip(output_tokens, can_del_word), + _skip_encoder_out(model.encoder, encoder_out, can_del_word), + ) + word_del_score = F.log_softmax(word_del_out, 2) + word_del_score_avg.append(word_del_score) + word_del_attn_avg.append(word_del_attn) + word_del_score_avg = torch.logsumexp( + torch.stack(word_del_score_avg, dim=0), dim=0 + ) - math.log(len(self.models)) + word_del_pred = word_del_score_avg.max(-1)[1].bool() + if word_del_attn_avg[0] is not None: + word_del_attn_avg = torch.stack(word_del_attn_avg, dim=0) / len(self.models) + else: + word_del_attn_avg = None + + _tokens, _scores, _attn = _apply_del_words( + output_tokens[can_del_word], + output_scores[can_del_word], + word_del_attn_avg, + word_del_pred, + self.pad, + self.bos, + self.eos, + ) + output_tokens = _fill(output_tokens, can_del_word, _tokens, self.pad) + output_scores = _fill(output_scores, can_del_word, _scores, 0) + attn = _fill(attn, can_del_word, _attn, 0.0) + return output_tokens, output_scores, attn + + def forward_mask_ins( + self, + encoder_outs, + output_tokens, + output_scores, + can_ins_mask, + eos_penalty, + max_lens, + ): + mask_ins_score_avg = [] + for model, encoder_out in zip(self.models, encoder_outs): + mask_ins_out, _ = model.decoder.forward_mask_ins( + _skip(output_tokens, can_ins_mask), + _skip_encoder_out(model.encoder, encoder_out, can_ins_mask), + ) + mask_ins_score = F.log_softmax(mask_ins_out, 2) + if eos_penalty > 0.0: + mask_ins_score[:, :, 0] -= eos_penalty + mask_ins_score_avg.append(mask_ins_score) + mask_ins_score_avg = torch.logsumexp( + torch.stack(mask_ins_score_avg, dim=0), dim=0 + ) - math.log(len(self.models)) + mask_ins_pred = mask_ins_score_avg.max(-1)[1] + mask_ins_pred = torch.min( + mask_ins_pred, max_lens[can_ins_mask, None].expand_as(mask_ins_pred) + ) + _tokens, _scores = _apply_ins_masks( + output_tokens[can_ins_mask], + output_scores[can_ins_mask], + mask_ins_pred, + self.pad, + self.unk, + self.eos, + ) + output_tokens = _fill(output_tokens, can_ins_mask, _tokens, self.pad) + output_scores = _fill(output_scores, can_ins_mask, _scores, 0) + return output_tokens, output_scores + + def forward_word_ins( + self, encoder_outs, output_tokens, output_scores, attn, can_ins_word + ): + word_ins_score_avg = [] + word_ins_attn_avg = [] + for model, encoder_out in zip(self.models, encoder_outs): + word_ins_out, word_ins_attn = model.decoder.forward_word_ins( + _skip(output_tokens, can_ins_word), + _skip_encoder_out(model.encoder, encoder_out, can_ins_word), + ) + word_ins_score = F.log_softmax(word_ins_out, 2) + word_ins_score_avg.append(word_ins_score) + word_ins_attn_avg.append(word_ins_attn) + word_ins_score_avg = torch.logsumexp( + torch.stack(word_ins_score_avg, dim=0), dim=0 + ) - math.log(len(self.models)) + if word_ins_attn_avg[0] is not None: + word_ins_attn_avg = torch.stack(word_ins_attn_avg, dim=0) / len(self.models) + else: + word_ins_attn_avg = None + word_ins_score_max, word_ins_pred = word_ins_score_avg.max(-1) + + _tokens, _scores = _apply_ins_words( + output_tokens[can_ins_word], + output_scores[can_ins_word], + word_ins_pred, + word_ins_score_max, + self.unk, + ) + + output_tokens = _fill(output_tokens, can_ins_word, _tokens, self.pad) + output_scores = _fill(output_scores, can_ins_word, _scores, 0) + attn = _fill(attn, can_ins_word, word_ins_attn, 0.0) + return output_tokens, output_scores, attn + + def initialize_output_tokens(self, encoder_outs, src_tokens): + # LevT doesn't do length prediction. + return self.models[0].initialize_output_tokens(encoder_outs[0], src_tokens) diff --git a/fairseq/models/nat/nonautoregressive_transformer.py b/fairseq/models/nat/nonautoregressive_transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..d114202d25fbd1dca66c7abebb0b0a8bffbe094d --- /dev/null +++ b/fairseq/models/nat/nonautoregressive_transformer.py @@ -0,0 +1,456 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn.functional as F +from fairseq import utils +from fairseq.iterative_refinement_generator import DecoderOut +from fairseq.models import register_model, register_model_architecture +from fairseq.models.nat import FairseqNATDecoder, FairseqNATModel, ensemble_decoder +from fairseq.models.transformer import Embedding +from fairseq.modules.transformer_sentence_encoder import init_bert_params + + +def _mean_pooling(enc_feats, src_masks): + # enc_feats: T x B x C + # src_masks: B x T or None + if src_masks is None: + enc_feats = enc_feats.mean(0) + else: + src_masks = (~src_masks).transpose(0, 1).type_as(enc_feats) + enc_feats = ( + (enc_feats / src_masks.sum(0)[None, :, None]) * src_masks[:, :, None] + ).sum(0) + return enc_feats + + +def _argmax(x, dim): + return (x == x.max(dim, keepdim=True)[0]).type_as(x) + + +def _uniform_assignment(src_lens, trg_lens): + max_trg_len = trg_lens.max() + steps = (src_lens.float() - 1) / (trg_lens.float() - 1) # step-size + # max_trg_len + index_t = utils.new_arange(trg_lens, max_trg_len).float() + index_t = steps[:, None] * index_t[None, :] # batch_size X max_trg_len + index_t = torch.round(index_t).long().detach() + return index_t + + +@register_model("nonautoregressive_transformer") +class NATransformerModel(FairseqNATModel): + @property + def allow_length_beam(self): + return True + + @staticmethod + def add_args(parser): + FairseqNATModel.add_args(parser) + + # length prediction + parser.add_argument( + "--src-embedding-copy", + action="store_true", + help="copy encoder word embeddings as the initial input of the decoder", + ) + parser.add_argument( + "--pred-length-offset", + action="store_true", + help="predicting the length difference between the target and source sentences", + ) + parser.add_argument( + "--sg-length-pred", + action="store_true", + help="stop the gradients back-propagated from the length predictor", + ) + parser.add_argument( + "--length-loss-factor", + type=float, + help="weights on the length prediction loss", + ) + + @classmethod + def build_decoder(cls, args, tgt_dict, embed_tokens): + decoder = NATransformerDecoder(args, tgt_dict, embed_tokens) + if getattr(args, "apply_bert_init", False): + decoder.apply(init_bert_params) + return decoder + + def forward( + self, src_tokens, src_lengths, prev_output_tokens, tgt_tokens, **kwargs + ): + # encoding + encoder_out = self.encoder(src_tokens, src_lengths=src_lengths, **kwargs) + + # length prediction + length_out = self.decoder.forward_length( + normalize=False, encoder_out=encoder_out + ) + length_tgt = self.decoder.forward_length_prediction( + length_out, encoder_out, tgt_tokens + ) + + # decoding + word_ins_out = self.decoder( + normalize=False, + prev_output_tokens=prev_output_tokens, + encoder_out=encoder_out, + ) + + return { + "word_ins": { + "out": word_ins_out, + "tgt": tgt_tokens, + "mask": tgt_tokens.ne(self.pad), + "ls": self.args.label_smoothing, + "nll_loss": True, + }, + "length": { + "out": length_out, + "tgt": length_tgt, + "factor": self.decoder.length_loss_factor, + }, + } + + def forward_decoder(self, decoder_out, encoder_out, decoding_format=None, **kwargs): + step = decoder_out.step + output_tokens = decoder_out.output_tokens + output_scores = decoder_out.output_scores + history = decoder_out.history + + # execute the decoder + output_masks = output_tokens.ne(self.pad) + _scores, _tokens = self.decoder( + normalize=True, + prev_output_tokens=output_tokens, + encoder_out=encoder_out, + step=step, + ).max(-1) + + output_tokens.masked_scatter_(output_masks, _tokens[output_masks]) + output_scores.masked_scatter_(output_masks, _scores[output_masks]) + if history is not None: + history.append(output_tokens.clone()) + + return decoder_out._replace( + output_tokens=output_tokens, + output_scores=output_scores, + attn=None, + history=history, + ) + + def initialize_output_tokens(self, encoder_out, src_tokens): + # length prediction + length_tgt = self.decoder.forward_length_prediction( + self.decoder.forward_length(normalize=True, encoder_out=encoder_out), + encoder_out=encoder_out, + ) + + max_length = length_tgt.clamp_(min=2).max() + idx_length = utils.new_arange(src_tokens, max_length) + + initial_output_tokens = src_tokens.new_zeros( + src_tokens.size(0), max_length + ).fill_(self.pad) + initial_output_tokens.masked_fill_( + idx_length[None, :] < length_tgt[:, None], self.unk + ) + initial_output_tokens[:, 0] = self.bos + initial_output_tokens.scatter_(1, length_tgt[:, None] - 1, self.eos) + + initial_output_scores = initial_output_tokens.new_zeros( + *initial_output_tokens.size() + ).type_as(encoder_out["encoder_out"][0]) + + return DecoderOut( + output_tokens=initial_output_tokens, + output_scores=initial_output_scores, + attn=None, + step=0, + max_step=0, + history=None, + ) + + def regenerate_length_beam(self, decoder_out, beam_size): + output_tokens = decoder_out.output_tokens + length_tgt = output_tokens.ne(self.pad).sum(1) + length_tgt = ( + length_tgt[:, None] + + utils.new_arange(length_tgt, 1, beam_size) + - beam_size // 2 + ) + length_tgt = length_tgt.view(-1).clamp_(min=2) + max_length = length_tgt.max() + idx_length = utils.new_arange(length_tgt, max_length) + + initial_output_tokens = output_tokens.new_zeros( + length_tgt.size(0), max_length + ).fill_(self.pad) + initial_output_tokens.masked_fill_( + idx_length[None, :] < length_tgt[:, None], self.unk + ) + initial_output_tokens[:, 0] = self.bos + initial_output_tokens.scatter_(1, length_tgt[:, None] - 1, self.eos) + + initial_output_scores = initial_output_tokens.new_zeros( + *initial_output_tokens.size() + ).type_as(decoder_out.output_scores) + + return decoder_out._replace( + output_tokens=initial_output_tokens, output_scores=initial_output_scores + ) + + +class NATransformerDecoder(FairseqNATDecoder): + def __init__(self, args, dictionary, embed_tokens, no_encoder_attn=False): + super().__init__( + args, dictionary, embed_tokens, no_encoder_attn=no_encoder_attn + ) + self.dictionary = dictionary + self.bos = dictionary.bos() + self.unk = dictionary.unk() + self.eos = dictionary.eos() + + self.encoder_embed_dim = args.encoder_embed_dim + self.sg_length_pred = getattr(args, "sg_length_pred", False) + self.pred_length_offset = getattr(args, "pred_length_offset", False) + self.length_loss_factor = getattr(args, "length_loss_factor", 0.1) + self.src_embedding_copy = getattr(args, "src_embedding_copy", False) + self.embed_length = Embedding(256, self.encoder_embed_dim, None) + + @ensemble_decoder + def forward(self, normalize, encoder_out, prev_output_tokens, step=0, **unused): + features, _ = self.extract_features( + prev_output_tokens, + encoder_out=encoder_out, + embedding_copy=(step == 0) & self.src_embedding_copy, + ) + decoder_out = self.output_layer(features) + return F.log_softmax(decoder_out, -1) if normalize else decoder_out + + @ensemble_decoder + def forward_length(self, normalize, encoder_out): + enc_feats = encoder_out["encoder_out"][0] # T x B x C + if len(encoder_out["encoder_padding_mask"]) > 0: + src_masks = encoder_out["encoder_padding_mask"][0] # B x T + else: + src_masks = None + enc_feats = _mean_pooling(enc_feats, src_masks) + if self.sg_length_pred: + enc_feats = enc_feats.detach() + length_out = F.linear(enc_feats, self.embed_length.weight) + return F.log_softmax(length_out, -1) if normalize else length_out + + def extract_features( + self, + prev_output_tokens, + encoder_out=None, + early_exit=None, + embedding_copy=False, + **unused + ): + """ + Similar to *forward* but only return features. + + Inputs: + prev_output_tokens: Tensor(B, T) + encoder_out: a dictionary of hidden states and masks + + Returns: + tuple: + - the decoder's features of shape `(batch, tgt_len, embed_dim)` + - a dictionary with any model-specific outputs + the LevenshteinTransformer decoder has full-attention to all generated tokens + """ + # embedding + if embedding_copy: + src_embd = encoder_out["encoder_embedding"][0] + if len(encoder_out["encoder_padding_mask"]) > 0: + src_mask = encoder_out["encoder_padding_mask"][0] + else: + src_mask = None + src_mask = ( + ~src_mask + if src_mask is not None + else prev_output_tokens.new_ones(*src_embd.size()[:2]).bool() + ) + + x, decoder_padding_mask = self.forward_embedding( + prev_output_tokens, + self.forward_copying_source( + src_embd, src_mask, prev_output_tokens.ne(self.padding_idx) + ), + ) + + else: + + x, decoder_padding_mask = self.forward_embedding(prev_output_tokens) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + attn = None + inner_states = [x] + + # decoder layers + for i, layer in enumerate(self.layers): + + # early exit from the decoder. + if (early_exit is not None) and (i >= early_exit): + break + + x, attn, _ = layer( + x, + encoder_out["encoder_out"][0] + if (encoder_out is not None and len(encoder_out["encoder_out"]) > 0) + else None, + encoder_out["encoder_padding_mask"][0] + if ( + encoder_out is not None + and len(encoder_out["encoder_padding_mask"]) > 0 + ) + else None, + self_attn_mask=None, + self_attn_padding_mask=decoder_padding_mask, + ) + inner_states.append(x) + + if self.layer_norm: + x = self.layer_norm(x) + + # T x B x C -> B x T x C + x = x.transpose(0, 1) + + if self.project_out_dim is not None: + x = self.project_out_dim(x) + + return x, {"attn": attn, "inner_states": inner_states} + + def forward_embedding(self, prev_output_tokens, states=None): + # embed positions + positions = ( + self.embed_positions(prev_output_tokens) + if self.embed_positions is not None + else None + ) + + # embed tokens and positions + if states is None: + x = self.embed_scale * self.embed_tokens(prev_output_tokens) + if self.project_in_dim is not None: + x = self.project_in_dim(x) + else: + x = states + + if positions is not None: + x += positions + x = self.dropout_module(x) + decoder_padding_mask = prev_output_tokens.eq(self.padding_idx) + return x, decoder_padding_mask + + def forward_copying_source(self, src_embeds, src_masks, tgt_masks): + length_sources = src_masks.sum(1) + length_targets = tgt_masks.sum(1) + mapped_inputs = _uniform_assignment(length_sources, length_targets).masked_fill( + ~tgt_masks, 0 + ) + copied_embedding = torch.gather( + src_embeds, + 1, + mapped_inputs.unsqueeze(-1).expand( + *mapped_inputs.size(), src_embeds.size(-1) + ), + ) + return copied_embedding + + def forward_length_prediction(self, length_out, encoder_out, tgt_tokens=None): + enc_feats = encoder_out["encoder_out"][0] # T x B x C + if len(encoder_out["encoder_padding_mask"]) > 0: + src_masks = encoder_out["encoder_padding_mask"][0] # B x T + else: + src_masks = None + if self.pred_length_offset: + if src_masks is None: + src_lengs = enc_feats.new_ones(enc_feats.size(1)).fill_( + enc_feats.size(0) + ) + else: + src_lengs = (~src_masks).transpose(0, 1).type_as(enc_feats).sum(0) + src_lengs = src_lengs.long() + + if tgt_tokens is not None: + # obtain the length target + tgt_lengs = tgt_tokens.ne(self.padding_idx).sum(1).long() + if self.pred_length_offset: + length_tgt = tgt_lengs - src_lengs + 128 + else: + length_tgt = tgt_lengs + length_tgt = length_tgt.clamp(min=0, max=255) + + else: + # predict the length target (greedy for now) + # TODO: implementing length-beam + pred_lengs = length_out.max(-1)[1] + if self.pred_length_offset: + length_tgt = pred_lengs - 128 + src_lengs + else: + length_tgt = pred_lengs + + return length_tgt + + +@register_model_architecture( + "nonautoregressive_transformer", "nonautoregressive_transformer" +) +def base_architecture(args): + args.encoder_embed_path = getattr(args, "encoder_embed_path", None) + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 2048) + args.encoder_layers = getattr(args, "encoder_layers", 6) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 8) + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False) + args.encoder_learned_pos = getattr(args, "encoder_learned_pos", False) + args.decoder_embed_path = getattr(args, "decoder_embed_path", None) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", args.encoder_embed_dim) + args.decoder_ffn_embed_dim = getattr( + args, "decoder_ffn_embed_dim", args.encoder_ffn_embed_dim + ) + args.decoder_layers = getattr(args, "decoder_layers", 6) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 8) + args.decoder_normalize_before = getattr(args, "decoder_normalize_before", False) + args.decoder_learned_pos = getattr(args, "decoder_learned_pos", False) + args.attention_dropout = getattr(args, "attention_dropout", 0.0) + args.activation_dropout = getattr(args, "activation_dropout", 0.0) + args.activation_fn = getattr(args, "activation_fn", "relu") + args.dropout = getattr(args, "dropout", 0.1) + args.adaptive_softmax_cutoff = getattr(args, "adaptive_softmax_cutoff", None) + args.adaptive_softmax_dropout = getattr(args, "adaptive_softmax_dropout", 0) + args.share_decoder_input_output_embed = getattr( + args, "share_decoder_input_output_embed", False + ) + args.share_all_embeddings = getattr(args, "share_all_embeddings", False) + args.no_token_positional_embeddings = getattr( + args, "no_token_positional_embeddings", False + ) + args.adaptive_input = getattr(args, "adaptive_input", False) + args.apply_bert_init = getattr(args, "apply_bert_init", False) + + args.decoder_output_dim = getattr( + args, "decoder_output_dim", args.decoder_embed_dim + ) + args.decoder_input_dim = getattr(args, "decoder_input_dim", args.decoder_embed_dim) + + # --- special arguments --- + args.sg_length_pred = getattr(args, "sg_length_pred", False) + args.pred_length_offset = getattr(args, "pred_length_offset", False) + args.length_loss_factor = getattr(args, "length_loss_factor", 0.1) + args.src_embedding_copy = getattr(args, "src_embedding_copy", False) + + +@register_model_architecture( + "nonautoregressive_transformer", "nonautoregressive_transformer_wmt_en_de" +) +def nonautoregressive_transformer_wmt_en_de(args): + base_architecture(args) diff --git a/fairseq/models/roberta/__init__.py b/fairseq/models/roberta/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..4cd723ae96aec8e3182773483f123109d23b620e --- /dev/null +++ b/fairseq/models/roberta/__init__.py @@ -0,0 +1,11 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .hub_interface import * # noqa +from .model import * # noqa +from .enc_dec import * # noqa +from .model_camembert import * # noqa +from .model_gottbert import * # noqa +from .model_xlmr import * # noqa diff --git a/fairseq/models/roberta/alignment_utils.py b/fairseq/models/roberta/alignment_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..ccc7f74cb94d5b8baa2d4e9dfd44f653d47ee43e --- /dev/null +++ b/fairseq/models/roberta/alignment_utils.py @@ -0,0 +1,118 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from collections import Counter +from typing import List + +import torch + + +def align_bpe_to_words(roberta, bpe_tokens: torch.LongTensor, other_tokens: List[str]): + """ + Helper to align GPT-2 BPE to other tokenization formats (e.g., spaCy). + + Args: + roberta (RobertaHubInterface): RoBERTa instance + bpe_tokens (torch.LongTensor): GPT-2 BPE tokens of shape `(T_bpe)` + other_tokens (List[str]): other tokens of shape `(T_words)` + + Returns: + List[str]: mapping from *other_tokens* to corresponding *bpe_tokens*. + """ + assert bpe_tokens.dim() == 1 + assert bpe_tokens[0] == 0 + + def clean(text): + return text.strip() + + # remove whitespaces to simplify alignment + bpe_tokens = [roberta.task.source_dictionary.string([x]) for x in bpe_tokens] + bpe_tokens = [ + clean(roberta.bpe.decode(x) if x not in {"<s>", ""} else x) for x in bpe_tokens + ] + other_tokens = [clean(str(o)) for o in other_tokens] + + # strip leading <s> + bpe_tokens = bpe_tokens[1:] + assert "".join(bpe_tokens) == "".join(other_tokens) + + # create alignment from every word to a list of BPE tokens + alignment = [] + bpe_toks = filter(lambda item: item[1] != "", enumerate(bpe_tokens, start=1)) + j, bpe_tok = next(bpe_toks) + for other_tok in other_tokens: + bpe_indices = [] + while True: + if other_tok.startswith(bpe_tok): + bpe_indices.append(j) + other_tok = other_tok[len(bpe_tok) :] + try: + j, bpe_tok = next(bpe_toks) + except StopIteration: + j, bpe_tok = None, None + elif bpe_tok.startswith(other_tok): + # other_tok spans multiple BPE tokens + bpe_indices.append(j) + bpe_tok = bpe_tok[len(other_tok) :] + other_tok = "" + else: + raise Exception('Cannot align "{}" and "{}"'.format(other_tok, bpe_tok)) + if other_tok == "": + break + assert len(bpe_indices) > 0 + alignment.append(bpe_indices) + assert len(alignment) == len(other_tokens) + + return alignment + + +def align_features_to_words(roberta, features, alignment): + """ + Align given features to words. + + Args: + roberta (RobertaHubInterface): RoBERTa instance + features (torch.Tensor): features to align of shape `(T_bpe x C)` + alignment: alignment between BPE tokens and words returned by + func:`align_bpe_to_words`. + """ + assert features.dim() == 2 + + bpe_counts = Counter(j for bpe_indices in alignment for j in bpe_indices) + assert bpe_counts[0] == 0 # <s> shouldn't be aligned + denom = features.new([bpe_counts.get(j, 1) for j in range(len(features))]) + weighted_features = features / denom.unsqueeze(-1) + + output = [weighted_features[0]] + largest_j = -1 + for bpe_indices in alignment: + output.append(weighted_features[bpe_indices].sum(dim=0)) + largest_j = max(largest_j, *bpe_indices) + for j in range(largest_j + 1, len(features)): + output.append(weighted_features[j]) + output = torch.stack(output) + assert torch.all(torch.abs(output.sum(dim=0) - features.sum(dim=0)) < 1e-4) + return output + + +def spacy_nlp(): + if getattr(spacy_nlp, "_nlp", None) is None: + try: + from spacy.lang.en import English + + spacy_nlp._nlp = English() + except ImportError: + raise ImportError("Please install spacy with: pip install spacy") + return spacy_nlp._nlp + + +def spacy_tokenizer(): + if getattr(spacy_tokenizer, "_tokenizer", None) is None: + try: + nlp = spacy_nlp() + spacy_tokenizer._tokenizer = nlp.Defaults.create_tokenizer(nlp) + except ImportError: + raise ImportError("Please install spacy with: pip install spacy") + return spacy_tokenizer._tokenizer diff --git a/fairseq/models/roberta/enc_dec.py b/fairseq/models/roberta/enc_dec.py new file mode 100644 index 0000000000000000000000000000000000000000..e538dee0aa5984b1a3d02ce81117d2046c030593 --- /dev/null +++ b/fairseq/models/roberta/enc_dec.py @@ -0,0 +1,192 @@ +import argparse +import logging + +import torch.nn as nn +import fairseq.checkpoint_utils +from fairseq.models import ( + FairseqEncoderDecoderModel, + register_model, + register_model_architecture, +) +from fairseq.models.transformer import TransformerDecoder +from fairseq.models.roberta import model as roberta + +logger = logging.getLogger(__name__) + + +@register_model("roberta_enc_dec") +class RobertaEncDecModel(FairseqEncoderDecoderModel): + @staticmethod + def add_args(parser): + parser.add_argument( + "--pretrained-mlm-checkpoint", + default=None, + type=str, + metavar="PRETRAINED", + help="path to pretrained mlm checkpoint", + ) + parser.add_argument( + "--pretrained-decoder", action="store_true", help="reload decoder" + ) + parser.add_argument( + "--hack-layernorm-embedding", + action="store_true", + help="hack to reload old models trained with encoder-normalize-before=False (no equivalent to encoder-normalize-before=False and layernorm_embedding=False", + ) + parser.add_argument( + "--share-decoder-input-output-embed", + action="store_true", + help="share decoder input and output embeddings", + ) + parser.add_argument( + "--share-all-embeddings", + action="store_true", + help="share encoder, decoder and output embeddings" + " (requires shared dictionary and embed dim)", + ) + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + + # make sure all arguments are present + base_enc_dec_architecture(args) + if args.pretrained_mlm_checkpoint: + arg_overrides = None + if args.hack_layernorm_embedding: + arg_overrides = {"layernorm_embedding": False} + loaded = fairseq.checkpoint_utils.load_model_ensemble_and_task( + [args.pretrained_mlm_checkpoint], arg_overrides=arg_overrides + ) + ([roberta_enc], _cfg, _task) = loaded + else: + # Do we need to edit untie_weights here ? + share_in_out = ( + args.share_decoder_input_output_embed or args.share_all_embeddings + ) + args.untie_weights_roberta = not share_in_out + if args.hack_layernorm_embedding: + args.layernorm_embedding = False + args.encoder_normalize_before = False + roberta_enc = roberta.RobertaModel.build_model(args, task) + + return cls.from_roberta(roberta_enc, args, task.source_dictionary) + + @staticmethod + def from_roberta(roberta_enc: roberta.RobertaModel, args, dictionary): + encoder = roberta_enc.encoder.sentence_encoder + vocab_size, embed_dim = encoder.embed_tokens.weight.shape + + if args.share_all_embeddings: + lm_head = roberta_enc.encoder.lm_head + assert encoder.embed_tokens.weight is lm_head.weight, ( + "Can't use --share-all-embeddings with a model " + "that was pretraiend with --untie-weights-roberta_enc" + ) + else: + lm_head = roberta.RobertaLMHead( + embed_dim, vocab_size, roberta_enc.args.activation_fn + ) + + dec_embs = nn.Embedding(vocab_size, embed_dim, dictionary.pad()) + if args.share_all_embeddings or args.share_decoder_input_output_embed: + # Note: I wasn't able to use Embedding _weight parameter to achive this sharing. + dec_embs.weight = lm_head.weight + + decoder = TransformerDecoder( + RobertaEncDecModel.read_args_from_roberta(roberta_enc.args), + dictionary, + dec_embs, + no_encoder_attn=False, + output_projection=lm_head, + ) + if getattr(args, "pretrained_decoder", False): + decoder_dict = encoder.state_dict() + + # TODO: hide setting "encoder_attn" layers behind a flag. + for k, w in list(decoder_dict.items()): + if ".self_attn" in k: + k_enc_attn = k.replace(".self_attn", ".encoder_attn") + decoder_dict[k_enc_attn] = w.detach().clone() + + for k, w in lm_head.state_dict().items(): + decoder_dict["output_projection." + k] = w + + missing_keys, unexpected_keys = decoder.load_state_dict( + decoder_dict, strict=False + ) + # missing_keys = [m for m in missing_keys if ".encoder_attn" not in m] + assert not missing_keys and not unexpected_keys, ( + "Failed to load state dict. " + f"Missing keys: {missing_keys}. " + f"Unexpected keys: {unexpected_keys}." + ) + + if args.share_all_embeddings: + assert decoder.output_projection.weight is decoder.embed_tokens.weight + assert encoder.embed_tokens.weight is decoder.embed_tokens.weight + elif args.share_decoder_input_output_embed: + assert decoder.output_projection.weight is decoder.embed_tokens.weight + assert encoder.embed_tokens.weight is not decoder.embed_tokens.weight + else: + assert decoder.output_projection.weight is not decoder.embed_tokens.weight + assert encoder.embed_tokens.weight is not decoder.embed_tokens.weight + + return RobertaEncDecModel(encoder, decoder) + + @staticmethod + def read_args_from_roberta(roberta_args: argparse.Namespace): + # TODO: this would become easier if encoder/decoder where using a similar + # TransformerConfig object + args = argparse.Namespace(**vars(roberta_args)) + attr_map = [ + ("encoder_attention_heads", "decoder_attention_heads"), + ("encoder_embed_dim", "decoder_embed_dim"), + ("encoder_embed_dim", "decoder_output_dim"), + ("encoder_normalize_before", "decoder_normalize_before"), + ("encoder_layers_to_keep", "decoder_layers_to_keep"), + ("encoder_ffn_embed_dim", "decoder_ffn_embed_dim"), + ("encoder_layerdrop", "decoder_layerdrop"), + ("encoder_layers", "decoder_layers"), + ("encoder_learned_pos", "decoder_learned_pos"), + # should this be set from here ? + ("max_positions", "max_target_positions"), + ] + for k1, k2 in attr_map: + setattr(args, k2, getattr(roberta_args, k1)) + + args.adaptive_softmax_cutoff = getattr(args, "adaptive_softmax_cutoff", None) + args.adaptive_softmax_dropout = getattr(args, "adaptive_softmax_dropout", 0) + args.share_decoder_input_output_embed = not roberta_args.untie_weights_roberta + return args + + def upgrade_state_dict_named(self, state_dict, name): + prefix = name + "." if name != "" else "" + super().upgrade_state_dict_named(state_dict, name) + old_keys = list(state_dict.keys()) + + # rename decoder -> encoder before upgrading children modules + for k in old_keys: + if k.startswith(prefix + "encoder.lm_head"): + state_dict.pop(k) + continue + new_k = k + new_k = new_k.replace(".sentence_encoder.", ".") + new_k = new_k.replace("decoder.lm_head.", "decoder.output_projection.") + if k == new_k: + continue + # print(k, "->", new_k) + state_dict[new_k] = state_dict.pop(k) + + +@register_model_architecture("roberta_enc_dec", "roberta_enc_dec") +def base_enc_dec_architecture(args): + args.hack_layernorm_embedding = getattr(args, "hack_layernorm_embedding", False) + args.pretrained_mlm_checkpoint = getattr(args, "pretrained_mlm_checkpoint", None) + args.pretrained_decoder = getattr(args, "pretrained_decoder", None) + args.share_all_embeddings = getattr(args, "share_all_embeddings", False) + args.share_decoder_input_output_embed = getattr( + args, "share_decoder_input_output_embed", False + ) + + roberta.base_architecture(args) diff --git a/fairseq/models/roberta/hub_interface.py b/fairseq/models/roberta/hub_interface.py new file mode 100644 index 0000000000000000000000000000000000000000..c9af434bde61f399a4eebaafd5811be9a37d538e --- /dev/null +++ b/fairseq/models/roberta/hub_interface.py @@ -0,0 +1,235 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import numpy as np +import torch +import torch.nn as nn +import torch.nn.functional as F +from fairseq import utils +from fairseq.data import encoders + + +class RobertaHubInterface(nn.Module): + """A simple PyTorch Hub interface to RoBERTa. + + Usage: https://github.com/pytorch/fairseq/tree/master/examples/roberta + """ + + def __init__(self, cfg, task, model): + super().__init__() + self.cfg = cfg + self.task = task + self.model = model + + self.bpe = encoders.build_bpe(cfg.bpe) + + # this is useful for determining the device + self.register_buffer("_float_tensor", torch.tensor([0], dtype=torch.float)) + + @property + def device(self): + return self._float_tensor.device + + def encode( + self, sentence: str, *addl_sentences, no_separator=False + ) -> torch.LongTensor: + """ + BPE-encode a sentence (or multiple sentences). + + Every sequence begins with a beginning-of-sentence (`<s>`) symbol. + Every sentence ends with an end-of-sentence (`</s>`) and we use an + extra end-of-sentence (`</s>`) as a separator. + + Example (single sentence): `<s> a b c </s>` + Example (sentence pair): `<s> d e f </s> </s> 1 2 3 </s>` + + The BPE encoding follows GPT-2. One subtle detail is that the GPT-2 BPE + requires leading spaces. For example:: + + >>> roberta.encode('Hello world').tolist() + [0, 31414, 232, 2] + >>> roberta.encode(' world').tolist() + [0, 232, 2] + >>> roberta.encode('world').tolist() + [0, 8331, 2] + """ + bpe_sentence = "<s> " + self.bpe.encode(sentence) + " </s>" + for s in addl_sentences: + bpe_sentence += " </s>" if not no_separator else "" + bpe_sentence += " " + self.bpe.encode(s) + " </s>" + tokens = self.task.source_dictionary.encode_line( + bpe_sentence, append_eos=False, add_if_not_exist=False + ) + return tokens.long() + + def decode(self, tokens: torch.LongTensor): + assert tokens.dim() == 1 + tokens = tokens.numpy() + if tokens[0] == self.task.source_dictionary.bos(): + tokens = tokens[1:] # remove <s> + eos_mask = tokens == self.task.source_dictionary.eos() + doc_mask = eos_mask[1:] & eos_mask[:-1] + sentences = np.split(tokens, doc_mask.nonzero()[0] + 1) + sentences = [ + self.bpe.decode(self.task.source_dictionary.string(s)) for s in sentences + ] + if len(sentences) == 1: + return sentences[0] + return sentences + + def extract_features( + self, tokens: torch.LongTensor, return_all_hiddens: bool = False + ) -> torch.Tensor: + if tokens.dim() == 1: + tokens = tokens.unsqueeze(0) + if tokens.size(-1) > self.model.max_positions(): + raise ValueError( + "tokens exceeds maximum length: {} > {}".format( + tokens.size(-1), self.model.max_positions() + ) + ) + features, extra = self.model( + tokens.to(device=self.device), + features_only=True, + return_all_hiddens=return_all_hiddens, + ) + if return_all_hiddens: + # convert from T x B x C -> B x T x C + inner_states = extra["inner_states"] + return [inner_state.transpose(0, 1) for inner_state in inner_states] + else: + return features # just the last layer's features + + def register_classification_head( + self, name: str, num_classes: int = None, embedding_size: int = None, **kwargs + ): + self.model.register_classification_head( + name, num_classes=num_classes, embedding_size=embedding_size, **kwargs + ) + + def predict(self, head: str, tokens: torch.LongTensor, return_logits: bool = False): + features = self.extract_features(tokens.to(device=self.device)) + logits = self.model.classification_heads[head](features) + if return_logits: + return logits + return F.log_softmax(logits, dim=-1) + + def extract_features_aligned_to_words( + self, sentence: str, return_all_hiddens: bool = False + ) -> torch.Tensor: + """Extract RoBERTa features, aligned to spaCy's word-level tokenizer.""" + from fairseq.models.roberta import alignment_utils + from spacy.tokens import Doc + + nlp = alignment_utils.spacy_nlp() + tokenizer = alignment_utils.spacy_tokenizer() + + # tokenize both with GPT-2 BPE and spaCy + bpe_toks = self.encode(sentence) + spacy_toks = tokenizer(sentence) + spacy_toks_ws = [t.text_with_ws for t in tokenizer(sentence)] + alignment = alignment_utils.align_bpe_to_words(self, bpe_toks, spacy_toks_ws) + + # extract features and align them + features = self.extract_features( + bpe_toks, return_all_hiddens=return_all_hiddens + ) + features = features.squeeze(0) + aligned_feats = alignment_utils.align_features_to_words( + self, features, alignment + ) + + # wrap in spaCy Doc + doc = Doc( + nlp.vocab, + words=["<s>"] + [x.text for x in spacy_toks] + ["</s>"], + spaces=[True] + + [x.endswith(" ") for x in spacy_toks_ws[:-1]] + + [True, False], + ) + assert len(doc) == aligned_feats.size(0) + doc.user_token_hooks["vector"] = lambda token: aligned_feats[token.i] + return doc + + def fill_mask(self, masked_input: str, topk: int = 5): + masked_token = "<mask>" + assert ( + masked_token in masked_input and masked_input.count(masked_token) == 1 + ), "Please add one {0} token for the input, eg: 'He is a {0} guy'".format( + masked_token + ) + + text_spans = masked_input.split(masked_token) + text_spans_bpe = ( + (" {0} ".format(masked_token)) + .join([self.bpe.encode(text_span.rstrip()) for text_span in text_spans]) + .strip() + ) + tokens = self.task.source_dictionary.encode_line( + "<s> " + text_spans_bpe + " </s>", + append_eos=False, + add_if_not_exist=False, + ) + + masked_index = (tokens == self.task.mask_idx).nonzero(as_tuple=False) + if tokens.dim() == 1: + tokens = tokens.unsqueeze(0) + + with utils.model_eval(self.model): + features, extra = self.model( + tokens.long().to(device=self.device), + features_only=False, + return_all_hiddens=False, + ) + logits = features[0, masked_index, :].squeeze() + prob = logits.softmax(dim=0) + values, index = prob.topk(k=topk, dim=0) + topk_predicted_token_bpe = self.task.source_dictionary.string(index) + + topk_filled_outputs = [] + for index, predicted_token_bpe in enumerate( + topk_predicted_token_bpe.split(" ") + ): + predicted_token = self.bpe.decode(predicted_token_bpe) + # Quick hack to fix https://github.com/pytorch/fairseq/issues/1306 + if predicted_token_bpe.startswith("\u2581"): + predicted_token = " " + predicted_token + if " {0}".format(masked_token) in masked_input: + topk_filled_outputs.append( + ( + masked_input.replace( + " {0}".format(masked_token), predicted_token + ), + values[index].item(), + predicted_token, + ) + ) + else: + topk_filled_outputs.append( + ( + masked_input.replace(masked_token, predicted_token), + values[index].item(), + predicted_token, + ) + ) + return topk_filled_outputs + + def disambiguate_pronoun(self, sentence: str) -> bool: + """ + Usage:: + + >>> disambiguate_pronoun('The _trophy_ would not fit in the brown suitcase because [it] was too big.') + True + + >>> disambiguate_pronoun('The trophy would not fit in the brown suitcase because [it] was too big.') + 'The trophy' + """ + assert hasattr( + self.task, "disambiguate_pronoun" + ), "roberta.disambiguate_pronoun() requires a model trained with the WSC task." + with utils.model_eval(self.model): + return self.task.disambiguate_pronoun( + self.model, sentence, use_cuda=self.device.type == "cuda" + ) diff --git a/fairseq/models/roberta/model.py b/fairseq/models/roberta/model.py new file mode 100644 index 0000000000000000000000000000000000000000..d9d0f324cf708a49a9d97ef05621dd1eb9bdefc8 --- /dev/null +++ b/fairseq/models/roberta/model.py @@ -0,0 +1,582 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +RoBERTa: A Robustly Optimized BERT Pretraining Approach. +""" + +import logging + +import torch +import torch.nn as nn +import torch.nn.functional as F +from fairseq import utils +from fairseq.models import ( + FairseqEncoder, + FairseqEncoderModel, + register_model, + register_model_architecture, +) +from fairseq.models.transformer import DEFAULT_MIN_PARAMS_TO_WRAP, TransformerEncoder +from fairseq.modules import LayerNorm +from fairseq.modules.quant_noise import quant_noise as apply_quant_noise_ +from fairseq.modules.transformer_sentence_encoder import init_bert_params + +from .hub_interface import RobertaHubInterface + + +logger = logging.getLogger(__name__) + + +@register_model("roberta") +class RobertaModel(FairseqEncoderModel): + @classmethod + def hub_models(cls): + return { + "roberta.base": "http://dl.fbaipublicfiles.com/fairseq/models/roberta.base.tar.gz", + "roberta.large": "http://dl.fbaipublicfiles.com/fairseq/models/roberta.large.tar.gz", + "roberta.large.mnli": "http://dl.fbaipublicfiles.com/fairseq/models/roberta.large.mnli.tar.gz", + "roberta.large.wsc": "http://dl.fbaipublicfiles.com/fairseq/models/roberta.large.wsc.tar.gz", + } + + def __init__(self, args, encoder): + super().__init__(encoder) + self.args = args + + # We follow BERT's random weight initialization + self.apply(init_bert_params) + + self.classification_heads = nn.ModuleDict() + + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + parser.add_argument( + "--encoder-layers", type=int, metavar="L", help="num encoder layers" + ) + parser.add_argument( + "--encoder-embed-dim", + type=int, + metavar="H", + help="encoder embedding dimension", + ) + parser.add_argument( + "--encoder-ffn-embed-dim", + type=int, + metavar="F", + help="encoder embedding dimension for FFN", + ) + parser.add_argument( + "--encoder-attention-heads", + type=int, + metavar="A", + help="num encoder attention heads", + ) + parser.add_argument( + "--activation-fn", + choices=utils.get_available_activation_fns(), + help="activation function to use", + ) + parser.add_argument( + "--pooler-activation-fn", + choices=utils.get_available_activation_fns(), + help="activation function to use for pooler layer", + ) + parser.add_argument( + "--encoder-normalize-before", + action="store_true", + help="apply layernorm before each encoder block", + ) + parser.add_argument( + "--layernorm-embedding", + action="store_true", + help="add layernorm to embedding", + ) + parser.add_argument( + "--dropout", type=float, metavar="D", help="dropout probability" + ) + parser.add_argument( + "--attention-dropout", + type=float, + metavar="D", + help="dropout probability for attention weights", + ) + parser.add_argument( + "--activation-dropout", + type=float, + metavar="D", + help="dropout probability after activation in FFN", + ) + parser.add_argument( + "--pooler-dropout", + type=float, + metavar="D", + help="dropout probability in the masked_lm pooler layers", + ) + parser.add_argument( + "--max-positions", type=int, help="number of positional embeddings to learn" + ) + parser.add_argument( + "--load-checkpoint-heads", + action="store_true", + help="(re-)register and load heads when loading checkpoints", + ) + parser.add_argument( + "--untie-weights-roberta", + action="store_true", + help="Untie weights between embeddings and classifiers in RoBERTa", + ) + # args for "Reducing Transformer Depth on Demand with Structured Dropout" (Fan et al., 2019) + parser.add_argument( + "--encoder-layerdrop", + type=float, + metavar="D", + default=0, + help="LayerDrop probability for encoder", + ) + parser.add_argument( + "--encoder-layers-to-keep", + default=None, + help="which layers to *keep* when pruning as a comma-separated list", + ) + # args for Training with Quantization Noise for Extreme Model Compression ({Fan*, Stock*} et al., 2020) + parser.add_argument( + "--quant-noise-pq", + type=float, + metavar="D", + default=0, + help="iterative PQ quantization noise at training time", + ) + parser.add_argument( + "--quant-noise-pq-block-size", + type=int, + metavar="D", + default=8, + help="block size of quantization noise at training time", + ) + parser.add_argument( + "--quant-noise-scalar", + type=float, + metavar="D", + default=0, + help="scalar quantization noise and scalar quantization at training time", + ) + # args for "Better Fine-Tuning by Reducing Representational Collapse" (Aghajanyan et al. 2020) + parser.add_argument( + "--spectral-norm-classification-head", + action="store_true", + default=False, + help="Apply spectral normalization on the classification head", + ) + # args for Fully Sharded Data Parallel (FSDP) training + parser.add_argument( + "--min-params-to-wrap", + type=int, + metavar="D", + default=DEFAULT_MIN_PARAMS_TO_WRAP, + help=( + "minimum number of params for a layer to be wrapped with FSDP() when " + "training with --ddp-backend=fully_sharded. Smaller values will " + "improve memory efficiency, but may make torch.distributed " + "communication less efficient due to smaller input sizes. This option " + "is set to 0 (i.e., always wrap) when --checkpoint-activations or " + "--offload-activations are passed." + ) + ) + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + + # make sure all arguments are present + base_architecture(args) + + if not hasattr(args, "max_positions"): + args.max_positions = args.tokens_per_sample + + encoder = RobertaEncoder(args, task.source_dictionary) + return cls(args, encoder) + + def forward( + self, + src_tokens, + features_only=False, + return_all_hiddens=False, + classification_head_name=None, + **kwargs, + ): + if classification_head_name is not None: + features_only = True + + x, extra = self.encoder(src_tokens, features_only, return_all_hiddens, **kwargs) + + if classification_head_name is not None: + x = self.classification_heads[classification_head_name](x) + return x, extra + + def get_normalized_probs(self, net_output, log_probs, sample=None): + """Get normalized probabilities (or log probs) from a net's output.""" + logits = net_output[0].float() + if log_probs: + return F.log_softmax(logits, dim=-1) + else: + return F.softmax(logits, dim=-1) + + def register_classification_head( + self, name, num_classes=None, inner_dim=None, **kwargs + ): + """Register a classification head.""" + if name in self.classification_heads: + prev_num_classes = self.classification_heads[name].out_proj.out_features + prev_inner_dim = self.classification_heads[name].dense.out_features + if num_classes != prev_num_classes or inner_dim != prev_inner_dim: + logger.warning( + 're-registering head "{}" with num_classes {} (prev: {}) ' + "and inner_dim {} (prev: {})".format( + name, num_classes, prev_num_classes, inner_dim, prev_inner_dim + ) + ) + self.classification_heads[name] = RobertaClassificationHead( + input_dim=self.args.encoder_embed_dim, + inner_dim=inner_dim or self.args.encoder_embed_dim, + num_classes=num_classes, + activation_fn=self.args.pooler_activation_fn, + pooler_dropout=self.args.pooler_dropout, + q_noise=self.args.quant_noise_pq, + qn_block_size=self.args.quant_noise_pq_block_size, + do_spectral_norm=self.args.spectral_norm_classification_head, + ) + + @property + def supported_targets(self): + return {"self"} + + @classmethod + def from_pretrained( + cls, + model_name_or_path, + checkpoint_file="model.pt", + data_name_or_path=".", + bpe="gpt2", + **kwargs, + ): + from fairseq import hub_utils + + x = hub_utils.from_pretrained( + model_name_or_path, + checkpoint_file, + data_name_or_path, + archive_map=cls.hub_models(), + bpe=bpe, + load_checkpoint_heads=True, + **kwargs, + ) + + logger.info(x["args"]) + return RobertaHubInterface(x["args"], x["task"], x["models"][0]) + + def upgrade_state_dict_named(self, state_dict, name): + prefix = name + "." if name != "" else "" + + # rename decoder -> encoder before upgrading children modules + for k in list(state_dict.keys()): + if k.startswith(prefix + "decoder"): + new_k = prefix + "encoder" + k[len(prefix + "decoder") :] + state_dict[new_k] = state_dict[k] + del state_dict[k] + + # rename emb_layer_norm -> layernorm_embedding + for k in list(state_dict.keys()): + if ".emb_layer_norm." in k: + new_k = k.replace(".emb_layer_norm.", ".layernorm_embedding.") + state_dict[new_k] = state_dict[k] + del state_dict[k] + + # upgrade children modules + super().upgrade_state_dict_named(state_dict, name) + + # Handle new classification heads present in the state dict. + current_head_names = ( + [] + if not hasattr(self, "classification_heads") + else self.classification_heads.keys() + ) + keys_to_delete = [] + for k in state_dict.keys(): + if not k.startswith(prefix + "classification_heads."): + continue + + head_name = k[len(prefix + "classification_heads.") :].split(".")[0] + num_classes = state_dict[ + prefix + "classification_heads." + head_name + ".out_proj.weight" + ].size(0) + inner_dim = state_dict[ + prefix + "classification_heads." + head_name + ".dense.weight" + ].size(0) + + if getattr(self.args, "load_checkpoint_heads", False): + if head_name not in current_head_names: + self.register_classification_head(head_name, num_classes, inner_dim) + else: + if head_name not in current_head_names: + logger.warning( + "deleting classification head ({}) from checkpoint " + "not present in current model: {}".format(head_name, k) + ) + keys_to_delete.append(k) + elif ( + num_classes + != self.classification_heads[head_name].out_proj.out_features + or inner_dim + != self.classification_heads[head_name].dense.out_features + ): + logger.warning( + "deleting classification head ({}) from checkpoint " + "with different dimensions than current model: {}".format( + head_name, k + ) + ) + keys_to_delete.append(k) + for k in keys_to_delete: + del state_dict[k] + + # Copy any newly-added classification heads into the state dict + # with their current weights. + if hasattr(self, "classification_heads"): + cur_state = self.classification_heads.state_dict() + for k, v in cur_state.items(): + if prefix + "classification_heads." + k not in state_dict: + logger.info("Overwriting " + prefix + "classification_heads." + k) + state_dict[prefix + "classification_heads." + k] = v + + +class RobertaLMHead(nn.Module): + """Head for masked language modeling.""" + + def __init__(self, embed_dim, output_dim, activation_fn, weight=None): + super().__init__() + self.dense = nn.Linear(embed_dim, embed_dim) + self.activation_fn = utils.get_activation_fn(activation_fn) + self.layer_norm = LayerNorm(embed_dim) + + if weight is None: + weight = nn.Linear(embed_dim, output_dim, bias=False).weight + self.weight = weight + self.bias = nn.Parameter(torch.zeros(output_dim)) + + def forward(self, features, masked_tokens=None, **kwargs): + # Only project the masked tokens while training, + # saves both memory and computation + if masked_tokens is not None: + features = features[masked_tokens, :] + + x = self.dense(features) + x = self.activation_fn(x) + x = self.layer_norm(x) + # project back to size of vocabulary with bias + x = F.linear(x, self.weight) + self.bias + return x + + +class RobertaClassificationHead(nn.Module): + """Head for sentence-level classification tasks.""" + + def __init__( + self, + input_dim, + inner_dim, + num_classes, + activation_fn, + pooler_dropout, + q_noise=0, + qn_block_size=8, + do_spectral_norm=False, + ): + super().__init__() + self.dense = nn.Linear(input_dim, inner_dim) + self.activation_fn = utils.get_activation_fn(activation_fn) + self.dropout = nn.Dropout(p=pooler_dropout) + self.out_proj = apply_quant_noise_( + nn.Linear(inner_dim, num_classes), q_noise, qn_block_size + ) + if do_spectral_norm: + if q_noise != 0: + raise NotImplementedError( + "Attempting to use Spectral Normalization with Quant Noise. This is not officially supported" + ) + self.out_proj = torch.nn.utils.spectral_norm(self.out_proj) + + def forward(self, features, **kwargs): + x = features[:, 0, :] # take <s> token (equiv. to [CLS]) + x = self.dropout(x) + x = self.dense(x) + x = self.activation_fn(x) + x = self.dropout(x) + x = self.out_proj(x) + return x + + +class RobertaEncoder(FairseqEncoder): + """RoBERTa encoder.""" + + def __init__(self, args, dictionary): + super().__init__(dictionary) + + # set any missing default values + base_architecture(args) + self.args = args + + if args.encoder_layers_to_keep: + args.encoder_layers = len(args.encoder_layers_to_keep.split(",")) + + embed_tokens = self.build_embedding( + len(dictionary), args.encoder_embed_dim, dictionary.pad() + ) + + self.sentence_encoder = self.build_encoder(args, dictionary, embed_tokens) + + self.lm_head = self.build_lm_head( + embed_dim=args.encoder_embed_dim, + output_dim=len(dictionary), + activation_fn=args.activation_fn, + weight=( + self.sentence_encoder.embed_tokens.weight + if not args.untie_weights_roberta + else None + ), + ) + + def build_embedding(self, vocab_size, embedding_dim, padding_idx): + return nn.Embedding(vocab_size, embedding_dim, padding_idx) + + def build_encoder(self, args, dictionary, embed_tokens): + encoder = TransformerEncoder(args, dictionary, embed_tokens) + encoder.apply(init_bert_params) + return encoder + + def build_lm_head(self, embed_dim, output_dim, activation_fn, weight): + return RobertaLMHead(embed_dim, output_dim, activation_fn, weight) + + def forward( + self, + src_tokens, + features_only=False, + return_all_hiddens=False, + masked_tokens=None, + **unused, + ): + """ + Args: + src_tokens (LongTensor): input tokens of shape `(batch, src_len)` + features_only (bool, optional): skip LM head and just return + features. If True, the output will be of shape + `(batch, src_len, embed_dim)`. + return_all_hiddens (bool, optional): also return all of the + intermediate hidden states (default: False). + + Returns: + tuple: + - the LM output of shape `(batch, src_len, vocab)` + - a dictionary of additional data, where 'inner_states' + is a list of hidden states. Note that the hidden + states have shape `(src_len, batch, vocab)`. + """ + x, extra = self.extract_features( + src_tokens, return_all_hiddens=return_all_hiddens + ) + if not features_only: + x = self.output_layer(x, masked_tokens=masked_tokens) + return x, extra + + def extract_features(self, src_tokens, return_all_hiddens=False, **kwargs): + encoder_out = self.sentence_encoder( + src_tokens, + return_all_hiddens=return_all_hiddens, + token_embeddings=kwargs.get("token_embeddings", None), + ) + # T x B x C -> B x T x C + features = encoder_out["encoder_out"][0].transpose(0, 1) + inner_states = encoder_out["encoder_states"] if return_all_hiddens else None + return features, {"inner_states": inner_states} + + def output_layer(self, features, masked_tokens=None, **unused): + return self.lm_head(features, masked_tokens) + + def max_positions(self): + """Maximum output length supported by the encoder.""" + return self.args.max_positions + + +@register_model_architecture("roberta", "roberta") +def base_architecture(args): + args.encoder_layers = getattr(args, "encoder_layers", 12) + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 768) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 3072) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 12) + + args.dropout = getattr(args, "dropout", 0.1) + args.attention_dropout = getattr(args, "attention_dropout", 0.1) + args.activation_dropout = getattr(args, "activation_dropout", 0.0) + args.pooler_dropout = getattr(args, "pooler_dropout", 0.0) + + args.max_source_positions = getattr(args, "max_positions", 512) + args.no_token_positional_embeddings = getattr( + args, "no_token_positional_embeddings", False + ) + + # BERT has a few structural differences compared to the original Transformer + args.encoder_learned_pos = getattr(args, "encoder_learned_pos", True) + args.layernorm_embedding = getattr(args, "layernorm_embedding", True) + args.no_scale_embedding = getattr(args, "no_scale_embedding", True) + args.activation_fn = getattr(args, "activation_fn", "gelu") + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False) + args.pooler_activation_fn = getattr(args, "pooler_activation_fn", "tanh") + args.untie_weights_roberta = getattr(args, "untie_weights_roberta", False) + + # Adaptive input config + args.adaptive_input = getattr(args, "adaptive_input", False) + + # LayerDrop config + args.encoder_layerdrop = getattr(args, "encoder_layerdrop", 0.0) + args.encoder_layers_to_keep = getattr(args, "encoder_layers_to_keep", None) + + # Quantization noise config + args.quant_noise_pq = getattr(args, "quant_noise_pq", 0) + args.quant_noise_pq_block_size = getattr(args, "quant_noise_pq_block_size", 8) + args.quant_noise_scalar = getattr(args, "quant_noise_scalar", 0) + + # R4F config + args.spectral_norm_classification_head = getattr( + args, "spectral_norm_classification_head", False + ) + + +@register_model_architecture("roberta", "roberta_prenorm") +def roberta_prenorm_architecture(args): + args.layernorm_embedding = getattr(args, "layernorm_embedding", False) + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", True) + base_architecture(args) + + +@register_model_architecture("roberta", "roberta_base") +def roberta_base_architecture(args): + base_architecture(args) + + +@register_model_architecture("roberta", "roberta_large") +def roberta_large_architecture(args): + args.encoder_layers = getattr(args, "encoder_layers", 24) + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 1024) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 4096) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 16) + base_architecture(args) + + +@register_model_architecture("roberta", "xlm") +def xlm_architecture(args): + args.encoder_layers = getattr(args, "encoder_layers", 16) + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 1280) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 1280 * 4) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 16) + base_architecture(args) diff --git a/fairseq/models/roberta/model_camembert.py b/fairseq/models/roberta/model_camembert.py new file mode 100644 index 0000000000000000000000000000000000000000..46447546fafb4a0a887b481022cac07631047c80 --- /dev/null +++ b/fairseq/models/roberta/model_camembert.py @@ -0,0 +1,50 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +CamemBERT: a Tasty French Language Model +""" + +from fairseq.models import register_model + +from .hub_interface import RobertaHubInterface +from .model import RobertaModel + + +@register_model("camembert") +class CamembertModel(RobertaModel): + @classmethod + def hub_models(cls): + return { + "camembert": "http://dl.fbaipublicfiles.com/fairseq/models/camembert-base.tar.gz", + "camembert.v0": "http://dl.fbaipublicfiles.com/fairseq/models/camembert-base.tar.gz", + "camembert-base": "http://dl.fbaipublicfiles.com/fairseq/models/camembert-base.tar.gz", + "camembert-large": "http://dl.fbaipublicfiles.com/fairseq/models/camembert-large.tar.gz", + "camembert-base-ccnet": "http://dl.fbaipublicfiles.com/fairseq/models/camembert-base-ccnet.tar.gz", + "camembert-base-ccnet-4gb": "http://dl.fbaipublicfiles.com/fairseq/models/camembert-base-ccnet-4gb.tar.gz", + "camembert-base-wikipedia-4gb": "http://dl.fbaipublicfiles.com/fairseq/models/camembert-base-wikipedia-4gb.tar.gz", + "camembert-base-oscar-4gb": "http://dl.fbaipublicfiles.com/fairseq/models/camembert-base-oscar-4gb.tar.gz", + } + + @classmethod + def from_pretrained( + cls, + model_name_or_path, + checkpoint_file="model.pt", + data_name_or_path=".", + bpe="sentencepiece", + **kwargs + ): + from fairseq import hub_utils + + x = hub_utils.from_pretrained( + model_name_or_path, + checkpoint_file, + data_name_or_path, + archive_map=cls.hub_models(), + bpe=bpe, + load_checkpoint_heads=True, + **kwargs, + ) + return RobertaHubInterface(x["args"], x["task"], x["models"][0]) diff --git a/fairseq/models/roberta/model_gottbert.py b/fairseq/models/roberta/model_gottbert.py new file mode 100644 index 0000000000000000000000000000000000000000..2e8c66354ac7ce7309226bb091a7baa4776fbfdc --- /dev/null +++ b/fairseq/models/roberta/model_gottbert.py @@ -0,0 +1,49 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +GottBERT: a pure German Language Model +""" + +from fairseq.models import register_model + +from .hub_interface import RobertaHubInterface +from .model import RobertaModel + + +@register_model('gottbert') +class GottbertModel(RobertaModel): + + @classmethod + def hub_models(cls): + return { + 'gottbert-base': 'https://dl.gottbert.de/fairseq/models/gottbert-base.tar.gz', + } + + @classmethod + def from_pretrained(cls, + model_name_or_path, + checkpoint_file='model.pt', + data_name_or_path='.', + bpe='hf_byte_bpe', + bpe_vocab='vocab.json', + bpe_merges='merges.txt', + bpe_add_prefix_space=False, + **kwargs + ): + from fairseq import hub_utils + + x = hub_utils.from_pretrained( + model_name_or_path, + checkpoint_file, + data_name_or_path, + archive_map=cls.hub_models(), + bpe=bpe, + load_checkpoint_heads=True, + bpe_vocab=bpe_vocab, + bpe_merges=bpe_merges, + bpe_add_prefix_space=bpe_add_prefix_space, + **kwargs, + ) + return RobertaHubInterface(x['args'], x['task'], x['models'][0]) diff --git a/fairseq/models/roberta/model_xlmr.py b/fairseq/models/roberta/model_xlmr.py new file mode 100644 index 0000000000000000000000000000000000000000..cf6e354d53b918dd4c7c78bfcd38ac0d63cab3bd --- /dev/null +++ b/fairseq/models/roberta/model_xlmr.py @@ -0,0 +1,46 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +Unsupervised Cross-lingual Representation Learning at Scale +""" + +from fairseq.models import register_model + +from .hub_interface import RobertaHubInterface +from .model import RobertaModel + + +@register_model("xlmr") +class XLMRModel(RobertaModel): + @classmethod + def hub_models(cls): + return { + "xlmr.base": "http://dl.fbaipublicfiles.com/fairseq/models/xlmr.base.tar.gz", + "xlmr.large": "http://dl.fbaipublicfiles.com/fairseq/models/xlmr.large.tar.gz", + "xlmr.xl": "http://dl.fbaipublicfiles.com/fairseq/models/xlmr/xlmr.xl.tar.gz", + "xlmr.xxl": "http://dl.fbaipublicfiles.com/fairseq/models/xlmr/xlmr.xxl.tar.gz", + } + + @classmethod + def from_pretrained( + cls, + model_name_or_path, + checkpoint_file="model.pt", + data_name_or_path=".", + bpe="sentencepiece", + **kwargs + ): + from fairseq import hub_utils + + x = hub_utils.from_pretrained( + model_name_or_path, + checkpoint_file, + data_name_or_path, + archive_map=cls.hub_models(), + bpe=bpe, + load_checkpoint_heads=True, + **kwargs, + ) + return RobertaHubInterface(x["args"], x["task"], x["models"][0]) diff --git a/fairseq/models/speech_to_text/__init__.py b/fairseq/models/speech_to_text/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..c6ae9b17ba37a228163fddcb6fed199e61ef02c8 --- /dev/null +++ b/fairseq/models/speech_to_text/__init__.py @@ -0,0 +1,8 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .berard import * # noqa +from .convtransformer import * # noqa +from .s2t_transformer import * # noqa diff --git a/fairseq/models/speech_to_text/berard.py b/fairseq/models/speech_to_text/berard.py new file mode 100644 index 0000000000000000000000000000000000000000..c505e3acaa84e5f3263ccbfaf9556f77123f09fc --- /dev/null +++ b/fairseq/models/speech_to_text/berard.py @@ -0,0 +1,606 @@ +#!/usr/bin/env python3 + +from ast import literal_eval +from typing import List, Tuple + +import torch +import torch.nn as nn +import torch.nn.functional as F +from fairseq import checkpoint_utils, utils +from fairseq.data.data_utils import lengths_to_padding_mask +from fairseq.models import ( + FairseqEncoder, + FairseqEncoderDecoderModel, + FairseqIncrementalDecoder, + register_model, + register_model_architecture, +) + + +@register_model("s2t_berard") +class BerardModel(FairseqEncoderDecoderModel): + """Implementation of a model similar to https://arxiv.org/abs/1802.04200 + + Paper title: End-to-End Automatic Speech Translation of Audiobooks + An implementation is available in tensorflow at + https://github.com/eske/seq2seq + Relevant files in this implementation are the config + (https://github.com/eske/seq2seq/blob/master/config/LibriSpeech/AST.yaml) + and the model code + (https://github.com/eske/seq2seq/blob/master/translate/models.py). + The encoder and decoder try to be close to the original implementation. + The attention is an MLP as in Bahdanau et al. + (https://arxiv.org/abs/1409.0473). + There is no state initialization by averaging the encoder outputs. + """ + + def __init__(self, encoder, decoder): + super().__init__(encoder, decoder) + + @staticmethod + def add_args(parser): + parser.add_argument( + "--input-layers", + type=str, + metavar="EXPR", + help="List of linear layer dimensions. These " + "layers are applied to the input features and " + "are followed by tanh and possibly dropout.", + ) + parser.add_argument( + "--dropout", + type=float, + metavar="D", + help="Dropout probability to use in the encoder/decoder. " + "Note that this parameters control dropout in various places, " + "there is no fine-grained control for dropout for embeddings " + "vs LSTM layers for example.", + ) + parser.add_argument( + "--in-channels", + type=int, + metavar="N", + help="Number of encoder input channels. " "Typically value is 1.", + ) + parser.add_argument( + "--conv-layers", + type=str, + metavar="EXPR", + help="List of conv layers " "(format: (channels, kernel, stride)).", + ) + parser.add_argument( + "--num-blstm-layers", + type=int, + metavar="N", + help="Number of encoder bi-LSTM layers.", + ) + parser.add_argument( + "--lstm-size", type=int, metavar="N", help="LSTM hidden size." + ) + parser.add_argument( + "--decoder-embed-dim", + type=int, + metavar="N", + help="Embedding dimension of the decoder target tokens.", + ) + parser.add_argument( + "--decoder-hidden-dim", + type=int, + metavar="N", + help="Decoder LSTM hidden dimension.", + ) + parser.add_argument( + "--decoder-num-layers", + type=int, + metavar="N", + help="Number of decoder LSTM layers.", + ) + parser.add_argument( + "--attention-dim", + type=int, + metavar="N", + help="Hidden layer dimension in MLP attention.", + ) + parser.add_argument( + "--output-layer-dim", + type=int, + metavar="N", + help="Hidden layer dim for linear layer prior to output projection.", + ) + parser.add_argument( + "--load-pretrained-encoder-from", + type=str, + metavar="STR", + help="model to take encoder weights from (for initialization)", + ) + parser.add_argument( + "--load-pretrained-decoder-from", + type=str, + metavar="STR", + help="model to take decoder weights from (for initialization)", + ) + + @classmethod + def build_encoder(cls, args, task): + encoder = BerardEncoder( + input_layers=literal_eval(args.input_layers), + conv_layers=literal_eval(args.conv_layers), + in_channels=args.input_channels, + input_feat_per_channel=args.input_feat_per_channel, + num_blstm_layers=args.num_blstm_layers, + lstm_size=args.lstm_size, + dropout=args.dropout, + ) + if getattr(args, "load_pretrained_encoder_from", None): + encoder = checkpoint_utils.load_pretrained_component_from_model( + component=encoder, checkpoint=args.load_pretrained_encoder_from + ) + return encoder + + @classmethod + def build_decoder(cls, args, task): + decoder = LSTMDecoder( + dictionary=task.target_dictionary, + embed_dim=args.decoder_embed_dim, + num_layers=args.decoder_num_layers, + hidden_size=args.decoder_hidden_dim, + dropout=args.dropout, + encoder_output_dim=2 * args.lstm_size, # bidirectional + attention_dim=args.attention_dim, + output_layer_dim=args.output_layer_dim, + ) + if getattr(args, "load_pretrained_decoder_from", None): + decoder = checkpoint_utils.load_pretrained_component_from_model( + component=decoder, checkpoint=args.load_pretrained_decoder_from + ) + return decoder + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + encoder = cls.build_encoder(args, task) + decoder = cls.build_decoder(args, task) + + return cls(encoder, decoder) + + def get_normalized_probs(self, net_output, log_probs, sample=None): + # net_output['encoder_out'] is a (B, T, D) tensor + lprobs = super().get_normalized_probs(net_output, log_probs, sample) + # lprobs is a (B, T, D) tensor + lprobs.batch_first = True + return lprobs + + +class BerardEncoder(FairseqEncoder): + def __init__( + self, + input_layers: List[int], + conv_layers: List[Tuple[int]], + in_channels: int, + input_feat_per_channel: int, + num_blstm_layers: int, + lstm_size: int, + dropout: float, + ): + """ + Args: + input_layers: list of linear layer dimensions. These layers are + applied to the input features and are followed by tanh and + possibly dropout. + conv_layers: list of conv2d layer configurations. A configuration is + a tuple (out_channels, conv_kernel_size, stride). + in_channels: number of input channels. + input_feat_per_channel: number of input features per channel. These + are speech features, typically 40 or 80. + num_blstm_layers: number of bidirectional LSTM layers. + lstm_size: size of the LSTM hidden (and cell) size. + dropout: dropout probability. Dropout can be applied after the + linear layers and LSTM layers but not to the convolutional + layers. + """ + super().__init__(None) + + self.input_layers = nn.ModuleList() + in_features = input_feat_per_channel + for out_features in input_layers: + if dropout > 0: + self.input_layers.append( + nn.Sequential( + nn.Linear(in_features, out_features), nn.Dropout(p=dropout) + ) + ) + else: + self.input_layers.append(nn.Linear(in_features, out_features)) + in_features = out_features + + self.in_channels = in_channels + self.input_dim = input_feat_per_channel + self.conv_kernel_sizes_and_strides = [] + self.conv_layers = nn.ModuleList() + lstm_input_dim = input_layers[-1] + for conv_layer in conv_layers: + out_channels, conv_kernel_size, conv_stride = conv_layer + self.conv_layers.append( + nn.Conv2d( + in_channels, + out_channels, + conv_kernel_size, + stride=conv_stride, + padding=conv_kernel_size // 2, + ) + ) + self.conv_kernel_sizes_and_strides.append((conv_kernel_size, conv_stride)) + in_channels = out_channels + lstm_input_dim //= conv_stride + + lstm_input_dim *= conv_layers[-1][0] + self.lstm_size = lstm_size + self.num_blstm_layers = num_blstm_layers + self.lstm = nn.LSTM( + input_size=lstm_input_dim, + hidden_size=lstm_size, + num_layers=num_blstm_layers, + dropout=dropout, + bidirectional=True, + ) + self.output_dim = 2 * lstm_size # bidirectional + if dropout > 0: + self.dropout = nn.Dropout(p=dropout) + else: + self.dropout = None + + def forward(self, src_tokens, src_lengths=None, **kwargs): + """ + Args + src_tokens: padded tensor (B, T, C * feat) + src_lengths: tensor of original lengths of input utterances (B,) + """ + bsz, max_seq_len, _ = src_tokens.size() + # (B, C, T, feat) + x = ( + src_tokens.view(bsz, max_seq_len, self.in_channels, self.input_dim) + .transpose(1, 2) + .contiguous() + ) + + for input_layer in self.input_layers: + x = input_layer(x) + x = torch.tanh(x) + + for conv_layer in self.conv_layers: + x = conv_layer(x) + + bsz, _, output_seq_len, _ = x.size() + + # (B, C, T, feat) -> (B, T, C, feat) -> (T, B, C, feat) -> + # (T, B, C * feat) + x = x.transpose(1, 2).transpose(0, 1).contiguous().view(output_seq_len, bsz, -1) + + input_lengths = src_lengths.clone() + for k, s in self.conv_kernel_sizes_and_strides: + p = k // 2 + input_lengths = (input_lengths.float() + 2 * p - k) / s + 1 + input_lengths = input_lengths.floor().long() + + packed_x = nn.utils.rnn.pack_padded_sequence(x, input_lengths) + + h0 = x.new(2 * self.num_blstm_layers, bsz, self.lstm_size).zero_() + c0 = x.new(2 * self.num_blstm_layers, bsz, self.lstm_size).zero_() + packed_outs, _ = self.lstm(packed_x, (h0, c0)) + + # unpack outputs and apply dropout + x, output_lengths = nn.utils.rnn.pad_packed_sequence(packed_outs) + if self.dropout is not None: + x = self.dropout(x) + + encoder_padding_mask = ( + lengths_to_padding_mask(output_lengths).to(src_tokens.device).t() + ) + + return { + "encoder_out": x, # (T, B, C) + "encoder_padding_mask": encoder_padding_mask, # (T, B) + } + + def reorder_encoder_out(self, encoder_out, new_order): + encoder_out["encoder_out"] = encoder_out["encoder_out"].index_select( + 1, new_order + ) + encoder_out["encoder_padding_mask"] = encoder_out[ + "encoder_padding_mask" + ].index_select(1, new_order) + return encoder_out + + +class MLPAttention(nn.Module): + """The original attention from Badhanau et al. (2014) + + https://arxiv.org/abs/1409.0473, based on a Multi-Layer Perceptron. + The attention score between position i in the encoder and position j in the + decoder is: alpha_ij = V_a * tanh(W_ae * enc_i + W_ad * dec_j + b_a) + """ + + def __init__(self, decoder_hidden_state_dim, context_dim, attention_dim): + super().__init__() + + self.context_dim = context_dim + self.attention_dim = attention_dim + # W_ae and b_a + self.encoder_proj = nn.Linear(context_dim, self.attention_dim, bias=True) + # W_ad + self.decoder_proj = nn.Linear( + decoder_hidden_state_dim, self.attention_dim, bias=False + ) + # V_a + self.to_scores = nn.Linear(self.attention_dim, 1, bias=False) + + def forward(self, decoder_state, source_hids, encoder_padding_mask): + """The expected input dimensions are: + decoder_state: bsz x decoder_hidden_state_dim + source_hids: src_len x bsz x context_dim + encoder_padding_mask: src_len x bsz + """ + src_len, bsz, _ = source_hids.size() + # (src_len*bsz) x context_dim (to feed through linear) + flat_source_hids = source_hids.view(-1, self.context_dim) + # (src_len*bsz) x attention_dim + encoder_component = self.encoder_proj(flat_source_hids) + # src_len x bsz x attention_dim + encoder_component = encoder_component.view(src_len, bsz, self.attention_dim) + # 1 x bsz x attention_dim + decoder_component = self.decoder_proj(decoder_state).unsqueeze(0) + # Sum with broadcasting and apply the non linearity + # src_len x bsz x attention_dim + hidden_att = torch.tanh( + (decoder_component + encoder_component).view(-1, self.attention_dim) + ) + # Project onto the reals to get attentions scores (src_len x bsz) + attn_scores = self.to_scores(hidden_att).view(src_len, bsz) + + # Mask + softmax (src_len x bsz) + if encoder_padding_mask is not None: + attn_scores = ( + attn_scores.float() + .masked_fill_(encoder_padding_mask, float("-inf")) + .type_as(attn_scores) + ) # FP16 support: cast to float and back + # srclen x bsz + normalized_masked_attn_scores = F.softmax(attn_scores, dim=0) + + # Sum weighted sources (bsz x context_dim) + attn_weighted_context = ( + source_hids * normalized_masked_attn_scores.unsqueeze(2) + ).sum(dim=0) + + return attn_weighted_context, normalized_masked_attn_scores + + +class LSTMDecoder(FairseqIncrementalDecoder): + def __init__( + self, + dictionary, + embed_dim, + num_layers, + hidden_size, + dropout, + encoder_output_dim, + attention_dim, + output_layer_dim, + ): + """ + Args: + dictionary: target text dictionary. + embed_dim: embedding dimension for target tokens. + num_layers: number of LSTM layers. + hidden_size: hidden size for LSTM layers. + dropout: dropout probability. Dropout can be applied to the + embeddings, the LSTM layers, and the context vector. + encoder_output_dim: encoder output dimension (hidden size of + encoder LSTM). + attention_dim: attention dimension for MLP attention. + output_layer_dim: size of the linear layer prior to output + projection. + """ + super().__init__(dictionary) + self.num_layers = num_layers + self.hidden_size = hidden_size + num_embeddings = len(dictionary) + padding_idx = dictionary.pad() + self.embed_tokens = nn.Embedding(num_embeddings, embed_dim, padding_idx) + if dropout > 0: + self.dropout = nn.Dropout(p=dropout) + else: + self.dropout = None + + self.layers = nn.ModuleList() + for layer_id in range(num_layers): + input_size = embed_dim if layer_id == 0 else encoder_output_dim + self.layers.append( + nn.LSTMCell(input_size=input_size, hidden_size=hidden_size) + ) + + self.context_dim = encoder_output_dim + self.attention = MLPAttention( + decoder_hidden_state_dim=hidden_size, + context_dim=encoder_output_dim, + attention_dim=attention_dim, + ) + + self.deep_output_layer = nn.Linear( + hidden_size + encoder_output_dim + embed_dim, output_layer_dim + ) + self.output_projection = nn.Linear(output_layer_dim, num_embeddings) + + def forward( + self, prev_output_tokens, encoder_out=None, incremental_state=None, **kwargs + ): + encoder_padding_mask = encoder_out["encoder_padding_mask"] + encoder_outs = encoder_out["encoder_out"] + + if incremental_state is not None: + prev_output_tokens = prev_output_tokens[:, -1:] + bsz, seqlen = prev_output_tokens.size() + + srclen = encoder_outs.size(0) + + # embed tokens + embeddings = self.embed_tokens(prev_output_tokens) + x = embeddings + if self.dropout is not None: + x = self.dropout(x) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + + # initialize previous states (or get from cache during incremental + # generation) + cached_state = utils.get_incremental_state( + self, incremental_state, "cached_state" + ) + if cached_state is not None: + prev_hiddens, prev_cells = cached_state + else: + prev_hiddens = [encoder_out["encoder_out"].mean(dim=0)] * self.num_layers + prev_cells = [x.new_zeros(bsz, self.hidden_size)] * self.num_layers + + attn_scores = x.new_zeros(bsz, srclen) + attention_outs = [] + outs = [] + for j in range(seqlen): + input = x[j, :, :] + attention_out = None + for i, layer in enumerate(self.layers): + # the previous state is one layer below except for the bottom + # layer where the previous state is the state emitted by the + # top layer + hidden, cell = layer( + input, + ( + prev_hiddens[(i - 1) % self.num_layers], + prev_cells[(i - 1) % self.num_layers], + ), + ) + if self.dropout is not None: + hidden = self.dropout(hidden) + prev_hiddens[i] = hidden + prev_cells[i] = cell + if attention_out is None: + attention_out, attn_scores = self.attention( + hidden, encoder_outs, encoder_padding_mask + ) + if self.dropout is not None: + attention_out = self.dropout(attention_out) + attention_outs.append(attention_out) + input = attention_out + + # collect the output of the top layer + outs.append(hidden) + + # cache previous states (no-op except during incremental generation) + utils.set_incremental_state( + self, incremental_state, "cached_state", (prev_hiddens, prev_cells) + ) + + # collect outputs across time steps + x = torch.cat(outs, dim=0).view(seqlen, bsz, self.hidden_size) + attention_outs_concat = torch.cat(attention_outs, dim=0).view( + seqlen, bsz, self.context_dim + ) + + # T x B x C -> B x T x C + x = x.transpose(0, 1) + attention_outs_concat = attention_outs_concat.transpose(0, 1) + + # concat LSTM output, attention output and embedding + # before output projection + x = torch.cat((x, attention_outs_concat, embeddings), dim=2) + x = self.deep_output_layer(x) + x = torch.tanh(x) + if self.dropout is not None: + x = self.dropout(x) + # project back to size of vocabulary + x = self.output_projection(x) + + # to return the full attn_scores tensor, we need to fix the decoder + # to account for subsampling input frames + # return x, attn_scores + return x, None + + def reorder_incremental_state(self, incremental_state, new_order): + super().reorder_incremental_state(incremental_state, new_order) + cached_state = utils.get_incremental_state( + self, incremental_state, "cached_state" + ) + if cached_state is None: + return + + def reorder_state(state): + if isinstance(state, list): + return [reorder_state(state_i) for state_i in state] + return state.index_select(0, new_order) + + new_state = tuple(map(reorder_state, cached_state)) + utils.set_incremental_state(self, incremental_state, "cached_state", new_state) + + +@register_model_architecture(model_name="s2t_berard", arch_name="s2t_berard") +def berard(args): + """The original version: "End-to-End Automatic Speech Translation of + Audiobooks" (https://arxiv.org/abs/1802.04200) + """ + args.input_layers = getattr(args, "input_layers", "[256, 128]") + args.conv_layers = getattr(args, "conv_layers", "[(16, 3, 2), (16, 3, 2)]") + args.num_blstm_layers = getattr(args, "num_blstm_layers", 3) + args.lstm_size = getattr(args, "lstm_size", 256) + args.dropout = getattr(args, "dropout", 0.2) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 128) + args.decoder_num_layers = getattr(args, "decoder_num_layers", 2) + args.decoder_hidden_dim = getattr(args, "decoder_hidden_dim", 512) + args.attention_dim = getattr(args, "attention_dim", 512) + args.output_layer_dim = getattr(args, "output_layer_dim", 128) + args.load_pretrained_encoder_from = getattr( + args, "load_pretrained_encoder_from", None + ) + args.load_pretrained_decoder_from = getattr( + args, "load_pretrained_decoder_from", None + ) + + +@register_model_architecture(model_name="s2t_berard", arch_name="s2t_berard_256_3_3") +def berard_256_3_3(args): + """Used in + * "Harnessing Indirect Training Data for End-to-End Automatic Speech + Translation: Tricks of the Trade" (https://arxiv.org/abs/1909.06515) + * "CoVoST: A Diverse Multilingual Speech-To-Text Translation Corpus" + (https://arxiv.org/pdf/2002.01320.pdf) + * "Self-Supervised Representations Improve End-to-End Speech Translation" + (https://arxiv.org/abs/2006.12124) + """ + args.decoder_num_layers = getattr(args, "decoder_num_layers", 3) + berard(args) + + +@register_model_architecture(model_name="s2t_berard", arch_name="s2t_berard_512_3_2") +def berard_512_3_2(args): + args.num_blstm_layers = getattr(args, "num_blstm_layers", 3) + args.lstm_size = getattr(args, "lstm_size", 512) + args.dropout = getattr(args, "dropout", 0.3) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 256) + args.decoder_num_layers = getattr(args, "decoder_num_layers", 2) + args.decoder_hidden_dim = getattr(args, "decoder_hidden_dim", 1024) + args.attention_dim = getattr(args, "attention_dim", 512) + args.output_layer_dim = getattr(args, "output_layer_dim", 256) + berard(args) + + +@register_model_architecture(model_name="s2t_berard", arch_name="s2t_berard_512_5_3") +def berard_512_5_3(args): + args.num_blstm_layers = getattr(args, "num_blstm_layers", 5) + args.lstm_size = getattr(args, "lstm_size", 512) + args.dropout = getattr(args, "dropout", 0.3) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 256) + args.decoder_num_layers = getattr(args, "decoder_num_layers", 3) + args.decoder_hidden_dim = getattr(args, "decoder_hidden_dim", 1024) + args.attention_dim = getattr(args, "attention_dim", 512) + args.output_layer_dim = getattr(args, "output_layer_dim", 256) + berard(args) diff --git a/fairseq/models/speech_to_text/convtransformer.py b/fairseq/models/speech_to_text/convtransformer.py new file mode 100644 index 0000000000000000000000000000000000000000..eba000d7b0826d2ecf5dc471156f8f8cc9f5e402 --- /dev/null +++ b/fairseq/models/speech_to_text/convtransformer.py @@ -0,0 +1,448 @@ +#!/usr/bin/env python3 + +import logging +import math +from typing import Dict, List, Optional, Tuple + +import torch +import torch.nn as nn +import torch.nn.functional as F +from fairseq import checkpoint_utils, utils +from fairseq.data.data_utils import lengths_to_padding_mask +from fairseq.models import ( + FairseqEncoder, + FairseqEncoderDecoderModel, + register_model, + register_model_architecture, +) +from fairseq.models.transformer import Embedding, TransformerDecoder +from fairseq.modules import LayerNorm, PositionalEmbedding, TransformerEncoderLayer +from torch import Tensor + +logger = logging.getLogger(__name__) + + +@register_model("convtransformer") +class ConvTransformerModel(FairseqEncoderDecoderModel): + """ + Transformer-based Speech translation model from ESPNet-ST + https://arxiv.org/abs/2004.10234 + """ + + def __init__(self, encoder, decoder): + super().__init__(encoder, decoder) + + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + parser.add_argument( + "--input-feat-per-channel", + type=int, + metavar="N", + help="encoder input dimension per input channel", + ) + parser.add_argument( + "--activation-fn", + choices=utils.get_available_activation_fns(), + help="activation function to use", + ) + parser.add_argument( + "--dropout", type=float, metavar="D", help="dropout probability" + ) + parser.add_argument( + "--attention-dropout", + type=float, + metavar="D", + help="dropout probability for attention weights", + ) + parser.add_argument( + "--activation-dropout", + "--relu-dropout", + type=float, + metavar="D", + help="dropout probability after activation in FFN.", + ) + parser.add_argument( + "--encoder-embed-dim", + type=int, + metavar="N", + help="encoder embedding dimension", + ) + parser.add_argument( + "--encoder-ffn-embed-dim", + type=int, + metavar="N", + help="encoder embedding dimension for FFN", + ) + parser.add_argument( + "--encoder-layers", type=int, metavar="N", help="num encoder layers" + ) + parser.add_argument( + "--encoder-attention-heads", + type=int, + metavar="N", + help="num encoder attention heads", + ) + parser.add_argument( + "--encoder-normalize-before", + action="store_true", + help="apply layernorm before each encoder block", + ) + parser.add_argument( + "--decoder-embed-dim", + type=int, + metavar="N", + help="decoder embedding dimension", + ) + parser.add_argument( + "--decoder-ffn-embed-dim", + type=int, + metavar="N", + help="decoder embedding dimension for FFN", + ) + parser.add_argument( + "--decoder-layers", type=int, metavar="N", help="num decoder layers" + ) + parser.add_argument( + "--decoder-attention-heads", + type=int, + metavar="N", + help="num decoder attention heads", + ) + parser.add_argument( + "--decoder-normalize-before", + action="store_true", + help="apply layernorm before each decoder block", + ) + parser.add_argument( + "--decoder-output-dim", + type=int, + metavar="N", + help="decoder output dimension (extra linear layer if different from decoder embed dim)", + ) + parser.add_argument( + "--share-decoder-input-output-embed", + action="store_true", + help="share decoder input and output embeddings", + ) + parser.add_argument( + "--layernorm-embedding", + action="store_true", + help="add layernorm to embedding", + ) + parser.add_argument( + "--no-scale-embedding", + action="store_true", + help="if True, dont scale embeddings", + ) + parser.add_argument( + "--load-pretrained-encoder-from", + type=str, + metavar="STR", + help="model to take encoder weights from (for initialization)", + ) + parser.add_argument( + "--load-pretrained-decoder-from", + type=str, + metavar="STR", + help="model to take decoder weights from (for initialization)", + ) + parser.add_argument( + "--conv-out-channels", + type=int, + metavar="INT", + help="the number of output channels of conv layer", + ) + + @classmethod + def build_encoder(cls, args): + encoder = ConvTransformerEncoder(args) + if getattr(args, "load_pretrained_encoder_from", None): + encoder = checkpoint_utils.load_pretrained_component_from_model( + component=encoder, checkpoint=args.load_pretrained_encoder_from + ) + return encoder + + @classmethod + def build_decoder(cls, args, task, embed_tokens): + decoder = TransformerDecoderNoExtra(args, task.target_dictionary, embed_tokens) + if getattr(args, "load_pretrained_decoder_from", None): + decoder = checkpoint_utils.load_pretrained_component_from_model( + component=decoder, checkpoint=args.load_pretrained_decoder_from + ) + return decoder + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + + # make sure all arguments are present in older models + base_architecture(args) + + def build_embedding(dictionary, embed_dim): + num_embeddings = len(dictionary) + padding_idx = dictionary.pad() + return Embedding(num_embeddings, embed_dim, padding_idx) + + decoder_embed_tokens = build_embedding( + task.target_dictionary, args.decoder_embed_dim + ) + encoder = cls.build_encoder(args) + decoder = cls.build_decoder(args, task, decoder_embed_tokens) + return cls(encoder, decoder) + + @staticmethod + @torch.jit.unused + def set_batch_first(lprobs): + lprobs.batch_first = True + + def get_normalized_probs( + self, + net_output: Tuple[Tensor, Optional[Dict[str, List[Optional[Tensor]]]]], + log_probs: bool, + sample: Optional[Dict[str, Tensor]] = None, + ): + # net_output['encoder_out'] is a (B, T, D) tensor + lprobs = self.get_normalized_probs_scriptable(net_output, log_probs, sample) + if self.training: + self.set_batch_first(lprobs) + return lprobs + + def output_layout(self): + return "BTD" + + """ + The forward method inherited from the base class has a **kwargs argument in + its input, which is not supported in torchscript. This method overrites the forward + method definition without **kwargs. + """ + + def forward(self, src_tokens, src_lengths, prev_output_tokens): + encoder_out = self.encoder(src_tokens=src_tokens, src_lengths=src_lengths) + decoder_out = self.decoder( + prev_output_tokens=prev_output_tokens, encoder_out=encoder_out + ) + return decoder_out + + +class ConvTransformerEncoder(FairseqEncoder): + """Conv + Transformer encoder""" + + def __init__(self, args): + """Construct an Encoder object.""" + super().__init__(None) + + self.dropout = args.dropout + self.embed_scale = ( + 1.0 if args.no_scale_embedding else math.sqrt(args.encoder_embed_dim) + ) + self.padding_idx = 1 + self.in_channels = 1 + self.input_dim = args.input_feat_per_channel + self.conv = torch.nn.Sequential( + torch.nn.Conv2d(1, args.conv_out_channels, 3, stride=2, padding=3 // 2), + torch.nn.ReLU(), + torch.nn.Conv2d( + args.conv_out_channels, + args.conv_out_channels, + 3, + stride=2, + padding=3 // 2, + ), + torch.nn.ReLU(), + ) + transformer_input_dim = self.infer_conv_output_dim( + self.in_channels, self.input_dim, args.conv_out_channels + ) + self.out = torch.nn.Linear(transformer_input_dim, args.encoder_embed_dim) + self.embed_positions = PositionalEmbedding( + args.max_source_positions, + args.encoder_embed_dim, + self.padding_idx, + learned=False, + ) + + self.transformer_layers = nn.ModuleList([]) + self.transformer_layers.extend( + [TransformerEncoderLayer(args) for i in range(args.encoder_layers)] + ) + if args.encoder_normalize_before: + self.layer_norm = LayerNorm(args.encoder_embed_dim) + else: + self.layer_norm = None + + def pooling_ratio(self): + return 4 + + def infer_conv_output_dim(self, in_channels, input_dim, out_channels): + sample_seq_len = 200 + sample_bsz = 10 + x = torch.randn(sample_bsz, in_channels, sample_seq_len, input_dim) + x = torch.nn.Conv2d(1, out_channels, 3, stride=2, padding=3 // 2)(x) + x = torch.nn.Conv2d(out_channels, out_channels, 3, stride=2, padding=3 // 2)(x) + x = x.transpose(1, 2) + mb, seq = x.size()[:2] + return x.contiguous().view(mb, seq, -1).size(-1) + + def forward(self, src_tokens, src_lengths): + """Encode input sequence. + :param torch.Tensor xs: input tensor + :param torch.Tensor masks: input mask + :return: position embedded tensor and mask + :rtype Tuple[torch.Tensor, torch.Tensor]: + """ + bsz, max_seq_len, _ = src_tokens.size() + x = ( + src_tokens.view(bsz, max_seq_len, self.in_channels, self.input_dim) + .transpose(1, 2) + .contiguous() + ) + x = self.conv(x) + bsz, _, output_seq_len, _ = x.size() + x = x.transpose(1, 2).transpose(0, 1).contiguous().view(output_seq_len, bsz, -1) + x = self.out(x) + x = self.embed_scale * x + + subsampling_factor = int(max_seq_len * 1.0 / output_seq_len + 0.5) + input_len_0 = (src_lengths.float() / subsampling_factor).ceil().long() + input_len_1 = x.size(0) * torch.ones([src_lengths.size(0)]).long().to( + input_len_0.device + ) + input_lengths = torch.min(input_len_0, input_len_1) + + encoder_padding_mask = lengths_to_padding_mask(input_lengths) + + positions = self.embed_positions(encoder_padding_mask).transpose(0, 1) + x += positions + x = F.dropout(x, p=self.dropout, training=self.training) + + for layer in self.transformer_layers: + x = layer(x, encoder_padding_mask) + + if not encoder_padding_mask.any(): + maybe_encoder_padding_mask = None + else: + maybe_encoder_padding_mask = encoder_padding_mask + + return { + "encoder_out": [x], + "encoder_padding_mask": [maybe_encoder_padding_mask] + if maybe_encoder_padding_mask is not None + else [], + "encoder_embedding": [], + "encoder_states": [], + "src_tokens": [], + "src_lengths": [], + } + + @torch.jit.export + def reorder_encoder_out(self, encoder_out: Dict[str, List[Tensor]], new_order): + """ + Reorder encoder output according to *new_order*. + + Args: + encoder_out: output from the ``forward()`` method + new_order (LongTensor): desired order + + Returns: + *encoder_out* rearranged according to *new_order* + """ + new_encoder_out = [encoder_out["encoder_out"][0].index_select(1, new_order)] + if len(encoder_out["encoder_padding_mask"]) == 0: + new_encoder_padding_mask = [] + else: + new_encoder_padding_mask = [ + (encoder_out["encoder_padding_mask"][0]).index_select(0, new_order) + ] + if len(encoder_out["encoder_embedding"]) == 0: + new_encoder_embedding = [] + else: + new_encoder_embedding = [ + (encoder_out["encoder_embedding"][0]).index_select(0, new_order) + ] + encoder_states = encoder_out["encoder_states"] + if len(encoder_states) > 0: + for idx, state in enumerate(encoder_states): + encoder_states[idx] = state.index_select(1, new_order) + + return { + "encoder_out": new_encoder_out, + "encoder_padding_mask": new_encoder_padding_mask, + "encoder_embedding": new_encoder_embedding, + "encoder_states": encoder_states, + "src_tokens": [], + "src_lengths": [], + } + + +class TransformerDecoderNoExtra(TransformerDecoder): + def extract_features( + self, + prev_output_tokens, + encoder_out: Optional[Dict[str, List[Tensor]]], + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + full_context_alignment: bool = False, + alignment_layer: Optional[int] = None, + alignment_heads: Optional[int] = None, + ): + # call scriptable method from parent class + x, _ = self.extract_features_scriptable( + prev_output_tokens, + encoder_out, + incremental_state, + full_context_alignment, + alignment_layer, + alignment_heads, + ) + return x, None + + +@register_model_architecture(model_name="convtransformer", arch_name="convtransformer") +def base_architecture(args): + args.input_feat_per_channel = getattr(args, "input_feat_per_channel", 80) + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 2048) + args.encoder_layers = getattr(args, "encoder_layers", 6) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 8) + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", args.encoder_embed_dim) + args.decoder_ffn_embed_dim = getattr( + args, "decoder_ffn_embed_dim", args.encoder_ffn_embed_dim + ) + args.decoder_layers = getattr(args, "decoder_layers", 6) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 8) + args.decoder_normalize_before = getattr(args, "decoder_normalize_before", False) + args.decoder_learned_pos = getattr(args, "decoder_learned_pos", False) + args.attention_dropout = getattr(args, "attention_dropout", 0.0) + args.activation_dropout = getattr(args, "activation_dropout", 0.0) + args.activation_fn = getattr(args, "activation_fn", "relu") + args.dropout = getattr(args, "dropout", 0.1) + args.adaptive_softmax_cutoff = getattr(args, "adaptive_softmax_cutoff", None) + args.adaptive_softmax_dropout = getattr(args, "adaptive_softmax_dropout", 0) + args.share_decoder_input_output_embed = getattr( + args, "share_decoder_input_output_embed", False + ) + args.no_token_positional_embeddings = getattr( + args, "no_token_positional_embeddings", False + ) + args.adaptive_input = getattr(args, "adaptive_input", False) + args.decoder_layerdrop = getattr(args, "decoder_layerdrop", 0.0) + + args.decoder_output_dim = getattr( + args, "decoder_output_dim", args.decoder_embed_dim + ) + args.decoder_input_dim = getattr(args, "decoder_input_dim", args.decoder_embed_dim) + args.no_scale_embedding = getattr(args, "no_scale_embedding", False) + args.quant_noise_pq = getattr(args, "quant_noise_pq", 0) + args.max_source_positions = getattr(args, "max_source_positions", 3000) + args.max_target_positions = getattr(args, "max_target_positions", 1024) + args.tie_adaptive_weights = getattr(args, "tie_adaptive_weights", False) + args.conv_out_channels = getattr(args, "conv_out_channels", args.encoder_embed_dim) + + +@register_model_architecture("convtransformer", "convtransformer_espnet") +def convtransformer_espnet(args): + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 256) + args.encoder_layers = getattr(args, "encoder_layers", 12) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 4) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 4) diff --git a/fairseq/models/speech_to_text/modules/augmented_memory_attention.py b/fairseq/models/speech_to_text/modules/augmented_memory_attention.py new file mode 100644 index 0000000000000000000000000000000000000000..e7465bc889fd1ba6ca2c60905a2eb6ff5cc62b9d --- /dev/null +++ b/fairseq/models/speech_to_text/modules/augmented_memory_attention.py @@ -0,0 +1,488 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from typing import Tuple, List + +import torch +import torch.nn.functional as F +from fairseq.models import FairseqEncoder +from fairseq.models.speech_to_text import ( + ConvTransformerEncoder, +) +from fairseq.models.speech_to_text.utils import attention_suppression +from fairseq.models.speech_to_text.utils import ( + lengths_to_encoder_padding_mask, + segments_to_sequence, + sequence_to_segments, +) +from fairseq.modules import MultiheadAttention, TransformerEncoderLayer +from torch import nn, Tensor + +# ------------------------------------------------------------------------------ +# AugmentedMemoryConvTransformerEncoder +# ------------------------------------------------------------------------------ + + +class AugmentedMemoryConvTransformerEncoder(ConvTransformerEncoder): + def __init__(self, args): + super().__init__(args) + + args.encoder_stride = self.stride() + + self.left_context = args.left_context // args.encoder_stride + + self.right_context = args.right_context // args.encoder_stride + + self.left_context_after_stride = args.left_context // args.encoder_stride + self.right_context_after_stride = args.right_context // args.encoder_stride + + self.transformer_layers = nn.ModuleList([]) + self.transformer_layers.extend( + [ + AugmentedMemoryTransformerEncoderLayer(args) + for i in range(args.encoder_layers) + ] + ) + + def stride(self): + # Hard coded here. Should infer from convs in future + stride = 4 + return stride + + def forward(self, src_tokens, src_lengths, states=None): + """Encode input sequence. + :param torch.Tensor xs: input tensor + :param torch.Tensor masks: input mask + :return: position embedded tensor and mask + :rtype Tuple[torch.Tensor, torch.Tensor]: + """ + bsz, max_seq_len, _ = src_tokens.size() + x = ( + src_tokens.view(bsz, max_seq_len, self.in_channels, self.input_dim) + .transpose(1, 2) + .contiguous() + ) + x = self.conv(x) + bsz, _, output_seq_len, _ = x.size() + x = x.transpose(1, 2).transpose(0, 1).contiguous().view(output_seq_len, bsz, -1) + x = self.out(x) + x = self.embed_scale * x + + subsampling_factor = 1.0 * max_seq_len / output_seq_len + input_lengths = torch.max( + (src_lengths.float() / subsampling_factor).ceil().long(), + x.size(0) * src_lengths.new_ones([src_lengths.size(0)]).long(), + ) + + encoder_padding_mask, _ = lengths_to_encoder_padding_mask( + input_lengths, batch_first=True + ) + + # TODO: fix positional embedding + positions = self.embed_positions(encoder_padding_mask).transpose(0, 1) + + x += positions + x = F.dropout(x, p=self.dropout, training=self.training) + + # State to store memory banks etc. + if states is None: + states = [ + {"memory_banks": None, "encoder_states": None} + for i in range(len(self.transformer_layers)) + ] + + for i, layer in enumerate(self.transformer_layers): + # x size: + # (self.left_size + self.segment_size + self.right_size) + # / self.stride, num_heads, dim + # TODO: Consider mask here + x = layer(x, states[i]) + states[i]["encoder_states"] = x[ + self.left_context_after_stride : -self.right_context_after_stride + ] + + lengths = ( + ( + ~encoder_padding_mask[ + :, self.left_context_after_stride : -self.right_context_after_stride + ] + ) + .sum(dim=1, keepdim=True) + .long() + ) + + return states[-1]["encoder_states"], lengths, states + + +# ------------------------------------------------------------------------------ +# AugmentedMemoryTransformerEncoderLayer +# ------------------------------------------------------------------------------ +class AugmentedMemoryTransformerEncoderLayer(TransformerEncoderLayer): + def __init__(self, args): + super().__init__(args) + + self.left_context = args.left_context // args.encoder_stride + self.right_context = args.right_context // args.encoder_stride + + def forward(self, x, state): + + length, batch_size, x_dim = x.size() + + residual = x + + if self.normalize_before: + x = self.self_attn_layer_norm(x) + + # init_state + if state.get("memory_banks", None) is None: + state["memory_banks"] = [] + + # TODO reseach new sum_query method + seg_start = self.left_context + seg_end = length - self.right_context + if seg_start < seg_end: + summarization_query = torch.mean(x[seg_start:seg_end], keepdim=True, dim=0) + else: + summarization_query = x.new_zeros(1, batch_size, x_dim) + + x = torch.cat([x, summarization_query], dim=0) + + x = self.self_attn(input_and_summary=x, state=state) + + x = self.dropout_module(x) + x = residual + x + + if not self.normalize_before: + x = self.self_attn_layer_norm(x) + + residual = x + if self.normalize_before: + x = self.final_layer_norm(x) + + x = self.activation_fn(self.fc1(x)) + x = self.activation_dropout_module(x) + x = self.fc2(x) + x = self.dropout_module(x) + x = residual + x + if not self.normalize_before: + x = self.final_layer_norm(x) + + return x + + def build_self_attention(self, embed_dim, args): + return AugmentedMemoryMultiheadAttention( + embed_dim=embed_dim, + num_heads=args.encoder_attention_heads, + dropout=args.attention_dropout, + self_attention=True, + q_noise=self.quant_noise, + qn_block_size=self.quant_noise_block_size, + tanh_on_mem=True, + max_memory_size=args.max_memory_size, + ) + + +# ------------------------------------------------------------------------------ +# AugmentedMemoryMultiheadAttention +# ------------------------------------------------------------------------------ +class AugmentedMemoryMultiheadAttention(MultiheadAttention): + """ + Augmented Memory Attention from + Streaming Transformer-based Acoustic Models + Using Self-attention with Augmented Memory + https://arxiv.org/abs/2005.08042 + """ + + def __init__( + self, + embed_dim, + num_heads, + kdim=None, + vdim=None, + dropout=0.0, + bias=True, + add_bias_kv=False, + add_zero_attn=False, + self_attention=False, + encoder_decoder_attention=False, + q_noise=0.0, + qn_block_size=8, + tanh_on_mem=False, + memory_dim=None, + std_scale=0.5, # 0.5 based on https://arxiv.org/abs/2005.09137 + max_memory_size=-1, + disable_mem_on_mem_attn=True, + ): + super().__init__( + embed_dim, + num_heads, + kdim, + vdim, + dropout, + bias, + add_bias_kv, + add_zero_attn, + self_attention, + encoder_decoder_attention, + q_noise, + qn_block_size, + ) + + self.memory_dim = memory_dim if memory_dim is not None else embed_dim + self.std_scale = std_scale + self.disable_mem_on_mem_attn = disable_mem_on_mem_attn + + # This Operator was used for factorization in PySpeech + self.v2e = lambda x: x + + if tanh_on_mem: + self.squash_mem = torch.tanh + self.nonlinear_squash_mem = True + else: + self.squash_mem = lambda x: x + self.nonlinear_squash_mem = False + + self.max_memory_size = max_memory_size + + def forward(self, input_and_summary, state): + """ + input: Encoder states of current segment with left or right context, + plus one summarization query + + """ + + length, batch_size, _ = input_and_summary.shape + length = length - 1 # not include sum_query, last index + + memory = state["memory_banks"] + # TODO: positional embedding on memory + + if self.max_memory_size > -1 and len(memory) > self.max_memory_size: + # TODO: need to fix here + if self.max_memory_size == 0: + memory = memory.new_zeros(1, memory.size(1), self.memory_dim) + else: + memory = memory[-self.max_memory_size :] + + memory_and_input = torch.cat(memory + [input_and_summary[:-1]], dim=0) + input_and_sum_query = input_and_summary + + q = self.q_proj(self.v2e(input_and_sum_query)) + k = self.k_proj(self.v2e(memory_and_input)) + v = self.v_proj(self.v2e(memory_and_input)) + + q = ( + q.contiguous() + .view(-1, batch_size * self.num_heads, self.head_dim) + .transpose(0, 1) + * self.scaling + ) + k = ( + k.contiguous() + .view(-1, batch_size * self.num_heads, self.head_dim) + .transpose(0, 1) + ) + + v = ( + v.contiguous() + .view(-1, batch_size * self.num_heads, self.head_dim) + .transpose(0, 1) + ) + + attention_weights = torch.bmm(q, k.transpose(1, 2)) + + if self.disable_mem_on_mem_attn: + attention_weights = self.suppress_mem_on_mem_attention( + batch_size, self.num_heads, len(memory), attention_weights + ) + + if self.std_scale is not None: + attention_weights = attention_suppression(attention_weights, self.std_scale) + + assert list(attention_weights.shape) == [ + batch_size * self.num_heads, + length + 1, + length + len(memory), + ] + + attention_weights = torch.nn.functional.softmax( + attention_weights.float(), dim=-1 + ).type_as(attention_weights) + + attention_probs = self.dropout_module(attention_weights) + + # [T, T, B, n_head] + [T, B, n_head, d_head] -> [T, B, n_head, d_head] + attention = torch.bmm(attention_probs, v) + + assert list(attention.shape) == [ + batch_size * self.num_heads, + length + 1, + self.head_dim, + ] + + attention = ( + attention.transpose(0, 1) + .contiguous() + .view(length + 1, batch_size, self.embed_dim) + ) + + output_and_memory = self.out_proj(attention) + + next_m = output_and_memory[-1:] + next_m = self.squash_mem(next_m) + output = output_and_memory[:-1] + + state["memory_banks"].append(next_m) + + return output + + def suppress_mem_on_mem_attention( + self, B: int, num_heads: int, mem_size: int, attention_weight: Tensor + ): + """ + Arguments: + - B: batch size + - num_heads: number of attention heads + - mem_size: size of memory bank + - attention_weight: a [B*num_heads, T + 1, T + mem_size] vector + + Return: + modified attention_weight with [B*num_heads, -1, :mem_size] = -inf + """ + attention_weight[:, -1, :mem_size] = float("-inf") + return attention_weight + + +# ------------------------------------------------------------------------------ +# SequenceEncoder +# ------------------------------------------------------------------------------ +class SequenceEncoder(FairseqEncoder): + """ + SequenceEncoder encodes sequences. + + More specifically, `src_tokens` and `src_lengths` in `forward()` should + describe a batch of "complete" sequences rather than segments. + + Segment-by-segment inference can be triggered by `segment_size`: + 1) `segment_size` is None: + SequenceEncoder treats the input sequence as one single segment. + 2) `segment_size` is not None (some int instead): + SequenceEncoder does the following: + 1. breaks the input sequence into several segments + 2. inference on each segment and collect the outputs + 3. concatanete segment outputs into the output sequence. + Note that `segment_size` here shouldn't include additional left/right + contexts needed, for example if we wish to infer with LC-BLSTM where the + middle chunk size is 100 and right context is 20, `segment_size` should be + 100. + """ + + def __init__(self, args, module): + super().__init__(None) + + self.module = module + self.input_time_axis = 1 + self.output_time_axis = 0 + self.segment_size = args.segment_size + self.left_context = args.left_context + self.right_context = args.right_context + + def forward( + self, + src_tokens: Tensor, + src_lengths: Tensor, + states=None, + ): + + seg_src_tokens_lengths = sequence_to_segments( + sequence=src_tokens, + time_axis=self.input_time_axis, + lengths=src_lengths, + segment_size=self.segment_size, + extra_left_context=self.left_context, + extra_right_context=self.right_context, + ) + + seg_encoder_states_lengths: List[Tuple[Tensor, Tensor]] = [] + + for seg_src_tokens, seg_src_lengths in seg_src_tokens_lengths: + (seg_encoder_states, seg_enc_lengths, states) = self.module( + seg_src_tokens, + seg_src_lengths, + states=states, + ) + + seg_encoder_states_lengths.append((seg_encoder_states, seg_enc_lengths)) + + encoder_out, enc_lengths = segments_to_sequence( + segments=seg_encoder_states_lengths, time_axis=self.output_time_axis + ) + + encoder_padding_mask, _ = lengths_to_encoder_padding_mask( + enc_lengths, batch_first=True + ) + + if not encoder_padding_mask.any(): + encoder_padding_mask = None + + return { + "encoder_out": [encoder_out], + "encoder_padding_mask": [encoder_padding_mask], + "encoder_embedding": [], + "encoder_states": [states], + "src_tokens": [], + "src_lengths": [], + } + + def incremental_encode( + self, + seg_src_tokens: Tensor, + seg_src_lengths: Tensor, + states=None, + ): + """ + Different from forward function, this function takes segmented speech + as input, and append encoder states to previous states + """ + (seg_encoder_states, seg_enc_lengths, states) = self.module( + seg_src_tokens, + seg_src_lengths, + states=states, + ) + return seg_encoder_states, seg_enc_lengths, states + + +# ------------------------------------------------------------------------------ +# Augmented memory model decorator +# ------------------------------------------------------------------------------ +def augmented_memory(klass): + class StreamSeq2SeqModel(klass): + @staticmethod + def add_args(parser): + super(StreamSeq2SeqModel, StreamSeq2SeqModel).add_args(parser) + parser.add_argument( + "--segment-size", type=int, required=True, help="Length of the segment." + ) + parser.add_argument( + "--left-context", + type=int, + default=0, + help="Left context for the segment.", + ) + parser.add_argument( + "--right-context", + type=int, + default=0, + help="Right context for the segment.", + ) + parser.add_argument( + "--max-memory-size", + type=int, + default=-1, + help="Right context for the segment.", + ) + + StreamSeq2SeqModel.__name__ = klass.__name__ + return StreamSeq2SeqModel diff --git a/fairseq/models/speech_to_text/modules/emformer.py b/fairseq/models/speech_to_text/modules/emformer.py new file mode 100644 index 0000000000000000000000000000000000000000..6ef76bd012ba40b0395fec2ca9ae9e9c136ffe40 --- /dev/null +++ b/fairseq/models/speech_to_text/modules/emformer.py @@ -0,0 +1,1837 @@ +#!/usr/bin/env python3 +# Copyright (c) 2017-present, Facebook, Inc. +# All rights reserved. +# +# This source code is licensed under the license found in the LICENSE file in +# the root directory of this source tree. An additional grant of patent rights +# can be found in the PATENTS file in the same directory. + + +import math +import re +from functools import partial +from typing import List, Optional, Tuple + +import torch +import torch.nn as nn +from fairseq.models import ( + FairseqEncoder, +) +from fairseq.models.speech_to_text.utils import ( + NoOp, + lengths_to_padding_mask, + segments_to_sequence, +) +from fairseq.models.speech_to_text.utils import ( + attention_suppression, + layer_norm_backward_hook, +) +from torch import Tensor, device as Device +from torch.quantization.qconfig import ( + default_dynamic_qconfig, + per_channel_dynamic_qconfig, +) + + +class RelativePositionEmbedding(nn.Module): + """ + Implementation according to https://arxiv.org/abs/1803.02155 + """ + + def __init__(self, head_dim, max_position, norm_init=True): + super().__init__() + self.head_dim = head_dim + self.max_position = max_position + self.embeddings = nn.Parameter(torch.Tensor(max_position * 2 + 1, head_dim)) + if norm_init: + nn.init.xavier_normal_(self.embeddings) + else: + nn.init.xavier_uniform_(self.embeddings) + + def forward(self, input: Tensor): + output = nn.functional.embedding(input.long(), self.embeddings) + return output + + +class Fp32LayerNorm(nn.Module): + def __init__( + self, + input_dim, + clamp_grad=True, + max_grad_value=256, + eps=1e-5, + elementwise_affine=True, + ): + super().__init__() + self.torch_module = torch.nn.LayerNorm( + input_dim, eps=eps, elementwise_affine=elementwise_affine + ) + if clamp_grad: + hook = partial(layer_norm_backward_hook, clamp_value=max_grad_value) + self.torch_module.register_backward_hook(hook) + + def forward(self, input): + output = torch.nn.functional.layer_norm( + input.float(), + self.torch_module.normalized_shape, + self.torch_module.weight.float() + if self.torch_module.weight is not None + else None, + self.torch_module.bias.float() + if self.torch_module.bias is not None + else None, + self.torch_module.eps, + ).type_as(input) + return output + + +# ------------------------------------------------------------------------------ +# PositionwiseFF +# ------------------------------------------------------------------------------ + + +class PositionwiseFF(nn.Module): + """ + FFN layer in transformer. + + Args: + input_dim: input embedding dimension + ffn_dim: FFN layer inner dimension + dropout_on_fc1: dropout for first linear layer + dropout_on_fc2: dropout fr second linear layer + activation_fn: activation function used after first linear layer. \ + Only relu or gelu is supported. + + """ + + def __init__( + self, input_dim, ffn_dim, dropout_on_fc1, dropout_on_fc2, activation_fn + ): + super(PositionwiseFF, self).__init__() + + self.input_dim = input_dim + self.ffn_dim = ffn_dim + if activation_fn == "relu": + ac = nn.ReLU() + elif activation_fn == "gelu": + ac = nn.GELU() + else: + raise ValueError("Unsupported activation_fn = ({})".format(activation_fn)) + + # fc1 -> ac -> dropout -> fc2 -> dropout + self.module = nn.Sequential( + nn.Linear(input_dim, ffn_dim), + ac, + nn.Dropout(dropout_on_fc1), + nn.Linear(ffn_dim, input_dim), + nn.Dropout(dropout_on_fc2), + ) + + self.layer_norm = Fp32LayerNorm(input_dim) + + def forward(self, input): + module_out = self.module(self.layer_norm(input)) + output = module_out + input + + return output + + def quantize_(self, params=None): + if params and "per_channel" in params and params["per_channel"]: + qconfig = per_channel_dynamic_qconfig + else: + qconfig = default_dynamic_qconfig + torch.quantization.quantize_dynamic( + self, {torch.nn.Linear: qconfig}, dtype=torch.qint8, inplace=True + ) + return self + + +# ------------------------------------------------------------------------------ +# SummarizationLayer +# ------------------------------------------------------------------------------ + + +class SummarizationLayer(nn.Module): + def __init__(self, method, segment_size, embedding_dim): + super(SummarizationLayer, self).__init__() + self.segment_size = segment_size + self.embedding_dim = embedding_dim + nonlin_match = re.match(r"nonlinear\((?P<act>[a-z]+),(?P<dim>[0-9]+)\)", method) + self.method = method + if method == "mean": + self.module = nn.AvgPool1d( + kernel_size=segment_size, + stride=segment_size, + ceil_mode=True, + ) + elif method == "max": + self.module = nn.MaxPool1d( + kernel_size=segment_size, + stride=segment_size, + ceil_mode=True, + ) + elif method == "linear": + self.module = nn.Linear(segment_size, 1) + elif nonlin_match: + nonlin_args = nonlin_match.groupdict() + act_type = nonlin_args["act"] + hid_dim = int(nonlin_args["dim"]) + if act_type == "relu": + act = nn.ReLU() + elif act_type == "gelu": + act = nn.GELU() + else: + raise ValueError("Unsupported activation_fn = ({})".format(act_type)) + self.module = nn.Sequential( + nn.Linear(segment_size, hid_dim), + act, + nn.Linear(hid_dim, 1), + ) + else: + raise ValueError("Unsupported summarization method = ({})".format(method)) + + def forward(self, input): + # T, B, D -> B, D, T + input = input.permute(1, 2, 0) + + if self.method == "mean" or self.method == "max": + output = self.module(input) + output = output.permute(2, 0, 1) + return output + + full_seg_length = input.size(2) // self.segment_size * self.segment_size + if full_seg_length > 0: + # at least one seg is full + B = input.size(0) + D = input.size(1) + input_todo = ( + input[:, :, :full_seg_length] + .contiguous() + .view(B, -1, self.segment_size) + ) + output = self.module(input_todo) + output = output.view(B, D, -1) + else: + output = input.new_zeros(input.size(0), input.size(1), 0) + left = input.size(2) - full_seg_length + if left > 0: + # when last seg is not full, use zeros as last memory placeholder + zeros = input.new_zeros(input.size(0), input.size(1), 1) + output = torch.cat([output, zeros], dim=2) + output = output.permute(2, 0, 1) + return output + + +# ------------------------------------------------------------------------------ +# NoSegAugmentedMemoryMultiheadAttentionBmm +# ------------------------------------------------------------------------------ + + +class NoSegAugmentedMemoryMultiheadAttentionBmm(nn.Module): + """ + Whole utterance augmented memory multihead attention using BMM. + + Different with previous augmented memory multihead attention where + the utterance is chunked into segments. Here we use attention mask + achieve so. The input embedding [right_context, utterance, summary] + is a concatenation of right context, utterance and summary. + + Right context block is the concatenation of all the right context for + each segments. [right_context_0, right_context_1, ..., right_context_n] + For example, if we have utterance = [v0, v1, v2, ...., v20]. segment + size 8, right_context size 4. Then the right context blocks = + [v8, v9, v10, v11, v16, v17, v18, v19, 0, 0, 0, 0], where v8, v9, v10, + and v11 are the right context for first segment. v16, v17, v18 and v19 + are the right context for second segment. 0, 0, 0 and 0 are right context + for the last segment. + + utterance is corresponding to input embedding sequence + + summary is concatenation of average of each segments. [summary_0, + summary_1, ..., ]. + + In augmented memory multihead attention, the query is [right_context, + utterance, summary], key is [memory, right_context, utterance]. Different + with AugmentedMemoryMultiheadAttentionBmm, memory here is passed from + previous attention layer. For the first attention layer, memory is average + of each segment. + + Memory is a concatenation of memory from each segments in previous attention + layer. For example, current layer is i, then memory is [m_0, m_1, ..., m_n]. + Each m_k is the output from seg_k in layer i-1. + + args: + input_dim: input embedding dimension + num_heads: number of heads in multihead self-attention + dropout: attention dropout + std_scale: if std_scale is not None. The weak attention suppression is + turned on. For std_scale = 0.5, all the attention smaller than + mean + 0.5 * std will be suppressed. + scaled_init: whether to use scaled init for linear weight + tanh_on_mem: whether to use tanh on memory output + use_mem: whether to use memory or not. When max_memory_size is 0, then + we don't have memory anymore. + layer_index: current self-attention layer index that is used in depth + initialization + max_relative_position: max relative position used in relative position + embedding + rpe_old_option: To be compatible with previous model. The previous model + was trained with attention += attention + rpe. The correct equation + should be attention = attention + rpe + + """ + + def __init__( + self, + input_dim, + num_heads, + dropout=0.0, + std_scale=None, + scaled_init=False, + tanh_on_mem=False, + use_mem=True, + mini_batches=False, + negative_inf="-inf", + layer_index=-1, + max_relative_position=0, + rpe_old_option=True, + ): + if input_dim % num_heads: + raise ValueError( + "input_dim ({}) must be divisible by num_heads ({})".format( + input_dim, num_heads + ) + ) + + super().__init__() + + embed_dim = input_dim + self.e2h_kv = torch.nn.Linear(input_dim, 2 * input_dim, bias=True) + self.e2h_q = torch.nn.Linear(input_dim, input_dim, bias=True) + self.rpe_old_option = rpe_old_option + if max_relative_position > 0: + self.use_rpe = True + self.rpe_k = RelativePositionEmbedding( + head_dim=input_dim // num_heads, + max_position=max_relative_position, + ) + self.rpe_v = RelativePositionEmbedding( + head_dim=input_dim // num_heads, + max_position=max_relative_position, + ) + else: + self.use_rpe = False + self.rpe_k = None + self.rpe_v = None + if scaled_init: + if layer_index == -1: + gain = 1.0 / math.sqrt(2) + else: + # https://arxiv.org/abs/2005.09684 depthwise initialization + # stablize the training greatly. Use depthwise initialization to + # replace incremental loss. + gain = 1.0 / math.sqrt(layer_index + 1) + torch.nn.init.xavier_uniform_(self.e2h_kv.weight, gain=gain) + torch.nn.init.xavier_uniform_(self.e2h_q.weight, gain=gain) + + self.out_proj = torch.nn.Linear(embed_dim, embed_dim, bias=True) + + self.embed_dim = embed_dim + self.num_heads = num_heads + self.dropout = dropout + + self.head_dim = embed_dim // num_heads + self.scaling = self.head_dim ** -0.5 + + self.std_scale = std_scale + self.use_mem = use_mem + self.mini_batches = mini_batches + self.negative_inf = negative_inf + + if tanh_on_mem: + self.squash_mem = torch.tanh + self.nonlinear_squash_mem = True + else: + self.squash_mem = NoOp() + self.nonlinear_squash_mem = False + + def prepare_qkv( + self, + input: Tensor, + mems: Tensor, + lengths: Tensor, + summary_length: int, + lc_length: int, + ): + # T: right_context length + utterance_length + summary_length + T, B, D = input.shape + mem_length = mems.size(0) + utterance_length = torch.max(lengths) + + right_context_blocks_length = T - utterance_length - summary_length + rc_block = input[:right_context_blocks_length, :, :] + utterance_block = input[right_context_blocks_length : T - summary_length, :, :] + + if B == 1: + padding_mask = None + else: + klengths = lengths + mem_length + right_context_blocks_length + lc_length + padding_mask = lengths_to_padding_mask(lengths=klengths) + + mem_rc_input = torch.cat([mems, rc_block, utterance_block], dim=0) + + # In training lc_length = 0 + key_length = mem_rc_input.size(0) + lc_length + rc_input_sum = input + q = self.e2h_q(rc_input_sum) + kv = self.e2h_kv(mem_rc_input) + k, v = kv.chunk(chunks=2, dim=2) + result_qkv = (q, k, v) + input_shape = (T, B, D) + result_lengths_info = ( + mem_length, + utterance_length, + right_context_blocks_length, + key_length, + ) + if padding_mask is not None: + assert padding_mask.size(0) == B + assert padding_mask.size(1) == key_length + + return result_qkv, input_shape, result_lengths_info, padding_mask + + def prepare_attention_weights( + self, + q: Tensor, + new_k: Tensor, + new_v: Tensor, + input_shape: Tuple[int, int, int], + rpe: Optional[Tensor], + ) -> Tuple[Tensor, Tensor, Tensor]: + T, B, D = input_shape + q = ( + q.contiguous().view(-1, B * self.num_heads, self.head_dim).transpose(0, 1) + * self.scaling + ) + + k = ( + new_k.contiguous() + .view(-1, B * self.num_heads, self.head_dim) + .transpose(0, 1) + ) + + v = ( + new_v.contiguous() + .view(-1, B * self.num_heads, self.head_dim) + .transpose(0, 1) + ) + + attention_weights = torch.bmm(q, k.transpose(1, 2)) + if self.use_rpe and rpe is not None and self.rpe_v is not None: + r_k = self.rpe_k(rpe) + # [q, B*h, d] * [q, k, d] -> [B*h, q, k] + attention_weights_rpe = torch.matmul( + q.transpose(0, 1), r_k.transpose(1, 2) + ).transpose(0, 1) + attention_weights = attention_weights + attention_weights_rpe + attention_weights_float = attention_weights.float() + + return attention_weights, attention_weights_float, v + + def prepare_attention_output( + self, + attention_weights: Tensor, + attention_weights_float: Tensor, + v: Tensor, + input_shape: Tuple[int, int, int], + key_length: int, + padding_mask: Optional[Tensor], + rpe: Optional[Tensor], + ) -> Tensor: + T, B, D = input_shape + if padding_mask is not None: + attention_weights_float = attention_weights_float.view( + B, self.num_heads, T, key_length + ) + attention_weights_float = attention_weights_float.masked_fill( + padding_mask.unsqueeze(1).unsqueeze(2).to(torch.bool), float("-inf") + ) + attention_weights_float = attention_weights_float.view( + B * self.num_heads, T, key_length + ) + + if self.std_scale is not None: + attention_weights_float = attention_suppression( + attention_weights_float, self.std_scale + ) + + attention_weights_float = torch.nn.functional.softmax( + attention_weights_float, dim=-1 + ) + attention_weights = attention_weights_float.type_as(attention_weights) + + attention_probs = torch.nn.functional.dropout( + attention_weights, p=self.dropout, training=self.training + ) + + # [T, key_length, B, n_head]+ [key_length, B, n_head, d_head] + # -> [T, B, n_head, d_head] + attention = torch.bmm(attention_probs, v) + if self.use_rpe and rpe is not None and self.rpe_v is not None: + r_v = self.rpe_v(rpe) + attention_rpe = torch.matmul( + attention_probs.transpose(0, 1), r_v + ).transpose(0, 1) + + if self.rpe_old_option: + attention += attention + attention_rpe + else: + attention = attention + attention_rpe + + assert list(attention.shape) == [B * self.num_heads, T, self.head_dim] + + attention = attention.transpose(0, 1).contiguous().view(T, B, self.embed_dim) + + rc_output_memory = self.out_proj(attention) + return rc_output_memory + + @torch.jit.unused + def forward( + self, + input: Tensor, + lengths: Tensor, + mems: Tensor, + attention_mask: Tensor, + pre_mems: Optional[Tensor] = None, + left_context_key: Optional[Tensor] = None, + left_context_val: Optional[Tensor] = None, + rpe: Optional[Tensor] = None, + ) -> Tuple[Tensor, Tensor, Tensor, Tensor]: + """ + forward function for NoSegAugmentedMemoryMultiheadAttentionBmm in training. + + args: + input: formed in the following way + [right_context_0, right_contex_1, ..., seg_0, seg_1, + ..., summary_0, summary_1,..] + lengths: the length of query which is [seg_0, seg_1, ....] + mems: [mem_0, mem_1, ...]. + attention_mask: attention mask for query = [right_context, query, summary] + key = [mem, right_context, query]. This is only used for traing. + + """ + if self.use_mem: + mem_length = mems.size(0) + summary_length = mem_length + 1 + if pre_mems is not None: + mems = torch.cat([pre_mems, mems], dim=0) + else: + mem_length = 0 + summary_length = 0 + + # In training, lc_length = 0 + if left_context_key is not None: + lc_length = left_context_key.size(0) + else: + lc_length = 0 + results = self.prepare_qkv( + input=input, + mems=mems, + lengths=lengths, + summary_length=summary_length, + lc_length=lc_length, + ) + result_qkv, input_shape, result_lengths_info, padding_mask = results + q, k, v = result_qkv + ( + mem_length, + utterance_length, + right_context_blocks_length, + key_length, + ) = result_lengths_info + + if left_context_key is not None: + # add the cache key and value + new_k = torch.cat( + [ + k[: mem_length + right_context_blocks_length, :, :], + left_context_key, + k[-utterance_length:, :, :], + ], + dim=0, + ) + new_v = torch.cat( + [ + v[: mem_length + right_context_blocks_length, :, :], + left_context_val, + v[-utterance_length:, :, :], + ], + dim=0, + ) + next_k = new_k[mem_length + right_context_blocks_length :, :, :] + next_v = new_v[mem_length + right_context_blocks_length :, :, :] + else: + new_k = k + new_v = v + next_k = None + next_v = None + + attention_weights, attention_weights_float, v = self.prepare_attention_weights( + q=q, + new_k=new_k, + new_v=new_v, + input_shape=input_shape, + rpe=rpe, + ) + + # mask attention + attention_mask = attention_mask.unsqueeze(0) + attention_weights_float = attention_weights_float.masked_fill( + attention_mask, float(self.negative_inf) + ) + + rc_output_memory = self.prepare_attention_output( + attention_weights=attention_weights, + attention_weights_float=attention_weights_float, + v=v, + input_shape=input_shape, + key_length=key_length, + padding_mask=padding_mask, + rpe=rpe, + ) + + if self.use_mem: + # next_m length equals to summary length - 1 + # last memory is ignored + if self.mini_batches: + next_m = rc_output_memory[-summary_length:] + else: + next_m = rc_output_memory[-summary_length:-1] + + next_m = self.squash_mem(next_m) + # rc and output + rc_output = rc_output_memory[:-summary_length] + if not self.nonlinear_squash_mem: + next_m = torch.clamp(next_m, min=-10, max=10) + else: + next_m = mems + rc_output = rc_output_memory + + return rc_output, next_m, next_k, next_v + + @torch.jit.export + def forward_jit( + self, + input: Tensor, + lengths: Tensor, + mems: Tensor, + left_context_key: Tensor, + left_context_val: Tensor, + rpe: Optional[Tensor], + ) -> Tuple[Tensor, Tensor, Tensor, Tensor]: + """ + forward function for NoSegAugmentedMemoryMultiheadAttentionBmm in decoding. + + args: + input: formed in the following way + [right_context_0, right_contex_1, ..., seg_0, seg_1, + ..., summary_0, summary_1,..] + lengths: the length of query which is [seg_0, seg_1, ....] + mems: [mem_0, mem_1, ...]. + left_context_key: left_context for key part. This is only used for online + decoding. In training, this is empty tensor + left_context_val: left_context for value part. This is only used for online + decoding. In training, this is empty tensor + + """ + lc_length = left_context_key.size(0) + + # In decoding, summary_length = 1 or 0 + if self.use_mem: + summary_length = 1 + else: + summary_length = 0 + + results = self.prepare_qkv( + input=input, + mems=mems, + lengths=lengths, + summary_length=summary_length, + lc_length=lc_length, + ) + result_qkv, input_shape, result_lengths_info, padding_mask = results + q, k, v = result_qkv + ( + mem_length, + utterance_length, + right_context_blocks_length, + key_length, + ) = result_lengths_info + + # add the cache key and value + new_k = torch.cat( + [ + k[: mem_length + right_context_blocks_length, :, :], + left_context_key, + k[-utterance_length:, :, :], + ], + dim=0, + ) + new_v = torch.cat( + [ + v[: mem_length + right_context_blocks_length, :, :], + left_context_val, + v[-utterance_length:, :, :], + ], + dim=0, + ) + next_k = new_k[mem_length + right_context_blocks_length :, :, :] + next_v = new_v[mem_length + right_context_blocks_length :, :, :] + + attention_weights, attention_weights_float, v = self.prepare_attention_weights( + q=q, + new_k=new_k, + new_v=new_v, + input_shape=input_shape, + rpe=rpe, + ) + # In online decoding, we don't have attention mask. But we still need + # to disable the attention from summary query to memory + attention_weights_float[:, -1, :mem_length] = float(self.negative_inf) + rc_output_memory = self.prepare_attention_output( + attention_weights=attention_weights, + attention_weights_float=attention_weights_float, + v=v, + input_shape=input_shape, + key_length=key_length, + padding_mask=padding_mask, + rpe=rpe, + ) + + # In decoding, summary length is 1 + if self.use_mem: + next_m = rc_output_memory[-1:] + next_m = self.squash_mem(next_m) + # rc and output + rc_output = rc_output_memory[:-1] + if not self.nonlinear_squash_mem: + next_m = torch.clamp(next_m, min=-10, max=10) + else: + rc_output = rc_output_memory + # empty tensor as input mems + next_m = mems + + return rc_output, next_m, next_k, next_v + + def quantize_(self, params=None): + if params and "per_channel" in params and params["per_channel"]: + qconfig = per_channel_dynamic_qconfig + else: + qconfig = default_dynamic_qconfig + torch.quantization.quantize_dynamic( + self, {torch.nn.Linear: qconfig}, dtype=torch.qint8, inplace=True + ) + return self + + +class NoSegAugmentedMemoryTransformer(nn.Module): + """ + Whole utterance augmented memory transformer. + + This is not pyspeech nn layer. It is used as a module in a master layer where + multiple transformers is used. + """ + + def __init__( + self, + input_dim, + num_heads, + ffn_dim, + dropout_in_attn=0.0, + dropout_on_attn=None, + dropout_on_fc1=None, + dropout_on_fc2=None, + activation_fn="relu", + tanh_on_mem=False, + std_scale=None, + scaled_init=False, + segment_size=128, + use_mem=True, + mini_batches=False, + negative_inf="-inf", + layer_index=-1, + summarization_method="mean", + max_relative_position=0, + rpe_old_option=True, + ): + super(NoSegAugmentedMemoryTransformer, self).__init__() + + self.attention = NoSegAugmentedMemoryMultiheadAttentionBmm( + input_dim=input_dim, + num_heads=num_heads, + dropout=dropout_in_attn, + scaled_init=scaled_init, + tanh_on_mem=tanh_on_mem, + std_scale=std_scale, + use_mem=use_mem, + mini_batches=mini_batches, + negative_inf=negative_inf, + layer_index=layer_index, + max_relative_position=max_relative_position, + ) + self.dropout = nn.Dropout(dropout_on_attn) + self.pos_ff = PositionwiseFF( + input_dim=input_dim, + ffn_dim=ffn_dim, + dropout_on_fc1=dropout_on_fc1, + dropout_on_fc2=dropout_on_fc2, + activation_fn=activation_fn, + ) + self.layer_norm_pre = Fp32LayerNorm(input_dim) + self.layer_norm = Fp32LayerNorm(input_dim) + self.segment_size = segment_size + self.use_mem = use_mem + + self.memory_op = SummarizationLayer( + summarization_method, segment_size, input_dim + ) + + def set_mini_batches(self, mini_batches): + self.attention.mini_batches = mini_batches + + def gen_summary_queries(self, input): + sum_input = self.memory_op(input) + return sum_input + + def pre_attention_ops(self, input, right_context_blocks): + rc_length = right_context_blocks.size(0) + input_length = input.size(0) + + rc_and_input = torch.cat([right_context_blocks, input], dim=0) + residual_input = rc_and_input + rc_and_input = self.layer_norm_pre(rc_and_input) + + query_input = rc_and_input[-input_length:, :, :] + return rc_length, input_length, residual_input, query_input, rc_and_input + + def after_attention_ops(self, attention_output, residual_input): + output = self.dropout(attention_output) + output = output + residual_input + output = self.pos_ff(output) + output = self.layer_norm(output) + return output + + @torch.jit.export + def forward_jit( + self, + input: Tensor, + lengths: Tensor, + mems: Tensor, + left_context_key: Tensor, + left_context_val: Tensor, + right_context_blocks: Tensor, + rpe: Optional[Tensor], + ) -> Tuple[Tensor, Tensor, Tensor, Tensor, Tensor]: + + results = self.pre_attention_ops(input, right_context_blocks) + rc_length, input_length, residual_input, query_input, rc_and_input = results + + # In online decoding, the summary query size is always 1 or 0 + if self.use_mem: + summary_query = self.gen_summary_queries(query_input) + summary_query = summary_query[0:1, :, :] + rc_qu_su = torch.cat([rc_and_input, summary_query], dim=0) + else: + rc_qu_su = rc_and_input + + rc_output, next_m, next_k, next_v = self.attention.forward_jit( + input=rc_qu_su, + lengths=lengths, + mems=mems, + left_context_key=left_context_key, + left_context_val=left_context_val, + rpe=rpe, + ) + rc_output = self.after_attention_ops(rc_output, residual_input) + results = ( + rc_output[-input_length:, :, :], + next_m, + rc_output[0:rc_length, :, :], + next_k, + next_v, + ) + return results + + @torch.jit.unused + def forward( + self, + input, + lengths, + mems, + right_context_blocks, + attention_mask, + pre_mems, + left_context_key, + left_context_val, + rpe, + ): + + results = self.pre_attention_ops(input, right_context_blocks) + rc_length, input_length, residual_input, query_input, rc_and_input = results + if self.use_mem: + summary_query = self.gen_summary_queries(query_input) + rc_qu_su = torch.cat([rc_and_input, summary_query], dim=0) + else: + rc_qu_su = rc_and_input + + rc_output, next_m, next_k, next_v = self.attention( + input=rc_qu_su, + lengths=lengths, + mems=mems, + attention_mask=attention_mask, + pre_mems=pre_mems, + left_context_key=left_context_key, + left_context_val=left_context_val, + rpe=rpe, + ) + + # [TODO] Note memory did not go through pos_ff. What happen if we pass + # memory through the pos_ff as well? + rc_output = self.after_attention_ops(rc_output, residual_input) + results = ( + rc_output[-input_length:, :, :], + next_m, + rc_output[0:rc_length, :, :], + next_k, + next_v, + ) + + return results + + +class NoSegAugmentedMemoryTransformerEncoderLayer(FairseqEncoder): + """ + Whole utterance augmented memory transformer encoder layer. This is a master layer + where we can define multiple augmented memory transformers. There are two reasons + to setup the master layer. + 1. We only need to define once about the attention mask. All the layers in the master + layer share the same mask. + 2. pyspeech nn layer has special input and output format. Defining one master layer is + easier to passing memory between different layes inside the master layer + + args: + input_dim: input embedding dimension + num_heads: number of heads in multihead self-attention + ffn_dim: ffn dimension in FFN layer + num_layers: number of augmented memory transformer layers + dropout_in_attn: dropout used in multi-head self-attention + dropout_on_attn: dropout used for output from te multihead self-attention + dropout_on_fc1: dropout used in FFN layer for the first linear layer + dropout_on_fc2: dropout used in FFN layer for the second linear layer + segment_size: segment size for each segment + context_config: (left_context_size, right_context_size) defines the surround context size + for each segment + max_memory_size: maximum memory size used for each segment + scaled_init: whether use scaled init for weight initialization in attention layer + std_scale: if std_scale is not None. The weak attention suppression is + turned on. For std_scale = 0.5, all the attention smaller than + mean + 0.5 * std will be suppressed. + activation_fn: activation function used in FFN layer. [ReLU, GELU] supported + tanh_on_mem: whether use tanh on memory + mini_batches: use mini-btach training + negative_inf: the negative infinity value used in attention masking. default is "-inf". + For some situation, e.g. LM. it is better to use "-1e8" to avoid nan issue. + summarization_method: method to generate segment summrization embedding + max_relative_position: max relatie position for relative position embedding + rpe_old_option: To be compatible with previous model. The previous model + was trained with attention += attention + rpe. The correct equation + should be attention = attention + rpe + [TODO]: remove the rpe_old_option by the end of 2021 Q1. + + """ + + def __init__( + self, + input_dim, + num_heads, + ffn_dim, + num_layers=1, + dropout_in_attn=0.0, + dropout_on_attn=0.0, + dropout_on_fc1=0.0, + dropout_on_fc2=0.0, + segment_size=128, + context_config=(0, 0), + max_memory_size=0, + scaled_init=True, + std_scale=None, + activation_fn="relu", + tanh_on_mem=False, + mini_batches=False, + negative_inf="-inf", + deep_init=True, + summarization_method="mean", + max_relative_position=0, + rpe_old_option=True, + ): + super().__init__(None) + if input_dim % num_heads: + raise ValueError( + "input_dim ({}) must be divisible by num_heads ({})".format( + input_dim, num_heads + ) + ) + + # we used to support growing memory size. However, it will cause + # cross stream batching failure. Now we need to have exact max memory size + if max_memory_size < 0: + raise ValueError("max_memory_size must be >= 0") + + # Only assign right_context. In decoding, left context will be cached. + # No need to let the online decoder to re-assign the left context + self.left_context, self.right_context = context_config + self.segment_size = segment_size + self.memory_dim = input_dim + self.max_memory_size = max_memory_size + self.mini_batches = mini_batches + if self.max_memory_size != 0: + self.use_mem = True + else: + self.use_mem = False + + self.memory_op = SummarizationLayer( + summarization_method, segment_size, input_dim + ) + + self.layers = torch.nn.ModuleList() + self.num_layers = num_layers + self.max_relative_position = max_relative_position + if self.max_relative_position > 0: + self.use_rpe = True + else: + self.use_rpe = False + for i in range(self.num_layers): + if deep_init: + layer_index = i + else: + layer_index = -1 + + self.layers.append( + NoSegAugmentedMemoryTransformer( + num_heads=num_heads, + input_dim=input_dim, + ffn_dim=ffn_dim, + dropout_in_attn=dropout_in_attn, + dropout_on_attn=dropout_on_attn, + dropout_on_fc1=dropout_on_fc1, + dropout_on_fc2=dropout_on_fc2, + segment_size=segment_size, + std_scale=std_scale, + activation_fn=activation_fn, + tanh_on_mem=tanh_on_mem, + scaled_init=scaled_init, + use_mem=self.use_mem, + mini_batches=mini_batches, + negative_inf=negative_inf, + layer_index=layer_index, + summarization_method=summarization_method, + max_relative_position=max_relative_position, + rpe_old_option=rpe_old_option, + ) + ) + + def set_mini_batches(self, mini_batches): + # handy function only used for unit test + self.mini_batches = mini_batches + for layer in self.layers: + layer.set_mini_batches(mini_batches) + + def _get_relative_position( + self, + input: Tensor, + max_relative_position: int, + left_context_length: int, + past_length: int, + is_decoding: bool, + ): + # For training, we copy the right context to the start of the utterance + # First dimension in distance is corresponding to query. + # [right context, utterance, summary vector] + # Second dimension in distance is corresponding to key. + # [Memory bank, right context, utterance] + # For summary vector in query part, the distance with + # all other position is 2*max_position. For memory bank in key, + # the distance with all other positions is 0. + + T, B, D = input.shape + num_segs = math.ceil((T - self.right_context) / self.segment_size) + + # utterance + u_st = past_length * self.segment_size + u_ed = u_st + T + utterance_ranges = torch.arange(u_st, u_ed - self.right_context) + + # left context. Only in minibatch or decoding + left_context_ranges = torch.arange(u_st - left_context_length, u_st) + + # Right context block + # right context + utterance + right_context_blocks = [] + for i in range(0, num_segs - 1): + st = (i + 1) * self.segment_size + u_st + ed = st + self.right_context + assert ed < u_ed + temp = torch.arange(st, ed) + right_context_blocks.append(temp) + right_context_blocks.append(torch.arange(u_ed - self.right_context, u_ed)) + right_context_ranges = torch.cat(right_context_blocks) + + if self.use_mem: + # Memory bank + # The position for memory -n, .., -1 + if is_decoding: + memory_size = min(past_length, self.max_memory_size) + else: + memory_size = num_segs + past_length - 1 + memory_bank_ranges = torch.arange( + -max_relative_position - 1, -max_relative_position - 1 - memory_size, -1 + ) + + # summary vector + # The position for summary vector as the T+max_relative_position+1. + # After the clamping, the relative position is max_relative_position + summary_pos_st = u_ed + max_relative_position + 1 + summary_vector_ranges = torch.arange( + summary_pos_st, summary_pos_st + num_segs + ) + + key_ranges = torch.cat( + [ + memory_bank_ranges, + right_context_ranges, + left_context_ranges, + utterance_ranges, + ] + ) + + query_ranges = torch.cat( + [right_context_ranges, utterance_ranges, summary_vector_ranges] + ) + else: + key_ranges = torch.cat( + [right_context_ranges, left_context_ranges, utterance_ranges] + ) + + query_ranges = torch.cat([right_context_ranges, utterance_ranges]) + + distance = key_ranges[None, :] - query_ranges[:, None] + distance_clamp = ( + torch.clamp(distance, -max_relative_position, max_relative_position) + + max_relative_position + ) + distance_clamp = distance_clamp.to(input.device).long().detach() + return distance_clamp + + def _get_attention_mask(self, input, past_length=0, left_context_cache=0): + # attention mask for each query contains three parts: + # 1. memory part + # 2. left_context + segment + # 3. right_context_block + # so for each segment and its correspoinding right context block, + # the attention matrix is formed by 9 parts: + # [0, m, 0, 0, right_context, 0, 0, seg, 0] + # [before memory, memory, after memory, before right context, right_context, + # after right context, before seg, seg, after seg] + # + # Query is formed in the way as [right_context_blocks, utterance, summary] + # + # Note: put m and right_context before segment is convenient + # for padding_mask operation. + # Key lengths = m_length + right_context_block_length + lengths + utterance_length, batch_size, _ = input.shape + summary_length = math.ceil(utterance_length / self.segment_size) + num_segs = summary_length + rc_length = self.right_context * num_segs + rc = self.right_context + lc = self.left_context + + # using mini-batches, there is left context cache available for current + # sequence. + lcc = left_context_cache + + # max_memory_size is 0 then we don't have memory and summary + # past_length is the memory carry from previous sequence + if self.use_mem: + mem_length = num_segs - 1 + past_length + else: + mem_length = 0 + rc_mask = [] + query_mask = [] + summary_mask = [] + for j in range(0, num_segs): + ssize = min(self.segment_size, utterance_length - j * self.segment_size) + + rc_size = rc + rc_mat = [] + q_mat = [] + s_mat = [] + m_start = max(j + past_length - self.max_memory_size, 0) + + # max_memory_size is 0, then we don't use memory + if self.use_mem: + # part 0: before memory + rc_mat.append(input.new_zeros(rc_size, m_start)) + q_mat.append(input.new_zeros(ssize, m_start)) + s_mat.append(input.new_zeros(1, m_start)) + + # part 1: memory + col_1 = j + past_length - m_start + rc_mat.append(torch.ones(rc_size, col_1, device=input.device)) + q_mat.append(torch.ones(ssize, col_1, device=input.device)) + # based on D22875746, disable summary query attention + # on memeory is better for long form utterance + s_mat.append(input.new_zeros(1, col_1)) + + # part 2: after memory + col_2 = mem_length - (j + past_length) + rc_mat.append(input.new_zeros(rc_size, col_2)) + q_mat.append(input.new_zeros(ssize, col_2)) + s_mat.append(input.new_zeros(1, col_2)) + + # part 3: before right context + rc_start = j * rc + rc_mat.append(input.new_zeros(rc_size, rc_start)) + q_mat.append(input.new_zeros(ssize, rc_start)) + s_mat.append(input.new_zeros(1, rc_start)) + + # part 4: right context + rc_end = rc_start + rc + col_4 = rc + rc_mat.append(torch.ones(rc_size, col_4, device=input.device)) + q_mat.append(torch.ones(ssize, col_4, device=input.device)) + s_mat.append(torch.ones(1, col_4, device=input.device)) + + # part 5: after right context + col_5 = rc_length - rc_end + rc_mat.append(input.new_zeros(rc_size, col_5)) + q_mat.append(input.new_zeros(ssize, col_5)) + s_mat.append(input.new_zeros(1, col_5)) + + # part 6: before query segment + seg_start = max(j * self.segment_size + lcc - lc, 0) + rc_mat.append(input.new_zeros(rc_size, seg_start)) + q_mat.append(input.new_zeros(ssize, seg_start)) + s_mat.append(input.new_zeros(1, seg_start)) + + # part 7: query segment + # note: right context is put in right context block + # here we only need to consider about left context + seg_end = min((j + 1) * self.segment_size + lcc, utterance_length + lcc) + col_7 = seg_end - seg_start + rc_mat.append(torch.ones(rc_size, col_7, device=input.device)) + q_mat.append(torch.ones(ssize, col_7, device=input.device)) + s_mat.append(torch.ones(1, col_7, device=input.device)) + + # part 8: after query segment + col_8 = utterance_length + lcc - seg_end + rc_mat.append(input.new_zeros(rc_size, col_8)) + q_mat.append(input.new_zeros(ssize, col_8)) + s_mat.append(input.new_zeros(1, col_8)) + + rc_mask.append(torch.cat(rc_mat, dim=1)) + query_mask.append(torch.cat(q_mat, dim=1)) + summary_mask.append(torch.cat(s_mat, dim=1)) + + # no memory, then we don't need summary either + if self.use_mem: + attention_mask = ( + 1 + - torch.cat( + [ + torch.cat(rc_mask, dim=0), + torch.cat(query_mask, dim=0), + torch.cat(summary_mask, dim=0), + ], + dim=0, + ) + ).to(torch.bool) + else: + attention_mask = ( + 1 + - torch.cat( + [torch.cat(rc_mask, dim=0), torch.cat(query_mask, dim=0)], dim=0 + ) + ).to(torch.bool) + + return attention_mask + + @torch.jit.export + def init_state( + self, batch_size: int, device: Optional[Device] = None + ) -> List[Tensor]: + empty_memory = torch.zeros( + self.num_layers, + self.max_memory_size, + batch_size, + self.memory_dim, + device=device, + ) + left_context_key = torch.zeros( + self.num_layers, + self.left_context, + batch_size, + self.memory_dim, + device=device, + ) + left_context_val = torch.zeros( + self.num_layers, + self.left_context, + batch_size, + self.memory_dim, + device=device, + ) + past_length = torch.zeros(1, batch_size, dtype=torch.int32, device=device) + + return [empty_memory, left_context_key, left_context_val, past_length] + + @torch.jit.export + def batch_state(self, states: List[List[Tensor]]) -> List[Tensor]: + if len(states) == 0: + return [] + batched_m = [] + batched_lc_key = [] + batched_lc_val = [] + batched_past_length = [] + for state in states: + if len(state) == 0: + continue + m, lc_key, lc_val, past_length = state + batched_m.append(m) + batched_lc_key.append(lc_key) + batched_lc_val.append(lc_val) + batched_past_length.append(past_length) + + if ( + (len(batched_m) == 0) + or (len(batched_lc_key) == 0) + or (len(batched_lc_val) == 0) + or (len(batched_past_length) == 0) + ): + return [ + torch.tensor([]), + torch.tensor([]), + torch.tensor([]), + torch.tensor([]), + ] + + batched_m = torch.cat(batched_m, dim=2) + batched_lc_key = torch.cat(batched_lc_key, dim=2) + batched_lc_val = torch.cat(batched_lc_val, dim=2) + batched_past_length = torch.cat(batched_past_length, dim=1) + return [batched_m, batched_lc_key, batched_lc_val, batched_past_length] + + @torch.jit.export + def reorder_state(self, state: List[Tensor], indices: Tensor) -> List[Tensor]: + if len(state) == 0: + return [] + m, lc_key, lc_val, past_length = state + indices = indices.to(device=m.device) + reord_m = torch.index_select(m, 2, indices) + reord_lc_key = torch.index_select(lc_key, 2, indices) + reord_lc_val = torch.index_select(lc_val, 2, indices) + reord_past_length = torch.index_select(past_length, 1, indices) + return [reord_m, reord_lc_key, reord_lc_val, reord_past_length] + + @torch.jit.export + def reset_state(self, state: List[Tensor], indices: Tensor) -> List[Tensor]: + m, lc_key, lc_val, past_length = state + m = m.index_fill(dim=2, index=indices, value=0.0) + lc_key = lc_key.index_fill(dim=2, index=indices, value=0.0) + lc_val = lc_val.index_fill(dim=2, index=indices, value=0.0) + past_length = past_length.index_fill(dim=1, index=indices, value=0) + + return [m, lc_key, lc_val, past_length] + + @torch.jit.export + def state_size(self) -> int: + return 4 + + @torch.jit.export + def batch_size_in_state( + self, state: Optional[List[Tensor]], sloppy: bool = True + ) -> Optional[int]: + if state is None: + return None + return state[0].size(2) + + def gen_summary_queries(self, input): + sum_input = self.memory_op(input) + return sum_input + + def _gen_right_context_padded_input(self, input): + # This function deals with input that is already + # padded with right context (e.g. minibatch training) + right_context_blocks = [] + T, B, D = input.shape + num_segs = math.ceil((T - self.right_context) / self.segment_size) + for i in range(0, num_segs - 1): + st = (i + 1) * self.segment_size + ed = st + self.right_context + assert ed < T + temp = input[st:ed, :, :] + right_context_blocks.append(temp) + + # last segment right context is already available + right_context_blocks.append(input[T - self.right_context :, :, :]) + return torch.cat(right_context_blocks, dim=0) + + def _gen_segs_right_context(self, input, lengths): + segments = [] + T, B, D = input.size() + nT = T - self.right_context + + # assume input is right context padded + num_segs = math.ceil(nT / self.segment_size) + # pad zeros to the utterance to make sure each + # segment has the same right context. For the + for i in range(0, num_segs - 1): + st = i * self.segment_size + ed = min(T, st + self.segment_size + self.right_context) + temp = input[st:ed, :, :] + rest_lengths = torch.clamp( + lengths - self.segment_size, min=0, max=nT - (i + 1) * self.segment_size + ) + segments.append((temp, lengths - rest_lengths + self.right_context)) + lengths = rest_lengths + + last_seg = input[st + self.segment_size :, :, :] + segments.append((last_seg, rest_lengths + self.right_context)) + + return segments + + @torch.jit.unused + def forward( + self, input: Tensor, padding_masks: Tensor, state: Optional[List[Tensor]] = None + ) -> Tuple[Tensor, Tensor, List[Tensor], List[Tensor]]: + # Xutai: originally the second argument is lengths. + lengths = (~padding_masks).sum(dim=1).long() + # mini batch training. + if self.mini_batches: + return self.forward_mini_batches(input, lengths, state) + + # regular full sequence training. Note, assume the right context in provided + # in the input. + T, B, D = input.size() + right_context_blocks = self._gen_right_context_padded_input(input) + + # generate the relative positional embedding + if self.use_rpe: + rpe = self._get_relative_position( + input=input, + max_relative_position=self.max_relative_position, + left_context_length=0, + past_length=0, + is_decoding=False, + ) + else: + rpe = None + input = input[: T - self.right_context, :, :] + + attention_mask = self._get_attention_mask(input) + + # firt layer use each segment mean as memory + # ignore the last one seg average + if self.use_mem: + mems = self.gen_summary_queries(input)[:-1, :, :] + else: + mems = torch.zeros(0, input.size(1), input.size(2), device=input.device) + mems = mems.type_as(input) + + output = input + all_outputs = [] + + for layer in self.layers: + output, mems, right_context_blocks, _, _ = layer( + input=output, + lengths=lengths, + attention_mask=attention_mask, + mems=mems, + right_context_blocks=right_context_blocks, + pre_mems=None, + left_context_key=None, + left_context_val=None, + rpe=rpe, + ) + all_outputs.append(output) + return output, padding_masks, [], all_outputs + + def forward_jit_mini_batch_init( + self, + seg: Tensor, + state: Optional[List[Tensor]] = None, + is_decoding: bool = False, + ): + # Prepare state. In whole sequence training, state is ignored. + # For minibatch training, we need to prepare state + if state is None: + state = self.init_state(batch_size=seg.size(1), device=seg.device) + if seg.dtype == torch.half: + state = [state[0].half(), state[1].half(), state[2].half(), state[3]] + + if self.use_mem: + # note input average only on seg, not on right context + # first layer use each segmetn mean as memory. the last + # one segment average is used in state + full_mems = self.gen_summary_queries(seg) + if is_decoding: + mems = full_mems[0:1, :, :] + state_mems = torch.cat([state[0][0], mems], dim=0) + else: + mems = full_mems[:-1, :, :] + state_mems = torch.cat([state[0][0], full_mems], dim=0) + else: + mems = state[0][0] + state_mems = mems + + # track processed segment number or memory number + # the same batch as the same bumber of past length + past_length = state[3][0][0].item() + past_left_context = min(past_length * self.segment_size, self.left_context) + past_length = min(self.max_memory_size, past_length) + + return state, mems, state_mems, past_length, past_left_context + + def state_update_before( + self, layer: int, state: List[Tensor], past_length: int, past_left_context: int + ): + pre_mems = state[0][layer][self.max_memory_size - past_length :, :, :] + lc_key = state[1][layer][self.left_context - past_left_context :, :, :] + lc_val = state[2][layer][self.left_context - past_left_context :, :, :] + return pre_mems, lc_key, lc_val + + def state_update_after( + self, + layer: int, + state: List[Tensor], + mems: Tensor, + next_key: Tensor, + next_val: Tensor, + mems_list: List[Tensor], + lc_key_list: List[Tensor], + lc_val_list: List[Tensor], + ): + # mems is used for next layer + if layer < self.num_layers - 1: + state_mems = torch.cat([state[0][layer + 1], mems], dim=0) + mems_list.append(state_mems[-self.max_memory_size :, :, :]) + + # when mems pass to next sequence, we need the last memory. when mems + # use for the next layer, we can ignore the last memory + mems = mems[:-1, :, :] + + # note state[1][i] and state[2][i] original length equals to self.left_context + new_k = torch.cat([state[1][layer], next_key], dim=0) + new_v = torch.cat([state[2][layer], next_val], dim=0) + lc_key_list.append(new_k[-self.left_context :, :, :]) + lc_val_list.append(new_v[-self.left_context :, :, :]) + return mems_list, lc_key_list, lc_val_list, mems + + def state_update_after_loop( + self, + state: List[Tensor], + mems_list: List[Tensor], + lc_key_list: List[Tensor], + lc_val_list: List[Tensor], + update_length: int, + ): + state[0] = torch.stack(mems_list, dim=0) + state[1] = torch.stack(lc_key_list, dim=0) + state[2] = torch.stack(lc_val_list, dim=0) + state[3] = state[3] + update_length + return state + + @torch.jit.unused + def forward_mini_batches( + self, input: Tensor, lengths: Tensor, state: Optional[List[Tensor]] = None + ) -> Tuple[Tensor, Tensor, List[Tensor], List[Tensor]]: + T, B, D = input.size() + + # input without right context + seg = input[: T - self.right_context, :, :] + + # get right context blocks + right_context_blocks = self._gen_right_context_padded_input(input) + + mems_list = [] + lc_key_list = [] + lc_val_list = [] + results = self.forward_jit_mini_batch_init(seg, state, False) + state, mems, state_mems, past_length, past_left_context = results + + # relative position embedding + if self.use_rpe: + rpe = self._get_relative_position( + input=input, + max_relative_position=self.max_relative_position, + left_context_length=past_left_context, + past_length=past_length, + is_decoding=False, + ) + else: + rpe = None + + # get attention mask based on seg (not include right context) and available + # left context + attention_mask = self._get_attention_mask(seg, past_length, past_left_context) + mems_list.append(state_mems[-self.max_memory_size :, :, :]) + output = seg + i = 0 + all_outputs = [] + for layer in self.layers: + # In order to make cross stream batching work, mem, left context key + # and left context value in the state should always be the same shape. + # We use the past length to track the processed segment number. In this + # way, we take out the essential memory, left context key and left + # context val from the state. After finish the forward for current segment + # we add the new memory, left context key and left context value into the + # staate and trim out the oldest part to keep the shape consistent. + pre_mems, lc_key, lc_val = self.state_update_before( + i, state, past_length, past_left_context + ) + + output, mems, right_context_blocks, next_key, next_val = layer.forward( + input=output, + lengths=lengths, + attention_mask=attention_mask, + mems=mems, + right_context_blocks=right_context_blocks, + pre_mems=pre_mems, + left_context_key=lc_key, + left_context_val=lc_val, + rpe=rpe, + ) + all_outputs.append(output) + mems_list, lc_key_list, lc_val_list, mems = self.state_update_after( + layer=i, + state=state, + mems=mems, + next_key=next_key, + next_val=next_val, + mems_list=mems_list, + lc_key_list=lc_key_list, + lc_val_list=lc_val_list, + ) + + i += 1 + + # update state + update_length = math.ceil((T - self.right_context) / self.segment_size) + state = self.state_update_after_loop( + state=state, + mems_list=mems_list, + lc_key_list=lc_key_list, + lc_val_list=lc_val_list, + update_length=update_length, + ) + + return output, lengths, state, all_outputs + + def forward_jit_test( + self, input: Tensor, lengths: Tensor, state: Optional[List[Tensor]] = None + ) -> Tuple[Tensor, Tensor, List[Tensor]]: + """ + This one simulate sequence encoder forward jit. This is for unit test purpose. + It is not used in training or decoding. Note, extra_right_context is set in + the model. In unit test, input = [utterance, right_context], lengths = + [utterance_length]. + args: + input: input utterance + lengths: utterance input length + state: None here. input is whole utterance + """ + # [TODO] sequence_to_segment has bug in lengths. + seg_src_tokens_lengths = self._gen_segs_right_context(input, lengths) + + seg_enc_tokens_lengths: List[Tuple[Tensor, Tensor]] = [] + state: Optional[List[Tensor]] = None + for seg_src_tokens, seg_src_lengths in seg_src_tokens_lengths: + seg_enc_tokens, seg_enc_lengths, state = self.forward_jit( + input=seg_src_tokens, lengths=seg_src_lengths, state=state + ) + seg_enc_tokens_lengths.append((seg_enc_tokens, seg_enc_lengths)) + + enc_tokens, enc_lengths = segments_to_sequence( + segments=seg_enc_tokens_lengths, time_axis=0 + ) + + state = [] # returns trivial state + + return enc_tokens, enc_lengths, state + + @torch.jit.export + def forward_jit( + self, input: Tensor, lengths: Tensor, state: Optional[List[Tensor]] = None + ) -> Tuple[Tensor, Tensor, List[Tensor]]: + """ + Forward helper for online decoding. + + args: + input: [seg, right_context]. We assume in online we + always padding the right context to the preset right context size. + For the last segment, we may have short segment size, but right + context size is the same as other segments + lengths: utterance input length is the utterance segment length and + right context size + state: [memory, left_context_key, left_context_val]. To improve throughput, + in addition to memory, we also cache key and value for left_context in + multihead self-attention + """ + # In online decoding, input = [segment, right_context] + # Lengths = [segment_length, right_context_length] + # so we need strip right context in output + T, B, D = input.size() + rc_str = T - self.right_context + rc_end = T + right_context_blocks = input[rc_str:rc_end, :, :] + seg = input[:rc_str, :, :] + lengths = torch.clamp(lengths - self.right_context, min=0) + mems_list = [] + lc_key_list = [] + lc_val_list = [] + + results = self.forward_jit_mini_batch_init(seg, state, True) + state, mems, state_mems, past_length, past_left_context = results + + # relative position embedding + if self.use_rpe: + rpe = self._get_relative_position( + input=input, + max_relative_position=self.max_relative_position, + left_context_length=past_left_context, + past_length=past_length, + is_decoding=True, + ) + else: + rpe = None + + # memory for first layer. + mems_list.append(state_mems[-self.max_memory_size :, :, :]) + output = seg + i = 0 + for layer in self.layers: + # In order to make cross stream batching work, mem, left context key + # and left context value in the state should always be the same shape. + # We use the past length to track the processed segment number. In this + # way, we take out the essential memory, left context key and left + # context val from the state. After finish the forward for current segment + # we add the new memory, left context key and left context value into the + # staate and trim out the oldest part to keep the shape consistent. + true_mems, lc_key, lc_val = self.state_update_before( + layer=i, + state=state, + past_length=past_length, + past_left_context=past_left_context, + ) + + output, mems, right_context_blocks, next_key, next_val = layer.forward_jit( + input=output, + lengths=lengths, + mems=true_mems, + right_context_blocks=right_context_blocks, + left_context_key=lc_key, + left_context_val=lc_val, + rpe=rpe, + ) + # mems is used for next layer + mems_list, lc_key_list, lc_val_list, _ = self.state_update_after( + layer=i, + state=state, + mems_list=mems_list, + mems=mems, + next_key=next_key, + next_val=next_val, + lc_key_list=lc_key_list, + lc_val_list=lc_val_list, + ) + i += 1 + + # update state + state = self.state_update_after_loop( + state=state, + mems_list=mems_list, + lc_key_list=lc_key_list, + lc_val_list=lc_val_list, + update_length=1, + ) + + return output, lengths, state + + def quantize_(self, params=None): + if params and "per_channel" in params and params["per_channel"]: + qconfig = per_channel_dynamic_qconfig + else: + qconfig = default_dynamic_qconfig + torch.quantization.quantize_dynamic( + self, {torch.nn.Linear: qconfig}, dtype=torch.qint8, inplace=True + ) + return self + + +# ------------------------------------------------------------------------------ +# Emformer encoder for seq2seq model +# This is a wrapper over the original emformer +# ------------------------------------------------------------------------------ +def emformer_encoder(klass): + class SpeechEncoder(klass): + def __init__(self, args): + super().__init__(args) + stride = SpeechEncoder.conv_layer_stride(args) + trf_left_context = args.segment_left_context // stride + trf_right_context = args.segment_right_context // stride + context_config = [trf_left_context, trf_right_context] + self.transformer_layers = nn.ModuleList( + [ + NoSegAugmentedMemoryTransformerEncoderLayer( + input_dim=args.encoder_embed_dim, + num_heads=args.encoder_attention_heads, + ffn_dim=args.encoder_ffn_embed_dim, + num_layers=args.encoder_layers, + dropout_in_attn=args.dropout, + dropout_on_attn=args.dropout, + dropout_on_fc1=args.dropout, + dropout_on_fc2=args.dropout, + activation_fn=args.activation_fn, + context_config=context_config, + segment_size=args.segment_length, + max_memory_size=args.max_memory_size, + scaled_init=True, # TODO: use constant for now. + tanh_on_mem=args.amtrf_tanh_on_mem, + ) + ] + ) + + def forward(self, src_tokens, src_lengths): + encoder_out = super().forward(src_tokens, src_lengths) + output = encoder_out["encoder_out"][0] + encoder_padding_masks = encoder_out["encoder_padding_mask"][0] + + # This is because that in the original implementation + # the output didn't consider the last segment as right context. + encoder_padding_masks = encoder_padding_masks[:, : output.size(0)] + + return { + "encoder_out": [output], + "encoder_padding_mask": [encoder_padding_masks], + "encoder_embedding": [], + "encoder_states": [], + "src_tokens": [], + "src_lengths": [], + } + + @staticmethod + def conv_layer_stride(args): + # TODO: make it configurable from the args + return 4 + + SpeechEncoder.__name__ = klass.__name__ + return SpeechEncoder diff --git a/fairseq/models/speech_to_text/s2t_transformer.py b/fairseq/models/speech_to_text/s2t_transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..5c935efaf5ef5fbf03479db6280f60aeeea5e6eb --- /dev/null +++ b/fairseq/models/speech_to_text/s2t_transformer.py @@ -0,0 +1,496 @@ +#!/usr/bin/env python3 + +import logging +import math +from typing import Dict, List, Optional, Tuple +from pathlib import Path + +import torch +import torch.nn as nn +from fairseq import checkpoint_utils, utils +from fairseq.data.data_utils import lengths_to_padding_mask +from fairseq.models import ( + FairseqEncoder, + FairseqEncoderDecoderModel, + register_model, + register_model_architecture, +) +from fairseq.models.transformer import Embedding, TransformerDecoder +from fairseq.modules import ( + FairseqDropout, + LayerNorm, + PositionalEmbedding, + TransformerEncoderLayer, +) +from torch import Tensor + + +logger = logging.getLogger(__name__) + + +class Conv1dSubsampler(nn.Module): + """Convolutional subsampler: a stack of 1D convolution (along temporal + dimension) followed by non-linear activation via gated linear units + (https://arxiv.org/abs/1911.08460) + + Args: + in_channels (int): the number of input channels + mid_channels (int): the number of intermediate channels + out_channels (int): the number of output channels + kernel_sizes (List[int]): the kernel size for each convolutional layer + """ + + def __init__( + self, + in_channels: int, + mid_channels: int, + out_channels: int, + kernel_sizes: List[int] = (3, 3), + ): + super(Conv1dSubsampler, self).__init__() + self.n_layers = len(kernel_sizes) + self.conv_layers = nn.ModuleList( + nn.Conv1d( + in_channels if i == 0 else mid_channels // 2, + mid_channels if i < self.n_layers - 1 else out_channels * 2, + k, + stride=2, + padding=k // 2, + ) + for i, k in enumerate(kernel_sizes) + ) + + def get_out_seq_lens_tensor(self, in_seq_lens_tensor): + out = in_seq_lens_tensor.clone() + for _ in range(self.n_layers): + out = ((out.float() - 1) / 2 + 1).floor().long() + return out + + def forward(self, src_tokens, src_lengths): + bsz, in_seq_len, _ = src_tokens.size() # B x T x (C x D) + x = src_tokens.transpose(1, 2).contiguous() # -> B x (C x D) x T + for conv in self.conv_layers: + x = conv(x) + x = nn.functional.glu(x, dim=1) + _, _, out_seq_len = x.size() + x = x.transpose(1, 2).transpose(0, 1).contiguous() # -> T x B x (C x D) + return x, self.get_out_seq_lens_tensor(src_lengths) + + +@register_model("s2t_transformer") +class S2TTransformerModel(FairseqEncoderDecoderModel): + """Adapted Transformer model (https://arxiv.org/abs/1706.03762) for + speech-to-text tasks. The Transformer encoder/decoder remains the same. + A trainable input subsampler is prepended to the Transformer encoder to + project inputs into the encoder dimension as well as downsample input + sequence for computational efficiency.""" + + def __init__(self, encoder, decoder): + super().__init__(encoder, decoder) + + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + # input + parser.add_argument( + "--conv-kernel-sizes", + type=str, + metavar="N", + help="kernel sizes of Conv1d subsampling layers", + ) + parser.add_argument( + "--conv-channels", + type=int, + metavar="N", + help="# of channels in Conv1d subsampling layers", + ) + # Transformer + parser.add_argument( + "--activation-fn", + type=str, + default="relu", + choices=utils.get_available_activation_fns(), + help="activation function to use", + ) + parser.add_argument( + "--dropout", type=float, metavar="D", help="dropout probability" + ) + parser.add_argument( + "--attention-dropout", + type=float, + metavar="D", + help="dropout probability for attention weights", + ) + parser.add_argument( + "--activation-dropout", + "--relu-dropout", + type=float, + metavar="D", + help="dropout probability after activation in FFN.", + ) + parser.add_argument( + "--encoder-embed-dim", + type=int, + metavar="N", + help="encoder embedding dimension", + ) + parser.add_argument( + "--encoder-ffn-embed-dim", + type=int, + metavar="N", + help="encoder embedding dimension for FFN", + ) + parser.add_argument( + "--encoder-layers", type=int, metavar="N", help="num encoder layers" + ) + parser.add_argument( + "--encoder-attention-heads", + type=int, + metavar="N", + help="num encoder attention heads", + ) + parser.add_argument( + "--encoder-normalize-before", + action="store_true", + help="apply layernorm before each encoder block", + ) + parser.add_argument( + "--decoder-embed-dim", + type=int, + metavar="N", + help="decoder embedding dimension", + ) + parser.add_argument( + "--decoder-ffn-embed-dim", + type=int, + metavar="N", + help="decoder embedding dimension for FFN", + ) + parser.add_argument( + "--decoder-layers", type=int, metavar="N", help="num decoder layers" + ) + parser.add_argument( + "--decoder-attention-heads", + type=int, + metavar="N", + help="num decoder attention heads", + ) + parser.add_argument( + "--decoder-normalize-before", + action="store_true", + help="apply layernorm before each decoder block", + ) + parser.add_argument( + "--share-decoder-input-output-embed", + action="store_true", + help="share decoder input and output embeddings", + ) + parser.add_argument( + "--layernorm-embedding", + action="store_true", + help="add layernorm to embedding", + ) + parser.add_argument( + "--no-scale-embedding", + action="store_true", + help="if True, dont scale embeddings", + ) + parser.add_argument( + "--load-pretrained-encoder-from", + type=str, + metavar="STR", + help="model to take encoder weights from (for initialization)", + ) + parser.add_argument( + '--encoder-freezing-updates', + type=int, + metavar='N', + help='freeze encoder for first N updates' + ) + + @classmethod + def build_encoder(cls, args): + encoder = S2TTransformerEncoder(args) + pretraining_path = getattr(args, "load_pretrained_encoder_from", None) + if pretraining_path is not None: + if not Path(pretraining_path).exists(): + logger.warning( + f"skipped pretraining because {pretraining_path} does not exist" + ) + else: + encoder = checkpoint_utils.load_pretrained_component_from_model( + component=encoder, checkpoint=pretraining_path + ) + logger.info(f"loaded pretrained encoder from: {pretraining_path}") + return encoder + + @classmethod + def build_decoder(cls, args, task, embed_tokens): + return TransformerDecoderScriptable(args, task.target_dictionary, embed_tokens) + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + + # make sure all arguments are present in older models + base_architecture(args) + + def build_embedding(dictionary, embed_dim): + num_embeddings = len(dictionary) + padding_idx = dictionary.pad() + return Embedding(num_embeddings, embed_dim, padding_idx) + + decoder_embed_tokens = build_embedding( + task.target_dictionary, args.decoder_embed_dim + ) + encoder = cls.build_encoder(args) + decoder = cls.build_decoder(args, task, decoder_embed_tokens) + return cls(encoder, decoder) + + def get_normalized_probs( + self, + net_output: Tuple[Tensor, Optional[Dict[str, List[Optional[Tensor]]]]], + log_probs: bool, + sample: Optional[Dict[str, Tensor]] = None, + ): + # net_output['encoder_out'] is a (B, T, D) tensor + lprobs = self.get_normalized_probs_scriptable(net_output, log_probs, sample) + lprobs.batch_first = True + return lprobs + + def forward(self, src_tokens, src_lengths, prev_output_tokens): + """ + The forward method inherited from the base class has a **kwargs + argument in its input, which is not supported in torchscript. This + method overwrites the forward method definition without **kwargs. + """ + encoder_out = self.encoder(src_tokens=src_tokens, src_lengths=src_lengths) + decoder_out = self.decoder( + prev_output_tokens=prev_output_tokens, encoder_out=encoder_out + ) + return decoder_out + + +class S2TTransformerEncoder(FairseqEncoder): + """Speech-to-text Transformer encoder that consists of input subsampler and + Transformer encoder.""" + + def __init__(self, args): + super().__init__(None) + + self.encoder_freezing_updates = args.encoder_freezing_updates + self.num_updates = 0 + + self.dropout_module = FairseqDropout( + p=args.dropout, module_name=self.__class__.__name__ + ) + self.embed_scale = math.sqrt(args.encoder_embed_dim) + if args.no_scale_embedding: + self.embed_scale = 1.0 + self.padding_idx = 1 + + self.subsample = Conv1dSubsampler( + args.input_feat_per_channel * args.input_channels, + args.conv_channels, + args.encoder_embed_dim, + [int(k) for k in args.conv_kernel_sizes.split(",")], + ) + + self.embed_positions = PositionalEmbedding( + args.max_source_positions, args.encoder_embed_dim, self.padding_idx + ) + + self.transformer_layers = nn.ModuleList( + [TransformerEncoderLayer(args) for _ in range(args.encoder_layers)] + ) + if args.encoder_normalize_before: + self.layer_norm = LayerNorm(args.encoder_embed_dim) + else: + self.layer_norm = None + + def _forward(self, src_tokens, src_lengths): + x, input_lengths = self.subsample(src_tokens, src_lengths) + x = self.embed_scale * x + + encoder_padding_mask = lengths_to_padding_mask(input_lengths) + positions = self.embed_positions(encoder_padding_mask).transpose(0, 1) + x += positions + x = self.dropout_module(x) + + for layer in self.transformer_layers: + x = layer(x, encoder_padding_mask) + + if self.layer_norm is not None: + x = self.layer_norm(x) + + return { + "encoder_out": [x], # T x B x C + "encoder_padding_mask": [encoder_padding_mask] if encoder_padding_mask.any() else [], # B x T + "encoder_embedding": [], # B x T x C + "encoder_states": [], # List[T x B x C] + "src_tokens": [], + "src_lengths": [], + } + + def forward(self, src_tokens, src_lengths): + if self.num_updates < self.encoder_freezing_updates: + with torch.no_grad(): + x = self._forward(src_tokens, src_lengths) + else: + x = self._forward(src_tokens, src_lengths) + return x + + def reorder_encoder_out(self, encoder_out, new_order): + new_encoder_out = ( + [] if len(encoder_out["encoder_out"]) == 0 + else [x.index_select(1, new_order) for x in encoder_out["encoder_out"]] + ) + + new_encoder_padding_mask = ( + [] if len(encoder_out["encoder_padding_mask"]) == 0 + else [x.index_select(0, new_order) for x in encoder_out["encoder_padding_mask"]] + ) + + new_encoder_embedding = ( + [] if len(encoder_out["encoder_embedding"]) == 0 + else [x.index_select(0, new_order) for x in encoder_out["encoder_embedding"]] + ) + + encoder_states = encoder_out["encoder_states"] + if len(encoder_states) > 0: + for idx, state in enumerate(encoder_states): + encoder_states[idx] = state.index_select(1, new_order) + + return { + "encoder_out": new_encoder_out, # T x B x C + "encoder_padding_mask": new_encoder_padding_mask, # B x T + "encoder_embedding": new_encoder_embedding, # B x T x C + "encoder_states": encoder_states, # List[T x B x C] + "src_tokens": [], # B x T + "src_lengths": [], # B x 1 + } + + def set_num_updates(self, num_updates): + super().set_num_updates(num_updates) + self.num_updates = num_updates + + +class TransformerDecoderScriptable(TransformerDecoder): + def extract_features( + self, + prev_output_tokens, + encoder_out: Optional[Dict[str, List[Tensor]]] = None, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + full_context_alignment: bool = False, + alignment_layer: Optional[int] = None, + alignment_heads: Optional[int] = None, + ): + # call scriptable method from parent class + x, _ = self.extract_features_scriptable( + prev_output_tokens, + encoder_out, + incremental_state, + full_context_alignment, + alignment_layer, + alignment_heads, + ) + return x, None + + +@register_model_architecture(model_name="s2t_transformer", arch_name="s2t_transformer") +def base_architecture(args): + args.encoder_freezing_updates = getattr(args, "encoder_freezing_updates", 0) + # Convolutional subsampler + args.conv_kernel_sizes = getattr(args, "conv_kernel_sizes", "5,5") + args.conv_channels = getattr(args, "conv_channels", 1024) + # Transformer + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 2048) + args.encoder_layers = getattr(args, "encoder_layers", 12) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 8) + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", True) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", args.encoder_embed_dim) + args.decoder_ffn_embed_dim = getattr( + args, "decoder_ffn_embed_dim", args.encoder_ffn_embed_dim + ) + args.decoder_layers = getattr(args, "decoder_layers", 6) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 8) + args.decoder_normalize_before = getattr(args, "decoder_normalize_before", True) + args.decoder_learned_pos = getattr(args, "decoder_learned_pos", False) + args.dropout = getattr(args, "dropout", 0.1) + args.attention_dropout = getattr(args, "attention_dropout", args.dropout) + args.activation_dropout = getattr(args, "activation_dropout", args.dropout) + args.activation_fn = getattr(args, "activation_fn", "relu") + args.adaptive_softmax_cutoff = getattr(args, "adaptive_softmax_cutoff", None) + args.adaptive_softmax_dropout = getattr(args, "adaptive_softmax_dropout", 0) + args.share_decoder_input_output_embed = getattr( + args, "share_decoder_input_output_embed", False + ) + args.no_token_positional_embeddings = getattr( + args, "no_token_positional_embeddings", False + ) + args.adaptive_input = getattr(args, "adaptive_input", False) + args.decoder_layerdrop = getattr(args, "decoder_layerdrop", 0.0) + args.decoder_output_dim = getattr( + args, "decoder_output_dim", args.decoder_embed_dim + ) + args.decoder_input_dim = getattr(args, "decoder_input_dim", args.decoder_embed_dim) + args.no_scale_embedding = getattr(args, "no_scale_embedding", False) + args.quant_noise_pq = getattr(args, "quant_noise_pq", 0) + + +@register_model_architecture("s2t_transformer", "s2t_transformer_s") +def s2t_transformer_s(args): + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 256) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 256 * 8) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 4) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 4) + args.dropout = getattr(args, "dropout", 0.1) + base_architecture(args) + + +@register_model_architecture("s2t_transformer", "s2t_transformer_xs") +def s2t_transformer_xs(args): + args.encoder_layers = getattr(args, "encoder_layers", 6) + args.decoder_layers = getattr(args, "decoder_layers", 3) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 256 * 4) + args.dropout = getattr(args, "dropout", 0.3) + s2t_transformer_s(args) + + +@register_model_architecture("s2t_transformer", "s2t_transformer_sp") +def s2t_transformer_sp(args): + args.encoder_layers = getattr(args, "encoder_layers", 16) + s2t_transformer_s(args) + + +@register_model_architecture("s2t_transformer", "s2t_transformer_m") +def s2t_transformer_m(args): + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 512 * 4) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 8) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 8) + args.dropout = getattr(args, "dropout", 0.15) + base_architecture(args) + + +@register_model_architecture("s2t_transformer", "s2t_transformer_mp") +def s2t_transformer_mp(args): + args.encoder_layers = getattr(args, "encoder_layers", 16) + s2t_transformer_m(args) + + +@register_model_architecture("s2t_transformer", "s2t_transformer_l") +def s2t_transformer_l(args): + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 1024) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 1024 * 4) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 16) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 16) + args.dropout = getattr(args, "dropout", 0.2) + base_architecture(args) + + +@register_model_architecture("s2t_transformer", "s2t_transformer_lp") +def s2t_transformer_lp(args): + args.encoder_layers = getattr(args, "encoder_layers", 16) + s2t_transformer_l(args) diff --git a/fairseq/models/speech_to_text/utils.py b/fairseq/models/speech_to_text/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..573f8537c9b5940f3eff1fef5e732c6ae7e7fdc0 --- /dev/null +++ b/fairseq/models/speech_to_text/utils.py @@ -0,0 +1,564 @@ +#!/usr/bin/env python3 +# Copyright (c) 2017-present, Facebook, Inc. +# All rights reserved. +# +# This source code is licensed under the license found in the LICENSE file in +# the root directory of this source tree. An additional grant of patent rights +# can be found in the PATENTS file in the same directory. + + +import logging +from collections.abc import Iterable +from itertools import repeat +from typing import List, Optional, Tuple + +import torch +from torch import Tensor + + +# ------------------------------------------------------------------------------ +# assert_equal() +# ------------------------------------------------------------------------------ + + +def assert_equal(value1, value2, name1=None, name2=None): + """Asserts two values are equal otherwise raise an error.""" + + str_name1 = "" if name1 is None else "{} ".format(name1) + str_name2 = "" if name2 is None else "{} ".format(name2) + if value1 != value2: + str_value1 = "{}" if name1 is None else "({})" + str_value1 = str_value1.format(value1) + str_value2 = "{}" if name2 is None else "({})" + str_value2 = str_value2.format(value2) + raise ValueError( + "Expected {}{} == {}{}".format(str_name1, str_value1, str_name2, str_value2) + ) + + +def fill_config(config, key, value): + if value is not None: + if key not in config or config[key] is None: + config[key] = value + assert_equal(value, config[key], "value", f'config["{key}"]') + + +# ------------------------------------------------------------------------------ +# check_and_return_expected() +# ------------------------------------------------------------------------------ + + +def check_and_return_expected(value, undefined_value, expected_value, name=None): + """ + Return the expected value while checking if the given value is undefined or + equal to the expected value. + """ + if (undefined_value is None and value is None) or (undefined_value == value): + return expected_value + if value != expected_value: + str_name = "" if name is None else "{} ".format(name) + str_value = "{}" if name is None else "({})" + str_value = str_value.format(value) + raise ValueError( + "Expected {}{} == {}".format(str_name, str_value, expected_value) + ) + return expected_value + + +# ------------------------------------------------------------------------------ +# get_time_axis() +# ------------------------------------------------------------------------------ + + +def get_time_axis(layout): + """ + Extract the time axis from the layout, for example for breaking sequence into + segments. + """ + if layout in ["TB", "TBD"]: + return 0 + if layout in ["BT", "BTD"]: + return 1 + if layout in ["BCTD"]: + return 2 + raise ValueError("Unsupported layout = {}".format(layout)) + + +# ------------------------------------------------------------------------------ +# get_batch_axis() +# ------------------------------------------------------------------------------ + + +def get_batch_axis(layout): + """ + Extract the batch axis from the layout + """ + if layout in ["TB", "TBD"]: + return 1 + if layout in ["BT", "BTD", "BCTD"]: + return 0 + raise ValueError("Unsupported layout = {}".format(layout)) + + +# ------------------------------------------------------------------------------ +# monotonically_increasing_and_bounded() +# ------------------------------------------------------------------------------ + + +def monotonically_increasing_and_bounded(iterable, min=None, max=None): + """ + Check if the elements in the given iterable are monotonically increasing and + bounded by upper/lower bounds. + """ + if not isinstance(iterable, Iterable): + raise TypeError( + "Expected iterable to be of type Iterable, got ({})".format( + iterable.__class__.__name__ + ) + ) + for i in range(len(iterable)): + if min is not None and iterable[i] < min: + return False + if max is not None and iterable[i] > max: + return False + if i > 0 and iterable[i] <= iterable[i - 1]: + return False + return True + + +# ------------------------------------------------------------------------------ +# to_pair() +# ------------------------------------------------------------------------------ + + +def to_pair(value, name): + """Make a pair (of type tuple) of given value.""" + if isinstance(value, Iterable): + if len(value) != 2: + raise ValueError( + "Expected `{}` to have exactly 2 elements, got: ({})".format( + name, value + ) + ) + return value + return tuple(repeat(value, 2)) + + +# ------------------------------------------------------------------------------ +# infer_conv_output_attrs() +# ------------------------------------------------------------------------------ + + +# TODO(cfyeh): figure out if we can get `output_dim` without calling the module. +def infer_conv_output_attrs( + module, input_channels, input_dim, batch_size=1, max_length=8 +): + """Get output attributes of a module with input.""" + input = torch.randn(batch_size, input_channels, max_length, input_dim) + output = module(input) + output_channels = output.shape[1] + output_dim = output.shape[-1] + return output_channels, output_dim + + +# ------------------------------------------------------------------------------ +# NoOp +# ------------------------------------------------------------------------------ + + +class NoOp(torch.nn.Module): + """ + NoOp simply passes the input as the output. + """ + + def __init__(self): + super().__init__() + + def forward(self, input: Tensor) -> Tensor: + return input + + +# ------------------------------------------------------------------------------ +# Permute: a torch.nn.Module applies permutation on the input tensor. +# ------------------------------------------------------------------------------ + + +class Permute(torch.nn.Module): + def __init__(self, dims): + super().__init__() + self.dims = dims + + def forward(self, input: Tensor) -> Tensor: + return input.permute(self.dims).contiguous() + + +# ------------------------------------------------------------------------------ +# lengths_to_padding_mask() +# ------------------------------------------------------------------------------ + + +def lengths_to_padding_mask(lengths: Tensor) -> Tensor: + """Convert lengths of shape (B, ) to padding mask.""" + batch_size = lengths.shape[0] + max_length = int(torch.max(lengths).item()) + padding_mask = torch.arange( # [0, ..., T-1] + max_length, device=lengths.device, dtype=lengths.dtype + ).expand(batch_size, max_length) >= lengths.unsqueeze(1) + + return padding_mask + + +# ------------------------------------------------------------------------------ +# lengths_to_attention_mask() +# ------------------------------------------------------------------------------ + + +def lengths_to_attention_mask( + lengths: Tensor, + left_context: Optional[int] = None, + right_context: Optional[int] = None, +) -> Optional[Tensor]: + """ + Generate attention mask based on (lengths, left_context, right_context). + left_context is None means unlimited left context. + right_context is None means unlimited right context. + """ + + if left_context is None and right_context is None: + return None + + max_length = int(torch.max(lengths).item()) + + # For example, with `max_length` == 5, + # indices = tensor([ + # [ 0, 1, 2, 3, 4, 5], + # [-1, 0, 1, 2, 3, 4], + # [-2, -1, 0, 1, 2, 3], + # [-3, -2, -1, 0, 1, 2], + # [-4, -3, -2, -1, 0, 1], + # [-5, -4, -3, -2, -1, 0], + # ]) + + # In some cases the second torch.arange is created on cpu which causes a + # failure. Adding the device option to guard against it. + indices = torch.arange( + max_length, device=lengths.device, dtype=lengths.dtype + ).expand(max_length, max_length) - torch.arange( + max_length, device=lengths.device + ).view( + max_length, -1 + ) + + # For example, with `max_length` == 5, + # bool_mask = tensor([ + # [True, True, True, True, True], + # [True, True, True, True, True], + # [True, True, True, True, True], + # [True, True, True, True, True], + # [True, True, True, True, True], + # ]) + bool_mask = ( + torch.tensor([True]).to(device=lengths.device).expand(max_length, max_length) + ) + + # For example, with `max_length` == 5, left_context == 2 + # left_mask = tensor([ + # [ True, True, True, True, True], + # [ True, True, True, True, True], + # [ True, True, True, True, True], + # [False, True, True, True, True], + # [False, False, True, True, True], + # ]) + if left_context is not None: + left_mask = indices >= -left_context + bool_mask = bool_mask & left_mask + + # For example, with `max_length` == 5, right_context == 1 + # right_mask = tensor([ + # [True, True, False, False, False], + # [True, True, True, False, False], + # [True, True, True, True, False], + # [True, True, True, True, True], + # [True, True, True, True, True], + # ]) + if right_context is not None: + right_mask = indices <= right_context + bool_mask = bool_mask & right_mask + + bool_mask = (~bool_mask).to(device=lengths.device) + return bool_mask + + +# ------------------------------------------------------------------------------ +# infer_output_norm() +# ------------------------------------------------------------------------------ + + +def infer_output_norm(module, output_norm=None): + """ + Infer the output norm (string and module) needed on the module gvien desired + output normalization. + """ + if output_norm == module.output_norm(): + # output_norm already matches module.output_norm(). + return (None, NoOp()) + + if output_norm is None and module.output_norm() is not None: + logger = logging.getLogger("infer_output_norm()") + logger.warning( + "trying to set output_norm ({}) ".format(output_norm) + + "but got module.output_norm() ({}), ".format(module.output_norm()) + + "the combined output_norm() will be ({})".format(module.output_norm()) + ) + return (None, NoOp()) + + if output_norm == "log_softmax": + if module.output_norm() is not None: + raise ValueError( + "incompatible output_norm ({}) ".format(output_norm) + + "and module.output_norm() ({})".format(module.output_norm()) + ) + else: + return ("log_softmax", torch.nn.LogSoftmax(dim=-1)) + + if output_norm == "softmax": + if module.output_norm() is not None: + raise ValueError( + "incompatible output_norm ({}) ".format(output_norm) + + "and module.output_norm() ({})".format(module.output_norm()) + ) + else: + return ("softmax", torch.nn.Softmax(dim=-1)) + + raise ValueError( + "output_norm ({}) not in ".format(output_norm) + + "supported list = [None, softmax, log_softmax]" + ) + + +# ------------------------------------------------------------------------------ +# infer_channels_from_layout() +# ------------------------------------------------------------------------------ + + +def infer_channels_from_layout(layout, channels): + """Extract the number of channels from the layout.""" + if layout in ("TBD", "BTD"): + if channels is not None and channels != 1: + raise ValueError( + "Expected channels ({}) to be 1 for layout = {}".format( + channels, layout + ) + ) + if channels is None: + return 1 + return channels + + +# ------------------------------------------------------------------------------ +# pad_sequence() +# ------------------------------------------------------------------------------ + + +@torch.jit.export +def pad_sequence( + sequence: Tensor, + time_axis: int, + extra_left_context: int = 0, + extra_right_context: int = 0, +) -> Tensor: + """Pad extra left/right contexts to the sequence.""" + + if extra_left_context == 0 and extra_right_context == 0: + return sequence + + tensors_to_concat = [] + + if extra_left_context: + size = (extra_left_context,) + fill_value = 0 + indices = torch.full( + size=size, + fill_value=fill_value, + dtype=torch.long, + device=sequence.device, + ) + left_padding = torch.index_select(sequence, time_axis, indices) + tensors_to_concat.append(left_padding) + + tensors_to_concat.append(sequence) + + # NOTE(cfyeh): for efficiency reason we pad 0 instead of the last frame for + # extra right contexts. + if extra_right_context: + size = list(sequence.shape) + size[time_axis] = extra_right_context + right_padding = torch.zeros(size, dtype=sequence.dtype, device=sequence.device) + tensors_to_concat.append(right_padding) + + padded_sequence = torch.cat(tensors_to_concat, dim=time_axis) + return padded_sequence + + +# ------------------------------------------------------------------------------ +# sequence_to_segments() +# ------------------------------------------------------------------------------ + + +@torch.jit.export +def sequence_to_segments( + sequence: Tensor, + time_axis: int, + lengths: Tensor, + segment_size: Optional[int] = None, + extra_left_context: int = 0, + extra_right_context: int = 0, +) -> List[Tuple[Tensor, Tensor]]: + """Breaks sequence into segments.""" + + sequence = pad_sequence( + sequence=sequence, + time_axis=time_axis, + extra_left_context=extra_left_context, + extra_right_context=extra_right_context, + ) + + lengths = lengths + extra_left_context + extra_right_context + + segments: List[Tuple[Tensor, Tensor]] = [] + + if segment_size is None: + segments.append((sequence, lengths)) + return segments + + offset = 0 + end = sequence.shape[time_axis] + step = segment_size + size = extra_left_context + segment_size + extra_right_context + + while offset + extra_left_context + extra_right_context < end: + clamped_size = min(size, end - offset) + segment_lengths = torch.clamp(lengths - offset, min=0, max=clamped_size) + indices = torch.arange( + start=offset, + end=(offset + clamped_size), + step=1, + dtype=torch.long, + device=sequence.device, + ) + segment_tensor = torch.index_select(sequence, time_axis, indices) + segments.append((segment_tensor, segment_lengths)) + offset = offset + step + + return segments + + +# ------------------------------------------------------------------------------ +# segments_to_sequence() +# ------------------------------------------------------------------------------ + + +@torch.jit.export +def segments_to_sequence( + segments: List[Tuple[Tensor, Tensor]], time_axis: int +) -> Tuple[Tensor, Tensor]: + """Concatenate segments into a full sequence.""" + if len(segments) == 1: + return segments[0] + + tensors_to_concat: List[Tensor] = [] + lengths_to_stack: List[Tensor] = [] + + for tensor, lengths in segments: + tensors_to_concat.append(tensor) + lengths_to_stack.append(lengths) + + sequence = torch.cat(tensors_to_concat, dim=time_axis) + lengths = torch.stack(lengths_to_stack, dim=0) + lengths = torch.sum(lengths, dim=0) + + return sequence, lengths + + +def lengths_to_encoder_padding_mask(lengths, batch_first: bool = False): + """ + convert lengths (a 1-D Long/Int tensor) to 2-D binary tensor + + Args: + lengths: a (B, )-shaped tensor + batch_first: whether to return a (B, T) tensor + + Return: + max_length: maximum length of B sequences + encoder_padding_mask: a (max_length, B) binary mask, where + [t, b] = False for t < lengths[b] and True otherwise + + TODO: + kernelize this function if benchmarking shows this function is slow + """ + max_lengths = torch.max(lengths).item() + bsz = lengths.size(0) + encoder_padding_mask = torch.arange( + max_lengths + ).to( # a (T, ) tensor with [0, ..., T-1] + lengths.device + ).view( # move to the right device + 1, max_lengths + ).expand( # reshape to (1, T)-shaped tensor + bsz, -1 + ) > lengths.view( # expand to (B, T)-shaped tensor + bsz, 1 + ).expand( + -1, max_lengths + ) + if not batch_first: + return encoder_padding_mask.t(), max_lengths + else: + return encoder_padding_mask, max_lengths + + +# ------------------------------------------------------------------------------ +# attention suppression +# ------------------------------------------------------------------------------ + + +def attention_suppression(attention_weights: Tensor, scale: float): + # B, H, qlen, klen -> B, H, qlen, 1 + attention_prob = torch.nn.functional.softmax(attention_weights.float(), dim=-1) + attention_nozeros = attention_prob.to(torch.bool) + nozeros_sum = torch.sum(attention_nozeros.to(torch.float), dim=-1, keepdim=True) + + # For very sparse situation, we need get round about 0s + key_sum = torch.sum(attention_prob, dim=-1, keepdim=True) + + # nozeros_sum should > 1 + key_mean = key_sum / (nozeros_sum + 1e-8) + + # std calculation + dis = (attention_prob - key_mean) * (attention_prob - key_mean) + + # if attention_prob[i] < threshold, then dis_masked[i] = 0; for all i + dis_masked = torch.where( + attention_nozeros, dis, attention_prob.new_zeros(attention_prob.size()) + ) + + key_var = torch.sum(dis_masked, dim=-1, keepdim=True) + key_var = key_var / (nozeros_sum - 1.0 + 1e-8) + key_std = torch.sqrt(key_var) + key_thread = key_mean - scale * key_std + + # if attention_prob[i] >= key_thread, then attention_prob[i] + # , otherwise "-inf" + inf_tensor = attention_prob.new_zeros(attention_prob.size()).detach() + inf_tensor[:] = float("-inf") + attention_weights_float = torch.where( + attention_prob < key_thread, + inf_tensor, + attention_weights.float(), + ) + + return attention_weights_float.type_as(attention_weights) + + +def layer_norm_backward_hook(module, grad_input, grad_output, clamp_value): + return tuple(torch.clamp(v, min=-clamp_value, max=clamp_value) for v in grad_input) diff --git a/fairseq/models/transformer.py b/fairseq/models/transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..f4f6bea27bb4c021aaea33e86f5e481edbb3facc --- /dev/null +++ b/fairseq/models/transformer.py @@ -0,0 +1,1187 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +from typing import Any, Dict, List, Optional, Tuple + +import torch +import torch.nn as nn +from fairseq import utils +from fairseq.distributed import fsdp_wrap +from fairseq.models import ( + FairseqEncoder, + FairseqEncoderDecoderModel, + FairseqIncrementalDecoder, + register_model, + register_model_architecture, +) +from fairseq.modules import ( + AdaptiveSoftmax, + BaseLayer, + FairseqDropout, + LayerDropModuleList, + LayerNorm, + PositionalEmbedding, + SinusoidalPositionalEmbedding, + TransformerDecoderLayer, + TransformerEncoderLayer, +) +from fairseq.modules.checkpoint_activations import checkpoint_wrapper +from fairseq.modules.quant_noise import quant_noise as apply_quant_noise_ +from torch import Tensor + + +DEFAULT_MAX_SOURCE_POSITIONS = 1024 +DEFAULT_MAX_TARGET_POSITIONS = 1024 + + +DEFAULT_MIN_PARAMS_TO_WRAP = int(1e8) + + +@register_model("transformer") +class TransformerModel(FairseqEncoderDecoderModel): + """ + Transformer model from `"Attention Is All You Need" (Vaswani, et al, 2017) + <https://arxiv.org/abs/1706.03762>`_. + + Args: + encoder (TransformerEncoder): the encoder + decoder (TransformerDecoder): the decoder + + The Transformer model provides the following named architectures and + command-line arguments: + + .. argparse:: + :ref: fairseq.models.transformer_parser + :prog: + """ + + @classmethod + def hub_models(cls): + # fmt: off + + def moses_subword(path): + return { + 'path': path, + 'tokenizer': 'moses', + 'bpe': 'subword_nmt', + } + + def moses_fastbpe(path): + return { + 'path': path, + 'tokenizer': 'moses', + 'bpe': 'fastbpe', + } + + def spm(path): + return { + 'path': path, + 'bpe': 'sentencepiece', + 'tokenizer': 'space', + } + + return { + 'transformer.wmt14.en-fr': moses_subword('https://dl.fbaipublicfiles.com/fairseq/models/wmt14.en-fr.joined-dict.transformer.tar.bz2'), + 'transformer.wmt16.en-de': 'https://dl.fbaipublicfiles.com/fairseq/models/wmt16.en-de.joined-dict.transformer.tar.bz2', + 'transformer.wmt18.en-de': moses_subword('https://dl.fbaipublicfiles.com/fairseq/models/wmt18.en-de.ensemble.tar.gz'), + 'transformer.wmt19.en-de': moses_fastbpe('https://dl.fbaipublicfiles.com/fairseq/models/wmt19.en-de.joined-dict.ensemble.tar.gz'), + 'transformer.wmt19.en-ru': moses_fastbpe('https://dl.fbaipublicfiles.com/fairseq/models/wmt19.en-ru.ensemble.tar.gz'), + 'transformer.wmt19.de-en': moses_fastbpe('https://dl.fbaipublicfiles.com/fairseq/models/wmt19.de-en.joined-dict.ensemble.tar.gz'), + 'transformer.wmt19.ru-en': moses_fastbpe('https://dl.fbaipublicfiles.com/fairseq/models/wmt19.ru-en.ensemble.tar.gz'), + 'transformer.wmt19.en-de.single_model': moses_fastbpe('https://dl.fbaipublicfiles.com/fairseq/models/wmt19.en-de.joined-dict.single_model.tar.gz'), + 'transformer.wmt19.en-ru.single_model': moses_fastbpe('https://dl.fbaipublicfiles.com/fairseq/models/wmt19.en-ru.single_model.tar.gz'), + 'transformer.wmt19.de-en.single_model': moses_fastbpe('https://dl.fbaipublicfiles.com/fairseq/models/wmt19.de-en.joined-dict.single_model.tar.gz'), + 'transformer.wmt19.ru-en.single_model': moses_fastbpe('https://dl.fbaipublicfiles.com/fairseq/models/wmt19.ru-en.single_model.tar.gz'), + 'transformer.wmt20.en-ta': spm('https://dl.fbaipublicfiles.com/fairseq/models/wmt20.en-ta.single.tar.gz'), + 'transformer.wmt20.en-iu.news': spm('https://dl.fbaipublicfiles.com/fairseq/models/wmt20.en-iu.news.single.tar.gz'), + 'transformer.wmt20.en-iu.nh': spm('https://dl.fbaipublicfiles.com/fairseq/models/wmt20.en-iu.nh.single.tar.gz'), + 'transformer.wmt20.ta-en': spm('https://dl.fbaipublicfiles.com/fairseq/models/wmt20.ta-en.single.tar.gz'), + 'transformer.wmt20.iu-en.news': spm('https://dl.fbaipublicfiles.com/fairseq/models/wmt20.iu-en.news.single.tar.gz'), + 'transformer.wmt20.iu-en.nh': spm('https://dl.fbaipublicfiles.com/fairseq/models/wmt20.iu-en.nh.single.tar.gz'), + 'transformer.flores101.mm100.615M': spm('https://dl.fbaipublicfiles.com/flores101/pretrained_models/flores101_mm100_615M.tar.gz'), + 'transformer.flores101.mm100.175M': spm('https://dl.fbaipublicfiles.com/flores101/pretrained_models/flores101_mm100_175M.tar.gz'), + } + # fmt: on + + def __init__(self, args, encoder, decoder): + super().__init__(encoder, decoder) + self.args = args + self.supports_align_args = True + + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + # fmt: off + parser.add_argument('--activation-fn', + choices=utils.get_available_activation_fns(), + help='activation function to use') + parser.add_argument('--dropout', type=float, metavar='D', + help='dropout probability') + parser.add_argument('--attention-dropout', type=float, metavar='D', + help='dropout probability for attention weights') + parser.add_argument('--activation-dropout', '--relu-dropout', type=float, metavar='D', + help='dropout probability after activation in FFN.') + parser.add_argument('--encoder-embed-path', type=str, metavar='STR', + help='path to pre-trained encoder embedding') + parser.add_argument('--encoder-embed-dim', type=int, metavar='N', + help='encoder embedding dimension') + parser.add_argument('--encoder-ffn-embed-dim', type=int, metavar='N', + help='encoder embedding dimension for FFN') + parser.add_argument('--encoder-layers', type=int, metavar='N', + help='num encoder layers') + parser.add_argument('--encoder-attention-heads', type=int, metavar='N', + help='num encoder attention heads') + parser.add_argument('--encoder-normalize-before', action='store_true', + help='apply layernorm before each encoder block') + parser.add_argument('--encoder-learned-pos', action='store_true', + help='use learned positional embeddings in the encoder') + parser.add_argument('--decoder-embed-path', type=str, metavar='STR', + help='path to pre-trained decoder embedding') + parser.add_argument('--decoder-embed-dim', type=int, metavar='N', + help='decoder embedding dimension') + parser.add_argument('--decoder-ffn-embed-dim', type=int, metavar='N', + help='decoder embedding dimension for FFN') + parser.add_argument('--decoder-layers', type=int, metavar='N', + help='num decoder layers') + parser.add_argument('--decoder-attention-heads', type=int, metavar='N', + help='num decoder attention heads') + parser.add_argument('--decoder-learned-pos', action='store_true', + help='use learned positional embeddings in the decoder') + parser.add_argument('--decoder-normalize-before', action='store_true', + help='apply layernorm before each decoder block') + parser.add_argument('--decoder-output-dim', type=int, metavar='N', + help='decoder output dimension (extra linear layer ' + 'if different from decoder embed dim') + parser.add_argument('--share-decoder-input-output-embed', action='store_true', + help='share decoder input and output embeddings') + parser.add_argument('--share-all-embeddings', action='store_true', + help='share encoder, decoder and output embeddings' + ' (requires shared dictionary and embed dim)') + parser.add_argument('--no-token-positional-embeddings', default=False, action='store_true', + help='if set, disables positional embeddings (outside self attention)') + parser.add_argument('--adaptive-softmax-cutoff', metavar='EXPR', + help='comma separated list of adaptive softmax cutoff points. ' + 'Must be used with adaptive_loss criterion'), + parser.add_argument('--adaptive-softmax-dropout', type=float, metavar='D', + help='sets adaptive softmax dropout for the tail projections') + parser.add_argument('--layernorm-embedding', action='store_true', + help='add layernorm to embedding') + parser.add_argument('--no-scale-embedding', action='store_true', + help='if True, dont scale embeddings') + parser.add_argument('--checkpoint-activations', action='store_true', + help='checkpoint activations at each layer, which saves GPU ' + 'memory usage at the cost of some additional compute') + parser.add_argument('--offload-activations', action='store_true', + help='checkpoint activations at each layer, then save to gpu. Sets --checkpoint-activations.') + # args for "Cross+Self-Attention for Transformer Models" (Peitz et al., 2019) + parser.add_argument('--no-cross-attention', default=False, action='store_true', + help='do not perform cross-attention') + parser.add_argument('--cross-self-attention', default=False, action='store_true', + help='perform cross+self-attention') + # args for "Reducing Transformer Depth on Demand with Structured Dropout" (Fan et al., 2019) + parser.add_argument('--encoder-layerdrop', type=float, metavar='D', default=0, + help='LayerDrop probability for encoder') + parser.add_argument('--decoder-layerdrop', type=float, metavar='D', default=0, + help='LayerDrop probability for decoder') + parser.add_argument('--encoder-layers-to-keep', default=None, + help='which layers to *keep* when pruning as a comma-separated list') + parser.add_argument('--decoder-layers-to-keep', default=None, + help='which layers to *keep* when pruning as a comma-separated list') + # args for Training with Quantization Noise for Extreme Model Compression ({Fan*, Stock*} et al., 2020) + parser.add_argument('--quant-noise-pq', type=float, metavar='D', default=0, + help='iterative PQ quantization noise at training time') + parser.add_argument('--quant-noise-pq-block-size', type=int, metavar='D', default=8, + help='block size of quantization noise at training time') + parser.add_argument('--quant-noise-scalar', type=float, metavar='D', default=0, + help='scalar quantization noise and scalar quantization at training time') + # args for Fully Sharded Data Parallel (FSDP) training + parser.add_argument( + '--min-params-to-wrap', type=int, metavar='D', default=DEFAULT_MIN_PARAMS_TO_WRAP, + help=( + 'minimum number of params for a layer to be wrapped with FSDP() when ' + 'training with --ddp-backend=fully_sharded. Smaller values will ' + 'improve memory efficiency, but may make torch.distributed ' + 'communication less efficient due to smaller input sizes. This option ' + 'is set to 0 (i.e., always wrap) when --checkpoint-activations or ' + '--offload-activations are passed.' + ) + ) + # fmt: on + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + + # make sure all arguments are present in older models + base_architecture(args) + + if args.encoder_layers_to_keep: + args.encoder_layers = len(args.encoder_layers_to_keep.split(",")) + if args.decoder_layers_to_keep: + args.decoder_layers = len(args.decoder_layers_to_keep.split(",")) + + if getattr(args, "max_source_positions", None) is None: + args.max_source_positions = DEFAULT_MAX_SOURCE_POSITIONS + if getattr(args, "max_target_positions", None) is None: + args.max_target_positions = DEFAULT_MAX_TARGET_POSITIONS + + src_dict, tgt_dict = task.source_dictionary, task.target_dictionary + + if args.share_all_embeddings: + if src_dict != tgt_dict: + raise ValueError("--share-all-embeddings requires a joined dictionary") + if args.encoder_embed_dim != args.decoder_embed_dim: + raise ValueError( + "--share-all-embeddings requires --encoder-embed-dim to match --decoder-embed-dim" + ) + if args.decoder_embed_path and ( + args.decoder_embed_path != args.encoder_embed_path + ): + raise ValueError( + "--share-all-embeddings not compatible with --decoder-embed-path" + ) + encoder_embed_tokens = cls.build_embedding( + args, src_dict, args.encoder_embed_dim, args.encoder_embed_path + ) + decoder_embed_tokens = encoder_embed_tokens + args.share_decoder_input_output_embed = True + else: + encoder_embed_tokens = cls.build_embedding( + args, src_dict, args.encoder_embed_dim, args.encoder_embed_path + ) + decoder_embed_tokens = cls.build_embedding( + args, tgt_dict, args.decoder_embed_dim, args.decoder_embed_path + ) + if getattr(args, "offload_activations", False): + args.checkpoint_activations = True # offloading implies checkpointing + encoder = cls.build_encoder(args, src_dict, encoder_embed_tokens) + decoder = cls.build_decoder(args, tgt_dict, decoder_embed_tokens) + if not args.share_all_embeddings: + min_params_to_wrap = getattr( + args, "min_params_to_wrap", DEFAULT_MIN_PARAMS_TO_WRAP + ) + # fsdp_wrap is a no-op when --ddp-backend != fully_sharded + encoder = fsdp_wrap(encoder, min_num_params=min_params_to_wrap) + decoder = fsdp_wrap(decoder, min_num_params=min_params_to_wrap) + return cls(args, encoder, decoder) + + @classmethod + def build_embedding(cls, args, dictionary, embed_dim, path=None): + num_embeddings = len(dictionary) + padding_idx = dictionary.pad() + + emb = Embedding(num_embeddings, embed_dim, padding_idx) + # if provided, load from preloaded dictionaries + if path: + embed_dict = utils.parse_embedding(path) + utils.load_embedding(embed_dict, dictionary, emb) + return emb + + @classmethod + def build_encoder(cls, args, src_dict, embed_tokens): + return TransformerEncoder(args, src_dict, embed_tokens) + + @classmethod + def build_decoder(cls, args, tgt_dict, embed_tokens): + return TransformerDecoder( + args, + tgt_dict, + embed_tokens, + no_encoder_attn=getattr(args, "no_cross_attention", False), + ) + + # TorchScript doesn't support optional arguments with variable length (**kwargs). + # Current workaround is to add union of all arguments in child classes. + def forward( + self, + src_tokens, + src_lengths, + prev_output_tokens, + return_all_hiddens: bool = True, + features_only: bool = False, + alignment_layer: Optional[int] = None, + alignment_heads: Optional[int] = None, + ): + """ + Run the forward pass for an encoder-decoder model. + + Copied from the base class, but without ``**kwargs``, + which are not supported by TorchScript. + """ + encoder_out = self.encoder( + src_tokens, src_lengths=src_lengths, return_all_hiddens=return_all_hiddens + ) + decoder_out = self.decoder( + prev_output_tokens, + encoder_out=encoder_out, + features_only=features_only, + alignment_layer=alignment_layer, + alignment_heads=alignment_heads, + src_lengths=src_lengths, + return_all_hiddens=return_all_hiddens, + ) + return decoder_out + + # Since get_normalized_probs is in the Fairseq Model which is not scriptable, + # I rewrite the get_normalized_probs from Base Class to call the + # helper function in the Base Class. + @torch.jit.export + def get_normalized_probs( + self, + net_output: Tuple[Tensor, Optional[Dict[str, List[Optional[Tensor]]]]], + log_probs: bool, + sample: Optional[Dict[str, Tensor]] = None, + ): + """Get normalized probabilities (or log probs) from a net's output.""" + return self.get_normalized_probs_scriptable(net_output, log_probs, sample) + + +class TransformerEncoder(FairseqEncoder): + """ + Transformer encoder consisting of *args.encoder_layers* layers. Each layer + is a :class:`TransformerEncoderLayer`. + + Args: + args (argparse.Namespace): parsed command-line arguments + dictionary (~fairseq.data.Dictionary): encoding dictionary + embed_tokens (torch.nn.Embedding): input embedding + """ + + def __init__(self, args, dictionary, embed_tokens): + self.args = args + super().__init__(dictionary) + self.register_buffer("version", torch.Tensor([3])) + + self.dropout_module = FairseqDropout( + args.dropout, module_name=self.__class__.__name__ + ) + self.encoder_layerdrop = args.encoder_layerdrop + + embed_dim = embed_tokens.embedding_dim + self.padding_idx = embed_tokens.padding_idx + self.max_source_positions = args.max_source_positions + + self.embed_tokens = embed_tokens + + self.embed_scale = 1.0 if args.no_scale_embedding else math.sqrt(embed_dim) + + self.embed_positions = ( + PositionalEmbedding( + args.max_source_positions, + embed_dim, + self.padding_idx, + learned=args.encoder_learned_pos, + ) + if not args.no_token_positional_embeddings + else None + ) + export = getattr(args, "export", False) + if getattr(args, "layernorm_embedding", False): + self.layernorm_embedding = LayerNorm(embed_dim, export=export) + else: + self.layernorm_embedding = None + + if not args.adaptive_input and args.quant_noise_pq > 0: + self.quant_noise = apply_quant_noise_( + nn.Linear(embed_dim, embed_dim, bias=False), + args.quant_noise_pq, + args.quant_noise_pq_block_size, + ) + else: + self.quant_noise = None + + if self.encoder_layerdrop > 0.0: + self.layers = LayerDropModuleList(p=self.encoder_layerdrop) + else: + self.layers = nn.ModuleList([]) + self.layers.extend( + [self.build_encoder_layer(args) for i in range(args.encoder_layers)] + ) + self.num_layers = len(self.layers) + + if args.encoder_normalize_before: + self.layer_norm = LayerNorm(embed_dim, export=export) + else: + self.layer_norm = None + + def build_encoder_layer(self, args): + layer = TransformerEncoderLayer(args) + checkpoint = getattr(args, "checkpoint_activations", False) + if checkpoint: + offload_to_cpu = getattr(args, "offload_activations", False) + layer = checkpoint_wrapper(layer, offload_to_cpu=offload_to_cpu) + # if we are checkpointing, enforce that FSDP always wraps the + # checkpointed layer, regardless of layer size + min_params_to_wrap = ( + getattr(args, "min_params_to_wrap", DEFAULT_MIN_PARAMS_TO_WRAP) + if not checkpoint + else 0 + ) + layer = fsdp_wrap(layer, min_num_params=min_params_to_wrap) + return layer + + def forward_embedding( + self, src_tokens, token_embedding: Optional[torch.Tensor] = None + ): + # embed tokens and positions + if token_embedding is None: + token_embedding = self.embed_tokens(src_tokens) + x = embed = self.embed_scale * token_embedding + if self.embed_positions is not None: + x = embed + self.embed_positions(src_tokens) + if self.layernorm_embedding is not None: + x = self.layernorm_embedding(x) + x = self.dropout_module(x) + if self.quant_noise is not None: + x = self.quant_noise(x) + return x, embed + + def forward( + self, + src_tokens, + src_lengths: Optional[torch.Tensor] = None, + return_all_hiddens: bool = False, + token_embeddings: Optional[torch.Tensor] = None, + ): + """ + Args: + src_tokens (LongTensor): tokens in the source language of shape + `(batch, src_len)` + src_lengths (torch.LongTensor): lengths of each source sentence of + shape `(batch)` + return_all_hiddens (bool, optional): also return all of the + intermediate hidden states (default: False). + token_embeddings (torch.Tensor, optional): precomputed embeddings + default `None` will recompute embeddings + + Returns: + dict: + - **encoder_out** (Tensor): the last encoder layer's output of + shape `(src_len, batch, embed_dim)` + - **encoder_padding_mask** (ByteTensor): the positions of + padding elements of shape `(batch, src_len)` + - **encoder_embedding** (Tensor): the (scaled) embedding lookup + of shape `(batch, src_len, embed_dim)` + - **encoder_states** (List[Tensor]): all intermediate + hidden states of shape `(src_len, batch, embed_dim)`. + Only populated if *return_all_hiddens* is True. + """ + return self.forward_scriptable( + src_tokens, src_lengths, return_all_hiddens, token_embeddings + ) + + # TorchScript doesn't support super() method so that the scriptable Subclass + # can't access the base class model in Torchscript. + # Current workaround is to add a helper function with different name and + # call the helper function from scriptable Subclass. + def forward_scriptable( + self, + src_tokens, + src_lengths: Optional[torch.Tensor] = None, + return_all_hiddens: bool = False, + token_embeddings: Optional[torch.Tensor] = None, + ): + """ + Args: + src_tokens (LongTensor): tokens in the source language of shape + `(batch, src_len)` + src_lengths (torch.LongTensor): lengths of each source sentence of + shape `(batch)` + return_all_hiddens (bool, optional): also return all of the + intermediate hidden states (default: False). + token_embeddings (torch.Tensor, optional): precomputed embeddings + default `None` will recompute embeddings + + Returns: + dict: + - **encoder_out** (Tensor): the last encoder layer's output of + shape `(src_len, batch, embed_dim)` + - **encoder_padding_mask** (ByteTensor): the positions of + padding elements of shape `(batch, src_len)` + - **encoder_embedding** (Tensor): the (scaled) embedding lookup + of shape `(batch, src_len, embed_dim)` + - **encoder_states** (List[Tensor]): all intermediate + hidden states of shape `(src_len, batch, embed_dim)`. + Only populated if *return_all_hiddens* is True. + """ + # compute padding mask + encoder_padding_mask = src_tokens.eq(self.padding_idx) + has_pads = src_tokens.device.type == "xla" or encoder_padding_mask.any() + + x, encoder_embedding = self.forward_embedding(src_tokens, token_embeddings) + + # account for padding while computing the representation + if has_pads: + x = x * (1 - encoder_padding_mask.unsqueeze(-1).type_as(x)) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + + encoder_states = [] + + if return_all_hiddens: + encoder_states.append(x) + + # encoder layers + for layer in self.layers: + x = layer( + x, encoder_padding_mask=encoder_padding_mask if has_pads else None + ) + if return_all_hiddens: + assert encoder_states is not None + encoder_states.append(x) + + if self.layer_norm is not None: + x = self.layer_norm(x) + + # The Pytorch Mobile lite interpreter does not supports returning NamedTuple in + # `forward` so we use a dictionary instead. + # TorchScript does not support mixed values so the values are all lists. + # The empty list is equivalent to None. + return { + "encoder_out": [x], # T x B x C + "encoder_padding_mask": [encoder_padding_mask], # B x T + "encoder_embedding": [encoder_embedding], # B x T x C + "encoder_states": encoder_states, # List[T x B x C] + "src_tokens": [], + "src_lengths": [], + } + + @torch.jit.export + def reorder_encoder_out(self, encoder_out: Dict[str, List[Tensor]], new_order): + """ + Reorder encoder output according to *new_order*. + + Args: + encoder_out: output from the ``forward()`` method + new_order (LongTensor): desired order + + Returns: + *encoder_out* rearranged according to *new_order* + """ + if len(encoder_out["encoder_out"]) == 0: + new_encoder_out = [] + else: + new_encoder_out = [encoder_out["encoder_out"][0].index_select(1, new_order)] + if len(encoder_out["encoder_padding_mask"]) == 0: + new_encoder_padding_mask = [] + else: + new_encoder_padding_mask = [ + encoder_out["encoder_padding_mask"][0].index_select(0, new_order) + ] + if len(encoder_out["encoder_embedding"]) == 0: + new_encoder_embedding = [] + else: + new_encoder_embedding = [ + encoder_out["encoder_embedding"][0].index_select(0, new_order) + ] + + if len(encoder_out["src_tokens"]) == 0: + src_tokens = [] + else: + src_tokens = [(encoder_out["src_tokens"][0]).index_select(0, new_order)] + + if len(encoder_out["src_lengths"]) == 0: + src_lengths = [] + else: + src_lengths = [(encoder_out["src_lengths"][0]).index_select(0, new_order)] + + encoder_states = encoder_out["encoder_states"] + if len(encoder_states) > 0: + for idx, state in enumerate(encoder_states): + encoder_states[idx] = state.index_select(1, new_order) + + return { + "encoder_out": new_encoder_out, # T x B x C + "encoder_padding_mask": new_encoder_padding_mask, # B x T + "encoder_embedding": new_encoder_embedding, # B x T x C + "encoder_states": encoder_states, # List[T x B x C] + "src_tokens": src_tokens, # B x T + "src_lengths": src_lengths, # B x 1 + } + + def max_positions(self): + """Maximum input length supported by the encoder.""" + if self.embed_positions is None: + return self.max_source_positions + return min(self.max_source_positions, self.embed_positions.max_positions) + + def upgrade_state_dict_named(self, state_dict, name): + """Upgrade a (possibly old) state dict for new versions of fairseq.""" + if isinstance(self.embed_positions, SinusoidalPositionalEmbedding): + weights_key = "{}.embed_positions.weights".format(name) + if weights_key in state_dict: + print("deleting {0}".format(weights_key)) + del state_dict[weights_key] + state_dict[ + "{}.embed_positions._float_tensor".format(name) + ] = torch.FloatTensor(1) + for i in range(self.num_layers): + # update layer norms + self.layers[i].upgrade_state_dict_named( + state_dict, "{}.layers.{}".format(name, i) + ) + + version_key = "{}.version".format(name) + if utils.item(state_dict.get(version_key, torch.Tensor([1]))[0]) < 2: + # earlier checkpoints did not normalize after the stack of layers + self.layer_norm = None + self.normalize = False + state_dict[version_key] = torch.Tensor([1]) + return state_dict + + +class TransformerDecoder(FairseqIncrementalDecoder): + """ + Transformer decoder consisting of *args.decoder_layers* layers. Each layer + is a :class:`TransformerDecoderLayer`. + + Args: + args (argparse.Namespace): parsed command-line arguments + dictionary (~fairseq.data.Dictionary): decoding dictionary + embed_tokens (torch.nn.Embedding): output embedding + no_encoder_attn (bool, optional): whether to attend to encoder outputs + (default: False). + """ + + def __init__( + self, + args, + dictionary, + embed_tokens, + no_encoder_attn=False, + output_projection=None, + ): + self.args = args + super().__init__(dictionary) + self.register_buffer("version", torch.Tensor([3])) + self._future_mask = torch.empty(0) + + self.dropout_module = FairseqDropout( + args.dropout, module_name=self.__class__.__name__ + ) + self.decoder_layerdrop = args.decoder_layerdrop + self.share_input_output_embed = args.share_decoder_input_output_embed + + input_embed_dim = embed_tokens.embedding_dim + embed_dim = args.decoder_embed_dim + self.embed_dim = embed_dim + self.output_embed_dim = args.decoder_output_dim + + self.padding_idx = embed_tokens.padding_idx + self.max_target_positions = args.max_target_positions + + self.embed_tokens = embed_tokens + + self.embed_scale = 1.0 if args.no_scale_embedding else math.sqrt(embed_dim) + + if not args.adaptive_input and args.quant_noise_pq > 0: + self.quant_noise = apply_quant_noise_( + nn.Linear(embed_dim, embed_dim, bias=False), + args.quant_noise_pq, + args.quant_noise_pq_block_size, + ) + else: + self.quant_noise = None + + self.project_in_dim = ( + Linear(input_embed_dim, embed_dim, bias=False) + if embed_dim != input_embed_dim + else None + ) + self.embed_positions = ( + PositionalEmbedding( + self.max_target_positions, + embed_dim, + self.padding_idx, + learned=args.decoder_learned_pos, + ) + if not args.no_token_positional_embeddings + else None + ) + export = getattr(args, "export", False) + if getattr(args, "layernorm_embedding", False): + self.layernorm_embedding = LayerNorm(embed_dim, export=export) + else: + self.layernorm_embedding = None + + self.cross_self_attention = getattr(args, "cross_self_attention", False) + + if self.decoder_layerdrop > 0.0: + self.layers = LayerDropModuleList(p=self.decoder_layerdrop) + else: + self.layers = nn.ModuleList([]) + self.layers.extend( + [ + self.build_decoder_layer(args, no_encoder_attn) + for _ in range(args.decoder_layers) + ] + ) + self.num_layers = len(self.layers) + + if args.decoder_normalize_before and not getattr( + args, "no_decoder_final_norm", False + ): + self.layer_norm = LayerNorm(embed_dim, export=export) + else: + self.layer_norm = None + + self.project_out_dim = ( + Linear(embed_dim, self.output_embed_dim, bias=False) + if embed_dim != self.output_embed_dim and not args.tie_adaptive_weights + else None + ) + + self.adaptive_softmax = None + self.output_projection = output_projection + if self.output_projection is None: + self.build_output_projection(args, dictionary, embed_tokens) + + def build_output_projection(self, args, dictionary, embed_tokens): + if args.adaptive_softmax_cutoff is not None: + self.adaptive_softmax = AdaptiveSoftmax( + len(dictionary), + self.output_embed_dim, + utils.eval_str_list(args.adaptive_softmax_cutoff, type=int), + dropout=args.adaptive_softmax_dropout, + adaptive_inputs=embed_tokens if args.tie_adaptive_weights else None, + factor=args.adaptive_softmax_factor, + tie_proj=args.tie_adaptive_proj, + ) + elif self.share_input_output_embed: + self.output_projection = nn.Linear( + self.embed_tokens.weight.shape[1], + self.embed_tokens.weight.shape[0], + bias=False, + ) + self.output_projection.weight = self.embed_tokens.weight + else: + self.output_projection = nn.Linear( + self.output_embed_dim, len(dictionary), bias=False + ) + nn.init.normal_( + self.output_projection.weight, mean=0, std=self.output_embed_dim ** -0.5 + ) + num_base_layers = getattr(args, "base_layers", 0) + for i in range(num_base_layers): + self.layers.insert( + ((i + 1) * args.decoder_layers) // (num_base_layers + 1), + BaseLayer(args), + ) + + def build_decoder_layer(self, args, no_encoder_attn=False): + layer = TransformerDecoderLayer(args, no_encoder_attn) + checkpoint = getattr(args, "checkpoint_activations", False) + if checkpoint: + offload_to_cpu = getattr(args, "offload_activations", False) + layer = checkpoint_wrapper(layer, offload_to_cpu=offload_to_cpu) + # if we are checkpointing, enforce that FSDP always wraps the + # checkpointed layer, regardless of layer size + min_params_to_wrap = ( + getattr(args, "min_params_to_wrap", DEFAULT_MIN_PARAMS_TO_WRAP) + if not checkpoint + else 0 + ) + layer = fsdp_wrap(layer, min_num_params=min_params_to_wrap) + return layer + + def forward( + self, + prev_output_tokens, + encoder_out: Optional[Dict[str, List[Tensor]]] = None, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + features_only: bool = False, + full_context_alignment: bool = False, + alignment_layer: Optional[int] = None, + alignment_heads: Optional[int] = None, + src_lengths: Optional[Any] = None, + return_all_hiddens: bool = False, + ): + """ + Args: + prev_output_tokens (LongTensor): previous decoder outputs of shape + `(batch, tgt_len)`, for teacher forcing + encoder_out (optional): output from the encoder, used for + encoder-side attention, should be of size T x B x C + incremental_state (dict): dictionary used for storing state during + :ref:`Incremental decoding` + features_only (bool, optional): only return features without + applying output layer (default: False). + full_context_alignment (bool, optional): don't apply + auto-regressive mask to self-attention (default: False). + + Returns: + tuple: + - the decoder's output of shape `(batch, tgt_len, vocab)` + - a dictionary with any model-specific outputs + """ + + x, extra = self.extract_features( + prev_output_tokens, + encoder_out=encoder_out, + incremental_state=incremental_state, + full_context_alignment=full_context_alignment, + alignment_layer=alignment_layer, + alignment_heads=alignment_heads, + ) + + if not features_only: + x = self.output_layer(x) + return x, extra + + def extract_features( + self, + prev_output_tokens, + encoder_out: Optional[Dict[str, List[Tensor]]], + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + full_context_alignment: bool = False, + alignment_layer: Optional[int] = None, + alignment_heads: Optional[int] = None, + ): + return self.extract_features_scriptable( + prev_output_tokens, + encoder_out, + incremental_state, + full_context_alignment, + alignment_layer, + alignment_heads, + ) + + """ + A scriptable subclass of this class has an extract_features method and calls + super().extract_features, but super() is not supported in torchscript. A copy of + this function is made to be used in the subclass instead. + """ + + def extract_features_scriptable( + self, + prev_output_tokens, + encoder_out: Optional[Dict[str, List[Tensor]]], + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + full_context_alignment: bool = False, + alignment_layer: Optional[int] = None, + alignment_heads: Optional[int] = None, + ): + """ + Similar to *forward* but only return features. + + Includes several features from "Jointly Learning to Align and + Translate with Transformer Models" (Garg et al., EMNLP 2019). + + Args: + full_context_alignment (bool, optional): don't apply + auto-regressive mask to self-attention (default: False). + alignment_layer (int, optional): return mean alignment over + heads at this layer (default: last layer). + alignment_heads (int, optional): only average alignment over + this many heads (default: all heads). + + Returns: + tuple: + - the decoder's features of shape `(batch, tgt_len, embed_dim)` + - a dictionary with any model-specific outputs + """ + bs, slen = prev_output_tokens.size() + if alignment_layer is None: + alignment_layer = self.num_layers - 1 + + enc: Optional[Tensor] = None + padding_mask: Optional[Tensor] = None + if encoder_out is not None and len(encoder_out["encoder_out"]) > 0: + enc = encoder_out["encoder_out"][0] + assert ( + enc.size()[1] == bs + ), f"Expected enc.shape == (t, {bs}, c) got {enc.shape}" + if encoder_out is not None and len(encoder_out["encoder_padding_mask"]) > 0: + padding_mask = encoder_out["encoder_padding_mask"][0] + + # embed positions + positions = None + if self.embed_positions is not None: + positions = self.embed_positions( + prev_output_tokens, incremental_state=incremental_state + ) + + if incremental_state is not None: + prev_output_tokens = prev_output_tokens[:, -1:] + if positions is not None: + positions = positions[:, -1:] + + # embed tokens and positions + x = self.embed_scale * self.embed_tokens(prev_output_tokens) + + if self.quant_noise is not None: + x = self.quant_noise(x) + + if self.project_in_dim is not None: + x = self.project_in_dim(x) + + if positions is not None: + x += positions + + if self.layernorm_embedding is not None: + x = self.layernorm_embedding(x) + + x = self.dropout_module(x) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + + self_attn_padding_mask: Optional[Tensor] = None + if self.cross_self_attention or prev_output_tokens.eq(self.padding_idx).any(): + self_attn_padding_mask = prev_output_tokens.eq(self.padding_idx) + + # decoder layers + attn: Optional[Tensor] = None + inner_states: List[Optional[Tensor]] = [x] + for idx, layer in enumerate(self.layers): + if incremental_state is None and not full_context_alignment: + self_attn_mask = self.buffered_future_mask(x) + else: + self_attn_mask = None + + x, layer_attn, _ = layer( + x, + enc, + padding_mask, + incremental_state, + self_attn_mask=self_attn_mask, + self_attn_padding_mask=self_attn_padding_mask, + need_attn=bool((idx == alignment_layer)), + need_head_weights=bool((idx == alignment_layer)), + ) + inner_states.append(x) + if layer_attn is not None and idx == alignment_layer: + attn = layer_attn.float().to(x) + + if attn is not None: + if alignment_heads is not None: + attn = attn[:alignment_heads] + + # average probabilities over heads + attn = attn.mean(dim=0) + + if self.layer_norm is not None: + x = self.layer_norm(x) + + # T x B x C -> B x T x C + x = x.transpose(0, 1) + + if self.project_out_dim is not None: + x = self.project_out_dim(x) + + return x, {"attn": [attn], "inner_states": inner_states} + + def output_layer(self, features): + """Project features to the vocabulary size.""" + if self.adaptive_softmax is None: + # project back to size of vocabulary + return self.output_projection(features) + else: + return features + + def max_positions(self): + """Maximum output length supported by the decoder.""" + if self.embed_positions is None: + return self.max_target_positions + return min(self.max_target_positions, self.embed_positions.max_positions) + + def buffered_future_mask(self, tensor): + dim = tensor.size(0) + # self._future_mask.device != tensor.device is not working in TorchScript. This is a workaround. + if ( + self._future_mask.size(0) == 0 + or (not self._future_mask.device == tensor.device) + or self._future_mask.size(0) < dim + ): + self._future_mask = torch.triu( + utils.fill_with_neg_inf(torch.zeros([dim, dim])), 1 + ) + self._future_mask = self._future_mask.to(tensor) + return self._future_mask[:dim, :dim] + + def upgrade_state_dict_named(self, state_dict, name): + """Upgrade a (possibly old) state dict for new versions of fairseq.""" + if isinstance(self.embed_positions, SinusoidalPositionalEmbedding): + weights_key = "{}.embed_positions.weights".format(name) + if weights_key in state_dict: + del state_dict[weights_key] + state_dict[ + "{}.embed_positions._float_tensor".format(name) + ] = torch.FloatTensor(1) + + if f"{name}.output_projection.weight" not in state_dict: + if self.share_input_output_embed: + embed_out_key = f"{name}.embed_tokens.weight" + else: + embed_out_key = f"{name}.embed_out" + if embed_out_key in state_dict: + state_dict[f"{name}.output_projection.weight"] = state_dict[ + embed_out_key + ] + if not self.share_input_output_embed: + del state_dict[embed_out_key] + + for i in range(self.num_layers): + # update layer norms + layer_norm_map = { + "0": "self_attn_layer_norm", + "1": "encoder_attn_layer_norm", + "2": "final_layer_norm", + } + for old, new in layer_norm_map.items(): + for m in ("weight", "bias"): + k = "{}.layers.{}.layer_norms.{}.{}".format(name, i, old, m) + if k in state_dict: + state_dict[ + "{}.layers.{}.{}.{}".format(name, i, new, m) + ] = state_dict[k] + del state_dict[k] + + version_key = "{}.version".format(name) + if utils.item(state_dict.get(version_key, torch.Tensor([1]))[0]) <= 2: + # earlier checkpoints did not normalize after the stack of layers + self.layer_norm = None + self.normalize = False + state_dict[version_key] = torch.Tensor([1]) + + return state_dict + + +def Embedding(num_embeddings, embedding_dim, padding_idx): + m = nn.Embedding(num_embeddings, embedding_dim, padding_idx=padding_idx) + nn.init.normal_(m.weight, mean=0, std=embedding_dim ** -0.5) + nn.init.constant_(m.weight[padding_idx], 0) + return m + + +def Linear(in_features, out_features, bias=True): + m = nn.Linear(in_features, out_features, bias) + nn.init.xavier_uniform_(m.weight) + if bias: + nn.init.constant_(m.bias, 0.0) + return m + + +@register_model_architecture("transformer", "transformer_tiny") +def tiny_architecture(args): + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 64) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 64) + args.encoder_layers = getattr(args, "encoder_layers", 2) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 2) + args.decoder_layers = getattr(args, "decoder_layers", 2) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 2) + return base_architecture(args) + + +@register_model_architecture("transformer", "transformer") +def base_architecture(args): + args.encoder_embed_path = getattr(args, "encoder_embed_path", None) + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 2048) + args.encoder_layers = getattr(args, "encoder_layers", 6) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 8) + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False) + args.encoder_learned_pos = getattr(args, "encoder_learned_pos", False) + args.decoder_embed_path = getattr(args, "decoder_embed_path", None) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", args.encoder_embed_dim) + args.decoder_ffn_embed_dim = getattr( + args, "decoder_ffn_embed_dim", args.encoder_ffn_embed_dim + ) + args.decoder_layers = getattr(args, "decoder_layers", 6) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 8) + args.decoder_normalize_before = getattr(args, "decoder_normalize_before", False) + args.decoder_learned_pos = getattr(args, "decoder_learned_pos", False) + args.attention_dropout = getattr(args, "attention_dropout", 0.0) + args.activation_dropout = getattr(args, "activation_dropout", 0.0) + args.activation_fn = getattr(args, "activation_fn", "relu") + args.dropout = getattr(args, "dropout", 0.1) + args.adaptive_softmax_cutoff = getattr(args, "adaptive_softmax_cutoff", None) + args.adaptive_softmax_dropout = getattr(args, "adaptive_softmax_dropout", 0) + args.share_decoder_input_output_embed = getattr( + args, "share_decoder_input_output_embed", False + ) + args.share_all_embeddings = getattr(args, "share_all_embeddings", False) + args.no_token_positional_embeddings = getattr( + args, "no_token_positional_embeddings", False + ) + args.adaptive_input = getattr(args, "adaptive_input", False) + args.no_cross_attention = getattr(args, "no_cross_attention", False) + args.cross_self_attention = getattr(args, "cross_self_attention", False) + + args.decoder_output_dim = getattr( + args, "decoder_output_dim", args.decoder_embed_dim + ) + args.decoder_input_dim = getattr(args, "decoder_input_dim", args.decoder_embed_dim) + + args.no_scale_embedding = getattr(args, "no_scale_embedding", False) + args.layernorm_embedding = getattr(args, "layernorm_embedding", False) + args.tie_adaptive_weights = getattr(args, "tie_adaptive_weights", False) + args.checkpoint_activations = getattr(args, "checkpoint_activations", False) + args.offload_activations = getattr(args, "offload_activations", False) + if args.offload_activations: + args.checkpoint_activations = True + args.encoder_layers_to_keep = getattr(args, "encoder_layers_to_keep", None) + args.decoder_layers_to_keep = getattr(args, "decoder_layers_to_keep", None) + args.encoder_layerdrop = getattr(args, "encoder_layerdrop", 0) + args.decoder_layerdrop = getattr(args, "decoder_layerdrop", 0) + args.quant_noise_pq = getattr(args, "quant_noise_pq", 0) + args.quant_noise_pq_block_size = getattr(args, "quant_noise_pq_block_size", 8) + args.quant_noise_scalar = getattr(args, "quant_noise_scalar", 0) + + +@register_model_architecture("transformer", "transformer_iwslt_de_en") +def transformer_iwslt_de_en(args): + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 1024) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 4) + args.encoder_layers = getattr(args, "encoder_layers", 6) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 512) + args.decoder_ffn_embed_dim = getattr(args, "decoder_ffn_embed_dim", 1024) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 4) + args.decoder_layers = getattr(args, "decoder_layers", 6) + base_architecture(args) + + +@register_model_architecture("transformer", "transformer_wmt_en_de") +def transformer_wmt_en_de(args): + base_architecture(args) + + +# parameters used in the "Attention Is All You Need" paper (Vaswani et al., 2017) +@register_model_architecture("transformer", "transformer_vaswani_wmt_en_de_big") +def transformer_vaswani_wmt_en_de_big(args): + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 1024) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 4096) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 16) + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 1024) + args.decoder_ffn_embed_dim = getattr(args, "decoder_ffn_embed_dim", 4096) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 16) + args.dropout = getattr(args, "dropout", 0.3) + base_architecture(args) + + +@register_model_architecture("transformer", "transformer_vaswani_wmt_en_fr_big") +def transformer_vaswani_wmt_en_fr_big(args): + args.dropout = getattr(args, "dropout", 0.1) + transformer_vaswani_wmt_en_de_big(args) + + +@register_model_architecture("transformer", "transformer_wmt_en_de_big") +def transformer_wmt_en_de_big(args): + args.attention_dropout = getattr(args, "attention_dropout", 0.1) + transformer_vaswani_wmt_en_de_big(args) + + +# default parameters used in tensor2tensor implementation +@register_model_architecture("transformer", "transformer_wmt_en_de_big_t2t") +def transformer_wmt_en_de_big_t2t(args): + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", True) + args.decoder_normalize_before = getattr(args, "decoder_normalize_before", True) + args.attention_dropout = getattr(args, "attention_dropout", 0.1) + args.activation_dropout = getattr(args, "activation_dropout", 0.1) + transformer_vaswani_wmt_en_de_big(args) diff --git a/fairseq/models/transformer_align.py b/fairseq/models/transformer_align.py new file mode 100644 index 0000000000000000000000000000000000000000..eaf585bd10e630ae6cd89920f197cd165f55ad58 --- /dev/null +++ b/fairseq/models/transformer_align.py @@ -0,0 +1,93 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq.models import register_model, register_model_architecture +from fairseq.models.transformer import ( + TransformerModel, + base_architecture, + transformer_wmt_en_de_big, +) + + +@register_model("transformer_align") +class TransformerAlignModel(TransformerModel): + """ + See "Jointly Learning to Align and Translate with Transformer + Models" (Garg et al., EMNLP 2019). + """ + + def __init__(self, encoder, decoder, args): + super().__init__(args, encoder, decoder) + self.alignment_heads = args.alignment_heads + self.alignment_layer = args.alignment_layer + self.full_context_alignment = args.full_context_alignment + + @staticmethod + def add_args(parser): + # fmt: off + super(TransformerAlignModel, TransformerAlignModel).add_args(parser) + parser.add_argument('--alignment-heads', type=int, metavar='D', + help='Number of cross attention heads per layer to supervised with alignments') + parser.add_argument('--alignment-layer', type=int, metavar='D', + help='Layer number which has to be supervised. 0 corresponding to the bottommost layer.') + parser.add_argument('--full-context-alignment', action='store_true', + help='Whether or not alignment is supervised conditioned on the full target context.') + # fmt: on + + @classmethod + def build_model(cls, args, task): + # set any default arguments + transformer_align(args) + + transformer_model = TransformerModel.build_model(args, task) + return TransformerAlignModel( + transformer_model.encoder, transformer_model.decoder, args + ) + + def forward(self, src_tokens, src_lengths, prev_output_tokens): + encoder_out = self.encoder(src_tokens, src_lengths) + return self.forward_decoder(prev_output_tokens, encoder_out) + + def forward_decoder( + self, + prev_output_tokens, + encoder_out=None, + incremental_state=None, + features_only=False, + **extra_args, + ): + attn_args = { + "alignment_layer": self.alignment_layer, + "alignment_heads": self.alignment_heads, + } + decoder_out = self.decoder(prev_output_tokens, encoder_out, **attn_args) + + if self.full_context_alignment: + attn_args["full_context_alignment"] = self.full_context_alignment + _, alignment_out = self.decoder( + prev_output_tokens, + encoder_out, + features_only=True, + **attn_args, + **extra_args, + ) + decoder_out[1]["attn"] = alignment_out["attn"] + + return decoder_out + + +@register_model_architecture("transformer_align", "transformer_align") +def transformer_align(args): + args.alignment_heads = getattr(args, "alignment_heads", 1) + args.alignment_layer = getattr(args, "alignment_layer", 4) + args.full_context_alignment = getattr(args, "full_context_alignment", False) + base_architecture(args) + + +@register_model_architecture("transformer_align", "transformer_wmt_en_de_big_align") +def transformer_wmt_en_de_big_align(args): + args.alignment_heads = getattr(args, "alignment_heads", 1) + args.alignment_layer = getattr(args, "alignment_layer", 4) + transformer_wmt_en_de_big(args) diff --git a/fairseq/models/transformer_from_pretrained_xlm.py b/fairseq/models/transformer_from_pretrained_xlm.py new file mode 100644 index 0000000000000000000000000000000000000000..236d9942e1fb0238cc92e2b4f160520b5cdd6504 --- /dev/null +++ b/fairseq/models/transformer_from_pretrained_xlm.py @@ -0,0 +1,152 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import os +from typing import Any, Dict + +from fairseq import checkpoint_utils +from fairseq.data.legacy.masked_lm_dictionary import MaskedLMDictionary +from fairseq.models import register_model, register_model_architecture +from fairseq.models.transformer import ( + TransformerDecoder, + TransformerEncoder, + TransformerModel, + base_architecture as transformer_base_architecture, +) + + +@register_model("transformer_from_pretrained_xlm") +class TransformerFromPretrainedXLMModel(TransformerModel): + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + TransformerModel.add_args(parser) + parser.add_argument( + "--pretrained-xlm-checkpoint", + type=str, + metavar="STR", + help="XLM model to use for initializing transformer encoder and/or decoder", + ) + parser.add_argument( + "--init-encoder-only", + action="store_true", + help="if set, don't load the XLM weights and embeddings into decoder", + ) + parser.add_argument( + "--init-decoder-only", + action="store_true", + help="if set, don't load the XLM weights and embeddings into encoder", + ) + + @classmethod + def build_model(self, args, task, cls_dictionary=MaskedLMDictionary): + assert hasattr(args, "pretrained_xlm_checkpoint"), ( + "You must specify a path for --pretrained-xlm-checkpoint to use " + "--arch transformer_from_pretrained_xlm" + ) + assert isinstance(task.source_dictionary, cls_dictionary) and isinstance( + task.target_dictionary, cls_dictionary + ), ( + "You should use a MaskedLMDictionary when using --arch " + "transformer_from_pretrained_xlm because the pretrained XLM model " + "was trained using data binarized with MaskedLMDictionary. " + "For translation, you may want to use --task " + "translation_from_pretrained_xlm" + ) + assert not ( + getattr(args, "init_encoder_only", False) + and getattr(args, "init_decoder_only", False) + ), "Only one of --init-encoder-only and --init-decoder-only can be set." + return super().build_model(args, task) + + @classmethod + def build_encoder(cls, args, src_dict, embed_tokens): + return TransformerEncoderFromPretrainedXLM(args, src_dict, embed_tokens) + + @classmethod + def build_decoder(cls, args, tgt_dict, embed_tokens): + return TransformerDecoderFromPretrainedXLM(args, tgt_dict, embed_tokens) + + +def upgrade_state_dict_with_xlm_weights( + state_dict: Dict[str, Any], pretrained_xlm_checkpoint: str +) -> Dict[str, Any]: + """ + Load XLM weights into a Transformer encoder or decoder model. + + Args: + state_dict: state dict for either TransformerEncoder or + TransformerDecoder + pretrained_xlm_checkpoint: checkpoint to load XLM weights from + + Raises: + AssertionError: If architecture (num layers, attention heads, etc.) + does not match between the current Transformer encoder or + decoder and the pretrained_xlm_checkpoint + """ + if not os.path.exists(pretrained_xlm_checkpoint): + raise IOError("Model file not found: {}".format(pretrained_xlm_checkpoint)) + + state = checkpoint_utils.load_checkpoint_to_cpu(pretrained_xlm_checkpoint) + xlm_state_dict = state["model"] + for key in xlm_state_dict.keys(): + + for search_key in ["embed_tokens", "embed_positions", "layers"]: + if search_key in key: + subkey = key[key.find(search_key) :] + assert subkey in state_dict, ( + "{} Transformer encoder / decoder " + "state_dict does not contain {}. Cannot " + "load {} from pretrained XLM checkpoint " + "{} into Transformer.".format( + str(state_dict.keys()), subkey, key, pretrained_xlm_checkpoint + ) + ) + + state_dict[subkey] = xlm_state_dict[key] + return state_dict + + +class TransformerEncoderFromPretrainedXLM(TransformerEncoder): + def __init__(self, args, dictionary, embed_tokens): + super().__init__(args, dictionary, embed_tokens) + if getattr(args, "init_decoder_only", False): + # Don't load XLM weights for encoder if --init-decoder-only + return + + assert hasattr(args, "pretrained_xlm_checkpoint"), ( + "--pretrained-xlm-checkpoint must be specified to load Transformer " + "encoder from pretrained XLM" + ) + xlm_loaded_state_dict = upgrade_state_dict_with_xlm_weights( + state_dict=self.state_dict(), + pretrained_xlm_checkpoint=args.pretrained_xlm_checkpoint, + ) + self.load_state_dict(xlm_loaded_state_dict, strict=True) + + +class TransformerDecoderFromPretrainedXLM(TransformerDecoder): + def __init__(self, args, dictionary, embed_tokens, no_encoder_attn=False): + super().__init__(args, dictionary, embed_tokens, no_encoder_attn) + if getattr(args, "init_encoder_only", False): + # Don't load XLM weights for decoder if --init-encoder-only + return + assert hasattr(args, "pretrained_xlm_checkpoint"), ( + "--pretrained-xlm-checkpoint must be specified to load Transformer " + "decoder from pretrained XLM" + ) + + xlm_loaded_state_dict = upgrade_state_dict_with_xlm_weights( + state_dict=self.state_dict(), + pretrained_xlm_checkpoint=args.pretrained_xlm_checkpoint, + ) + self.load_state_dict(xlm_loaded_state_dict, strict=True) + + +@register_model_architecture( + "transformer_from_pretrained_xlm", "transformer_from_pretrained_xlm" +) +def base_architecture(args): + transformer_base_architecture(args) diff --git a/fairseq/models/transformer_lm.py b/fairseq/models/transformer_lm.py new file mode 100644 index 0000000000000000000000000000000000000000..a546776912b24f8aec4011f52b5ac1884112634e --- /dev/null +++ b/fairseq/models/transformer_lm.py @@ -0,0 +1,544 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +from dataclasses import dataclass, field +from typing import Optional + +from fairseq import options, utils +from fairseq.dataclass import ChoiceEnum, FairseqDataclass +from fairseq.models import ( + FairseqLanguageModel, + register_model, + register_model_architecture, +) +from fairseq.models.transformer import ( + DEFAULT_MIN_PARAMS_TO_WRAP, Embedding, TransformerDecoder +) +from fairseq.modules import AdaptiveInput, CharacterTokenEmbedder +from omegaconf import II + + +DEFAULT_MAX_TARGET_POSITIONS = 1024 + + +@dataclass +class TransformerLanguageModelConfig(FairseqDataclass): + activation_fn: ChoiceEnum(utils.get_available_activation_fns()) = field( + default="relu", metadata={"help": "activation function to use"} + ) + dropout: float = field(default=0.1, metadata={"help": "dropout probability"}) + attention_dropout: float = field( + default=0.0, metadata={"help": "dropout probability for attention weights"} + ) + activation_dropout: float = field( + default=0.0, metadata={"help": "dropout probability after activation in FFN."} + ) + relu_dropout: float = field( + default=0.0, metadata={"help": "dropout probability after activation in FFN."} + ) + decoder_embed_dim: int = field( + default=512, metadata={"help": "decoder embedding dimension"} + ) + decoder_output_dim: int = field( + default=512, metadata={"help": "decoder output dimension"} + ) + decoder_input_dim: int = field( + default=512, metadata={"help": "decoder input dimension"} + ) + decoder_ffn_embed_dim: int = field( + default=2048, metadata={"help": "decoder embedding dimension for FFN"} + ) + decoder_layers: int = field(default=6, metadata={"help": "num decoder layers"}) + decoder_attention_heads: int = field( + default=8, metadata={"help": "num decoder attention heads"} + ) + decoder_normalize_before: bool = field( + default=False, metadata={"help": "apply layernorm before each decoder block"} + ) + no_decoder_final_norm: bool = field( + default=False, + metadata={"help": "don't add an extra layernorm after the last decoder block"}, + ) + adaptive_softmax_cutoff: Optional[str] = field( + default=None, + metadata={ + "help": "comma separated list of adaptive softmax cutoff points. " + "Must be used with adaptive_loss criterion" + }, + ) + adaptive_softmax_dropout: float = field( + default=0, + metadata={"help": "sets adaptive softmax dropout for the tail projections"}, + ) + adaptive_softmax_factor: float = field( + default=4, metadata={"help": "adaptive input factor"} + ) + no_token_positional_embeddings: bool = field( + default=False, + metadata={ + "help": "if set, disables positional embeddings (outside self attention)" + }, + ) + share_decoder_input_output_embed: bool = field( + default=False, metadata={"help": "share decoder input and output embeddings"} + ) + character_embeddings: bool = field( + default=False, + metadata={ + "help": "if set, uses character embedding convolutions to produce token embeddings" + }, + ) + character_filters: str = field( + default="[(1, 64), (2, 128), (3, 192), (4, 256), (5, 256), (6, 256), (7, 256)]", + metadata={"help": "size of character embeddings"}, + ) + character_embedding_dim: int = field( + default=4, metadata={"help": "size of character embeddings"} + ) + char_embedder_highway_layers: int = field( + default=2, + metadata={"help": "number of highway layers for character token embeddder"}, + ) + adaptive_input: bool = field( + default=False, metadata={"help": "if set, uses adaptive input"} + ) + adaptive_input_factor: float = field( + default=4, metadata={"help": "adaptive input factor"} + ) + adaptive_input_cutoff: Optional[str] = field( + default=None, + metadata={"help": "comma separated list of adaptive input cutoff points."}, + ) + tie_adaptive_weights: bool = field( + default=False, + metadata={ + "help": "if set, ties the weights of adaptive softmax and adaptive input" + }, + ) + tie_adaptive_proj: bool = field( + default=False, + metadata={ + "help": "if set, ties the projection weights of adaptive softmax and adaptive input" + }, + ) + decoder_learned_pos: bool = field( + default=False, + metadata={"help": "use learned positional embeddings in the decoder"}, + ) + layernorm_embedding: bool = field( + default=False, metadata={"help": "add layernorm to embedding"} + ) + no_scale_embedding: bool = field( + default=False, metadata={"help": "if True, dont scale embeddings"} + ) + checkpoint_activations: bool = field( + default=False, metadata={"help": "checkpoint activations at each layer"} + ) + offload_activations: bool = field( + default=False, + metadata={"help": "move checkpointed activations to CPU after they are used."}, + ) + # config for "Reducing Transformer Depth on Demand with Structured Dropout" (Fan et al., 2019) + decoder_layerdrop: float = field( + default=0.0, metadata={"help": "LayerDrop probability for decoder"} + ) + decoder_layers_to_keep: Optional[str] = field( + default=None, + metadata={ + "help": "which layers to *keep* when pruning as a comma-separated list" + }, + ) + # config for Training with Quantization Noise for Extreme Model Compression ({Fan*, Stock*} et al., 2020) + quant_noise_pq: float = field( + default=0.0, + metadata={"help": "iterative PQ quantization noise at training time"}, + ) + quant_noise_pq_block_size: int = field( + default=8, + metadata={"help": "block size of quantization noise at training time"}, + ) + quant_noise_scalar: float = field( + default=0.0, + metadata={ + "help": "scalar quantization noise and scalar quantization at training time" + }, + ) + # config for Fully Sharded Data Parallel (FSDP) training + min_params_to_wrap: int = field( + default=DEFAULT_MIN_PARAMS_TO_WRAP, + metadata={ + "help": ( + "minimum number of params for a layer to be wrapped with FSDP() when " + "training with --ddp-backend=fully_sharded. Smaller values will " + "improve memory efficiency, but may make torch.distributed " + "communication less efficient due to smaller input sizes. This option " + "is set to 0 (i.e., always wrap) when --checkpoint-activations or " + "--offload-activations are passed." + ) + } + ) + # config for "BASE Layers: Simplifying Training of Large, Sparse Models" + base_layers: Optional[int] = field( + default=0, metadata={"help": "number of BASE layers in total"} + ) + base_sublayers: Optional[int] = field( + default=1, metadata={"help": "number of sublayers in each BASE layer"} + ) + base_shuffle: Optional[int] = field( + default=1, metadata={"help": "shuffle tokens between workers before computing assignment"} + ) + # options from other parts of the config + add_bos_token: bool = II("task.add_bos_token") + tokens_per_sample: int = II("task.tokens_per_sample") + max_target_positions: Optional[int] = II("task.max_target_positions") + tpu: bool = II("common.tpu") + + +@register_model("transformer_lm", dataclass=TransformerLanguageModelConfig) +class TransformerLanguageModel(FairseqLanguageModel): + @classmethod + def hub_models(cls): + def moses_fastbpe(path): + return {"path": path, "tokenizer": "moses", "bpe": "fastbpe"} + + def spm(path): + return {"path": path, "tokenizer": "space", "bpe": "sentencepiece"} + + return { + "transformer_lm.gbw.adaptive_huge": "https://dl.fbaipublicfiles.com/fairseq/models/lm/adaptive_lm_gbw_huge.tar.bz2", + "transformer_lm.wiki103.adaptive": "https://dl.fbaipublicfiles.com/fairseq/models/lm/adaptive_lm_wiki103.v2.tar.bz2", + "transformer_lm.wmt19.en": moses_fastbpe( + "https://dl.fbaipublicfiles.com/fairseq/models/lm/wmt19.en.tar.bz2" + ), + "transformer_lm.wmt19.de": moses_fastbpe( + "https://dl.fbaipublicfiles.com/fairseq/models/lm/wmt19.de.tar.bz2" + ), + "transformer_lm.wmt19.ru": moses_fastbpe( + "https://dl.fbaipublicfiles.com/fairseq/models/lm/wmt19.ru.tar.bz2" + ), + "transformer_lm.wmt20.en": spm( + "https://dl.fbaipublicfiles.com/fairseq/models/lm/wmt20.en.tar.gz" + ), + "transformer_lm.wmt20.ta": spm( + "https://dl.fbaipublicfiles.com/fairseq/models/lm/wmt20.ta.tar.gz" + ), + "transformer_lm.wmt20.iu.news": spm( + "https://dl.fbaipublicfiles.com/fairseq/models/lm/wmt20.iu.news.tar.gz" + ), + "transformer_lm.wmt20.iu.nh": spm( + "https://dl.fbaipublicfiles.com/fairseq/models/lm/wmt20.iu.nh.tar.gz" + ), + } + + def __init__(self, decoder): + super().__init__(decoder) + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + + if args.decoder_layers_to_keep: + args.decoder_layers = len(args.decoder_layers_to_keep.split(",")) + + if getattr(args, "max_target_positions", None) is None: + args.max_target_positions = getattr( + args, "tokens_per_sample", DEFAULT_MAX_TARGET_POSITIONS + ) + + if args.character_embeddings: + embed_tokens = CharacterTokenEmbedder( + task.source_dictionary, + eval(args.character_filters), + args.character_embedding_dim, + args.decoder_embed_dim, + args.char_embedder_highway_layers, + ) + elif args.adaptive_input: + embed_tokens = AdaptiveInput( + len(task.source_dictionary), + task.source_dictionary.pad(), + args.decoder_input_dim, + args.adaptive_input_factor, + args.decoder_embed_dim, + options.eval_str_list(args.adaptive_input_cutoff, type=int), + args.quant_noise_pq, + args.quant_noise_pq_block_size, + ) + else: + embed_tokens = cls.build_embedding( + args, task.source_dictionary, args.decoder_input_dim + ) + + if args.tie_adaptive_weights: + assert args.adaptive_input + assert args.adaptive_input_factor == args.adaptive_softmax_factor + assert ( + args.adaptive_softmax_cutoff == args.adaptive_input_cutoff + ), "{} != {}".format( + args.adaptive_softmax_cutoff, args.adaptive_input_cutoff + ) + assert args.decoder_input_dim == args.decoder_output_dim + + decoder = TransformerDecoder( + args, task.target_dictionary, embed_tokens, no_encoder_attn=True + ) + return cls(decoder) + + @classmethod + def build_embedding(cls, args, dictionary, embed_dim, path=None): + embed_tokens = Embedding(len(dictionary), embed_dim, dictionary.pad()) + return embed_tokens + + +def base_lm_architecture(args): + # backward compatibility for older model checkpoints + if hasattr(args, "no_tie_adaptive_proj"): + # previous models defined --no-tie-adaptive-proj, so use the existence of + # that option to determine if this is an "old" model checkpoint + args.no_decoder_final_norm = True # old models always set this to True + if args.no_tie_adaptive_proj is False: + args.tie_adaptive_proj = True + if hasattr(args, "decoder_final_norm"): + args.no_decoder_final_norm = not args.decoder_final_norm + + args.dropout = getattr(args, "dropout", 0.1) + args.attention_dropout = getattr(args, "attention_dropout", 0.0) + + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 512) + args.decoder_ffn_embed_dim = getattr(args, "decoder_ffn_embed_dim", 2048) + args.decoder_layers = getattr(args, "decoder_layers", 6) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 8) + args.adaptive_softmax_cutoff = getattr(args, "adaptive_softmax_cutoff", None) + args.adaptive_softmax_dropout = getattr(args, "adaptive_softmax_dropout", 0) + args.adaptive_softmax_factor = getattr(args, "adaptive_softmax_factor", 4) + args.decoder_learned_pos = getattr(args, "decoder_learned_pos", False) + args.activation_fn = getattr(args, "activation_fn", "relu") + + args.decoder_layerdrop = getattr(args, "decoder_layerdrop", 0) + args.decoder_layers_to_keep = getattr(args, "decoder_layers_to_keep", None) + args.quant_noise_pq = getattr(args, "quant_noise_pq", 0) + args.quant_noise_pq_block_size = getattr(args, "quant_noise_pq_block_size", 8) + args.quant_noise_scalar = getattr(args, "quant_noise_scalar", 0) + + args.base_layers = getattr(args, "base_layers", 0) + args.base_sublayers = getattr(args, "base_sublayers", 1) + args.base_shuffle = getattr(args, "base_shuffle", False) + + args.add_bos_token = getattr(args, "add_bos_token", False) + args.no_token_positional_embeddings = getattr( + args, "no_token_positional_embeddings", False + ) + args.share_decoder_input_output_embed = getattr( + args, "share_decoder_input_output_embed", False + ) + args.character_embeddings = getattr(args, "character_embeddings", False) + + args.decoder_output_dim = getattr( + args, "decoder_output_dim", args.decoder_embed_dim + ) + args.decoder_input_dim = getattr(args, "decoder_input_dim", args.decoder_embed_dim) + + # Model training is not stable without this + args.decoder_normalize_before = True + args.no_decoder_final_norm = getattr(args, "no_decoder_final_norm", False) + + args.adaptive_input = getattr(args, "adaptive_input", False) + args.adaptive_input_factor = getattr(args, "adaptive_input_factor", 4) + args.adaptive_input_cutoff = getattr(args, "adaptive_input_cutoff", None) + + args.tie_adaptive_weights = getattr(args, "tie_adaptive_weights", False) + args.tie_adaptive_proj = getattr(args, "tie_adaptive_proj", False) + + args.no_scale_embedding = getattr(args, "no_scale_embedding", False) + args.layernorm_embedding = getattr(args, "layernorm_embedding", False) + args.checkpoint_activations = getattr(args, "checkpoint_activations", False) + args.offload_activations = getattr(args, "offload_activations", False) + if args.offload_activations: + args.checkpoint_activations = True + + +@register_model_architecture("transformer_lm", "transformer_lm_big") +def transformer_lm_big(args): + args.decoder_layers = getattr(args, "decoder_layers", 12) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 1024) + args.decoder_ffn_embed_dim = getattr(args, "decoder_ffn_embed_dim", 4096) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 16) + base_lm_architecture(args) + + +@register_model_architecture("transformer_lm", "transformer_lm_wiki103") +@register_model_architecture("transformer_lm", "transformer_lm_baevski_wiki103") +def transformer_lm_baevski_wiki103(args): + args.decoder_layers = getattr(args, "decoder_layers", 16) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 8) + args.dropout = getattr(args, "dropout", 0.3) + args.adaptive_input = getattr(args, "adaptive_input", True) + args.tie_adaptive_weights = getattr(args, "tie_adaptive_weights", True) + args.adaptive_input_cutoff = getattr(args, "adaptive_input_cutoff", "20000,60000") + args.adaptive_softmax_cutoff = getattr( + args, "adaptive_softmax_cutoff", "20000,60000" + ) + args.adaptive_softmax_dropout = getattr(args, "adaptive_softmax_dropout", 0.2) + args.attention_dropout = getattr(args, "attention_dropout", 0.1) + args.activation_dropout = getattr(args, "activation_dropout", 0.1) + args.no_decoder_final_norm = getattr(args, "no_decoder_final_norm", True) + args.tie_adaptive_proj = getattr(args, "tie_adaptive_proj", True) + transformer_lm_big(args) + + +@register_model_architecture("transformer_lm", "transformer_lm_gbw") +@register_model_architecture("transformer_lm", "transformer_lm_baevski_gbw") +def transformer_lm_baevski_gbw(args): + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 512) + args.dropout = getattr(args, "dropout", 0.1) + args.attention_dropout = getattr(args, "attention_dropout", 0.1) + args.no_decoder_final_norm = getattr(args, "no_decoder_final_norm", True) + transformer_lm_big(args) + + +@register_model_architecture("transformer_lm", "transformer_lm_gpt") +def transformer_lm_gpt(args): + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 768) + args.decoder_ffn_embed_dim = getattr(args, "decoder_ffn_embed_dim", 3072) + args.decoder_layers = getattr(args, "decoder_layers", 12) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 12) + args.dropout = getattr(args, "dropout", 0.1) + args.attention_dropout = getattr(args, "attention_dropout", 0.1) + args.activation_fn = getattr(args, "activation_fn", "gelu") + base_lm_architecture(args) + + +@register_model_architecture("transformer_lm", "transformer_lm_gpt2_small") +def transformer_lm_gpt2_small(args): + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 1024) + args.decoder_ffn_embed_dim = getattr(args, "decoder_ffn_embed_dim", 4096) + args.decoder_layers = getattr(args, "decoder_layers", 24) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 16) + args.dropout = getattr(args, "dropout", 0.1) + args.attention_dropout = getattr(args, "attention_dropout", 0.1) + args.activation_fn = getattr(args, "activation_fn", "gelu") + base_lm_architecture(args) + + +@register_model_architecture("transformer_lm", "transformer_lm_gpt2_tiny") +def transformer_lm_gpt2_tiny(args): + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 64) + args.decoder_ffn_embed_dim = getattr(args, "decoder_ffn_embed_dim", 64) + args.decoder_layers = getattr(args, "decoder_layers", 2) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 1) + args.dropout = getattr(args, "dropout", 0.1) + args.attention_dropout = getattr(args, "attention_dropout", 0.1) + args.activation_fn = getattr(args, "activation_fn", "gelu") + base_lm_architecture(args) + + +@register_model_architecture("transformer_lm", "transformer_lm_gpt2_medium") +def transformer_lm_gpt2_medium(args): + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 1280) + args.decoder_ffn_embed_dim = getattr(args, "decoder_ffn_embed_dim", 5120) + args.decoder_layers = getattr(args, "decoder_layers", 36) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 20) + args.dropout = getattr(args, "dropout", 0.1) + args.attention_dropout = getattr(args, "attention_dropout", 0.1) + args.activation_fn = getattr(args, "activation_fn", "gelu") + base_lm_architecture(args) + + +@register_model_architecture("transformer_lm", "transformer_lm_gpt2_big") +def transformer_lm_gpt2_big(args): + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 1600) + args.decoder_ffn_embed_dim = getattr(args, "decoder_ffn_embed_dim", 6400) + args.decoder_layers = getattr(args, "decoder_layers", 48) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 25) + args.dropout = getattr(args, "dropout", 0.1) + args.attention_dropout = getattr(args, "attention_dropout", 0.1) + args.activation_fn = getattr(args, "activation_fn", "gelu") + base_lm_architecture(args) + + +def base_gpt3_architecture(args): + args.decoder_input_dim = args.decoder_embed_dim + args.decoder_output_dim = args.decoder_embed_dim + args.decoder_ffn_embed_dim = getattr(args, "decoder_ffn_embed_dim", args.decoder_embed_dim * 4) + # GPT-3 used learned positional embeddings, rather than sinusoidal + args.decoder_learned_pos = getattr(args, "decoder_learned_pos", True) + args.dropout = getattr(args, "dropout", 0.0) + args.attention_dropout = getattr(args, "attention_dropout", 0.0) + args.activation_fn = getattr(args, "activation_fn", "gelu") + args.share_decoder_input_output_embed = True + base_lm_architecture(args) + + +@register_model_architecture("transformer_lm", "transformer_lm_gpt3_small") +def transformer_lm_gpt3_small(args): + # 125M params + args.decoder_layers = getattr(args, "decoder_layers", 12) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 768) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 12) + base_gpt3_architecture(args) + + +@register_model_architecture("transformer_lm", "transformer_lm_gpt3_medium") +def transformer_lm_gpt3_medium(args): + # 350M params + args.decoder_layers = getattr(args, "decoder_layers", 24) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 1024) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 16) + base_gpt3_architecture(args) + + +@register_model_architecture("transformer_lm", "transformer_lm_gpt3_large") +def transformer_lm_gpt3_large(args): + # 760M params + args.decoder_layers = getattr(args, "decoder_layers", 24) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 1536) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 16) + base_gpt3_architecture(args) + + +@register_model_architecture("transformer_lm", "transformer_lm_gpt3_xl") +def transformer_lm_gpt3_xl(args): + # 1.3B params + args.decoder_layers = getattr(args, "decoder_layers", 24) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 2048) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 32) + base_gpt3_architecture(args) + + +@register_model_architecture("transformer_lm", "transformer_lm_gpt3_2_7") +def transformer_lm_gpt3_2_7(args): + # 2.7B params + args.decoder_layers = getattr(args, "decoder_layers", 32) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 2560) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 32) + base_gpt3_architecture(args) + + +@register_model_architecture("transformer_lm", "transformer_lm_gpt3_6_7") +def transformer_lm_gpt3_6_7(args): + # 6.7B params + args.decoder_layers = getattr(args, "decoder_layers", 32) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 4096) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 32) + base_gpt3_architecture(args) + + +@register_model_architecture("transformer_lm", "transformer_lm_gpt3_13") +def transformer_lm_gpt3_13(args): + # 13B params + args.decoder_layers = getattr(args, "decoder_layers", 40) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 5120) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 40) + base_gpt3_architecture(args) + + +@register_model_architecture("transformer_lm", "transformer_lm_gpt3_175") +def transformer_lm_gpt3_175(args): + # 175B params + args.decoder_layers = getattr(args, "decoder_layers", 96) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 12288) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 96) + base_gpt3_architecture(args) diff --git a/fairseq/models/wav2vec/__init__.py b/fairseq/models/wav2vec/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..06cec18183ca14cd534d14558e8b44e25f3e69d5 --- /dev/null +++ b/fairseq/models/wav2vec/__init__.py @@ -0,0 +1,8 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .wav2vec import * # noqa +from .wav2vec2 import * # noqa +from .wav2vec2_asr import * # noqa diff --git a/fairseq/models/wav2vec/wav2vec.py b/fairseq/models/wav2vec/wav2vec.py new file mode 100644 index 0000000000000000000000000000000000000000..af6604da10f504baabff50bf14a6eb2214bffef3 --- /dev/null +++ b/fairseq/models/wav2vec/wav2vec.py @@ -0,0 +1,630 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from dataclasses import dataclass, field +import logging +import math +from typing import Optional, Tuple +from omegaconf import II +import sys + +import torch +import torch.nn as nn +import torch.nn.functional as F +from fairseq.dataclass import ChoiceEnum, FairseqDataclass +from fairseq.models import BaseFairseqModel, register_model +from fairseq.modules import ( + Fp32GroupNorm, + Fp32LayerNorm, + GumbelVectorQuantizer, + KmeansVectorQuantizer, + TransposeLast, +) +from fairseq.tasks import FairseqTask +from fairseq.utils import buffered_arange + + +logger = logging.getLogger(__name__) + + +AGGREGATOR_CHOICES = ChoiceEnum(["cnn", "gru"]) +PROJECT_FEATURES_CHOICES = ChoiceEnum(["none", "same", "new"]) +ACTIVATION_CHOICES = ChoiceEnum(["relu", "gelu"]) +VQ_TYPE_CHOICES = ChoiceEnum(["none", "gumbel", "kmeans"]) + + +@dataclass +class Wav2VecConfig(FairseqDataclass): + prediction_steps: int = field( + default=12, metadata={"help": "number of steps ahead to predict"} + ) + sample_distance: Optional[int] = field( + default=None, + metadata={ + "help": "sample distance from target. does not work properly with cross-sampling" + }, + ) + cross_sample_negatives: int = field( + default=0, metadata={"help": "num of cross sampled negatives"} + ) + num_negatives: int = field( + default=10, metadata={"help": "num of sampled negatives"} + ) + conv_feature_layers: str = field( + default="[(512, 10, 5), (512, 8, 4), (512, 4, 2), (512, 4, 2), (512, 4, 2), (512, 1, 1), (512, 1, 1), (512, 1, 1)]", + metadata={ + "help": "convolutional feature extraction layers [(dim, kernel_size, stride), ...]" + }, + ) + conv_aggregator_layers: str = field( + default="[(512, 2, 1), (512, 3, 1), (512, 4, 1), (512, 5, 1), (512, 6, 1), (512, 7, 1), (512, 8, 1), (512, 9, 1), (512, 10, 1), (512, 11, 1), (512, 12, 1), (512, 13, 1)]", + metadata={ + "help": "convolutional aggregator layers [(dim, kernel_size, stride), ...]" + }, + ) + dropout: float = field( + default=0.0, metadata={"help": "dropout to apply within the model"} + ) + dropout_features: float = field( + default=0.0, metadata={"help": "dropout to apply to the features"} + ) + dropout_agg: float = field( + default=0.0, metadata={"help": "dropout to apply after aggregation step"} + ) + aggregator: AGGREGATOR_CHOICES = field( + default="cnn", metadata={"help": "type of aggregator to use"} + ) + gru_dim: int = field(default=512, metadata={"help": "GRU dimensionality"}) + no_conv_bias: bool = field( + default=False, metadata={"help": "if set, does not learn bias for conv layers"} + ) + agg_zero_pad: bool = field( + default=False, + metadata={"help": "if set, zero pads in aggregator instead of repl pad"}, + ) + skip_connections_feat: bool = field( + default=False, + metadata={"help": "if set, adds skip connections to the feature extractor"}, + ) + skip_connections_agg: bool = field( + default=True, + metadata={"help": "if set, adds skip connections to the aggregator"}, + ) + residual_scale: float = field( + default=0.5, metadata={"help": "scales residual by sqrt(value)"} + ) + log_compression: bool = field( + default=True, + metadata={"help": "if set, adds a log compression to feature extractor"}, + ) + balanced_classes: bool = field( + default=False, + metadata={"help": "if set, loss is scaled to balance for number of negatives"}, + ) + project_features: PROJECT_FEATURES_CHOICES = field( + default="none", + metadata={ + "help": "if not none, features are projected using the (same or new) aggregator" + }, + ) + non_affine_group_norm: bool = field( + default=False, metadata={"help": "if set, group norm is not affine"} + ) + offset: str = field( + default="auto", + metadata={ + "help": "if set to 'auto', it is computed automatically from the receptive field, else set to int value" + }, + ) + activation: ACTIVATION_CHOICES = field( + default="relu", + metadata={ + "help": "if set to 'auto', it is computed automatically from the receptive field, else set to int value" + }, + ) + vq_type: VQ_TYPE_CHOICES = field( + default="none", metadata={"help": "which type of quantizer to use"} + ) + vq_vars: int = field( + default=320, + metadata={"help": "project to this many vector quantized variables per group"}, + ) + vq_groups: int = field( + default=2, metadata={"help": "number of groups of latent variables"} + ) + vq_dim: int = field( + default=0, + metadata={ + "help": "uses this dimensionality for quantized vectors. 0 to use model dim // groups" + }, + ) + vq_depth: int = field( + default=1, metadata={"help": "number of layers for vq weight projection"} + ) + combine_groups: bool = field( + default=False, metadata={"help": "if set, variables are shared among groups"} + ) + vq_temp: Tuple[float, float, float] = field( + default=(2.0, 0.5, 0.999995), + metadata={ + "help": "temperature for latent variable sampling with gumbel softmax. should be a tuple of 3 values (start, end, decay)" + }, + ) + vq_gamma: float = field( + default=0.25, + metadata={"help": "gamma parameter for kmeans style vector quantization"}, + ) + infonce: bool = II("criterion.infonce") + + +@register_model("wav2vec", dataclass=Wav2VecConfig) +class Wav2VecModel(BaseFairseqModel): + @classmethod + def build_model(cls, cfg: Wav2VecConfig, task: FairseqTask): + """Build a new model instance.""" + + model = Wav2VecModel(cfg) + logger.info(model) + return model + + def __init__(self, cfg: Wav2VecConfig): + super().__init__() + + self.prediction_steps = cfg.prediction_steps + offset = cfg.offset + + if cfg.activation == "relu": + activation = nn.ReLU() + elif cfg.activation == "gelu": + activation = nn.GELU() + else: + raise Exception("unknown activation " + cfg.activation) + + feature_enc_layers = eval(cfg.conv_feature_layers) + self.feature_extractor = ConvFeatureExtractionModel( + conv_layers=feature_enc_layers, + dropout=0.0, + log_compression=cfg.log_compression, + skip_connections=cfg.skip_connections_feat, + residual_scale=cfg.residual_scale, + non_affine_group_norm=cfg.non_affine_group_norm, + activation=activation, + ) + embed = feature_enc_layers[-1][0] + + self.vector_quantizer = None + if cfg.vq_type == "gumbel": + self.vector_quantizer = GumbelVectorQuantizer( + dim=embed, + num_vars=cfg.vq_vars, + temp=cfg.vq_temp, + groups=cfg.vq_groups, + combine_groups=cfg.combine_groups, + vq_dim=cfg.vq_dim if cfg.vq_dim > 0 else embed, + time_first=False, + activation=activation, + weight_proj_depth=cfg.vq_depth, + weight_proj_factor=2, + ) + elif cfg.vq_type == "kmeans": + self.vector_quantizer = KmeansVectorQuantizer( + dim=embed, + num_vars=cfg.vq_vars, + groups=cfg.vq_groups, + combine_groups=cfg.combine_groups, + vq_dim=cfg.vq_dim if cfg.vq_dim > 0 else embed, + time_first=False, + gamma=cfg.vq_gamma, + ) + else: + assert ( + cfg.vq_type == "none" or cfg.vq_type is None + ), "Unknown quantizer type" + + if cfg.offset == "auto": + jin = 0 + rin = 0 + for _, k, stride in feature_enc_layers: + if rin == 0: + rin = k + rin = rin + (k - 1) * jin + if jin == 0: + jin = stride + else: + jin *= stride + offset = math.ceil(rin / jin) + + offset = int(offset) + + def make_aggregator(): + if cfg.aggregator == "cnn": + agg_layers = eval(cfg.conv_aggregator_layers) + agg_dim = agg_layers[-1][0] + feature_aggregator = ConvAggegator( + conv_layers=agg_layers, + embed=embed, + dropout=cfg.dropout, + skip_connections=cfg.skip_connections_agg, + residual_scale=cfg.residual_scale, + non_affine_group_norm=cfg.non_affine_group_norm, + conv_bias=not cfg.no_conv_bias, + zero_pad=cfg.agg_zero_pad, + activation=activation, + ) + elif cfg.aggregator == "gru": + agg_dim = cfg.gru_dim + feature_aggregator = nn.Sequential( + TransposeLast(), + nn.GRU( + input_size=embed, + hidden_size=agg_dim, + num_layers=1, + dropout=cfg.dropout, + ), + TransposeLast(deconstruct_idx=0), + ) + else: + raise Exception("unknown aggregator type " + cfg.aggregator) + + return feature_aggregator, agg_dim + + self.feature_aggregator, agg_dim = make_aggregator() + + self.wav2vec_predictions = Wav2VecPredictionsModel( + in_dim=agg_dim, + out_dim=embed, + prediction_steps=cfg.prediction_steps, + n_negatives=cfg.num_negatives, + cross_sample_negatives=cfg.cross_sample_negatives, + sample_distance=cfg.sample_distance, + dropout=cfg.dropout, + offset=offset, + balanced_classes=cfg.balanced_classes, + infonce=cfg.infonce, + ) + + self.dropout_feats = nn.Dropout(p=cfg.dropout_features) + self.dropout_agg = nn.Dropout(p=cfg.dropout_agg) + + if cfg.project_features == "none": + self.project_features = None + elif cfg.project_features == "same": + self.project_features = self.feature_aggregator + elif cfg.project_features == "new": + self.project_features, _ = make_aggregator() + + def forward(self, source): + result = {} + + features = self.feature_extractor(source) + if self.vector_quantizer: + q_res = self.vector_quantizer(features) + features = q_res["x"] + for k in q_res.keys(): + if k != "x": + result[k] = q_res[k] + + x = self.dropout_feats(features) + x = self.feature_aggregator(x) + x = self.dropout_agg(x) + + if self.project_features is not None: + features = self.project_features(features) + x, targets = self.wav2vec_predictions(x, features) + result["cpc_logits"] = x + result["cpc_targets"] = targets + + return result + + def upgrade_state_dict_named(self, state_dict, name): + super().upgrade_state_dict_named(state_dict, name) + + def max_positions(self): + """Maximum length supported by the model.""" + return sys.maxsize + + def get_logits(self, net_output): + logits = net_output["cpc_logits"] + return logits + + def get_targets(self, sample, net_output): + t = net_output["cpc_targets"] + if isinstance(t, tuple): + t = t[0] + return t.contiguous() + + def get_target_weights(self, targets, net_output): + targets = net_output["cpc_targets"] + if isinstance(targets, tuple) and targets[-1] is not None: + return targets[-1] + return None + + def get_extra_losses(self, net_output): + loss = None + if "prob_perplexity" in net_output: + loss = net_output["num_vars"] - net_output["prob_perplexity"] + elif "kmeans_loss" in net_output: + loss = net_output["kmeans_loss"] + + return loss + + +def norm_block(is_layer_norm, dim, affine=True): + if is_layer_norm: + mod = nn.Sequential( + TransposeLast(), + Fp32LayerNorm(dim, elementwise_affine=affine), + TransposeLast(), + ) + else: + mod = Fp32GroupNorm(1, dim, affine=affine) + + return mod + + +class ConvFeatureExtractionModel(nn.Module): + def __init__( + self, + conv_layers, + dropout, + log_compression, + skip_connections, + residual_scale, + non_affine_group_norm, + activation, + ): + super().__init__() + + def block(n_in, n_out, k, stride): + return nn.Sequential( + nn.Conv1d(n_in, n_out, k, stride=stride, bias=False), + nn.Dropout(p=dropout), + norm_block( + is_layer_norm=False, dim=n_out, affine=not non_affine_group_norm + ), + activation, + ) + + in_d = 1 + self.conv_layers = nn.ModuleList() + for dim, k, stride in conv_layers: + self.conv_layers.append(block(in_d, dim, k, stride)) + in_d = dim + + self.log_compression = log_compression + self.skip_connections = skip_connections + self.residual_scale = math.sqrt(residual_scale) + + def forward(self, x): + # BxT -> BxCxT + x = x.unsqueeze(1) + + for conv in self.conv_layers: + residual = x + x = conv(x) + if self.skip_connections and x.size(1) == residual.size(1): + tsz = x.size(2) + r_tsz = residual.size(2) + residual = residual[..., :: r_tsz // tsz][..., :tsz] + x = (x + residual) * self.residual_scale + + if self.log_compression: + x = x.abs() + x = x + 1 + x = x.log() + + return x + + +class ZeroPad1d(nn.Module): + def __init__(self, pad_left, pad_right): + super().__init__() + self.pad_left = pad_left + self.pad_right = pad_right + + def forward(self, x): + return F.pad(x, (self.pad_left, self.pad_right)) + + +class ConvAggegator(nn.Module): + def __init__( + self, + conv_layers, + embed, + dropout, + skip_connections, + residual_scale, + non_affine_group_norm, + conv_bias, + zero_pad, + activation, + ): + super().__init__() + + def block(n_in, n_out, k, stride): + # padding dims only really make sense for stride = 1 + ka = k // 2 + kb = ka - 1 if k % 2 == 0 else ka + + pad = ( + ZeroPad1d(ka + kb, 0) if zero_pad else nn.ReplicationPad1d((ka + kb, 0)) + ) + + return nn.Sequential( + pad, + nn.Conv1d(n_in, n_out, k, stride=stride, bias=conv_bias), + nn.Dropout(p=dropout), + norm_block(False, n_out, affine=not non_affine_group_norm), + activation, + ) + + in_d = embed + self.conv_layers = nn.ModuleList() + self.residual_proj = nn.ModuleList() + for dim, k, stride in conv_layers: + if in_d != dim and skip_connections: + self.residual_proj.append(nn.Conv1d(in_d, dim, 1, bias=False)) + else: + self.residual_proj.append(None) + + self.conv_layers.append(block(in_d, dim, k, stride)) + in_d = dim + self.conv_layers = nn.Sequential(*self.conv_layers) + self.skip_connections = skip_connections + self.residual_scale = math.sqrt(residual_scale) + + def forward(self, x): + for rproj, conv in zip(self.residual_proj, self.conv_layers): + residual = x + x = conv(x) + if self.skip_connections: + if rproj is not None: + residual = rproj(residual) + x = (x + residual) * self.residual_scale + return x + + +class Wav2VecPredictionsModel(nn.Module): + def __init__( + self, + in_dim, + out_dim, + prediction_steps, + n_negatives, + cross_sample_negatives, + sample_distance, + dropout, + offset, + balanced_classes, + infonce, + ): + super().__init__() + + self.n_negatives = n_negatives + self.cross_sample_negatives = cross_sample_negatives + self.sample_distance = sample_distance + self.project_to_steps = nn.ConvTranspose2d( + in_dim, out_dim, (1, prediction_steps) + ) + self.dropout = nn.Dropout(p=dropout) + self.offset = offset + self.balanced_classes = balanced_classes + self.infonce = infonce + + def sample_negatives(self, y): + bsz, fsz, tsz = y.shape + + y = y.transpose(0, 1) # BCT -> CBT + y = y.contiguous().view(fsz, -1) # CBT => C(BxT) + + cross_high = tsz * bsz + high = tsz if self.sample_distance is None else min(tsz, self.sample_distance) + assert high > 1 + + neg_idxs = torch.randint(low=0, high=high, size=(bsz, self.n_negatives * tsz)) + + with torch.no_grad(): + if self.n_negatives > 0: + tszs = ( + buffered_arange(tsz) + .unsqueeze(-1) + .expand(-1, self.n_negatives) + .flatten() + ) + + neg_idxs = torch.randint( + low=0, high=high - 1, size=(bsz, self.n_negatives * tsz) + ) + neg_idxs[neg_idxs >= tszs] += 1 + + if self.cross_sample_negatives > 0: + tszs = ( + buffered_arange(tsz) + .unsqueeze(-1) + .expand(-1, self.cross_sample_negatives) + .flatten() + ) + + cross_neg_idxs = torch.randint( + low=0, + high=cross_high - 1, + size=(bsz, self.cross_sample_negatives * tsz), + ) + cross_neg_idxs[cross_neg_idxs >= tszs] += 1 + + if self.n_negatives > 0: + for i in range(1, bsz): + neg_idxs[i] += i * high + else: + neg_idxs = cross_neg_idxs + + if self.cross_sample_negatives > 0 and self.n_negatives > 0: + neg_idxs = torch.cat([neg_idxs, cross_neg_idxs], dim=1) + + negs = y[..., neg_idxs.view(-1)] + negs = negs.view( + fsz, bsz, self.n_negatives + self.cross_sample_negatives, tsz + ).permute( + 2, 1, 0, 3 + ) # to NxBxCxT + + return negs + + def forward(self, x, y): + + x = x.unsqueeze(-1) + x = self.project_to_steps(x) # BxCxTxS + x = self.dropout(x) + + negatives = self.sample_negatives(y) + y = y.unsqueeze(0) + targets = torch.cat([y, negatives], dim=0) # Copies x B x C x T + + copies = targets.size(0) + bsz, dim, tsz, steps = x.shape + steps = min(steps, tsz - self.offset) + + predictions = x.new( + bsz * copies * (tsz - self.offset + 1) * steps + - ((steps + 1) * steps // 2) * copies * bsz + ) + if self.infonce: + labels = predictions.new_full( + (predictions.shape[0] // copies,), 0, dtype=torch.long + ) + else: + labels = torch.zeros_like(predictions) + weights = ( + torch.full_like(labels, 1 / self.n_negatives) + if self.balanced_classes and not self.infonce + else None + ) + + start = end = 0 + for i in range(steps): + offset = i + self.offset + end = start + (tsz - offset) * bsz * copies + if self.infonce: + predictions[start:end] = torch.einsum( + "bct,nbct->tbn", x[..., :-offset, i], targets[..., offset:] + ).flatten() + else: + pos_num = (end - start) // copies + predictions[start:end] = torch.einsum( + "bct,nbct->nbt", x[..., :-offset, i], targets[..., offset:] + ).flatten() + labels[start : start + pos_num] = 1.0 + if weights is not None: + weights[start : start + pos_num] = 1.0 + start = end + assert end == predictions.numel(), "{} != {}".format(end, predictions.numel()) + + if self.infonce: + predictions = predictions.view(-1, copies) + else: + if weights is not None: + labels = (labels, weights) + + return predictions, labels diff --git a/fairseq/models/wav2vec/wav2vec2.py b/fairseq/models/wav2vec/wav2vec2.py new file mode 100644 index 0000000000000000000000000000000000000000..714fd3ab50443b8d15715b1cf5abd4eb517298c4 --- /dev/null +++ b/fairseq/models/wav2vec/wav2vec2.py @@ -0,0 +1,1016 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +from dataclasses import dataclass, field +from typing import List, Tuple + +import numpy as np +import torch +import torch.nn as nn +import torch.nn.functional as F +from fairseq import utils +from fairseq.data.data_utils import compute_mask_indices +from fairseq.dataclass import ChoiceEnum, FairseqDataclass +from fairseq.models import BaseFairseqModel, register_model +from fairseq.modules import ( + Fp32GroupNorm, + Fp32LayerNorm, + GradMultiply, + GumbelVectorQuantizer, + LayerNorm, + MultiheadAttention, + SamePad, + TransposeLast, +) +from fairseq.modules.transformer_sentence_encoder import init_bert_params +from fairseq.utils import buffered_arange, index_put, is_xla_tensor + + +EXTRACTOR_MODE_CHOICES = ChoiceEnum(["default", "layer_norm"]) +MASKING_DISTRIBUTION_CHOICES = ChoiceEnum(["static", "uniform", "normal", "poisson"]) + + +@dataclass +class Wav2Vec2Config(FairseqDataclass): + extractor_mode: EXTRACTOR_MODE_CHOICES = field( + default="default", + metadata={ + "help": "mode for feature extractor. default has a single group norm with d " + "groups in the first conv block, whereas layer_norm has layer norms in " + "every block (meant to use with normalize=True)" + }, + ) + encoder_layers: int = field( + default=12, metadata={"help": "num encoder layers in the transformer"} + ) + encoder_embed_dim: int = field( + default=768, metadata={"help": "encoder embedding dimension"} + ) + encoder_ffn_embed_dim: int = field( + default=3072, metadata={"help": "encoder embedding dimension for FFN"} + ) + encoder_attention_heads: int = field( + default=12, metadata={"help": "num encoder attention heads"} + ) + activation_fn: ChoiceEnum(utils.get_available_activation_fns()) = field( + default="gelu", metadata={"help": "activation function to use"} + ) + + # dropouts + dropout: float = field( + default=0.1, metadata={"help": "dropout probability for the transformer"} + ) + attention_dropout: float = field( + default=0.1, metadata={"help": "dropout probability for attention weights"} + ) + activation_dropout: float = field( + default=0.0, metadata={"help": "dropout probability after activation in FFN"} + ) + encoder_layerdrop: float = field( + default=0.0, metadata={"help": "probability of dropping a tarnsformer layer"} + ) + dropout_input: float = field( + default=0.0, + metadata={"help": "dropout to apply to the input (after feat extr)"}, + ) + dropout_features: float = field( + default=0.0, + metadata={"help": "dropout to apply to the features (after feat extr)"}, + ) + + final_dim: int = field( + default=0, + metadata={ + "help": "project final representations and targets to this many dimensions." + "set to encoder_embed_dim is <= 0" + }, + ) + layer_norm_first: bool = field( + default=False, metadata={"help": "apply layernorm first in the transformer"} + ) + conv_feature_layers: str = field( + default="[(512, 10, 5)] + [(512, 3, 2)] * 4 + [(512,2,2)] + [(512,2,2)]", + metadata={ + "help": "string describing convolutional feature extraction layers in form of a python list that contains " + "[(dim, kernel_size, stride), ...]" + }, + ) + conv_bias: bool = field( + default=False, metadata={"help": "include bias in conv encoder"} + ) + logit_temp: float = field( + default=0.1, metadata={"help": "temperature to divide logits by"} + ) + quantize_targets: bool = field( + default=False, metadata={"help": "use quantized targets"} + ) + quantize_input: bool = field( + default=False, metadata={"help": "use quantized inputs"} + ) + same_quantizer: bool = field( + default=False, metadata={"help": "use same quantizer for inputs and targets"} + ) + target_glu: bool = field( + default=False, metadata={"help": "adds projection + glu to targets"} + ) + feature_grad_mult: float = field( + default=1.0, metadata={"help": "multiply feature extractor var grads by this"} + ) + quantizer_depth: int = field( + default=1, + metadata={"help": "number of quantizer layers"}, + ) + quantizer_factor: int = field( + default=3, + metadata={ + "help": "dimensionality increase for inner quantizer layers (if depth > 1)" + }, + ) + latent_vars: int = field( + default=320, + metadata={"help": "number of latent variables V in each group of the codebook"}, + ) + latent_groups: int = field( + default=2, + metadata={"help": "number of groups G of latent variables in the codebook"}, + ) + latent_dim: int = field( + default=0, + metadata={ + "help": "if > 0, uses this dimensionality for latent variables. " + "otherwise uses final_dim / latent_groups" + }, + ) + + # masking + mask_length: int = field(default=10, metadata={"help": "mask length"}) + mask_prob: float = field( + default=0.65, metadata={"help": "probability of replacing a token with mask"} + ) + mask_selection: MASKING_DISTRIBUTION_CHOICES = field( + default="static", metadata={"help": "how to choose mask length"} + ) + mask_other: float = field( + default=0, + metadata={ + "help": "secondary mask argument (used for more complex distributions), " + "see help in compute_mask_indices" + }, + ) + no_mask_overlap: bool = field( + default=False, metadata={"help": "whether to allow masks to overlap"} + ) + mask_min_space: int = field( + default=1, + metadata={"help": "min space between spans (if no overlap is enabled)"}, + ) + + # channel masking + mask_channel_length: int = field( + default=10, metadata={"help": "length of the mask for features (channels)"} + ) + mask_channel_prob: float = field( + default=0.0, metadata={"help": "probability of replacing a feature with 0"} + ) + mask_channel_before: bool = False + mask_channel_selection: MASKING_DISTRIBUTION_CHOICES = field( + default="static", + metadata={"help": "how to choose mask length for channel masking"}, + ) + mask_channel_other: float = field( + default=0, + metadata={ + "help": "secondary mask argument (used for more complex distributions), " + "see help in compute_mask_indicesh" + }, + ) + no_mask_channel_overlap: bool = field( + default=False, metadata={"help": "whether to allow channel masks to overlap"} + ) + mask_channel_min_space: int = field( + default=1, + metadata={"help": "min space between spans (if no overlap is enabled)"}, + ) + + # negative selection + num_negatives: int = field( + default=100, + metadata={"help": "number of negative examples from the same sample"}, + ) + negatives_from_everywhere: bool = field( + default=False, + metadata={"help": "sample negatives from everywhere, not just masked states"}, + ) + cross_sample_negatives: int = field( + default=0, metadata={"help": "number of negative examples from the any sample"} + ) + codebook_negatives: int = field( + default=0, metadata={"help": "number of negative examples codebook"} + ) + + # positional embeddings + conv_pos: int = field( + default=128, + metadata={"help": "number of filters for convolutional positional embeddings"}, + ) + conv_pos_groups: int = field( + default=16, + metadata={"help": "number of groups for convolutional positional embedding"}, + ) + + latent_temp: Tuple[float, float, float] = field( + default=(2, 0.5, 0.999995), + metadata={ + "help": "temperature for latent variable sampling. " + "can be tuple of 3 values (start, end, decay)" + }, + ) + + +@register_model("wav2vec2", dataclass=Wav2Vec2Config) +class Wav2Vec2Model(BaseFairseqModel): + def __init__(self, cfg: Wav2Vec2Config): + super().__init__() + self.cfg = cfg + + feature_enc_layers = eval(cfg.conv_feature_layers) + self.embed = feature_enc_layers[-1][0] + + self.feature_extractor = ConvFeatureExtractionModel( + conv_layers=feature_enc_layers, + dropout=0.0, + mode=cfg.extractor_mode, + conv_bias=cfg.conv_bias, + ) + + self.post_extract_proj = ( + nn.Linear(self.embed, cfg.encoder_embed_dim) + if self.embed != cfg.encoder_embed_dim and not cfg.quantize_input + else None + ) + + self.mask_prob = cfg.mask_prob + self.mask_selection = cfg.mask_selection + self.mask_other = cfg.mask_other + self.mask_length = cfg.mask_length + self.no_mask_overlap = cfg.no_mask_overlap + self.mask_min_space = cfg.mask_min_space + + self.mask_channel_prob = cfg.mask_channel_prob + self.mask_channel_before = cfg.mask_channel_before + self.mask_channel_selection = cfg.mask_channel_selection + self.mask_channel_other = cfg.mask_channel_other + self.mask_channel_length = cfg.mask_channel_length + self.no_mask_channel_overlap = cfg.no_mask_channel_overlap + self.mask_channel_min_space = cfg.mask_channel_min_space + + self.dropout_input = nn.Dropout(cfg.dropout_input) + self.dropout_features = nn.Dropout(cfg.dropout_features) + + self.feature_grad_mult = cfg.feature_grad_mult + + self.quantizer = None + self.input_quantizer = None + + self.n_negatives = cfg.num_negatives + self.cross_sample_negatives = cfg.cross_sample_negatives + self.codebook_negatives = cfg.codebook_negatives + self.negatives_from_everywhere = cfg.negatives_from_everywhere + + self.logit_temp = cfg.logit_temp + + final_dim = cfg.final_dim if cfg.final_dim > 0 else cfg.encoder_embed_dim + + if cfg.quantize_targets: + vq_dim = cfg.latent_dim if cfg.latent_dim > 0 else final_dim + self.quantizer = GumbelVectorQuantizer( + dim=self.embed, + num_vars=cfg.latent_vars, + temp=cfg.latent_temp, + groups=cfg.latent_groups, + combine_groups=False, + vq_dim=vq_dim, + time_first=True, + weight_proj_depth=cfg.quantizer_depth, + weight_proj_factor=cfg.quantizer_factor, + ) + self.project_q = nn.Linear(vq_dim, final_dim) + else: + self.project_q = nn.Linear(self.embed, final_dim) + + if cfg.quantize_input: + if cfg.same_quantizer and self.quantizer is not None: + vq_dim = final_dim + self.input_quantizer = self.quantizer + else: + vq_dim = cfg.latent_dim if cfg.latent_dim > 0 else cfg.encoder_embed_dim + self.input_quantizer = GumbelVectorQuantizer( + dim=self.embed, + num_vars=cfg.latent_vars, + temp=cfg.latent_temp, + groups=cfg.latent_groups, + combine_groups=False, + vq_dim=vq_dim, + time_first=True, + weight_proj_depth=cfg.quantizer_depth, + weight_proj_factor=cfg.quantizer_factor, + ) + self.project_inp = nn.Linear(vq_dim, cfg.encoder_embed_dim) + + self.mask_emb = nn.Parameter( + torch.FloatTensor(cfg.encoder_embed_dim).uniform_() + ) + + self.encoder = TransformerEncoder(cfg) + self.layer_norm = LayerNorm(self.embed) + + self.target_glu = None + if cfg.target_glu: + self.target_glu = nn.Sequential( + nn.Linear(final_dim, final_dim * 2), nn.GLU() + ) + + self.final_proj = nn.Linear(cfg.encoder_embed_dim, final_dim) + + def upgrade_state_dict_named(self, state_dict, name): + super().upgrade_state_dict_named(state_dict, name) + """Upgrade a (possibly old) state dict for new versions of fairseq.""" + return state_dict + + @classmethod + def build_model(cls, cfg: Wav2Vec2Config, task=None): + """Build a new model instance.""" + + return cls(cfg) + + def apply_mask( + self, + x, + padding_mask, + mask_indices=None, + mask_channel_indices=None, + ): + B, T, C = x.shape + + if self.mask_channel_prob > 0 and self.mask_channel_before: + mask_channel_indices = compute_mask_indices( + (B, C), + None, + self.mask_channel_prob, + self.mask_channel_length, + self.mask_channel_selection, + self.mask_channel_other, + no_overlap=self.no_mask_channel_overlap, + min_space=self.mask_channel_min_space, + ) + mask_channel_indices = ( + torch.from_numpy(mask_channel_indices) + .to(x.device) + .unsqueeze(1) + .expand(-1, T, -1) + ) + x[mask_channel_indices] = 0 + + if self.mask_prob > 0: + if mask_indices is None: + mask_indices = compute_mask_indices( + (B, T), + padding_mask, + self.mask_prob, + self.mask_length, + self.mask_selection, + self.mask_other, + min_masks=2, + no_overlap=self.no_mask_overlap, + min_space=self.mask_min_space, + ) + mask_indices = torch.from_numpy(mask_indices).to(x.device) + x = index_put(x, mask_indices, self.mask_emb) + else: + mask_indices = None + + if self.mask_channel_prob > 0 and not self.mask_channel_before: + if mask_channel_indices is None: + mask_channel_indices = compute_mask_indices( + (B, C), + None, + self.mask_channel_prob, + self.mask_channel_length, + self.mask_channel_selection, + self.mask_channel_other, + no_overlap=self.no_mask_channel_overlap, + min_space=self.mask_channel_min_space, + ) + mask_channel_indices = ( + torch.from_numpy(mask_channel_indices) + .to(x.device) + .unsqueeze(1) + .expand(-1, T, -1) + ) + x = index_put(x, mask_channel_indices, 0) + + return x, mask_indices + + def sample_negatives(self, y, num, padding_count=None): + + if self.n_negatives == 0 and self.cross_sample_negatives == 0: + return y.new(0) + + bsz, tsz, fsz = y.shape + y = y.view(-1, fsz) # BTC => (BxT)C + + # FIXME: what happens if padding_count is specified? + cross_high = tsz * bsz + high = tsz - (padding_count or 0) + with torch.no_grad(): + assert high > 1, f"{bsz,tsz,fsz}" + + if self.n_negatives > 0: + tszs = ( + buffered_arange(num) + .unsqueeze(-1) + .expand(-1, self.n_negatives) + .flatten() + ) + + neg_idxs = torch.randint( + low=0, high=high - 1, size=(bsz, self.n_negatives * num) + ) + neg_idxs[neg_idxs >= tszs] += 1 + + if self.cross_sample_negatives > 0: + tszs = ( + buffered_arange(num) + .unsqueeze(-1) + .expand(-1, self.cross_sample_negatives) + .flatten() + ) + + cross_neg_idxs = torch.randint( + low=0, + high=cross_high - 1, + size=(bsz, self.cross_sample_negatives * num), + ) + cross_neg_idxs[cross_neg_idxs >= tszs] += 1 + + if self.n_negatives > 0: + for i in range(1, bsz): + neg_idxs[i] += i * high + else: + neg_idxs = cross_neg_idxs + + if self.cross_sample_negatives > 0 and self.n_negatives > 0: + neg_idxs = torch.cat([neg_idxs, cross_neg_idxs], dim=1) + + negs = y[neg_idxs.view(-1)] + negs = negs.view( + bsz, num, self.n_negatives + self.cross_sample_negatives, fsz + ).permute( + 2, 0, 1, 3 + ) # to NxBxTxC + return negs, neg_idxs + + def compute_preds(self, x, y, negatives): + + neg_is_pos = (y == negatives).all(-1) + y = y.unsqueeze(0) + targets = torch.cat([y, negatives], dim=0) + + logits = torch.cosine_similarity(x.float(), targets.float(), dim=-1).type_as(x) + + logits = logits / self.logit_temp + + if is_xla_tensor(logits) or neg_is_pos.any(): + fillval = -float(2 ** 30) + if not hasattr(self, "_inftensor"): + self._inftensor = ( + torch.tensor(fillval).to(x.device) + if is_xla_tensor(logits) + else float("-inf") + ) + logits[1:] = index_put(logits[1:], neg_is_pos, self._inftensor) + + return logits + + def _get_feat_extract_output_lengths(self, input_lengths: torch.LongTensor): + """ + Computes the output length of the convolutional layers + """ + + def _conv_out_length(input_length, kernel_size, stride): + return torch.floor((input_length - kernel_size) / stride + 1) + + conv_cfg_list = eval(self.cfg.conv_feature_layers) + + for i in range(len(conv_cfg_list)): + input_lengths = _conv_out_length( + input_lengths, conv_cfg_list[i][1], conv_cfg_list[i][2] + ) + + return input_lengths.to(torch.long) + + def forward( + self, + source, + padding_mask=None, + mask=True, + features_only=False, + layer=None, + mask_indices=None, + mask_channel_indices=None, + padding_count=None, + ): + + if self.feature_grad_mult > 0: + features = self.feature_extractor(source) + if self.feature_grad_mult != 1.0: + features = GradMultiply.apply(features, self.feature_grad_mult) + else: + with torch.no_grad(): + features = self.feature_extractor(source) + + features_pen = features.float().pow(2).mean() + + features = features.transpose(1, 2) + features = self.layer_norm(features) + unmasked_features = features.clone() + + if padding_mask is not None and padding_mask.any(): + input_lengths = (1 - padding_mask.long()).sum(-1) + # apply conv formula to get real output_lengths + output_lengths = self._get_feat_extract_output_lengths(input_lengths) + + padding_mask = torch.zeros( + features.shape[:2], dtype=features.dtype, device=features.device + ) + + # these two operations makes sure that all values + # before the output lengths indices are attended to + padding_mask[ + ( + torch.arange(padding_mask.shape[0], device=padding_mask.device), + output_lengths - 1, + ) + ] = 1 + padding_mask = (1 - padding_mask.flip([-1]).cumsum(-1).flip([-1])).bool() + else: + padding_mask = None + + if self.post_extract_proj is not None: + features = self.post_extract_proj(features) + + features = self.dropout_input(features) + unmasked_features = self.dropout_features(unmasked_features) + + num_vars = None + code_ppl = None + prob_ppl = None + curr_temp = None + + if self.input_quantizer: + q = self.input_quantizer(features, produce_targets=False) + features = q["x"] + num_vars = q["num_vars"] + code_ppl = q["code_perplexity"] + prob_ppl = q["prob_perplexity"] + curr_temp = q["temp"] + features = self.project_inp(features) + + if mask: + x, mask_indices = self.apply_mask( + features, + padding_mask, + mask_indices=mask_indices, + mask_channel_indices=mask_channel_indices, + ) + if not is_xla_tensor(x) and mask_indices is not None: + # tpu-comment: reducing the size in a dynamic way causes + # too many recompilations on xla. + y = unmasked_features[mask_indices].view( + unmasked_features.size(0), -1, unmasked_features.size(-1) + ) + else: + y = unmasked_features + else: + x = features + y = unmasked_features + mask_indices = None + + x, layer_results = self.encoder(x, padding_mask=padding_mask, layer=layer) + + if features_only: + return { + "x": x, + "padding_mask": padding_mask, + "features": unmasked_features, + "layer_results": layer_results, + } + + if self.quantizer: + q = self.quantizer(y, produce_targets=False) + y = q["x"] + num_vars = q["num_vars"] + code_ppl = q["code_perplexity"] + prob_ppl = q["prob_perplexity"] + curr_temp = q["temp"] + + y = self.project_q(y) + + if self.negatives_from_everywhere: + neg_cands = self.quantizer(unmasked_features, produce_targets=False)[ + "x" + ] + negs, _ = self.sample_negatives( + neg_cands, + y.size(1), + padding_count=padding_count, + ) + negs = self.project_q(negs) + + else: + negs, _ = self.sample_negatives( + y, + y.size(1), + padding_count=padding_count, + ) + + if self.codebook_negatives > 0: + cb_negs = self.quantizer.sample_from_codebook( + y.size(0) * y.size(1), self.codebook_negatives + ) + cb_negs = cb_negs.view( + self.codebook_negatives, y.size(0), y.size(1), -1 + ) # order doesnt matter + cb_negs = self.project_q(cb_negs) + negs = torch.cat([negs, cb_negs], dim=0) + else: + y = self.project_q(y) + + if self.negatives_from_everywhere: + negs, _ = self.sample_negatives( + unmasked_features, + y.size(1), + padding_count=padding_count, + ) + negs = self.project_q(negs) + else: + negs, _ = self.sample_negatives( + y, + y.size(1), + padding_count=padding_count, + ) + + if not is_xla_tensor(x): + # tpu-comment: reducing the size in a dynamic way causes + # too many recompilations on xla. + x = x[mask_indices].view(x.size(0), -1, x.size(-1)) + + if self.target_glu: + y = self.target_glu(y) + negs = self.target_glu(negs) + + x = self.final_proj(x) + x = self.compute_preds(x, y, negs) + + result = { + "x": x, + "padding_mask": padding_mask, + "features_pen": features_pen, + } + + if prob_ppl is not None: + result["prob_perplexity"] = prob_ppl + result["code_perplexity"] = code_ppl + result["num_vars"] = num_vars + result["temp"] = curr_temp + + return result + + def quantize(self, x): + assert self.quantizer is not None + x = self.feature_extractor(x) + x = x.transpose(1, 2) + x = self.layer_norm(x) + return self.quantizer.forward_idx(x) + + def extract_features(self, source, padding_mask, mask=False, layer=None): + res = self.forward( + source, padding_mask, mask=mask, features_only=True, layer=layer + ) + return res + + def get_logits(self, net_output): + logits = net_output["x"] + logits = logits.transpose(0, 2) + logits = logits.reshape(-1, logits.size(-1)) + return logits + + def get_targets(self, sample, net_output, expand_steps=True): + x = net_output["x"] + return x.new_zeros(x.size(1) * x.size(2), dtype=torch.long) + + def get_extra_losses(self, net_output): + pen = [] + + if "prob_perplexity" in net_output: + pen.append( + (net_output["num_vars"] - net_output["prob_perplexity"]) + / net_output["num_vars"] + ) + + if "features_pen" in net_output: + pen.append(net_output["features_pen"]) + + return pen + + def remove_pretraining_modules(self): + self.quantizer = None + self.project_q = None + self.target_glu = None + self.final_proj = None + + +class ConvFeatureExtractionModel(nn.Module): + def __init__( + self, + conv_layers: List[Tuple[int, int, int]], + dropout: float = 0.0, + mode: str = "default", + conv_bias: bool = False, + ): + super().__init__() + + assert mode in {"default", "layer_norm"} + + def block( + n_in, + n_out, + k, + stride, + is_layer_norm=False, + is_group_norm=False, + conv_bias=False, + ): + def make_conv(): + conv = nn.Conv1d(n_in, n_out, k, stride=stride, bias=conv_bias) + nn.init.kaiming_normal_(conv.weight) + return conv + + assert ( + is_layer_norm and is_group_norm + ) == False, "layer norm and group norm are exclusive" + + if is_layer_norm: + return nn.Sequential( + make_conv(), + nn.Dropout(p=dropout), + nn.Sequential( + TransposeLast(), + Fp32LayerNorm(dim, elementwise_affine=True), + TransposeLast(), + ), + nn.GELU(), + ) + elif is_group_norm: + return nn.Sequential( + make_conv(), + nn.Dropout(p=dropout), + Fp32GroupNorm(dim, dim, affine=True), + nn.GELU(), + ) + else: + return nn.Sequential(make_conv(), nn.Dropout(p=dropout), nn.GELU()) + + in_d = 1 + self.conv_layers = nn.ModuleList() + for i, cl in enumerate(conv_layers): + assert len(cl) == 3, "invalid conv definition: " + str(cl) + (dim, k, stride) = cl + + self.conv_layers.append( + block( + in_d, + dim, + k, + stride, + is_layer_norm=mode == "layer_norm", + is_group_norm=mode == "default" and i == 0, + conv_bias=conv_bias, + ) + ) + in_d = dim + + def forward(self, x): + + # BxT -> BxCxT + x = x.unsqueeze(1) + + for conv in self.conv_layers: + x = conv(x) + + return x + + +class TransformerEncoder(nn.Module): + def __init__(self, args): + super().__init__() + + self.dropout = args.dropout + self.embedding_dim = args.encoder_embed_dim + + self.pos_conv = nn.Conv1d( + self.embedding_dim, + self.embedding_dim, + kernel_size=args.conv_pos, + padding=args.conv_pos // 2, + groups=args.conv_pos_groups, + ) + dropout = 0 + std = math.sqrt((4 * (1.0 - dropout)) / (args.conv_pos * self.embedding_dim)) + nn.init.normal_(self.pos_conv.weight, mean=0, std=std) + nn.init.constant_(self.pos_conv.bias, 0) + + self.pos_conv = nn.utils.weight_norm(self.pos_conv, name="weight", dim=2) + self.pos_conv = nn.Sequential(self.pos_conv, SamePad(args.conv_pos), nn.GELU()) + + self.layers = nn.ModuleList( + [ + TransformerSentenceEncoderLayer( + embedding_dim=self.embedding_dim, + ffn_embedding_dim=args.encoder_ffn_embed_dim, + num_attention_heads=args.encoder_attention_heads, + dropout=self.dropout, + attention_dropout=args.attention_dropout, + activation_dropout=args.activation_dropout, + activation_fn=args.activation_fn, + layer_norm_first=args.layer_norm_first, + ) + for _ in range(args.encoder_layers) + ] + ) + + self.layer_norm_first = args.layer_norm_first + self.layer_norm = LayerNorm(self.embedding_dim) + self.layerdrop = args.encoder_layerdrop + + self.apply(init_bert_params) + + def forward(self, x, padding_mask=None, layer=None): + x, layer_results = self.extract_features(x, padding_mask, layer) + + if self.layer_norm_first and layer is None: + x = self.layer_norm(x) + + return x, layer_results + + def extract_features(self, x, padding_mask=None, tgt_layer=None): + + if padding_mask is not None: + x = index_put(x, padding_mask, 0) + + x_conv = self.pos_conv(x.transpose(1, 2)) + x_conv = x_conv.transpose(1, 2) + x = x + x_conv + + if not self.layer_norm_first: + x = self.layer_norm(x) + + x = F.dropout(x, p=self.dropout, training=self.training) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + + layer_results = [] + r = None + for i, layer in enumerate(self.layers): + dropout_probability = np.random.random() + if not self.training or (dropout_probability > self.layerdrop): + x, z = layer(x, self_attn_padding_mask=padding_mask, need_weights=False) + if tgt_layer is not None: + layer_results.append((x, z)) + if i == tgt_layer: + r = x + break + + if r is not None: + x = r + + # T x B x C -> B x T x C + x = x.transpose(0, 1) + + return x, layer_results + + def max_positions(self): + """Maximum output length supported by the encoder.""" + return self.args.max_positions + + def upgrade_state_dict_named(self, state_dict, name): + """Upgrade a (possibly old) state dict for new versions of fairseq.""" + return state_dict + + +class TransformerSentenceEncoderLayer(nn.Module): + """ + Implements a Transformer Encoder Layer used in BERT/XLM style pre-trained + models. + """ + + def __init__( + self, + embedding_dim: float = 768, + ffn_embedding_dim: float = 3072, + num_attention_heads: float = 8, + dropout: float = 0.1, + attention_dropout: float = 0.1, + activation_dropout: float = 0.1, + activation_fn: str = "relu", + layer_norm_first: bool = False, + ) -> None: + + super().__init__() + # Initialize parameters + self.embedding_dim = embedding_dim + self.dropout = dropout + self.activation_dropout = activation_dropout + + # Initialize blocks + self.activation_fn = utils.get_activation_fn(activation_fn) + self.self_attn = MultiheadAttention( + self.embedding_dim, + num_attention_heads, + dropout=attention_dropout, + self_attention=True, + ) + + self.dropout1 = nn.Dropout(dropout) + self.dropout2 = nn.Dropout(self.activation_dropout) + self.dropout3 = nn.Dropout(dropout) + + self.layer_norm_first = layer_norm_first + + # layer norm associated with the self attention layer + self.self_attn_layer_norm = LayerNorm(self.embedding_dim) + self.fc1 = nn.Linear(self.embedding_dim, ffn_embedding_dim) + self.fc2 = nn.Linear(ffn_embedding_dim, self.embedding_dim) + + # layer norm associated with the position wise feed-forward NN + self.final_layer_norm = LayerNorm(self.embedding_dim) + + def forward( + self, + x: torch.Tensor, + self_attn_mask: torch.Tensor = None, + self_attn_padding_mask: torch.Tensor = None, + need_weights: bool = False, + att_args=None, + ): + """ + LayerNorm is applied either before or after the self-attention/ffn + modules similar to the original Transformer imlementation. + """ + residual = x + + if self.layer_norm_first: + x = self.self_attn_layer_norm(x) + x, attn = self.self_attn( + query=x, + key=x, + value=x, + key_padding_mask=self_attn_padding_mask, + attn_mask=self_attn_mask, + ) + x = self.dropout1(x) + x = residual + x + + residual = x + x = self.final_layer_norm(x) + x = self.activation_fn(self.fc1(x)) + x = self.dropout2(x) + x = self.fc2(x) + x = self.dropout3(x) + x = residual + x + else: + x, attn = self.self_attn( + query=x, + key=x, + value=x, + key_padding_mask=self_attn_padding_mask, + ) + + x = self.dropout1(x) + x = residual + x + + x = self.self_attn_layer_norm(x) + + residual = x + x = self.activation_fn(self.fc1(x)) + x = self.dropout2(x) + x = self.fc2(x) + x = self.dropout3(x) + x = residual + x + x = self.final_layer_norm(x) + + return x, attn diff --git a/fairseq/models/wav2vec/wav2vec2_asr.py b/fairseq/models/wav2vec/wav2vec2_asr.py new file mode 100644 index 0000000000000000000000000000000000000000..405d1e613a9bbf8294302c4526267f1330ffc5cd --- /dev/null +++ b/fairseq/models/wav2vec/wav2vec2_asr.py @@ -0,0 +1,655 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from argparse import Namespace +import contextlib +import copy +import math +import numpy as np +import torch +import torch.nn as nn +import torch.nn.functional as F +from dataclasses import dataclass, field +from omegaconf import MISSING, II, open_dict +from typing import Any, Optional + +from fairseq import checkpoint_utils, tasks, utils +from fairseq.dataclass import FairseqDataclass +from fairseq.dataclass.utils import convert_namespace_to_omegaconf +from fairseq.tasks import FairseqTask +from fairseq.models import ( + BaseFairseqModel, + FairseqEncoder, + FairseqEncoderDecoderModel, + FairseqIncrementalDecoder, + register_model, +) +from fairseq.models.wav2vec.wav2vec2 import MASKING_DISTRIBUTION_CHOICES +from fairseq.modules import ( + LayerNorm, + PositionalEmbedding, + TransformerDecoderLayer, +) + + +@dataclass +class Wav2Vec2AsrConfig(FairseqDataclass): + w2v_path: str = field( + default=MISSING, metadata={"help": "path to wav2vec 2.0 model"} + ) + no_pretrained_weights: bool = field( + default=False, metadata={"help": "if true, does not load pretrained weights"} + ) + dropout_input: float = field( + default=0.0, + metadata={"help": "dropout to apply to the input (after feat extr)"}, + ) + final_dropout: float = field( + default=0.0, + metadata={"help": "dropout after transformer and before final projection"}, + ) + dropout: float = field( + default=0.0, metadata={"help": "dropout probability inside wav2vec 2.0 model"} + ) + attention_dropout: float = field( + default=0.0, + metadata={ + "help": "dropout probability for attention weights inside wav2vec 2.0 model" + }, + ) + activation_dropout: float = field( + default=0.0, + metadata={ + "help": "dropout probability after activation in FFN inside wav2vec 2.0 model" + }, + ) + + # masking + apply_mask: bool = field( + default=False, metadata={"help": "apply masking during fine-tuning"} + ) + mask_length: int = field( + default=10, metadata={"help": "repeat the mask indices multiple times"} + ) + mask_prob: float = field( + default=0.5, + metadata={ + "help": "probability of replacing a token with mask (normalized by length)" + }, + ) + mask_selection: MASKING_DISTRIBUTION_CHOICES = field( + default="static", metadata={"help": "how to choose masks"} + ) + mask_other: float = field( + default=0, + metadata={ + "help": "secondary mask argument (used for more complex distributions), " + "see help in compute_mask_indices" + }, + ) + no_mask_overlap: bool = field( + default=False, metadata={"help": "whether to allow masks to overlap"} + ) + + # channel masking + mask_channel_length: int = field( + default=10, metadata={"help": "length of the mask for features (channels)"} + ) + mask_channel_prob: float = field( + default=0.0, metadata={"help": "probability of replacing a feature with 0"} + ) + mask_channel_selection: MASKING_DISTRIBUTION_CHOICES = field( + default="static", + metadata={"help": "how to choose mask length for channel masking"}, + ) + mask_channel_other: float = field( + default=0, + metadata={ + "help": "secondary mask argument (used for more complex distributions), " + "see help in compute_mask_indicesh" + }, + ) + no_mask_channel_overlap: bool = field( + default=False, metadata={"help": "whether to allow channel masks to overlap"} + ) + freeze_finetune_updates: int = field( + default=0, metadata={"help": "dont finetune wav2vec for this many updates"} + ) + feature_grad_mult: float = field( + default=0.0, metadata={"help": "reset feature grad mult in wav2vec 2.0 to this"} + ) + layerdrop: float = field( + default=0.0, metadata={"help": "probability of dropping a layer in wav2vec 2.0"} + ) + mask_channel_before: bool = False + normalize: bool = II("task.normalize") + data: str = II("task.data") + # this holds the loaded wav2vec args + w2v_args: Any = None + + +@dataclass +class Wav2Vec2CtcConfig(Wav2Vec2AsrConfig): + blank_weight: float = 0 + blank_mode: str = "add" + mask_min_space: Optional[int] = field( + default=1, + metadata={"help": "min space between spans (if no overlap is enabled)"}, + ) + mask_channel_min_space: Optional[int] = field( + default=1, + metadata={"help": "min space between spans (if no overlap is enabled)"}, + ) + conv_feature_layers: Optional[str] = field( + default="[(512, 10, 5)] + [(512, 3, 2)] * 4 + [(512,2,2)] + [(512,2,2)]", + metadata={ + "help": ( + "string describing convolutional feature extraction " + "layers in form of a python list that contains " + "[(dim, kernel_size, stride), ...]" + ), + }, + ) + encoder_embed_dim: Optional[int] = field( + default=768, metadata={"help": "encoder embedding dimension"} + ) + + +@register_model("wav2vec_ctc", dataclass=Wav2Vec2CtcConfig) +class Wav2VecCtc(BaseFairseqModel): + def __init__(self, cfg: Wav2Vec2CtcConfig, w2v_encoder: BaseFairseqModel): + super().__init__() + self.cfg = cfg + self.w2v_encoder = w2v_encoder + self.blank_weight = cfg.blank_weight + self.blank_mode = cfg.blank_mode + + def upgrade_state_dict_named(self, state_dict, name): + super().upgrade_state_dict_named(state_dict, name) + return state_dict + + @classmethod + def build_model(cls, cfg: Wav2Vec2CtcConfig, task: FairseqTask): + """Build a new model instance.""" + w2v_encoder = Wav2VecEncoder(cfg, len(task.target_dictionary)) + return cls(cfg, w2v_encoder) + + def get_logits(self, net_output, normalize=False): + logits = net_output["encoder_out"] + if self.blank_weight != 0: + if self.blank_mode == "add": + logits[..., 0] += self.blank_weight + elif self.blank_mode == "set": + logits[..., 0] = self.blank_weight + else: + raise Exception(f"invalid blank mode {self.blank_mode}") + + if net_output["padding_mask"] is not None and net_output["padding_mask"].any(): + logits[net_output["padding_mask"].T][..., 0] = float("inf") + logits[net_output["padding_mask"].T][..., 1:] = float("-inf") + + if normalize: + logits = utils.log_softmax(logits.float(), dim=-1) + + return logits + + def get_normalized_probs(self, net_output, log_probs): + """Get normalized probabilities (or log probs) from a net's output.""" + + logits = self.get_logits(net_output) + + if log_probs: + return utils.log_softmax(logits.float(), dim=-1) + else: + return utils.softmax(logits.float(), dim=-1) + + def forward(self, **kwargs): + x = self.w2v_encoder(**kwargs) + return x + + +@dataclass +class Wav2Vec2Seq2SeqConfig(Wav2Vec2AsrConfig): + decoder_embed_dim: int = field( + default=768, metadata={"help": "decoder embedding dimension"} + ) + decoder_ffn_embed_dim: int = field( + default=3072, metadata={"help": "decoder embedding dimension for FFN"} + ) + decoder_layers: int = field(default=6, metadata={"help": "num of decoder layers"}) + decoder_layerdrop: float = field( + default=0.0, metadata={"help": "decoder layerdrop chance"} + ) + decoder_attention_heads: int = field( + default=4, metadata={"help": "num decoder attention heads"} + ) + decoder_learned_pos: bool = field( + default=False, + metadata={"help": "use learned positional embeddings in the decoder"}, + ) + decoder_normalize_before: bool = field( + default=False, metadata={"help": "apply layernorm before each decoder block"} + ) + no_token_positional_embeddings: bool = field( + default=False, + metadata={ + "help": "if set, disables positional embeddings (outside self attention)" + }, + ) + decoder_dropout: float = field( + default=0.0, metadata={"help": "dropout probability in the decoder"} + ) + decoder_attention_dropout: float = field( + default=0.0, + metadata={ + "help": "dropout probability for attention weights inside the decoder" + }, + ) + decoder_activation_dropout: float = field( + default=0.0, + metadata={ + "help": "dropout probability after activation in FFN inside the decoder" + }, + ) + max_target_positions: int = field( + default=2048, metadata={"help": "max target positions"} + ) + share_decoder_input_output_embed: bool = field( + default=False, metadata={"help": "share decoder input and output embeddings"} + ) + autoregressive: bool = II("task.autoregressive") + + +@register_model("wav2vec_seq2seq", dataclass=Wav2Vec2Seq2SeqConfig) +class Wav2Vec2Seq2SeqModel(FairseqEncoderDecoderModel): + def __init__(self, encoder, decoder): + super().__init__(encoder, decoder) + + @classmethod + def build_model(cls, cfg: Wav2Vec2Seq2SeqConfig, task: FairseqTask): + """Build a new model instance.""" + + assert ( + cfg.autoregressive + ), "Please set task.autoregressive=true for seq2seq asr models" + + src_dict, tgt_dict = task.source_dictionary, task.target_dictionary + + def build_embedding(dictionary, embed_dim): + num_embeddings = len(dictionary) + padding_idx = dictionary.pad() + emb = Embedding(num_embeddings, embed_dim, padding_idx) + return emb + + decoder_embed_tokens = build_embedding(tgt_dict, cfg.decoder_embed_dim) + + encoder = cls.build_encoder(cfg) + decoder = cls.build_decoder(cfg, tgt_dict, decoder_embed_tokens) + + return Wav2Vec2Seq2SeqModel(encoder, decoder) + + @classmethod + def build_encoder(cls, cfg: Wav2Vec2AsrConfig): + return Wav2VecEncoder(cfg) + + @classmethod + def build_decoder(cls, cfg: Wav2Vec2Seq2SeqConfig, tgt_dict, embed_tokens): + return TransformerDecoder(cfg, tgt_dict, embed_tokens) + + def forward(self, **kwargs): + encoder_out = self.encoder(tbc=False, **kwargs) + decoder_out = self.decoder(encoder_out=encoder_out, **kwargs) + return decoder_out + + def upgrade_state_dict_named(self, state_dict, name): + super().upgrade_state_dict_named(state_dict, name) + return state_dict + + +class Wav2VecEncoder(FairseqEncoder): + def __init__(self, cfg: Wav2Vec2AsrConfig, output_size=None): + self.apply_mask = cfg.apply_mask + + arg_overrides = { + "dropout": cfg.dropout, + "activation_dropout": cfg.activation_dropout, + "dropout_input": cfg.dropout_input, + "attention_dropout": cfg.attention_dropout, + "mask_length": cfg.mask_length, + "mask_prob": cfg.mask_prob, + "mask_selection": cfg.mask_selection, + "mask_other": cfg.mask_other, + "no_mask_overlap": cfg.no_mask_overlap, + "mask_channel_length": cfg.mask_channel_length, + "mask_channel_prob": cfg.mask_channel_prob, + "mask_channel_before": cfg.mask_channel_before, + "mask_channel_selection": cfg.mask_channel_selection, + "mask_channel_other": cfg.mask_channel_other, + "no_mask_channel_overlap": cfg.no_mask_channel_overlap, + "encoder_layerdrop": cfg.layerdrop, + "feature_grad_mult": cfg.feature_grad_mult, + } + + if cfg.w2v_args is None: + state = checkpoint_utils.load_checkpoint_to_cpu(cfg.w2v_path, arg_overrides) + w2v_args = state.get("cfg", None) + if w2v_args is None: + w2v_args = convert_namespace_to_omegaconf(state["args"]) + cfg.w2v_args = w2v_args + else: + state = None + w2v_args = cfg.w2v_args + if isinstance(w2v_args, Namespace): + cfg.w2v_args = w2v_args = convert_namespace_to_omegaconf(w2v_args) + + assert cfg.normalize == w2v_args.task.normalize, ( + "Fine-tuning works best when data normalization is the same. " + "Please check that --normalize is set or unset for both pre-training and here" + ) + + w2v_args.task.data = cfg.data + task = tasks.setup_task(w2v_args.task) + model = task.build_model(w2v_args.model) + + if state is not None and not cfg.no_pretrained_weights: + model.load_state_dict(state["model"], strict=True) + + model.remove_pretraining_modules() + + super().__init__(task.source_dictionary) + + d = w2v_args.model.encoder_embed_dim + + self.w2v_model = model + + self.final_dropout = nn.Dropout(cfg.final_dropout) + self.freeze_finetune_updates = cfg.freeze_finetune_updates + self.num_updates = 0 + + targ_d = None + self.proj = None + + if output_size is not None: + targ_d = output_size + elif getattr(cfg, "decoder_embed_dim", d) != d: + targ_d = cfg.decoder_embed_dim + + if targ_d is not None: + self.proj = Linear(d, targ_d) + + def set_num_updates(self, num_updates): + """Set the number of parameters updates.""" + super().set_num_updates(num_updates) + self.num_updates = num_updates + + def forward(self, source, padding_mask, tbc=True, **kwargs): + w2v_args = { + "source": source, + "padding_mask": padding_mask, + "mask": self.apply_mask and self.training, + } + + ft = self.freeze_finetune_updates <= self.num_updates + + with torch.no_grad() if not ft else contextlib.ExitStack(): + res = self.w2v_model.extract_features(**w2v_args) + + x = res["x"] + padding_mask = res["padding_mask"] + + if tbc: + # BTC -> TBC + x = x.transpose(0, 1) + + x = self.final_dropout(x) + + if self.proj: + x = self.proj(x) + + return { + "encoder_out": x, # T x B x C + "encoder_padding_mask": padding_mask.transpose(0, 1) + if padding_mask is not None + else None, # T x B + "padding_mask": padding_mask, + "layer_results": res["layer_results"], + } + + def reorder_encoder_out(self, encoder_out, new_order): + if encoder_out["encoder_out"] is not None: + encoder_out["encoder_out"] = encoder_out["encoder_out"].index_select( + 1, new_order + ) + if encoder_out["encoder_padding_mask"] is not None: + encoder_out["encoder_padding_mask"] = encoder_out[ + "encoder_padding_mask" + ].index_select(0, new_order) + return encoder_out + + def max_positions(self): + """Maximum input length supported by the encoder.""" + return None + + def upgrade_state_dict_named(self, state_dict, name): + return state_dict + + +class TransformerDecoder(FairseqIncrementalDecoder): + """ + Transformer decoder consisting of *args.decoder_layers* layers. Each layer + is a :class:`TransformerDecoderLayer`. + + Args: + args (argparse.Namespace): parsed command-line arguments + dictionary (~fairseq.data.Dictionary): decoding dictionary + embed_tokens (torch.nn.Embedding): output embedding + no_encoder_attn (bool, optional): whether to attend to encoder outputs + (default: False). + """ + + def __init__( + self, + cfg: Wav2Vec2Seq2SeqConfig, + dictionary, + embed_tokens, + no_encoder_attn=False, + ): + super().__init__(dictionary) + + self.dropout = cfg.decoder_dropout + self.share_input_output_embed = cfg.share_decoder_input_output_embed + + input_embed_dim = embed_tokens.embedding_dim + embed_dim = cfg.decoder_embed_dim + self.output_embed_dim = cfg.decoder_embed_dim + + self.layerdrop = cfg.decoder_layerdrop + + padding_idx = embed_tokens.padding_idx + self.max_target_positions = cfg.max_target_positions + + self.embed_tokens = embed_tokens + self.embed_scale = math.sqrt(embed_dim) # todo: try with input_embed_dim + + self.project_in_dim = ( + Linear(input_embed_dim, embed_dim, bias=False) + if embed_dim != input_embed_dim + else None + ) + + self.embed_positions = ( + PositionalEmbedding( + cfg.max_target_positions, + embed_dim, + padding_idx, + learned=cfg.decoder_learned_pos, + ) + if not cfg.no_token_positional_embeddings + else None + ) + + # TODO: update this when transformer gets converted to dataclass configs + transformer_cfg = copy.deepcopy(cfg) + with open_dict(transformer_cfg): + transformer_cfg.dropout = transformer_cfg.decoder_dropout + transformer_cfg.attention_dropout = ( + transformer_cfg.decoder_attention_dropout + ) + transformer_cfg.activation_dropout = ( + transformer_cfg.decoder_activation_dropout + ) + + self.layers = nn.ModuleList([]) + self.layers.extend( + [ + TransformerDecoderLayer(transformer_cfg, no_encoder_attn) + for _ in range(transformer_cfg.decoder_layers) + ] + ) + + if not self.share_input_output_embed: + self.embed_out = nn.Parameter( + torch.Tensor(len(dictionary), self.output_embed_dim) + ) + nn.init.normal_(self.embed_out, mean=0, std=self.output_embed_dim ** -0.5) + + if transformer_cfg.decoder_normalize_before: + self.layer_norm = LayerNorm(embed_dim) + else: + self.layer_norm = None + + def forward( + self, prev_output_tokens, encoder_out=None, incremental_state=None, **unused + ): + """ + Args: + prev_output_tokens (LongTensor): previous decoder outputs of shape + `(batch, tgt_len)`, for teacher forcing + encoder_out (Tensor, optional): output from the encoder, used for + encoder-side attention + incremental_state (dict): dictionary used for storing state during + :ref:`Incremental decoding` + + Returns: + tuple: + - the decoder's output of shape `(batch, tgt_len, vocab)` + - a dictionary with any model-specific outputs + """ + prev_output_tokens = prev_output_tokens.long() + x, extra = self.extract_features( + prev_output_tokens, encoder_out, incremental_state + ) + x = self.output_layer(x) + return x, extra + + def extract_features( + self, prev_output_tokens, encoder_out=None, incremental_state=None, **unused + ): + """ + Similar to *forward* but only return features. + + Returns: + tuple: + - the decoder's features of shape `(batch, tgt_len, embed_dim)` + - a dictionary with any model-specific outputs + """ + + # embed positions + positions = ( + self.embed_positions( + prev_output_tokens, incremental_state=incremental_state + ) + if self.embed_positions is not None + else None + ) + + if incremental_state is not None: + prev_output_tokens = prev_output_tokens[:, -1:] + if positions is not None: + positions = positions[:, -1:] + + # embed tokens and positions + x = self.embed_scale * self.embed_tokens(prev_output_tokens) + + if self.project_in_dim is not None: + x = self.project_in_dim(x) + + if positions is not None: + x += positions + x = F.dropout(x, p=self.dropout, training=self.training) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + attn = None + + inner_states = [x] + + # decoder layers + for layer in self.layers: + dropout_probability = np.random.random() + if not self.training or (dropout_probability > self.layerdrop): + x, attn, _ = layer( + x, + encoder_out["encoder_out"] if encoder_out is not None else None, + encoder_out["padding_mask"] if encoder_out is not None else None, + incremental_state, + self_attn_mask=self.buffered_future_mask(x) + if incremental_state is None + else None, + ) + inner_states.append(x) + + if self.layer_norm: + x = self.layer_norm(x) + + # T x B x C -> B x T x C + x = x.transpose(0, 1) + + return x, {"attn": attn, "inner_states": inner_states} + + def output_layer(self, features, **kwargs): + """Project features to the vocabulary size.""" + # project back to size of vocabulary + if self.share_input_output_embed: + return F.linear(features, self.embed_tokens.weight) + else: + return F.linear(features, self.embed_out) + + def max_positions(self): + """Maximum output length supported by the decoder.""" + if self.embed_positions is None: + return self.max_target_positions + return min(self.max_target_positions, self.embed_positions.max_positions) + + def buffered_future_mask(self, tensor): + dim = tensor.size(0) + if ( + not hasattr(self, "_future_mask") + or self._future_mask is None + or self._future_mask.device != tensor.device + or self._future_mask.size(0) < dim + ): + self._future_mask = torch.triu( + utils.fill_with_neg_inf(tensor.new(dim, dim)), 1 + ) + return self._future_mask[:dim, :dim] + + def upgrade_state_dict_named(self, state_dict, name): + return state_dict + + +def Embedding(num_embeddings, embedding_dim, padding_idx): + m = nn.Embedding(num_embeddings, embedding_dim, padding_idx=padding_idx) + nn.init.normal_(m.weight, mean=0, std=embedding_dim ** -0.5) + nn.init.constant_(m.weight[padding_idx], 0) + return m + + +def Linear(in_features, out_features, bias=True): + m = nn.Linear(in_features, out_features, bias) + nn.init.xavier_uniform_(m.weight) + if bias: + nn.init.constant_(m.bias, 0.0) + return m diff --git a/fairseq/modules/__init__.py b/fairseq/modules/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..81930aa71c00ab8a6c36e362e8de3d356d0cf30a --- /dev/null +++ b/fairseq/modules/__init__.py @@ -0,0 +1,78 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +"""isort:skip_file""" + +from .adaptive_input import AdaptiveInput +from .adaptive_softmax import AdaptiveSoftmax +from .base_layer import BaseLayer +from .beamable_mm import BeamableMM +from .character_token_embedder import CharacterTokenEmbedder +from .conv_tbc import ConvTBC +from .cross_entropy import cross_entropy +from .downsampled_multihead_attention import DownsampledMultiHeadAttention +from .dynamic_convolution import DynamicConv, DynamicConv1dTBC +from .dynamic_crf_layer import DynamicCRF +from .fairseq_dropout import FairseqDropout +from .fp32_group_norm import Fp32GroupNorm +from .gelu import gelu, gelu_accurate +from .grad_multiply import GradMultiply +from .gumbel_vector_quantizer import GumbelVectorQuantizer +from .kmeans_vector_quantizer import KmeansVectorQuantizer +from .layer_drop import LayerDropModuleList +from .layer_norm import Fp32LayerNorm, LayerNorm +from .learned_positional_embedding import LearnedPositionalEmbedding +from .lightweight_convolution import LightweightConv, LightweightConv1dTBC +from .linearized_convolution import LinearizedConvolution +from .multihead_attention import MultiheadAttention +from .positional_embedding import PositionalEmbedding +from .same_pad import SamePad +from .scalar_bias import ScalarBias +from .sinusoidal_positional_embedding import SinusoidalPositionalEmbedding +from .transformer_sentence_encoder_layer import TransformerSentenceEncoderLayer +from .transformer_sentence_encoder import TransformerSentenceEncoder +from .transpose_last import TransposeLast +from .unfold import unfold1d +from .transformer_layer import TransformerDecoderLayer, TransformerEncoderLayer +from .vggblock import VGGBlock + +__all__ = [ + "AdaptiveInput", + "AdaptiveSoftmax", + "BaseLayer", + "BeamableMM", + "CharacterTokenEmbedder", + "ConvTBC", + "cross_entropy", + "DownsampledMultiHeadAttention", + "DynamicConv1dTBC", + "DynamicConv", + "DynamicCRF", + "FairseqDropout", + "Fp32GroupNorm", + "Fp32LayerNorm", + "gelu", + "gelu_accurate", + "GradMultiply", + "GumbelVectorQuantizer", + "KmeansVectorQuantizer", + "LayerDropModuleList", + "LayerNorm", + "LearnedPositionalEmbedding", + "LightweightConv1dTBC", + "LightweightConv", + "LinearizedConvolution", + "MultiheadAttention", + "PositionalEmbedding", + "SamePad", + "ScalarBias", + "SinusoidalPositionalEmbedding", + "TransformerSentenceEncoderLayer", + "TransformerSentenceEncoder", + "TransformerDecoderLayer", + "TransformerEncoderLayer", + "TransposeLast", + "VGGBlock", + "unfold1d", +] diff --git a/fairseq/modules/adaptive_input.py b/fairseq/modules/adaptive_input.py new file mode 100644 index 0000000000000000000000000000000000000000..446534a9f8b87337a4dd752944ea386ff7cf7965 --- /dev/null +++ b/fairseq/modules/adaptive_input.py @@ -0,0 +1,80 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +from typing import List + +import torch +from fairseq.modules.quant_noise import quant_noise +from torch import nn + + +class AdaptiveInput(nn.Module): + def __init__( + self, + vocab_size: int, + padding_idx: int, + initial_dim: int, + factor: float, + output_dim: int, + cutoff: List[int], + q_noise: float = 0, + qn_block_size: int = 8, + ): + super().__init__() + + if vocab_size > cutoff[-1]: + cutoff = cutoff + [vocab_size] + else: + assert ( + vocab_size == cutoff[-1] + ), "cannot specify cutoff larger than vocab size" + + self.cutoff = cutoff + self.embedding_dim = output_dim + self.padding_idx = padding_idx + + self.embeddings = nn.ModuleList() + for i in range(len(self.cutoff)): + prev = self.cutoff[i - 1] if i > 0 else 0 + size = self.cutoff[i] - prev + dim = int(initial_dim // (factor ** i)) + seq = nn.Sequential( + nn.Embedding(size, dim, self.padding_idx), + quant_noise( + nn.Linear(dim, output_dim, bias=False), q_noise, qn_block_size + ), + ) + + self.embeddings.append(seq) + self.padding_idx = None + self.padding_idx = padding_idx + + def init_weights(m): + if isinstance(m, nn.Embedding): + nn.init.normal_(m.weight, mean=0, std=m.weight.shape[1] ** -0.5) + nn.init.constant_(m.weight[padding_idx], 0) + elif hasattr(m, "weight"): + nn.init.xavier_uniform_(m.weight) + + self.apply(init_weights) + + self.register_buffer("_float_tensor", torch.FloatTensor(1)) + + def weights_for_band(self, band: int): + return self.embeddings[band][0].weight, self.embeddings[band][1].weight + + def forward(self, input: torch.Tensor): + result = self._float_tensor.new(input.shape + (self.embedding_dim,)) + for i in range(len(self.cutoff)): + mask = input.lt(self.cutoff[i]) + if i > 0: + mask.mul_(input.ge(self.cutoff[i - 1])) + chunk_input = input[mask] - self.cutoff[i - 1] + else: + chunk_input = input[mask] + if mask.any(): + result[mask] = self.embeddings[i](chunk_input) + return result diff --git a/fairseq/modules/adaptive_softmax.py b/fairseq/modules/adaptive_softmax.py new file mode 100644 index 0000000000000000000000000000000000000000..ae0c77ba0f6ee98501306d66cbc4a948b4ade0f7 --- /dev/null +++ b/fairseq/modules/adaptive_softmax.py @@ -0,0 +1,268 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import functools +import operator + +import torch +import torch.nn.functional as F +from fairseq.modules.fairseq_dropout import FairseqDropout +from fairseq.modules.quant_noise import quant_noise +from torch import nn + + +class TiedLinear(nn.Module): + def __init__(self, weight, transpose): + super().__init__() + self.weight = weight + self.transpose = transpose + + def forward(self, input): + return F.linear(input, self.weight.t() if self.transpose else self.weight) + + +class TiedHeadModule(nn.Module): + def __init__(self, weights, input_dim, num_classes, q_noise, qn_block_size): + super().__init__() + tied_emb, _ = weights + self.num_words, emb_dim = tied_emb.size() + + self.word_proj = quant_noise( + TiedLinear(tied_emb, transpose=False), q_noise, qn_block_size + ) + if input_dim != emb_dim: + self.word_proj = nn.Sequential( + quant_noise( + nn.Linear(input_dim, emb_dim, bias=False), q_noise, qn_block_size + ), + self.word_proj, + ) + + self.class_proj = quant_noise( + nn.Linear(input_dim, num_classes, bias=False), q_noise, qn_block_size + ) + self.out_dim = self.num_words + num_classes + + self.register_buffer("_float_tensor", torch.FloatTensor(1)) + + def forward(self, input): + inp_sz = functools.reduce(operator.mul, input.shape[:-1], 1) + out = self._float_tensor.new(inp_sz, self.out_dim) + out[:, : self.num_words] = self.word_proj(input.view(inp_sz, -1)) + out[:, self.num_words :] = self.class_proj(input.view(inp_sz, -1)) + return out + + +class AdaptiveSoftmax(nn.Module): + """ + This is an implementation of the efficient softmax approximation for + graphical processing units (GPU), described in the paper "Efficient softmax + approximation for GPUs" (http://arxiv.org/abs/1609.04309). + """ + + def __init__( + self, + vocab_size, + input_dim, + cutoff, + dropout, + factor=4.0, + adaptive_inputs=None, + tie_proj=False, + q_noise=0, + qn_block_size=8, + ): + super().__init__() + + if vocab_size > cutoff[-1]: + cutoff = cutoff + [vocab_size] + else: + assert ( + vocab_size == cutoff[-1] + ), "cannot specify cutoff larger than vocab size" + + output_dim = cutoff[0] + len(cutoff) - 1 + + self.vocab_size = vocab_size + self.cutoff = cutoff + self.dropout_module = FairseqDropout( + dropout, module_name=self.__class__.__name__ + ) + self.input_dim = input_dim + self.factor = factor + self.q_noise = q_noise + self.qn_block_size = qn_block_size + + self.lsm = nn.LogSoftmax(dim=1) + + if adaptive_inputs is not None: + self.head = TiedHeadModule( + adaptive_inputs.weights_for_band(0), + input_dim, + len(cutoff) - 1, + self.q_noise, + self.qn_block_size, + ) + else: + self.head = quant_noise( + nn.Linear(input_dim, output_dim, bias=False), + self.q_noise, + self.qn_block_size, + ) + + self._make_tail(adaptive_inputs, tie_proj) + + def init_weights(m): + if ( + hasattr(m, "weight") + and not isinstance(m, TiedLinear) + and not isinstance(m, TiedHeadModule) + ): + nn.init.xavier_uniform_(m.weight) + + self.apply(init_weights) + + self.register_buffer("version", torch.LongTensor([1])) + + def _make_tail(self, adaptive_inputs=None, tie_proj=False): + self.tail = nn.ModuleList() + for i in range(len(self.cutoff) - 1): + dim = int(self.input_dim // self.factor ** (i + 1)) + + tied_emb, tied_proj = ( + adaptive_inputs.weights_for_band(i + 1) + if adaptive_inputs is not None + else (None, None) + ) + + if tied_proj is not None: + if tie_proj: + proj = quant_noise( + TiedLinear(tied_proj, transpose=True), + self.q_noise, + self.qn_block_size, + ) + else: + proj = quant_noise( + nn.Linear(tied_proj.size(0), tied_proj.size(1), bias=False), + self.q_noise, + self.qn_block_size, + ) + else: + proj = quant_noise( + nn.Linear(self.input_dim, dim, bias=False), + self.q_noise, + self.qn_block_size, + ) + + if tied_emb is None: + out_proj = nn.Linear( + dim, self.cutoff[i + 1] - self.cutoff[i], bias=False + ) + else: + out_proj = TiedLinear(tied_emb, transpose=False) + + m = nn.Sequential( + proj, + nn.Dropout(self.dropout_module.p), + quant_noise(out_proj, self.q_noise, self.qn_block_size), + ) + + self.tail.append(m) + + def upgrade_state_dict_named(self, state_dict, name): + version_name = name + ".version" + if version_name not in state_dict: + raise Exception("This version of the model is no longer supported") + + def adapt_target(self, target): + """ + In order to be efficient, the AdaptiveSoftMax does not compute the + scores for all the word of the vocabulary for all the examples. It is + thus necessary to call the method adapt_target of the AdaptiveSoftMax + layer inside each forward pass. + """ + + target = target.view(-1) + new_target = [target.clone()] + target_idxs = [] + + for i in range(len(self.cutoff) - 1): + mask = target.ge(self.cutoff[i]).mul(target.lt(self.cutoff[i + 1])) + new_target[0][mask] = self.cutoff[0] + i + + if mask.any(): + target_idxs.append(mask.nonzero(as_tuple=False).squeeze(1)) + new_target.append(target[mask].add(-self.cutoff[i])) + else: + target_idxs.append(None) + new_target.append(None) + + return new_target, target_idxs + + def forward(self, input, target): + """ + Args: + input: (b x t x d) + target: (b x t) + Returns: + 2 lists: output for each cutoff section and new targets by cut off + """ + + input = input.contiguous().view(-1, input.size(-1)) + input = self.dropout_module(input) + + new_target, target_idxs = self.adapt_target(target) + output = [self.head(input)] + + for i in range(len(target_idxs)): + if target_idxs[i] is not None: + output.append(self.tail[i](input.index_select(0, target_idxs[i]))) + else: + output.append(None) + + return output, new_target + + def get_log_prob(self, input, target): + """ + Computes the log probabilities for all the words of the vocabulary, + given a 2D tensor of hidden vectors. + """ + + bsz, length, dim = input.size() + input = input.contiguous().view(-1, dim) + + if target is not None: + _, target_idxs = self.adapt_target(target) + else: + target_idxs = None + + head_y = self.head(input) + log_probs = head_y.new_zeros(input.size(0), self.vocab_size) + + head_sz = self.cutoff[0] + len(self.tail) + log_probs[:, :head_sz] = self.lsm(head_y) + tail_priors = log_probs[:, self.cutoff[0] : head_sz].clone() + + for i in range(len(self.tail)): + start = self.cutoff[i] + end = self.cutoff[i + 1] + + if target_idxs is None: + tail_out = log_probs[:, start:end] + tail_out.copy_(self.tail[i](input)) + log_probs[:, start:end] = self.lsm(tail_out).add_( + tail_priors[:, i, None] + ) + elif target_idxs[i] is not None: + idxs = target_idxs[i] + tail_out = log_probs[idxs, start:end] + tail_out.copy_(self.tail[i](input[idxs])) + log_probs[idxs, start:end] = self.lsm(tail_out).add_( + tail_priors[idxs, i, None] + ) + + log_probs = log_probs.view(bsz, length, -1) + return log_probs diff --git a/fairseq/modules/base_layer.py b/fairseq/modules/base_layer.py new file mode 100644 index 0000000000000000000000000000000000000000..e7ef155b25fc73e74780879f665288c9bc95fd80 --- /dev/null +++ b/fairseq/modules/base_layer.py @@ -0,0 +1,135 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch.nn as nn +import torch +import sys +from fairseq import utils +from fairseq.distributed import utils as distributed_utils +from fairseq.modules.layer_norm import LayerNorm + + +class BaseLayer(nn.Module): + + def __init__(self, args): + super().__init__() + self.num_workers = distributed_utils.get_data_parallel_world_size() + expert_centroids = torch.empty(self.num_workers, args.decoder_embed_dim) + torch.nn.init.orthogonal_(expert_centroids, gain=0.1) + self.register_parameter("expert_centroids", torch.nn.Parameter(expert_centroids)) + self.expert_network = nn.Sequential(*([BaseSublayer(args) for _ in range(args.base_sublayers)])) + self.expert_id = distributed_utils.get_data_parallel_rank() + self.shuffle = args.base_shuffle + self.cpp = self.load_assignment() + + # Add a special attribute to the expert parameters, so we know not to sync their gradients + for param in self.expert_network.parameters(): + param.expert = True + + def forward(self, input_features, *args, **kwargs): + features = input_features.reshape(-1, input_features.size(-1)) + is_training = input_features.requires_grad + + if self.shuffle and is_training: + # Send each token to a random worker, to break correlations within the batch + shuffle_sort = torch.randperm(features.size(0), device=features.device) + features = All2All.apply(features[shuffle_sort]) + + with torch.no_grad(): + # Compute similarity of each token to each expert, for routing + token_expert_affinities = features.matmul(self.expert_centroids.transpose(0, 1)) + + # Compute which token goes to which expert + sort_by_expert, input_splits, output_splits = self.balanced_assignment(token_expert_affinities) \ + if is_training else self.greedy_assignment(token_expert_affinities) + # Swap these tokens for the right ones for our expert + routed_features = All2All.apply(features[sort_by_expert], output_splits, input_splits) + + if routed_features.size(0) > 0: + # Mix in the expert network based on how appropriate it is for these tokens + alpha = torch.sigmoid(routed_features.mv(self.expert_centroids[self.expert_id])).unsqueeze(1) + routed_features = alpha * self.expert_network(routed_features) + (1 - alpha) * routed_features + # Return to original worker and ordering + result = All2All.apply(routed_features, input_splits, output_splits)[self.inverse_sort(sort_by_expert)] + + if self.shuffle and is_training: + # Undo shuffling + result = All2All.apply(result)[self.inverse_sort(shuffle_sort)] + + # Return additional Nones for compatibility with TransformerDecoderLayer + return result.view(input_features.size()), None, None + + def inverse_sort(self, order): + # Creates an index that undoes a sort: xs==xs[order][inverse_sort(order)] + return torch.empty_like(order).scatter_(0, order, torch.arange(0, order.size(0), device=order.device)) + + def balanced_assignment(self, scores): + ok = scores.isfinite() + if not ok.all(): + # NaNs here can break the assignment algorithm + scores[~ok] = scores[ok].min() + return self.cpp.balanced_assignment(scores), None, None + + # Assigns each token to the top k experts + def greedy_assignment(self, scores, k=1): + token_to_workers = torch.topk(scores, dim=1, k=k, largest=True).indices.view(-1) + token_to_workers, sort_ordering = torch.sort(token_to_workers) + worker2token = sort_ordering // k + + # Find how many tokens we're sending to each other worker (being careful for sending 0 tokens to some workers) + output_splits = torch.zeros((self.num_workers,), dtype=torch.long, device=scores.device) + workers, counts = torch.unique_consecutive(token_to_workers, return_counts=True) + output_splits[workers] = counts + # Tell other workers how many tokens to expect from us + input_splits = All2All.apply(output_splits) + return worker2token, input_splits.tolist(), output_splits.tolist() + + def load_assignment(self): + try: + from fairseq import libbase + + return libbase + + except ImportError as e: + sys.stderr.write( + "ERROR: missing libbase. run `python setup.py build_ext --inplace`\n" + ) + raise e + + +class BaseSublayer(nn.Module): + def __init__(self, args): + super().__init__() + self.activation_fn = utils.get_activation_fn( + activation=getattr(args, 'activation_fn', 'relu') or "relu" + ) + self.norm = LayerNorm(args.decoder_embed_dim, export=False) + self.ff1 = torch.nn.Linear(args.decoder_embed_dim, args.decoder_ffn_embed_dim) + self.ff2 = torch.nn.Linear(args.decoder_ffn_embed_dim, args.decoder_embed_dim) + self.ff2.weight.data.zero_() + + def forward(self, xs): + return xs + self.ff2(self.activation_fn(self.ff1(self.norm(xs)))) + + +# Wraps torch.distributed.all_to_all_single as a function that supports autograd +class All2All(torch.autograd.Function): + @staticmethod + def forward(ctx, xs, input_splits=None, output_splits=None): + ctx.input_splits = input_splits + ctx.output_splits = output_splits + + ys = torch.empty_like(xs) if output_splits is None else \ + xs.new_empty(size=[sum(output_splits)] + list(xs.size()[1:])) + torch.distributed.all_to_all_single(ys, xs, output_split_sizes=output_splits, input_split_sizes=input_splits) + return ys + + @staticmethod + def backward(ctx, grad_output): + result = torch.empty_like(grad_output) if ctx.input_splits is None else \ + grad_output.new_empty(size=[sum(ctx.input_splits)] + list(grad_output.size()[1:])) + torch.distributed.all_to_all_single(result, grad_output, + output_split_sizes=ctx.input_splits, input_split_sizes=ctx.output_splits) + return result, None, None diff --git a/fairseq/modules/beamable_mm.py b/fairseq/modules/beamable_mm.py new file mode 100644 index 0000000000000000000000000000000000000000..eff1a4607f600c71210e6b914985dc48731aae86 --- /dev/null +++ b/fairseq/modules/beamable_mm.py @@ -0,0 +1,49 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn as nn + + +class BeamableMM(nn.Module): + """This module provides an optimized MM for beam decoding with attention. + + It leverage the fact that the source-side of the input is replicated beam + times and the target-side of the input is of width one. This layer speeds up + inference by replacing the inputs {(bsz x 1 x nhu), (bsz x sz2 x nhu)} + with smaller inputs {(bsz/beam x beam x nhu), (bsz/beam x sz2 x nhu)}. + """ + + def __init__(self, beam_size=None): + super(BeamableMM, self).__init__() + self.beam_size = beam_size + + def forward(self, input1, input2): + if ( + not self.training + and self.beam_size is not None # test mode + and input1.dim() == 3 # beam size is set + and input1.size(1) # only support batched input + == 1 # single time step update + ): + bsz, beam = input1.size(0), self.beam_size + + # bsz x 1 x nhu --> bsz/beam x beam x nhu + input1 = input1[:, 0, :].unfold(0, beam, beam).transpose(2, 1) + + # bsz x sz2 x nhu --> bsz/beam x sz2 x nhu + input2 = input2.unfold(0, beam, beam)[:, :, :, 0] + + # use non batched operation if bsz = beam + if input1.size(0) == 1: + output = torch.mm(input1[0, :, :], input2[0, :, :]) + else: + output = input1.bmm(input2) + return output.view(bsz, 1, -1) + else: + return input1.bmm(input2) + + def set_beam_size(self, beam_size): + self.beam_size = beam_size diff --git a/fairseq/modules/character_token_embedder.py b/fairseq/modules/character_token_embedder.py new file mode 100644 index 0000000000000000000000000000000000000000..181221b61b9f76453b67e3b848b198620dce912c --- /dev/null +++ b/fairseq/modules/character_token_embedder.py @@ -0,0 +1,214 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +from typing import List, Tuple + +import torch +import torch.nn.functional as F +from fairseq.data import Dictionary +from torch import nn + + +CHAR_PAD_IDX = 0 +CHAR_EOS_IDX = 257 + + +logger = logging.getLogger(__name__) + + +class CharacterTokenEmbedder(torch.nn.Module): + def __init__( + self, + vocab: Dictionary, + filters: List[Tuple[int, int]], + char_embed_dim: int, + word_embed_dim: int, + highway_layers: int, + max_char_len: int = 50, + char_inputs: bool = False, + ): + super(CharacterTokenEmbedder, self).__init__() + + self.onnx_trace = False + self.embedding_dim = word_embed_dim + self.max_char_len = max_char_len + self.char_embeddings = nn.Embedding(257, char_embed_dim, padding_idx=0) + self.symbol_embeddings = nn.Parameter(torch.FloatTensor(2, word_embed_dim)) + self.eos_idx, self.unk_idx = 0, 1 + self.char_inputs = char_inputs + + self.convolutions = nn.ModuleList() + for width, out_c in filters: + self.convolutions.append( + nn.Conv1d(char_embed_dim, out_c, kernel_size=width) + ) + + last_dim = sum(f[1] for f in filters) + + self.highway = Highway(last_dim, highway_layers) if highway_layers > 0 else None + + self.projection = nn.Linear(last_dim, word_embed_dim) + + assert ( + vocab is not None or char_inputs + ), "vocab must be set if not using char inputs" + self.vocab = None + if vocab is not None: + self.set_vocab(vocab, max_char_len) + + self.reset_parameters() + + def prepare_for_onnx_export_(self): + self.onnx_trace = True + + def set_vocab(self, vocab, max_char_len): + word_to_char = torch.LongTensor(len(vocab), max_char_len) + + truncated = 0 + for i in range(len(vocab)): + if i < vocab.nspecial: + char_idxs = [0] * max_char_len + else: + chars = vocab[i].encode() + # +1 for padding + char_idxs = [c + 1 for c in chars] + [0] * (max_char_len - len(chars)) + if len(char_idxs) > max_char_len: + truncated += 1 + char_idxs = char_idxs[:max_char_len] + word_to_char[i] = torch.LongTensor(char_idxs) + + if truncated > 0: + logger.info( + "truncated {} words longer than {} characters".format( + truncated, max_char_len + ) + ) + + self.vocab = vocab + self.word_to_char = word_to_char + + @property + def padding_idx(self): + return Dictionary().pad() if self.vocab is None else self.vocab.pad() + + def reset_parameters(self): + nn.init.xavier_normal_(self.char_embeddings.weight) + nn.init.xavier_normal_(self.symbol_embeddings) + nn.init.xavier_uniform_(self.projection.weight) + + nn.init.constant_( + self.char_embeddings.weight[self.char_embeddings.padding_idx], 0.0 + ) + nn.init.constant_(self.projection.bias, 0.0) + + def forward( + self, + input: torch.Tensor, + ): + if self.char_inputs: + chars = input.view(-1, self.max_char_len) + pads = chars[:, 0].eq(CHAR_PAD_IDX) + eos = chars[:, 0].eq(CHAR_EOS_IDX) + if eos.any(): + if self.onnx_trace: + chars = torch.where(eos.unsqueeze(1), chars.new_zeros(1), chars) + else: + chars[eos] = 0 + + unk = None + else: + flat_words = input.view(-1) + chars = self.word_to_char[flat_words.type_as(self.word_to_char)].type_as( + input + ) + pads = flat_words.eq(self.vocab.pad()) + eos = flat_words.eq(self.vocab.eos()) + unk = flat_words.eq(self.vocab.unk()) + + word_embs = self._convolve(chars) + if self.onnx_trace: + if pads.any(): + word_embs = torch.where( + pads.unsqueeze(1), word_embs.new_zeros(1), word_embs + ) + if eos.any(): + word_embs = torch.where( + eos.unsqueeze(1), self.symbol_embeddings[self.eos_idx], word_embs + ) + if unk is not None and unk.any(): + word_embs = torch.where( + unk.unsqueeze(1), self.symbol_embeddings[self.unk_idx], word_embs + ) + else: + if pads.any(): + word_embs[pads] = 0 + if eos.any(): + word_embs[eos] = self.symbol_embeddings[self.eos_idx] + if unk is not None and unk.any(): + word_embs[unk] = self.symbol_embeddings[self.unk_idx] + + return word_embs.view(input.size()[:2] + (-1,)) + + def _convolve( + self, + char_idxs: torch.Tensor, + ): + char_embs = self.char_embeddings(char_idxs) + char_embs = char_embs.transpose(1, 2) # BTC -> BCT + + conv_result = [] + + for conv in self.convolutions: + x = conv(char_embs) + x, _ = torch.max(x, -1) + x = F.relu(x) + conv_result.append(x) + + x = torch.cat(conv_result, dim=-1) + + if self.highway is not None: + x = self.highway(x) + x = self.projection(x) + + return x + + +class Highway(torch.nn.Module): + """ + A `Highway layer <https://arxiv.org/abs/1505.00387>`_. + Adopted from the AllenNLP implementation. + """ + + def __init__(self, input_dim: int, num_layers: int = 1): + super(Highway, self).__init__() + self.input_dim = input_dim + self.layers = nn.ModuleList( + [nn.Linear(input_dim, input_dim * 2) for _ in range(num_layers)] + ) + self.activation = nn.ReLU() + + self.reset_parameters() + + def reset_parameters(self): + for layer in self.layers: + # As per comment in AllenNLP: + # We should bias the highway layer to just carry its input forward. We do that by + # setting the bias on `B(x)` to be positive, because that means `g` will be biased to + # be high, so we will carry the input forward. The bias on `B(x)` is the second half + # of the bias vector in each Linear layer. + nn.init.constant_(layer.bias[self.input_dim :], 1) + + nn.init.constant_(layer.bias[: self.input_dim], 0) + nn.init.xavier_normal_(layer.weight) + + def forward(self, x: torch.Tensor): + for layer in self.layers: + projection = layer(x) + proj_x, gate = projection.chunk(2, dim=-1) + proj_x = self.activation(proj_x) + gate = torch.sigmoid(gate) + x = gate * x + (gate.new_tensor([1]) - gate) * proj_x + return x diff --git a/fairseq/modules/checkpoint_activations.py b/fairseq/modules/checkpoint_activations.py new file mode 100644 index 0000000000000000000000000000000000000000..b44fc346cec1ab24d8056075b3df14020a86214b --- /dev/null +++ b/fairseq/modules/checkpoint_activations.py @@ -0,0 +1,236 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import functools +from typing import Any, Dict, List, Tuple, Union + +import torch +import torch.utils.checkpoint as checkpoint +from fairseq import utils + + +def checkpoint_wrapper(m, offload_to_cpu=False): + """ + A friendlier wrapper for performing activation checkpointing. + + Compared to the PyTorch version, this version: + - wraps an nn.Module, so that all subsequent calls will use checkpointing + - handles keyword arguments in the forward + - handles non-Tensor outputs from the forward + + Usage:: + + checkpointed_module = checkpoint_wrapper(my_module, offload_to_cpu=True) + a, b = checkpointed_module(x, y=3, z=torch.Tensor([1])) + """ + # should I check whether original_forward has already been set? + assert not hasattr( + m, "precheckpoint_forward" + ), "checkpoint function has already been applied?" + m.precheckpoint_forward = m.forward + m.forward = functools.partial( + _checkpointed_forward, + m.precheckpoint_forward, # original_forward + offload_to_cpu, + ) + return m + + +def unwrap_checkpoint(m: torch.nn.Module): + """ + unwrap a module and its children from checkpoint_wrapper + """ + for module in m.modules(): + if hasattr(module, "precheckpoint_forward"): + module.forward = module.precheckpoint_forward + del module.precheckpoint_forward + return m + + +def _checkpointed_forward(original_forward, offload_to_cpu, *args, **kwargs): + # Autograd Functions in PyTorch work best with positional args, since + # the backward must return gradients (or None) for every input argument. + # We can flatten keyword arguments to make this easier. + kwarg_keys, flat_args = pack_kwargs(*args, **kwargs) + parent_ctx_dict = {"offload": offload_to_cpu} + output = CheckpointFunction.apply( + original_forward, parent_ctx_dict, kwarg_keys, *flat_args + ) + if isinstance(output, torch.Tensor): + return output + else: + packed_non_tensor_outputs = parent_ctx_dict["packed_non_tensor_outputs"] + if packed_non_tensor_outputs: + output = unpack_non_tensors(output, packed_non_tensor_outputs) + return output + + +def pack_kwargs(*args, **kwargs) -> Tuple[List[str], List[Any]]: + """ + Usage:: + + kwarg_keys, flat_args = pack_kwargs(1, 2, a=3, b=4) + args, kwargs = unpack_kwargs(kwarg_keys, flat_args) + assert args == [1, 2] + assert kwargs == {"a": 3, "b": 4} + """ + kwarg_keys = [] + flat_args = list(args) + for k, v in kwargs.items(): + kwarg_keys.append(k) + flat_args.append(v) + return kwarg_keys, flat_args + + +def unpack_kwargs( + kwarg_keys: List[str], flat_args: List[Any] +) -> Tuple[List[Any], Dict[str, Any]]: + if len(kwarg_keys) == 0: + return flat_args, {} + args = flat_args[: -len(kwarg_keys)] + kwargs = {k: v for k, v in zip(kwarg_keys, flat_args[-len(kwarg_keys) :])} + return args, kwargs + + +def split_non_tensors( + mixed: Union[torch.Tensor, Tuple[Any]] +) -> Tuple[Tuple[torch.Tensor], Dict[str, List[Any]]]: + """ + Usage:: + + x = torch.Tensor([1]) + y = torch.Tensor([2]) + tensors, packed_non_tensors = split_non_tensors((x, y, None, 3)) + recon = unpack_non_tensors(tensors, packed_non_tensors) + assert recon == (x, y, None, 3) + """ + if isinstance(mixed, torch.Tensor): + return (mixed,), None + tensors = [] + packed_non_tensors = {"is_tensor": [], "objects": []} + for o in mixed: + if isinstance(o, torch.Tensor): + packed_non_tensors["is_tensor"].append(True) + tensors.append(o) + else: + packed_non_tensors["is_tensor"].append(False) + packed_non_tensors["objects"].append(o) + return tuple(tensors), packed_non_tensors + + +def unpack_non_tensors( + tensors: Tuple[torch.Tensor], + packed_non_tensors: Dict[str, List[Any]], +) -> Tuple[Any]: + if packed_non_tensors is None: + return tensors + assert isinstance(packed_non_tensors, dict) + mixed = [] + is_tensor_list = packed_non_tensors["is_tensor"] + objects = packed_non_tensors["objects"] + assert len(tensors) + len(objects) == len(is_tensor_list) + obj_i = tnsr_i = 0 + for is_tensor in is_tensor_list: + if is_tensor: + mixed.append(tensors[tnsr_i]) + tnsr_i += 1 + else: + mixed.append(objects[obj_i]) + obj_i += 1 + return tuple(mixed) + + +class CheckpointFunction(torch.autograd.Function): + """Similar to the torch version, but support non-Tensor outputs. + + The caller is expected to provide a dict (*parent_ctx_dict*) that will hold + the non-Tensor outputs. These should be combined with the Tensor *outputs* + by calling ``unpack_non_tensors``. + """ + + @staticmethod + def forward(ctx, run_function, parent_ctx_dict, kwarg_keys, *args): + if torch.is_grad_enabled(): # grad may be disabled, e.g., during validation + checkpoint.check_backward_validity(args) + + ctx.run_function = run_function + ctx.kwarg_keys = kwarg_keys + ctx.fwd_rng_state = utils.get_rng_state() + + tensor_inputs, packed_non_tensor_inputs = split_non_tensors(args) + if parent_ctx_dict["offload"]: + ctx.fwd_device = tuple(x.device for x in tensor_inputs) + ctx.grad_requirements = tuple(x.requires_grad for x in tensor_inputs) + tensor_inputs = tuple(x.cpu() for x in tensor_inputs) + + else: + ctx.fwd_device, ctx.grad_requirements = None, None + + ctx.save_for_backward(*tensor_inputs) + ctx.packed_non_tensor_inputs = packed_non_tensor_inputs + + with torch.no_grad(): + unpacked_args, unpacked_kwargs = unpack_kwargs(kwarg_keys, args) + outputs = run_function(*unpacked_args, **unpacked_kwargs) + + if isinstance(outputs, torch.Tensor): + return outputs + else: + # Autograd Functions don't like non-Tensor outputs. We can split the + # non-Tensor and Tensor outputs, returning the former by reference + # through *parent_ctx_dict* and returning the latter directly. + outputs, packed_non_tensor_outputs = split_non_tensors(outputs) + parent_ctx_dict["packed_non_tensor_outputs"] = packed_non_tensor_outputs + return outputs + + @staticmethod + def backward(ctx, *args): + if not torch.autograd._is_checkpoint_valid(): + raise RuntimeError( + "Checkpointing is not compatible with .grad(), please use .backward() if possible" + ) + + tensor_inputs: Tuple = ctx.saved_tensors + tensor_inputs = checkpoint.detach_variable(tensor_inputs) + if ctx.fwd_device is not None: + tensor_inputs = [ + t.to(ctx.fwd_device[i]) for i, t in enumerate(tensor_inputs) + ] + for i, need_grad in enumerate(ctx.grad_requirements): + tensor_inputs[i].requires_grad = need_grad + inputs = unpack_non_tensors(tensor_inputs, ctx.packed_non_tensor_inputs) + + # Store the current states. + bwd_rng_state = utils.get_rng_state() + + # Set the states to what it used to be before the forward pass. + utils.set_rng_state(ctx.fwd_rng_state) + + with torch.enable_grad(): + unpacked_args, unpacked_kwargs = unpack_kwargs(ctx.kwarg_keys, inputs) + outputs = ctx.run_function(*unpacked_args, **unpacked_kwargs) + tensor_outputs, _ = split_non_tensors(outputs) + # Set the states back to what it was at the start of this function. + utils.set_rng_state(bwd_rng_state) + + # Run backward() with only Tensors that require grad + outputs_with_grad = [] + args_with_grad = [] + for i in range(len(tensor_outputs)): + if tensor_outputs[i].requires_grad: + outputs_with_grad.append(tensor_outputs[i]) + args_with_grad.append(args[i]) + if len(outputs_with_grad) == 0: + raise RuntimeError( + "None of the outputs have requires_grad=True, " + "this checkpoint() is not necessary" + ) + + torch.autograd.backward(outputs_with_grad, args_with_grad) + + grads = tuple( + inp.grad if isinstance(inp, torch.Tensor) else None for inp in inputs + ) + return (None, None, None) + grads diff --git a/fairseq/modules/conv_tbc.py b/fairseq/modules/conv_tbc.py new file mode 100644 index 0000000000000000000000000000000000000000..65e17ec94f7e595cb657b3d2daaa1052a95d0677 --- /dev/null +++ b/fairseq/modules/conv_tbc.py @@ -0,0 +1,53 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +from torch import nn +from torch.nn.modules.utils import _single +from torch import Tensor + + +class ConvTBC(torch.nn.Module): + """1D convolution over an input of shape (time x batch x channel) + + The implementation uses gemm to perform the convolution. This implementation + is faster than cuDNN for small kernel sizes. + """ + + def __init__(self, in_channels, out_channels, kernel_size, padding=0): + super(ConvTBC, self).__init__() + self.in_channels = in_channels + self.out_channels = out_channels + self.kernel_size = _single(kernel_size) + self.padding = _single(padding) + + self.weight = torch.nn.Parameter( + torch.Tensor(self.kernel_size[0], in_channels, out_channels) + ) + self.bias = torch.nn.Parameter(torch.Tensor(out_channels)) + + self.reset_parameters() + + def reset_parameters(self): + nn.init.xavier_normal_(self.weight) + nn.init.zeros_(self.bias) + + def conv_tbc(self, input: Tensor): + return torch.conv_tbc( + input.contiguous(), self.weight, self.bias, self.padding[0] + ) + + def forward(self, input: Tensor): + return self.conv_tbc(input) + + def __repr__(self): + s = ( + "{name}({in_channels}, {out_channels}, kernel_size={kernel_size}" + ", padding={padding}" + ) + if self.bias is None: + s += ", bias=False" + s += ")" + return s.format(name=self.__class__.__name__, **self.__dict__) diff --git a/fairseq/modules/cross_entropy.py b/fairseq/modules/cross_entropy.py new file mode 100644 index 0000000000000000000000000000000000000000..6f33c24cb56e25f91595009af38e63784c2263a0 --- /dev/null +++ b/fairseq/modules/cross_entropy.py @@ -0,0 +1,61 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging + +import torch +import torch.nn.functional as F + + +logger = logging.getLogger(__name__) + + +def _cross_entropy_pytorch(logits, target, ignore_index=None, reduction="mean"): + lprobs = F.log_softmax(logits, dim=-1, dtype=torch.float32) + return F.nll_loss( + lprobs, + target, + ignore_index=ignore_index, + reduction=reduction, + ) + + +try: + import xentropy_cuda + from apex.contrib import xentropy + + def cross_entropy(logits, target, ignore_index=-100, reduction="mean"): + if logits.device == torch.device("cpu"): + return _cross_entropy_pytorch(logits, target, ignore_index, reduction) + else: + if not getattr(cross_entropy, "_has_logged_once", False): + logger.info("using fused cross entropy") + cross_entropy._has_logged_once = True + + half_to_float = logits.dtype == torch.half + losses = xentropy.SoftmaxCrossEntropyLoss.apply( + logits, + target, + 0.0, + ignore_index, + half_to_float, + ) + if reduction == "sum": + return losses.sum() + elif reduction == "mean": + if ignore_index >= 0: + return losses.sum() / target.ne(ignore_index).sum() + else: + return losses.mean() + elif reduction == "none": + return losses + else: + raise NotImplementedError + + +except ImportError: + + def cross_entropy(logits, target, ignore_index=-100, reduction="mean"): + return _cross_entropy_pytorch(logits, target, ignore_index, reduction) diff --git a/fairseq/modules/cuda_utils.cu b/fairseq/modules/cuda_utils.cu new file mode 100644 index 0000000000000000000000000000000000000000..516f1d92440e9e2c092f122e45d81b45cb135602 --- /dev/null +++ b/fairseq/modules/cuda_utils.cu @@ -0,0 +1,203 @@ +/** + * Copyright (c) Facebook, Inc. and its affiliates. + * + * This source code is licensed under the MIT license found in the + * LICENSE file in the root directory of this source tree. + */ + + +template <typename U, typename V> +constexpr __host__ __device__ auto divUp(U a, V b) -> decltype(a + b) { + return (a + b - 1) / b; +} + + +template<int FS, int SB, int padding_l, typename scalar_t> +__inline__ __device__ +void zeroSharedMem(scalar_t* data) { + /* + Given an array of length FS + SB, zero out the first padding_l and last + (FS - padding_l) values in the array + */ + + int tid = threadIdx.x; + + if (FS < SB) { + + // zero all if we have enough threads in a block to do all of them + if (tid < padding_l || tid > SB - FS + padding_l - 1) { + data[tid] = scalar_t(0.0); + } + } else { + + // otherwise zero out one block at a time + const int numIterations = divUp<int, int>(FS, SB); + for (int i = 0; i < numIterations; i++) { + int offset = i * SB; + if (tid + offset < padding_l) { + data[tid + offset] = scalar_t(0.0); + } else if (tid + offset < FS) { + data[SB + tid + offset] = scalar_t(0.0); + } + } + } +} + +template<typename scalar_t> +__inline__ __device__ +scalar_t warpReduce(scalar_t data) { + /* + Reduce an array within each warp. After processing all values in warp will + caontain the sum of all original values in that warp. + + data - pointer to data to reduce + */ + data += __shfl_xor_sync(SHFL_MASK, data, 16); + data += __shfl_xor_sync(SHFL_MASK, data, 8); + data += __shfl_xor_sync(SHFL_MASK, data, 4); + data += __shfl_xor_sync(SHFL_MASK, data, 2); + data += __shfl_xor_sync(SHFL_MASK, data, 1); + return data; +} + +template<typename scalar_t> +__inline__ __device__ +scalar_t blockReduce(scalar_t data) { + /* + Reduce an entire array on the block level. After processing, the + first value in the array will contain the reduced sum. + + data - pointer to data to reduce + */ + + static __shared__ scalar_t warpSum[32]; + const int tid = threadIdx.x; + int wid = tid / 32; + int lane = tid % 32; + + __syncthreads(); + + // reduce each warp then write to shared memory + scalar_t sum = warpReduce(data); + if (lane == 0) { + warpSum[wid] = sum; + } + + __syncthreads(); + + scalar_t v; + // perform final sum of partial warp sums + if (tid < blockDim.x / 32) { + v = warpSum[lane]; + } else { + v = scalar_t(0.0); + } + + if (wid == 0) { + v = warpReduce(v); + } + __syncthreads(); + + return v; +} + +void checkCudaStatus(cudaError_t status, int lineNumber = -1) { + + if (status != cudaSuccess) { + std::cout << cudaGetErrorString(status) + << " at line " << lineNumber << std::endl; + std::cout << "Exiting" << std::endl; + exit(1); + } +} + +template<int FS, int SB, int padding_l, typename scalar_t> +__device__ +void load_input_to_shared(const scalar_t* input, // global memory + int inputOffset, int sequenceLength, + int iteration, int numIterations, + bool no_prev, scalar_t* output /* shared memory */) { + /* + Load a block size of input into shared memory with + right and left overhang of total size FS. If previously + loaded memory, overlap will be shifted over to reduce + global memory access + + input - pointer to start of channel sequence + inputOffset - how far in the sequence to start loading + sequenceLength - total length of sequence + iteration - which block of sequence we are loading + numIterations - total number of blocks to load + no_prev - whether to load the whole block if the previous block + wasn't loaded + output - shared memory to write input to + */ + + const int tid = threadIdx.x; + + // Load the left "overhang" of input + if (iteration > 0) { + if (padding_l < SB) { + + // load all at once + if (tid < padding_l) { + output[tid] = (no_prev) ? input[inputOffset - padding_l + tid] : output[tid + SB]; + } + } else { + + // load in chunks of size SB + int numIterations = divUp<int, int>(padding_l, SB); + for (int i = 0; i < numIterations; i++) { + int offset = i * SB; + if ((tid + offset) < padding_l) { + output[tid + offset] = (no_prev) ? input[inputOffset - padding_l + tid + offset] : output[tid + offset + SB]; + } + } + } + } + + // Load the right "overhang" of input + if (iteration < (numIterations - 1)) { + const int elementsLeft = sequenceLength - (iteration+1) * SB; + + if ((FS - padding_l) < SB) { + + // load all at once + if (tid < (FS - padding_l)) { + output[padding_l + SB + tid] = (tid < elementsLeft) ? input[inputOffset + SB + tid] : scalar_t(0.0); + } + } else { + + // load in chunks of size SB + int numIterations = divUp<int, int>(FS - padding_l, SB); + for (int i = 0; i < numIterations; i++) { + int offset = i * SB; + if ((tid + offset) < (FS - padding_l)) { + output[padding_l + SB + tid + offset] = ((tid + offset) < elementsLeft) ? input[inputOffset + SB + tid + offset] : scalar_t(0.0); + } + } + } + } + + // We should also clear out the right "overhang" + if (iteration == (numIterations - 1)) { + if ((FS - padding_l) < SB) { + + // clear out all at once + if (tid < (FS - padding_l)) { + output[padding_l + SB + tid] = scalar_t(0.0); + } + } else { + + // clear in chunks of size SB + int numIterations = divUp<int, int>(FS - padding_l, SB); + for (int i = 0; i < numIterations; i++) { + int offset = i * SB; + if ((tid + offset) < (FS - padding_l)) { + output[padding_l + SB + tid + offset] = scalar_t(0.0); + } + } + } + } + output[tid + padding_l] = ((inputOffset + tid) < sequenceLength) ? input[inputOffset + tid] : scalar_t(0.0); +} diff --git a/fairseq/modules/downsampled_multihead_attention.py b/fairseq/modules/downsampled_multihead_attention.py new file mode 100644 index 0000000000000000000000000000000000000000..2cdece3f7fca2b830eb72999ce93f58667ed595b --- /dev/null +++ b/fairseq/modules/downsampled_multihead_attention.py @@ -0,0 +1,316 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +# + +import math + +import torch +import torch.nn as nn +import torch.nn.functional as F +from fairseq.modules.fairseq_dropout import FairseqDropout +from fairseq.modules.scalar_bias import scalar_bias + + +class SingleHeadAttention(nn.Module): + """ + Single-head attention that supports Gating and Downsampling + """ + + def __init__( + self, + out_channels, + embed_dim, + head_dim, + head_index, + dropout=0.0, + bias=True, + project_input=True, + gated=False, + downsample=False, + num_heads=1, + ): + super().__init__() + self.embed_dim = embed_dim + self.dropout_module = FairseqDropout( + dropout, module_name=self.__class__.__name__ + ) + self.head_index = head_index + self.head_dim = head_dim + self.project_input = project_input + self.gated = gated + self.downsample = downsample + self.num_heads = num_heads + self.projection = None + + k_layers = [] + v_layers = [] + if self.downsample: + k_layers.append(Downsample(self.head_index)) + v_layers.append(Downsample(self.head_index)) + out_proj_size = self.head_dim + else: + out_proj_size = self.head_dim * self.num_heads + if self.gated: + k_layers.append(GatedLinear(self.embed_dim, out_proj_size, bias=bias)) + self.in_proj_q = GatedLinear(self.embed_dim, out_proj_size, bias=bias) + v_layers.append(GatedLinear(self.embed_dim, out_proj_size, bias=bias)) + else: + k_layers.append(Linear(self.embed_dim, out_proj_size, bias=bias)) + self.in_proj_q = Linear(self.embed_dim, out_proj_size, bias=bias) + v_layers.append(Linear(self.embed_dim, out_proj_size, bias=bias)) + + self.in_proj_k = nn.Sequential(*k_layers) + self.in_proj_v = nn.Sequential(*v_layers) + + if self.downsample: + self.out_proj = Linear(out_proj_size, self.head_dim, bias=bias) + else: + self.out_proj = Linear(out_proj_size, out_channels, bias=bias) + + self.scaling = self.head_dim ** -0.5 + + def forward( + self, + query, + key, + value, + mask_future_timesteps=False, + key_padding_mask=None, + use_scalar_bias=False, + ): + """Input shape: Time x Batch x Channel + Self-attention can be implemented by passing in the same arguments for + query, key and value. Future timesteps can be masked with the + `mask_future_timesteps` argument. Padding elements can be excluded from + the key by passing a binary ByteTensor (`key_padding_mask`) with shape: + batch x src_len, where padding elements are indicated by 1s. + """ + src_len, bsz, out_channels = key.size() + tgt_len = query.size(0) + assert list(query.size()) == [tgt_len, bsz, out_channels] + assert key.size() == value.size() + + if key_padding_mask is not None: + assert key_padding_mask.size(0) == bsz + assert key_padding_mask.size(1) == src_len + + if self.downsample: + size = bsz + else: + size = bsz * self.num_heads + + k = key + v = value + q = query + if self.project_input: + q = self.in_proj_q(q) + k = self.in_proj_k(k) + v = self.in_proj_v(v) + src_len = k.size()[0] + q *= self.scaling + + if not self.downsample: + q = q.view(tgt_len, size, self.head_dim) + k = k.view(src_len, size, self.head_dim) + v = v.view(src_len, size, self.head_dim) + + q = q.transpose(0, 1) + k = k.transpose(0, 1) + v = v.transpose(0, 1) + + attn_weights = torch.bmm(q, k.transpose(1, 2)) + if mask_future_timesteps: + assert ( + query.size() == key.size() + ), "mask_future_timesteps only applies to self-attention" + attn_weights *= torch.tril( + attn_weights.data.new([1]).expand(tgt_len, tgt_len).clone(), + diagonal=-1, + )[:, :: self.head_index + 1 if self.downsample else 1].unsqueeze(0) + attn_weights += torch.triu( + attn_weights.data.new([-math.inf]).expand(tgt_len, tgt_len).clone(), + diagonal=0, + )[:, :: self.head_index + 1 if self.downsample else 1].unsqueeze(0) + tgt_size = tgt_len + if use_scalar_bias: + attn_weights = scalar_bias(attn_weights, 2) + v = scalar_bias(v, 1) + tgt_size += 1 + + if key_padding_mask is not None: + # don't attend to padding symbols + if key_padding_mask.max() > 0: + if self.downsample: + attn_weights = attn_weights.view(bsz, 1, tgt_len, src_len) + else: + attn_weights = attn_weights.view( + size, self.num_heads, tgt_len, src_len + ) + attn_weights = attn_weights.masked_fill( + key_padding_mask.unsqueeze(1).unsqueeze(2), + -math.inf, + ) + attn_weights = attn_weights.view(size, tgt_len, src_len) + attn_weights = F.softmax(attn_weights, dim=-1) + attn_weights = self.dropout_module(attn_weights) + + attn = torch.bmm(attn_weights, v) + if self.downsample: + attn = attn.transpose(0, 1).contiguous().view(tgt_len, bsz, self.head_dim) + else: + attn = attn.transpose(0, 1).contiguous().view(tgt_len, bsz, self.embed_dim) + + attn = self.out_proj(attn) + + return attn, attn_weights + + +class DownsampledMultiHeadAttention(nn.ModuleList): + """ + Multi-headed attention with Gating and Downsampling + """ + + def __init__( + self, + out_channels, + embed_dim, + num_heads, + dropout=0.0, + bias=True, + project_input=True, + gated=False, + downsample=False, + ): + self.embed_dim = embed_dim + self.num_heads = num_heads + self.head_dim = embed_dim // num_heads + self.downsample = downsample + self.gated = gated + self.project_input = project_input + assert self.head_dim * num_heads == embed_dim + + if self.downsample: + attention_heads = [] + for index in range(self.num_heads): + attention_heads.append( + SingleHeadAttention( + out_channels, + self.embed_dim, + self.head_dim, + index, + dropout, + bias, + self.project_input, + self.gated, + self.downsample, + self.num_heads, + ) + ) + super().__init__(modules=attention_heads) + self.out_proj = Linear(embed_dim, out_channels, bias=bias) + else: + # either we have a list of attention heads, or just one attention head + # if not being downsampled, we can do the heads with one linear layer instead of separate ones + super().__init__() + self.attention_module = SingleHeadAttention( + out_channels, + self.embed_dim, + self.head_dim, + 1, + dropout, + bias, + self.project_input, + self.gated, + self.downsample, + self.num_heads, + ) + + def forward( + self, + query, + key, + value, + mask_future_timesteps=False, + key_padding_mask=None, + use_scalar_bias=False, + ): + src_len, bsz, embed_dim = key.size() + tgt_len = query.size(0) + assert embed_dim == self.embed_dim + assert list(query.size()) == [tgt_len, bsz, embed_dim] + assert key.size() == value.size() + + tgt_size = tgt_len + if use_scalar_bias: + tgt_size += 1 + + attn = [] + attn_weights = [] + if self.downsample: + for attention_head_number in range(self.num_heads): + # call the forward of each attention head + _attn, _attn_weight = self[attention_head_number]( + query, + key, + value, + mask_future_timesteps, + key_padding_mask, + use_scalar_bias, + ) + attn.append(_attn) + attn_weights.append(_attn_weight) + full_attn = torch.cat(attn, dim=2) + full_attn = self.out_proj(full_attn) + return full_attn, attn_weights[0].clone() + else: + _attn, _attn_weight = self.attention_module( + query, + key, + value, + mask_future_timesteps, + key_padding_mask, + use_scalar_bias, + ) + attn.append(_attn) + attn_weights.append(_attn_weight) + full_attn = torch.cat(attn, dim=2) + full_attn_weights = torch.cat(attn_weights) + full_attn_weights = full_attn_weights.view( + bsz, self.num_heads, tgt_size, src_len + ) + full_attn_weights = full_attn_weights.sum(dim=1) / self.num_heads + return full_attn, full_attn_weights + + +class Downsample(nn.Module): + """ + Selects every nth element, where n is the index + """ + + def __init__(self, index): + super().__init__() + self.index = index + + def forward(self, x): + return x[:: self.index + 1] + + +def Linear(in_features, out_features, dropout=0.0, bias=True): + """Weight-normalized Linear layer (input: B x T x C)""" + m = nn.Linear(in_features, out_features, bias=bias) + m.weight.data.normal_(mean=0, std=math.sqrt((1 - dropout) / in_features)) + m.bias.data.zero_() + return nn.utils.weight_norm(m) + + +def GatedLinear(in_features, out_features, dropout=0.0, bias=True): + """Weight-normalized Linear layer (input: B x T x C) with interspersed GLU units""" + return nn.Sequential( + Linear(in_features, out_features * 4, dropout, bias), + nn.GLU(), + Linear(out_features * 2, out_features * 2, dropout, bias), + nn.GLU(), + Linear(out_features, out_features, dropout, bias), + ) diff --git a/fairseq/modules/dynamic_convolution.py b/fairseq/modules/dynamic_convolution.py new file mode 100644 index 0000000000000000000000000000000000000000..0121d453b9e026f5128dd41fce691aa1b4486448 --- /dev/null +++ b/fairseq/modules/dynamic_convolution.py @@ -0,0 +1,310 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn as nn +import torch.nn.functional as F +from fairseq import utils +from fairseq.incremental_decoding_utils import with_incremental_state +from fairseq.modules.fairseq_dropout import FairseqDropout + +from .unfold import unfold1d + + +def DynamicConv( + input_size, + kernel_size=1, + padding_l=None, + num_heads=1, + weight_dropout=0.0, + weight_softmax=False, + renorm_padding=False, + bias=False, + conv_bias=False, + query_size=None, + in_proj=False, +): + if torch.cuda.is_available(): + try: + from fairseq.modules.dynamicconv_layer import DynamicconvLayer + + return DynamicconvLayer( + input_size, + kernel_size=kernel_size, + padding_l=padding_l, + num_heads=num_heads, + weight_dropout=weight_dropout, + weight_softmax=weight_softmax, + renorm_padding=renorm_padding, + bias=bias, + conv_bias=conv_bias, + query_size=query_size, + ) + except ImportError as e: + print(e) + return DynamicConv1dTBC( + input_size, + kernel_size=kernel_size, + padding_l=padding_l, + num_heads=num_heads, + weight_dropout=weight_dropout, + weight_softmax=weight_softmax, + renorm_padding=renorm_padding, + bias=bias, + conv_bias=conv_bias, + query_size=query_size, + ) + + +def Linear(in_features, out_features, bias=True): + m = nn.Linear(in_features, out_features, bias) + nn.init.xavier_uniform_(m.weight) + if bias: + nn.init.constant_(m.bias, 0.0) + return m + + +@with_incremental_state +class DynamicConv1dTBC(nn.Module): + """Dynamic lightweight convolution taking T x B x C inputs + Args: + input_size: # of channels of the input + kernel_size: convolution channels + padding_l: padding to the left when using "same" padding + num_heads: number of heads used. The weight is of shape (num_heads, 1, kernel_size) + weight_dropout: the drop rate of the DropConnect to drop the weight + weight_softmax: normalize the weight with softmax before the convolution + renorm_padding: re-normalize the filters to ignore the padded part (only the non-padding parts sum up to 1) + bias: use bias + conv_bias: bias of the convolution + query_size: specified when feeding a different input as the query + in_proj: project the input and generate the filter together + + Shape: + Input: TxBxC, i.e. (timesteps, batch_size, input_size) + Output: TxBxC, i.e. (timesteps, batch_size, input_size) + + Attributes: + weight: the learnable weights of the module of shape + `(num_heads, 1, kernel_size)` + bias: the learnable bias of the module of shape `(input_size)` + """ + + def __init__( + self, + input_size, + kernel_size=1, + padding_l=None, + num_heads=1, + weight_dropout=0.0, + weight_softmax=False, + renorm_padding=False, + bias=False, + conv_bias=False, + query_size=None, + in_proj=False, + ): + super().__init__() + self.input_size = input_size + self.query_size = input_size if query_size is None else query_size + self.kernel_size = kernel_size + self.padding_l = padding_l + self.num_heads = num_heads + self.weight_dropout_module = FairseqDropout( + weight_dropout, module_name=self.__class__.__name__ + ) + self.weight_softmax = weight_softmax + self.renorm_padding = renorm_padding + + if in_proj: + self.weight_linear = Linear( + self.input_size, self.input_size + num_heads * kernel_size * 1 + ) + else: + self.weight_linear = Linear( + self.query_size, num_heads * kernel_size * 1, bias=bias + ) + if conv_bias: + self.conv_bias = nn.Parameter(torch.Tensor(input_size)) + else: + self.conv_bias = None + self.reset_parameters() + + @property + def in_proj(self): + return ( + self.weight_linear.out_features + == self.input_size + self.num_heads * self.kernel_size + ) + + def reset_parameters(self): + self.weight_linear.reset_parameters() + if self.conv_bias is not None: + nn.init.constant_(self.conv_bias, 0.0) + + def forward(self, x, incremental_state=None, query=None, unfold=None): + """Assuming the input, x, of the shape T x B x C and producing an output in the shape T x B x C + args: + x: Input of shape T x B x C, i.e. (timesteps, batch_size, input_size) + incremental_state: A dict to keep the state + unfold: unfold the input or not. If not, we use the matrix trick instead + query: use the specified query to predict the conv filters + """ + unfold = ( + x.size(0) > 512 if unfold is None else unfold + ) # use unfold mode as default for long sequence to save memory + unfold = unfold or (incremental_state is not None) + assert query is None or not self.in_proj + + if query is None: + query = x + if unfold: + output = self._forward_unfolded(x, incremental_state, query) + else: + output = self._forward_expanded(x, incremental_state, query) + + if self.conv_bias is not None: + output = output + self.conv_bias.view(1, 1, -1) + return output + + def _forward_unfolded(self, x, incremental_state, query): + """The conventional implementation of convolutions. + Unfolding the input by having a window shifting to the right.""" + T, B, C = x.size() + K, H = self.kernel_size, self.num_heads + R = C // H + assert R * H == C == self.input_size + + if self.in_proj: + proj = self.weight_linear(x) + x = proj.narrow(2, 0, self.input_size).contiguous() + weight = ( + proj.narrow(2, self.input_size, H * K).contiguous().view(T * B * H, -1) + ) + else: + weight = self.weight_linear(query).view(T * B * H, -1) + + # renorm_padding is only implemented in _forward_expanded + assert not self.renorm_padding or incremental_state is not None + + if incremental_state is not None: + input_buffer = self._get_input_buffer(incremental_state) + if input_buffer is None: + input_buffer = x.new() + x_unfold = torch.cat([input_buffer, x.unsqueeze(3)], dim=3) + if self.kernel_size > 1: + self._set_input_buffer( + incremental_state, x_unfold[:, :, :, -self.kernel_size + 1 :] + ) + x_unfold = x_unfold.view(T * B * H, R, -1) + else: + padding_l = self.padding_l + if K > T and padding_l == K - 1: + weight = weight.narrow(1, K - T, T) + K, padding_l = T, T - 1 + # unfold the input: T x B x C --> T' x B x C x K + x_unfold = unfold1d(x, K, padding_l, 0) + x_unfold = x_unfold.view(T * B * H, R, K) + + if self.weight_softmax and not self.renorm_padding: + weight = F.softmax(weight, dim=1) + weight = weight.narrow(1, 0, K) + + if incremental_state is not None: + weight = weight[:, -x_unfold.size(2) :] + K = weight.size(1) + + if self.weight_softmax and self.renorm_padding: + weight = F.softmax(weight, dim=1) + + weight = self.weight_dropout_module(weight, inplace=False) + + output = torch.bmm(x_unfold, weight.unsqueeze(2)) # T*B*H x R x 1 + output = output.view(T, B, C) + return output + + def _forward_expanded(self, x, incremental_stat, query): + """Turn the convolution filters into band matrices and do matrix multiplication. + This is faster when the sequence is short, but less memory efficient. + This is not used in the decoder during inference. + """ + T, B, C = x.size() + K, H = self.kernel_size, self.num_heads + R = C // H + assert R * H == C == self.input_size + if self.in_proj: + proj = self.weight_linear(x) + x = proj.narrow(2, 0, self.input_size).contiguous() + weight = ( + proj.narrow(2, self.input_size, H * K).contiguous().view(T * B * H, -1) + ) + else: + weight = self.weight_linear(query).view(T * B * H, -1) + + if not self.renorm_padding: + if self.weight_softmax: + weight = F.softmax(weight, dim=1) + weight = self.weight_dropout_module(weight, inplace=False) + weight = weight.narrow(1, 0, K).contiguous() + weight = weight.view(T, B * H, K).transpose(0, 1) + + x = x.view(T, B * H, R).transpose(0, 1) + if self.weight_softmax and self.renorm_padding: + # turn the convolution filters into band matrices + weight_expanded = weight.new(B * H, T, T + K - 1).fill_(float("-inf")) + weight_expanded.as_strided( + (B * H, T, K), (T * (T + K - 1), T + K, 1) + ).copy_(weight) + weight_expanded = weight_expanded.narrow(2, self.padding_l, T) + # normalize the weight over valid positions like self-attention + weight_expanded = F.softmax(weight_expanded, dim=2) + weight_expanded = self.weight_dropout_module(weight_expanded, inplace=False) + else: + P = self.padding_l + # For efficiency, we cut the kernel size and reduce the padding when the kernel is larger than the length + if K > T and P == K - 1: + weight = weight.narrow(2, K - T, T) + K, P = T, T - 1 + # turn the convolution filters into band matrices + weight_expanded = weight.new_zeros(B * H, T, T + K - 1, requires_grad=False) + weight_expanded.as_strided( + (B * H, T, K), (T * (T + K - 1), T + K, 1) + ).copy_(weight) + weight_expanded = weight_expanded.narrow(2, P, T) # B*H x T x T + output = torch.bmm(weight_expanded, x) + output = output.transpose(0, 1).contiguous().view(T, B, C) + return output + + def reorder_incremental_state(self, incremental_state, new_order): + input_buffer = self._get_input_buffer(incremental_state) + if input_buffer is not None: + input_buffer = input_buffer.index_select(1, new_order) + self._set_input_buffer(incremental_state, input_buffer) + + def _get_input_buffer(self, incremental_state): + return utils.get_incremental_state(self, incremental_state, "input_buffer") + + def _set_input_buffer(self, incremental_state, new_buffer): + return utils.set_incremental_state( + self, incremental_state, "input_buffer", new_buffer + ) + + def extra_repr(self): + s = "{}, kernel_size={}, padding_l={}, num_heads={}, weight_softmax={}, conv_bias={}, renorm_padding={}, in_proj={}".format( + self.input_size, + self.kernel_size, + self.padding_l, + self.num_heads, + self.weight_softmax, + self.conv_bias is not None, + self.renorm_padding, + self.in_proj, + ) + + if self.query_size != self.input_size: + s += ", query_size={}".format(self.query_size) + if self.weight_dropout_module.p > 0.0: + s += ", weight_dropout={}".format(self.weight_dropout_module.p) + return s diff --git a/fairseq/modules/dynamic_crf_layer.py b/fairseq/modules/dynamic_crf_layer.py new file mode 100644 index 0000000000000000000000000000000000000000..8fcc6b8d2672d2eacc6d01b9688bac44d5e1ce26 --- /dev/null +++ b/fairseq/modules/dynamic_crf_layer.py @@ -0,0 +1,189 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +""" +This file is to re-implemented the low-rank and beam approximation of CRF layer +Proposed by: + +Sun, Zhiqing, et al. +Fast Structured Decoding for Sequence Models +https://arxiv.org/abs/1910.11555 + +The CRF implementation is mainly borrowed from +https://github.com/kmkurn/pytorch-crf/blob/master/torchcrf/__init__.py + +""" + +import numpy as np +import torch +import torch.nn as nn + + +def logsumexp(x, dim=1): + return torch.logsumexp(x.float(), dim=dim).type_as(x) + + +class DynamicCRF(nn.Module): + """Dynamic CRF layer is used to approximate the traditional + Conditional Random Fields (CRF) + $P(y | x) = 1/Z(x) exp(sum_i s(y_i, x) + sum_i t(y_{i-1}, y_i, x))$ + + where in this function, we assume the emition scores (s) are given, + and the transition score is a |V| x |V| matrix $M$ + + in the following two aspects: + (1) it used a low-rank approximation for the transition matrix: + $M = E_1 E_2^T$ + (2) it used a beam to estimate the normalizing factor Z(x) + """ + + def __init__(self, num_embedding, low_rank=32, beam_size=64): + super().__init__() + + self.E1 = nn.Embedding(num_embedding, low_rank) + self.E2 = nn.Embedding(num_embedding, low_rank) + + self.vocb = num_embedding + self.rank = low_rank + self.beam = beam_size + + def extra_repr(self): + return "vocab_size={}, low_rank={}, beam_size={}".format( + self.vocb, self.rank, self.beam + ) + + def forward(self, emissions, targets, masks, beam=None): + """ + Compute the conditional log-likelihood of a sequence of target tokens given emission scores + + Args: + emissions (`~torch.Tensor`): Emission score are usually the unnormalized decoder output + ``(batch_size, seq_len, vocab_size)``. We assume batch-first + targets (`~torch.LongTensor`): Sequence of target token indices + ``(batch_size, seq_len) + masks (`~torch.ByteTensor`): Mask tensor with the same size as targets + + Returns: + `~torch.Tensor`: approximated log-likelihood + """ + numerator = self._compute_score(emissions, targets, masks) + denominator = self._compute_normalizer(emissions, targets, masks, beam) + return numerator - denominator + + def forward_decoder(self, emissions, masks=None, beam=None): + """ + Find the most likely output sequence using Viterbi algorithm. + + Args: + emissions (`~torch.Tensor`): Emission score are usually the unnormalized decoder output + ``(batch_size, seq_len, vocab_size)``. We assume batch-first + masks (`~torch.ByteTensor`): Mask tensor with the same size as targets + + Returns: + `~torch.LongTensor`: decoded sequence from the CRF model + """ + return self._viterbi_decode(emissions, masks, beam) + + def _compute_score(self, emissions, targets, masks=None): + batch_size, seq_len = targets.size() + emission_scores = emissions.gather(2, targets[:, :, None])[:, :, 0] # B x T + transition_scores = (self.E1(targets[:, :-1]) * self.E2(targets[:, 1:])).sum(2) + + scores = emission_scores + scores[:, 1:] += transition_scores + + if masks is not None: + scores = scores * masks.type_as(scores) + return scores.sum(-1) + + def _compute_normalizer(self, emissions, targets=None, masks=None, beam=None): + # HACK: we include "target" which is a hueristic for training + # HACK: we use a beam of tokens to approximate the normalizing factor (which is bad?) + + beam = beam if beam is not None else self.beam + batch_size, seq_len = emissions.size()[:2] + if targets is not None: + _emissions = emissions.scatter(2, targets[:, :, None], np.float("inf")) + beam_targets = _emissions.topk(beam, 2)[1] + beam_emission_scores = emissions.gather(2, beam_targets) + else: + beam_emission_scores, beam_targets = emissions.topk(beam, 2) + beam_transition_score1 = self.E1(beam_targets[:, :-1]) # B x (T-1) x K x D + beam_transition_score2 = self.E2(beam_targets[:, 1:]) # B x (T-1) x K x D + beam_transition_matrix = torch.bmm( + beam_transition_score1.view(-1, beam, self.rank), + beam_transition_score2.view(-1, beam, self.rank).transpose(1, 2), + ) + beam_transition_matrix = beam_transition_matrix.view(batch_size, -1, beam, beam) + + # compute the normalizer in the log-space + score = beam_emission_scores[:, 0] # B x K + for i in range(1, seq_len): + next_score = score[:, :, None] + beam_transition_matrix[:, i - 1] + next_score = logsumexp(next_score, dim=1) + beam_emission_scores[:, i] + + if masks is not None: + score = torch.where(masks[:, i : i + 1], next_score, score) + else: + score = next_score + + # Sum (log-sum-exp) over all possible tags + return logsumexp(score, dim=1) + + def _viterbi_decode(self, emissions, masks=None, beam=None): + # HACK: we use a beam of tokens to approximate the normalizing factor (which is bad?) + + beam = beam if beam is not None else self.beam + batch_size, seq_len = emissions.size()[:2] + beam_emission_scores, beam_targets = emissions.topk(beam, 2) + beam_transition_score1 = self.E1(beam_targets[:, :-1]) # B x (T-1) x K x D + beam_transition_score2 = self.E2(beam_targets[:, 1:]) # B x (T-1) x K x D + beam_transition_matrix = torch.bmm( + beam_transition_score1.view(-1, beam, self.rank), + beam_transition_score2.view(-1, beam, self.rank).transpose(1, 2), + ) + beam_transition_matrix = beam_transition_matrix.view(batch_size, -1, beam, beam) + + traj_tokens, traj_scores = [], [] + finalized_tokens, finalized_scores = [], [] + + # compute the normalizer in the log-space + score = beam_emission_scores[:, 0] # B x K + dummy = ( + torch.arange(beam, device=score.device).expand(*score.size()).contiguous() + ) + + for i in range(1, seq_len): + traj_scores.append(score) + _score = score[:, :, None] + beam_transition_matrix[:, i - 1] + _score, _index = _score.max(dim=1) + _score = _score + beam_emission_scores[:, i] + + if masks is not None: + score = torch.where(masks[:, i : i + 1], _score, score) + index = torch.where(masks[:, i : i + 1], _index, dummy) + else: + score, index = _score, _index + traj_tokens.append(index) + + # now running the back-tracing and find the best + best_score, best_index = score.max(dim=1) + finalized_tokens.append(best_index[:, None]) + finalized_scores.append(best_score[:, None]) + + for idx, scs in zip(reversed(traj_tokens), reversed(traj_scores)): + previous_index = finalized_tokens[-1] + finalized_tokens.append(idx.gather(1, previous_index)) + finalized_scores.append(scs.gather(1, previous_index)) + + finalized_tokens.reverse() + finalized_tokens = torch.cat(finalized_tokens, 1) + finalized_tokens = beam_targets.gather(2, finalized_tokens[:, :, None])[:, :, 0] + + finalized_scores.reverse() + finalized_scores = torch.cat(finalized_scores, 1) + finalized_scores[:, 1:] = finalized_scores[:, 1:] - finalized_scores[:, :-1] + + return finalized_scores, finalized_tokens diff --git a/fairseq/modules/dynamicconv_layer/__init__.py b/fairseq/modules/dynamicconv_layer/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..22dc6f403d2a0ecdb1b9e7e69ed96bd560e93b2c --- /dev/null +++ b/fairseq/modules/dynamicconv_layer/__init__.py @@ -0,0 +1,6 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .dynamicconv_layer import DynamicconvLayer # noqa diff --git a/fairseq/modules/dynamicconv_layer/cuda_function_gen.py b/fairseq/modules/dynamicconv_layer/cuda_function_gen.py new file mode 100644 index 0000000000000000000000000000000000000000..9304f99eb8169a614f39babc830c84cac80e080b --- /dev/null +++ b/fairseq/modules/dynamicconv_layer/cuda_function_gen.py @@ -0,0 +1,223 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +def gen_forward(): + + kernels = [3, 5, 7, 15, 31, 63, 127, 255] + blocks = [32, 64, 128, 256] + + head = """ +/** + * Copyright (c) Facebook, Inc. and its affiliates. + * + * This source code is licensed under the MIT license found in the + * LICENSE file in the root directory of this source tree. + */ + +#include "dynamicconv_cuda.cuh" + +std::vector<at::Tensor> dynamicconv_cuda_forward(at::Tensor input, at::Tensor weight, int padding_l) { + + at::DeviceGuard g(input.device()); + const auto minibatch = input.size(0); + const auto numFeatures = input.size(1); + const auto sequenceLength = input.size(2); + + const auto numHeads = weight.size(1); + const auto filterSize = weight.size(2); + + const auto numFiltersInBlock = numFeatures / numHeads; + const dim3 blocks(minibatch, numFeatures); + + auto output = at::zeros_like(input); + auto stream = at::cuda::getCurrentCUDAStream(); +""" + + switch = """ + switch(filterSize) { +""" + + case_k = """ + case {k}: +""" + + main_block = """ + if (padding_l == {pad}) {{ + AT_DISPATCH_FLOATING_TYPES_AND_HALF(input.scalar_type(), "dynamicconv_forward", ([&] {{ + dynamicconv_forward_kernel<{k}, {b_size}, {pad}, scalar_t> + <<<blocks, {b_size}, 0, stream>>>( + input.data<scalar_t>(), + weight.data<scalar_t>(), + minibatch, + sequenceLength, + numFeatures, + numFiltersInBlock, + numHeads, + output.data<scalar_t>()); + }})); + }} else +""" + + bad_padding = """ + { + std::cout << "WARNING: Unsupported padding size - skipping forward pass" << std::endl; + } + break;\n +""" + + end = """ + default: + std::cout << "WARNING: Unsupported filter length passed - skipping forward pass" << std::endl; + } + + return {output}; +} +""" + + with open("dynamicconv_cuda_forward.cu", "w") as forward: + forward.write(head) + forward.write(switch) + for k in kernels: + b_size = 32 + for b in blocks: + if b > k: + b_size = b + break + forward.write(case_k.format(k=k)) + for pad in [k // 2, k - 1]: + forward.write(main_block.format(k=k, b_size=b_size, pad=pad)) + forward.write(bad_padding) + forward.write(end) + + +def gen_backward(): + + kernels = [3, 5, 7, 15, 31, 63, 127, 255] + thresh = [512, 512, 512, 512, 512, 380, 256, 256] + min_block = [64, 64, 64, 64, 64, 64, 128, 256] + seqs = [32 * x for x in [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16]] + + head = """ +/** + * Copyright (c) Facebook, Inc. and its affiliates. + * + * This source code is licensed under the MIT license found in the + * LICENSE file in the root directory of this source tree. + */ + +#include "dynamicconv_cuda.cuh" + +std::vector<at::Tensor> dynamicconv_cuda_backward(at::Tensor gradOutput, int padding_l, at::Tensor input, at::Tensor weight) { + + at::DeviceGuard g(input.device()); + const auto minibatch = input.size(0); + const auto numFeatures = input.size(1); + const auto sequenceLength = input.size(2); + + const auto numHeads = weight.size(1); + const auto filterSize = weight.size(2); + + const auto numFiltersInBlock = numFeatures / numHeads; + auto numChunks = 1; + + auto gradInput = at::zeros_like(input); + auto gradWeight = at::zeros_like(weight); + auto stream = at::cuda::getCurrentCUDAStream(); + + dim3 blocks(minibatch, numHeads, numChunks); +""" + + sequence_if = """ + if (sequenceLength < {seq}) {{ + switch(filterSize) {{ +""" + + case_k = """ + case {k}: +""" + + chunks_reset = """ + numChunks = int(ceilf(sequenceLength/float({b_size}))); + blocks = dim3(minibatch, numHeads, numChunks); +""" + + main_block = """ + if (padding_l == {p}) {{ + AT_DISPATCH_FLOATING_TYPES_AND_HALF(gradOutput.scalar_type(), "dynamicconv_backward", ([&] {{ + dynamicconv_backward_kernel<{k}, {b_size}, {p}, scalar_t> + <<<blocks, {b_size}, 0, stream>>>( + gradOutput.data<scalar_t>(), + input.data<scalar_t>(), + weight.data<scalar_t>(), + minibatch, + sequenceLength, + numFeatures, + numFiltersInBlock, + numHeads, + gradWeight.data<scalar_t>(), + gradInput.data<scalar_t>()); + }})); + }} else +""" + + bad_padding = """ + { + std::cout << "WARNING: Unsupported padding size - skipping backward pass" << std::endl; + } + break;\n +""" + + bad_filter = """ + default: + std::cout << "WARNING: Unsupported filter length passed - skipping backward pass" << std::endl; + } +""" + + con_else = """ + } else +""" + + final_else = """ + { + switch(filterSize) { +""" + + last_return = """ + } + return {gradInput, gradWeight}; +} +""" + + with open("dynamicconv_cuda_backward.cu", "w") as backward: + backward.write(head) + for seq in seqs: + backward.write(sequence_if.format(seq=seq)) + for k, t, m in zip(kernels, thresh, min_block): + backward.write(case_k.format(k=k)) + if seq <= t: + b_size = seq + else: + b_size = m + backward.write(chunks_reset.format(b_size=b_size)) + for p in [k // 2, k - 1]: + backward.write(main_block.format(k=k, b_size=b_size, p=p)) + backward.write(bad_padding) + backward.write(bad_filter) + backward.write(con_else) + backward.write(final_else) + for k, m in zip(kernels, min_block): + backward.write(case_k.format(k=k)) + backward.write(chunks_reset.format(b_size=m)) + for p in [k // 2, k - 1]: + backward.write(main_block.format(k=k, b_size=m, p=p)) + backward.write(bad_padding) + backward.write(bad_filter) + backward.write(last_return) + + +if __name__ == "__main__": + gen_forward() + gen_backward() diff --git a/fairseq/modules/dynamicconv_layer/dynamicconv_cuda.cpp b/fairseq/modules/dynamicconv_layer/dynamicconv_cuda.cpp new file mode 100644 index 0000000000000000000000000000000000000000..ebd4df0e9608d769f31eadc6e0b487505f11b279 --- /dev/null +++ b/fairseq/modules/dynamicconv_layer/dynamicconv_cuda.cpp @@ -0,0 +1,56 @@ +/** + * Copyright (c) Facebook, Inc. and its affiliates. + * + * This source code is licensed under the MIT license found in the + * LICENSE file in the root directory of this source tree. + */ + +#include <torch/extension.h> +#include <vector> + +std::vector<at::Tensor> dynamicconv_cuda_forward( + at::Tensor input, + at::Tensor filters, + int padding_l); + +std::vector<at::Tensor> dynamicconv_cuda_backward( + at::Tensor gradOutput, + int padding_l, + at::Tensor input, + at::Tensor filters); + + +#define CHECK_CUDA(x) AT_ASSERTM(x.type().is_cuda(), #x " must be a CUDA tensor") +#define CHECK_CONTIGUOUS(x) AT_ASSERTM(x.is_contiguous(), #x " must be contiguous") +#define CHECK_INPUT(x) CHECK_CUDA(x); CHECK_CONTIGUOUS(x) + +std::vector<at::Tensor> dynamicconv_forward( + at::Tensor input, + at::Tensor filters, + int padding_l) { + + CHECK_INPUT(input); + CHECK_INPUT(filters); + + return dynamicconv_cuda_forward(input, filters, + padding_l); +} + +std::vector<at::Tensor> dynamicconv_backward( + at::Tensor gradOutput, + int padding_l, + at::Tensor input, + at::Tensor filters) { + + CHECK_INPUT(gradOutput); + CHECK_INPUT(input); + CHECK_INPUT(filters); + + return dynamicconv_cuda_backward(gradOutput, padding_l, + input, filters); +} + +PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { + m.def("forward", &dynamicconv_forward, "dynamicconv forward (CUDA)"); + m.def("backward", &dynamicconv_backward, "dynamicconv backward (CUDA)"); +} diff --git a/fairseq/modules/dynamicconv_layer/dynamicconv_cuda.cuh b/fairseq/modules/dynamicconv_layer/dynamicconv_cuda.cuh new file mode 100644 index 0000000000000000000000000000000000000000..2196259433aefc88f96cd5bbcae57740a9a8c2dc --- /dev/null +++ b/fairseq/modules/dynamicconv_layer/dynamicconv_cuda.cuh @@ -0,0 +1,51 @@ +/** + * Copyright (c) Facebook, Inc. and its affiliates. + * + * This source code is licensed under the MIT license found in the + * LICENSE file in the root directory of this source tree. + */ + +#include <ATen/ATen.h> +#include <c10/cuda/CUDAStream.h> + +#include <cuda.h> +#include <cuda_fp16.h> +#include <cuda_runtime.h> + +#include <algorithm> +#include <functional> +#include <iostream> +#include <stdexcept> +#include <utility> +#include <vector> + +#include <stdlib.h> +#include <assert.h> +#include <math.h> + +#define SHFL_MASK 0xffffffff + +template<int FS, int SB, int padding_l, typename scalar_t> +__global__ +void dynamicconv_forward_kernel(const scalar_t* input, + const scalar_t* weight, + int minibatch, + int sequenceLength, + int numFeatures, + int numFiltersInBlock, + int numHeads, + scalar_t* output); + +template<int FS, int SB, int padding_l, typename scalar_t> +__global__ +void dynamicconv_backward_kernel( + const scalar_t* gradOutput, // B * C * T + const scalar_t* input, // B * C * T + const scalar_t* weight, + int minibatch, + int sequenceLength, + int numFeatures, + int numFiltersInBlock, + int numHeads, + scalar_t* gradWeight, + scalar_t* gradInput); // B * H * k * T diff --git a/fairseq/modules/dynamicconv_layer/dynamicconv_cuda_kernel.cu b/fairseq/modules/dynamicconv_layer/dynamicconv_cuda_kernel.cu new file mode 100644 index 0000000000000000000000000000000000000000..300d35b6478080a9594a22e335988c321d43127f --- /dev/null +++ b/fairseq/modules/dynamicconv_layer/dynamicconv_cuda_kernel.cu @@ -0,0 +1,168 @@ +/** + * Copyright (c) Facebook, Inc. and its affiliates. + * + * This source code is licensed under the MIT license found in the + * LICENSE file in the root directory of this source tree. + */ + +#include "dynamicconv_cuda.cuh" +#include "dynamicconv_cuda_forward.cu" +#include "dynamicconv_cuda_backward.cu" +#include "../cuda_utils.cu" + +// FS is filter size and kernels are specialized for filter sizes +template<int FS, int SB, int padding_l, typename scalar_t> +__global__ +void dynamicconv_forward_kernel(const scalar_t* input, + const scalar_t* weight, + int minibatch, + int sequenceLength, + int numFeatures, + int numFiltersInBlock, + int numHeads, + scalar_t* output) { + assert(blockDim.x == SB); + + const int tid = threadIdx.x; + const int batchIdx = blockIdx.x; + const int featureIdx = blockIdx.y; + const int head = featureIdx / numFiltersInBlock; + + const int IOOffset = batchIdx * numFeatures * sequenceLength + + featureIdx * sequenceLength; + const scalar_t* inputFeature = &input[IOOffset]; + scalar_t* outputFeature = &output[IOOffset]; + + scalar_t filter[FS]; + + __shared__ scalar_t tempInput[SB + FS]; + zeroSharedMem<FS, SB, padding_l>(tempInput); + + const int numIterations = divUp<int, int>(sequenceLength, SB); + + for (int i = 0; i < numIterations; ++i) { + __syncthreads(); + const int inputOffset = i * SB; + load_input_to_shared<FS, SB, padding_l>(inputFeature, inputOffset, + sequenceLength, i, + numIterations, false, tempInput); + __syncthreads(); + if (inputOffset + tid < sequenceLength) { + + #pragma unroll + for (int k = 0; k < FS; ++k) { + const int filterOffset = batchIdx * numHeads * FS * sequenceLength + + head * FS * sequenceLength + + k * sequenceLength + + i * SB + tid; + filter[k] = weight[filterOffset]; + } + + scalar_t out = scalar_t(0.0); + #pragma unroll + for (int k = 0; k < FS; ++k) { + out += filter[k] * tempInput[tid + k]; + } + + outputFeature[inputOffset + tid] = out; + + } + } +} + +template<int FS, int SB, int padding_l, typename scalar_t> +__global__ +void dynamicconv_backward_kernel( + const scalar_t* gradOutput, // B * C * T + const scalar_t* input, // B * C * T + const scalar_t* weight, + int minibatch, + int sequenceLength, + int numFeatures, + int numFiltersInBlock, + int numHeads, + scalar_t* gradWeight, + scalar_t* gradInput) { // B * H * k * T + + assert(blockDim.x == SB); + + // each block operates on a single batch and filter head + const int tid = threadIdx.x; + const int batchIdx = blockIdx.x; + const int headIdx = blockIdx.y; + const int chunkIdx = blockIdx.z; + + const int numChunks = divUp<int, int>(sequenceLength, SB); + const int inputOffset = chunkIdx * SB; + + // initialize shared memory for output gradient and input + __shared__ scalar_t tempGradOutput[SB + FS]; + __shared__ scalar_t tempInput[SB + FS]; + const int padding = FS - padding_l - 1; + + zeroSharedMem<FS, SB, padding>(tempGradOutput); + zeroSharedMem<FS, SB, padding_l>(tempInput); + + // initialize local filter and weight gradient sum arrays + scalar_t tempGradSum[FS]; + scalar_t bfilter[FS]; + for (int k = 0; k < FS; ++k) { + tempGradSum[k] = scalar_t(0.0); + + int idxOffset = inputOffset + tid + k - padding; + if (idxOffset >= 0 && idxOffset < sequenceLength) { + int bfilterOffset = batchIdx * numHeads * FS * sequenceLength + + headIdx * FS * sequenceLength + + (FS - k - 1) * sequenceLength + + idxOffset; + bfilter[k] = weight[bfilterOffset]; + } else { + bfilter[k] = scalar_t(0.0); + } + } + + + // iterate over filter block + for (int featureIdx = 0; featureIdx < numFiltersInBlock; ++featureIdx) { + __syncthreads(); + + // load input and output gradient for this channel and chunk + const int IOOffset = batchIdx * numFeatures * sequenceLength + + (headIdx * numFiltersInBlock + featureIdx) * sequenceLength; + const scalar_t* inputFeature = &input[IOOffset]; + const scalar_t* gradOutputFeature = &gradOutput[IOOffset]; + scalar_t* gradInputFeature = &gradInput[IOOffset]; + + load_input_to_shared<FS, SB, padding>(gradOutputFeature, inputOffset, + sequenceLength, chunkIdx, + numChunks, true, tempGradOutput); + load_input_to_shared<FS, SB, padding_l>(inputFeature, inputOffset, + sequenceLength, chunkIdx, + numChunks, true, tempInput); + __syncthreads(); + + // sum input and weight gradients + scalar_t out = scalar_t(0.0); + #pragma unroll + for (int k = 0; k < FS; ++k) { + tempGradSum[k] += tempInput[tid + k] * tempGradOutput[tid + padding]; + out += bfilter[k] * tempGradOutput[tid + k]; + } + + if (inputOffset + tid < sequenceLength) { + gradInputFeature[inputOffset + tid] = out; + } + } + + const int gradOffset = batchIdx * numHeads * FS * sequenceLength + + headIdx * FS * sequenceLength; + scalar_t *gradWeightFeature = &gradWeight[gradOffset]; + + // write weight gradient + if (inputOffset + tid < sequenceLength) { + for (int k = 0; k < FS; ++k) { + const int outputOffset = k * sequenceLength + inputOffset + tid; + gradWeightFeature[outputOffset] = tempGradSum[k]; + } + } +} diff --git a/fairseq/modules/dynamicconv_layer/dynamicconv_layer.py b/fairseq/modules/dynamicconv_layer/dynamicconv_layer.py new file mode 100644 index 0000000000000000000000000000000000000000..711ed03483f4089dbe91964a89021b49eeffbedc --- /dev/null +++ b/fairseq/modules/dynamicconv_layer/dynamicconv_layer.py @@ -0,0 +1,227 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import dynamicconv_cuda +import torch +import torch.nn.functional as F +from fairseq import utils +from fairseq.incremental_decoding_utils import with_incremental_state +from fairseq.modules.fairseq_dropout import FairseqDropout +from fairseq.modules.unfold import unfold1d +from torch import nn +from torch.autograd import Function + + +class dynamicconvFunction(Function): + @staticmethod + def forward(ctx, x, weights, padding_l): + ctx.padding_l = padding_l + outputs = dynamicconv_cuda.forward(x, weights, padding_l) + variables = [x, weights] + ctx.save_for_backward(*variables) + return outputs[0] + + @staticmethod + def backward(ctx, grad_output): + outputs = dynamicconv_cuda.backward( + grad_output.contiguous(), ctx.padding_l, *ctx.saved_tensors + ) + grad_input, grad_weights = outputs + return grad_input, grad_weights, None + + +@with_incremental_state +class DynamicconvLayer(nn.Module): + def __init__( + self, + input_size, + kernel_size=1, + padding_l=None, + weight_softmax=False, + num_heads=1, + weight_dropout=0.0, + bias=False, + renorm_padding=False, + conv_bias=False, + query_size=None, + ): + + super(DynamicconvLayer, self).__init__() + self.input_size = input_size + self.query_size = input_size if query_size is None else query_size + self.kernel_size = kernel_size + self.padding_l = padding_l + self.num_heads = num_heads + self.weight_softmax = weight_softmax + self.weight_dropout_module = FairseqDropout( + weight_dropout, module_name=self.__class__.__name__ + ) + self.renorm_padding = renorm_padding + self.bias = bias + + self.weight_linear = nn.Linear(input_size, num_heads * kernel_size, bias) + if conv_bias: + self.conv_bias = nn.Parameter(torch.Tensor(input_size)) + else: + self.conv_bias = None + self.reset_parameters() + + def reset_parameters(self): + nn.init.xavier_uniform_(self.weight_linear.weight) + if self.conv_bias is not None: + nn.init.constant_(self.conv_bias, 0.0) + nn.init.constant_(self.weight_linaer.bias, 0.0) + + def forward(self, x, incremental_state=None, query=None, unfold=None): + + T, B, C = x.size() + K, H = self.kernel_size, self.num_heads + # R = C // H + + # during inference time, incremental BMM is faster + if incremental_state is not None: + unfold = ( + x.size(0) > 512 if unfold is None else unfold + ) # use unfold mode as default for long sequence to save memory + unfold = unfold or (incremental_state is not None) + assert query is None + + if query is None: + query = x + if unfold: + output = self._forward_unfolded(x, incremental_state, query) + else: + output = self._forward_expanded(x, incremental_state, query) + + if self.conv_bias is not None: + output = output + self.conv_bias.view(1, 1, -1) + + return output + + # during training time, use CUDA kernel + else: + weight = self.weight_linear(x).view(T, B, H, K) + if self.weight_softmax: + weight = F.softmax(weight, dim=-1) + if self.weight_dropout_module.p: + weight = self.weight_dropout_module(weight) + + weight = weight.permute(1, 2, 3, 0).contiguous() + self.filters = weight + x = x.permute(1, 2, 0).contiguous() + output = dynamicconvFunction.apply(x, weight, self.padding_l).permute( + 2, 0, 1 + ) + if self.conv_bias is not None: + output = output + self.conv_bias.view(1, 1, -1) + return output + + def reorder_incremental_state(self, incremental_state, new_order): + input_buffer = self._get_input_buffer(incremental_state) + if input_buffer is not None: + input_buffer = input_buffer.index_select(1, new_order) + self._set_input_buffer(incremental_state, input_buffer) + + def _get_input_buffer(self, incremental_state): + return utils.get_incremental_state(self, incremental_state, "input_buffer") + + def _set_input_buffer(self, incremental_state, new_buffer): + return utils.set_incremental_state( + self, incremental_state, "input_buffer", new_buffer + ) + + def _forward_unfolded(self, x, incremental_state, query): + """The conventional implementation of convolutions. + Unfolding the input by having a window shifting to the right.""" + T, B, C = x.size() + K, H = self.kernel_size, self.num_heads + R = C // H + assert R * H == C == self.input_size + + weight = self.weight_linear(query).view(T * B * H, -1) + + # renorm_padding is only implemented in _forward_expanded + assert not self.renorm_padding or incremental_state is not None + + if incremental_state is not None: + input_buffer = self._get_input_buffer(incremental_state) + if input_buffer is None: + input_buffer = x.new() + x_unfold = torch.cat([input_buffer, x.unsqueeze(3)], dim=3) + if self.kernel_size > 1: + self._set_input_buffer( + incremental_state, x_unfold[:, :, :, -self.kernel_size + 1 :] + ) + x_unfold = x_unfold.view(T * B * H, R, -1) + else: + padding_l = self.padding_l + if K > T and padding_l == K - 1: + weight = weight.narrow(1, K - T, T) + K, padding_l = T, T - 1 + # unfold the input: T x B x C --> T' x B x C x K + x_unfold = unfold1d(x, K, padding_l, 0) + x_unfold = x_unfold.view(T * B * H, R, K) + + if self.weight_softmax and not self.renorm_padding: + weight = F.softmax(weight, dim=1) + weight = weight.narrow(1, 0, K) + + if incremental_state is not None: + weight = weight[:, -x_unfold.size(2) :] + K = weight.size(1) + + if self.weight_softmax and self.renorm_padding: + weight = F.softmax(weight, dim=1) + + weight = self.weight_dropout_module(weight, inplace=False) + + output = torch.bmm(x_unfold, weight.unsqueeze(2)) # T*B*H x R x 1 + output = output.view(T, B, C) + return output + + def _forward_expanded(self, x, incremental_stat, query): + """Turn the convolution filters into band matrices and do matrix multiplication. + This is faster when the sequence is short, but less memory efficient. + This is not used in the decoder during inference. + """ + T, B, C = x.size() + K, H = self.kernel_size, self.num_heads + R = C // H + assert R * H == C == self.input_size + weight = self.weight_linear(query).view(T * B * H, -1) + + if not self.renorm_padding: + if self.weight_softmax: + weight = F.softmax(weight, dim=1) + weight = self.weight_dropout_module(weight, inplace=False) + weight = weight.narrow(1, 0, K).contiguous() + weight = weight.view(T, B * H, K).transpose(0, 1) + + x = x.view(T, B * H, R).transpose(0, 1) + if self.weight_softmax and self.renorm_padding: + # turn the convolution filters into band matrices + weight_expanded = weight.new(B * H, T, T + K - 1).fill_(float("-inf")) + weight_expanded.as_strided( + (B * H, T, K), (T * (T + K - 1), T + K, 1) + ).copy_(weight) + weight_expanded = weight_expanded.narrow(2, self.padding_l, T) + # normalize the weight over valid positions like self-attention + weight_expanded = F.softmax(weight_expanded, dim=2) + weight_expanded = self.weight_dropout_module(weight_expanded, inplace=False) + else: + P = self.padding_l + # For efficiency, we cut the kernel size and reduce the padding when the kernel is larger than the length + if K > T and P == K - 1: + weight = weight.narrow(2, K - T, T) + K, P = T, T - 1 + # turn the convolution filters into band matrices + weight_expanded = weight.new_zeros(B * H, T, T + K - 1, requires_grad=False) + weight_expanded.as_strided( + (B * H, T, K), (T * (T + K - 1), T + K, 1) + ).copy_(weight) + weight_expanded = weight_expanded.narrow(2, P, T) # B*H x T x T + output = torch.bmm(weight_expanded, x) + output = output.transpose(0, 1).contiguous().view(T, B, C) + return output diff --git a/fairseq/modules/dynamicconv_layer/dynamiconv_cpu.cpp b/fairseq/modules/dynamicconv_layer/dynamiconv_cpu.cpp new file mode 100644 index 0000000000000000000000000000000000000000..8a6af4285da3c40a01383541acf1f455ffc060fb --- /dev/null +++ b/fairseq/modules/dynamicconv_layer/dynamiconv_cpu.cpp @@ -0,0 +1,35 @@ +#include <torch/torch.h> +#include <vector> + +std::vector<float*> dynamicconv_cpu_forward( + float* input, + float* filters, + int padding_l); + +std::vector<float*> dynamicconv_cpu_backward( + float* gradOutput, + int padding_l, + float* input, + float* filters); + +std::vector<float*> dynamicconv_forward( + float* input, + float* filters, + int padding_l) { + + return dynamicconv_cpu_forward(input, filters, padding_l); +} + +std::vector<float*> dynamicconv_backward( + float* gradOutput, + int padding_l, + float* input, + float* filters) { + + return dynamicconv_cpu_backward(gradOutput, padding_l, input, filters); +} + +PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { + m.def("forward", &dynamicconv_forward, "dynamicconv forward (CPU)"); + m.def("backward", &dynamicconv_backward, "dynamicconv backward (CPU)"); +} diff --git a/fairseq/modules/dynamicconv_layer/setup.py b/fairseq/modules/dynamicconv_layer/setup.py new file mode 100644 index 0000000000000000000000000000000000000000..6a21f7e2ee0840a3b251522275a0b32a856951d7 --- /dev/null +++ b/fairseq/modules/dynamicconv_layer/setup.py @@ -0,0 +1,23 @@ +#!/usr/bin/env python3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from setuptools import setup +from torch.utils.cpp_extension import BuildExtension, CUDAExtension + + +setup( + name="dynamicconv_layer", + ext_modules=[ + CUDAExtension( + name="dynamicconv_cuda", + sources=[ + "dynamicconv_cuda.cpp", + "dynamicconv_cuda_kernel.cu", + ], + ), + ], + cmdclass={"build_ext": BuildExtension}, +) diff --git a/fairseq/modules/fairseq_dropout.py b/fairseq/modules/fairseq_dropout.py new file mode 100644 index 0000000000000000000000000000000000000000..3cddca77186f5ddd5cfb9c0ed6def9bafdf3bf1e --- /dev/null +++ b/fairseq/modules/fairseq_dropout.py @@ -0,0 +1,51 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +from typing import List, Optional + +import torch.nn as nn +import torch.nn.functional as F + + +logger = logging.getLogger(__name__) + + +class FairseqDropout(nn.Module): + def __init__(self, p, module_name=None): + super().__init__() + self.p = p + self.module_name = module_name + self.apply_during_inference = False + + def forward(self, x, inplace: bool = False): + if self.p > 0 and (self.training or self.apply_during_inference): + return F.dropout(x, p=self.p, training=True, inplace=inplace) + else: + return x + + def make_generation_fast_( + self, + name: str, + retain_dropout: bool = False, + retain_dropout_modules: Optional[List[str]] = None, + **kwargs + ): + if retain_dropout: + if retain_dropout_modules is not None and self.module_name is None: + logger.warning( + "Cannot enable dropout during inference for module {} " + "because module_name was not set".format(name) + ) + elif ( + retain_dropout_modules is None # if None, apply to all modules + or self.module_name in retain_dropout_modules + ): + logger.info( + "Enabling dropout during inference for module: {}".format(name) + ) + self.apply_during_inference = True + else: + logger.info("Disabling dropout for module: {}".format(name)) diff --git a/fairseq/modules/fp32_group_norm.py b/fairseq/modules/fp32_group_norm.py new file mode 100644 index 0000000000000000000000000000000000000000..d03aac022e30c8c14a600062d1d86429504ba003 --- /dev/null +++ b/fairseq/modules/fp32_group_norm.py @@ -0,0 +1,25 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +Layer norm done in fp32 (for fp16 training) +""" + +import torch.nn as nn +import torch.nn.functional as F + + +class Fp32GroupNorm(nn.GroupNorm): + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + + def forward(self, input): + output = F.group_norm( + input.float(), + self.num_groups, + self.weight.float() if self.weight is not None else None, + self.bias.float() if self.bias is not None else None, + self.eps, + ) + return output.type_as(input) diff --git a/fairseq/modules/gelu.py b/fairseq/modules/gelu.py new file mode 100644 index 0000000000000000000000000000000000000000..a2f1ecff4a3ae3de3eb7d327b9163c46b18a15ed --- /dev/null +++ b/fairseq/modules/gelu.py @@ -0,0 +1,25 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +See "Gaussian Error Linear Units (GELUs)" by Dan Hendrycks and Kevin Gimpel with +the corresponding GitHub repo: https://github.com/hendrycks/GELUs +""" + +import math + +import torch +import torch.nn as nn + + +def gelu_accurate(x): + if not hasattr(gelu_accurate, "_a"): + gelu_accurate._a = math.sqrt(2 / math.pi) + return ( + 0.5 * x * (1 + torch.tanh(gelu_accurate._a * (x + 0.044715 * torch.pow(x, 3)))) + ) + + +def gelu(x: torch.Tensor) -> torch.Tensor: + return torch.nn.functional.gelu(x.float()).type_as(x) diff --git a/fairseq/modules/grad_multiply.py b/fairseq/modules/grad_multiply.py new file mode 100644 index 0000000000000000000000000000000000000000..08d15f55dfda9c61a1cf8641ea31424fe1d97f57 --- /dev/null +++ b/fairseq/modules/grad_multiply.py @@ -0,0 +1,18 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch + + +class GradMultiply(torch.autograd.Function): + @staticmethod + def forward(ctx, x, scale): + ctx.scale = scale + res = x.new(x) + return res + + @staticmethod + def backward(ctx, grad): + return grad * ctx.scale, None diff --git a/fairseq/modules/gumbel_vector_quantizer.py b/fairseq/modules/gumbel_vector_quantizer.py new file mode 100644 index 0000000000000000000000000000000000000000..71134388889d7f224655957256e78fd6c02d72a3 --- /dev/null +++ b/fairseq/modules/gumbel_vector_quantizer.py @@ -0,0 +1,202 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn as nn +import torch.nn.functional as F + + +class GumbelVectorQuantizer(nn.Module): + def __init__( + self, + dim, + num_vars, + temp, + groups, + combine_groups, + vq_dim, + time_first, + activation=nn.GELU(), + weight_proj_depth=1, + weight_proj_factor=1, + ): + """Vector quantization using gumbel softmax + + Args: + dim: input dimension (channels) + num_vars: number of quantized vectors per group + temp: temperature for training. this should be a tuple of 3 elements: (start, stop, decay factor) + groups: number of groups for vector quantization + combine_groups: whether to use the vectors for all groups + vq_dim: dimensionality of the resulting quantized vector + time_first: if true, expect input in BxTxC format, otherwise in BxCxT + activation: what activation to use (should be a module). this is only used if weight_proj_depth is > 1 + weight_proj_depth: number of layers (with activation in between) to project input before computing logits + weight_proj_factor: this is used only if weight_proj_depth is > 1. scales the inner dimensionality of + projections by this factor + """ + super().__init__() + + self.groups = groups + self.combine_groups = combine_groups + self.input_dim = dim + self.num_vars = num_vars + self.time_first = time_first + + assert ( + vq_dim % groups == 0 + ), f"dim {vq_dim} must be divisible by groups {groups} for concatenation" + + var_dim = vq_dim // groups + num_groups = groups if not combine_groups else 1 + + self.vars = nn.Parameter(torch.FloatTensor(1, num_groups * num_vars, var_dim)) + nn.init.uniform_(self.vars) + + if weight_proj_depth > 1: + + def block(input_dim, output_dim): + return nn.Sequential(nn.Linear(input_dim, output_dim), activation) + + inner_dim = self.input_dim * weight_proj_factor + self.weight_proj = nn.Sequential( + *[ + block(self.input_dim if i == 0 else inner_dim, inner_dim) + for i in range(weight_proj_depth - 1) + ], + nn.Linear(inner_dim, groups * num_vars), + ) + else: + self.weight_proj = nn.Linear(self.input_dim, groups * num_vars) + nn.init.normal_(self.weight_proj.weight, mean=0, std=1) + nn.init.zeros_(self.weight_proj.bias) + + if isinstance(temp, str): + import ast + temp = ast.literal_eval(temp) + assert len(temp) == 3, f"{temp}, {len(temp)}" + + self.max_temp, self.min_temp, self.temp_decay = temp + self.curr_temp = self.max_temp + self.codebook_indices = None + + def set_num_updates(self, num_updates): + self.curr_temp = max( + self.max_temp * self.temp_decay ** num_updates, self.min_temp + ) + + def get_codebook_indices(self): + if self.codebook_indices is None: + from itertools import product + + p = [range(self.num_vars)] * self.groups + inds = list(product(*p)) + self.codebook_indices = torch.tensor( + inds, dtype=torch.long, device=self.vars.device + ).flatten() + + if not self.combine_groups: + self.codebook_indices = self.codebook_indices.view( + self.num_vars ** self.groups, -1 + ) + for b in range(1, self.groups): + self.codebook_indices[:, b] += self.num_vars * b + self.codebook_indices = self.codebook_indices.flatten() + return self.codebook_indices + + def codebook(self): + indices = self.get_codebook_indices() + return ( + self.vars.squeeze(0) + .index_select(0, indices) + .view(self.num_vars ** self.groups, -1) + ) + + def sample_from_codebook(self, b, n): + indices = self.get_codebook_indices() + indices = indices.view(-1, self.groups) + cb_size = indices.size(0) + assert ( + n < cb_size + ), f"sample size {n} is greater than size of codebook {cb_size}" + sample_idx = torch.randint(low=0, high=cb_size, size=(b * n,)) + indices = indices[sample_idx] + + z = self.vars.squeeze(0).index_select(0, indices.flatten()).view(b, n, -1) + return z + + def to_codebook_index(self, indices): + res = indices.new_full(indices.shape[:-1], 0) + for i in range(self.groups): + exponent = self.groups - i - 1 + res += indices[..., i] * (self.num_vars ** exponent) + return res + + def forward_idx(self, x): + res = self.forward(x, produce_targets=True) + return res["x"], res["targets"] + + def forward(self, x, produce_targets=False): + + result = {"num_vars": self.num_vars * self.groups} + + if not self.time_first: + x = x.transpose(1, 2) + + bsz, tsz, fsz = x.shape + x = x.reshape(-1, fsz) + x = self.weight_proj(x) + x = x.view(bsz * tsz * self.groups, -1) + + _, k = x.max(-1) + hard_x = ( + x.new_zeros(*x.shape) + .scatter_(-1, k.view(-1, 1), 1.0) + .view(bsz * tsz, self.groups, -1) + ) + hard_probs = torch.mean(hard_x.float(), dim=0) + result["code_perplexity"] = torch.exp( + -torch.sum(hard_probs * torch.log(hard_probs + 1e-7), dim=-1) + ).sum() + + avg_probs = torch.softmax( + x.view(bsz * tsz, self.groups, -1).float(), dim=-1 + ).mean(dim=0) + result["prob_perplexity"] = torch.exp( + -torch.sum(avg_probs * torch.log(avg_probs + 1e-7), dim=-1) + ).sum() + + result["temp"] = self.curr_temp + + if self.training: + x = F.gumbel_softmax(x.float(), tau=self.curr_temp, hard=True).type_as(x) + else: + x = hard_x + + x = x.view(bsz * tsz, -1) + + vars = self.vars + if self.combine_groups: + vars = vars.repeat(1, self.groups, 1) + + if produce_targets: + result["targets"] = ( + x.view(bsz * tsz * self.groups, -1) + .argmax(dim=-1) + .view(bsz, tsz, self.groups) + .detach() + ) + + x = x.unsqueeze(-1) * vars + x = x.view(bsz * tsz, self.groups, self.num_vars, -1) + x = x.sum(-2) + x = x.view(bsz, tsz, -1) + + if not self.time_first: + x = x.transpose(1, 2) # BTC -> BCT + + result["x"] = x + + return result diff --git a/fairseq/modules/kmeans_vector_quantizer.py b/fairseq/modules/kmeans_vector_quantizer.py new file mode 100644 index 0000000000000000000000000000000000000000..040db1e83e775a3bb59d5263d22aae9276a83f22 --- /dev/null +++ b/fairseq/modules/kmeans_vector_quantizer.py @@ -0,0 +1,127 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn as nn +from fairseq.modules import Fp32GroupNorm + + +class KmeansVectorQuantizer(nn.Module): + def __init__( + self, dim, num_vars, groups, combine_groups, vq_dim, time_first, gamma=0.25 + ): + """Vector quantization using straight pass-through estimator (i.e. kmeans) + + Args: + dim: input dimension (channels) + num_vars: number of quantized vectors per group + groups: number of groups for vector quantization + combine_groups: whether to use the vectors for all groups + vq_dim: dimensionality of the resulting quantized vector + time_first: if true, expect input in BxTxC format, otherwise in BxCxT + gamma: commitment loss coefficient + """ + super().__init__() + + self.groups = groups + self.combine_groups = combine_groups + self.input_dim = dim + self.num_vars = num_vars + self.vq_dim = vq_dim + self.time_first = time_first + + assert ( + vq_dim % groups == 0 + ), f"dim {vq_dim} must be divisible by groups {groups} for concatenation" + + self.var_dim = vq_dim // groups + num_groups = groups if not combine_groups else 1 + + self.embedding = nn.Parameter( + 0.01 * torch.randn(num_vars, num_groups, self.var_dim) + ) + self.projection = nn.Sequential( + nn.Conv1d(dim, dim, kernel_size=1, groups=groups, bias=False), + Fp32GroupNorm(groups, dim), + ) + self.gamma = gamma + self.mse_mean = nn.MSELoss(reduction="mean") + + def _pass_grad(self, x, y): + """Manually set gradient for backward pass. + for y = f(x), ensure that during the backward pass, + dL/dy = dL/dx regardless of f(x). + Returns: + y, with the gradient forced to be dL/dy = dL/dx. + """ + + return y.detach() + (x - x.detach()) + + @property + def expand_embedding(self): + if self.combine_groups: + return self.embedding.expand(self.num_vars, self.groups, self.var_dim) + return self.embedding + + def forward_idx(self, x): + res = self.forward(x, produce_targets=True) + return res["x"], res["targets"] + + def forward(self, x, produce_targets=False): + + result = {"num_vars": self.num_vars} + + if self.time_first: + x = x.transpose(1, 2) + + bsz, fsz, tsz = x.shape + + ze = self.projection(x) + ze_ = ze.view(bsz, self.groups, self.var_dim, tsz).permute(0, 3, 1, 2) + d = ( + (ze_.unsqueeze(0) - self.expand_embedding.unsqueeze(1).unsqueeze(1)) + .view(self.num_vars, bsz, tsz, self.groups, -1) + .norm(dim=-1, p=2) + ) + idx = d.argmin(dim=0) + zq = ( + torch.stack( + [ + self.expand_embedding[idx[..., group], group] + for group in range(self.groups) + ], + dim=-2, + ) + .view(bsz, tsz, self.groups * self.var_dim) + .permute(0, 2, 1) + ) + assert ze.shape == zq.shape, (ze.shape, zq.shape) + x = self._pass_grad(ze, zq) + + hard_x = ( + idx.new_zeros(bsz * tsz * self.groups, self.num_vars) + .scatter_(-1, idx.view(-1, 1), 1.0) + .view(bsz * tsz, self.groups, -1) + ) + hard_probs = torch.mean(hard_x.float(), dim=0) + result["code_perplexity"] = torch.exp( + -torch.sum(hard_probs * torch.log(hard_probs + 1e-7), dim=-1) + ).sum() + + if produce_targets: + result["targets"] = idx + + if self.time_first: + x = x.transpose(1, 2) # BCT -> BTC + result["x"] = x + + ze = ze.float() + zq = zq.float() + latent_loss = self.mse_mean(zq, ze.detach()) + commitment_loss = self.mse_mean(ze, zq.detach()) + + result["kmeans_loss"] = latent_loss + self.gamma * commitment_loss + + return result diff --git a/fairseq/modules/layer_drop.py b/fairseq/modules/layer_drop.py new file mode 100644 index 0000000000000000000000000000000000000000..8961d8bcbc492c40c6b30973234416ce5a414f5a --- /dev/null +++ b/fairseq/modules/layer_drop.py @@ -0,0 +1,44 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +LayerDrop as described in https://arxiv.org/abs/1909.11556. +""" + +import torch +import torch.nn as nn + + +class LayerDropModuleList(nn.ModuleList): + """ + A LayerDrop implementation based on :class:`torch.nn.ModuleList`. + + We refresh the choice of which layers to drop every time we iterate + over the LayerDropModuleList instance. During evaluation we always + iterate over all layers. + + Usage:: + + layers = LayerDropList(p=0.5, modules=[layer1, layer2, layer3]) + for layer in layers: # this might iterate over layers 1 and 3 + x = layer(x) + for layer in layers: # this might iterate over all layers + x = layer(x) + for layer in layers: # this might not iterate over any layers + x = layer(x) + + Args: + p (float): probability of dropping out each layer + modules (iterable, optional): an iterable of modules to add + """ + + def __init__(self, p, modules=None): + super().__init__(modules) + self.p = p + + def __iter__(self): + dropout_probs = torch.empty(len(self)).uniform_() + for i, m in enumerate(super().__iter__()): + if not self.training or (dropout_probs[i] > self.p): + yield m diff --git a/fairseq/modules/layer_norm.py b/fairseq/modules/layer_norm.py new file mode 100644 index 0000000000000000000000000000000000000000..234609d9e213a650e0032aaa0ca0462a818bfead --- /dev/null +++ b/fairseq/modules/layer_norm.py @@ -0,0 +1,50 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn as nn +import torch.nn.functional as F + + +try: + from apex.normalization import FusedLayerNorm as _FusedLayerNorm + + has_fused_layernorm = True + + class FusedLayerNorm(_FusedLayerNorm): + @torch.jit.unused + def forward(self, x): + if not x.is_cuda: + return super().forward(x) + else: + with torch.cuda.device(x.device): + return super().forward(x) + + +except ImportError: + has_fused_layernorm = False + + +def LayerNorm(normalized_shape, eps=1e-5, elementwise_affine=True, export=False): + if torch.jit.is_scripting(): + export = True + if not export and torch.cuda.is_available() and has_fused_layernorm: + return FusedLayerNorm(normalized_shape, eps, elementwise_affine) + return torch.nn.LayerNorm(normalized_shape, eps, elementwise_affine) + + +class Fp32LayerNorm(nn.LayerNorm): + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + + def forward(self, input): + output = F.layer_norm( + input.float(), + self.normalized_shape, + self.weight.float() if self.weight is not None else None, + self.bias.float() if self.bias is not None else None, + self.eps, + ) + return output.type_as(input) diff --git a/fairseq/modules/learned_positional_embedding.py b/fairseq/modules/learned_positional_embedding.py new file mode 100644 index 0000000000000000000000000000000000000000..378d0f707183dd344dbb9288dda394b11053acf0 --- /dev/null +++ b/fairseq/modules/learned_positional_embedding.py @@ -0,0 +1,61 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from typing import Dict, Optional + +import torch +import torch.nn as nn +import torch.nn.functional as F +from fairseq import utils +from torch import Tensor + + +class LearnedPositionalEmbedding(nn.Embedding): + """ + This module learns positional embeddings up to a fixed maximum size. + Padding ids are ignored by either offsetting based on padding_idx + or by setting padding_idx to None and ensuring that the appropriate + position ids are passed to the forward function. + """ + + def __init__(self, num_embeddings: int, embedding_dim: int, padding_idx: int): + super().__init__(num_embeddings, embedding_dim, padding_idx) + self.onnx_trace = False + if self.padding_idx is not None: + self.max_positions = self.num_embeddings - self.padding_idx - 1 + else: + self.max_positions = self.num_embeddings + + def forward( + self, + input: Tensor, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + positions: Optional[Tensor] = None, + ): + """Input is expected to be of size [bsz x seqlen].""" + assert (positions is None) or ( + self.padding_idx is None + ), "If positions is pre-computed then padding_idx should not be set." + + if positions is None: + if incremental_state is not None: + # positions is the same for every token when decoding a single step + # Without the int() cast, it doesn't work in some cases when exporting to ONNX + positions = torch.zeros( + (1, 1), device=input.device, dtype=input.dtype + ).fill_(int(self.padding_idx + input.size(1))) + else: + positions = utils.make_positions( + input, self.padding_idx, onnx_trace=self.onnx_trace + ) + return F.embedding( + positions, + self.weight, + self.padding_idx, + self.max_norm, + self.norm_type, + self.scale_grad_by_freq, + self.sparse, + ) diff --git a/fairseq/modules/lightconv_layer/__init__.py b/fairseq/modules/lightconv_layer/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..3b2a99c1227f827768911e5e22e79f6865ffbfd3 --- /dev/null +++ b/fairseq/modules/lightconv_layer/__init__.py @@ -0,0 +1,6 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .lightconv_layer import LightconvLayer # noqa diff --git a/fairseq/modules/lightconv_layer/cuda_function_gen.py b/fairseq/modules/lightconv_layer/cuda_function_gen.py new file mode 100644 index 0000000000000000000000000000000000000000..a25433dd8edae2f0b52d7d0eeeb829cabc6b4b89 --- /dev/null +++ b/fairseq/modules/lightconv_layer/cuda_function_gen.py @@ -0,0 +1,289 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +def gen_forward(): + + kernels = [3, 5, 7, 15, 31, 63, 127, 255] + seqs = [32 * x for x in [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16]] + + head = """ +/** + * Copyright (c) Facebook, Inc. and its affiliates. + * + * This source code is licensed under the MIT license found in the + * LICENSE file in the root directory of this source tree. + */ + +#include "lightconv_cuda.cuh" + +std::vector<at::Tensor> lightconv_cuda_forward(at::Tensor input, at::Tensor filters, int padding_l) { + + at::DeviceGuard g(input.device()); + const auto minibatch = input.size(0); + const auto numFeatures = input.size(1); + const auto sequenceLength = input.size(2); + + const auto numHeads = filters.size(0); + const auto filterSize = filters.size(1); + + const auto numFiltersInBlock = numFeatures / numHeads; + + const dim3 blocks(minibatch, numFeatures); + + auto output = at::zeros_like(input); + auto stream = at::cuda::getCurrentCUDAStream(); +""" + + sequence_if = """ + if (sequenceLength <= {seq}) {{ + switch(filterSize) {{ +""" + + case_k = """ + case {k}: +""" + + main_block = """ + if (padding_l == {pad}) {{ + AT_DISPATCH_FLOATING_TYPES_AND_HALF(input.scalar_type(), "lightconv_forward", ([&] {{ + lightconv_forward_kernel<{k}, {b_size}, {pad}, scalar_t> + <<<blocks, {b_size}, 0, stream>>>( + input.data<scalar_t>(), + filters.data<scalar_t>(), + minibatch, + sequenceLength, + numFeatures, + numFiltersInBlock, + output.data<scalar_t>()); + }})); + }} else +""" + + bad_padding = """ + { + std::cout << "WARNING: Unsupported padding size - skipping forward pass" << std::endl; + } + break; +""" + + bad_filter = """ + default: + std::cout << "WARNING: Unsupported filter length passed - skipping forward pass" << std::endl; + } +""" + + con_else = """ + } else +""" + + final_else = """ + { + switch(filterSize) { +""" + + final_return = """ + } + + return {output}; +} +""" + + with open("lightconv_cuda_forward.cu", "w") as forward: + forward.write(head) + for seq in seqs: + forward.write(sequence_if.format(seq=seq)) + for k in kernels: + forward.write(case_k.format(k=k)) + for pad in [k // 2, k - 1]: + forward.write(main_block.format(k=k, b_size=seq, pad=pad)) + forward.write(bad_padding) + forward.write(bad_filter) + forward.write(con_else) + + forward.write(final_else) + for k in kernels: + forward.write(case_k.format(k=k)) + for pad in [k // 2, k - 1]: + forward.write(main_block.format(k=k, b_size=seq, pad=pad)) + forward.write(bad_padding) + forward.write(bad_filter) + forward.write(final_return) + + +def gen_backward(): + + head = """ +/** + * Copyright (c) Facebook, Inc. and its affiliates. + * + * This source code is licensed under the MIT license found in the + * LICENSE file in the root directory of this source tree. + */ + +#include "lightconv_cuda.cuh" + +std::vector<at::Tensor> lightconv_cuda_backward( + at::Tensor gradOutput, + int padding_l, + at::Tensor input, + at::Tensor filters) { + + // gradWrtInput + const int minibatch = input.size(0); + const int numFeatures = input.size(1); + const int sequenceLength = input.size(2); + + const int numHeads = filters.size(0); + const int filterSize = filters.size(1); + + const dim3 gradBlocks(minibatch, numFeatures); + const dim3 weightGradFirstpassShortBlocks(minibatch, numHeads); + const dim3 weightGradSecondpassBlocks(numHeads, filterSize); + + const int numFiltersInBlock = numFeatures / numHeads; + + auto gradInput = at::zeros_like(input); + auto gradFilters = at::zeros_like(filters); + + at::DeviceGuard g(input.device()); + auto stream = at::cuda::getCurrentCUDAStream(); + + switch(filterSize) { +""" + + sequence_if = """ + if (sequenceLength <= {seq}) {{ +""" + + case_k = """ + case {k}: +""" + + main_block = """ + if (padding_l == {p}) {{ + AT_DISPATCH_FLOATING_TYPES_AND_HALF(input.scalar_type(), "lightconv_backward", ([&] {{ + lightconv_grad_wrt_input_kernel<{k}, {b_size}, {p}, scalar_t> + <<<gradBlocks, {b_size}, 0, stream>>>( + gradOutput.data<scalar_t>(), + filters.data<scalar_t>(), + minibatch, + sequenceLength, + numFeatures, + numFiltersInBlock, + gradInput.data<scalar_t>()); + +""" + + weight_grad_short = """ + at::Tensor tempSumGradFilters = at::zeros({{minibatch, numHeads, filterSize}}, input.options().dtype(at::kFloat)); + lightconv_grad_wrt_weights_firstpass_short_kernel<{k}, {b_size}, {p}, scalar_t> + <<<weightGradFirstpassShortBlocks, {b_size}, 0, stream>>>( + input.data<scalar_t>(), + gradOutput.data<scalar_t>(), + minibatch, + sequenceLength, + numFeatures, + numFiltersInBlock, + numHeads, + tempSumGradFilters.data<float>() + ); + + lightconv_grad_wrt_weights_secondpass_short_kernel<{k}, {b_size}, scalar_t> + <<<weightGradSecondpassBlocks, {b_size}, 0, stream>>>( + tempSumGradFilters.data<float>(), + minibatch, + numFiltersInBlock, + gradFilters.data<scalar_t>() + ); + }})); + }} else +""" + + weight_grad = """ + at::Tensor tempSumGradFilters = at::zeros({{minibatch, numFeatures, filterSize}}, input.options().dtype(at::kFloat)); + lightconv_grad_wrt_weights_firstpass_kernel<{k}, {b_size}, {p}, scalar_t> + <<<gradBlocks, {b_size}, 0, stream>>>( + input.data<scalar_t>(), + gradOutput.data<scalar_t>(), + minibatch, + sequenceLength, + numFeatures, + numFiltersInBlock, + tempSumGradFilters.data<float>() + ); + + lightconv_grad_wrt_weights_secondpass_kernel<{k}, {b_size}, scalar_t> + <<<weightGradSecondpassBlocks, {b_size}, 0, stream>>>( + tempSumGradFilters.data<float>(), + minibatch, + numFiltersInBlock, + gradFilters.data<scalar_t>() + ); + }})); + }} else +""" + + bad_padding = """ + { + std::cout << "WARNING: Unsupported padding size - skipping backward pass" << std::endl; + } +""" + + breakout = """ + break; +""" + + bad_filter = """ + default: + std::cout << "WARNING: Unsupported filter length passed - skipping backward pass" << std::endl; +""" + + con_else = """ + } else +""" + + final_else = """ + { + switch(filterSize) { +""" + + last_return = """ + } + return {gradInput, gradFilters}; +} +""" + + kernels = [3, 5, 7, 15, 31, 63, 127, 255] + seqs = [32 * x for x in [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16]] + thresh = [32, 32, 64, 128, 256, -1, -1, -1] + max_mem = [-1, -1, -1, -1, -1, 192, 96, 64] + + with open("lightconv_cuda_backward.cu", "w") as backward: + backward.write(head) + for (k, t, mem) in zip(kernels, thresh, max_mem): + backward.write(case_k.format(k=k)) + for seq in seqs: + if (t == -1 or seq <= t) and (mem == -1 or seq < mem): + backward.write(sequence_if.format(seq=seq)) + for p in [k // 2, k - 1]: + backward.write(main_block.format(k=k, b_size=seq, p=p)) + backward.write(weight_grad_short.format(k=k, b_size=seq, p=p)) + backward.write(bad_padding) + else: + for p in [k // 2, k - 1]: + backward.write(main_block.format(k=k, b_size=32, p=p)) + backward.write(weight_grad.format(k=k, b_size=32, p=p)) + backward.write(bad_padding) + backward.write(breakout) + break + backward.write(con_else) + backward.write(bad_filter) + backward.write(last_return) + + +if __name__ == "__main__": + gen_forward() + gen_backward() diff --git a/fairseq/modules/lightconv_layer/lightconv_cuda.cpp b/fairseq/modules/lightconv_layer/lightconv_cuda.cpp new file mode 100644 index 0000000000000000000000000000000000000000..4bf6b5ad365d604bd91eda384bb422857b640744 --- /dev/null +++ b/fairseq/modules/lightconv_layer/lightconv_cuda.cpp @@ -0,0 +1,54 @@ +/** + * Copyright (c) Facebook, Inc. and its affiliates. + * + * This source code is licensed under the MIT license found in the + * LICENSE file in the root directory of this source tree. + */ + +#include <torch/extension.h> +#include <vector> + +std::vector<at::Tensor> lightconv_cuda_forward( + at::Tensor input, + at::Tensor filters, + int padding_l); + +std::vector<at::Tensor> lightconv_cuda_backward( + at::Tensor gradOutput, + int padding_l, + at::Tensor input, + at::Tensor filters); + + +#define CHECK_CUDA(x) AT_ASSERTM(x.type().is_cuda(), #x " must be a CUDA tensor") +#define CHECK_CONTIGUOUS(x) AT_ASSERTM(x.is_contiguous(), #x " must be contiguous") +#define CHECK_INPUT(x) CHECK_CUDA(x); CHECK_CONTIGUOUS(x) + +std::vector<at::Tensor> lightconv_forward( + at::Tensor input, + at::Tensor filters, + int padding_l) { + + CHECK_INPUT(input); + CHECK_INPUT(filters); + + return lightconv_cuda_forward(input, filters, padding_l); +} + +std::vector<at::Tensor> lightconv_backward( + at::Tensor gradOutput, + int padding_l, + at::Tensor input, + at::Tensor filters) { + + CHECK_INPUT(gradOutput); + CHECK_INPUT(input); + CHECK_INPUT(filters); + + return lightconv_cuda_backward(gradOutput, padding_l, input, filters); +} + +PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { + m.def("forward", &lightconv_forward, "lighconv forward (CUDA)"); + m.def("backward", &lightconv_backward, "lighconv backward (CUDA)"); +} diff --git a/fairseq/modules/lightconv_layer/lightconv_cuda.cuh b/fairseq/modules/lightconv_layer/lightconv_cuda.cuh new file mode 100644 index 0000000000000000000000000000000000000000..3cae57b68fc96872a5047a7a0d081b78456e8fae --- /dev/null +++ b/fairseq/modules/lightconv_layer/lightconv_cuda.cuh @@ -0,0 +1,83 @@ +/** + * Copyright (c) Facebook, Inc. and its affiliates. + * + * This source code is licensed under the MIT license found in the + * LICENSE file in the root directory of this source tree. + */ + +#include <ATen/ATen.h> +#include <c10/cuda/CUDAStream.h> + +#include <cuda.h> +#include <cuda_runtime.h> + +#include <algorithm> +#include <functional> +#include <iostream> +#include <stdexcept> +#include <utility> +#include <vector> + +#include <stdlib.h> +#include <assert.h> + +#define SHFL_MASK 0xffffffff + +template<int FS, int SB, int padding_l, typename scalar_t> +__global__ +void lightconv_forward_kernel(const scalar_t* input, + const scalar_t* filters, + int minibatch, int sequenceLength, + int numFeatures, int numFiltersInBlock, + scalar_t* output); + +template<int FS, int SB, int padding_l, typename scalar_t> +__global__ +void lightconv_grad_wrt_input_kernel( + const scalar_t* input, + const scalar_t* filters, + int minibatch, + int sequenceLength, + int numFeatures, + int numFiltersInBlock, + scalar_t* output); + +template<int FS, int SB, int padding_l, typename scalar_t> +__global__ +void lightconv_grad_wrt_weights_firstpass_short_kernel( + const scalar_t* input, + const scalar_t* gradInput, + int minibatch, + int sequenceLength, + int numFeatures, + int numFiltersInBlock, + int numHeads, + float* output); + +template<int FS, int SB, typename scalar_t> +__global__ +void lightconv_grad_wrt_weights_secondpass_short_kernel( + const float* input, + const int minibatch, + const int numFiltersInBlock, + scalar_t* output); + +template<int FS, int SB, int padding_l, typename scalar_t> +__global__ +void lightconv_grad_wrt_weights_firstpass_kernel( + const scalar_t* input, + const scalar_t* gradInput, + int minibatch, + int sequenceLength, + int numFeatures, + int numFiltersInBlock, + float* output); + +template<int FS, int SB, typename scalar_t> +__global__ +void lightconv_grad_wrt_weights_secondpass_kernel( + const float* input, + const int minibatch, + const int numFiltersInBlock, + scalar_t* output); + diff --git a/fairseq/modules/lightconv_layer/lightconv_cuda_kernel.cu b/fairseq/modules/lightconv_layer/lightconv_cuda_kernel.cu new file mode 100644 index 0000000000000000000000000000000000000000..8ee83a56c89754c2abbe717b269d07ca9e64eef2 --- /dev/null +++ b/fairseq/modules/lightconv_layer/lightconv_cuda_kernel.cu @@ -0,0 +1,375 @@ +/** + * Copyright (c) Facebook, Inc. and its affiliates. + * + * This source code is licensed under the MIT license found in the + * LICENSE file in the root directory of this source tree. + */ + +#include "lightconv_cuda.cuh" +#include "lightconv_cuda_forward.cu" +#include "lightconv_cuda_backward.cu" +#include "../cuda_utils.cu" + +template<int FS, int SB, int padding_l, typename scalar_t> +__global__ +void lightconv_forward_kernel(const scalar_t* input, + const scalar_t* filters, + int minibatch, int sequenceLength, + int numFeatures, int numFiltersInBlock, + scalar_t* output) { + + const int tid = threadIdx.x; + const int batchIdx = blockIdx.x; + const int featureIdx = blockIdx.y; + const int filterIdx = featureIdx / numFiltersInBlock; + + const int IOOffset = numFeatures * sequenceLength * batchIdx + featureIdx * sequenceLength; + const scalar_t* inputFeature = &input[IOOffset]; + scalar_t* outputFeature = &output[IOOffset]; + const scalar_t* inputFilter = &filters[filterIdx * FS]; + + assert(blockDim.x == SB); + + scalar_t filter[FS]; + #pragma unroll + for (int i = 0; i < FS; ++i) { + filter[i] = inputFilter[i]; + } + + __shared__ scalar_t temp[SB + FS]; + zeroSharedMem<FS, SB, padding_l>(temp); + + const int numIterations = divUp<int, int>(sequenceLength, SB); + + for (int i = 0; i < numIterations; ++i) { + // Read input into shared memory + const int inputOffset = i * SB; + + load_input_to_shared<FS, SB, padding_l>(inputFeature, inputOffset, sequenceLength, + i, numIterations, (numIterations == 1), temp); + + __syncthreads(); + + scalar_t out = 0; + #pragma unroll + for (int j = 0; j < FS; ++j) { + out += filter[j] * temp[tid + j]; + } + + // Write output + const int outputOffset = inputOffset; + if ((outputOffset + tid) < sequenceLength) { + outputFeature[outputOffset + tid] = out; + } + + __syncthreads(); + } +} + +template<int FS, int SB, int padding_l, typename scalar_t> +__global__ +void lightconv_grad_wrt_input_kernel( + const scalar_t* input, + const scalar_t* filters, + int minibatch, + int sequenceLength, + int numFeatures, + int numFiltersInBlock, + scalar_t* output) { + + // input grad kernel is similar to forward kernel + const int tid = threadIdx.x; + const int batchIdx = blockIdx.x; + const int featureIdx = blockIdx.y; + const int filterIdx = featureIdx / numFiltersInBlock; + + const int IOOffset = numFeatures * sequenceLength * batchIdx + featureIdx * sequenceLength; + const scalar_t* inputFeature = &input[IOOffset]; + scalar_t* outputFeature = &output[IOOffset]; + const scalar_t* inputFilter = &filters[filterIdx * FS]; + + assert(blockDim.x == SB); + + scalar_t filter[FS]; + + // The only change is loading the filter in reverse + #pragma unroll + for (int i = 0; i < FS; ++i) { + filter[i] = inputFilter[FS - i - 1]; + } + + __shared__ scalar_t temp[SB + FS]; + const int padding = FS - padding_l - 1; + zeroSharedMem<FS, SB, padding>(temp); + + __syncthreads(); + + const int numIterations = divUp<int, int>(sequenceLength, SB); + + for (int i = 0; i < numIterations; ++i) { + // Read input into shared memory + const int inputOffset = i * SB; + + load_input_to_shared<FS, SB, padding>(inputFeature, inputOffset, sequenceLength, + i, numIterations, false, temp); + + __syncthreads(); + + scalar_t out = 0; + #pragma unroll + for (int j = 0; j < FS; ++j) { + out += filter[j] * temp[tid + j]; + } + + // Write output + const int outputOffset = inputOffset; + if ((outputOffset + tid) < sequenceLength) { + outputFeature[outputOffset + tid] = out; + } + + __syncthreads(); + } +} + +// This is by far the most expensive kernel in terms of time taken. +// Can be 16x slower than the forward or grad_wrt_input when filter size is 31 +template<int FS, int SB, int padding_l, typename scalar_t> +__global__ +void lightconv_grad_wrt_weights_firstpass_short_kernel( + const scalar_t* input, + const scalar_t* gradInput, + int minibatch, + int sequenceLength, + int numFeatures, + int numFiltersInBlock, + int numHeads, + float* output) { + + const int tid = threadIdx.x; + const int batchIdx = blockIdx.x; + const int filterIdx = blockIdx.y; + + const int numIterations = divUp<int, int>(sequenceLength, SB); + + float* tempOutputGradWeight = &output[filterIdx * FS * minibatch]; + + assert(blockDim.x == SB); + + __shared__ scalar_t tempInput[SB + FS]; + __shared__ scalar_t tempGradInput[SB + FS]; + + // local weight accumulation + float accumWeights[FS]; + + // Initialize memory + for (int i = 0; i < FS; ++i) { + accumWeights[i] = float(0.0); + } + + + // loop over each sequence within filterblock + for (int idxInFilterBlock = 0; idxInFilterBlock < numFiltersInBlock; ++idxInFilterBlock) { + + const int featureOffset = batchIdx * numFeatures * sequenceLength + (filterIdx * numFiltersInBlock + idxInFilterBlock) * sequenceLength; + const scalar_t* inputFeature = &input[featureOffset]; + const scalar_t* gradInputFeature = &gradInput[featureOffset]; + + zeroSharedMem<FS, SB, padding_l>(tempInput); + zeroSharedMem<FS, SB, (FS/2)>(tempGradInput); + __syncthreads(); + + for (int i = 0; i < numIterations; ++i) { + + const int inputOffset = i * SB; + + load_input_to_shared<FS, SB, padding_l>(inputFeature, inputOffset, sequenceLength, + i, numIterations, false, tempInput); + load_input_to_shared<FS, SB, (FS/2)>(gradInputFeature, inputOffset, sequenceLength, + i, numIterations, false, tempGradInput); + + __syncthreads(); + + const int gradIndex = (FS/2) + tid; + scalar_t tempGrad = tempGradInput[gradIndex]; + + #pragma unroll + for (int j = 0; j < FS; j++) { + const int inputIndex = tid + j; + accumWeights[j] += tempInput[inputIndex] * tempGrad; + } + + __syncthreads(); + + } + + } + + // Row-major sum + for (int filterWeightIdx = 0; filterWeightIdx < FS; ++filterWeightIdx) { + + float temp; + if (tid < sequenceLength) { + temp = accumWeights[filterWeightIdx]; + } else { + temp = float(0.0); + } + + const int outputOffset = filterWeightIdx * minibatch + batchIdx; + + temp = blockReduce(temp); + + if (tid == 0) { + tempOutputGradWeight[outputOffset] = temp; + } + } +} + +template<int FS, int SB, typename scalar_t> +__global__ +void lightconv_grad_wrt_weights_secondpass_short_kernel( + const float* input, + const int minibatch, + const int numFiltersInBlock, + scalar_t* output) { + + assert(blockDim.x == SB); + + const int tid = threadIdx.x; + + const int filterIdx = blockIdx.x; + const int filterWeightIdx = blockIdx.y; + + const int inputOffset = filterIdx * FS * minibatch + + filterWeightIdx * minibatch; + const float* tempInput = &input[inputOffset]; + + // read into shared memory for reduction + int readIndex = tid; + + float sum = 0.0; + while (readIndex < minibatch) { + sum += tempInput[readIndex]; + readIndex += SB; + } + + float temp = blockReduce(sum); + + if (tid == 0) { + output[blockIdx.x * FS + blockIdx.y] = temp; + } +} + +// This is by far the most expensive kernel in terms of time taken. +// Can be 16x slower than the forward or grad_wrt_input when filter size is 31 +template<int FS, int SB, int padding_l, typename scalar_t> +__global__ +void lightconv_grad_wrt_weights_firstpass_kernel( + const scalar_t* input, + const scalar_t* gradInput, + int minibatch, + int sequenceLength, + int numFeatures, + int numFiltersInBlock, + float* output) { + + assert(blockDim.x == SB); + + const int tid = threadIdx.x; + const int batchIdx = blockIdx.x; + const int featureIdx = blockIdx.y; + const int filterIdx = featureIdx / numFiltersInBlock; + const int idxInFilterBlock = featureIdx % numFiltersInBlock; + + const int numIterations = divUp<int, int>(sequenceLength, SB); + + float temp; + + __shared__ scalar_t tempInput[SB + FS]; + __shared__ scalar_t tempGradInput[SB + FS]; + zeroSharedMem<FS, SB, padding_l>(tempInput); + zeroSharedMem<FS, SB, (FS/2)>(tempGradInput); + __syncthreads(); + + float accumWeights[FS]; + + for (int i = 0; i < FS; ++i) { + accumWeights[i] = float(0.0); + } + + const int IOOffset = batchIdx * numFeatures * sequenceLength + featureIdx * sequenceLength; + const scalar_t* inputFeature = &input[IOOffset]; + const scalar_t* gradInputFeature = &gradInput[IOOffset]; + float* tempOutputGradWeight = &output[filterIdx * FS * minibatch * numFiltersInBlock]; + + for (int i = 0; i < numIterations; ++i) { + const int inputOffset = i * SB; + + load_input_to_shared<FS, SB, padding_l>(inputFeature, inputOffset, sequenceLength, + i, numIterations, false, tempInput); + load_input_to_shared<FS, SB, (FS/2)>(gradInputFeature, inputOffset, sequenceLength, + i, numIterations, false, tempGradInput); + __syncthreads(); + + #pragma unroll + for (int j = 0; j < FS; ++j) { + accumWeights[j] += tempInput[tid + j] * tempGradInput[tid + (FS/2)]; + } + + __syncthreads(); + } + + // Row-major sum + for (int filterWeightIdx = 0; filterWeightIdx < FS; ++filterWeightIdx) { + + // Write to shared memory before reduction + if (tid < sequenceLength) { + temp = accumWeights[filterWeightIdx]; + } else { + temp = float(0.0); + } + + temp = blockReduce(temp); + + const int outputOffset = filterWeightIdx * minibatch * numFiltersInBlock + + batchIdx * numFiltersInBlock + + idxInFilterBlock; + + if (tid == 0) { + tempOutputGradWeight[outputOffset] = temp; + } + } +} + +template<int FS, int SB, typename scalar_t> +__global__ +void lightconv_grad_wrt_weights_secondpass_kernel( + const float* input, + const int minibatch, + const int numFiltersInBlock, + scalar_t* output) { + + assert(blockDim.x == SB); + const int tid = threadIdx.x; + + // What is the id within a minibatch + const int filterIdx = blockIdx.x; + const int filterWeightIdx = blockIdx.y; + + const int inputOffset = filterIdx * FS * minibatch * numFiltersInBlock + + filterWeightIdx * minibatch * numFiltersInBlock; + const float* tempInput = &input[inputOffset]; + + int readIndex = tid; + + float sum = float(0.0); + while (readIndex < (minibatch * numFiltersInBlock)) { + sum += tempInput[readIndex]; + readIndex += SB; + } + + float temp = blockReduce(sum); + + if (tid == 0) { + output[blockIdx.x * FS + blockIdx.y] = temp; + } +} diff --git a/fairseq/modules/lightconv_layer/lightconv_layer.py b/fairseq/modules/lightconv_layer/lightconv_layer.py new file mode 100644 index 0000000000000000000000000000000000000000..e7e597f4749c591b057d776aacec39b44d99c037 --- /dev/null +++ b/fairseq/modules/lightconv_layer/lightconv_layer.py @@ -0,0 +1,137 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import lightconv_cuda +import torch +import torch.nn.functional as F +from fairseq import utils +from fairseq.incremental_decoding_utils import with_incremental_state +from fairseq.modules.fairseq_dropout import FairseqDropout +from torch import nn +from torch.autograd import Function + + +class lightconvFunction(Function): + @staticmethod + def forward(ctx, x, weights, padding_l): + ctx.padding_l = padding_l + outputs = lightconv_cuda.forward(x, weights, padding_l) + variables = [x, weights] + ctx.save_for_backward(*variables) + return outputs[0] + + @staticmethod + def backward(ctx, grad_output): + outputs = lightconv_cuda.backward( + grad_output.contiguous(), ctx.padding_l, *ctx.saved_tensors + ) + grad_input, grad_weights = outputs + return grad_input, grad_weights, None + + +@with_incremental_state +class LightconvLayer(nn.Module): + def __init__( + self, + input_size, + kernel_size=1, + padding_l=None, + weight_softmax=False, + num_heads=1, + weight_dropout=0.0, + bias=False, + ): + super(LightconvLayer, self).__init__() + self.input_size = input_size + self.kernel_size = kernel_size + self.padding_l = padding_l + self.num_heads = num_heads + self.weight_softmax = weight_softmax + self.weight_dropout_module = FairseqDropout( + weight_dropout, module_name=self.__class__.__name__ + ) + + self.weight = nn.Parameter(torch.Tensor(num_heads, kernel_size)) + if bias: + self.bias = nn.Parameter(torch.Tensor(input_size)) + else: + self.bias = None + self.reset_parameters() + + def upgrade_state_dict_named(self, state_dict, name): + prefix = name + "." if name != "" else "" + for k, v in state_dict.items(): + if k.endswith(prefix + "weight"): + if v.dim() == 3 and v.size(1) == 1: + state_dict[k] = v.squeeze(1) + + def reset_parameters(self): + nn.init.xavier_uniform_(self.weight) + if self.bias is not None: + nn.init.constant_(self.bias, 0.0) + + def forward(self, x, incremental_state=None): + + # during inference time, incremental BMM is faster + if incremental_state is not None: + T, B, C = x.size() + K, H = self.kernel_size, self.num_heads + R = C // H + input_buffer = self._get_input_buffer(incremental_state) + if input_buffer is None: + input_buffer = x.new() + x_unfold = torch.cat([input_buffer, x.unsqueeze(3)], dim=3) + if self.kernel_size > 1: + self._set_input_buffer( + incremental_state, x_unfold[:, :, :, -self.kernel_size + 1 :] + ) + x_unfold = x_unfold.view(T * B * H, R, -1) + + weight = self.weight + if self.weight_softmax: + weight = F.softmax(weight.float(), dim=1).type_as(weight) + + weight = weight[:, -x_unfold.size(2) :] + + K = weight.size(1) + + weight = ( + weight.view(1, H, K) + .expand(T * B, H, K) + .contiguous() + .view(T * B * H, K, 1) + ) + + weight = self.weight_dropout_module(weight) + output = torch.bmm(x_unfold, weight) # T*B*H x R x 1 + output = output.view(T, B, C) + return output + + # during training time, use CUDA kernel + else: + x = x.permute(1, 2, 0).contiguous() + weight = self.weight + if self.weight_softmax: + weight = F.softmax(self.weight, -1) + if self.weight_dropout_module.p: + weight = self.weight_dropout_module(weight) + return lightconvFunction.apply(x, weight, self.padding_l).permute(2, 0, 1) + + def reorder_incremental_state(self, incremental_state, new_order): + input_buffer = self._get_input_buffer(incremental_state) + if input_buffer is not None: + input_buffer = input_buffer.index_select(1, new_order) + self._set_input_buffer(incremental_state, input_buffer) + + def _get_input_buffer(self, incremental_state): + return utils.get_incremental_state(self, incremental_state, "input_buffer") + + def _set_input_buffer(self, incremental_state, new_buffer): + return utils.set_incremental_state( + self, incremental_state, "input_buffer", new_buffer + ) + + def half(self): + return self._apply(lambda t: t.half() if t.is_floating_point() else t) diff --git a/fairseq/modules/lightconv_layer/setup.py b/fairseq/modules/lightconv_layer/setup.py new file mode 100644 index 0000000000000000000000000000000000000000..052635be79b466d0ad56cf5cf607bd10c2297ecf --- /dev/null +++ b/fairseq/modules/lightconv_layer/setup.py @@ -0,0 +1,23 @@ +#!/usr/bin/env python3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from setuptools import setup +from torch.utils.cpp_extension import BuildExtension, CUDAExtension + + +setup( + name="lightconv_layer", + ext_modules=[ + CUDAExtension( + "lightconv_cuda", + [ + "lightconv_cuda.cpp", + "lightconv_cuda_kernel.cu", + ], + ), + ], + cmdclass={"build_ext": BuildExtension}, +) diff --git a/fairseq/modules/lightweight_convolution.py b/fairseq/modules/lightweight_convolution.py new file mode 100644 index 0000000000000000000000000000000000000000..ec11a9507951c9e8f3564753841dd9c74a4900e0 --- /dev/null +++ b/fairseq/modules/lightweight_convolution.py @@ -0,0 +1,310 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn as nn +import torch.nn.functional as F +from fairseq import utils +from fairseq.incremental_decoding_utils import with_incremental_state +from fairseq.modules.fairseq_dropout import FairseqDropout +from fairseq.modules.unfold import unfold1d + + +def LightweightConv( + input_size, + kernel_size=1, + padding_l=None, + num_heads=1, + weight_dropout=0.0, + weight_softmax=False, + bias=False, +): + if torch.cuda.is_available(): + try: + from fairseq.modules.lightconv_layer import LightconvLayer + + return LightconvLayer( + input_size, + kernel_size=kernel_size, + padding_l=padding_l, + num_heads=num_heads, + weight_dropout=weight_dropout, + weight_softmax=weight_softmax, + bias=bias, + ) + except ImportError as e: + print(e) + return LightweightConv1dTBC( + input_size, + kernel_size=kernel_size, + padding_l=padding_l, + num_heads=num_heads, + weight_dropout=weight_dropout, + weight_softmax=weight_softmax, + bias=bias, + ) + + +class LightweightConv1d(nn.Module): + """Lightweight Convolution assuming the input is BxCxT + This is just an example that explains LightConv clearer than the TBC version. + We don't use this module in the model. + + Args: + input_size: # of channels of the input and output + kernel_size: convolution channels + padding: padding + num_heads: number of heads used. The weight is of shape + `(num_heads, 1, kernel_size)` + weight_softmax: normalize the weight with softmax before the convolution + + Shape: + Input: BxCxT, i.e. (batch_size, input_size, timesteps) + Output: BxCxT, i.e. (batch_size, input_size, timesteps) + + Attributes: + weight: the learnable weights of the module of shape + `(num_heads, 1, kernel_size)` + bias: the learnable bias of the module of shape `(input_size)` + """ + + def __init__( + self, + input_size, + kernel_size=1, + padding=0, + num_heads=1, + weight_softmax=False, + bias=False, + weight_dropout=0.0, + ): + super().__init__() + self.input_size = input_size + self.kernel_size = kernel_size + self.num_heads = num_heads + self.padding = padding + self.weight_softmax = weight_softmax + self.weight = nn.Parameter(torch.Tensor(num_heads, 1, kernel_size)) + + if bias: + self.bias = nn.Parameter(torch.Tensor(input_size)) + else: + self.bias = None + self.weight_dropout_module = FairseqDropout( + weight_dropout, module_name=self.__class__.__name__ + ) + self.reset_parameters() + + def reset_parameters(self): + nn.init.xavier_uniform_(self.weight) + if self.bias is not None: + nn.init.constant_(self.bias, 0.0) + + def forward(self, input): + """ + input size: B x C x T + output size: B x C x T + """ + B, C, T = input.size() + H = self.num_heads + + weight = self.weight + if self.weight_softmax: + weight = F.softmax(weight, dim=-1) + + weight = self.weight_dropout_module(weight) + # Merge every C/H entries into the batch dimension (C = self.input_size) + # B x C x T -> (B * C/H) x H x T + # One can also expand the weight to C x 1 x K by a factor of C/H + # and do not reshape the input instead, which is slow though + input = input.view(-1, H, T) + output = F.conv1d(input, weight, padding=self.padding, groups=self.num_heads) + output = output.view(B, C, T) + if self.bias is not None: + output = output + self.bias.view(1, -1, 1) + + return output + + +@with_incremental_state +class LightweightConv1dTBC(nn.Module): + """Lightweight Convolution assuming the input is TxBxC + Args: + input_size: # of channels of the input + kernel_size: convolution channels + padding_l: padding to the left when using "same" padding + num_heads: number of heads used. The weight is of shape (num_heads, 1, kernel_size) + weight_dropout: the drop rate of the DropConnect to drop the weight + weight_softmax: normalize the weight with softmax before the convolution + bias: use bias + + Shape: + Input: TxBxC, i.e. (timesteps, batch_size, input_size) + Output: TxBxC, i.e. (timesteps, batch_size, input_size) + + Attributes: + weight: the learnable weights of the module of shape + `(num_heads, 1, kernel_size)` + bias: the learnable bias of the module of shape `(input_size)` + """ + + def __init__( + self, + input_size, + kernel_size=1, + padding_l=None, + num_heads=1, + weight_dropout=0.0, + weight_softmax=False, + bias=False, + ): + super().__init__() + self.input_size = input_size + self.kernel_size = kernel_size + self.padding_l = padding_l + self.num_heads = num_heads + self.weight_dropout_module = FairseqDropout( + weight_dropout, module_name=self.__class__.__name__ + ) + self.weight_softmax = weight_softmax + + self.weight = nn.Parameter(torch.Tensor(num_heads, 1, kernel_size)) + if bias: + self.bias = nn.Parameter(torch.Tensor(input_size)) + else: + self.bias = None + + self.reset_parameters() + self.onnx_trace = False + + def reset_parameters(self): + nn.init.xavier_uniform_(self.weight) + if self.bias is not None: + nn.init.constant_(self.bias, 0.0) + + def forward(self, x, incremental_state=None, unfold=False): + """Assuming the input, x, of the shape T x B x C and producing an output in the shape T x B x C + args: + x: Input of shape T x B x C, i.e. (timesteps, batch_size, input_size) + incremental_state: A dict to keep the state + unfold: unfold the input or not. If not, we use the matrix trick instead + """ + unfold = unfold or (incremental_state is not None) + + if unfold: + output = self._forward_unfolded(x, incremental_state) + else: + output = self._forward_expanded(x, incremental_state) + + if self.bias is not None: + output = output + self.bias.view(1, 1, -1) + return output + + def prepare_for_onnx_export_(self): + self.onnx_trace = True + + def _forward_unfolded(self, x, incremental_state): + """The conventional implementation of convolutions. + Unfolding the input by having a window shifting to the right.""" + T, B, C = x.size() + K, H = self.kernel_size, self.num_heads + R = C // H + assert R * H == C == self.input_size + + weight = self.weight.view(H, K) + if incremental_state is not None: + input_buffer = self._get_input_buffer(incremental_state) + if input_buffer is None: + input_buffer = x.new() + x_unfold = torch.cat([input_buffer, x.unsqueeze(3)], dim=3) + if self.kernel_size > 1: + self._set_input_buffer( + incremental_state, x_unfold[:, :, :, -self.kernel_size + 1 :] + ) + x_unfold = x_unfold.view(T * B * H, R, -1) + else: + # unfold the input: T x B x C --> T' x B x C x K + x_unfold = unfold1d(x, self.kernel_size, self.padding_l, 0) + x_unfold = x_unfold.view(T * B * H, R, K) + + if self.weight_softmax: + weight = utils.softmax(weight, dim=1, onnx_trace=self.onnx_trace).type_as( + weight + ) + + if incremental_state is not None: + weight = weight[:, -x_unfold.size(2) :] + K = weight.size(1) + + weight = ( + weight.view(1, H, K).expand(T * B, H, K).contiguous().view(T * B * H, K, 1) + ) + + weight = self.weight_dropout_module(weight) + output = torch.bmm(x_unfold, weight) # T*B*H x R x 1 + output = output.view(T, B, C) + return output + + def _forward_expanded(self, x, incremental_state): + """Turn the convolution filters into band matrices and do matrix multiplication. + This is faster when the sequence is short, but less memory efficient. + This is not used in the decoder during inference. + """ + T, B, C = x.size() + K, H = self.kernel_size, self.num_heads + R = C // H + assert R * H == C == self.input_size + + weight = self.weight.view(H, K) + if self.weight_softmax: + weight = utils.softmax(weight, dim=1, onnx_trace=self.onnx_trace).type_as( + weight + ) + weight = weight.view(1, H, K).expand(T * B, H, K).contiguous() + weight = weight.view(T, B * H, K).transpose(0, 1) + + x = x.view(T, B * H, R).transpose(0, 1) + P = self.padding_l + if K > T and P == K - 1: + weight = weight.narrow(2, K - T, T) + K, P = T, T - 1 + # turn the convolution filters into band matrices + weight_expanded = weight.new_zeros(B * H, T, T + K - 1, requires_grad=False) + weight_expanded.as_strided((B * H, T, K), (T * (T + K - 1), T + K, 1)).copy_( + weight + ) + weight_expanded = weight_expanded.narrow(2, P, T) + weight_expanded = self.weight_dropout_module(weight_expanded) + + output = torch.bmm(weight_expanded, x) + output = output.transpose(0, 1).contiguous().view(T, B, C) + return output + + def reorder_incremental_state(self, incremental_state, new_order): + input_buffer = self._get_input_buffer(incremental_state) + if input_buffer is not None: + input_buffer = input_buffer.index_select(1, new_order) + self._set_input_buffer(incremental_state, input_buffer) + + def _get_input_buffer(self, incremental_state): + return utils.get_incremental_state(self, incremental_state, "input_buffer") + + def _set_input_buffer(self, incremental_state, new_buffer): + return utils.set_incremental_state( + self, incremental_state, "input_buffer", new_buffer + ) + + def extra_repr(self): + s = "{}, kernel_size={}, padding_l={}, num_heads={}, weight_softmax={}, bias={}".format( + self.input_size, + self.kernel_size, + self.padding_l, + self.num_heads, + self.weight_softmax, + self.bias is not None, + ) + if self.weight_dropout_module.p > 0.0: + s += ", weight_dropout={}".format(self.weight_dropout_module.p) + return s diff --git a/fairseq/modules/linearized_convolution.py b/fairseq/modules/linearized_convolution.py new file mode 100644 index 0000000000000000000000000000000000000000..f7e156cb0c75cb375447859c8b6749311372c35e --- /dev/null +++ b/fairseq/modules/linearized_convolution.py @@ -0,0 +1,110 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn.functional as F +from fairseq import utils +from fairseq.incremental_decoding_utils import with_incremental_state + +from .conv_tbc import ConvTBC + +from typing import Dict, Optional +from torch import Tensor + +@with_incremental_state +class LinearizedConvolution(ConvTBC): + """An optimized version of nn.Conv1d. + + At training time, this module uses ConvTBC, which is an optimized version + of Conv1d. At inference time, it optimizes incremental generation (i.e., + one time step at a time) by replacing the convolutions with linear layers. + Note that the input order changes from training to inference. + """ + + def __init__(self, in_channels, out_channels, kernel_size, **kwargs): + super().__init__(in_channels, out_channels, kernel_size, **kwargs) + self._linearized_weight = None + self.register_backward_hook(self._clear_linearized_weight) + + def state_dict(self, destination=None, prefix="", keep_vars=False): + state = ConvTBC.state_dict(self, destination, prefix, keep_vars=keep_vars) + # don't store redundant _linearized_weight in checkpoints + if prefix + "_linearized_weight" in state: + del state[prefix + "_linearized_weight"] + return state + + def upgrade_state_dict_named(self, state_dict, name): + prefix = name + "." if name != "" else "" + if prefix + "_linearized_weight" in state_dict: + del state_dict[prefix + "_linearized_weight"] + + @torch.jit.export + def forward(self, input, incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None): + """ + Args: + incremental_state: Used to buffer signal; if not None, then input is + expected to contain a single frame. If the input order changes + between time steps, call reorder_incremental_state. + Input: + Time x Batch x Channel during training + Batch x Time x Channel during inference + """ + if incremental_state is None: + output = self.conv_tbc(input) + if self.kernel_size[0] > 1 and self.padding[0] > 0: + # remove future timesteps added by padding + output = output[: -self.padding[0], :, :] + return output + + # reshape weight + weight = self._get_linearized_weight() + kw = self.kernel_size[0] + + bsz = input.size(0) # input: bsz x len x dim + if kw > 1: + input = input.data + input_buffer = self._get_input_buffer(incremental_state) + if input_buffer is None: + input_buffer = input.new(bsz, kw, input.size(2)).zero_() + self._set_input_buffer(incremental_state, input_buffer) + else: + # shift buffer + input_buffer[:, :-1, :] = input_buffer[:, 1:, :].clone() + # append next input + input_buffer[:, -1, :] = input[:, -1, :] + input = input_buffer + with torch.no_grad(): + output = F.linear(input.view(bsz, -1), weight, self.bias) + return output.view(bsz, 1, -1) + + @torch.jit.unused + def reorder_incremental_state(self, incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]], new_order): + input_buffer = self._get_input_buffer(incremental_state) + if input_buffer is not None: + input_buffer = input_buffer.index_select(0, new_order) + self._set_input_buffer(incremental_state, input_buffer) + + @torch.jit.unused + def _get_input_buffer(self, incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]]): + return utils.get_incremental_state(self, incremental_state, "input_buffer") + + @torch.jit.unused + def _set_input_buffer(self, incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]], new_buffer): + return utils.set_incremental_state( + self, incremental_state, "input_buffer", new_buffer + ) + + @torch.jit.unused + def _get_linearized_weight(self): + if self._linearized_weight is None: + kw = self.kernel_size[0] + weight = self.weight.transpose(2, 1).transpose(1, 0).contiguous() + assert weight.size() == (self.out_channels, kw, self.in_channels) + return weight.view(self.out_channels, -1) + return self._linearized_weight + + @torch.jit.unused + def _clear_linearized_weight(self, *args): + self._linearized_weight = None diff --git a/fairseq/modules/multihead_attention.py b/fairseq/modules/multihead_attention.py new file mode 100644 index 0000000000000000000000000000000000000000..9bdca0f6af43a0a89e9225594ba5b6fbc5ee04c1 --- /dev/null +++ b/fairseq/modules/multihead_attention.py @@ -0,0 +1,500 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +from typing import Dict, Optional, Tuple + +import torch +import torch.nn.functional as F +from fairseq import utils +from fairseq.incremental_decoding_utils import with_incremental_state +from fairseq.modules.fairseq_dropout import FairseqDropout +from fairseq.modules.quant_noise import quant_noise +from torch import Tensor, nn +from torch.nn import Parameter + + +@with_incremental_state +class MultiheadAttention(nn.Module): + """Multi-headed attention. + + See "Attention Is All You Need" for more details. + """ + + def __init__( + self, + embed_dim, + num_heads, + kdim=None, + vdim=None, + dropout=0.0, + bias=True, + add_bias_kv=False, + add_zero_attn=False, + self_attention=False, + encoder_decoder_attention=False, + q_noise=0.0, + qn_block_size=8, + ): + super().__init__() + self.embed_dim = embed_dim + self.kdim = kdim if kdim is not None else embed_dim + self.vdim = vdim if vdim is not None else embed_dim + self.qkv_same_dim = self.kdim == embed_dim and self.vdim == embed_dim + + self.num_heads = num_heads + self.dropout_module = FairseqDropout( + dropout, module_name=self.__class__.__name__ + ) + + self.head_dim = embed_dim // num_heads + assert ( + self.head_dim * num_heads == self.embed_dim + ), "embed_dim must be divisible by num_heads" + self.scaling = self.head_dim ** -0.5 + + self.self_attention = self_attention + self.encoder_decoder_attention = encoder_decoder_attention + + assert not self.self_attention or self.qkv_same_dim, ( + "Self-attention requires query, key and " "value to be of the same size" + ) + + self.k_proj = quant_noise( + nn.Linear(self.kdim, embed_dim, bias=bias), q_noise, qn_block_size + ) + self.v_proj = quant_noise( + nn.Linear(self.vdim, embed_dim, bias=bias), q_noise, qn_block_size + ) + self.q_proj = quant_noise( + nn.Linear(embed_dim, embed_dim, bias=bias), q_noise, qn_block_size + ) + + self.out_proj = quant_noise( + nn.Linear(embed_dim, embed_dim, bias=bias), q_noise, qn_block_size + ) + + if add_bias_kv: + self.bias_k = Parameter(torch.Tensor(1, 1, embed_dim)) + self.bias_v = Parameter(torch.Tensor(1, 1, embed_dim)) + else: + self.bias_k = self.bias_v = None + + self.add_zero_attn = add_zero_attn + + self.reset_parameters() + + self.onnx_trace = False + + def prepare_for_onnx_export_(self): + self.onnx_trace = True + + def reset_parameters(self): + if self.qkv_same_dim: + # Empirically observed the convergence to be much better with + # the scaled initialization + nn.init.xavier_uniform_(self.k_proj.weight, gain=1 / math.sqrt(2)) + nn.init.xavier_uniform_(self.v_proj.weight, gain=1 / math.sqrt(2)) + nn.init.xavier_uniform_(self.q_proj.weight, gain=1 / math.sqrt(2)) + else: + nn.init.xavier_uniform_(self.k_proj.weight) + nn.init.xavier_uniform_(self.v_proj.weight) + nn.init.xavier_uniform_(self.q_proj.weight) + + nn.init.xavier_uniform_(self.out_proj.weight) + if self.out_proj.bias is not None: + nn.init.constant_(self.out_proj.bias, 0.0) + if self.bias_k is not None: + nn.init.xavier_normal_(self.bias_k) + if self.bias_v is not None: + nn.init.xavier_normal_(self.bias_v) + + def forward( + self, + query, + key: Optional[Tensor], + value: Optional[Tensor], + key_padding_mask: Optional[Tensor] = None, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + need_weights: bool = True, + static_kv: bool = False, + attn_mask: Optional[Tensor] = None, + before_softmax: bool = False, + need_head_weights: bool = False, + ) -> Tuple[Tensor, Optional[Tensor]]: + """Input shape: Time x Batch x Channel + + Args: + key_padding_mask (ByteTensor, optional): mask to exclude + keys that are pads, of shape `(batch, src_len)`, where + padding elements are indicated by 1s. + need_weights (bool, optional): return the attention weights, + averaged over heads (default: False). + attn_mask (ByteTensor, optional): typically used to + implement causal attention, where the mask prevents the + attention from looking forward in time (default: None). + before_softmax (bool, optional): return the raw attention + weights and values before the attention softmax. + need_head_weights (bool, optional): return the attention + weights for each head. Implies *need_weights*. Default: + return the average attention weights over all heads. + """ + if need_head_weights: + need_weights = True + + is_tpu = query.device.type == "xla" + + tgt_len, bsz, embed_dim = query.size() + src_len = tgt_len + assert embed_dim == self.embed_dim + assert list(query.size()) == [tgt_len, bsz, embed_dim] + if key is not None: + src_len, key_bsz, _ = key.size() + if not torch.jit.is_scripting(): + assert key_bsz == bsz + assert value is not None + assert src_len, bsz == value.shape[:2] + + if ( + not self.onnx_trace + and not is_tpu # don't use PyTorch version on TPUs + and incremental_state is None + and not static_kv + # A workaround for quantization to work. Otherwise JIT compilation + # treats bias in linear module as method. + and not torch.jit.is_scripting() + ): + assert key is not None and value is not None + return F.multi_head_attention_forward( + query, + key, + value, + self.embed_dim, + self.num_heads, + torch.empty([0]), + torch.cat((self.q_proj.bias, self.k_proj.bias, self.v_proj.bias)), + self.bias_k, + self.bias_v, + self.add_zero_attn, + self.dropout_module.p, + self.out_proj.weight, + self.out_proj.bias, + self.training or self.dropout_module.apply_during_inference, + key_padding_mask, + need_weights, + attn_mask, + use_separate_proj_weight=True, + q_proj_weight=self.q_proj.weight, + k_proj_weight=self.k_proj.weight, + v_proj_weight=self.v_proj.weight, + ) + + if incremental_state is not None: + saved_state = self._get_input_buffer(incremental_state) + if saved_state is not None and "prev_key" in saved_state: + # previous time steps are cached - no need to recompute + # key and value if they are static + if static_kv: + assert self.encoder_decoder_attention and not self.self_attention + key = value = None + else: + saved_state = None + + if self.self_attention: + q = self.q_proj(query) + k = self.k_proj(query) + v = self.v_proj(query) + elif self.encoder_decoder_attention: + # encoder-decoder attention + q = self.q_proj(query) + if key is None: + assert value is None + k = v = None + else: + k = self.k_proj(key) + v = self.v_proj(key) + + else: + assert key is not None and value is not None + q = self.q_proj(query) + k = self.k_proj(key) + v = self.v_proj(value) + q *= self.scaling + + if self.bias_k is not None: + assert self.bias_v is not None + k = torch.cat([k, self.bias_k.repeat(1, bsz, 1)]) + v = torch.cat([v, self.bias_v.repeat(1, bsz, 1)]) + if attn_mask is not None: + attn_mask = torch.cat( + [attn_mask, attn_mask.new_zeros(attn_mask.size(0), 1)], dim=1 + ) + if key_padding_mask is not None: + key_padding_mask = torch.cat( + [ + key_padding_mask, + key_padding_mask.new_zeros(key_padding_mask.size(0), 1), + ], + dim=1, + ) + + q = ( + q.contiguous() + .view(tgt_len, bsz * self.num_heads, self.head_dim) + .transpose(0, 1) + ) + if k is not None: + k = ( + k.contiguous() + .view(-1, bsz * self.num_heads, self.head_dim) + .transpose(0, 1) + ) + if v is not None: + v = ( + v.contiguous() + .view(-1, bsz * self.num_heads, self.head_dim) + .transpose(0, 1) + ) + + if saved_state is not None: + # saved states are stored with shape (bsz, num_heads, seq_len, head_dim) + if "prev_key" in saved_state: + _prev_key = saved_state["prev_key"] + assert _prev_key is not None + prev_key = _prev_key.view(bsz * self.num_heads, -1, self.head_dim) + if static_kv: + k = prev_key + else: + assert k is not None + k = torch.cat([prev_key, k], dim=1) + src_len = k.size(1) + if "prev_value" in saved_state: + _prev_value = saved_state["prev_value"] + assert _prev_value is not None + prev_value = _prev_value.view(bsz * self.num_heads, -1, self.head_dim) + if static_kv: + v = prev_value + else: + assert v is not None + v = torch.cat([prev_value, v], dim=1) + prev_key_padding_mask: Optional[Tensor] = None + if "prev_key_padding_mask" in saved_state: + prev_key_padding_mask = saved_state["prev_key_padding_mask"] + assert k is not None and v is not None + key_padding_mask = MultiheadAttention._append_prev_key_padding_mask( + key_padding_mask=key_padding_mask, + prev_key_padding_mask=prev_key_padding_mask, + batch_size=bsz, + src_len=k.size(1), + static_kv=static_kv, + ) + + saved_state["prev_key"] = k.view(bsz, self.num_heads, -1, self.head_dim) + saved_state["prev_value"] = v.view(bsz, self.num_heads, -1, self.head_dim) + saved_state["prev_key_padding_mask"] = key_padding_mask + # In this branch incremental_state is never None + assert incremental_state is not None + incremental_state = self._set_input_buffer(incremental_state, saved_state) + assert k is not None + assert k.size(1) == src_len + + # This is part of a workaround to get around fork/join parallelism + # not supporting Optional types. + if key_padding_mask is not None and key_padding_mask.dim() == 0: + key_padding_mask = None + + if key_padding_mask is not None: + assert key_padding_mask.size(0) == bsz + assert key_padding_mask.size(1) == src_len + + if self.add_zero_attn: + assert v is not None + src_len += 1 + k = torch.cat([k, k.new_zeros((k.size(0), 1) + k.size()[2:])], dim=1) + v = torch.cat([v, v.new_zeros((v.size(0), 1) + v.size()[2:])], dim=1) + if attn_mask is not None: + attn_mask = torch.cat( + [attn_mask, attn_mask.new_zeros(attn_mask.size(0), 1)], dim=1 + ) + if key_padding_mask is not None: + key_padding_mask = torch.cat( + [ + key_padding_mask, + torch.zeros(key_padding_mask.size(0), 1).type_as( + key_padding_mask + ), + ], + dim=1, + ) + + attn_weights = torch.bmm(q, k.transpose(1, 2)) + attn_weights = self.apply_sparse_mask(attn_weights, tgt_len, src_len, bsz) + + assert list(attn_weights.size()) == [bsz * self.num_heads, tgt_len, src_len] + + if attn_mask is not None: + attn_mask = attn_mask.unsqueeze(0) + if self.onnx_trace: + attn_mask = attn_mask.repeat(attn_weights.size(0), 1, 1) + attn_weights += attn_mask + + if key_padding_mask is not None: + # don't attend to padding symbols + attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + if not is_tpu: + attn_weights = attn_weights.masked_fill( + key_padding_mask.unsqueeze(1).unsqueeze(2).to(torch.bool), + float("-inf"), + ) + else: + attn_weights = attn_weights.transpose(0, 2) + attn_weights = attn_weights.masked_fill(key_padding_mask, float("-inf")) + attn_weights = attn_weights.transpose(0, 2) + attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len) + + if before_softmax: + return attn_weights, v + + attn_weights_float = utils.softmax( + attn_weights, dim=-1, onnx_trace=self.onnx_trace + ) + attn_weights = attn_weights_float.type_as(attn_weights) + attn_probs = self.dropout_module(attn_weights) + + assert v is not None + attn = torch.bmm(attn_probs, v) + assert list(attn.size()) == [bsz * self.num_heads, tgt_len, self.head_dim] + if self.onnx_trace and attn.size(1) == 1: + # when ONNX tracing a single decoder step (sequence length == 1) + # the transpose is a no-op copy before view, thus unnecessary + attn = attn.contiguous().view(tgt_len, bsz, embed_dim) + else: + attn = attn.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim) + attn = self.out_proj(attn) + attn_weights: Optional[Tensor] = None + if need_weights: + attn_weights = attn_weights_float.view( + bsz, self.num_heads, tgt_len, src_len + ).transpose(1, 0) + if not need_head_weights: + # average attention weights over heads + attn_weights = attn_weights.mean(dim=0) + + return attn, attn_weights + + @staticmethod + def _append_prev_key_padding_mask( + key_padding_mask: Optional[Tensor], + prev_key_padding_mask: Optional[Tensor], + batch_size: int, + src_len: int, + static_kv: bool, + ) -> Optional[Tensor]: + # saved key padding masks have shape (bsz, seq_len) + if prev_key_padding_mask is not None and static_kv: + new_key_padding_mask = prev_key_padding_mask + elif prev_key_padding_mask is not None and key_padding_mask is not None: + new_key_padding_mask = torch.cat( + [prev_key_padding_mask.float(), key_padding_mask.float()], dim=1 + ) + # During incremental decoding, as the padding token enters and + # leaves the frame, there will be a time when prev or current + # is None + elif prev_key_padding_mask is not None: + if src_len > prev_key_padding_mask.size(1): + filler = torch.zeros( + (batch_size, src_len - prev_key_padding_mask.size(1)), + device=prev_key_padding_mask.device, + ) + new_key_padding_mask = torch.cat( + [prev_key_padding_mask.float(), filler.float()], dim=1 + ) + else: + new_key_padding_mask = prev_key_padding_mask.float() + elif key_padding_mask is not None: + if src_len > key_padding_mask.size(1): + filler = torch.zeros( + (batch_size, src_len - key_padding_mask.size(1)), + device=key_padding_mask.device, + ) + new_key_padding_mask = torch.cat( + [filler.float(), key_padding_mask.float()], dim=1 + ) + else: + new_key_padding_mask = key_padding_mask.float() + else: + new_key_padding_mask = prev_key_padding_mask + return new_key_padding_mask + + @torch.jit.export + def reorder_incremental_state( + self, + incremental_state: Dict[str, Dict[str, Optional[Tensor]]], + new_order: Tensor, + ): + """Reorder buffered internal state (for incremental generation).""" + input_buffer = self._get_input_buffer(incremental_state) + if input_buffer is not None: + for k in input_buffer.keys(): + input_buffer_k = input_buffer[k] + if input_buffer_k is not None: + if self.encoder_decoder_attention and input_buffer_k.size( + 0 + ) == new_order.size(0): + break + input_buffer[k] = input_buffer_k.index_select(0, new_order) + incremental_state = self._set_input_buffer(incremental_state, input_buffer) + return incremental_state + + def _get_input_buffer( + self, incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] + ) -> Dict[str, Optional[Tensor]]: + result = self.get_incremental_state(incremental_state, "attn_state") + if result is not None: + return result + else: + empty_result: Dict[str, Optional[Tensor]] = {} + return empty_result + + def _set_input_buffer( + self, + incremental_state: Dict[str, Dict[str, Optional[Tensor]]], + buffer: Dict[str, Optional[Tensor]], + ): + return self.set_incremental_state(incremental_state, "attn_state", buffer) + + def apply_sparse_mask(self, attn_weights, tgt_len: int, src_len: int, bsz: int): + return attn_weights + + def upgrade_state_dict_named(self, state_dict, name): + prefix = name + "." if name != "" else "" + items_to_add = {} + keys_to_remove = [] + for k in state_dict.keys(): + if k.endswith(prefix + "in_proj_weight"): + # in_proj_weight used to be q + k + v with same dimensions + dim = int(state_dict[k].shape[0] / 3) + items_to_add[prefix + "q_proj.weight"] = state_dict[k][:dim] + items_to_add[prefix + "k_proj.weight"] = state_dict[k][dim : 2 * dim] + items_to_add[prefix + "v_proj.weight"] = state_dict[k][2 * dim :] + + keys_to_remove.append(k) + + k_bias = prefix + "in_proj_bias" + if k_bias in state_dict.keys(): + dim = int(state_dict[k].shape[0] / 3) + items_to_add[prefix + "q_proj.bias"] = state_dict[k_bias][:dim] + items_to_add[prefix + "k_proj.bias"] = state_dict[k_bias][ + dim : 2 * dim + ] + items_to_add[prefix + "v_proj.bias"] = state_dict[k_bias][2 * dim :] + + keys_to_remove.append(prefix + "in_proj_bias") + + for k in keys_to_remove: + del state_dict[k] + + for key, value in items_to_add.items(): + state_dict[key] = value diff --git a/fairseq/modules/positional_embedding.py b/fairseq/modules/positional_embedding.py new file mode 100644 index 0000000000000000000000000000000000000000..8e94e35edb46bf9dea911fe74577d8ecbe9b5ff1 --- /dev/null +++ b/fairseq/modules/positional_embedding.py @@ -0,0 +1,35 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch.nn as nn + +from .learned_positional_embedding import LearnedPositionalEmbedding +from .sinusoidal_positional_embedding import SinusoidalPositionalEmbedding + + +def PositionalEmbedding( + num_embeddings: int, + embedding_dim: int, + padding_idx: int, + learned: bool = False, +): + if learned: + # if padding_idx is specified then offset the embedding ids by + # this index and adjust num_embeddings appropriately + # TODO: The right place for this offset would be inside + # LearnedPositionalEmbedding. Move this there for a cleaner implementation. + if padding_idx is not None: + num_embeddings = num_embeddings + padding_idx + 1 + m = LearnedPositionalEmbedding(num_embeddings, embedding_dim, padding_idx) + nn.init.normal_(m.weight, mean=0, std=embedding_dim ** -0.5) + if padding_idx is not None: + nn.init.constant_(m.weight[padding_idx], 0) + else: + m = SinusoidalPositionalEmbedding( + embedding_dim, + padding_idx, + init_size=num_embeddings + padding_idx + 1, + ) + return m diff --git a/fairseq/modules/quant_noise.py b/fairseq/modules/quant_noise.py new file mode 100644 index 0000000000000000000000000000000000000000..d777dfbb6c1bf6a9b769dfdaec35d5ef084c8a8b --- /dev/null +++ b/fairseq/modules/quant_noise.py @@ -0,0 +1,107 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn as nn + + +def quant_noise(module, p, block_size): + """ + Wraps modules and applies quantization noise to the weights for + subsequent quantization with Iterative Product Quantization as + described in "Training with Quantization Noise for Extreme Model Compression" + + Args: + - module: nn.Module + - p: amount of Quantization Noise + - block_size: size of the blocks for subsequent quantization with iPQ + + Remarks: + - Module weights must have the right sizes wrt the block size + - Only Linear, Embedding and Conv2d modules are supported for the moment + - For more detail on how to quantize by blocks with convolutional weights, + see "And the Bit Goes Down: Revisiting the Quantization of Neural Networks" + - We implement the simplest form of noise here as stated in the paper + which consists in randomly dropping blocks + """ + + # if no quantization noise, don't register hook + if p <= 0: + return module + + # supported modules + assert isinstance(module, (nn.Linear, nn.Embedding, nn.Conv2d)) + + # test whether module.weight has the right sizes wrt block_size + is_conv = module.weight.ndim == 4 + + # 2D matrix + if not is_conv: + assert ( + module.weight.size(1) % block_size == 0 + ), "Input features must be a multiple of block sizes" + + # 4D matrix + else: + # 1x1 convolutions + if module.kernel_size == (1, 1): + assert ( + module.in_channels % block_size == 0 + ), "Input channels must be a multiple of block sizes" + # regular convolutions + else: + k = module.kernel_size[0] * module.kernel_size[1] + assert k % block_size == 0, "Kernel size must be a multiple of block size" + + def _forward_pre_hook(mod, input): + # no noise for evaluation + if mod.training: + if not is_conv: + # gather weight and sizes + weight = mod.weight + in_features = weight.size(1) + out_features = weight.size(0) + + # split weight matrix into blocks and randomly drop selected blocks + mask = torch.zeros( + in_features // block_size * out_features, device=weight.device + ) + mask.bernoulli_(p) + mask = mask.repeat_interleave(block_size, -1).view(-1, in_features) + + else: + # gather weight and sizes + weight = mod.weight + in_channels = mod.in_channels + out_channels = mod.out_channels + + # split weight matrix into blocks and randomly drop selected blocks + if mod.kernel_size == (1, 1): + mask = torch.zeros( + int(in_channels // block_size * out_channels), + device=weight.device, + ) + mask.bernoulli_(p) + mask = mask.repeat_interleave(block_size, -1).view(-1, in_channels) + else: + mask = torch.zeros( + weight.size(0), weight.size(1), device=weight.device + ) + mask.bernoulli_(p) + mask = ( + mask.unsqueeze(2) + .unsqueeze(3) + .repeat(1, 1, mod.kernel_size[0], mod.kernel_size[1]) + ) + + # scale weights and apply mask + mask = mask.to( + torch.bool + ) # x.bool() is not currently supported in TorchScript + s = 1 / (1 - p) + mod.weight.data = s * weight.masked_fill(mask, 0) + + module.register_forward_pre_hook(_forward_pre_hook) + return module diff --git a/fairseq/modules/quantization/__init__.py b/fairseq/modules/quantization/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/fairseq/modules/quantization/pq/__init__.py b/fairseq/modules/quantization/pq/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..5b10b51b1b0ca21aaec96344f86a0ab9df0c22f8 --- /dev/null +++ b/fairseq/modules/quantization/pq/__init__.py @@ -0,0 +1,6 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .utils import SizeTracker, quantize_model_ # NOQA diff --git a/fairseq/modules/quantization/pq/em.py b/fairseq/modules/quantization/pq/em.py new file mode 100644 index 0000000000000000000000000000000000000000..6f15c3e46bd052b1e00929e7ece9355fb03846c7 --- /dev/null +++ b/fairseq/modules/quantization/pq/em.py @@ -0,0 +1,211 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import os +import random +from collections import Counter + +import torch + + +class EM: + """ + EM algorithm used to quantize the columns of W to minimize + + ||W - W_hat||^2 + + Args: + - W: weight matrix of size (in_features x out_features) + - n_iter: number of k-means iterations + - n_centroids: number of centroids (size of codebook) + - eps: for cluster reassignment when an empty cluster is found + - max_tentatives for cluster reassignment when an empty cluster is found + - verbose: print error after each iteration + + Remarks: + - If one cluster is empty, the most populated cluster is split into + two clusters + - All the relevant dimensions are specified in the code + """ + + def __init__( + self, W, n_centroids=256, n_iter=20, eps=1e-6, max_tentatives=30, verbose=True + ): + self.W = W + self.n_centroids = n_centroids + self.n_iter = n_iter + self.eps = eps + self.max_tentatives = max_tentatives + self.verbose = verbose + self.centroids = torch.Tensor() + self.assignments = torch.Tensor() + self.objective = [] + + def initialize_centroids(self): + """ + Initializes the centroids by sampling random columns from W. + """ + + in_features, out_features = self.W.size() + indices = torch.randint( + low=0, high=out_features, size=(self.n_centroids,) + ).long() + self.centroids = self.W[:, indices].t() # (n_centroids x in_features) + + def step(self, i): + """ + There are two standard steps for each iteration: expectation (E) and + minimization (M). The E-step (assignment) is performed with an exhaustive + search and the M-step (centroid computation) is performed with + the exact solution. + + Args: + - i: step number + + Remarks: + - The E-step heavily uses PyTorch broadcasting to speed up computations + and reduce the memory overhead + """ + + # assignments (E-step) + distances = self.compute_distances() # (n_centroids x out_features) + self.assignments = torch.argmin(distances, dim=0) # (out_features) + n_empty_clusters = self.resolve_empty_clusters() + + # centroids (M-step) + for k in range(self.n_centroids): + W_k = self.W[:, self.assignments == k] # (in_features x size_of_cluster_k) + self.centroids[k] = W_k.mean(dim=1) # (in_features) + + # book-keeping + obj = (self.centroids[self.assignments].t() - self.W).norm(p=2).item() + self.objective.append(obj) + if self.verbose: + logging.info( + f"Iteration: {i},\t" + f"objective: {obj:.6f},\t" + f"resolved empty clusters: {n_empty_clusters}" + ) + + def resolve_empty_clusters(self): + """ + If one cluster is empty, the most populated cluster is split into + two clusters by shifting the respective centroids. This is done + iteratively for a fixed number of tentatives. + """ + + # empty clusters + counts = Counter(map(lambda x: x.item(), self.assignments)) + empty_clusters = set(range(self.n_centroids)) - set(counts.keys()) + n_empty_clusters = len(empty_clusters) + + tentatives = 0 + while len(empty_clusters) > 0: + # given an empty cluster, find most populated cluster and split it into two + k = random.choice(list(empty_clusters)) + m = counts.most_common(1)[0][0] + e = torch.randn_like(self.centroids[m]) * self.eps + self.centroids[k] = self.centroids[m].clone() + self.centroids[k] += e + self.centroids[m] -= e + + # recompute assignments + distances = self.compute_distances() # (n_centroids x out_features) + self.assignments = torch.argmin(distances, dim=0) # (out_features) + + # check for empty clusters + counts = Counter(map(lambda x: x.item(), self.assignments)) + empty_clusters = set(range(self.n_centroids)) - set(counts.keys()) + + # increment tentatives + if tentatives == self.max_tentatives: + logging.info( + f"Could not resolve all empty clusters, {len(empty_clusters)} remaining" + ) + raise EmptyClusterResolveError + tentatives += 1 + + return n_empty_clusters + + def compute_distances(self): + """ + For every centroid m, computes + + ||M - m[None, :]||_2 + + Remarks: + - We rely on PyTorch's broadcasting to speed up computations + and reduce the memory overhead + - Without chunking, the sizes in the broadcasting are modified as: + (n_centroids x n_samples x out_features) -> (n_centroids x out_features) + - The broadcasting computation is automatically chunked so that + the tensors fit into the memory of the GPU + """ + + nb_centroids_chunks = 1 + + while True: + try: + return torch.cat( + [ + (self.W[None, :, :] - centroids_c[:, :, None]).norm(p=2, dim=1) + for centroids_c in self.centroids.chunk( + nb_centroids_chunks, dim=0 + ) + ], + dim=0, + ) + except RuntimeError: + nb_centroids_chunks *= 2 + + def assign(self): + """ + Assigns each column of W to its closest centroid, thus essentially + performing the E-step in train(). + + Remarks: + - The function must be called after train() or after loading + centroids using self.load(), otherwise it will return empty tensors + """ + + distances = self.compute_distances() # (n_centroids x out_features) + self.assignments = torch.argmin(distances, dim=0) # (out_features) + + def save(self, path, layer): + """ + Saves centroids and assignments. + + Args: + - path: folder used to save centroids and assignments + """ + + torch.save(self.centroids, os.path.join(path, "{}_centroids.pth".format(layer))) + torch.save( + self.assignments, os.path.join(path, "{}_assignments.pth".format(layer)) + ) + torch.save(self.objective, os.path.join(path, "{}_objective.pth".format(layer))) + + def load(self, path, layer): + """ + Loads centroids and assignments from a given path + + Args: + - path: folder use to load centroids and assignments + """ + + self.centroids = torch.load( + os.path.join(path, "{}_centroids.pth".format(layer)) + ) + self.assignments = torch.load( + os.path.join(path, "{}_assignments.pth".format(layer)) + ) + self.objective = torch.load( + os.path.join(path, "{}_objective.pth".format(layer)) + ) + + +class EmptyClusterResolveError(Exception): + pass diff --git a/fairseq/modules/quantization/pq/modules/__init__.py b/fairseq/modules/quantization/pq/modules/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..b67c8e8ad691aa01e9e10e904d69d94595387668 --- /dev/null +++ b/fairseq/modules/quantization/pq/modules/__init__.py @@ -0,0 +1,8 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .qconv import PQConv2d # NOQA +from .qemb import PQEmbedding # NOQA +from .qlinear import PQLinear # NOQA diff --git a/fairseq/modules/quantization/pq/modules/qconv.py b/fairseq/modules/quantization/pq/modules/qconv.py new file mode 100644 index 0000000000000000000000000000000000000000..d15ec192e8cda6265a198e583a9bf7fb194dd129 --- /dev/null +++ b/fairseq/modules/quantization/pq/modules/qconv.py @@ -0,0 +1,115 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import numpy as np +import torch +import torch.nn as nn +import torch.nn.functional as F +from torch.nn.modules.utils import _pair + + +class PQConv2d(nn.Module): + """ + Quantized counterpart of nn.Conv2d module. Stores the centroid, the assignments + and the non-quantized biases. The full weight is re-instantiated at each forward + pass and autograd automatically computes the gradients with respect to the + centroids. + + Args: + - centroids: centroids of size n_centroids x block_size + - assignments: assignments of the centroids to the subvectors + of size self.out_channels x n_blocks + - bias: the non-quantized bias, must be either torch.Tensor or None + + Remarks: + - We refer the reader to the official documentation of the nn.Conv2d module + for the other arguments and the behavior of the module. + - Performance tests on GPU show that this implementation is 10% slower than + the non-quantized nn.Conv2d module for a standard training loop. + - During the backward, the gradients are averaged by cluster and not summed. + This explains the hook registered to the centroids. + """ + + def __init__( + self, + centroids, + assignments, + bias, + in_channels, + out_channels, + kernel_size, + stride=1, + padding=0, + dilation=1, + groups=1, + padding_mode="zeros", + ): + super(PQConv2d, self).__init__() + self.block_size = centroids.size(1) + self.n_centroids = centroids.size(0) + self.in_channels = in_channels + self.out_channels = out_channels + self.kernel_size = _pair(kernel_size) + self.stride = _pair(stride) + self.padding = _pair(padding) + self.dilation = _pair(dilation) + self.groups = groups + self.padding_mode = padding_mode + # check compatibility + if in_channels // groups * np.prod(self.kernel_size) % self.block_size != 0: + raise ValueError("Wrong PQ sizes") + if len(assignments) % out_channels != 0: + raise ValueError("Wrong PQ sizes") + if in_channels % groups != 0: + raise ValueError("in_channels must be divisible by groups") + if out_channels % groups != 0: + raise ValueError("out_channels must be divisible by groups") + # define parameters + self.centroids = nn.Parameter(centroids, requires_grad=True) + self.register_buffer("assignments", assignments) + self.register_buffer("counts", torch.bincount(assignments).type_as(centroids)) + if bias is not None: + self.bias = nn.Parameter(bias) + else: + self.register_parameter("bias", None) + # register hook for averaging gradients per centroids instead of summing + self.centroids.register_hook(lambda x: x / self.counts[:, None]) + + @property + def weight(self): + return ( + self.centroids[self.assignments] + .reshape(-1, self.out_channels, self.block_size) + .permute(1, 0, 2) + .reshape( + self.out_channels, self.in_channels // self.groups, *self.kernel_size + ) + ) + + def forward(self, x): + return F.conv2d( + x, + self.weight, + self.bias, + self.stride, + self.padding, + self.dilation, + self.groups, + ) + + def extra_repr(self): + s = "{in_channels}, {out_channels}, kernel_size={kernel_size}, stride={stride}" + if self.padding != (0,) * len(self.padding): + s += ", padding={padding}" + if self.dilation != (1,) * len(self.dilation): + s += ", dilation={dilation}" + if self.groups != 1: + s += ", groups={groups}" + if self.bias is None: + s += ", bias=False" + if self.padding_mode != "zeros": + s += ", padding_mode={padding_mode}" + s += ", n_centroids={n_centroids}, block_size={block_size}" + return s.format(**self.__dict__) diff --git a/fairseq/modules/quantization/pq/modules/qemb.py b/fairseq/modules/quantization/pq/modules/qemb.py new file mode 100644 index 0000000000000000000000000000000000000000..3a74ad3c4c7c9d3203d26e7885864ba578951bfe --- /dev/null +++ b/fairseq/modules/quantization/pq/modules/qemb.py @@ -0,0 +1,107 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn as nn +import torch.nn.functional as F + + +class PQEmbedding(nn.Module): + """ + Quantized counterpart of nn.Embedding module. Stores the centroids and + the assignments. The full weight is re-instantiated at each forward + pass. + + Args: + - centroids: centroids of size n_centroids x block_size + - assignments: assignments of the centroids to the subvectors + of size self.out_features x n_blocks + - bias: the non-quantized bias + + Remarks: + - We refer the reader to the official documentation of the nn.Embedding module + for the other arguments and the behavior of the module + - Performance tests on GPU show that this implementation is 10% slower than + the non-quantized nn.Embedding module for a standard training loop. + """ + + def __init__( + self, + centroids, + assignments, + num_embeddings, + embedding_dim, + padding_idx=None, + max_norm=None, + norm_type=2.0, + scale_grad_by_freq=False, + sparse=False, + _weight=None, + ): + super(PQEmbedding, self).__init__() + self.block_size = centroids.size(1) + self.n_centroids = centroids.size(0) + self.num_embeddings = num_embeddings + self.embedding_dim = embedding_dim + if padding_idx is not None: + if padding_idx > 0: + assert ( + padding_idx < self.num_embeddings + ), "Padding_idx must be within num_embeddings" + elif padding_idx < 0: + assert ( + padding_idx >= -self.num_embeddings + ), "Padding_idx must be within num_embeddings" + padding_idx = self.num_embeddings + padding_idx + self.padding_idx = padding_idx + self.max_norm = max_norm + self.norm_type = norm_type + self.scale_grad_by_freq = scale_grad_by_freq + self.sparse = sparse + # check compatibility + if self.embedding_dim % self.block_size != 0: + raise ValueError("Wrong PQ sizes") + if len(assignments) % self.num_embeddings != 0: + raise ValueError("Wrong PQ sizes") + # define parameters + self.centroids = nn.Parameter(centroids, requires_grad=True) + self.register_buffer("assignments", assignments) + self.register_buffer("counts", torch.bincount(assignments).type_as(centroids)) + + @property + def weight(self): + return ( + self.centroids[self.assignments] + .reshape(-1, self.num_embeddings, self.block_size) + .permute(1, 0, 2) + .flatten(1, 2) + ) + + def forward(self, input): + return F.embedding( + input, + self.weight, + self.padding_idx, + self.max_norm, + self.norm_type, + self.scale_grad_by_freq, + self.sparse, + ) + + def extra_repr(self): + s = "{num_embeddings}, {embedding_dim}" + if self.padding_idx is not None: + s += ", padding_idx={padding_idx}" + if self.max_norm is not None: + s += ", max_norm={max_norm}" + if self.norm_type != 2: + s += ", norm_type={norm_type}" + if self.scale_grad_by_freq is not False: + s += ", scale_grad_by_freq={scale_grad_by_freq}" + if self.sparse is not False: + s += ", sparse=True" + s += ", n_centroids={n_centroids}, block_size={block_size}" + + return s.format(**self.__dict__) diff --git a/fairseq/modules/quantization/pq/modules/qlinear.py b/fairseq/modules/quantization/pq/modules/qlinear.py new file mode 100644 index 0000000000000000000000000000000000000000..9bdd25a8685bb7c7b32e1f02372aaeb26d8ba53a --- /dev/null +++ b/fairseq/modules/quantization/pq/modules/qlinear.py @@ -0,0 +1,71 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn as nn +import torch.nn.functional as F + + +class PQLinear(nn.Module): + """ + Quantized counterpart of nn.Linear module. Stores the centroid, the assignments + and the non-quantized biases. The full weight is re-instantiated at each forward + pass. + + Args: + - centroids: centroids of size n_centroids x block_size + - assignments: assignments of the centroids to the subvectors + of size self.out_features x n_blocks + - bias: the non-quantized bias + + Remarks: + - We refer the reader to the official documentation of the nn.Linear module + for the other arguments and the behavior of the module + - Performance tests on GPU show that this implementation is 15% slower than + the non-quantized nn.Linear module for a standard training loop. + """ + + def __init__(self, centroids, assignments, bias, in_features, out_features): + super(PQLinear, self).__init__() + self.block_size = centroids.size(1) + self.n_centroids = centroids.size(0) + self.in_features = in_features + self.out_features = out_features + # check compatibility + if self.in_features % self.block_size != 0: + raise ValueError("Wrong PQ sizes") + if len(assignments) % self.out_features != 0: + raise ValueError("Wrong PQ sizes") + # define parameters + self.centroids = nn.Parameter(centroids, requires_grad=True) + self.register_buffer("assignments", assignments) + self.register_buffer("counts", torch.bincount(assignments).type_as(centroids)) + if bias is not None: + self.bias = nn.Parameter(bias) + else: + self.register_parameter("bias", None) + + @property + def weight(self): + return ( + self.centroids[self.assignments] + .reshape(-1, self.out_features, self.block_size) + .permute(1, 0, 2) + .flatten(1, 2) + ) + + def forward(self, x): + return F.linear( + x, + self.weight, + self.bias, + ) + + def extra_repr(self): + return f"in_features={self.in_features},\ + out_features={self.out_features},\ + n_centroids={self.n_centroids},\ + block_size={self.block_size},\ + bias={self.bias is not None}" diff --git a/fairseq/modules/quantization/pq/pq.py b/fairseq/modules/quantization/pq/pq.py new file mode 100644 index 0000000000000000000000000000000000000000..eddc2eb34602403f10979f54cd23a45bc2f104d5 --- /dev/null +++ b/fairseq/modules/quantization/pq/pq.py @@ -0,0 +1,128 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .em import EM, EmptyClusterResolveError + + +class PQ(EM): + """ + Quantizes the layer weights W with the standard Product Quantization + technique. This learns a codebook of codewords or centroids of size + block_size from W. For further reference on using PQ to quantize + neural networks, see "And the Bit Goes Down: Revisiting the Quantization + of Neural Networks", Stock et al., ICLR 2020. + + PQ is performed in two steps: + (1) The matrix W (weights or fully-connected or convolutional layer) + is reshaped to (block_size, -1). + - If W is fully-connected (2D), its columns are split into + blocks of size block_size. + - If W is convolutional (4D), its filters are split along the + spatial dimension. + (2) We apply the standard EM/k-means algorithm to the resulting reshaped matrix. + + Args: + - W: weight matrix to quantize of size (in_features x out_features) + - block_size: size of the blocks (subvectors) + - n_centroids: number of centroids + - n_iter: number of k-means iterations + - eps: for cluster reassignment when an empty cluster is found + - max_tentatives for cluster reassignment when an empty cluster is found + - verbose: print information after each iteration + + Remarks: + - block_size be compatible with the shape of W + """ + + def __init__( + self, + W, + block_size, + n_centroids=256, + n_iter=20, + eps=1e-6, + max_tentatives=30, + verbose=True, + ): + self.block_size = block_size + W_reshaped = self._reshape(W) + super(PQ, self).__init__( + W_reshaped, + n_centroids=n_centroids, + n_iter=n_iter, + eps=eps, + max_tentatives=max_tentatives, + verbose=verbose, + ) + + def _reshape(self, W): + """ + Reshapes the matrix W as expained in step (1). + """ + + # fully connected: by convention the weight has size out_features x in_features + if len(W.size()) == 2: + self.out_features, self.in_features = W.size() + assert ( + self.in_features % self.block_size == 0 + ), "Linear: n_blocks must be a multiple of in_features" + return ( + W.reshape(self.out_features, -1, self.block_size) + .permute(2, 1, 0) + .flatten(1, 2) + ) + + # convolutional: we reshape along the spatial dimension + elif len(W.size()) == 4: + self.out_channels, self.in_channels, self.k_h, self.k_w = W.size() + assert ( + self.in_channels * self.k_h * self.k_w + ) % self.block_size == 0, ( + "Conv2d: n_blocks must be a multiple of in_channels * k_h * k_w" + ) + return ( + W.reshape(self.out_channels, -1, self.block_size) + .permute(2, 1, 0) + .flatten(1, 2) + ) + # not implemented + else: + raise NotImplementedError(W.size()) + + def encode(self): + """ + Performs self.n_iter EM steps. + """ + + self.initialize_centroids() + for i in range(self.n_iter): + try: + self.step(i) + except EmptyClusterResolveError: + break + + def decode(self): + """ + Returns the encoded full weight matrix. Must be called after + the encode function. + """ + + # fully connected case + if "k_h" not in self.__dict__: + return ( + self.centroids[self.assignments] + .reshape(-1, self.out_features, self.block_size) + .permute(1, 0, 2) + .flatten(1, 2) + ) + + # convolutional case + else: + return ( + self.centroids[self.assignments] + .reshape(-1, self.out_channels, self.block_size) + .permute(1, 0, 2) + .reshape(self.out_channels, self.in_channels, self.k_h, self.k_w) + ) diff --git a/fairseq/modules/quantization/pq/utils.py b/fairseq/modules/quantization/pq/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..03b15e4b1b58c9a1e6d42052b3bd5457df9a6e2e --- /dev/null +++ b/fairseq/modules/quantization/pq/utils.py @@ -0,0 +1,337 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import re +from operator import attrgetter, itemgetter + +import numpy as np +import torch.distributed as dist +import torch.nn as nn + +from .modules import PQConv2d, PQEmbedding, PQLinear +from .pq import PQ + + +def quantize_model_( + model, + size_tracker, + layers_to_quantize, + block_sizes_config, + n_centroids_config, + step=0, + n_iter=15, + eps=1e-6, + max_tentatives=100, + verbose=True, +): + """ + Quantize a model in-place by stages. All the targeted + layers are replaced by their quantized counterpart, + and the model is ready for the finetuning of the + centroids in a standard training loop (no modifications + required). Note that we do not quantize biases. + + Args: + - model: a nn.Module + - size_tracker: useful for tracking quatization statistics + - layers_to_quantize: a list containing regexps for + filtering the layers to quantize at each stage according + to their name (as in model.named_parameters()) + - block_sizes_config: dict like + { + 'Conv2d': ('kernel_size', {'(3, 3)': 9, '(1, 1)': 4}), + 'Linear': ('in_features', {'*': 8}) + } + For instance, all conv2d layers with kernel size 3x3 have + a block size of 9 and all Linear layers are quantized with + a block size of 8, irrespective of their size. + - n_centroids_config: dict like + { + 'Conv2d': ('kernel_size', {'*': 256}), + 'Linear': ('in_features', {'*': 256}) + } + For instance, all conv2d layers are quantized with 256 centroids + - step: the layers to quantize inplace corresponding + to layers_to_quantize[step] + """ + + quantized_layers = get_layers(model, layers_to_quantize[step]) + + for layer in quantized_layers: + + # book-keeping + is_master_process = (not dist.is_initialized()) or ( + dist.is_initialized() and dist.get_rank() == 0 + ) + verbose = verbose and is_master_process + + # get block size and centroids + module = attrgetter(layer)(model) + block_size = get_param(module, layer, block_sizes_config) + n_centroids = get_param(module, layer, n_centroids_config) + if verbose: + logging.info( + f"Quantizing layer {layer} with block size {block_size} and {n_centroids} centroids" + ) + + # quantize layer + weight = module.weight.data.clone() + is_bias = "bias" in [x[0] for x in module.named_parameters()] + bias = module.bias.data.clone() if is_bias else None + quantizer = PQ( + weight, + block_size, + n_centroids=n_centroids, + n_iter=n_iter, + eps=eps, + max_tentatives=max_tentatives, + verbose=verbose, + ) + + # quantization performed on all GPUs with same seed + quantizer.encode() + centroids = quantizer.centroids.contiguous() + assignments = quantizer.assignments.contiguous() + + # broadcast results to make sure weights are up-to-date + if dist.is_initialized(): + dist.broadcast(centroids, 0) + dist.broadcast(assignments, 0) + + # instantiate the quantized counterpart + if isinstance(module, nn.Linear): + out_features, in_features = map( + lambda k: module.__dict__[k], ["out_features", "in_features"] + ) + quantized_module = PQLinear( + centroids, assignments, bias, in_features, out_features + ) + elif isinstance(module, nn.Embedding): + num_embeddings, embedding_dim = map( + lambda k: module.__dict__[k], ["num_embeddings", "embedding_dim"] + ) + quantized_module = PQEmbedding( + centroids, assignments, num_embeddings, embedding_dim + ) + elif isinstance(module, nn.Conv2d): + out_channels, in_channels, kernel_size = map( + lambda k: module.__dict__[k], + ["out_channels", "in_channels", "kernel_size"], + ) + stride, padding, dilation, groups, padding_mode = map( + lambda k: module.__dict__[k], + ["stride", "padding", "dilation", "groups", "padding_mode"], + ) + + quantized_module = PQConv2d( + centroids, + assignments, + bias, + in_channels, + out_channels, + kernel_size, + stride=stride, + padding=padding, + dilation=dilation, + groups=groups, + padding_mode=padding_mode, + ) + else: + raise ValueError(f"Module {module} not yet supported for quantization") + + # replace layer by its quantized counterpart + attrsetter(layer)(model, quantized_module) + + # update statistics + size_tracker.update(weight, block_size, n_centroids) + + # return name of quantized layers + return quantized_layers + + +def get_layers(model, filter_regexp): + """ + Filters out the layers according to a regexp. Note that + we omit biases. + + Args: + - model: a nn.Module + - filter_regexp: a regexp to filter the layers to keep + according to their name in model.named_parameters(). + For instance, the regexp: + + down_layers\\.[123456]\\.(conv[12]|identity\\.conv)) + + is keeping blocks down_layers from 1 to 6, and inside + each block is keeping conv1, conv2 and identity.conv. + + Remarks: + - We add (module\\.)? at the beginning of the regexp to + account for the possible use of nn.parallel.DataParallel + """ + + # get all parameter names + all_layers = map(itemgetter(0), model.named_parameters()) + + # remove biases + all_layers = filter(lambda x: "bias" not in x, all_layers) + + # remove .weight in all other names (or .weight_orig is spectral norm) + all_layers = map(lambda x: x.replace(".weight_orig", ""), all_layers) + all_layers = map(lambda x: x.replace(".weight", ""), all_layers) + + # return filtered layers + filter_regexp = "(module\\.)?" + "(" + filter_regexp + ")" + r = re.compile(filter_regexp) + + return list(filter(r.match, all_layers)) + + +def get_param(module, layer_name, param_config): + """ + Given a quantization configuration, get the right parameter + for the module to be quantized. + + Args: + - module: a nn.Module + - layer_name: the name of the layer + - param_config: a dict like + { + 'Conv2d': ('kernel_size', {'(3, 3)': 9, '(1, 1)': 4}), + 'Linear': ('in_features', {'*': 8}) + } + For instance, all conv2d layers with kernel size 3x3 have + a block size of 9 and all Linear layers are quantized with + a block size of 8, irrespective of their size. + + Remarks: + - if 'fuzzy_name' is passed as a parameter, layers whose layer_name + include 'fuzzy_name' will be assigned the given parameter. + In the following example, conv.expand layers will have a block + size of 9 while conv.reduce will have a block size of 4 and all + other layers will have a block size of 2. + { + 'Conv2d': ('fuzzy_name', {'expand': 9, 'reduce': 4, '*': 2}), + 'Linear': ('fuzzy_name', {'classifier': 8, 'projection': 4}) + } + + """ + + layer_type = module.__class__.__name__ + + if layer_type not in param_config: + raise KeyError(f"Layer type {layer_type} not in config for layer {module}") + + feature, params = param_config[module.__class__.__name__] + + if feature != "fuzzy_name": + feature_value = str(getattr(module, feature)) + if feature_value not in params: + if "*" in params: + feature_value = "*" + else: + raise KeyError( + f"{feature}={feature_value} not in config for layer {module}" + ) + else: + feature_values = [name for name in params if name in layer_name] + if len(feature_values) == 0: + if "*" in params: + feature_value = "*" + else: + raise KeyError(f"name={layer_name} not in config for {module}") + else: + feature_value = feature_values[0] + + return params[feature_value] + + +class SizeTracker(object): + """ + Class to keep track of the compressed network size with iPQ. + + Args: + - model: a nn.Module + + Remarks: + - The compressed size is the sum of three components + for each layer in the network: + (1) Storing the centroids given by iPQ in fp16 + (2) Storing the assignments of the blocks in int8 + (3) Storing all non-compressed elements such as biases + - This cost in only valid if we use 256 centroids (then + indexing can indeed by done with int8). + """ + + def __init__(self, model): + self.model = model + self.size_non_compressed_model = self.compute_size() + self.size_non_quantized = self.size_non_compressed_model + self.size_index = 0 + self.size_centroids = 0 + self.n_quantized_layers = 0 + + def compute_size(self): + """ + Computes the size of the model (in MB). + """ + + res = 0 + for _, p in self.model.named_parameters(): + res += p.numel() + return res * 4 / 1024 / 1024 + + def update(self, W, block_size, n_centroids): + """ + Updates the running statistics when quantizing a new layer. + """ + + # bits per weights + bits_per_weight = np.log2(n_centroids) / block_size + self.n_quantized_layers += 1 + + # size of indexing the subvectors of size block_size (in MB) + size_index_layer = bits_per_weight * W.numel() / 8 / 1024 / 1024 + self.size_index += size_index_layer + + # size of the centroids stored in float16 (in MB) + size_centroids_layer = n_centroids * block_size * 2 / 1024 / 1024 + self.size_centroids += size_centroids_layer + + # size of non-compressed layers, e.g. LayerNorms or biases (in MB) + size_uncompressed_layer = W.numel() * 4 / 1024 / 1024 + self.size_non_quantized -= size_uncompressed_layer + + def __repr__(self): + size_compressed = ( + self.size_index + self.size_centroids + self.size_non_quantized + ) + compression_ratio = self.size_non_compressed_model / size_compressed # NOQA + return ( + f"Non-compressed model size: {self.size_non_compressed_model:.2f} MB. " + f"After quantizing {self.n_quantized_layers} layers, size " + f"(indexing + centroids + other): {self.size_index:.2f} MB + " + f"{self.size_centroids:.2f} MB + {self.size_non_quantized:.2f} MB = " + f"{size_compressed:.2f} MB, compression ratio: {compression_ratio:.2f}x" + ) + + +def attrsetter(*items): + def resolve_attr(obj, attr): + attrs = attr.split(".") + head = attrs[:-1] + tail = attrs[-1] + + for name in head: + obj = getattr(obj, name) + return obj, tail + + def g(obj, val): + for attr in items: + resolved_obj, resolved_attr = resolve_attr(obj, attr) + setattr(resolved_obj, resolved_attr, val) + + return g diff --git a/fairseq/modules/quantization/quantization_options.py b/fairseq/modules/quantization/quantization_options.py new file mode 100644 index 0000000000000000000000000000000000000000..b46d682c0edaeaaf2a230e51d50da2a32d4bda98 --- /dev/null +++ b/fairseq/modules/quantization/quantization_options.py @@ -0,0 +1,44 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +def parse_config_yaml(yaml_data): + # Initialize to default options. + quantization_options = { + "n_centroids": { + "Linear": ["in_features", {"*": 256}], + "Embedding": ["embedding_dim", {"*": 256}], + }, + "block_sizes": { + "Linear": ["fuzzy_name", {"fc": 8, "attn": 4, "emb": 4}], + "Embedding": ["fuzzy_name", {"emb": 8}], + }, + "layers_to_quantize": [ + "decoder\\.layers\\.\\d+\\.fc[12]", + "decoder\\.embed_tokens\\.embeddings\\.[012]\\.[01]", + "decoder\\.layers\\.\\d+\\.self_attn\\.(k_proj|v_proj|q_proj|out_proj)", + ], + } + + if "n_centroids" in yaml_data: + quantization_options["n_centroids"] = { + layer: convert_yaml_to_tuple(layer_data) + for layer, layer_data in yaml_data["n_centroids"].items() + } + if "block_sizes" in yaml_data: + quantization_options["block_sizes"] = { + layer: convert_yaml_to_tuple(layer_data) + for layer, layer_data in yaml_data["block_sizes"].items() + } + if "layers_to_quantize" in yaml_data: + quantization_options["layers_to_quantize"] = yaml_data["layers_to_quantize"] + + return quantization_options + + +def convert_yaml_to_tuple(yaml_dictionary): + """Converts a yaml dictionary with two keys: `key` and `value` into a two + argument tuple of those values.""" + return (yaml_dictionary["key"], yaml_dictionary["value"]) diff --git a/fairseq/modules/quantization/scalar/__init__.py b/fairseq/modules/quantization/scalar/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..143834f3d036780eb6844c82f0c6f2d10cfe2f61 --- /dev/null +++ b/fairseq/modules/quantization/scalar/__init__.py @@ -0,0 +1,6 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .utils import quantize_model_ # NOQA diff --git a/fairseq/modules/quantization/scalar/modules/__init__.py b/fairseq/modules/quantization/scalar/modules/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..8031d9cdb23f2bc72596f8bc9cfa4965f96e3e6c --- /dev/null +++ b/fairseq/modules/quantization/scalar/modules/__init__.py @@ -0,0 +1,9 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .qact import ActivationQuantizer # NOQA +from .qconv import IntConv2d # NOQA +from .qemb import IntEmbedding # NOQA +from .qlinear import IntLinear # NOQA diff --git a/fairseq/modules/quantization/scalar/modules/qact.py b/fairseq/modules/quantization/scalar/modules/qact.py new file mode 100644 index 0000000000000000000000000000000000000000..c5dd1d63362423ab0cfc381dddabb547a3b44c72 --- /dev/null +++ b/fairseq/modules/quantization/scalar/modules/qact.py @@ -0,0 +1,88 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch + +from ..ops import emulate_int + + +class ActivationQuantizer: + """ + Fake scalar quantization of the activations using a forward hook. + + Args: + - module. a nn.Module for which we quantize the *post-activations* + - p: proportion of activations to quantize, set by default to 1 + - update_step: to recompute quantization parameters + - bits: number of bits for quantization + - method: choose among {"tensor", "histogram", "channel"} + - clamp_threshold: to prevent gradients overflow + + Remarks: + - Parameters scale and zero_point are recomputed every update_step + forward pass to reduce the overhead + - For the list of quantization methods and number of bits, see ops.py + - To remove the hook from the module, simply call self.handle.remove() + - At test time, the activations are fully quantized + - We use the straight-through estimator so that the gradients + back-propagate nicely in the network, this is implemented with + the detach() trick + - The activations are hard-clamped in [-clamp_threshold, clamp_threshold] + to prevent overflow during the backward pass + """ + + def __init__( + self, + module, + p=1, + update_step=1000, + bits=8, + method="histogram", + clamp_threshold=5, + ): + self.module = module + self.p = p + self.update_step = update_step + self.counter = 0 + self.bits = bits + self.method = method + self.clamp_threshold = clamp_threshold + self.handle = None + self.register_hook() + + def register_hook(self): + # forward hook + def quantize_hook(module, x, y): + + # update parameters every 1000 iterations + if self.counter % self.update_step == 0: + self.scale = None + self.zero_point = None + self.counter += 1 + + # train with QuantNoise and evaluate the fully quantized network + p = self.p if self.module.training else 1 + + # quantize activations + y_q, self.scale, self.zero_point = emulate_int( + y.detach(), + bits=self.bits, + method=self.method, + scale=self.scale, + zero_point=self.zero_point, + ) + + # mask to apply noise + mask = torch.zeros_like(y) + mask.bernoulli_(1 - p) + noise = (y_q - y).masked_fill(mask.bool(), 0) + + # using straight-through estimator (STE) + clamp_low = -self.scale * self.zero_point + clamp_high = self.scale * (2 ** self.bits - 1 - self.zero_point) + return torch.clamp(y, clamp_low.item(), clamp_high.item()) + noise.detach() + + # register hook + self.handle = self.module.register_forward_hook(quantize_hook) diff --git a/fairseq/modules/quantization/scalar/modules/qconv.py b/fairseq/modules/quantization/scalar/modules/qconv.py new file mode 100644 index 0000000000000000000000000000000000000000..83788c6f71fd41e61fd115681a22d53ce8b8362c --- /dev/null +++ b/fairseq/modules/quantization/scalar/modules/qconv.py @@ -0,0 +1,149 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn.functional as F +from torch.nn.modules.conv import _ConvNd +from torch.nn.modules.utils import _pair + +from ..ops import emulate_int + + +class IntConv2d(_ConvNd): + """ + Quantized counterpart of the nn.Conv2d module that applies QuantNoise during training. + + Args: + - standard nn.Conv2d parameters + - p: amount of noise to inject (0 = no quantization, 1 = quantize all the weights) + - bits: number of bits + - method: choose among {"tensor", "histogram", "channel"} + - update_step: recompute scale and zero_point every update_steps iterations + + Remarks: + - We use the straight-thgourh estimator so that the gradients + back-propagate nicely in the network, this is implemented with + the detach() trick + - Parameters scale and zero_point are recomputed every update_step + forward pass to reduce the overhead + - At test time, the weights are fully quantized + """ + + def __init__( + self, + in_channels, + out_channels, + kernel_size, + stride=1, + padding=0, + dilation=1, + groups=1, + bias=True, + padding_mode="zeros", + p=0, + bits=8, + method="histogram", + update_step=1000, + ): + kernel_size = _pair(kernel_size) + stride = _pair(stride) + padding = _pair(padding) + dilation = _pair(dilation) + super(IntConv2d, self).__init__( + in_channels, + out_channels, + kernel_size, + stride, + padding, + dilation, + False, + _pair(0), + groups, + bias, + padding_mode, + ) + + # quantization parameters + self.p = p + self.bits = bits + self.method = method + self.update_step = update_step + self.counter = 0 + + def _conv_forward(self, input, weight): + if self.padding_mode != "zeros": + return F.conv2d( + F.pad(input, self._padding_repeated_twice, mode=self.padding_mode), + weight, + self.bias, + self.stride, + _pair(0), + self.dilation, + self.groups, + ) + return F.conv2d( + input, + weight, + self.bias, + self.stride, + self.padding, + self.dilation, + self.groups, + ) + + def forward(self, input): + # train with QuantNoise and evaluate the fully quantized network + p = self.p if self.training else 1 + + # update parameters every 100 iterations + if self.counter % self.update_step == 0: + self.scale = None + self.zero_point = None + self.counter += 1 + + # quantize weight + weight_quantized, self.scale, self.zero_point = emulate_int( + self.weight.detach(), + bits=self.bits, + method=self.method, + scale=self.scale, + zero_point=self.zero_point, + ) + + # mask to apply noise + mask = torch.zeros_like(self.weight) + mask.bernoulli_(1 - p) + noise = (weight_quantized - self.weight).masked_fill(mask.bool(), 0) + + # using straight-through estimator (STE) + clamp_low = -self.scale * self.zero_point + clamp_high = self.scale * (2 ** self.bits - 1 - self.zero_point) + weight = ( + torch.clamp(self.weight, clamp_low.item(), clamp_high.item()) + + noise.detach() + ) + + # return output + output = self._conv_forward(input, weight) + return output + + def extra_repr(self): + return ( + "in_channels={}, out_channels={}, kernel_size={}, stride={}, " + "padding={}, dilation={}, groups={}, bias={}, quant_noise={}, " + "bits={}, method={}".format( + self.in_channels, + self.out_channels, + self.kernel_size, + self.stride, + self.padding, + self.dilation, + self.groups, + self.bias is not None, + self.p, + self.bits, + self.method, + ) + ) diff --git a/fairseq/modules/quantization/scalar/modules/qemb.py b/fairseq/modules/quantization/scalar/modules/qemb.py new file mode 100644 index 0000000000000000000000000000000000000000..d6cf06e5872cb86e5c2e726153c7a80c78db9d1e --- /dev/null +++ b/fairseq/modules/quantization/scalar/modules/qemb.py @@ -0,0 +1,147 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn as nn +import torch.nn.functional as F + +from ..ops import emulate_int + + +class IntEmbedding(nn.Module): + """ + Quantized counterpart of the nn.Embedding module that applies QuantNoise during training. + + Args: + - num_embeddings: number of tokens + - embedding_dim: embedding dimension + - p: amount of noise to inject (0 = no quantization, 1 = quantize all the weights) + - bits: number of bits + - method: choose among {"tensor", "histogram", "channel"} + - update_step: recompute scale and zero_point every update_steps iterations + + Remarks: + - We use the straight-through estimator so that the gradients + back-propagate nicely in the network, this is implemented with + the detach() trick + - Parameters scale and zero_point are recomputed every update_step + forward pass to reduce the overhead + - At test time, the weights are fully quantized + """ + + def __init__( + self, + num_embeddings, + embedding_dim, + padding_idx=None, + max_norm=None, + norm_type=2.0, + scale_grad_by_freq=False, + sparse=False, + _weight=None, + p=0, + update_step=1000, + bits=8, + method="histogram", + ): + super(IntEmbedding, self).__init__() + self.num_embeddings = num_embeddings + self.embedding_dim = embedding_dim + if padding_idx is not None: + if padding_idx > 0: + assert ( + padding_idx < self.num_embeddings + ), "Padding_idx must be within num_embeddings" + elif padding_idx < 0: + assert ( + padding_idx >= -self.num_embeddings + ), "Padding_idx must be within num_embeddings" + padding_idx = self.num_embeddings + padding_idx + self.padding_idx = padding_idx + self.max_norm = max_norm + self.norm_type = norm_type + self.scale_grad_by_freq = scale_grad_by_freq + if _weight is None: + self.weight = nn.Parameter(torch.Tensor(num_embeddings, embedding_dim)) + self.reset_parameters() + else: + assert list(_weight.shape) == [ + num_embeddings, + embedding_dim, + ], "Shape of weight does not match num_embeddings and embedding_dim" + self.weight = nn.Parameter(_weight) + self.sparse = sparse + + # quantization parameters + self.p = p + self.bits = bits + self.method = method + self.update_step = update_step + self.counter = 0 + + def reset_parameters(self): + nn.init.normal_(self.weight) + if self.padding_idx is not None: + with torch.no_grad(): + self.weight[self.padding_idx].fill_(0) + + def forward(self, input): + # train with QuantNoise and evaluate the fully quantized network + p = self.p if self.training else 1 + + # update parameters every 1000 iterations + if self.counter % self.update_step == 0: + self.scale = None + self.zero_point = None + self.counter += 1 + + # quantize weight + weight_quantized, self.scale, self.zero_point = emulate_int( + self.weight.detach(), + bits=self.bits, + method=self.method, + scale=self.scale, + zero_point=self.zero_point, + ) + + # mask to apply noise + mask = torch.zeros_like(self.weight) + mask.bernoulli_(1 - p) + noise = (weight_quantized - self.weight).masked_fill(mask.bool(), 0) + + # using straight-through estimator (STE) + clamp_low = -self.scale * self.zero_point + clamp_high = self.scale * (2 ** self.bits - 1 - self.zero_point) + weight = ( + torch.clamp(self.weight, clamp_low.item(), clamp_high.item()) + + noise.detach() + ) + + # return output + output = F.embedding( + input, + weight, + self.padding_idx, + self.max_norm, + self.norm_type, + self.scale_grad_by_freq, + self.sparse, + ) + return output + + def extra_repr(self): + s = "{num_embeddings}, {embedding_dim}" + if self.padding_idx is not None: + s += ", padding_idx={padding_idx}" + if self.max_norm is not None: + s += ", max_norm={max_norm}" + if self.norm_type != 2: + s += ", norm_type={norm_type}" + if self.scale_grad_by_freq is not False: + s += ", scale_grad_by_freq={scale_grad_by_freq}" + if self.sparse is not False: + s += ", sparse=True" + s += "quant_noise={p}, bits={bits}, method={method}" + return s.format(**self.__dict__) diff --git a/fairseq/modules/quantization/scalar/modules/qlinear.py b/fairseq/modules/quantization/scalar/modules/qlinear.py new file mode 100644 index 0000000000000000000000000000000000000000..9db1559386bce286301d31435851dc4ea76687a5 --- /dev/null +++ b/fairseq/modules/quantization/scalar/modules/qlinear.py @@ -0,0 +1,113 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn as nn +import torch.nn.functional as F + +from ..ops import emulate_int + + +class IntLinear(nn.Module): + """ + Quantized counterpart of the nn.Linear module that applies QuantNoise during training. + + Args: + - in_features: input features + - out_features: output features + - bias: bias or not + - p: amount of noise to inject (0 = no quantization, 1 = quantize all the weights) + - bits: number of bits + - method: choose among {"tensor", "histogram", "channel"} + - update_step: recompute scale and zero_point every update_steps iterations + + Remarks: + - We use the straight-through estimator so that the gradients + back-propagate nicely in the network, this is implemented with + the detach() trick. + - Parameters scale and zero_point are recomputed every update_step + forward pass to reduce the overhead + - At test time, the weights are fully quantized + """ + + def __init__( + self, + in_features, + out_features, + bias=True, + p=0, + update_step=3000, + bits=8, + method="histogram", + ): + super(IntLinear, self).__init__() + self.in_features = int(in_features) + self.out_features = int(out_features) + self.weight = torch.nn.Parameter(torch.Tensor(out_features, in_features)) + self.chosen_bias = bias + if self.chosen_bias: + self.bias = torch.nn.Parameter(torch.Tensor(out_features)) + else: + self.register_parameter("bias", None) + self.reset_parameters() + + # quantization parameters + self.p = p + self.bits = bits + self.method = method + self.update_step = update_step + self.counter = 0 + + def reset_parameters(self): + nn.init.xavier_uniform_(self.weight) + if self.chosen_bias: + nn.init.constant_(self.bias, 0.0) + return + + def forward(self, input): + # train with QuantNoise and evaluate the fully quantized network + p = self.p if self.training else 1 + + # update parameters every 100 iterations + if self.counter % self.update_step == 0: + self.scale = None + self.zero_point = None + self.counter += 1 + + # quantize weight + weight_quantized, self.scale, self.zero_point = emulate_int( + self.weight.detach(), + bits=self.bits, + method=self.method, + scale=self.scale, + zero_point=self.zero_point, + ) + + # mask to apply noise + mask = torch.zeros_like(self.weight) + mask.bernoulli_(1 - p) + noise = (weight_quantized - self.weight).masked_fill(mask.bool(), 0) + + # using straight-through estimator (STE) + clamp_low = -self.scale * self.zero_point + clamp_high = self.scale * (2 ** self.bits - 1 - self.zero_point) + weight = ( + torch.clamp(self.weight, clamp_low.item(), clamp_high.item()) + + noise.detach() + ) + + # return output + output = F.linear(input, weight, self.bias) + return output + + def extra_repr(self): + return "in_features={}, out_features={}, bias={}, quant_noise={}, bits={}, method={}".format( + self.in_features, + self.out_features, + self.bias is not None, + self.p, + self.bits, + self.method, + ) diff --git a/fairseq/modules/quantization/scalar/ops.py b/fairseq/modules/quantization/scalar/ops.py new file mode 100644 index 0000000000000000000000000000000000000000..2a855159be2795bdad45f1365e202d9abd26433b --- /dev/null +++ b/fairseq/modules/quantization/scalar/ops.py @@ -0,0 +1,49 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch + + +def emulate_int(w, bits, method, scale=None, zero_point=None): + q = globals()[f"emulate_int{bits}_{method}"] + return q(w, scale=scale, zero_point=zero_point) + + +def quantize(w, scale, zero_point): + return ( + torch.clamp(torch.round(w / scale + zero_point), 0, 255) - zero_point + ) * scale + + +def emulate_int8_histogram(w, scale=None, zero_point=None): + if scale is None: + obs = torch.quantization.observer.HistogramObserver() + _ = obs(w.float()) + scale, zero_point = obs.calculate_qparams() + scale = scale.cuda().type_as(w) + zero_point = zero_point.cuda().type_as(w) + return quantize(w, scale, zero_point), scale, zero_point + + +def emulate_int8_channel(w, scale=None, zero_point=None): + if scale is None: + obs = torch.quantization.observer.PerChannelMinMaxObserver( + ch_axis=-1, qscheme=torch.per_channel_symmetric + ) + _ = obs(w) + scale, zero_point, ch_axis = obs.get_qparams() + scale = scale.cuda().type_as(w) + zero_point = zero_point.cuda().type_as(w) + return quantize(w, scale, zero_point), scale, zero_point + + +def emulate_int8_tensor(w, scale=None, zero_point=None): + if scale is None: + obs = torch.quantization.observer.MinMaxObserver() + _ = obs(w) + scale, zero_point = obs.calculate_qparams() + scale = scale.cuda().type_as(w) + zero_point = zero_point.cuda().type_as(w) + return quantize(w, scale, zero_point), scale, zero_point diff --git a/fairseq/modules/quantization/scalar/utils.py b/fairseq/modules/quantization/scalar/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..32cf616568160004bd97a673f2d85923974c1fae --- /dev/null +++ b/fairseq/modules/quantization/scalar/utils.py @@ -0,0 +1,77 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +from operator import attrgetter + +import torch.distributed as dist +import torch.nn as nn + +from ..pq.utils import attrsetter, get_layers +from .modules import ActivationQuantizer, IntConv2d, IntEmbedding, IntLinear + + +MAPPING = {nn.Linear: IntLinear, nn.Embedding: IntEmbedding, nn.Conv2d: IntConv2d} + + +def quantize_model_(model, p=0.2, bits=8, update_step=3000): + """ + Replaces all modules with their scalar quantized counterpart and + registers hooks to quantize the post-ativations of those modules. + + Args: + - model: a nn.Module + - p: amount of noise (0 for no noise, 1 to quantize all the weights/activations) + - bits: number of bits + - update_step: update quantization parameters every update_step steps + """ + + # quantize all layers + quantized_layers = get_layers(model, "(.*?)") + + for layer in quantized_layers: + + # book-keeping + is_master_process = (not dist.is_initialized()) or ( + dist.is_initialized() and dist.get_rank() == 0 + ) + + # recover module + module = attrgetter(layer)(model) + if is_master_process: + logging.info( + f"Quantizing layer {layer} with bits={bits} and QuantNoise={p}" + ) + + # quantization params + q_params = { + "p": p, + "update_step": update_step, + "bits": bits, + "method": "histogram", + "counter": 0, + } + + # instantiate the quantized counterpart + if isinstance(module, tuple(MAPPING.keys())): + QuantizedModule = MAPPING[module.__class__] + quantized_module = QuantizedModule.__new__(QuantizedModule) + params = module.__dict__ + params.update(q_params) + quantized_module.__dict__.update(params) + + else: + if is_master_process: + logging.info(f"Module {module} not yet supported for quantization") + continue + + # activation quantization + a_q = ActivationQuantizer(quantized_module, p=0, bits=bits, method="histogram") + + # replace layer by its quantized counterpart + attrsetter(layer)(model, quantized_module) + + # return name of quantized layers + return quantized_layers diff --git a/fairseq/modules/same_pad.py b/fairseq/modules/same_pad.py new file mode 100644 index 0000000000000000000000000000000000000000..4c04990ea6fdb291f162ee8ac3d17a92483daf8e --- /dev/null +++ b/fairseq/modules/same_pad.py @@ -0,0 +1,21 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +from torch import nn + + +class SamePad(nn.Module): + def __init__(self, kernel_size, causal=False): + super().__init__() + if causal: + self.remove = kernel_size - 1 + else: + self.remove = 1 if kernel_size % 2 == 0 else 0 + + def forward(self, x): + if self.remove > 0: + x = x[:, :, : -self.remove] + return x diff --git a/fairseq/modules/scalar_bias.py b/fairseq/modules/scalar_bias.py new file mode 100644 index 0000000000000000000000000000000000000000..c96247c75914fabb8a2b7ff731bb82b588f72690 --- /dev/null +++ b/fairseq/modules/scalar_bias.py @@ -0,0 +1,31 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +# + +import torch + + +class ScalarBias(torch.autograd.Function): + """ + Adds a vector of scalars, used in self-attention mechanism to allow + the model to optionally attend to this vector instead of the past + """ + + @staticmethod + def forward(ctx, input, dim, bias_init): + size = list(input.size()) + size[dim] += 1 + output = input.new(*size).fill_(bias_init) + output.narrow(dim, 1, size[dim] - 1).copy_(input) + ctx.dim = dim + return output + + @staticmethod + def backward(ctx, grad): + return grad.narrow(ctx.dim, 1, grad.size(ctx.dim) - 1), None, None + + +def scalar_bias(input, dim, bias_init=0): + return ScalarBias.apply(input, dim, bias_init) diff --git a/fairseq/modules/sinusoidal_positional_embedding.py b/fairseq/modules/sinusoidal_positional_embedding.py new file mode 100644 index 0000000000000000000000000000000000000000..4793ecfb522d0729fc2d24a3ddf0c6a774d67773 --- /dev/null +++ b/fairseq/modules/sinusoidal_positional_embedding.py @@ -0,0 +1,105 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +from typing import Any, Optional + +import torch +import torch.onnx.operators +from fairseq import utils +from torch import Tensor, nn + + +class SinusoidalPositionalEmbedding(nn.Module): + """This module produces sinusoidal positional embeddings of any length. + + Padding symbols are ignored. + """ + + def __init__(self, embedding_dim, padding_idx, init_size=1024): + super().__init__() + self.embedding_dim = embedding_dim + self.padding_idx = padding_idx if padding_idx is not None else 0 + self.weights = SinusoidalPositionalEmbedding.get_embedding( + init_size, embedding_dim, padding_idx + ) + self.onnx_trace = False + self.register_buffer("_float_tensor", torch.FloatTensor(1)) + self.max_positions = int(1e5) + + def prepare_for_onnx_export_(self): + self.onnx_trace = True + + @staticmethod + def get_embedding( + num_embeddings: int, embedding_dim: int, padding_idx: Optional[int] = None + ): + """Build sinusoidal embeddings. + + This matches the implementation in tensor2tensor, but differs slightly + from the description in Section 3.5 of "Attention Is All You Need". + """ + half_dim = embedding_dim // 2 + emb = math.log(10000) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=torch.float) * -emb) + emb = torch.arange(num_embeddings, dtype=torch.float).unsqueeze( + 1 + ) * emb.unsqueeze(0) + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1).view( + num_embeddings, -1 + ) + if embedding_dim % 2 == 1: + # zero pad + emb = torch.cat([emb, torch.zeros(num_embeddings, 1)], dim=1) + if padding_idx is not None: + emb[padding_idx, :] = 0 + return emb + + def forward( + self, + input, + incremental_state: Optional[Any] = None, + timestep: Optional[Tensor] = None, + positions: Optional[Any] = None, + ): + """Input is expected to be of size [bsz x seqlen].""" + bspair = torch.onnx.operators.shape_as_tensor(input) + bsz, seq_len = bspair[0], bspair[1] + max_pos = self.padding_idx + 1 + seq_len + if self.weights is None or max_pos > self.weights.size(0): + # recompute/expand embeddings if needed + self.weights = SinusoidalPositionalEmbedding.get_embedding( + max_pos, self.embedding_dim, self.padding_idx + ) + self.weights = self.weights.to(self._float_tensor) + + if incremental_state is not None: + # positions is the same for every token when decoding a single step + pos = timestep.view(-1)[0] + 1 if timestep is not None else seq_len + if self.onnx_trace: + return ( + self.weights.index_select(index=self.padding_idx + pos, dim=0) + .unsqueeze(1) + .repeat(bsz, 1, 1) + ) + return self.weights[self.padding_idx + pos, :].expand(bsz, 1, -1) + + positions = utils.make_positions( + input, self.padding_idx, onnx_trace=self.onnx_trace + ) + if self.onnx_trace: + flat_embeddings = self.weights.detach().index_select(0, positions.view(-1)) + embedding_shape = torch.cat( + (bsz.view(1), seq_len.view(1), torch.tensor([-1], dtype=torch.long)) + ) + embeddings = torch.onnx.operators.reshape_from_tensor_shape( + flat_embeddings, embedding_shape + ) + return embeddings + return ( + self.weights.index_select(0, positions.view(-1)) + .view(bsz, seq_len, -1) + .detach() + ) diff --git a/fairseq/modules/sparse_multihead_attention.py b/fairseq/modules/sparse_multihead_attention.py new file mode 100644 index 0000000000000000000000000000000000000000..3cbd9d6785886e319aab0601517e27df733b6f97 --- /dev/null +++ b/fairseq/modules/sparse_multihead_attention.py @@ -0,0 +1,140 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +import torch + +from .multihead_attention import MultiheadAttention + + +class SparseMultiheadAttention(MultiheadAttention): + """Sparse Multi-Headed Attention. + + "Generating Long Sequences with Sparse Transformers". Implements + fixed factorized self attention, where l=stride and c=expressivity. + A(1) includes all words in the stride window and A(2) takes a summary of c + words from the end of each stride window. + If is_bidirectional=False, we do not include any words past the current word, + as in the paper. + """ + + def __init__( + self, + embed_dim, + num_heads, + kdim=None, + vdim=None, + dropout=0.0, + bias=True, + add_bias_kv=False, + add_zero_attn=False, + self_attention=False, + encoder_decoder_attention=False, + stride=32, + expressivity=8, + is_bidirectional=True, + ): + + super().__init__( + embed_dim, + num_heads, + kdim, + vdim, + dropout, + bias, + add_bias_kv, + add_zero_attn, + self_attention, + encoder_decoder_attention, + ) + + self.is_bidirectional = is_bidirectional + self.stride = stride + self.expressivity = expressivity + assert self.stride > 0 and self.stride >= self.expressivity + + # Used for Ai(2) calculations - beginning of [l-c, l] range + def compute_checkpoint(self, word_index): + if word_index % self.stride == 0 and word_index != 0: + checkpoint_index = word_index - self.expressivity + else: + checkpoint_index = ( + math.floor(word_index / self.stride) * self.stride + + self.stride + - self.expressivity + ) + return checkpoint_index + + # Computes Ai(2) + def compute_subset_summaries(self, absolute_max): + checkpoint_index = self.compute_checkpoint(0) + subset_two = set() + while checkpoint_index <= absolute_max - 1: + summary = set( + range( + checkpoint_index, + min(checkpoint_index + self.expressivity + 1, absolute_max), + ) + ) + subset_two = subset_two.union(summary) + checkpoint_index = self.compute_checkpoint(checkpoint_index + self.stride) + return subset_two + + # Sparse Transformer Fixed Attention Pattern: https://arxiv.org/pdf/1904.10509.pdf + def compute_fixed_attention_subset(self, word_index, tgt_len): + # +1s account for range function; [min, max) -> [min, max] + if not self.is_bidirectional: + absolute_max = word_index + 1 + else: + absolute_max = tgt_len + + # Subset 1 - whole window + rounded_index = ( + math.floor((word_index + self.stride) / self.stride) * self.stride + ) + if word_index % self.stride == 0 and word_index != 0: + subset_one = set( + range(word_index - self.stride, min(absolute_max, word_index + 1)) + ) + else: + subset_one = set( + range( + max(0, rounded_index - self.stride), + min(absolute_max, rounded_index + 1), + ) + ) + + # Subset 2 - summary per window + # If bidirectional, subset 2 is the same for every index + subset_two = set() + if not self.is_bidirectional: + subset_two = self.compute_subset_summaries(absolute_max) + + return subset_one.union(subset_two) + + # Compute sparse mask - if bidirectional, can pre-compute and store + def buffered_sparse_mask(self, tensor, tgt_len, src_len): + assert tgt_len > self.stride + sparse_mask = torch.empty((tgt_len, src_len)).float().fill_(float("-inf")) + + # If bidirectional, subset 2 is the same for every index + subset_summaries = set() + if self.is_bidirectional: + subset_summaries = self.compute_subset_summaries(tgt_len) + + for i in range(tgt_len): + fixed_attention_subset = self.compute_fixed_attention_subset(i, tgt_len) + fixed_attention_subset = fixed_attention_subset.union(subset_summaries) + included_word_indices = torch.LongTensor(list(fixed_attention_subset)) + sparse_mask[i].index_fill_(0, included_word_indices, 0) + return sparse_mask.type_as(tensor) + + def apply_sparse_mask(self, attn_weights, tgt_len, src_len, bsz): + sparse_mask = self.buffered_sparse_mask(attn_weights, tgt_len, src_len) + sparse_mask = sparse_mask.unsqueeze(0).expand( + bsz * self.num_heads, tgt_len, src_len + ) + attn_weights += sparse_mask diff --git a/fairseq/modules/sparse_transformer_sentence_encoder.py b/fairseq/modules/sparse_transformer_sentence_encoder.py new file mode 100644 index 0000000000000000000000000000000000000000..f41ec09327fe80b50d20674e7482794ce45c531c --- /dev/null +++ b/fairseq/modules/sparse_transformer_sentence_encoder.py @@ -0,0 +1,96 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch.nn as nn +from fairseq.modules import TransformerSentenceEncoder +from fairseq.modules.sparse_transformer_sentence_encoder_layer import ( + SparseTransformerSentenceEncoderLayer, +) + + +class SparseTransformerSentenceEncoder(TransformerSentenceEncoder): + """ + Sparse implementation of the TransformerSentenceEncoder + - see SparseMultiheadAttention + """ + + def __init__( + self, + padding_idx: int, + vocab_size: int, + num_encoder_layers: int = 6, + embedding_dim: int = 768, + ffn_embedding_dim: int = 3072, + num_attention_heads: int = 8, + dropout: float = 0.1, + attention_dropout: float = 0.1, + activation_dropout: float = 0.1, + max_seq_len: int = 256, + num_segments: int = 2, + use_position_embeddings: bool = True, + offset_positions_by_padding: bool = True, + encoder_normalize_before: bool = False, + apply_bert_init: bool = False, + activation_fn: str = "relu", + learned_pos_embedding: bool = True, + embed_scale: float = None, + freeze_embeddings: bool = False, + n_trans_layers_to_freeze: int = 0, + export: bool = False, + is_bidirectional: bool = True, + stride: int = 32, + expressivity: int = 8, + ) -> None: + + super().__init__( + padding_idx, + vocab_size, + num_encoder_layers, + embedding_dim, + ffn_embedding_dim, + num_attention_heads, + dropout, + attention_dropout, + activation_dropout, + max_seq_len, + num_segments, + use_position_embeddings, + offset_positions_by_padding, + encoder_normalize_before, + apply_bert_init, + activation_fn, + learned_pos_embedding, + embed_scale, + freeze_embeddings, + n_trans_layers_to_freeze, + export, + ) + + self.layers = nn.ModuleList( + [ + SparseTransformerSentenceEncoderLayer( + embedding_dim=self.embedding_dim, + ffn_embedding_dim=ffn_embedding_dim, + num_attention_heads=num_attention_heads, + dropout=dropout, + attention_dropout=attention_dropout, + activation_dropout=activation_dropout, + activation_fn=activation_fn, + export=export, + is_bidirectional=is_bidirectional, + stride=stride, + expressivity=expressivity, + ) + for _ in range(num_encoder_layers) + ] + ) + + def freeze_module_params(m): + if m is not None: + for p in m.parameters(): + p.requires_grad = False + + for layer in range(n_trans_layers_to_freeze): + freeze_module_params(self.layers[layer]) diff --git a/fairseq/modules/sparse_transformer_sentence_encoder_layer.py b/fairseq/modules/sparse_transformer_sentence_encoder_layer.py new file mode 100644 index 0000000000000000000000000000000000000000..d95da59c2471bfa858fd627605196d7f41f9ec12 --- /dev/null +++ b/fairseq/modules/sparse_transformer_sentence_encoder_layer.py @@ -0,0 +1,51 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq.modules import TransformerSentenceEncoderLayer +from fairseq.modules.sparse_multihead_attention import SparseMultiheadAttention + + +class SparseTransformerSentenceEncoderLayer(TransformerSentenceEncoderLayer): + """ + Implements a Sprase Transformer Encoder Layer (see SparseMultiheadAttention) + """ + + def __init__( + self, + embedding_dim: int = 768, + ffn_embedding_dim: int = 3072, + num_attention_heads: int = 8, + dropout: float = 0.1, + attention_dropout: float = 0.1, + activation_dropout: float = 0.1, + activation_fn: str = "relu", + export: bool = False, + is_bidirectional: bool = True, + stride: int = 32, + expressivity: int = 8, + ) -> None: + + super().__init__( + embedding_dim, + ffn_embedding_dim, + num_attention_heads, + dropout, + attention_dropout, + activation_dropout, + activation_fn, + export, + ) + + self.self_attn = SparseMultiheadAttention( + self.embedding_dim, + num_attention_heads, + dropout=attention_dropout, + add_bias_kv=False, + add_zero_attn=False, + self_attention=True, + is_bidirectional=is_bidirectional, + stride=stride, + expressivity=expressivity, + ) diff --git a/fairseq/modules/transformer_layer.py b/fairseq/modules/transformer_layer.py new file mode 100644 index 0000000000000000000000000000000000000000..4f9ea22a9b9e27d78d4b66ce1268379b7b158002 --- /dev/null +++ b/fairseq/modules/transformer_layer.py @@ -0,0 +1,414 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from typing import Dict, List, Optional + +import torch +import torch.nn as nn +from fairseq import utils +from fairseq.modules import LayerNorm, MultiheadAttention +from fairseq.modules.fairseq_dropout import FairseqDropout +from fairseq.modules.quant_noise import quant_noise +from torch import Tensor + + +class TransformerEncoderLayer(nn.Module): + """Encoder layer block. + + In the original paper each operation (multi-head attention or FFN) is + postprocessed with: `dropout -> add residual -> layernorm`. In the + tensor2tensor code they suggest that learning is more robust when + preprocessing each layer with layernorm and postprocessing with: + `dropout -> add residual`. We default to the approach in the paper, but the + tensor2tensor approach can be enabled by setting + *args.encoder_normalize_before* to ``True``. + + Args: + args (argparse.Namespace): parsed command-line arguments + """ + + def __init__(self, args): + super().__init__() + self.args = args + self.embed_dim = args.encoder_embed_dim + self.quant_noise = getattr(args, 'quant_noise_pq', 0) + self.quant_noise_block_size = getattr(args, 'quant_noise_pq_block_size', 8) or 8 + self.self_attn = self.build_self_attention(self.embed_dim, args) + export = getattr(args, "export", False) + self.self_attn_layer_norm = LayerNorm(self.embed_dim, export=export) + self.dropout_module = FairseqDropout( + args.dropout, module_name=self.__class__.__name__ + ) + self.activation_fn = utils.get_activation_fn( + activation=getattr(args, 'activation_fn', 'relu') or "relu" + ) + activation_dropout_p = getattr(args, "activation_dropout", 0) or 0 + if activation_dropout_p == 0: + # for backwards compatibility with models that use args.relu_dropout + activation_dropout_p = getattr(args, "relu_dropout", 0) or 0 + self.activation_dropout_module = FairseqDropout( + float(activation_dropout_p), module_name=self.__class__.__name__ + ) + self.normalize_before = args.encoder_normalize_before + self.fc1 = self.build_fc1( + self.embed_dim, + args.encoder_ffn_embed_dim, + self.quant_noise, + self.quant_noise_block_size, + ) + self.fc2 = self.build_fc2( + args.encoder_ffn_embed_dim, + self.embed_dim, + self.quant_noise, + self.quant_noise_block_size, + ) + + self.final_layer_norm = LayerNorm(self.embed_dim, export=export) + + def build_fc1(self, input_dim, output_dim, q_noise, qn_block_size): + return quant_noise( + nn.Linear(input_dim, output_dim), p=q_noise, block_size=qn_block_size + ) + + def build_fc2(self, input_dim, output_dim, q_noise, qn_block_size): + return quant_noise( + nn.Linear(input_dim, output_dim), p=q_noise, block_size=qn_block_size + ) + + def build_self_attention(self, embed_dim, args): + return MultiheadAttention( + embed_dim, + args.encoder_attention_heads, + dropout=args.attention_dropout, + self_attention=True, + q_noise=self.quant_noise, + qn_block_size=self.quant_noise_block_size, + ) + + def residual_connection(self, x, residual): + return residual + x + + def upgrade_state_dict_named(self, state_dict, name): + """ + Rename layer norm states from `...layer_norms.0.weight` to + `...self_attn_layer_norm.weight` and `...layer_norms.1.weight` to + `...final_layer_norm.weight` + """ + layer_norm_map = {"0": "self_attn_layer_norm", "1": "final_layer_norm"} + for old, new in layer_norm_map.items(): + for m in ("weight", "bias"): + k = "{}.layer_norms.{}.{}".format(name, old, m) + if k in state_dict: + state_dict["{}.{}.{}".format(name, new, m)] = state_dict[k] + del state_dict[k] + + def forward(self, x, encoder_padding_mask: Optional[Tensor], attn_mask: Optional[Tensor] = None): + """ + Args: + x (Tensor): input to the layer of shape `(seq_len, batch, embed_dim)` + encoder_padding_mask (ByteTensor): binary ByteTensor of shape + `(batch, seq_len)` where padding elements are indicated by ``1``. + attn_mask (ByteTensor): binary tensor of shape `(tgt_len, src_len)`, + where `tgt_len` is the length of output and `src_len` is the + length of input, though here both are equal to `seq_len`. + `attn_mask[tgt_i, src_j] = 1` means that when calculating the + embedding for `tgt_i`, we exclude (mask out) `src_j`. This is + useful for strided self-attention. + + Returns: + encoded output of shape `(seq_len, batch, embed_dim)` + """ + # anything in original attn_mask = 1, becomes -1e8 + # anything in original attn_mask = 0, becomes 0 + # Note that we cannot use -inf here, because at some edge cases, + # the attention weight (before softmax) for some padded element in query + # will become -inf, which results in NaN in model parameters + if attn_mask is not None: + attn_mask = attn_mask.masked_fill(attn_mask.to(torch.bool), -1e8) + + residual = x + if self.normalize_before: + x = self.self_attn_layer_norm(x) + x, _ = self.self_attn( + query=x, + key=x, + value=x, + key_padding_mask=encoder_padding_mask, + need_weights=False, + attn_mask=attn_mask, + ) + x = self.dropout_module(x) + x = self.residual_connection(x, residual) + if not self.normalize_before: + x = self.self_attn_layer_norm(x) + + residual = x + if self.normalize_before: + x = self.final_layer_norm(x) + x = self.activation_fn(self.fc1(x)) + x = self.activation_dropout_module(x) + x = self.fc2(x) + x = self.dropout_module(x) + x = self.residual_connection(x, residual) + if not self.normalize_before: + x = self.final_layer_norm(x) + return x + + +class TransformerDecoderLayer(nn.Module): + """Decoder layer block. + + In the original paper each operation (multi-head attention, encoder + attention or FFN) is postprocessed with: `dropout -> add residual -> + layernorm`. In the tensor2tensor code they suggest that learning is more + robust when preprocessing each layer with layernorm and postprocessing with: + `dropout -> add residual`. We default to the approach in the paper, but the + tensor2tensor approach can be enabled by setting + *args.decoder_normalize_before* to ``True``. + + Args: + args (argparse.Namespace): parsed command-line arguments + no_encoder_attn (bool, optional): whether to attend to encoder outputs + (default: False). + """ + + def __init__( + self, args, no_encoder_attn=False, add_bias_kv=False, add_zero_attn=False + ): + super().__init__() + self.embed_dim = args.decoder_embed_dim + self.dropout_module = FairseqDropout( + args.dropout, module_name=self.__class__.__name__ + ) + self.quant_noise = getattr(args, "quant_noise_pq", 0) + self.quant_noise_block_size = getattr(args, "quant_noise_pq_block_size", 8) + + self.cross_self_attention = getattr(args, "cross_self_attention", False) + + self.self_attn = self.build_self_attention( + self.embed_dim, + args, + add_bias_kv=add_bias_kv, + add_zero_attn=add_zero_attn, + ) + + self.activation_fn = utils.get_activation_fn( + activation=str(args.activation_fn) + if getattr(args, "activation_fn", None) is not None + else "relu" + ) + activation_dropout_p = getattr(args, "activation_dropout", 0) or 0 + if activation_dropout_p == 0: + # for backwards compatibility with models that use args.relu_dropout + activation_dropout_p = getattr(args, "relu_dropout", 0) or 0 + self.activation_dropout_module = FairseqDropout( + float(activation_dropout_p), module_name=self.__class__.__name__ + ) + self.normalize_before = args.decoder_normalize_before + + export = getattr(args, "export", False) + self.self_attn_layer_norm = LayerNorm(self.embed_dim, export=export) + + if no_encoder_attn: + self.encoder_attn = None + self.encoder_attn_layer_norm = None + else: + self.encoder_attn = self.build_encoder_attention(self.embed_dim, args) + self.encoder_attn_layer_norm = LayerNorm(self.embed_dim, export=export) + + self.fc1 = self.build_fc1( + self.embed_dim, + args.decoder_ffn_embed_dim, + self.quant_noise, + self.quant_noise_block_size, + ) + self.fc2 = self.build_fc2( + args.decoder_ffn_embed_dim, + self.embed_dim, + self.quant_noise, + self.quant_noise_block_size, + ) + + self.final_layer_norm = LayerNorm(self.embed_dim, export=export) + self.need_attn = True + + self.onnx_trace = False + + def build_fc1(self, input_dim, output_dim, q_noise, qn_block_size): + return quant_noise(nn.Linear(input_dim, output_dim), q_noise, qn_block_size) + + def build_fc2(self, input_dim, output_dim, q_noise, qn_block_size): + return quant_noise(nn.Linear(input_dim, output_dim), q_noise, qn_block_size) + + def build_self_attention( + self, embed_dim, args, add_bias_kv=False, add_zero_attn=False + ): + return MultiheadAttention( + embed_dim, + args.decoder_attention_heads, + dropout=args.attention_dropout, + add_bias_kv=add_bias_kv, + add_zero_attn=add_zero_attn, + self_attention=not getattr(args, "cross_self_attention", False), + q_noise=self.quant_noise, + qn_block_size=self.quant_noise_block_size, + ) + + def build_encoder_attention(self, embed_dim, args): + return MultiheadAttention( + embed_dim, + args.decoder_attention_heads, + kdim=getattr(args, "encoder_embed_dim", None), + vdim=getattr(args, "encoder_embed_dim", None), + dropout=args.attention_dropout, + encoder_decoder_attention=True, + q_noise=self.quant_noise, + qn_block_size=self.quant_noise_block_size, + ) + + def prepare_for_onnx_export_(self): + self.onnx_trace = True + + def residual_connection(self, x, residual): + return residual + x + + def forward( + self, + x, + encoder_out: Optional[torch.Tensor] = None, + encoder_padding_mask: Optional[torch.Tensor] = None, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + prev_self_attn_state: Optional[List[torch.Tensor]] = None, + prev_attn_state: Optional[List[torch.Tensor]] = None, + self_attn_mask: Optional[torch.Tensor] = None, + self_attn_padding_mask: Optional[torch.Tensor] = None, + need_attn: bool = False, + need_head_weights: bool = False, + ): + """ + Args: + x (Tensor): input to the layer of shape `(seq_len, batch, embed_dim)` + encoder_padding_mask (ByteTensor, optional): binary + ByteTensor of shape `(batch, src_len)` where padding + elements are indicated by ``1``. + need_attn (bool, optional): return attention weights + need_head_weights (bool, optional): return attention weights + for each head (default: return average over heads). + + Returns: + encoded output of shape `(seq_len, batch, embed_dim)` + """ + if need_head_weights: + need_attn = True + + residual = x + if self.normalize_before: + x = self.self_attn_layer_norm(x) + if prev_self_attn_state is not None: + prev_key, prev_value = prev_self_attn_state[:2] + saved_state: Dict[str, Optional[Tensor]] = { + "prev_key": prev_key, + "prev_value": prev_value, + } + if len(prev_self_attn_state) >= 3: + saved_state["prev_key_padding_mask"] = prev_self_attn_state[2] + assert incremental_state is not None + self.self_attn._set_input_buffer(incremental_state, saved_state) + _self_attn_input_buffer = self.self_attn._get_input_buffer(incremental_state) + if self.cross_self_attention and not ( + incremental_state is not None + and _self_attn_input_buffer is not None + and "prev_key" in _self_attn_input_buffer + ): + if self_attn_mask is not None: + assert encoder_out is not None + self_attn_mask = torch.cat( + (x.new_zeros(x.size(0), encoder_out.size(0)), self_attn_mask), dim=1 + ) + if self_attn_padding_mask is not None: + if encoder_padding_mask is None: + assert encoder_out is not None + encoder_padding_mask = self_attn_padding_mask.new_zeros( + encoder_out.size(1), encoder_out.size(0) + ) + self_attn_padding_mask = torch.cat( + (encoder_padding_mask, self_attn_padding_mask), dim=1 + ) + assert encoder_out is not None + y = torch.cat((encoder_out, x), dim=0) + else: + y = x + + x, attn = self.self_attn( + query=x, + key=y, + value=y, + key_padding_mask=self_attn_padding_mask, + incremental_state=incremental_state, + need_weights=False, + attn_mask=self_attn_mask, + ) + x = self.dropout_module(x) + x = self.residual_connection(x, residual) + if not self.normalize_before: + x = self.self_attn_layer_norm(x) + + if self.encoder_attn is not None and encoder_out is not None: + residual = x + if self.normalize_before: + x = self.encoder_attn_layer_norm(x) + if prev_attn_state is not None: + prev_key, prev_value = prev_attn_state[:2] + saved_state: Dict[str, Optional[Tensor]] = { + "prev_key": prev_key, + "prev_value": prev_value, + } + if len(prev_attn_state) >= 3: + saved_state["prev_key_padding_mask"] = prev_attn_state[2] + assert incremental_state is not None + self.encoder_attn._set_input_buffer(incremental_state, saved_state) + + x, attn = self.encoder_attn( + query=x, + key=encoder_out, + value=encoder_out, + key_padding_mask=encoder_padding_mask, + incremental_state=incremental_state, + static_kv=True, + need_weights=need_attn or (not self.training and self.need_attn), + need_head_weights=need_head_weights, + ) + x = self.dropout_module(x) + x = self.residual_connection(x, residual) + if not self.normalize_before: + x = self.encoder_attn_layer_norm(x) + + residual = x + if self.normalize_before: + x = self.final_layer_norm(x) + + x = self.activation_fn(self.fc1(x)) + x = self.activation_dropout_module(x) + x = self.fc2(x) + x = self.dropout_module(x) + x = self.residual_connection(x, residual) + if not self.normalize_before: + x = self.final_layer_norm(x) + if self.onnx_trace and incremental_state is not None: + saved_state = self.self_attn._get_input_buffer(incremental_state) + assert saved_state is not None + if self_attn_padding_mask is not None: + self_attn_state = [ + saved_state["prev_key"], + saved_state["prev_value"], + saved_state["prev_key_padding_mask"], + ] + else: + self_attn_state = [saved_state["prev_key"], saved_state["prev_value"]] + return x, attn, self_attn_state + return x, attn, None + + def make_generation_fast_(self, need_attn: bool = False, **kwargs): + self.need_attn = need_attn diff --git a/fairseq/modules/transformer_sentence_encoder.py b/fairseq/modules/transformer_sentence_encoder.py new file mode 100644 index 0000000000000000000000000000000000000000..d0540d69229fb994b9e573a5016c9f239b7929e2 --- /dev/null +++ b/fairseq/modules/transformer_sentence_encoder.py @@ -0,0 +1,291 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from typing import Optional, Tuple + +import torch +import torch.nn as nn +from fairseq.modules import ( + FairseqDropout, + LayerDropModuleList, + LayerNorm, + MultiheadAttention, + PositionalEmbedding, + TransformerSentenceEncoderLayer, +) +from fairseq.modules.quant_noise import quant_noise as apply_quant_noise_ + + +def init_bert_params(module): + """ + Initialize the weights specific to the BERT Model. + This overrides the default initializations depending on the specified arguments. + 1. If normal_init_linear_weights is set then weights of linear + layer will be initialized using the normal distribution and + bais will be set to the specified value. + 2. If normal_init_embed_weights is set then weights of embedding + layer will be initialized using the normal distribution. + 3. If normal_init_proj_weights is set then weights of + in_project_weight for MultiHeadAttention initialized using + the normal distribution (to be validated). + """ + + def normal_(data): + # with FSDP, module params will be on CUDA, so we cast them back to CPU + # so that the RNG is consistent with and without FSDP + data.copy_( + data.cpu().normal_(mean=0.0, std=0.02).to(data.device) + ) + + if isinstance(module, nn.Linear): + normal_(module.weight.data) + if module.bias is not None: + module.bias.data.zero_() + if isinstance(module, nn.Embedding): + normal_(module.weight.data) + if module.padding_idx is not None: + module.weight.data[module.padding_idx].zero_() + if isinstance(module, MultiheadAttention): + normal_(module.q_proj.weight.data) + normal_(module.k_proj.weight.data) + normal_(module.v_proj.weight.data) + + +class TransformerSentenceEncoder(nn.Module): + """ + Implementation for a Bi-directional Transformer based Sentence Encoder used + in BERT/XLM style pre-trained models. + + This first computes the token embedding using the token embedding matrix, + position embeddings (if specified) and segment embeddings + (if specified). After applying the specified number of + TransformerEncoderLayers, it outputs all the internal states of the + encoder as well as the final representation associated with the first + token (usually CLS token). + + Input: + - tokens: B x T matrix representing sentences + - segment_labels: B x T matrix representing segment label for tokens + + Output: + - a tuple of the following: + - a list of internal model states used to compute the + predictions where each tensor has shape T x B x C + - sentence representation associated with first input token + in format B x C. + """ + + def __init__( + self, + padding_idx: int, + vocab_size: int, + num_encoder_layers: int = 6, + embedding_dim: int = 768, + ffn_embedding_dim: int = 3072, + num_attention_heads: int = 8, + dropout: float = 0.1, + attention_dropout: float = 0.1, + activation_dropout: float = 0.1, + layerdrop: float = 0.0, + max_seq_len: int = 256, + num_segments: int = 2, + use_position_embeddings: bool = True, + offset_positions_by_padding: bool = True, + encoder_normalize_before: bool = False, + apply_bert_init: bool = False, + activation_fn: str = "relu", + learned_pos_embedding: bool = True, + embed_scale: float = None, + freeze_embeddings: bool = False, + n_trans_layers_to_freeze: int = 0, + export: bool = False, + traceable: bool = False, + q_noise: float = 0.0, + qn_block_size: int = 8, + ) -> None: + + super().__init__() + self.padding_idx = padding_idx + self.vocab_size = vocab_size + self.dropout_module = FairseqDropout( + dropout, module_name=self.__class__.__name__ + ) + self.layerdrop = layerdrop + self.max_seq_len = max_seq_len + self.embedding_dim = embedding_dim + self.num_segments = num_segments + self.use_position_embeddings = use_position_embeddings + self.apply_bert_init = apply_bert_init + self.learned_pos_embedding = learned_pos_embedding + self.traceable = traceable + + self.embed_tokens = self.build_embedding( + self.vocab_size, self.embedding_dim, self.padding_idx + ) + self.embed_scale = embed_scale + + if q_noise > 0: + self.quant_noise = apply_quant_noise_( + nn.Linear(self.embedding_dim, self.embedding_dim, bias=False), + q_noise, + qn_block_size, + ) + else: + self.quant_noise = None + + self.segment_embeddings = ( + nn.Embedding(self.num_segments, self.embedding_dim, padding_idx=None) + if self.num_segments > 0 + else None + ) + + self.embed_positions = ( + PositionalEmbedding( + self.max_seq_len, + self.embedding_dim, + padding_idx=(self.padding_idx if offset_positions_by_padding else None), + learned=self.learned_pos_embedding, + ) + if self.use_position_embeddings + else None + ) + + if encoder_normalize_before: + self.emb_layer_norm = LayerNorm(self.embedding_dim, export=export) + else: + self.emb_layer_norm = None + + if self.layerdrop > 0.0: + self.layers = LayerDropModuleList(p=self.layerdrop) + else: + self.layers = nn.ModuleList([]) + self.layers.extend( + [ + self.build_transformer_sentence_encoder_layer( + embedding_dim=self.embedding_dim, + ffn_embedding_dim=ffn_embedding_dim, + num_attention_heads=num_attention_heads, + dropout=self.dropout_module.p, + attention_dropout=attention_dropout, + activation_dropout=activation_dropout, + activation_fn=activation_fn, + export=export, + q_noise=q_noise, + qn_block_size=qn_block_size, + ) + for _ in range(num_encoder_layers) + ] + ) + + # Apply initialization of model params after building the model + if self.apply_bert_init: + self.apply(init_bert_params) + + def freeze_module_params(m): + if m is not None: + for p in m.parameters(): + p.requires_grad = False + + if freeze_embeddings: + freeze_module_params(self.embed_tokens) + freeze_module_params(self.segment_embeddings) + freeze_module_params(self.embed_positions) + freeze_module_params(self.emb_layer_norm) + + for layer in range(n_trans_layers_to_freeze): + freeze_module_params(self.layers[layer]) + + def build_embedding(self, vocab_size, embedding_dim, padding_idx): + return nn.Embedding(vocab_size, embedding_dim, padding_idx) + + def build_transformer_sentence_encoder_layer( + self, + embedding_dim, + ffn_embedding_dim, + num_attention_heads, + dropout, + attention_dropout, + activation_dropout, + activation_fn, + export, + q_noise, + qn_block_size, + ): + return TransformerSentenceEncoderLayer( + embedding_dim=embedding_dim, + ffn_embedding_dim=ffn_embedding_dim, + num_attention_heads=num_attention_heads, + dropout=dropout, + attention_dropout=attention_dropout, + activation_dropout=activation_dropout, + activation_fn=activation_fn, + export=export, + q_noise=q_noise, + qn_block_size=qn_block_size, + ) + + def forward( + self, + tokens: torch.Tensor, + segment_labels: torch.Tensor = None, + last_state_only: bool = False, + positions: Optional[torch.Tensor] = None, + token_embeddings: Optional[torch.Tensor] = None, + attn_mask: Optional[torch.Tensor] = None, + ) -> Tuple[torch.Tensor, torch.Tensor]: + is_tpu = tokens.device.type == "xla" + + # compute padding mask. This is needed for multi-head attention + padding_mask = tokens.eq(self.padding_idx) + if not self.traceable and not is_tpu and not padding_mask.any(): + padding_mask = None + + if token_embeddings is not None: + x = token_embeddings + else: + x = self.embed_tokens(tokens) + + if self.embed_scale is not None: + x = x * self.embed_scale + + if self.embed_positions is not None: + x = x + self.embed_positions(tokens, positions=positions) + + if self.segment_embeddings is not None and segment_labels is not None: + x = x + self.segment_embeddings(segment_labels) + + if self.quant_noise is not None: + x = self.quant_noise(x) + + if self.emb_layer_norm is not None: + x = self.emb_layer_norm(x) + + x = self.dropout_module(x) + + # account for padding while computing the representation + if padding_mask is not None: + x = x * (1 - padding_mask.unsqueeze(-1).type_as(x)) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + + inner_states = [] + if not last_state_only: + inner_states.append(x) + + for layer in self.layers: + x, _ = layer(x, self_attn_padding_mask=padding_mask, self_attn_mask=attn_mask) + if not last_state_only: + inner_states.append(x) + + sentence_rep = x[0, :, :] + + if last_state_only: + inner_states = [x] + + if self.traceable: + return torch.stack(inner_states), sentence_rep + else: + return inner_states, sentence_rep diff --git a/fairseq/modules/transformer_sentence_encoder_layer.py b/fairseq/modules/transformer_sentence_encoder_layer.py new file mode 100644 index 0000000000000000000000000000000000000000..f869c4b2f8fb15f96a292e39bd293df7898a4fce --- /dev/null +++ b/fairseq/modules/transformer_sentence_encoder_layer.py @@ -0,0 +1,139 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from typing import Callable, Optional + +import torch +import torch.nn as nn +from fairseq import utils +from fairseq.modules import LayerNorm, MultiheadAttention +from fairseq.modules.fairseq_dropout import FairseqDropout +from fairseq.modules.quant_noise import quant_noise + + +class TransformerSentenceEncoderLayer(nn.Module): + """ + Implements a Transformer Encoder Layer used in BERT/XLM style pre-trained + models. + """ + + def __init__( + self, + embedding_dim: int = 768, + ffn_embedding_dim: int = 3072, + num_attention_heads: int = 8, + dropout: float = 0.1, + attention_dropout: float = 0.1, + activation_dropout: float = 0.1, + activation_fn: str = "relu", + export: bool = False, + q_noise: float = 0.0, + qn_block_size: int = 8, + init_fn: Callable = None, + ) -> None: + super().__init__() + + if init_fn is not None: + init_fn() + + # Initialize parameters + self.embedding_dim = embedding_dim + self.num_attention_heads = num_attention_heads + self.attention_dropout = attention_dropout + self.q_noise = q_noise + self.qn_block_size = qn_block_size + + self.dropout_module = FairseqDropout( + dropout, module_name=self.__class__.__name__ + ) + self.activation_dropout_module = FairseqDropout( + activation_dropout, module_name=self.__class__.__name__ + ) + + # Initialize blocks + self.activation_fn = utils.get_activation_fn(activation_fn) + self.self_attn = self.build_self_attention( + self.embedding_dim, + num_attention_heads, + dropout=attention_dropout, + self_attention=True, + q_noise=q_noise, + qn_block_size=qn_block_size, + ) + + # layer norm associated with the self attention layer + self.self_attn_layer_norm = LayerNorm(self.embedding_dim, export=export) + + self.fc1 = self.build_fc1( + self.embedding_dim, + ffn_embedding_dim, + q_noise=q_noise, + qn_block_size=qn_block_size, + ) + self.fc2 = self.build_fc2( + ffn_embedding_dim, + self.embedding_dim, + q_noise=q_noise, + qn_block_size=qn_block_size, + ) + + # layer norm associated with the position wise feed-forward NN + self.final_layer_norm = LayerNorm(self.embedding_dim, export=export) + + def build_fc1(self, input_dim, output_dim, q_noise, qn_block_size): + return quant_noise(nn.Linear(input_dim, output_dim), q_noise, qn_block_size) + + def build_fc2(self, input_dim, output_dim, q_noise, qn_block_size): + return quant_noise(nn.Linear(input_dim, output_dim), q_noise, qn_block_size) + + def build_self_attention( + self, + embed_dim, + num_attention_heads, + dropout, + self_attention, + q_noise, + qn_block_size, + ): + return MultiheadAttention( + embed_dim, + num_attention_heads, + dropout=dropout, + self_attention=True, + q_noise=q_noise, + qn_block_size=qn_block_size, + ) + + def forward( + self, + x: torch.Tensor, + self_attn_mask: Optional[torch.Tensor] = None, + self_attn_padding_mask: Optional[torch.Tensor] = None, + ): + """ + LayerNorm is applied either before or after the self-attention/ffn + modules similar to the original Transformer implementation. + """ + residual = x + x, attn = self.self_attn( + query=x, + key=x, + value=x, + key_padding_mask=self_attn_padding_mask, + need_weights=False, + attn_mask=self_attn_mask, + ) + x = self.dropout_module(x) + x = residual + x + x = self.self_attn_layer_norm(x) + + residual = x + x = self.activation_fn(self.fc1(x)) + x = self.activation_dropout_module(x) + x = self.fc2(x) + x = self.dropout_module(x) + x = residual + x + x = self.final_layer_norm(x) + return x, attn diff --git a/fairseq/modules/transpose_last.py b/fairseq/modules/transpose_last.py new file mode 100644 index 0000000000000000000000000000000000000000..e578b3ec5097bfac5c976b207ea46bec1d9bd4f5 --- /dev/null +++ b/fairseq/modules/transpose_last.py @@ -0,0 +1,20 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +transpose last 2 dimensions of the input +""" + +import torch.nn as nn + + +class TransposeLast(nn.Module): + def __init__(self, deconstruct_idx=None): + super().__init__() + self.deconstruct_idx = deconstruct_idx + + def forward(self, x): + if self.deconstruct_idx is not None: + x = x[self.deconstruct_idx] + return x.transpose(-2, -1) diff --git a/fairseq/modules/unfold.py b/fairseq/modules/unfold.py new file mode 100644 index 0000000000000000000000000000000000000000..138272f1ef4f673b29e36aed4531106f7ce95968 --- /dev/null +++ b/fairseq/modules/unfold.py @@ -0,0 +1,19 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch.nn.functional as F + + +def unfold1d(x, kernel_size, padding_l, pad_value=0): + """unfold T x B x C to T x B x C x K""" + if kernel_size > 1: + T, B, C = x.size() + x = F.pad( + x, (0, 0, 0, 0, padding_l, kernel_size - 1 - padding_l), value=pad_value + ) + x = x.as_strided((T, B, C, kernel_size), (B * C, C, 1, B * C)) + else: + x = x.unsqueeze(3) + return x diff --git a/fairseq/modules/vggblock.py b/fairseq/modules/vggblock.py new file mode 100644 index 0000000000000000000000000000000000000000..ee5ee19a34816c7350c21fba7c4907fec8ca7a61 --- /dev/null +++ b/fairseq/modules/vggblock.py @@ -0,0 +1,116 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from __future__ import absolute_import, division, print_function, unicode_literals + +from collections.abc import Iterable +from itertools import repeat + +import torch +import torch.nn as nn + + +def _pair(v): + if isinstance(v, Iterable): + assert len(v) == 2, "len(v) != 2" + return v + return tuple(repeat(v, 2)) + + +def infer_conv_output_dim(conv_op, input_dim, sample_inchannel): + sample_seq_len = 200 + sample_bsz = 10 + x = torch.randn(sample_bsz, sample_inchannel, sample_seq_len, input_dim) + # N x C x H x W + # N: sample_bsz, C: sample_inchannel, H: sample_seq_len, W: input_dim + x = conv_op(x) + # N x C x H x W + x = x.transpose(1, 2) + # N x H x C x W + bsz, seq = x.size()[:2] + per_channel_dim = x.size()[3] + # bsz: N, seq: H, CxW the rest + return x.contiguous().view(bsz, seq, -1).size(-1), per_channel_dim + + +class VGGBlock(torch.nn.Module): + """ + VGG motibated cnn module https://arxiv.org/pdf/1409.1556.pdf + + Args: + in_channels: (int) number of input channels (typically 1) + out_channels: (int) number of output channels + conv_kernel_size: convolution channels + pooling_kernel_size: the size of the pooling window to take a max over + num_conv_layers: (int) number of convolution layers + input_dim: (int) input dimension + conv_stride: the stride of the convolving kernel. + Can be a single number or a tuple (sH, sW) Default: 1 + padding: implicit paddings on both sides of the input. + Can be a single number or a tuple (padH, padW). Default: None + layer_norm: (bool) if layer norm is going to be applied. Default: False + + Shape: + Input: BxCxTxfeat, i.e. (batch_size, input_size, timesteps, features) + Output: BxCxTxfeat, i.e. (batch_size, input_size, timesteps, features) + """ + + def __init__( + self, + in_channels, + out_channels, + conv_kernel_size, + pooling_kernel_size, + num_conv_layers, + input_dim, + conv_stride=1, + padding=None, + layer_norm=False, + ): + assert ( + input_dim is not None + ), "Need input_dim for LayerNorm and infer_conv_output_dim" + super(VGGBlock, self).__init__() + self.in_channels = in_channels + self.out_channels = out_channels + self.conv_kernel_size = _pair(conv_kernel_size) + self.pooling_kernel_size = _pair(pooling_kernel_size) + self.num_conv_layers = num_conv_layers + self.padding = ( + tuple(e // 2 for e in self.conv_kernel_size) + if padding is None + else _pair(padding) + ) + self.conv_stride = _pair(conv_stride) + + self.layers = nn.ModuleList() + for layer in range(num_conv_layers): + conv_op = nn.Conv2d( + in_channels if layer == 0 else out_channels, + out_channels, + self.conv_kernel_size, + stride=self.conv_stride, + padding=self.padding, + ) + self.layers.append(conv_op) + if layer_norm: + conv_output_dim, per_channel_dim = infer_conv_output_dim( + conv_op, input_dim, in_channels if layer == 0 else out_channels + ) + self.layers.append(nn.LayerNorm(per_channel_dim)) + input_dim = per_channel_dim + self.layers.append(nn.ReLU()) + + if self.pooling_kernel_size is not None: + pool_op = nn.MaxPool2d(kernel_size=self.pooling_kernel_size, ceil_mode=True) + self.layers.append(pool_op) + self.total_output_dim, self.output_dim = infer_conv_output_dim( + pool_op, input_dim, out_channels + ) + + def forward(self, x): + for i, _ in enumerate(self.layers): + x = self.layers[i](x) + return x diff --git a/fairseq/nan_detector.py b/fairseq/nan_detector.py new file mode 100644 index 0000000000000000000000000000000000000000..faa8031d4666c9ba9837919fe1c884dacf47ac3a --- /dev/null +++ b/fairseq/nan_detector.py @@ -0,0 +1,108 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging + +import torch + + +logger = logging.getLogger(__name__) + + +class NanDetector: + """ + Detects the first NaN or Inf in forward and/or backward pass and logs, together with the module name + """ + + def __init__(self, model, forward=True, backward=True): + self.bhooks = [] + self.fhooks = [] + self.forward = forward + self.backward = backward + self.named_parameters = list(model.named_parameters()) + self.reset() + + for name, mod in model.named_modules(): + mod.__module_name = name + self.add_hooks(mod) + + def __enter__(self): + return self + + def __exit__(self, exc_type, exc_value, exc_traceback): + # Dump out all model gnorms to enable better debugging + norm = {} + gradients = {} + for name, param in self.named_parameters: + if param.grad is not None: + grad_norm = torch.norm(param.grad.data, p=2, dtype=torch.float32) + norm[name] = grad_norm.item() + if torch.isnan(grad_norm).any() or torch.isinf(grad_norm).any(): + gradients[name] = param.grad.data + if len(gradients) > 0: + logger.info("Detected nan/inf grad norm, dumping norms...") + logger.info(f"norms: {norm}") + logger.info(f"gradients: {gradients}") + + self.close() + + def add_hooks(self, module): + if self.forward: + self.fhooks.append(module.register_forward_hook(self.fhook_fn)) + if self.backward: + self.bhooks.append(module.register_backward_hook(self.bhook_fn)) + + def reset(self): + self.has_printed_f = False + self.has_printed_b = False + + def _detect(self, tensor, name, backward): + err = None + if ( + torch.is_floating_point(tensor) + # single value tensors (like the loss) will not provide much info + and tensor.numel() >= 2 + ): + with torch.no_grad(): + if torch.isnan(tensor).any(): + err = "NaN" + elif torch.isinf(tensor).any(): + err = "Inf" + if err is not None: + err = f"{err} detected in output of {name}, shape: {tensor.shape}, {'backward' if backward else 'forward'}" + return err + + def _apply(self, module, inp, x, backward): + if torch.is_tensor(x): + if isinstance(inp, tuple) and len(inp) > 0: + inp = inp[0] + err = self._detect(x, module.__module_name, backward) + if err is not None: + if torch.is_tensor(inp) and not backward: + err += ( + f" input max: {inp.max().item()}, input min: {inp.min().item()}" + ) + + has_printed_attr = "has_printed_b" if backward else "has_printed_f" + logger.warning(err) + setattr(self, has_printed_attr, True) + elif isinstance(x, dict): + for v in x.values(): + self._apply(module, inp, v, backward) + elif isinstance(x, list) or isinstance(x, tuple): + for v in x: + self._apply(module, inp, v, backward) + + def fhook_fn(self, module, inp, output): + if not self.has_printed_f: + self._apply(module, inp, output, backward=False) + + def bhook_fn(self, module, inp, output): + if not self.has_printed_b: + self._apply(module, inp, output, backward=True) + + def close(self): + for hook in self.fhooks + self.bhooks: + hook.remove() diff --git a/fairseq/ngram_repeat_block.py b/fairseq/ngram_repeat_block.py new file mode 100644 index 0000000000000000000000000000000000000000..854125149448a2d37ad2773cd1e6d614e73e0e79 --- /dev/null +++ b/fairseq/ngram_repeat_block.py @@ -0,0 +1,150 @@ +# Originally from Microsoft Corporation. +# Licensed under the MIT License. + +""" Wrapper for ngram_repeat_block cuda extension """ +import torch +from torch import nn + +import math +from typing import Dict, List, Optional +import warnings + +try: + from fairseq import ngram_repeat_block_cuda + + EXTENSION_BUILT = True +except ImportError: + EXTENSION_BUILT = False + + +def is_cuda_extension_usable() -> bool: + """Check whether ngram_repeat_block_cuda is built properly""" + if not EXTENSION_BUILT or not torch.cuda.is_available(): + return False + bsz = 2 + tokens = torch.tensor([[4, 4, 3, 2], [1, 2, 3, 4]], dtype=torch.long, device="cuda") + lprobs = torch.rand((8, 12), device="cuda") + try: + outputs = ngram_repeat_block_cuda.forward(tokens, lprobs, bsz, 3, 4, 3) + outputs = outputs + 4 # This line breaks if the extension is built incorrectly. + return True + except RuntimeError: + warnings.warn( + "NGramRepeatBlock extension must be rebuilt." + 'Run TORCH_CUDA_ARCH_LIST="6.0;6.1;7.0" python setup.py build_ext --inplace' + ) + return False + + +class NGramRepeatBlock(nn.Module): + """ Wrapper class for calling ngram_repeat_block cuda extension """ + + def __init__(self, no_repeat_ngram_size: int, use_extension: bool = True): + super().__init__() + self.use_extension = is_cuda_extension_usable() if use_extension else False + self.no_repeat_ngram_size = no_repeat_ngram_size + + def reset_parameters(self): + pass + + @torch.jit.unused + def call_cuda_extension( + self, + tokens, + lprobs, + bsz: int, + beam_size: int, + step: int, + ): + return ngram_repeat_block_cuda.forward( + tokens, lprobs, bsz, step, beam_size, self.no_repeat_ngram_size + ) + + def forward( + self, + tokens, + lprobs, + bsz: int, + beam_size: int, + step: int, + ): + """ + Args: + tokens(Tensor): Input tokens(Bsz*beam, seq_len) + lprobs(Tensor): likelihood probability, + Expected to be updated in place.(Bsz*beam, vocab_size) + bsz(int): batch size + step(int): current step + beam_size(int): beam size + no_repeat_ngram_size(int): Ngram size + """ + msg = f"expected {bsz *beam_size} got" + assert tokens.size(0) == bsz * beam_size, f"{msg} {tokens.size(0)}" + assert lprobs.size(0) == bsz * beam_size, f"{msg} {lprobs.size(0)}" + if self.use_extension: + return self.call_cuda_extension(tokens, lprobs, bsz, beam_size, step) + + else: + return self._no_repeat_ngram( + tokens, + lprobs, + bsz, + beam_size, + step, + ) + + def _no_repeat_ngram(self, tokens, lprobs, bsz: int, beam_size: int, step: int): + """For each hypothesis generate a list of previous ngrams and set associated lprobs to -inf""" + gen_ngrams: List[Dict[str, List[int]]] = [ + torch.jit.annotate(Dict[str, List[int]], {}) + for bbsz_idx in range(bsz * beam_size) + ] + cpu_tokens = tokens.cpu() + for bbsz_idx in range(bsz * beam_size): + gen_tokens: List[int] = cpu_tokens[bbsz_idx].tolist() + for ngram in self.transpose_list( + [gen_tokens[i:] for i in range(self.no_repeat_ngram_size)] + ): + key = ",".join([str(x) for x in ngram[:-1]]) + gen_ngrams[bbsz_idx][key] = gen_ngrams[bbsz_idx].get( + key, torch.jit.annotate(List[int], []) + ) + [ngram[-1]] + if step + 2 - self.no_repeat_ngram_size >= 0: + # no banned tokens if we haven't generated no_repeat_ngram_size tokens yet + banned_tokens = [ + self.calculate_banned_tokens( + tokens, step, gen_ngrams, self.no_repeat_ngram_size, bbsz_idx + ) + for bbsz_idx in range(bsz * beam_size) + ] + else: + banned_tokens = [ + torch.jit.annotate(List[int], []) for bbsz_idx in range(bsz * beam_size) + ] + for bbsz_idx in range(bsz * beam_size): + lprobs[bbsz_idx][ + torch.tensor(banned_tokens[bbsz_idx], dtype=torch.int64) + ] = torch.tensor(-math.inf).to(lprobs) + return lprobs + + @staticmethod + def calculate_banned_tokens( + tokens, + step: int, + gen_ngrams: List[Dict[str, List[int]]], + no_repeat_ngram_size: int, + bbsz_idx: int, + ): + tokens_list: List[int] = tokens[ + bbsz_idx, step + 2 - no_repeat_ngram_size : step + 1 + ].tolist() + # before decoding the next token, prevent decoding of ngrams that have already appeared + ngram_index = ",".join([str(x) for x in tokens_list]) + return gen_ngrams[bbsz_idx].get(ngram_index, torch.jit.annotate(List[int], [])) + + @staticmethod + def transpose_list(l: List[List[int]]): + # GeneratorExp aren't supported in TS so ignoring the lint + min_len = min([len(x) for x in l]) # noqa + l2 = [[row[i] for row in l] for i in range(min_len)] + return l2 diff --git a/fairseq/optim/__init__.py b/fairseq/optim/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..be783be896396ff659c0bd173a7acebb8a2d165d --- /dev/null +++ b/fairseq/optim/__init__.py @@ -0,0 +1,48 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +"""isort:skip_file""" + +import importlib +import os + +from fairseq import registry +from fairseq.optim.bmuf import FairseqBMUF # noqa +from fairseq.optim.fairseq_optimizer import ( # noqa + FairseqOptimizer, + LegacyFairseqOptimizer, +) +from fairseq.optim.amp_optimizer import AMPOptimizer +from fairseq.optim.fp16_optimizer import FP16Optimizer, MemoryEfficientFP16Optimizer +from fairseq.optim.shard import shard_ +from omegaconf import DictConfig + +__all__ = [ + "AMPOptimizer", + "FairseqOptimizer", + "FP16Optimizer", + "MemoryEfficientFP16Optimizer", + "shard_", +] + +( + _build_optimizer, + register_optimizer, + OPTIMIZER_REGISTRY, + OPTIMIZER_DATACLASS_REGISTRY, +) = registry.setup_registry("--optimizer", base_class=FairseqOptimizer, required=True) + + +def build_optimizer(cfg: DictConfig, params, *extra_args, **extra_kwargs): + if all(isinstance(p, dict) for p in params): + params = [t for p in params for t in p.values()] + params = list(filter(lambda p: p.requires_grad, params)) + return _build_optimizer(cfg, params, *extra_args, **extra_kwargs) + + +# automatically import any Python files in the optim/ directory +for file in sorted(os.listdir(os.path.dirname(__file__))): + if file.endswith(".py") and not file.startswith("_"): + file_name = file[: file.find(".py")] + importlib.import_module("fairseq.optim." + file_name) diff --git a/fairseq/optim/adadelta.py b/fairseq/optim/adadelta.py new file mode 100644 index 0000000000000000000000000000000000000000..f1a21549770f0904a6a40a42ff7eb52811f1bfbe --- /dev/null +++ b/fairseq/optim/adadelta.py @@ -0,0 +1,47 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch.optim + +from . import LegacyFairseqOptimizer, register_optimizer + + +@register_optimizer("adadelta") +class Adadelta(LegacyFairseqOptimizer): + def __init__(self, args, params): + super().__init__(args) + self._optimizer = torch.optim.Adadelta(params, **self.optimizer_config) + + @staticmethod + def add_args(parser): + """Add optimizer-specific arguments to the parser.""" + # fmt: off + parser.add_argument('--adadelta-rho', type=float, default=0.9, metavar='RHO', + help='coefficient used for computing a running average of squared gradients') + parser.add_argument('--adadelta-eps', type=float, default=1e-6, metavar='EPS', + help='term added to the denominator to improve numerical stability') + parser.add_argument('--weight-decay', '--wd', default=0.0, type=float, metavar='WD', + help='weight decay') + parser.add_argument('--anneal-eps', action='store_true', help='flag to anneal eps') + # fmt: on + + @property + def optimizer_config(self): + """ + Return a kwarg dictionary that will be used to override optimizer + args stored in checkpoints. This allows us to load a checkpoint and + resume training using a different set of optimizer args, e.g., with a + different learning rate. + """ + return { + "lr": self.args.lr[0], + "rho": self.args.adadelta_rho, + "eps": self.args.adadelta_eps, + "weight_decay": self.args.weight_decay, + } + + @property + def supports_flat_params(self): + return True diff --git a/fairseq/optim/adafactor.py b/fairseq/optim/adafactor.py new file mode 100644 index 0000000000000000000000000000000000000000..c969b9fbc0d229a25f2046ec67c53c57a433814b --- /dev/null +++ b/fairseq/optim/adafactor.py @@ -0,0 +1,268 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +import torch +import torch.optim + +from . import LegacyFairseqOptimizer, register_optimizer + + +@register_optimizer("adafactor") +class FairseqAdafactor(LegacyFairseqOptimizer): + def __init__(self, args, params): + super().__init__(args) + self._optimizer = Adafactor(params, **self.optimizer_config) + + @staticmethod + def add_args(parser): + """Add optimizer-specific arguments to the parser.""" + # fmt: off + parser.add_argument('--adafactor-eps', default='(1e-30, 1e-3)', metavar="E", + help='epsilons for Adafactor optimizer') + parser.add_argument('--clip-threshold', type=float, default=1.0, metavar="C", + help='threshold for clipping update root mean square') + parser.add_argument('--decay-rate', type=float, default=-0.8, metavar="D", + help='decay rate of the second moment estimator') + parser.add_argument('--beta1', type=float, default=None, metavar="B", + help='beta for first moment estimator. Optional') + parser.add_argument('--weight-decay', '--wd', default=0.0, type=float, metavar='WD', + help='weight decay') + parser.add_argument('--scale-parameter', action='store_true', + help='scale learning rate by root mean square of parameter') + parser.add_argument('--relative-step', action='store_true', + help='set learning rate to inverse square root of timestep,' + 'otherwise use external learning rate') + parser.add_argument('--warmup-init', action='store_true', + help='use relative step for warm-up learning rate schedule') + # fmt: on + + @property + def optimizer_config(self): + """ + Return a kwarg dictionary that will be used to override optimizer + args stored in checkpoints. This allows us to load a checkpoint and + resume training using a different set of optimizer args, e.g., with a + different learning rate. + Note : Convergence issues empirically observed with fp16 on. + Might require search for appropriate configuration. + """ + return { + "lr": self.args.lr[0], + "eps": eval(self.args.adafactor_eps), + "clip_threshold": self.args.clip_threshold, + "decay_rate": self.args.decay_rate, + "beta1": self.args.beta1, + "weight_decay": self.args.weight_decay, + "scale_parameter": self.args.scale_parameter, # defaults to False + "relative_step": self.args.relative_step, # defaults to False + "warmup_init": self.args.warmup_init, + } + + +class Adafactor(torch.optim.Optimizer): + """Implements Adafactor algorithm. + + This implementation is based on: + `Adafactor: Adaptive Learning Rates with Sublinear Memory Cost` + (see https://arxiv.org/abs/1804.04235) + + Note that this optimizer internally adjusts the learning rate + depending on the *scale_parameter*, *relative_step* and + *warmup_init* options. To use a manual (external) learning rate + schedule you should set `scale_parameter=False` and + `relative_step=False`. + + Args: + params (iterable): iterable of parameters to optimize or dicts defining + parameter groups + lr (float, optional): external learning rate (default: None) + eps (tuple[float, float]): regularization constans for square gradient + and parameter scale respectively (default: (1e-30, 1e-3)) + clip_threshold (float): threshold of root mean square of + final gradient update (default: 1.0) + decay_rate (float): coefficient used to compute running averages of square + gradient (default: -0.8) + beta1 (float): coefficient used for computing running averages of gradient + (default: None) + weight_decay (float, optional): weight decay (L2 penalty) (default: 0) + scale_parameter (bool): if True, learning rate is scaled by root mean square of + parameter (default: True) + relative_step (bool): if True, time-dependent learning rate is computed + instead of external learning rate (default: True) + warmup_init (bool): time-dependent learning rate computation depends on + whether warm-up initialization is being used (default: False) + """ + + def __init__( + self, + params, + lr=None, + eps=(1e-30, 1e-3), + clip_threshold=1.0, + decay_rate=-0.8, + beta1=None, + weight_decay=0.0, + scale_parameter=True, + relative_step=True, + warmup_init=False, + ): + if lr is not None and relative_step: + raise ValueError("Cannot combine manual lr and relative_step options") + if warmup_init and not relative_step: + raise ValueError("warmup_init requires relative_step=True") + + defaults = dict( + lr=lr, + eps=eps, + clip_threshold=clip_threshold, + decay_rate=decay_rate, + beta1=beta1, + weight_decay=weight_decay, + scale_parameter=scale_parameter, + relative_step=relative_step, + warmup_init=warmup_init, + ) + super(Adafactor, self).__init__(params, defaults) + + @property + def supports_memory_efficient_fp16(self): + return True + + @property + def supports_flat_params(self): + return False + + def _get_lr(self, param_group, param_state): + rel_step_sz = param_group["lr"] + if param_group["relative_step"]: + min_step = ( + 1e-6 * param_state["step"] if param_group["warmup_init"] else 1e-2 + ) + rel_step_sz = min(min_step, 1.0 / math.sqrt(param_state["step"])) + param_scale = 1.0 + if param_group["scale_parameter"]: + param_scale = max(param_group["eps"][1], param_state["RMS"]) + return param_scale * rel_step_sz + + def _get_options(self, param_group, param_shape): + factored = len(param_shape) >= 2 + use_first_moment = param_group["beta1"] is not None + return factored, use_first_moment + + def _rms(self, tensor): + return tensor.norm(2) / (tensor.numel() ** 0.5) + + def _approx_sq_grad(self, exp_avg_sq_row, exp_avg_sq_col): + r_factor = ( + (exp_avg_sq_row / exp_avg_sq_row.mean(dim=-1, keepdim=True)) + .rsqrt_() + .unsqueeze(-1) + ) + c_factor = exp_avg_sq_col.unsqueeze(-2).rsqrt() + return torch.mul(r_factor, c_factor) + + def step(self, closure=None): + """Performs a single optimization step. + + Args: + closure (callable, optional): A closure that reevaluates the model + and returns the loss. + """ + loss = None + if closure is not None: + loss = closure() + + for group in self.param_groups: + for p in group["params"]: + if p.grad is None: + continue + grad = p.grad.data + if grad.dtype in {torch.float16, torch.bfloat16}: + grad = grad.float() + if grad.is_sparse: + raise RuntimeError("Adafactor does not support sparse gradients.") + + state = self.state[p] + grad_shape = grad.shape + + factored, use_first_moment = self._get_options(group, grad_shape) + # State Initialization + if len(state) == 0: + state["step"] = 0 + + if use_first_moment: + # Exponential moving average of gradient values + state["exp_avg"] = torch.zeros_like(grad) + if factored: + state["exp_avg_sq_row"] = torch.zeros(grad_shape[:-1]).to(grad) + state["exp_avg_sq_col"] = torch.zeros( + grad_shape[:-2] + grad_shape[-1:] + ).to(grad) + else: + state["exp_avg_sq"] = torch.zeros_like(grad) + + state["RMS"] = 0 + else: + if use_first_moment: + state["exp_avg"] = state["exp_avg"].to(grad) + if factored: + state["exp_avg_sq_row"] = state["exp_avg_sq_row"].to(grad) + state["exp_avg_sq_col"] = state["exp_avg_sq_col"].to(grad) + else: + state["exp_avg_sq"] = state["exp_avg_sq"].to(grad) + + p_data_fp32 = p.data + if p.data.dtype in {torch.float16, torch.bfloat16}: + p_data_fp32 = p_data_fp32.float() + + state["step"] += 1 + state["RMS"] = self._rms(p_data_fp32) + group["lr"] = self._get_lr(group, state) + + beta2t = 1.0 - math.pow(state["step"], group["decay_rate"]) + update = (grad ** 2) + group["eps"][0] + if factored: + exp_avg_sq_row = state["exp_avg_sq_row"] + exp_avg_sq_col = state["exp_avg_sq_col"] + + exp_avg_sq_row.mul_(beta2t).add_( + update.mean(dim=-1), alpha=1.0 - beta2t + ) + exp_avg_sq_col.mul_(beta2t).add_( + update.mean(dim=-2), alpha=1.0 - beta2t + ) + + # Approximation of exponential moving average of square of gradient + update = self._approx_sq_grad(exp_avg_sq_row, exp_avg_sq_col) + update.mul_(grad) + else: + exp_avg_sq = state["exp_avg_sq"] + + exp_avg_sq.mul_(beta2t).add_(update, alpha=1.0 - beta2t) + update = exp_avg_sq.rsqrt().mul_(grad) + + update.div_( + (self._rms(update) / group["clip_threshold"]).clamp_(min=1.0) + ) + update.mul_(group["lr"]) + + if use_first_moment: + exp_avg = state["exp_avg"] + exp_avg.mul_(group["beta1"]).add_(update, alpha=1 - group["beta1"]) + update = exp_avg + + if group["weight_decay"] != 0: + p_data_fp32.add_( + p_data_fp32, alpha=-group["weight_decay"] * group["lr"] + ) + + p_data_fp32.add_(-update) + + if p.data.dtype in {torch.float16, torch.bfloat16}: + p.data.copy_(p_data_fp32) + + return loss diff --git a/fairseq/optim/adagrad.py b/fairseq/optim/adagrad.py new file mode 100644 index 0000000000000000000000000000000000000000..4f539541c1c91d8c822f7ce624fa6eabf744f60e --- /dev/null +++ b/fairseq/optim/adagrad.py @@ -0,0 +1,40 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch.optim + +from . import LegacyFairseqOptimizer, register_optimizer + + +@register_optimizer("adagrad") +class Adagrad(LegacyFairseqOptimizer): + def __init__(self, args, params): + super().__init__(args) + self._optimizer = torch.optim.Adagrad(params, **self.optimizer_config) + + @staticmethod + def add_args(parser): + """Add optimizer-specific arguments to the parser.""" + # fmt: off + parser.add_argument('--weight-decay', '--wd', default=0.0, type=float, metavar='WD', + help='weight decay') + # fmt: on + + @property + def optimizer_config(self): + """ + Return a kwarg dictionary that will be used to override optimizer + args stored in checkpoints. This allows us to load a checkpoint and + resume training using a different set of optimizer args, e.g., with a + different learning rate. + """ + return { + "lr": self.args.lr[0], + "weight_decay": self.args.weight_decay, + } + + @property + def supports_flat_params(self): + return False diff --git a/fairseq/optim/adam.py b/fairseq/optim/adam.py new file mode 100644 index 0000000000000000000000000000000000000000..6a31e53a6285b75a2e0ee03ae54a9ec94df00e9d --- /dev/null +++ b/fairseq/optim/adam.py @@ -0,0 +1,228 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import math +from collections.abc import Collection +from dataclasses import dataclass, field +from typing import Any, List + +import torch +import torch.distributed as dist +import torch.optim +from fairseq.dataclass import FairseqDataclass +from fairseq.optim import FairseqOptimizer, register_optimizer +from fairseq.optim.fused_adam import get_fused_adam_class +from omegaconf import II, OmegaConf + + +logger = logging.getLogger(__name__) + + +@dataclass +class FairseqAdamConfig(FairseqDataclass): + adam_betas: Any = field( + default=(0.9, 0.999), metadata={"help": "betas for Adam optimizer"} + ) + adam_eps: float = field( + default=1e-8, metadata={"help": "epsilon for Adam optimizer"} + ) + weight_decay: float = field(default=0.0, metadata={"help": "weight decay"}) + use_old_adam: bool = field( + default=False, metadata={"help": "Use fairseq.optim.adam.Adam"} + ) + # TODO common vars below in parent + tpu: bool = II("common.tpu") + lr: List[float] = II("optimization.lr") + + +@register_optimizer("adam", dataclass=FairseqAdamConfig) +class FairseqAdam(FairseqOptimizer): + """Adam optimizer for fairseq. + + Important note: this optimizer corresponds to the "AdamW" variant of + Adam in its weight decay behavior. As such, it is most closely + analogous to torch.optim.AdamW from PyTorch. + """ + + def __init__(self, cfg: FairseqAdamConfig, params): + super().__init__(cfg) + fused_adam_cls = get_fused_adam_class() + use_fused_adam = ( + not getattr(cfg, "use_old_adam", False) + and fused_adam_cls is not None + and torch.cuda.is_available() + ) + if getattr(cfg, "tpu", False): + # on TPUs we use the Adam defined here, since it + # automatically casts gradients to FP32 + self._optimizer = Adam(params, **self.optimizer_config) + elif use_fused_adam: + logger.info("using FusedAdam") + self._optimizer = fused_adam_cls(params, **self.optimizer_config) + else: + self._optimizer = Adam(params, **self.optimizer_config) + + @property + def optimizer_config(self): + """ + Return a kwarg dictionary that will be used to override optimizer + args stored in checkpoints. This allows us to load a checkpoint and + resume training using a different set of optimizer args, e.g., with a + different learning rate. + """ + return { + "lr": self.cfg.lr[0] + if isinstance(self.cfg.lr, Collection) + else self.cfg.lr, + "betas": eval(self.cfg.adam_betas) + if isinstance(self.cfg.adam_betas, str) + else OmegaConf.to_container(self.cfg.adam_betas), + "eps": self.cfg.adam_eps, + "weight_decay": self.cfg.weight_decay, + } + + def average_params(self): + """Reduce Params is only used during BMUF distributed training.""" + state_dict = self.optimizer.state_dict() + total_gpus = float(dist.get_world_size()) + + for _, value in state_dict["state"].items(): + value["exp_avg"] /= total_gpus + value["exp_avg_sq"] /= total_gpus + dist.all_reduce(value["exp_avg"], op=dist.ReduceOp.SUM) + dist.all_reduce(value["exp_avg_sq"], op=dist.ReduceOp.SUM) + + +class Adam(torch.optim.Optimizer): + r"""Implements Adam algorithm. + + This implementation is modified from torch.optim.Adam based on: + `Fixed Weight Decay Regularization in Adam` + (see https://arxiv.org/abs/1711.05101) + + It has been proposed in `Adam: A Method for Stochastic Optimization`_. + + Args: + params (iterable): iterable of parameters to optimize or dicts defining + parameter groups + lr (float, optional): learning rate (default: 1e-3) + betas (Tuple[float, float], optional): coefficients used for computing + running averages of gradient and its square (default: (0.9, 0.999)) + eps (float, optional): term added to the denominator to improve + numerical stability (default: 1e-8) + weight_decay (float, optional): weight decay (L2 penalty) (default: 0) + amsgrad (boolean, optional): whether to use the AMSGrad variant of this + algorithm from the paper `On the Convergence of Adam and Beyond`_ + + .. _Adam\: A Method for Stochastic Optimization: + https://arxiv.org/abs/1412.6980 + .. _On the Convergence of Adam and Beyond: + https://openreview.net/forum?id=ryQu7f-RZ + """ + + def __init__( + self, + params, + lr=1e-3, + betas=(0.9, 0.999), + eps=1e-8, + weight_decay=0, + amsgrad=False, + ): + defaults = dict( + lr=lr, betas=betas, eps=eps, weight_decay=weight_decay, amsgrad=amsgrad + ) + super(Adam, self).__init__(params, defaults) + + @property + def supports_memory_efficient_fp16(self): + return True + + @property + def supports_flat_params(self): + return True + + def step(self, closure=None): + """Performs a single optimization step. + + Args: + closure (callable, optional): A closure that reevaluates the model + and returns the loss. + """ + loss = None + if closure is not None: + loss = closure() + + for group in self.param_groups: + for p in group["params"]: + if p.grad is None: + continue + grad = p.grad.data + if grad.dtype in {torch.float16, torch.bfloat16}: + grad = grad.float() + if grad.is_sparse: + raise RuntimeError( + "Adam does not support sparse gradients, please consider SparseAdam instead" + ) + amsgrad = group.get("amsgrad", False) + + p_data_fp32 = p.data + if p.data.dtype in {torch.float16, torch.bfloat16}: + p_data_fp32 = p_data_fp32.float() + + state = self.state[p] + + # State initialization + if len(state) == 0: + state["step"] = 0 + # Exponential moving average of gradient values + state["exp_avg"] = torch.zeros_like(p_data_fp32) + # Exponential moving average of squared gradient values + state["exp_avg_sq"] = torch.zeros_like(p_data_fp32) + if amsgrad: + # Maintains max of all exp. moving avg. of sq. grad. values + state["max_exp_avg_sq"] = torch.zeros_like(p_data_fp32) + else: + state["exp_avg"] = state["exp_avg"].to(p_data_fp32) + state["exp_avg_sq"] = state["exp_avg_sq"].to(p_data_fp32) + if amsgrad: + state["max_exp_avg_sq"] = state["max_exp_avg_sq"].to( + p_data_fp32 + ) + + exp_avg, exp_avg_sq = state["exp_avg"], state["exp_avg_sq"] + if amsgrad: + max_exp_avg_sq = state["max_exp_avg_sq"] + beta1, beta2 = group["betas"] + + state["step"] += 1 + + # Decay the first and second moment running average coefficient + exp_avg.mul_(beta1).add_(grad, alpha=1 - beta1) + exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2) + if amsgrad: + # Maintains the maximum of all 2nd moment running avg. till now + torch.max(max_exp_avg_sq, exp_avg_sq, out=max_exp_avg_sq) + # Use the max. for normalizing running avg. of gradient + denom = max_exp_avg_sq.sqrt().add_(group["eps"]) + else: + denom = exp_avg_sq.sqrt().add_(group["eps"]) + + bias_correction1 = 1 - beta1 ** state["step"] + bias_correction2 = 1 - beta2 ** state["step"] + step_size = group["lr"] * math.sqrt(bias_correction2) / bias_correction1 + + if group["weight_decay"] != 0: + p_data_fp32.add_( + p_data_fp32, alpha=-group["weight_decay"] * group["lr"] + ) + + p_data_fp32.addcdiv_(exp_avg, denom, value=-step_size) + + if p.data.dtype in {torch.float16, torch.bfloat16}: + p.data.copy_(p_data_fp32) + + return loss diff --git a/fairseq/optim/adamax.py b/fairseq/optim/adamax.py new file mode 100644 index 0000000000000000000000000000000000000000..98ff8ad7ad6c12ab5efc53ca76db2f1663be7906 --- /dev/null +++ b/fairseq/optim/adamax.py @@ -0,0 +1,172 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.optim + +from . import LegacyFairseqOptimizer, register_optimizer + + +@register_optimizer("adamax") +class FairseqAdamax(LegacyFairseqOptimizer): + def __init__(self, args, params): + super().__init__(args) + self._optimizer = Adamax(params, **self.optimizer_config) + + @staticmethod + def add_args(parser): + """Add optimizer-specific arguments to the parser.""" + # fmt: off + parser.add_argument('--adamax-betas', default='(0.9, 0.999)', metavar='B', + help='betas for Adam optimizer') + parser.add_argument('--adamax-eps', type=float, default=1e-8, metavar='D', + help='epsilon for Adam optimizer') + parser.add_argument('--weight-decay', '--wd', default=0.0, type=float, metavar='WD', + help='weight decay') + parser.add_argument('--no-bias-correction', default=False, action='store_true', + help='disable bias correction') + # fmt: on + + @property + def optimizer_config(self): + """ + Return a kwarg dictionary that will be used to override optimizer + args stored in checkpoints. This allows us to load a checkpoint and + resume training using a different set of optimizer args, e.g., with a + different learning rate. + """ + return { + "lr": self.args.lr[0], + "betas": eval(self.args.adamax_betas), + "eps": self.args.adamax_eps, + "weight_decay": self.args.weight_decay, + "bias_correction": not self.args.no_bias_correction, + } + + +class Adamax(torch.optim.Optimizer): + """Implements Adamax algorithm (a variant of Adam based on infinity norm). + + It has been proposed in `Adam: A Method for Stochastic Optimization`__. + + Compared to the version in PyTorch, this version implements a fix for weight decay. + + Args: + params (iterable): iterable of parameters to optimize or dicts defining + parameter groups + lr (float, optional): learning rate (default: 2e-3) + betas (Tuple[float, float], optional): coefficients used for computing + running averages of gradient and its square + eps (float, optional): term added to the denominator to improve + numerical stability (default: 1e-8) + weight_decay (float, optional): weight decay (L2 penalty) (default: 0) + bias_correction (bool, optional): enable bias correction (default: True) + + __ https://arxiv.org/abs/1412.6980 + """ + + def __init__( + self, + params, + lr=2e-3, + betas=(0.9, 0.999), + eps=1e-8, + weight_decay=0, + bias_correction=True, + ): + if not 0.0 <= lr: + raise ValueError("Invalid learning rate: {}".format(lr)) + if not 0.0 <= eps: + raise ValueError("Invalid epsilon value: {}".format(eps)) + if not 0.0 <= betas[0] < 1.0: + raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0])) + if not 0.0 <= betas[1] < 1.0: + raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1])) + if not 0.0 <= weight_decay: + raise ValueError("Invalid weight_decay value: {}".format(weight_decay)) + + defaults = dict( + lr=lr, + betas=betas, + eps=eps, + weight_decay=weight_decay, + bias_correction=bias_correction, + ) + super(Adamax, self).__init__(params, defaults) + + @property + def supports_memory_efficient_fp16(self): + return True + + @property + def supports_flat_params(self): + return True + + def step(self, closure=None): + """Performs a single optimization step. + + Args: + closure (callable, optional): A closure that reevaluates the model + and returns the loss. + """ + loss = None + if closure is not None: + loss = closure() + + for group in self.param_groups: + for p in group["params"]: + if p.grad is None: + continue + grad = p.grad.data.float() + if grad.is_sparse: + raise RuntimeError("Adamax does not support sparse gradients") + + p_data_fp32 = p.data + if p.data.dtype in {torch.float16, torch.bfloat16}: + p_data_fp32 = p_data_fp32.float() + + state = self.state[p] + + # State initialization + if len(state) == 0: + state["step"] = 0 + state["exp_avg"] = torch.zeros_like(p_data_fp32) + state["exp_inf"] = torch.zeros_like(p_data_fp32) + else: + state["exp_avg"] = state["exp_avg"].to(p_data_fp32) + state["exp_inf"] = state["exp_inf"].to(p_data_fp32) + + exp_avg, exp_inf = state["exp_avg"], state["exp_inf"] + beta1, beta2 = group["betas"] + eps = group["eps"] + + state["step"] += 1 + + # Update biased first moment estimate. + exp_avg.mul_(beta1).add_(grad, alpha=1 - beta1) + + # Update the exponentially weighted infinity norm. + torch.max( + exp_inf.mul_(beta2), + grad.abs_(), + out=exp_inf, + ) + + step_size = group["lr"] + if group["bias_correction"]: + bias_correction = 1 - beta1 ** state["step"] + step_size /= bias_correction + + if group["weight_decay"] != 0: + p_data_fp32.add_( + p_data_fp32, alpha=-group["weight_decay"] * group["lr"] + ) + + p_data_fp32.addcdiv_(exp_avg, exp_inf.add(eps), value=-step_size) + + if p.data.dtype in {torch.float16, torch.bfloat16}: + p.data.copy_(p_data_fp32) + + return loss diff --git a/fairseq/optim/amp_optimizer.py b/fairseq/optim/amp_optimizer.py new file mode 100644 index 0000000000000000000000000000000000000000..3b7958e50ce444474c48d1f5aeff05d66c19e5b6 --- /dev/null +++ b/fairseq/optim/amp_optimizer.py @@ -0,0 +1,105 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging + +import torch +from fairseq import optim +from omegaconf import DictConfig + +logger = logging.getLogger(__name__) + + +class AMPOptimizer(optim.FairseqOptimizer): + """ + Wrap an *optimizer* to support AMP (automatic mixed precision) training. + """ + + def __init__(self, cfg: DictConfig, params, fp32_optimizer, **kwargs): + super().__init__(cfg.optimizer) + self.fp32_optimizer = fp32_optimizer + amp_kwargs = {"init_scale": cfg.common.fp16_init_scale} + if getattr(cfg.common, "amp_scale_window", None) is not None: + amp_kwargs["growth_interval"] = cfg.common.amp_init_scale + self._grad_scaler = torch.cuda.amp.GradScaler(**amp_kwargs) + self.min_loss_scale = cfg.common.min_loss_scale + + @classmethod + def build_optimizer(cls, cfg: DictConfig, params, **kwargs): + """ + Args: + cfg (omegaconf.DictConfig): fairseq args + params (iterable): iterable of parameters to optimize + """ + fp32_optimizer = optim.build_optimizer(cfg.optimizer, params) + return cls(cfg, params, fp32_optimizer, **kwargs) + + def backward(self, loss): + """Computes the sum of gradients of the given tensor w.r.t. graph leaves. + + Compared to :func:`fairseq.optim.FairseqOptimizer.backward`, this + function additionally dynamically scales the loss to avoid gradient + underflow. + """ + self._grad_scaler.scale(loss).backward() + + def step(self): + self.scaler.step(self.fp32_optimizer) + self.scaler.update() + + def clip_grad_norm(self, max_norm, aggregate_norm_fn=None): + """Clips gradient norm.""" + self.scaler.unscale_(self.optimizer) + grad_norm = self.fp32_optimizer.clip_grad_norm(max_norm, aggregate_norm_fn) + if not torch.isfinite(grad_norm).all(): + new_loss_scale = self.next_loss_scale + if new_loss_scale <= self.min_loss_scale: + raise FloatingPointError( + ( + "AMP: Minimum loss scale reached ({}). Your loss is probably exploding. " + "Try restarting training or use fp32. {}" + ).format(self.min_loss_scale, new_loss_scale) + ) + else: + logger.info("AMP: overflow detected, setting scale to " + f"to {new_loss_scale}") + return grad_norm + + @property + def scaler(self): + return self._grad_scaler + + @property + def next_loss_scale(self): + return self.scaler.get_scale() * self.scaler.get_backoff_factor() + + @property + def optimizer(self): + return self.fp32_optimizer.optimizer + + @optimizer.setter + def optimizer(self, optimizer): + self.fp32_optimizer.optimizer = optimizer + + @property + def lr_scheduler(self): + return getattr(self.fp32_optimizer, "lr_scheduler", None) + + @property + def optimizer_config(self): + return self.fp32_optimizer.optimizer_config + + def get_lr(self): + return self.fp32_optimizer.get_lr() + + def set_lr(self, lr): + self.fp32_optimizer.set_lr(lr) + + def all_reduce_grads(self, module): + self.fp32_optimizer.all_reduce_grads(module) + + @property + def supports_flat_params(self): + return self.fp32_optimizer.supports_flat_params diff --git a/fairseq/optim/bmuf.py b/fairseq/optim/bmuf.py new file mode 100644 index 0000000000000000000000000000000000000000..d6d0e04e86eb894efe59e13a78843d01ca9e651d --- /dev/null +++ b/fairseq/optim/bmuf.py @@ -0,0 +1,200 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from dataclasses import dataclass, field + +import torch +import torch.distributed as dist +from fairseq.dataclass.configs import FairseqBMUFConfig +from fairseq.dataclass.utils import gen_parser_from_dataclass +from fairseq.optim.fairseq_optimizer import FairseqOptimizer + + +class FairseqBMUF(FairseqOptimizer): + """ + Implements incremental block distributed data parallelism similar to + https://ieeexplore.ieee.org/document/7472805 + + Paper title: Scalable training of deep learning machines by incremental + block training with intra-block parallel optimization and blockwise + model-update filtering + """ + + def __init__(self, cfg: FairseqBMUFConfig, optimizer): + super().__init__(cfg) + self._optimizer = optimizer + self._num_updates = 0 + self.sync_iter = cfg.global_sync_iter + self.block_momentum = cfg.block_momentum + self.block_lr = cfg.block_lr + self._reset_local_data() + self.warmup_iteration = cfg.warmup_iterations + self.use_nbm = cfg.use_nbm + self.initial_state = self._optimizer.state_dict() + self.average_sync = self.cfg.average_sync + self.world_size = self.cfg.distributed_world_size + + @staticmethod + def add_args(parser): + """Add optimizer-specific arguments to the parser.""" + gen_parser_from_dataclass(parser, FairseqBMUFConfig()) + + @property + def optimizer(self): + return self._optimizer.optimizer + + @property + def optimizer_config(self): + return self._optimizer.optimizer_config + + def get_lr(self): + return self._optimizer.get_lr() + + def set_lr(self, lr): + self._optimizer.set_lr(lr) + + def state_dict(self): + return self._optimizer.state_dict() + + def load_state_dict(self, state_dict, optimizer_overrides=None): + self._optimizer.load_state_dict(state_dict, optimizer_overrides) + self.initial_state = self._optimizer.state_dict() + + def multiply_grads(self, c): + """Multiplies grads by a constant *c*.""" + self._optimizer.multiply_grads(c) + + def clip_grad_norm(self, max_norm, aggregate_norm_fn=None): + """Clips gradient norm.""" + return self._optimizer.clip_grad_norm(max_norm, aggregate_norm_fn) + + def average_params(self): + self._optimizer.average_params() + + def _block_sync(self): + if self.world_size <= 1: + return + # Update the global model using local models from all GPUs + # (Step-1) Calculate grad between previously synced model and + # currrent local model + if self.block_momentum != 0: + self._calc_grad() + + # (Step-2) Average gradient from all GPUs + self._avg_grad_from_all_gpus() + + # (Step-3) Calculate global momentum and update the global model + if self.block_momentum != 0: + self._update_global_model() + + # (Step-4) Average local optimizer params + if self.average_sync: + self.average_params() + + def _is_warmup_end(self): + # Check whether train iterations is equal to warmup iter + if self.get_num_updates() == self.warmup_iteration: + return True + return False + + def _is_bmuf_iter(self): + # Check whether train iterations is equal to bmuf sync iter + if (self.get_num_updates() > self.warmup_iteration) and ( + self.get_num_updates() % self.sync_iter == 0 + ): + return True + return False + + def _warmup_sync(self, root_rank=0): + if self.world_size <= 1: + return + # Broadcast the local model to all gpus + for param in self.params: + dist.broadcast(param.data, src=root_rank) + + # Update local optimizer state + if self.average_sync: + self._optimizer.average_params() + else: + self._optimizer.load_state_dict(self.initial_state) + + self._reset_local_data() + + def step(self, closure=None): + """Performs a single optimization step.""" + self._optimizer.step(closure) + self.set_num_updates(self.get_num_updates() + 1) + if self._is_warmup_end(): + self._warmup_sync() + elif self._is_bmuf_iter(): + self._block_sync() + + def zero_grad(self): + """Clears the gradients of all optimized parameters.""" + self._optimizer.zero_grad() + + def get_num_updates(self): + """Get the number of parameters updates.""" + return self._num_updates + + def set_num_updates(self, num_updates): + """Set the number of parameters updates.""" + self._num_updates = num_updates + + @torch.no_grad() + def _reset_local_data(self): + # (Step-0) Initialize global momentum parameters and store global copy on each gpu + self.global_params = [torch.zeros_like(p.data) for p in self.params] + self.smoothed_grads = [p.data.new_zeros(p.data.size()) for p in self.params] + self.grads = [p.data.new_zeros(p.data.size()) for p in self.params] + + # saving the global model locally for calculating gradient during bmuf sync + for param, global_param in zip(self.params, self.global_params): + global_param.copy_(param.data) + + @torch.no_grad() + def _calc_grad(self): + # global_params is basically the global copy from the previously finished + # synchronisation. param.data is local parameter after block_sync_freq + # for the local gpu. so grad is difference between previously synced + # model and currrent local model. + for index, (param, global_param) in enumerate( + zip(self.params, self.global_params) + ): + self.grads[index] = global_param - param.data + + def _avg_grad_from_all_gpus(self): + for index, param in enumerate(self.params): + sync_para = param.data if self.block_momentum == 0 else self.grads[index] + sync_para /= float(dist.get_world_size()) + dist.all_reduce(sync_para, op=dist.ReduceOp.SUM) + + @torch.no_grad() + def _update_global_model(self): + for index, (param, global_param, smoothed_grad, grad) in enumerate( + zip( + self.params, + self.global_params, + self.smoothed_grads, + # all gpus would share the same value of smoothed_grad, since it is + # always computed on synchronized gradients. + self.grads, + ) + ): + # global_param is basically last syncrhornized parameter. though + # smoothed_grad is local, all processes will have same value of + # smoothed_grad and hence param is globally synchronized copy. + # smoothed_grad(t) = BM * smoothed_grad(t-1) + BM_lr * grad(t) + smoothed_grad = self.block_momentum * smoothed_grad + self.block_lr * grad + param.data.copy_(global_param - smoothed_grad) + + # A Nesterov momentum here is to do a partial weight update before + # calculating the gradient + if self.use_nbm: + param.data.copy_(param.data - self.block_momentum * smoothed_grad) + + # backup for the next synchronization. + self.smoothed_grads[index] = smoothed_grad + global_param.copy_(param.data) diff --git a/fairseq/optim/composite.py b/fairseq/optim/composite.py new file mode 100644 index 0000000000000000000000000000000000000000..a5366d62434a4400ba9cc524f4286f99f733d121 --- /dev/null +++ b/fairseq/optim/composite.py @@ -0,0 +1,188 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +from collections import defaultdict +from dataclasses import dataclass, field +from typing import Dict, Any, List, Optional + +import torch.optim +from fairseq.dataclass import FairseqDataclass +from fairseq.optim import FairseqOptimizer, register_optimizer, _build_optimizer +from fairseq.optim.lr_scheduler import FairseqLRScheduler, build_lr_scheduler +from omegaconf import II, open_dict + + +logger = logging.getLogger(__name__) + + +@dataclass +class OptimizerAndSchedulerConfig(FairseqDataclass): + optimizer: Any = None + lr_scheduler: Optional[Any] = None + lr: List = II("optimization.lr") + lr_float: Optional[float] = None # this makes it easier to sweep on learning rate with auto sweepers + + +@dataclass +class CompositeOptimizerConfig(FairseqDataclass): + groups: Dict[str, Any] = field( + default_factory=lambda: {}, + metadata={ + "help": "optimizer name -> optimizer OptimizerAndSchedulerConfig. " + "Configures a different optimizer and (optionally) lr scheduler for each parameter group" + }, + ) + + +@register_optimizer("composite", dataclass=CompositeOptimizerConfig) +class FairseqCompositeOptimizer(FairseqOptimizer): + + optimizers: Dict[str, FairseqOptimizer] = {} + lr_schedulers: Dict[str, FairseqLRScheduler] = {} + lr_scheduler: FairseqLRScheduler = None + _optimizer: torch.optim.Optimizer + + def __init__(self, cfg: CompositeOptimizerConfig, params): + super().__init__(cfg) + + assert ( + len(params) > 1 + ), "Composite optimizer only works when there are multiple parameter groups (try fp16_no_flatten_grads: true)" + + groupped_params = defaultdict(list) + for p in params: + group = getattr(p, "param_group", "default") + groupped_params[group].append(p) + + assert groupped_params.keys() == cfg.groups.keys(), ( + f"Parameter groups {groupped_params.keys()} and optimizer groups {cfg.groups.keys()} are not the same! " + "Try setting 'param_group' on your parameters in the model." + ) + + for group, group_params in groupped_params.items(): + group_cfg = cfg.groups[group] + with open_dict(group_cfg): + if group_cfg.lr_float is not None: + group_cfg.optimizer.lr = [group_cfg.lr_float] + group_cfg.lr_scheduler.lr = [group_cfg.lr_float] + else: + group_cfg.optimizer.lr = group_cfg.lr + group_cfg.lr_scheduler.lr = group_cfg.lr + self.optimizers[group] = _build_optimizer(group_cfg.optimizer, group_params) + if group_cfg.lr_scheduler is not None: + self.lr_schedulers[group] = build_lr_scheduler( + group_cfg.lr_scheduler, self.optimizers[group] + ) + + if len(self.lr_schedulers) > 0: + assert len(self.lr_schedulers) == len(self.optimizers), ( + f"Please provide an lr scheduler for each optimizer to use pass_through scheduler. " + f"Optimizers: {self.optimizers}; Lr scheds: {self.lr_schedulers}" + ) + self.lr_scheduler = CompositeLRScheduler(self.lr_schedulers) + + self._optimizer = CompositeOptimizer(self.optimizers) + + @property + def supports_groups(self): + return True + + @property + def param_groups(self): + for opt in self.optimizers.values(): + for group in opt.param_groups: + yield group + + def get_lr(self): + """Return the current learning rate.""" + k = ( + "default" + if "default" in self.optimizers + else next(iter(self.optimizers.keys())) + ) + return self.optimizers[k].param_groups[0]["lr"] + + def state_dict(self): + """Return the LR scheduler state dict.""" + return {k: s.state_dict() for k, s in self.optimizers.items()} + + def load_state_dict(self, state_dict, optimizer_overrides=None): + """Load an LR scheduler state dict.""" + for k, state in state_dict.items(): + if k not in self.optimizers: + # skip extra keys like "loss_scale" added by fp16 optimizer + continue + + overrides = ( + optimizer_overrides[k] + if isinstance(optimizer_overrides, dict) and k in optimizer_overrides + else None + ) + self.optimizers[k].load_state_dict(state, optimizer_overrides=overrides) + + +class CompositeOptimizer(torch.optim.Optimizer): + def __init__(self, optimizers: Dict[str, FairseqOptimizer]): + self.optimizers = optimizers + + @property + def supports_memory_efficient_fp16(self): + return all(o.supports_memory_efficient_fp16 for o in self.optimizers.values()) + + @property + def supports_flat_params(self): + return all(o.supports_flat_params for o in self.optimizers.values()) + + def step(self, closure=None, groups=None): + """Performs a single optimization step. + + Args: + closure (callable, optional): A closure that reevaluates the model + and returns the loss. + """ + loss = None + if closure is not None: + loss = closure() + + for k, opt in self.optimizers.items(): + if groups is None or k in groups: + opt.step() + + return loss + + def zero_grad(self): + for opt in self.optimizers.values(): + opt.zero_grad() + + +class CompositeLRScheduler(FairseqLRScheduler): + def __init__(self, lr_schedulers): + super().__init__(None, None) + + self.lr_schedulers = lr_schedulers + + def state_dict(self): + """Return the LR scheduler state dict.""" + return {k: s.state_dict() for k, s in self.lr_schedulers.items()} + + def load_state_dict(self, state_dict): + """Load an LR scheduler state dict.""" + for k, state in state_dict.items(): + self.lr_schedulers[k].load_state_dict(state) + + def step_begin_epoch(self, epoch): + """Update the learning rate at the beginning of the given epoch.""" + for s in self.lr_schedulers.values(): + s.step_begin_epoch(epoch) + + def step(self, epoch, val_loss=None): + """Update the learning rate at the end of the given epoch.""" + for s in self.lr_schedulers.values(): + s.step(epoch) + + def step_update(self, num_updates): + """Update the learning rate after each update.""" + return {k: s.step_update(num_updates) for k, s in self.lr_schedulers.items()} diff --git a/fairseq/optim/cpu_adam.py b/fairseq/optim/cpu_adam.py new file mode 100644 index 0000000000000000000000000000000000000000..e36bccf123020d1d90acafdf6a641be1fd926b8b --- /dev/null +++ b/fairseq/optim/cpu_adam.py @@ -0,0 +1,200 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import importlib +from collections.abc import Collection +from dataclasses import dataclass, field +from typing import List + +import torch +from fairseq.dataclass import FairseqDataclass +from fairseq.optim import FairseqOptimizer, register_optimizer +from omegaconf import II, DictConfig + + +try: + import deepspeed + has_deepspeed = True +except ImportError as e: + has_deepspeed = False + + +def _get_cpu_adam(): + try: + from deepspeed.ops.op_builder import CPUAdamBuilder + return CPUAdamBuilder().load() + except ImportError: + # fbcode + from deepspeed.ops.adam import DeepSpeedCPUAdam as ds_opt_adam + return ds_opt_adam + +@dataclass +class FairseqCPUAdamConfig(FairseqDataclass): + adam_betas: str = field( + default="(0.9, 0.999)", metadata={"help": "betas for Adam optimizer"} + ) + adam_eps: float = field( + default=1e-8, metadata={"help": "epsilon for Adam optimizer"} + ) + weight_decay: float = field(default=0.0, metadata={"help": "weight decay"}) + fp16_adam_stats: bool = field( + default=False, metadata={"help": "use FP16 stats (with automatic scaling)"} + ) + # TODO common vars below in parent + lr: List[float] = II("optimization.lr") + + +@register_optimizer("cpu_adam", dataclass=FairseqCPUAdamConfig) +class FairseqCPUAdam(FairseqOptimizer): + """Adam optimizer for fairseq, optimized for CPU tensors. + + Important note: this optimizer corresponds to the "AdamW" variant of + Adam in its weight decay behavior. As such, it is most closely + analogous to torch.optim.AdamW from PyTorch. + """ + + def __init__(self, cfg: DictConfig, params): + super().__init__(cfg) + self._optimizer = CPUAdam(params, **self.optimizer_config) + + @property + def optimizer_config(self): + """ + Return a kwarg dictionary that will be used to override optimizer + args stored in checkpoints. This allows us to load a checkpoint and + resume training using a different set of optimizer args, e.g., with a + different learning rate. + """ + return { + "lr": self.cfg.lr[0] + if isinstance(self.cfg.lr, Collection) + else self.cfg.lr, + "betas": eval(self.cfg.adam_betas), + "eps": self.cfg.adam_eps, + "weight_decay": self.cfg.weight_decay, + "use_fp16_stats": self.cfg.fp16_adam_stats, + } + + +class CPUAdam(torch.optim.Optimizer): + + optimizer_id = 0 + + def __init__( + self, + params, + lr=1e-3, + bias_correction=True, + betas=(0.9, 0.999), + eps=1e-8, + weight_decay=0, + use_fp16_stats=False, + ): + defaults = { + "lr": lr, + "bias_correction": bias_correction, + "betas": betas, + "eps": eps, + "weight_decay": weight_decay, + } + super().__init__(params, defaults) + + self.use_fp16_stats = use_fp16_stats + self.FLOAT16_MAX = 65504.0 + + if not has_deepspeed: + raise ImportError("Please install DeepSpeed: pip install deepspeed") + + self.opt_id = CPUAdam.optimizer_id + CPUAdam.optimizer_id = CPUAdam.optimizer_id + 1 + + self.ds_opt_adam = _get_cpu_adam() + adamw_mode = True + self.ds_opt_adam.create_adam( + self.opt_id, lr, betas[0], betas[1], eps, weight_decay, adamw_mode + ) + + @property + def supports_flat_params(self): + return True + + @torch.no_grad() + def step(self, closure=None): + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + + for group_id, group in enumerate(self.param_groups): + for param_id, p in enumerate(group["params"]): + if p.grad is None: + continue + + state = self.state[p] + if len(state) == 0: + state["step"] = 0 + dtype = torch.float16 if self.use_fp16_stats else p.data.dtype + # gradient momentums + state["exp_avg"] = torch.zeros_like( + p.data, dtype=dtype, device="cpu" + ) + # gradient variances + state["exp_avg_sq"] = torch.zeros_like( + p.data, dtype=dtype, device="cpu" + ) + if self.use_fp16_stats: + assert torch.is_floating_point(p.data) + state["exp_avg_scale"] = 1.0 + state["exp_avg_sq_scale"] = 1.0 + + exp_avg, exp_avg_sq = state["exp_avg"], state["exp_avg_sq"] + + p_data_bak = p.data # backup of the original data pointer + + p.data = p.data.to(dtype=torch.float32, device="cpu") + p.grad.data = p.grad.data.to(dtype=torch.float32, device="cpu") + + if self.use_fp16_stats: + exp_avg = exp_avg.float() * state["exp_avg_scale"] + exp_avg_sq = exp_avg_sq.float() * state["exp_avg_sq_scale"] + + state["step"] += 1 + beta1, beta2 = group["betas"] + + self.ds_opt_adam.adam_update( + self.opt_id, + state["step"], + group["lr"], + beta1, + beta2, + group["eps"], + group["weight_decay"], + group["bias_correction"], + p.data, + p.grad.data, + exp_avg, + exp_avg_sq, + ) + + if p_data_bak.data_ptr() != p.data.data_ptr(): + p_data_bak.copy_(p.data) + p.data = p_data_bak + + if self.use_fp16_stats: + + def inf_norm(t): + return torch.norm(t, float("inf")) + + # from github.com/openai/jukebox/blob/master/jukebox/utils/fp16.py + state["exp_avg_scale"], state["exp_avg_sq_scale"] = ( + 1e-8 + inf_norm(exp_avg) / self.FLOAT16_MAX, + 1e-8 + inf_norm(exp_avg_sq) / self.FLOAT16_MAX, + ) + state["exp_avg"], state["exp_avg_sq"] = ( + (exp_avg / state["exp_avg_scale"]).half(), + (exp_avg_sq / state["exp_avg_sq_scale"]).half(), + ) + + return loss diff --git a/fairseq/optim/dynamic_loss_scaler.py b/fairseq/optim/dynamic_loss_scaler.py new file mode 100644 index 0000000000000000000000000000000000000000..43f9be37b9067c520cd794b9a941c57adae25e97 --- /dev/null +++ b/fairseq/optim/dynamic_loss_scaler.py @@ -0,0 +1,70 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +class DynamicLossScaler(object): + def __init__( + self, + init_scale=2.0 ** 15, + scale_factor=2.0, + scale_window=2000, + tolerance=0.0, + threshold=None, + min_loss_scale=1e-4, + ): + self.loss_scale = init_scale + self.scale_factor = scale_factor + self.scale_window = scale_window + self.tolerance = tolerance + self.threshold = threshold + self._iter = 0 + self._last_overflow_iter = -1 + self._last_rescale_iter = -1 + self._overflows_since_rescale = 0 + self.min_loss_scale = min_loss_scale + + def scale(self, outputs): + return self.loss_scale * outputs + + def update(self): + if (self._iter - self._last_overflow_iter) % self.scale_window == 0: + self.loss_scale *= self.scale_factor + self._last_rescale_iter = self._iter + self._iter += 1 + + def _decrease_loss_scale(self): + self.loss_scale /= self.scale_factor + if self.threshold is not None: + self.loss_scale = max(self.loss_scale, self.threshold) + + def check_overflow(self, grad_norm): + # detect inf and nan + if grad_norm == float("inf") or grad_norm != grad_norm: + # overflow has occured + prev_scale = self.loss_scale + iter_since_rescale = self._iter - self._last_rescale_iter + + self._last_overflow_iter = self._iter + self._overflows_since_rescale += 1 + pct_overflow = self._overflows_since_rescale / float(iter_since_rescale) + if pct_overflow >= self.tolerance: + self._decrease_loss_scale() + self._last_rescale_iter = self._iter + self._overflows_since_rescale = 0 + + if self.loss_scale <= self.min_loss_scale: + # Use FloatingPointError as an uncommon error that parent + # functions can safely catch to stop training. + self.loss_scale = prev_scale + raise FloatingPointError( + ( + "Minimum loss scale reached ({}). Your loss is probably exploding. " + "Try lowering the learning rate, using gradient clipping or " + "increasing the batch size." + ).format(self.min_loss_scale) + ) + + self._iter += 1 + raise OverflowError("setting loss scale to: " + str(self.loss_scale)) diff --git a/fairseq/optim/fairseq_optimizer.py b/fairseq/optim/fairseq_optimizer.py new file mode 100644 index 0000000000000000000000000000000000000000..7e5411753a2ba94f3a7a68316131530b8b17d22a --- /dev/null +++ b/fairseq/optim/fairseq_optimizer.py @@ -0,0 +1,179 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +from fairseq import utils +from fairseq.dataclass.utils import gen_parser_from_dataclass + + +class FairseqOptimizer(object): + def __init__(self, cfg): + super().__init__() + self.cfg = cfg + + @classmethod + def add_args(cls, parser): + """Add optimizer-specific arguments to the parser.""" + dc = getattr(cls, "__dataclass", None) + if dc is not None: + gen_parser_from_dataclass(parser, dc()) + + @property + def optimizer(self): + """Return a torch.optim.optimizer.Optimizer instance.""" + if not hasattr(self, "_optimizer"): + raise NotImplementedError + if not isinstance(self._optimizer, torch.optim.Optimizer): + raise ValueError("_optimizer must be an instance of torch.optim.Optimizer") + return self._optimizer + + @optimizer.setter + def optimizer(self, optimizer): + """Reset optimizer instance.""" + if not hasattr(self, "_optimizer"): + raise NotImplementedError + if not isinstance(self._optimizer, torch.optim.Optimizer): + raise ValueError("_optimizer must be an instance of torch.optim.Optimizer") + self._optimizer = optimizer + + @property + def optimizer_config(self): + """ + Return a kwarg dictionary that will be used to override optimizer + args stored in checkpoints. This allows us to load a checkpoint and + resume training using a different set of optimizer args, e.g., with a + different learning rate. + """ + raise NotImplementedError + + @property + def params(self): + """Return an iterable of the parameters held by the optimizer.""" + for param_group in self.param_groups: + for p in param_group["params"]: + yield p + + @property + def param_groups(self): + return self.optimizer.param_groups + + def __getstate__(self): + return self._optimizer.__getstate__() + + def get_lr(self): + """Return the current learning rate.""" + return self.param_groups[0]["lr"] + + def set_lr(self, lr): + """Set the learning rate.""" + for param_group in self.param_groups: + param_group["lr"] = lr + + def state_dict(self): + """Return the optimizer's state dict.""" + return self.optimizer.state_dict() + + def load_state_dict(self, state_dict, optimizer_overrides=None): + """Load an optimizer state dict. + + In general we should prefer the configuration of the existing optimizer + instance (e.g., learning rate) over that found in the state_dict. This + allows us to resume training from a checkpoint using a new set of + optimizer args. + """ + self.optimizer.load_state_dict(state_dict) + + if optimizer_overrides is not None and len(optimizer_overrides) > 0: + # override learning rate, momentum, etc. with latest values + for group in self.param_groups: + group.update(optimizer_overrides) + + def backward(self, loss): + """Computes the sum of gradients of the given tensor w.r.t. graph leaves.""" + loss.backward() + + def all_reduce_grads(self, module): + """Manually all-reduce gradients (if required).""" + if hasattr(module, "all_reduce_grads"): + module.all_reduce_grads() + + def multiply_grads(self, c): + """Multiplies grads by a constant *c*.""" + for p in self.params: + if p.grad is not None: + if torch.is_tensor(c): + c = c.to(p.grad.device) + p.grad.data.mul_(c) + + def clip_grad_norm(self, max_norm, aggregate_norm_fn=None): + """Clips gradient norm.""" + return utils.clip_grad_norm_(self.params, max_norm, aggregate_norm_fn) + + def step(self, closure=None, scale=1.0, groups=None): + """Performs a single optimization step.""" + if self.supports_step_with_scale: + if self.supports_groups: + self.optimizer.step(closure, scale=scale, groups=groups) + else: + self.optimizer.step(closure, scale=scale) + else: + if scale != 1.0: + self.multiply_grads(1.0 / scale) + if self.supports_groups: + self.optimizer.step(closure, groups=groups) + else: + self.optimizer.step(closure) + + def zero_grad(self): + """Clears the gradients of all optimized parameters.""" + for p in self.params: + p.grad = None + self.optimizer.zero_grad() + + @property + def supports_memory_efficient_fp16(self): + if hasattr(self.optimizer, "supports_memory_efficient_fp16"): + return self.optimizer.supports_memory_efficient_fp16 + return False + + @property + def supports_step_with_scale(self): + if hasattr(self.optimizer, "supports_step_with_scale"): + return self.optimizer.supports_step_with_scale + return False + + @property + def supports_groups(self): + if hasattr(self.optimizer, "supports_groups"): + return self.optimizer.supports_groups + return False + + @property + def supports_flat_params(self): + """ + Whether the optimizer supports collapsing of the model + parameters/gradients into a single contiguous Tensor. + """ + if hasattr(self.optimizer, "supports_flat_params"): + return self.optimizer.supports_flat_params + return False + + def average_params(self): + pass + + def broadcast_global_state_dict(self, state_dict): + """ + Broadcasts a global state dict to all ranks. + Useful for optimizers that shard state between ranks. + """ + if hasattr(self.optimizer, "broadcast_global_state_dict"): + return self.optimizer.broadcast_global_state_dict(state_dict) + else: + return state_dict + + +class LegacyFairseqOptimizer(FairseqOptimizer): + def __init__(self, args): + self.args = args diff --git a/fairseq/optim/fp16_optimizer.py b/fairseq/optim/fp16_optimizer.py new file mode 100644 index 0000000000000000000000000000000000000000..370a910102a43f34d0101717e4bd71f729f6e238 --- /dev/null +++ b/fairseq/optim/fp16_optimizer.py @@ -0,0 +1,546 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from collections import defaultdict +from itertools import chain + +import torch +from fairseq import optim +from omegaconf import DictConfig + +from .dynamic_loss_scaler import DynamicLossScaler + + +class _FP16OptimizerMixin(object): + def __init__(self, *args, **kwargs): + # forward __init__ call to the next class in mro(method resolution order) + super().__init__(*args, **kwargs) + self._multiply_factor = 1.0 + + @property + def has_flat_params(self): + return torch.is_tensor(self.fp32_params) or ( + isinstance(self.fp32_params, dict) + and all(torch.is_tensor(t) for t in self.fp32_params.values()) + ) + + @classmethod + def build_fp32_params(cls, args, params, flatten=True): + # create FP32 copy of parameters and grads + if flatten: + is_pipeline_parallel = getattr( + args, "pipeline_model_parallel", False + ) and getattr(args, "distributed_no_spawn", False) + total_param_size = sum(p.data.numel() for p in params) + devices = [torch.cuda.current_device()] + if is_pipeline_parallel: + devices = list(set(args.pipeline_devices)) + fp32_params = {} + for device in devices: + if is_pipeline_parallel: + device_param_size = sum( + p.data.numel() for p in params if p.device.index == device + ) + device_params = [p for p in params if p.device.index == device] + else: + device_param_size = total_param_size + device_params = params + fp32_params[device] = ( + device_params[0].new(0).float().new(device_param_size) + ) + offset = 0 + for p in device_params: + numel = p.data.numel() + fp32_params[device][offset : offset + numel].copy_(p.data.view(-1)) + offset += numel + fp32_params[device] = torch.nn.Parameter(fp32_params[device]) + fp32_params[device].grad = fp32_params[device].data.new( + device_param_size + ) + return fp32_params + else: + fp32_params = [] + for p in params: + p32 = torch.nn.Parameter(p.data.float()) + if hasattr(p, 'expert'): + p32.expert = True + p32.grad = torch.zeros_like(p32.data) + if hasattr(p, "param_group"): + p32.param_group = p.param_group + fp32_params.append(p32) + return fp32_params + + def state_dict(self): + """Return the optimizer's state dict.""" + state_dict = self.fp32_optimizer.state_dict() + if self.scaler is not None: + state_dict["loss_scale"] = self.scaler.loss_scale + return state_dict + + def load_state_dict(self, state_dict, optimizer_overrides=None): + """Load an optimizer state dict. + + In general we should prefer the configuration of the existing optimizer + instance (e.g., learning rate) over that found in the state_dict. This + allows us to resume training from a checkpoint using a new set of + optimizer args. + """ + if "loss_scale" in state_dict and self.scaler is not None: + self.scaler.loss_scale = state_dict["loss_scale"] + self.fp32_optimizer.load_state_dict(state_dict, optimizer_overrides) + + def backward(self, loss): + """Computes the sum of gradients of the given tensor w.r.t. graph leaves. + + Compared to :func:`fairseq.optim.FairseqOptimizer.backward`, this + function additionally dynamically scales the loss to avoid gradient + underflow. + """ + if self.scaler is not None: + loss = self.scaler.scale(loss) + loss.backward() + self._needs_sync = True + + def _sync_fp16_grads_to_fp32(self): + if self._needs_sync: + # copy FP16 grads to FP32 + if self.has_flat_params: + devices = list(self.fp32_params.keys()) + device_params_dict = defaultdict(list) + for p in self.fp16_params: + if p.requires_grad: + device_params_dict[p.device.index].append(p) + for device in devices: + device_params = device_params_dict[device] + offset = 0 + for p in device_params: + grad_data = ( + p.grad.data + if p.grad is not None + else p.data.new_zeros(p.data.shape) + ) + numel = grad_data.numel() + self.fp32_params[device].grad.data[ + offset : offset + numel + ].copy_(grad_data.view(-1)) + offset += numel + else: + for p, p32 in zip(self.fp16_params, self.fp32_params): + if not p.requires_grad: + continue + if p.grad is not None: + if p32.grad is None: + p32.grad = p.grad.data.float() + else: + p32.grad.data.copy_(p.grad.data) + else: + p32.grad = torch.zeros_like(p.data, dtype=torch.float) + + self._needs_sync = False + + def _sync_fp32_params_to_fp16(self): + # copy FP32 params back into FP16 model + if self.has_flat_params: + devices = list(self.fp32_params.keys()) + device_params_dict = defaultdict(list) + for p in self.fp16_params: + device_params_dict[p.device.index].append(p) + for device in devices: + device_params = device_params_dict[device] + offset = 0 + for p in device_params: + numel = p.data.numel() + p.data.copy_( + self.fp32_params[device] + .data[offset : offset + numel] + .view_as(p.data) + ) + offset += numel + else: + for p, p32 in zip(self.fp16_params, self.fp32_params): + if not p.requires_grad: + continue + p.data.copy_(p32.data) + + def _unscale_grads(self): + self._sync_fp16_grads_to_fp32() + if ( + # Skip the multiplication if it's a no-op (i.e., if _multiply_factor + # is 1.0). At the same time, we want to avoid the device-to-host + # transfer by comparing it to 1.0. Since _multiply_factor starts as + # a Python float, we roughly assume that if it's a tensor then it's + # probably not =1.0 anymore and we do the multiplication. Otherwise + # we can safely check the value without a D2H transfer. + torch.is_tensor(self._multiply_factor) + or self._multiply_factor != 1.0 + ): + self.fp32_optimizer.multiply_grads(self._multiply_factor) + self._multiply_factor = 1.0 + + def multiply_grads(self, c): + """Multiplies grads by a constant ``c``.""" + self._multiply_factor *= c + + def clip_grad_norm(self, max_norm, aggregate_norm_fn=None): + """Clips gradient norm and updates dynamic loss scaler.""" + self._sync_fp16_grads_to_fp32() + + grad_norm = self._multiply_factor * self.fp32_optimizer.clip_grad_norm( + 0, aggregate_norm_fn + ) + + if self.scaler is not None: + if grad_norm > max_norm > 0.0: + self._multiply_factor *= max_norm / grad_norm + + self.scaler.check_overflow(grad_norm) + elif max_norm > 0.0: + clip_coef = (max_norm / (grad_norm + 1e-6)).clamp_(max=1) + self._multiply_factor *= clip_coef + + return grad_norm + + def step(self, closure=None, groups=None): + """Performs a single optimization step.""" + self._sync_fp16_grads_to_fp32() + + if getattr(self, "supports_step_with_scale", False): + self.fp32_optimizer.step(closure, scale=(1.0 / self._multiply_factor), groups=groups) + else: + self._unscale_grads() + self.fp32_optimizer.step(closure, groups=groups) + + if self.scaler is not None: + self.scaler.update() + + self._sync_fp32_params_to_fp16() + + def zero_grad(self): + """Clears the gradients of all optimized parameters.""" + for p in self.fp16_params: + p.grad = None + if self.has_flat_params: + if torch.is_tensor(self.fp32_params): + self.fp32_params.grad.zero_() + elif isinstance(self.fp32_params, dict): + for fp32_params in self.fp32_params.values(): + fp32_params.grad.zero_() + else: + raise RuntimeError("self.fp32_params must be a tensor or dict") + else: + for p32 in self.fp32_params: + if p32.grad is not None: + p32.grad.zero_() + self._needs_sync = False + + if self.scaler is not None: + self._multiply_factor = 1.0 / float(self.scaler.loss_scale) + + +class FP16Optimizer(_FP16OptimizerMixin, optim.FairseqOptimizer): + """ + Wrap an *optimizer* to support FP16 (mixed precision) training. + """ + + def __init__(self, cfg: DictConfig, params, fp32_optimizer, fp32_params, **kwargs): + super().__init__(cfg.optimizer) + self.fp16_params = params + self.fp32_optimizer = fp32_optimizer + self.fp32_params = fp32_params + + if getattr(cfg.common, "fp16_scale_window", None) is None: + if len(cfg.optimization.update_freq) > 1: + raise ValueError( + "--fp16-scale-window must be given explicitly when using a " + "custom --update-freq schedule" + ) + data_parallel_size = int( + cfg.distributed_training.distributed_world_size + / cfg.common.model_parallel_size + ) + scale_window = int( + 2 ** 14 / data_parallel_size / cfg.optimization.update_freq[0] + ) + else: + scale_window = cfg.common.fp16_scale_window + + if not getattr(cfg.common, "bf16", False): + self.scaler = DynamicLossScaler( + init_scale=cfg.common.fp16_init_scale, + scale_window=scale_window, + tolerance=cfg.common.fp16_scale_tolerance, + threshold=cfg.common.threshold_loss_scale, + min_loss_scale=cfg.common.min_loss_scale, + ) + else: + # disable loss scaling for bfloat16 + self.scaler = None + + @classmethod + def build_optimizer(cls, cfg: DictConfig, params, **kwargs): + """ + Args: + cfg (omegaconf.DictConfig): fairseq args + params (iterable): iterable of parameters to optimize + """ + flatten = not getattr(cfg.common, "fp16_no_flatten_grads", False) + if getattr(cfg.common, "bf16", False): + flatten = False # mixed precision is faster on TPUs without flat grads + fp32_params = cls.build_fp32_params(cfg.optimizer, params, flatten=flatten) + if flatten: + fp32_optimizer = optim.build_optimizer(cfg.optimizer, [fp32_params]) + else: + fp32_optimizer = optim.build_optimizer(cfg.optimizer, fp32_params) + if flatten and not fp32_optimizer.supports_flat_params: + raise RuntimeError( + f"chosen optimizer {fp32_optimizer.__class__.__name__} does not support flat params, please set --fp16-no-flatten-grads" + ) + return cls(cfg, params, fp32_optimizer, fp32_params, **kwargs) + + @property + def optimizer(self): + return self.fp32_optimizer.optimizer + + @optimizer.setter + def optimizer(self, optimizer): + self.fp32_optimizer.optimizer = optimizer + + @property + def lr_scheduler(self): + return getattr(self.fp32_optimizer, "lr_scheduler", None) + + @property + def optimizer_config(self): + return self.fp32_optimizer.optimizer_config + + def get_lr(self): + return self.fp32_optimizer.get_lr() + + def set_lr(self, lr): + self.fp32_optimizer.set_lr(lr) + + def all_reduce_grads(self, module): + self.fp32_optimizer.all_reduce_grads(module) + + @property + def supports_flat_params(self): + return self.fp32_optimizer.supports_flat_params + + +class _MemoryEfficientFP16OptimizerMixin(object): + def __init__(self, *args, **kwargs): + # forward __init__ call to the next class in MRO (method resolution order) + super().__init__(*args, **kwargs) + self._multiply_factor = 1.0 + + @property + def has_flat_params(self): + return False + + def state_dict(self): + """Return the optimizer's state dict.""" + state_dict = self.wrapped_optimizer.state_dict() + if self.scaler is not None: + state_dict["loss_scale"] = self.scaler.loss_scale + return state_dict + + def load_state_dict(self, state_dict, optimizer_overrides=None): + """Load an optimizer state dict. + + In general we should prefer the configuration of the existing optimizer + instance (e.g., learning rate) over that found in the state_dict. This + allows us to resume training from a checkpoint using a new set of + optimizer args. + """ + if "loss_scale" in state_dict and self.scaler is not None: + self.scaler.loss_scale = state_dict["loss_scale"] + + self.wrapped_optimizer.load_state_dict(state_dict, optimizer_overrides) + + # Hack: PyTorch automatically casts the optimizer state to match the + # type of the current parameters. But with --memory-efficient-fp16 the + # params are FP16 while the optimizer state is FP32 and we don't want + # to cast. A workaround is to manually copy back the original state + # after the optimizer has been loaded. + if not getattr(self.optimizer, "disable_mem_eff_fp16_loading_hack", False): + groups = self.optimizer.param_groups + saved_groups = state_dict["param_groups"] + id_map = { + old_id: p + for old_id, p in zip( + chain(*(g["params"] for g in saved_groups)), + chain(*(g["params"] for g in groups)), + ) + } + for k, v in state_dict["state"].items(): + if k in id_map: + param = id_map[k] + self.optimizer.state[param] = v + + def backward(self, loss): + """Computes the sum of gradients of the given tensor w.r.t. graph leaves. + + Compared to :func:`fairseq.optim.FairseqOptimizer.backward`, this + function additionally dynamically scales the loss to avoid gradient + underflow. + """ + if self.scaler is not None: + loss = self.scaler.scale(loss) + loss.backward() + + def _unscale_grads(self): + if ( + # Skip the multiplication if it's a no-op (i.e., if _multiply_factor + # is 1.0). At the same time, we want to avoid the device-to-host + # transfer by comparing it to 1.0. Since _multiply_factor starts as + # a Python float, we roughly assume that if it's a tensor then it's + # probably not =1.0 anymore and we do the multiplication. Otherwise + # we can safely check the value without a D2H transfer. + torch.is_tensor(self._multiply_factor) + or self._multiply_factor != 1.0 + ): + self.wrapped_optimizer.multiply_grads(self._multiply_factor) + self._multiply_factor = 1.0 + + def multiply_grads(self, c): + """Multiplies grads by a constant *c*.""" + self._multiply_factor *= c + + def clip_grad_norm(self, max_norm, aggregate_norm_fn=None): + """Clips gradient norm and updates dynamic loss scaler.""" + max_norm = float(max_norm) + grad_norm = self._multiply_factor * self.wrapped_optimizer.clip_grad_norm( + 0, aggregate_norm_fn + ) + + if self.scaler is not None: + grad_norm_cpu = float(grad_norm) + if grad_norm_cpu > max_norm > 0.0: + self._multiply_factor *= max_norm / grad_norm_cpu + + # detect overflow and adjust loss scale + self.scaler.check_overflow(grad_norm_cpu) + elif max_norm > 0.0: + clip_coef = (max_norm / (grad_norm + 1e-6)).clamp_(max=1) + self._multiply_factor *= clip_coef + + return grad_norm + + def step(self, closure=None, groups=None): + """Performs a single optimization step.""" + if getattr(self, "supports_step_with_scale", False): + # NOTE(msb) optimizer divides by scale factor + self.wrapped_optimizer.step(closure, scale=(1.0 / self._multiply_factor), groups=groups) + else: + self._unscale_grads() + self.wrapped_optimizer.step(closure, groups=groups) + + if self.scaler is not None: + self.scaler.update() + + def zero_grad(self): + """Clears the gradients of all optimized parameters.""" + self.wrapped_optimizer.zero_grad() + if self.scaler is not None: + self._multiply_factor = 1.0 / float(self.scaler.loss_scale) + else: + self._multiply_factor = 1.0 + + @property + def supports_flat_params(self): + return self.wrapped_optimizer.supports_flat_params + + +class MemoryEfficientFP16Optimizer( + _MemoryEfficientFP16OptimizerMixin, optim.FairseqOptimizer +): + """ + Wrap an *optimizer* to support FP16 (mixed precision) training. + + Compared to :class:`fairseq.optim.FP16Optimizer`, this version does not + maintain an FP32 copy of the model. We instead expect the optimizer to + convert the gradients to FP32 internally and sync the results back to the + FP16 model params. This significantly reduces memory usage but slightly + increases the time spent in the optimizer. + + Since this wrapper depends on specific functionality in the wrapped + optimizer (i.e., on-the-fly conversion of grads to FP32), only certain + optimizers can be wrapped. This is determined by the + *supports_memory_efficient_fp16* property. + """ + + def __init__( + self, cfg: DictConfig, params, optimizer, allow_unsupported=False, **kwargs + ): + if not allow_unsupported and not optimizer.supports_memory_efficient_fp16: + raise ValueError( + "Unsupported optimizer: {}".format(optimizer.__class__.__name__) + ) + + super().__init__(cfg.optimizer) + self.wrapped_optimizer = optimizer + + if getattr(cfg.common, "fp16_scale_window", None) is None: + if len(cfg.optimization.update_freq) > 1: + raise ValueError( + "--fp16-scale-window must be given explicitly when using a " + "custom --update-freq schedule" + ) + data_parallel_size = int( + cfg.distributed_training.distributed_world_size + / cfg.common.model_parallel_size + ) + scale_window = int( + 2 ** 14 / data_parallel_size / cfg.optimization.update_freq[0] + ) + else: + scale_window = cfg.common.fp16_scale_window + + if not getattr(cfg.common, "bf16", False): + self.scaler = DynamicLossScaler( + init_scale=cfg.common.fp16_init_scale, + scale_window=scale_window, + tolerance=cfg.common.fp16_scale_tolerance, + threshold=cfg.common.threshold_loss_scale, + min_loss_scale=cfg.common.min_loss_scale, + ) + else: + # disable loss scaling for bfloat16 + self.scaler = None + + @classmethod + def build_optimizer(cls, cfg: DictConfig, params, **kwargs): + """ + Args: + args (argparse.Namespace): fairseq args + params (iterable): iterable of parameters to optimize + """ + fp16_optimizer = optim.build_optimizer(cfg.optimizer, params) + return cls(cfg, params, fp16_optimizer, **kwargs) + + @property + def optimizer(self): + return self.wrapped_optimizer.optimizer + + @optimizer.setter + def optimizer(self, optimizer): + self.wrapped_optimizer.optimizer = optimizer + + @property + def optimizer_config(self): + return self.wrapped_optimizer.optimizer_config + + @property + def lr_scheduler(self): + return getattr(self.wrapped_optimizer, "lr_scheduler", None) + + def get_lr(self): + return self.wrapped_optimizer.get_lr() + + def set_lr(self, lr): + self.wrapped_optimizer.set_lr(lr) + + def all_reduce_grads(self, module): + self.wrapped_optimizer.all_reduce_grads(module) diff --git a/fairseq/optim/fused_adam.py b/fairseq/optim/fused_adam.py new file mode 100644 index 0000000000000000000000000000000000000000..e2b8e1bcd12c3636ae83be1b21ded4fcd55ccc96 --- /dev/null +++ b/fairseq/optim/fused_adam.py @@ -0,0 +1,348 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import types + +import torch + + +def get_fused_adam_class(): + """ + Look for the FusedAdam optimizer from apex. We first try to load the + "contrib" interface, which is a bit faster than the main interface, + but is technically deprecated. + """ + try: + # The "deprecated" interface in recent versions of apex is a bit + # faster than the main interface, since we don't use the apex + # optimizer. This can be installed by passing the + # `--deprecated_fused_adam` option when building apex. + global fused_adam_cuda + import importlib + + fused_adam_cuda = importlib.import_module("fused_adam_cuda") + return FusedAdamV1 + except ImportError: + try: + # fallback to the newer interface + from apex.optimizers import FusedAdam as _FusedAdam # noqa + from apex.multi_tensor_apply import multi_tensor_applier + + if multi_tensor_applier.available: + return FusedAdamV2 + except ImportError: + pass + return None + + +class FusedAdamV1(torch.optim.Optimizer): + """ + Implements Adam algorithm. Currently GPU-only. Requires Apex to be installed via + ``python setup.py install --cuda_ext --cpp_ext``. + + It has been proposed in `Adam: A Method for Stochastic Optimization`_. + + Compared to the original version in Apex, the fairseq version casts grads + and params to FP32 internally to support ``--memory-efficient-fp16``. + + Args: + params (iterable): iterable of parameters to optimize or dicts defining + parameter groups. + lr (float, optional): learning rate. (default: 1e-3) + betas (Tuple[float, float], optional): coefficients used for computing + running averages of gradient and its square. (default: (0.9, 0.999)) + eps (float, optional): term added to the denominator to improve + numerical stability. (default: 1e-8) + weight_decay (float, optional): weight decay (L2 penalty) (default: 0) + amsgrad (boolean, optional): whether to use the AMSGrad variant of this + algorithm from the paper `On the Convergence of Adam and Beyond`_ + (default: False) NOT SUPPORTED in FusedAdam! + eps_inside_sqrt (boolean, optional): in the 'update parameters' step, + adds eps to the bias-corrected second moment estimate before + evaluating square root instead of adding it to the square root of + second moment estimate as in the original paper. (default: False) + .. _Adam: A Method for Stochastic Optimization: + https://arxiv.org/abs/1412.6980 + .. _On the Convergence of Adam and Beyond: + https://openreview.net/forum?id=ryQu7f-RZ + """ + + def __init__( + self, + params, + lr=1e-3, + bias_correction=True, + betas=(0.9, 0.999), + eps=1e-8, + eps_inside_sqrt=False, + weight_decay=0.0, + max_grad_norm=0.0, + amsgrad=False, + ): + global fused_adam_cuda + import importlib + + fused_adam_cuda = importlib.import_module("fused_adam_cuda") + + if amsgrad: + raise RuntimeError("FusedAdam does not support the AMSGrad variant.") + defaults = { + "lr": lr, + "bias_correction": bias_correction, + "betas": betas, + "eps": eps, + "weight_decay": weight_decay, + "max_grad_norm": max_grad_norm, + } + super().__init__(params, defaults) + self.eps_mode = 0 if eps_inside_sqrt else 1 + + @property + def supports_memory_efficient_fp16(self): + return True + + @property + def supports_flat_params(self): + return True + + @property + def supports_step_with_scale(self): + return True + + def step(self, closure=None, grads=None, scale=1.0, grad_norms=None): + """Performs a single optimization step. + Args: + closure (callable, optional): A closure that reevaluates the model + and returns the loss. + grads (list of tensors, optional): weight gradient to use for the + optimizer update. If gradients have type torch.half, parameters + are expected to be in type torch.float. (default: None) + output params (list of tensors, optional): A reduced precision copy + of the updated weights written out in addition to the regular + updated weights. Have to be of same type as gradients. (default: None) + scale (float, optional): factor to divide gradient tensor values + by before applying to weights. (default: 1) + """ + loss = None + if closure is not None: + loss = closure() + + if grads is None: + grads_group = [None] * len(self.param_groups) + # backward compatibility + # assuming a list/generator of parameter means single group + elif isinstance(grads, types.GeneratorType): + grads_group = [grads] + elif type(grads[0]) != list: + grads_group = [grads] + else: + grads_group = grads + + if grad_norms is None: + grad_norms = [None] * len(self.param_groups) + + for group, grads_this_group, grad_norm in zip( + self.param_groups, grads_group, grad_norms + ): + if grads_this_group is None: + grads_this_group = [None] * len(group["params"]) + + # compute combined scale factor for this group + combined_scale = scale + if group.get("max_grad_norm", 0) > 0: + # norm is in fact norm*scale + clip = ((grad_norm / scale) + 1e-6) / group["max_grad_norm"] + if clip > 1: + combined_scale = clip * scale + + bias_correction = 1 if group.get("bias_correction", 1) else 0 + + for p, grad in zip(group["params"], grads_this_group): + # note: p.grad should not ever be set for correct + # operation of mixed precision optimizer that sometimes + # sends None gradients + if p.grad is None and grad is None: + continue + if grad is None: + grad = p.grad.data + if grad.is_sparse: + raise RuntimeError( + "FusedAdam does not support sparse gradients, " + "please consider SparseAdam instead" + ) + + p_data_fp32 = p.data.float() + + state = self.state[p] + + # State initialization + if len(state) == 0: + state["step"] = 0 + # Exponential moving average of gradient values + state["exp_avg"] = torch.zeros_like(p_data_fp32) + # Exponential moving average of squared gradient values + state["exp_avg_sq"] = torch.zeros_like(p_data_fp32) + else: + state["exp_avg"] = state["exp_avg"].to(p_data_fp32) + state["exp_avg_sq"] = state["exp_avg_sq"].to(p_data_fp32) + + exp_avg = state["exp_avg"] + exp_avg_sq = state["exp_avg_sq"] + beta1, beta2 = group["betas"] + + state["step"] += 1 + + out_p = p.data + with torch.cuda.device(p.device): + fused_adam_cuda.adam( + p_data_fp32, + out_p, + exp_avg, + exp_avg_sq, + grad, + group["lr"], + beta1, + beta2, + group["eps"], + combined_scale, + state["step"], + self.eps_mode, + bias_correction, + group["weight_decay"], + ) + + return loss + + +try: + from apex.optimizers import FusedAdam + from apex.multi_tensor_apply import multi_tensor_applier + + class FusedAdamV2(FusedAdam): + """ + Compared to the original version in Apex, the fairseq version casts grads + and params to FP32 internally to support ``--memory-efficient-fp16``. + """ + + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + if not hasattr(self, "multi_tensor_adam"): + raise Exception( + "Apex installation is outdated. Please install an updated version of apex." + ) + + @property + def supports_memory_efficient_fp16(self): + return True + + @property + def supports_flat_params(self): + return True + + def step( + self, + closure=None, + grads=None, + output_params=None, + scale=None, + grad_norms=None, + ): + """Performs a single optimization step.""" + loss = None + if closure is not None: + loss = closure() + + for group in self.param_groups: + bias_correction = 1 if group["bias_correction"] else 0 + beta1, beta2 = group["betas"] + + # assume same step across group now to simplify things + # per parameter step can be easily support by making it tensor, or pass list into kernel + if "step" in group: + group["step"] += 1 + else: + group["step"] = 1 + + # create lists for multi-tensor apply + g_16, p_16, orig_p_16, m_16, v_16 = [], [], [], [], [] + g_32, p_32, m_32, v_32 = [], [], [], [] + + for p in group["params"]: + if p.grad is None: + continue + if p.grad.data.is_sparse: + raise RuntimeError( + "FusedAdam does not support sparse gradients, " + "please consider SparseAdam instead" + ) + + state = self.state[p] + # State initialization + if len(state) == 0: + # Exponential moving average of gradient values + state["exp_avg"] = torch.zeros_like(p.data, dtype=torch.float) + # Exponential moving average of squared gradient values + state["exp_avg_sq"] = torch.zeros_like( + p.data, dtype=torch.float + ) + else: + state["exp_avg"] = state["exp_avg"].to( + device=p.data.device, dtype=torch.float + ) + state["exp_avg_sq"] = state["exp_avg_sq"].to( + device=p.data.device, dtype=torch.float + ) + + if p.dtype == torch.float16: + g_16.append(p.grad.data.float()) + p_16.append(p.data.float()) + orig_p_16.append(p.data) + m_16.append(state["exp_avg"]) + v_16.append(state["exp_avg_sq"]) + elif p.dtype == torch.float32: + g_32.append(p.grad.data) + p_32.append(p.data) + m_32.append(state["exp_avg"]) + v_32.append(state["exp_avg_sq"]) + else: + raise RuntimeError("FusedAdam only support fp16 and fp32.") + + with torch.cuda.device(p.device): + if len(g_16) > 0: + multi_tensor_applier( + self.multi_tensor_adam, + self._dummy_overflow_buf, + [g_16, p_16, m_16, v_16], + group["lr"], + beta1, + beta2, + group["eps"], + group["step"], + self.adam_w_mode, + bias_correction, + group["weight_decay"], + ) + for orig_p, p in zip(orig_p_16, p_16): + orig_p.copy_(p.data) + if len(g_32) > 0: + multi_tensor_applier( + self.multi_tensor_adam, + self._dummy_overflow_buf, + [g_32, p_32, m_32, v_32], + group["lr"], + beta1, + beta2, + group["eps"], + group["step"], + self.adam_w_mode, + bias_correction, + group["weight_decay"], + ) + + return loss + + +except ImportError: + pass diff --git a/fairseq/optim/fused_lamb.py b/fairseq/optim/fused_lamb.py new file mode 100644 index 0000000000000000000000000000000000000000..f4f2bdb0c6c65f7758509b6d4d2f2c48cb6e8b4f --- /dev/null +++ b/fairseq/optim/fused_lamb.py @@ -0,0 +1,51 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq.optim import LegacyFairseqOptimizer, register_optimizer + + +@register_optimizer("lamb") +class FairseqLAMB(LegacyFairseqOptimizer): + """LAMB optimizer.""" + + def __init__(self, args, params): + super().__init__(args) + try: + from apex.optimizers import FusedLAMB + + self._optimizer = FusedLAMB(params, **self.optimizer_config) + except ImportError: + raise ImportError("Please install apex to use LAMB optimizer") + + @staticmethod + def add_args(parser): + """Add optimizer-specific arguments to the parser.""" + # fmt: off + parser.add_argument('--lamb-betas', default='(0.9, 0.999)', metavar='B', + help='betas for LAMB optimizer') + parser.add_argument('--lamb-eps', type=float, default=1e-8, metavar='D', + help='epsilon for LAMB optimizer') + parser.add_argument('--weight-decay', '--wd', default=0.0, type=float, metavar='WD', + help='weight decay') + # fmt: on + + @property + def optimizer_config(self): + """ + Return a kwarg dictionary that will be used to override optimizer + args stored in checkpoints. This allows us to load a checkpoint and + resume training using a different set of optimizer args, e.g., with a + different learning rate. + """ + return { + "lr": self.args.lr[0], + "betas": eval(self.args.lamb_betas), + "eps": self.args.lamb_eps, + "weight_decay": self.args.weight_decay, + } + + @property + def supports_flat_params(self): + return False diff --git a/fairseq/optim/lr_scheduler/__init__.py b/fairseq/optim/lr_scheduler/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..5b3dbc023aa4a6f7bfb8403b8204d71ca432f79c --- /dev/null +++ b/fairseq/optim/lr_scheduler/__init__.py @@ -0,0 +1,36 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +"""isort:skip_file""" + +import importlib +import os + +from fairseq import registry +from fairseq.optim.lr_scheduler.fairseq_lr_scheduler import ( # noqa + FairseqLRScheduler, + LegacyFairseqLRScheduler, +) +from omegaconf import DictConfig + + +( + build_lr_scheduler_, + register_lr_scheduler, + LR_SCHEDULER_REGISTRY, + LR_SCHEDULER_DATACLASS_REGISTRY, +) = registry.setup_registry( + "--lr-scheduler", base_class=FairseqLRScheduler, default="fixed" +) + + +def build_lr_scheduler(cfg: DictConfig, optimizer): + return build_lr_scheduler_(cfg, optimizer) + + +# automatically import any Python files in the optim/lr_scheduler/ directory +for file in sorted(os.listdir(os.path.dirname(__file__))): + if file.endswith(".py") and not file.startswith("_"): + file_name = file[: file.find(".py")] + importlib.import_module("fairseq.optim.lr_scheduler." + file_name) diff --git a/fairseq/optim/lr_scheduler/cosine_lr_scheduler.py b/fairseq/optim/lr_scheduler/cosine_lr_scheduler.py new file mode 100644 index 0000000000000000000000000000000000000000..51f58359eda387d67748f48217906ac6d16ccd08 --- /dev/null +++ b/fairseq/optim/lr_scheduler/cosine_lr_scheduler.py @@ -0,0 +1,147 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +from collections.abc import Collection +from dataclasses import dataclass, field +from typing import List + +from omegaconf import II + +from fairseq.dataclass import FairseqDataclass +from fairseq.optim.lr_scheduler import FairseqLRScheduler, register_lr_scheduler + + +@dataclass +class CosineLRScheduleConfig(FairseqDataclass): + warmup_updates: int = field( + default=0, + metadata={"help": "warmup the learning rate linearly for the first N updates"}, + ) + warmup_init_lr: float = field( + default=-1, + metadata={ + "help": "initial learning rate during warmup phase; default is cfg.lr" + }, + ) + lr: List[float] = field( + default=II("optimization.lr"), + metadata={"help": "max learning rate, must be more than cfg.min_lr"}, + ) + min_lr: float = field(default=0.0, metadata={"help": "min learning rate"}) + t_mult: float = field( + default=1.0, metadata={"help": "factor to grow the length of each period"} + ) + lr_period_updates: float = field( + default=-1, metadata={"help": "initial number of updates per period"} + ) + lr_shrink: float = field( + default=0.1, metadata={"help": "shrink factor for annealing"} + ) + # This is not required, but is for convenience in inferring lr_period_updates + max_update: int = II("optimization.max_update") + + +@register_lr_scheduler("cosine", dataclass=CosineLRScheduleConfig) +class CosineLRSchedule(FairseqLRScheduler): + """Assign LR based on a cyclical schedule that follows the cosine function. + + See https://arxiv.org/pdf/1608.03983.pdf for details. + + We also support a warmup phase where we linearly increase the learning rate + from some initial learning rate (``--warmup-init-lr``) until the configured + max learning rate (``--lr``). + + During warmup:: + + lrs = torch.linspace(cfg.warmup_init_lr, cfg.lr, cfg.warmup_updates) + lr = lrs[update_num] + + After warmup:: + + lr = cfg.min_lr + 0.5*(cfg.lr - cfg.min_lr)*(1 + cos(t_curr / t_i)) + + where ``t_curr`` is current percentage of updates within the current period + range and ``t_i`` is the current period range, which is scaled by ``t_mul`` + after every iteration. + """ + + def __init__(self, cfg: CosineLRScheduleConfig, fairseq_optimizer): + super().__init__(cfg, fairseq_optimizer) + if isinstance(cfg.lr, Collection) and len(cfg.lr) > 1: + raise ValueError( + "Cannot use a fixed learning rate schedule with cosine." + f" Consider --lr-scheduler=fixed instead. ({cfg.lr})" + ) + + self.max_lr = cfg.lr[0] if isinstance(cfg.lr, Collection) else cfg.lr + assert ( + self.max_lr > cfg.min_lr + ), f"max_lr (={cfg.lr}) must be more than min_lr (={cfg.min_lr})" + + warmup_end_lr = self.max_lr + if cfg.warmup_init_lr < 0: + cfg.warmup_init_lr = cfg.min_lr + + self.t_mult = cfg.t_mult + self.period = cfg.lr_period_updates + + if self.period <= 0: + assert ( + cfg.max_update > 0 + ), "Either --max_update or --lr-period-updates must be set" + self.period = cfg.max_update - cfg.warmup_updates + + if cfg.warmup_updates > 0: + # linearly warmup for the first cfg.warmup_updates + self.lr_step = (warmup_end_lr - cfg.warmup_init_lr) / cfg.warmup_updates + else: + self.lr_step = 1 + + self.warmup_updates = cfg.warmup_updates + self.lr_shrink = cfg.lr_shrink + + # initial learning rate + self.lr = cfg.warmup_init_lr + self.optimizer.set_lr(self.lr) + + def step(self, epoch, val_loss=None): + """Update the learning rate at the end of the given epoch.""" + super().step(epoch, val_loss) + # we don't change the learning rate at epoch boundaries + return self.optimizer.get_lr() + + def step_update(self, num_updates): + """Update the learning rate after each update.""" + if num_updates < self.cfg.warmup_updates: + self.lr = self.cfg.warmup_init_lr + num_updates * self.lr_step + else: + curr_updates = num_updates - self.cfg.warmup_updates + if self.t_mult != 1: + i = math.floor( + math.log( + 1 - curr_updates / self.period * (1 - self.t_mult), self.t_mult + ) + ) + t_i = self.t_mult ** i * self.period + t_curr = ( + curr_updates + - (1 - self.t_mult ** i) / (1 - self.t_mult) * self.period + ) + else: + i = math.floor(curr_updates / self.period) + t_i = self.period + t_curr = curr_updates - (self.period * i) + + lr_shrink = self.lr_shrink ** i + min_lr = self.cfg.min_lr * lr_shrink + max_lr = self.max_lr * lr_shrink + + self.lr = min_lr + 0.5 * (max_lr - min_lr) * ( + 1 + math.cos(math.pi * t_curr / t_i) + ) + + self.optimizer.set_lr(self.lr) + return self.lr diff --git a/fairseq/optim/lr_scheduler/fairseq_lr_scheduler.py b/fairseq/optim/lr_scheduler/fairseq_lr_scheduler.py new file mode 100644 index 0000000000000000000000000000000000000000..6c12fa56b825e81bcc3fc7a97d206777418260ef --- /dev/null +++ b/fairseq/optim/lr_scheduler/fairseq_lr_scheduler.py @@ -0,0 +1,59 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from argparse import Namespace + +from fairseq.dataclass.utils import gen_parser_from_dataclass +from fairseq.optim import FairseqOptimizer + + +class FairseqLRScheduler(object): + def __init__(self, cfg, optimizer): + super().__init__() + if optimizer is not None and not isinstance(optimizer, FairseqOptimizer): + raise ValueError("optimizer must be an instance of FairseqOptimizer") + self.cfg = cfg + self.optimizer = optimizer + self.best = None + + @classmethod + def add_args(cls, parser): + """Add arguments to the parser for this LR scheduler.""" + dc = getattr(cls, "__dataclass", None) + if dc is not None: + gen_parser_from_dataclass(parser, dc()) + + def state_dict(self): + """Return the LR scheduler state dict.""" + return {"best": self.best} + + def load_state_dict(self, state_dict): + """Load an LR scheduler state dict.""" + self.best = state_dict["best"] + + def step_begin_epoch(self, epoch): + """Update the learning rate at the beginning of the given epoch.""" + pass + + def step(self, epoch, val_loss=None): + """Update the learning rate at the end of the given epoch.""" + if val_loss is not None: + if self.best is None: + self.best = val_loss + else: + self.best = min(self.best, val_loss) + + def step_update(self, num_updates): + """Update the learning rate after each update.""" + return self.optimizer.get_lr() + + +class LegacyFairseqLRScheduler(FairseqLRScheduler): + def __init__(self, args: Namespace, optimizer): + if not isinstance(optimizer, FairseqOptimizer): + raise ValueError("optimizer must be an instance of FairseqOptimizer") + self.args = args + self.optimizer = optimizer + self.best = None diff --git a/fairseq/optim/lr_scheduler/fixed_schedule.py b/fairseq/optim/lr_scheduler/fixed_schedule.py new file mode 100644 index 0000000000000000000000000000000000000000..d0e7e14b7e72b1151f7d7f19094430bbab64f8f0 --- /dev/null +++ b/fairseq/optim/lr_scheduler/fixed_schedule.py @@ -0,0 +1,76 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from dataclasses import dataclass, field +from typing import Optional, List +from omegaconf import II + +from fairseq.dataclass import FairseqDataclass +from fairseq.optim.lr_scheduler import FairseqLRScheduler, register_lr_scheduler + + +@dataclass +class FixedLRScheduleConfig(FairseqDataclass): + force_anneal: Optional[int] = field( + default=None, + metadata={"help": "force annealing at specified epoch"}, + ) + lr_shrink: float = field( + default=0.1, + metadata={"help": "shrink factor for annealing, lr_new = (lr * lr_shrink)"}, + ) + warmup_updates: int = field( + default=0, + metadata={"help": "warmup the learning rate linearly for the first N updates"}, + ) + lr: List[float] = II("optimization.lr") + + +@register_lr_scheduler("fixed", dataclass=FixedLRScheduleConfig) +class FixedLRSchedule(FairseqLRScheduler): + """Decay the LR on a fixed schedule.""" + + def __init__(self, cfg: FixedLRScheduleConfig, optimizer): + super().__init__(cfg, optimizer) + + self.lr = cfg.lr[0] + if cfg.warmup_updates > 0: + self.warmup_factor = 1.0 / cfg.warmup_updates + else: + self.warmup_factor = 1 + + def state_dict(self): + return {"lr": self.lr} + + def load_state_dict(self, state_dict): + if "lr" in state_dict: + self.lr = state_dict["lr"] + + def get_next_lr(self, epoch): + lrs = self.cfg.lr + if self.cfg.force_anneal is None or epoch < self.cfg.force_anneal: + # use fixed LR schedule + next_lr = lrs[min(epoch - 1, len(lrs) - 1)] + else: + # annneal based on lr_shrink + next_lr = lrs[-1] * self.cfg.lr_shrink ** ( + epoch + 1 - self.cfg.force_anneal + ) + return next_lr + + def step_begin_epoch(self, epoch): + """Update the learning rate at the beginning of the given epoch.""" + self.lr = self.get_next_lr(epoch) + self.optimizer.set_lr(self.warmup_factor * self.lr) + return self.optimizer.get_lr() + + def step_update(self, num_updates): + """Update the learning rate after each update.""" + if self.cfg.warmup_updates > 0 and num_updates < self.cfg.warmup_updates: + self.warmup_factor = (num_updates + 1) / float(self.cfg.warmup_updates) + self.optimizer.set_lr(self.warmup_factor * self.lr) + else: + self.optimizer.set_lr(self.lr) + return self.optimizer.get_lr() diff --git a/fairseq/optim/lr_scheduler/inverse_square_root_schedule.py b/fairseq/optim/lr_scheduler/inverse_square_root_schedule.py new file mode 100644 index 0000000000000000000000000000000000000000..0f87bb5d7ed5c7eb8011d4c651f2ecbf0ae700ac --- /dev/null +++ b/fairseq/optim/lr_scheduler/inverse_square_root_schedule.py @@ -0,0 +1,85 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from collections.abc import Collection +from dataclasses import dataclass, field +from typing import List + +from omegaconf import II + +from fairseq.dataclass import FairseqDataclass +from fairseq.optim.lr_scheduler import FairseqLRScheduler, register_lr_scheduler + + +@dataclass +class InverseSquareRootLRScheduleConfig(FairseqDataclass): + warmup_updates: int = field( + default=4000, + metadata={"help": "warmup the learning rate linearly for the first N updates"}, + ) + warmup_init_lr: float = field( + default=-1, + metadata={ + "help": "initial learning rate during warmup phase; default is cfg.lr" + }, + ) + lr: List[float] = II("optimization.lr") + + +@register_lr_scheduler("inverse_sqrt", dataclass=InverseSquareRootLRScheduleConfig) +class InverseSquareRootSchedule(FairseqLRScheduler): + """Decay the LR based on the inverse square root of the update number. + + We also support a warmup phase where we linearly increase the learning rate + from some initial learning rate (``--warmup-init-lr``) until the configured + learning rate (``--lr``). Thereafter we decay proportional to the number of + updates, with a decay factor set to align with the configured learning rate. + + During warmup:: + + lrs = torch.linspace(cfg.warmup_init_lr, cfg.lr, cfg.warmup_updates) + lr = lrs[update_num] + + After warmup:: + + decay_factor = cfg.lr * sqrt(cfg.warmup_updates) + lr = decay_factor / sqrt(update_num) + """ + + def __init__(self, cfg: InverseSquareRootLRScheduleConfig, optimizer): + super().__init__(cfg, optimizer) + if isinstance(cfg.lr, Collection) and len(cfg.lr) > 1: + raise ValueError( + "Cannot use a fixed learning rate schedule with inverse_sqrt." + " Consider --lr-scheduler=fixed instead." + ) + warmup_end_lr = cfg.lr[0] if isinstance(cfg.lr, Collection) else cfg.lr + if cfg.warmup_init_lr < 0: + cfg.warmup_init_lr = 0 if cfg.warmup_updates > 0 else warmup_end_lr + + # linearly warmup for the first cfg.warmup_updates + self.lr_step = (warmup_end_lr - cfg.warmup_init_lr) / cfg.warmup_updates + + # then, decay prop. to the inverse square root of the update number + self.decay_factor = warmup_end_lr * cfg.warmup_updates ** 0.5 + + # initial learning rate + self.lr = cfg.warmup_init_lr + self.optimizer.set_lr(self.lr) + + def step(self, epoch, val_loss=None): + """Update the learning rate at the end of the given epoch.""" + super().step(epoch, val_loss) + # we don't change the learning rate at epoch boundaries + return self.optimizer.get_lr() + + def step_update(self, num_updates): + """Update the learning rate after each update.""" + if num_updates < self.cfg.warmup_updates: + self.lr = self.cfg.warmup_init_lr + num_updates * self.lr_step + else: + self.lr = self.decay_factor * num_updates ** -0.5 + self.optimizer.set_lr(self.lr) + return self.lr diff --git a/fairseq/optim/lr_scheduler/manual_lr_scheduler.py b/fairseq/optim/lr_scheduler/manual_lr_scheduler.py new file mode 100644 index 0000000000000000000000000000000000000000..0269a1e2853854745e23b07931294f37b67d0295 --- /dev/null +++ b/fairseq/optim/lr_scheduler/manual_lr_scheduler.py @@ -0,0 +1,110 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from . import LegacyFairseqLRScheduler, register_lr_scheduler +import logging +import ast + +logger = logging.getLogger(__name__) +logger.setLevel(logging.WARNING) + + +@register_lr_scheduler("manual") +class ManualSchedule(LegacyFairseqLRScheduler): + """Decay the LR on a manual schedule.""" + + def __init__(self, args, optimizer): + super().__init__(args, optimizer) + + self.epoch2lr = self.parse_manuallr_args(args.epoch2lr) + self.update2lr = self.parse_manuallr_args(args.update2lr) + logger.info("@@@ ManualSchedule epoch2lr={}".format(self.epoch2lr)) + logger.info("@@@ ManualSchedule update2lr={}".format(self.update2lr)) + + if 1 in self.epoch2lr: + self.lr = self.epoch2lr[1] + elif 1 in self.update2lr: + self.lr = self.update2lr[1] + else: + self.lr = args.lr[0] + self.optimizer.set_lr(self.lr) # Set the beginning of the epoch. + + def parse_manuallr_args(self, lr_args_str): + lr_dict = ast.literal_eval(lr_args_str.replace(' ', '')) + if not isinstance(lr_dict, dict): + raise ValueError("epoch2lr/update2lr must be abel to evaluated to a dict") + + lr_args = {} + logger.info("@@@ after parsing input dictionary lr_dict = {}".format(lr_dict)) + for key, val in lr_dict.items(): + if "," in key: + for k in key.split(","): + lr_args[int(k)] = float(val) + elif "-" in key: + s = int(key.split("-")[0]) + e = int(key.split("-")[1]) + for k in range(s, e + 1, 1): + lr_args[k] = float(val) + else: + lr_args[int(key)] = float(val) + + return lr_args + + @staticmethod + def add_args(parser): + """Add arguments to the parser for this LR scheduler.""" + # fmt: off + parser.add_argument( + "--epoch2lr", + type=str, + metavar="DICT", + default="{}", + help="a dictionary used to set lr for each epoch manually", + ) + parser.add_argument( + "--update2lr", + type=str, + metavar="DICT", + default="{}", + help="a dictionary used to set lr for each update manually", + ) + # fmt: on + + def state_dict(self): + return {"lr": self.lr} + + def load_state_dict(self, state_dict): + if "lr" in state_dict: + self.lr = state_dict["lr"] + + def get_next_lr(self, epoch): + manual_keys = [k for k in self.epoch2lr if k <= epoch] + if manual_keys: + manual_lr = self.epoch2lr[max(manual_keys)] + else: + logger.warning("@@@ epoch={} does not exist in manual lr input. epoch2lr={}...".format( + epoch, list(self.epoch2lr.items())[:min(10, len(self.epoch2lr.keys())-1)] + )) + manual_lr = self.optimizer.get_lr() + return manual_lr + + def step_begin_epoch(self, epoch): + """Update the learning rate at the beginning of the given epoch.""" + self.lr = self.get_next_lr(epoch) + self.optimizer.set_lr(self.lr) + return self.optimizer.get_lr() + + def step_update(self, num_updates): + """Update the learning rate after each update.""" + manual_keys = [k for k in self.update2lr if k <= num_updates] + if manual_keys: + manual_lr = self.update2lr[max(manual_keys)] + else: + logger.warning("epoch={} does not exist in manual lr input update2lr={}...".format( + num_updates, list(self.update2lr.items())[:min(10, len(self.update2lr.keys())-1)])) + manual_lr = self.optimizer.get_lr() + + self.optimizer.set_lr(manual_lr) + return self.optimizer.get_lr() diff --git a/fairseq/optim/lr_scheduler/pass_through.py b/fairseq/optim/lr_scheduler/pass_through.py new file mode 100644 index 0000000000000000000000000000000000000000..2f93db328c1de9b268e8ee1c0c1cad558fd089aa --- /dev/null +++ b/fairseq/optim/lr_scheduler/pass_through.py @@ -0,0 +1,39 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from dataclasses import dataclass + +from fairseq.dataclass import FairseqDataclass +from fairseq.optim.lr_scheduler import FairseqLRScheduler, register_lr_scheduler + + +@dataclass +class PassThroughScheduleConfig(FairseqDataclass): + pass + + +@register_lr_scheduler("pass_through", dataclass=PassThroughScheduleConfig) +class PassThroughScheduleSchedule(FairseqLRScheduler): + """Delegate lr scheduling to the optimizer.""" + + def __init__(self, cfg: PassThroughScheduleConfig, optimizer): + super().__init__(cfg, optimizer) + assert ( + hasattr(optimizer, "lr_scheduler") and optimizer.lr_scheduler is not None + ), "Pass-through schedule can only be used with optimizers with their own schedulers" + + def state_dict(self): + return self.optimizer.lr_scheduler.state_dict() + + def load_state_dict(self, state_dict): + self.optimizer.lr_scheduler.load_state_dict(state_dict) + + def step_begin_epoch(self, epoch): + """Update the learning rate at the beginning of the given epoch.""" + return self.optimizer.lr_scheduler.step_begin_epoch(epoch) + + def step_update(self, num_updates): + """Update the learning rate after each update.""" + return self.optimizer.lr_scheduler.step_update(num_updates) diff --git a/fairseq/optim/lr_scheduler/polynomial_decay_schedule.py b/fairseq/optim/lr_scheduler/polynomial_decay_schedule.py new file mode 100644 index 0000000000000000000000000000000000000000..b8109a7c1e79cd057c355504d07bac5615c02ea9 --- /dev/null +++ b/fairseq/optim/lr_scheduler/polynomial_decay_schedule.py @@ -0,0 +1,89 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from dataclasses import dataclass, field +from typing import Optional, List +from omegaconf import II + +from fairseq.dataclass import FairseqDataclass +from fairseq.optim.lr_scheduler import FairseqLRScheduler, register_lr_scheduler + + +@dataclass +class PolynomialDecayLRScheduleConfig(FairseqDataclass): + warmup_updates: int = field( + default=0, + metadata={"help": "warmup the learning rate linearly for the first N updates"}, + ) + force_anneal: Optional[int] = field( + default=None, + metadata={"help": "force annealing at specified epoch"}, + ) + end_learning_rate: float = field( + default=0.0, + metadata={"help": "learning rate to decay to"}, + ) + power: float = field( + default=1.0, + metadata={"help": "decay exponent"}, + ) + total_num_update: float = field( + default=II("optimization.max_update"), + metadata={"help": "total number of updates over which to decay learning rate"}, + ) + lr: List[float] = II("optimization.lr") + + +@register_lr_scheduler("polynomial_decay", dataclass=PolynomialDecayLRScheduleConfig) +class PolynomialDecayLRSchedule(FairseqLRScheduler): + """Decay the LR on a fixed schedule.""" + + def __init__(self, cfg: PolynomialDecayLRScheduleConfig, optimizer): + super().__init__(cfg, optimizer) + + assert cfg.total_num_update > 0 + + self.lr = cfg.lr[0] + if cfg.warmup_updates > 0: + self.warmup_factor = 1.0 / cfg.warmup_updates + else: + self.warmup_factor = 1 + self.end_learning_rate = cfg.end_learning_rate + self.total_num_update = cfg.total_num_update + self.power = cfg.power + self.optimizer.set_lr(self.warmup_factor * self.lr) + + def get_next_lr(self, epoch): + lrs = self.cfg.lr + if self.cfg.force_anneal is None or epoch < self.cfg.force_anneal: + # use fixed LR schedule + next_lr = lrs[min(epoch, len(lrs) - 1)] + else: + # annneal based on lr_shrink + next_lr = self.optimizer.get_lr() + return next_lr + + def step_begin_epoch(self, epoch): + """Update the learning rate at the beginning of the given epoch.""" + self.lr = self.get_next_lr(epoch) + self.optimizer.set_lr(self.warmup_factor * self.lr) + return self.optimizer.get_lr() + + def step_update(self, num_updates): + """Update the learning rate after each update.""" + if self.cfg.warmup_updates > 0 and num_updates <= self.cfg.warmup_updates: + self.warmup_factor = num_updates / float(self.cfg.warmup_updates) + lr = self.warmup_factor * self.lr + elif num_updates >= self.total_num_update: + lr = self.end_learning_rate + else: + warmup = self.cfg.warmup_updates + lr_range = self.lr - self.end_learning_rate + pct_remaining = 1 - (num_updates - warmup) / ( + self.total_num_update - warmup + ) + lr = lr_range * pct_remaining ** (self.power) + self.end_learning_rate + self.optimizer.set_lr(lr) + return self.optimizer.get_lr() diff --git a/fairseq/optim/lr_scheduler/reduce_lr_on_plateau.py b/fairseq/optim/lr_scheduler/reduce_lr_on_plateau.py new file mode 100644 index 0000000000000000000000000000000000000000..6e29ba79b6b848fda0dab103d05483bd623f3688 --- /dev/null +++ b/fairseq/optim/lr_scheduler/reduce_lr_on_plateau.py @@ -0,0 +1,143 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from dataclasses import dataclass, field +from typing import List + +import torch.optim.lr_scheduler +from omegaconf import II + +from fairseq.dataclass import FairseqDataclass +from fairseq.optim.lr_scheduler import FairseqLRScheduler, register_lr_scheduler + + +@dataclass +class ReduceLROnPlateauLRScheduleConfig(FairseqDataclass): + lr_shrink: float = field( + default=0.1, metadata={"help": "shrink factor for annealing"} + ) + lr_threshold: float = field( + default=1e-4, + metadata={ + "help": ( + "threshold for measuring the new optimum, to only focus on " + "significant changes" + ) + }, + ) + lr_patience: int = field( + default=0, + metadata={ + "help": ( + "number of epochs with no improvement after which learning rate will " + "be reduced" + ) + }, + ) + warmup_updates: int = field( + default=0, + metadata={"help": "warmup the learning rate linearly for the first N updates"}, + ) + warmup_init_lr: float = field( + default=-1, + metadata={ + "help": "initial learning rate during warmup phase; default is cfg.lr" + }, + ) + lr: List[float] = II("optimization.lr") + maximize_best_checkpoint_metric: bool = II( + "checkpoint.maximize_best_checkpoint_metric" + ) + + +@register_lr_scheduler( + "reduce_lr_on_plateau", dataclass=ReduceLROnPlateauLRScheduleConfig +) +class ReduceLROnPlateauLRSchedule(FairseqLRScheduler): + """ + Decay the LR by a factor every time the validation loss plateaus. + Also comes with optional warmup phase, where we linearly increase + the learning rate from some initial learning rate + (``--warmup-init-lr``) until the configured learning rate + (``--lr``). Thereafter the lr is adjusted according to original + reduce_on_plateau scheme. + + During warmup:: + + lrs = torch.linspace( + cfg.warmup_init_lr, cfg.lr, cfg.warmup_updates + ) + lr = lrs[update_num] + """ + + def __init__(self, cfg: ReduceLROnPlateauLRScheduleConfig, optimizer): + super().__init__(cfg, optimizer) + if len(cfg.lr) > 1: + raise ValueError( + "Cannot use a fixed learning rate schedule with reduce_lr_on_plateau." + " Consider --lr-scheduler=fixed instead." + ) + self.lr_scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau( + self.optimizer.optimizer, + patience=cfg.lr_patience, + factor=cfg.lr_shrink, + mode="max" if cfg.maximize_best_checkpoint_metric else "min", + threshold=cfg.lr_threshold, + ) + warmup_end_lr = cfg.lr[0] + # if no warm up, sets initial lr to be cfg.lr[0] + if cfg.warmup_init_lr < 0: + cfg.warmup_init_lr = 0 if cfg.warmup_updates > 0 else warmup_end_lr + + # linearly warmup for the first cfg.warmup_updates + if cfg.warmup_updates > 0: + self.lr_step = (warmup_end_lr - cfg.warmup_init_lr) / cfg.warmup_updates + + # this flag is either set from arg when no warm up, or set by + # step_update() when warmup finishes + self.warmup_end = True if cfg.warmup_updates <= 0 else False + + # initial learning rate + # this self.lr is used only during init and/or warm up period + self.lr = cfg.warmup_init_lr + self.optimizer.set_lr(self.lr) + + def state_dict(self): + """Return the LR scheduler state dict.""" + return { + "best": self.lr_scheduler.best, + "last_epoch": self.lr_scheduler.last_epoch, + } + + def load_state_dict(self, state_dict): + """Load an LR scheduler state dict.""" + self.lr_scheduler.best = state_dict["best"] + if "last_epoch" in state_dict: + self.lr_scheduler.last_epoch = state_dict["last_epoch"] + + def step(self, epoch, val_loss=None): + """ + Update the learning rate at the end of the given epoch if warmup + finishes otherwise no update of lr on epoch boundaries + """ + if val_loss is not None and self.warmup_end is True: + self.lr_scheduler.step(val_loss) + else: + self.lr_scheduler.last_epoch = epoch + return self.optimizer.get_lr() + + def step_update(self, num_updates): + """ + Update the learning rate after each update.""" + # if there is warmup + if self.cfg.warmup_updates > 0: + if num_updates <= self.cfg.warmup_updates: + self.lr = self.cfg.warmup_init_lr + num_updates * self.lr_step + self.optimizer.set_lr(self.lr) + else: + if self.warmup_end is False: + self.warmup_end = True + # else do nothing + return self.optimizer.get_lr() diff --git a/fairseq/optim/lr_scheduler/tri_stage_lr_scheduler.py b/fairseq/optim/lr_scheduler/tri_stage_lr_scheduler.py new file mode 100644 index 0000000000000000000000000000000000000000..4d5547c39b14f62acbd4f4b9ab3abfb3009c0e6d --- /dev/null +++ b/fairseq/optim/lr_scheduler/tri_stage_lr_scheduler.py @@ -0,0 +1,175 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +from dataclasses import dataclass, field +from typing import Optional, List, Tuple +from omegaconf import II + +from fairseq.dataclass import FairseqDataclass +from fairseq.optim.lr_scheduler import FairseqLRScheduler, register_lr_scheduler + + +@dataclass +class TriStageLRScheduleConfig(FairseqDataclass): + warmup_steps: int = field( + default=0, + metadata={"help": "warmup the learning rate linearly for the first N updates"}, + ) + hold_steps: int = field( + default=0, + metadata={"help": "steps in hold stage"}, + ) + decay_steps: int = field( + default=0, + metadata={"help": "steps in decay stages"}, + ) + phase_ratio: Optional[Tuple[float, float, float]] = field( + default=None, + metadata={ + "help": ( + "if set, automatically sets warmup/hold/decay steps to the ratio " + "specified here from max_updates. the ratios must add up to 1.0" + ) + }, + ) + init_lr_scale: float = field( + default=0.01, + metadata={"help": "initial learning rate scale during warmup phase"}, + ) + final_lr_scale: float = field( + default=0.01, + metadata={"help": "final learning rate scale"}, + ) + max_update: float = II("optimization.max_update") + lr: List[float] = II("optimization.lr") + + +@register_lr_scheduler("tri_stage", dataclass=TriStageLRScheduleConfig) +class TriStageLRSchedule(FairseqLRScheduler): + """Tristage learning rate schedulr + + Implement the learning rate scheduler in https://arxiv.org/pdf/1904.08779.pdf + + Similar to inverse_squre_root scheduler, but tri_stage learning rate employs + three stages LR scheduling: + + - warmup stage, starting from `lr` * `init_lr_scale`, linearly + increased to `lr` in `warmup_steps` iterations + + - hold stage, after `warmup_steps`, keep the LR as `lr` for `hold_steps` + iterations + + - decay stage, after hold stage, decay LR exponetially to + `lr` * `final_lr_scale` in `decay_steps`; + after that LR is keep as `final_lr_scale` * `lr` + + During warmup:: + + init_lr = cfg.init_lr_scale * cfg.lr + lrs = torch.linspace(init_lr, cfg.lr, cfg.warmup_steps) + lr = lrs[update_num] + + During hold:: + + lr = cfg.lr + + During decay:: + + decay_factor = - math.log(cfg.final_lr_scale) / cfg.decay_steps + lr = cfg.lr * exp(- (update_num - warmup_steps - decay_steps) * decay_factor) + + After that:: + + lr = cfg.lr * cfg.final_lr_scale + """ + + def __init__(self, cfg: TriStageLRScheduleConfig, optimizer): + super().__init__(cfg, optimizer) + if len(cfg.lr) > 1: + raise ValueError( + "Cannot use a fixed learning rate schedule with tri-stage lr." + " Consider --lr-scheduler=fixed instead." + ) + + # calculate LR at each point + self.peak_lr = cfg.lr[0] + self.init_lr = cfg.init_lr_scale * cfg.lr[0] + self.final_lr = cfg.final_lr_scale * cfg.lr[0] + + if cfg.phase_ratio is not None: + assert cfg.max_update > 0 + assert sum(cfg.phase_ratio) == 1, "phase ratios must add up to 1" + self.warmup_steps = int(cfg.max_update * cfg.phase_ratio[0]) + self.hold_steps = int(cfg.max_update * cfg.phase_ratio[1]) + self.decay_steps = int(cfg.max_update * cfg.phase_ratio[2]) + else: + self.warmup_steps = cfg.warmup_steps + self.hold_steps = cfg.hold_steps + self.decay_steps = cfg.decay_steps + + assert ( + self.warmup_steps + self.hold_steps + self.decay_steps > 0 + ), "please specify steps or phase_ratio" + + self.warmup_rate = ( + (self.peak_lr - self.init_lr) / self.warmup_steps + if self.warmup_steps != 0 + else 0 + ) + self.decay_factor = -math.log(cfg.final_lr_scale) / self.decay_steps + + # initial learning rate + self.lr = self.init_lr + self.optimizer.set_lr(self.lr) + + def _decide_stage(self, update_step): + """ + return stage, and the corresponding steps within the current stage + """ + if update_step < self.warmup_steps: + # warmup state + return 0, update_step + + offset = self.warmup_steps + + if update_step < offset + self.hold_steps: + # hold stage + return 1, update_step - offset + + offset += self.hold_steps + + if update_step <= offset + self.decay_steps: + # decay stage + return 2, update_step - offset + + offset += self.decay_steps + + # still here ? constant lr stage + return 3, update_step - offset + + def step(self, epoch, val_loss=None): + """Update the learning rate at the end of the given epoch.""" + super().step(epoch, val_loss) + # we don't change the learning rate at epoch boundaries + return self.optimizer.get_lr() + + def step_update(self, num_updates): + """Update the learning rate after each update.""" + stage, steps_in_stage = self._decide_stage(num_updates) + if stage == 0: + self.lr = self.init_lr + self.warmup_rate * steps_in_stage + elif stage == 1: + self.lr = self.peak_lr + elif stage == 2: + self.lr = self.peak_lr * math.exp(-self.decay_factor * steps_in_stage) + elif stage == 3: + self.lr = self.final_lr + else: + raise ValueError("Undefined stage") + + self.optimizer.set_lr(self.lr) + + return self.lr diff --git a/fairseq/optim/lr_scheduler/triangular_lr_scheduler.py b/fairseq/optim/lr_scheduler/triangular_lr_scheduler.py new file mode 100644 index 0000000000000000000000000000000000000000..bfe2a0d381f28525f90ee120b31a69210338eb1b --- /dev/null +++ b/fairseq/optim/lr_scheduler/triangular_lr_scheduler.py @@ -0,0 +1,83 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +from dataclasses import dataclass, field +from typing import List + +from omegaconf import II + +from fairseq.dataclass import FairseqDataclass +from fairseq.optim.lr_scheduler import FairseqLRScheduler, register_lr_scheduler + + +@dataclass +class TriangularLRScheduleConfig(FairseqDataclass): + max_lr: float = field( + default="???", metadata={"help": "max learning rate, must be more than cfg.lr"} + ) + lr_period_updates: float = field( + default=5000, + metadata={"help": "initial number of updates per period (cycle length)"}, + ) + lr_shrink: float = field( + default=0.1, metadata={"help": "shrink factor for annealing"} + ) + shrink_min: bool = field( + default=False, metadata={"help": "if set, also shrinks min lr"} + ) + lr: List[float] = II("optimization.lr") + + +@register_lr_scheduler("triangular", dataclass=TriangularLRScheduleConfig) +class TriangularLRSchedule(FairseqLRScheduler): + """Assign LR based on a triangular cyclical schedule. + + See https://arxiv.org/pdf/1506.01186.pdf for details. + """ + + def __init__(self, cfg: TriangularLRScheduleConfig, optimizer): + super().__init__(cfg, optimizer) + if len(cfg.lr) > 1: + raise ValueError( + "Cannot use a fixed learning rate schedule with triangular." + " Consider --lr-scheduler=fixed instead." + ) + + lr = cfg.lr[0] + + assert cfg.max_lr > lr, "max_lr must be more than lr" + self.min_lr = lr + self.max_lr = cfg.max_lr + self.stepsize = cfg.lr_period_updates // 2 + self.lr_shrink = cfg.lr_shrink + self.shrink_min = cfg.shrink_min + + # initial learning rate + self.lr = self.min_lr + self.optimizer.set_lr(self.lr) + + def step(self, epoch, val_loss=None): + """Update the learning rate at the end of the given epoch.""" + super().step(epoch, val_loss) + # we don't change the learning rate at epoch boundaries + return self.optimizer.get_lr() + + def step_update(self, num_updates): + """Update the learning rate after each update.""" + cycle = math.floor(num_updates / (2 * self.stepsize)) + + lr_shrink = self.lr_shrink ** cycle + max_lr = self.max_lr * lr_shrink + if self.shrink_min: + min_lr = self.min_lr * lr_shrink + else: + min_lr = self.min_lr + + x = abs(num_updates / self.stepsize - 2 * (cycle + 1) + 1) + self.lr = min_lr + (max_lr - min_lr) * max(0, (1 - x)) + + self.optimizer.set_lr(self.lr) + return self.lr diff --git a/fairseq/optim/nag.py b/fairseq/optim/nag.py new file mode 100644 index 0000000000000000000000000000000000000000..c30a6c0fb1e8d5dc7edd5b53ba15a6acd46ecbff --- /dev/null +++ b/fairseq/optim/nag.py @@ -0,0 +1,111 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from collections.abc import Collection +from dataclasses import dataclass, field +from typing import List + +import torch +from fairseq.dataclass import FairseqDataclass +from omegaconf import II, DictConfig +from torch.optim.optimizer import Optimizer, required + +from . import FairseqOptimizer, register_optimizer + + +@dataclass +class FairseqNAGConfig(FairseqDataclass): + momentum: float = field(default=0.99, metadata={"help": "momentum factor"}) + weight_decay: float = field(default=0.0, metadata={"help": "weight decay"}) + # TODO common vars in parent class + lr: List[float] = II("optimization.lr") + + +@register_optimizer("nag", dataclass=FairseqNAGConfig) +class FairseqNAG(FairseqOptimizer): + def __init__(self, cfg: DictConfig, params): + super().__init__(cfg) + self._optimizer = NAG(params, **self.optimizer_config) + + @property + def optimizer_config(self): + """ + Return a kwarg dictionary that will be used to override optimizer + args stored in checkpoints. This allows us to load a checkpoint and + resume training using a different set of optimizer args, e.g., with a + different learning rate. + """ + return { + "lr": self.cfg.lr[0] + if isinstance(self.cfg.lr, Collection) + else self.cfg.lr, + "momentum": self.cfg.momentum, + "weight_decay": self.cfg.weight_decay, + } + + +class NAG(Optimizer): + def __init__(self, params, lr=required, momentum=0, weight_decay=0): + defaults = dict(lr=lr, lr_old=lr, momentum=momentum, weight_decay=weight_decay) + super(NAG, self).__init__(params, defaults) + + @property + def supports_memory_efficient_fp16(self): + return True + + @property + def supports_flat_params(self): + return True + + def step(self, closure=None): + """Performs a single optimization step. + + Args: + closure (callable, optional): A closure that reevaluates the model + and returns the loss. + """ + loss = None + if closure is not None: + loss = closure() + + for group in self.param_groups: + weight_decay = group["weight_decay"] + momentum = group["momentum"] + lr = group["lr"] + lr_old = group.get("lr_old", lr) + lr_correct = lr / lr_old if lr_old > 0 else lr + + for p in group["params"]: + if p.grad is None: + continue + + p_data_fp32 = p.data + if p_data_fp32.dtype in {torch.float16, torch.bfloat16}: + p_data_fp32 = p_data_fp32.float() + + d_p = p.grad.data.float() + param_state = self.state[p] + if "momentum_buffer" not in param_state: + param_state["momentum_buffer"] = torch.zeros_like(d_p) + else: + param_state["momentum_buffer"] = param_state["momentum_buffer"].to( + d_p + ) + + buf = param_state["momentum_buffer"] + + if weight_decay != 0: + p_data_fp32.mul_(1 - lr * weight_decay) + p_data_fp32.add_(buf, alpha=momentum * momentum * lr_correct) + p_data_fp32.add_(d_p, alpha=-(1 + momentum) * lr) + + buf.mul_(momentum * lr_correct).add_(d_p, alpha=-lr) + + if p.data.dtype in {torch.float16, torch.bfloat16}: + p.data.copy_(p_data_fp32) + + group["lr_old"] = lr + + return loss diff --git a/fairseq/optim/sgd.py b/fairseq/optim/sgd.py new file mode 100644 index 0000000000000000000000000000000000000000..8e34fb99a18fff12ab76be5894a84cbbb2f48176 --- /dev/null +++ b/fairseq/optim/sgd.py @@ -0,0 +1,43 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch.optim + +from . import LegacyFairseqOptimizer, register_optimizer + + +@register_optimizer("sgd") +class SGD(LegacyFairseqOptimizer): + def __init__(self, args, params): + super().__init__(args) + self._optimizer = torch.optim.SGD(params, **self.optimizer_config) + + @staticmethod + def add_args(parser): + """Add optimizer-specific arguments to the parser.""" + # fmt: off + parser.add_argument('--momentum', default=0.0, type=float, metavar='M', + help='momentum factor') + parser.add_argument('--weight-decay', '--wd', default=0.0, type=float, metavar='WD', + help='weight decay') + # fmt: on + + @property + def optimizer_config(self): + """ + Return a kwarg dictionary that will be used to override optimizer + args stored in checkpoints. This allows us to load a checkpoint and + resume training using a different set of optimizer args, e.g., with a + different learning rate. + """ + return { + "lr": self.args.lr[0], + "momentum": self.args.momentum, + "weight_decay": self.args.weight_decay, + } + + @property + def supports_flat_params(self): + return True diff --git a/fairseq/optim/shard.py b/fairseq/optim/shard.py new file mode 100644 index 0000000000000000000000000000000000000000..9d7f2eb9e5de6086fe2435d432bde7521ebb8155 --- /dev/null +++ b/fairseq/optim/shard.py @@ -0,0 +1,58 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from typing import Any, Dict + +from fairseq.distributed import utils + + +try: + from fairscale.optim import OSS + + _has_fairscale = True +except ImportError: + _has_fairscale = False + + +def shard_(optimizer, group): + if not _has_fairscale: + raise ImportError( + "\n\nPlease install the fairscale package:" "\n\n pip install fairscale" + ) + + class FairseqOSS(OSS): + @property + def disable_mem_eff_fp16_loading_hack(self): + return True + + def __getattr__(self, name): + if name.startswith("supports") and hasattr(self.optim, name): + return getattr(self.optim, name) + raise AttributeError( + "'FairseqOSS' object has no attribute {0!r}".format(name) + ) + + def broadcast_global_state_dict( + self, state_dict: Dict[str, Any] + ) -> Dict[str, Any]: + """ + Broadcasts the entire state_dict to all other ranks + each rank is responsible to load their own partition of data + """ + return utils.broadcast_object( + state_dict, + src_rank=0, + group=self.group, + ) + + torch_optimizer = optimizer.optimizer + optim_cls = type(torch_optimizer) + + optimizer.optimizer = FairseqOSS( + torch_optimizer.param_groups, + optim_cls, + group=group, + **optimizer.optimizer_config + ) diff --git a/fairseq/options.py b/fairseq/options.py new file mode 100644 index 0000000000000000000000000000000000000000..2d9f8381a71415ebe9c14c13068abadf289a03e4 --- /dev/null +++ b/fairseq/options.py @@ -0,0 +1,381 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +from pathlib import Path +from typing import Callable, List, Optional, Union + +import torch +from fairseq import utils +from fairseq.data.indexed_dataset import get_available_dataset_impl +from fairseq.dataclass.configs import ( + CheckpointConfig, + CommonConfig, + CommonEvalConfig, + DatasetConfig, + DistributedTrainingConfig, + EvalLMConfig, + GenerationConfig, + InteractiveConfig, + OptimizationConfig, +) +from fairseq.dataclass.utils import gen_parser_from_dataclass + +# this import is for backward compatibility +from fairseq.utils import csv_str_list, eval_bool, eval_str_dict, eval_str_list # noqa + + +def get_preprocessing_parser(default_task="translation"): + parser = get_parser("Preprocessing", default_task) + add_preprocess_args(parser) + return parser + + +def get_training_parser(default_task="translation"): + parser = get_parser("Trainer", default_task) + add_dataset_args(parser, train=True) + add_distributed_training_args(parser) + add_model_args(parser) + add_optimization_args(parser) + add_checkpoint_args(parser) + return parser + + +def get_generation_parser(interactive=False, default_task="translation"): + parser = get_parser("Generation", default_task) + add_dataset_args(parser, gen=True) + add_distributed_training_args(parser, default_world_size=1) + add_generation_args(parser) + add_checkpoint_args(parser) + if interactive: + add_interactive_args(parser) + return parser + + +def get_interactive_generation_parser(default_task="translation"): + return get_generation_parser(interactive=True, default_task=default_task) + + +def get_eval_lm_parser(default_task="language_modeling"): + parser = get_parser("Evaluate Language Model", default_task) + add_dataset_args(parser, gen=True) + add_distributed_training_args(parser, default_world_size=1) + add_eval_lm_args(parser) + return parser + + +def get_validation_parser(default_task=None): + parser = get_parser("Validation", default_task) + add_dataset_args(parser, train=True) + add_distributed_training_args(parser, default_world_size=1) + group = parser.add_argument_group("Evaluation") + gen_parser_from_dataclass(group, CommonEvalConfig()) + return parser + + +def parse_args_and_arch( + parser: argparse.ArgumentParser, + input_args: List[str] = None, + parse_known: bool = False, + suppress_defaults: bool = False, + modify_parser: Optional[Callable[[argparse.ArgumentParser], None]] = None, +): + """ + Args: + parser (ArgumentParser): the parser + input_args (List[str]): strings to parse, defaults to sys.argv + parse_known (bool): only parse known arguments, similar to + `ArgumentParser.parse_known_args` + suppress_defaults (bool): parse while ignoring all default values + modify_parser (Optional[Callable[[ArgumentParser], None]]): + function to modify the parser, e.g., to set default values + """ + if suppress_defaults: + # Parse args without any default values. This requires us to parse + # twice, once to identify all the necessary task/model args, and a second + # time with all defaults set to None. + args = parse_args_and_arch( + parser, + input_args=input_args, + parse_known=parse_known, + suppress_defaults=False, + ) + suppressed_parser = argparse.ArgumentParser(add_help=False, parents=[parser]) + suppressed_parser.set_defaults(**{k: None for k, v in vars(args).items()}) + args = suppressed_parser.parse_args(input_args) + return argparse.Namespace( + **{k: v for k, v in vars(args).items() if v is not None} + ) + + from fairseq.models import ARCH_MODEL_REGISTRY, ARCH_CONFIG_REGISTRY, MODEL_REGISTRY + + # Before creating the true parser, we need to import optional user module + # in order to eagerly import custom tasks, optimizers, architectures, etc. + usr_parser = argparse.ArgumentParser(add_help=False, allow_abbrev=False) + usr_parser.add_argument("--user-dir", default=None) + usr_args, _ = usr_parser.parse_known_args(input_args) + utils.import_user_module(usr_args) + + if modify_parser is not None: + modify_parser(parser) + + # The parser doesn't know about model/criterion/optimizer-specific args, so + # we parse twice. First we parse the model/criterion/optimizer, then we + # parse a second time after adding the *-specific arguments. + # If input_args is given, we will parse those args instead of sys.argv. + args, _ = parser.parse_known_args(input_args) + + # Add model-specific args to parser. + if hasattr(args, "arch"): + model_specific_group = parser.add_argument_group( + "Model-specific configuration", + # Only include attributes which are explicitly given as command-line + # arguments or which have default values. + argument_default=argparse.SUPPRESS, + ) + if args.arch in ARCH_MODEL_REGISTRY: + ARCH_MODEL_REGISTRY[args.arch].add_args(model_specific_group) + elif args.arch in MODEL_REGISTRY: + MODEL_REGISTRY[args.arch].add_args(model_specific_group) + else: + raise RuntimeError() + + if hasattr(args, "task"): + from fairseq.tasks import TASK_REGISTRY + + TASK_REGISTRY[args.task].add_args(parser) + if getattr(args, "use_bmuf", False): + # hack to support extra args for block distributed data parallelism + from fairseq.optim.bmuf import FairseqBMUF + + FairseqBMUF.add_args(parser) + + # Add *-specific args to parser. + from fairseq.registry import REGISTRIES + + for registry_name, REGISTRY in REGISTRIES.items(): + choice = getattr(args, registry_name, None) + if choice is not None: + cls = REGISTRY["registry"][choice] + if hasattr(cls, "add_args"): + cls.add_args(parser) + elif hasattr(cls, "__dataclass"): + gen_parser_from_dataclass(parser, cls.__dataclass()) + + # Modify the parser a second time, since defaults may have been reset + if modify_parser is not None: + modify_parser(parser) + + # Parse a second time. + if parse_known: + args, extra = parser.parse_known_args(input_args) + else: + args = parser.parse_args(input_args) + extra = None + # Post-process args. + if ( + hasattr(args, "batch_size_valid") and args.batch_size_valid is None + ) or not hasattr(args, "batch_size_valid"): + args.batch_size_valid = args.batch_size + if hasattr(args, "max_tokens_valid") and args.max_tokens_valid is None: + args.max_tokens_valid = args.max_tokens + if getattr(args, "memory_efficient_fp16", False): + args.fp16 = True + if getattr(args, "memory_efficient_bf16", False): + args.bf16 = True + args.tpu = getattr(args, "tpu", False) + args.bf16 = getattr(args, "bf16", False) + if args.bf16: + args.tpu = True + if args.tpu and args.fp16: + raise ValueError("Cannot combine --fp16 and --tpu, use --bf16 on TPUs") + + if getattr(args, "seed", None) is None: + args.seed = 1 # default seed for training + args.no_seed_provided = True + else: + args.no_seed_provided = False + + # Apply architecture configuration. + if hasattr(args, "arch") and args.arch in ARCH_CONFIG_REGISTRY: + ARCH_CONFIG_REGISTRY[args.arch](args) + + if parse_known: + return args, extra + else: + return args + + +def get_parser(desc, default_task="translation"): + # Before creating the true parser, we need to import optional user module + # in order to eagerly import custom tasks, optimizers, architectures, etc. + usr_parser = argparse.ArgumentParser(add_help=False, allow_abbrev=False) + usr_parser.add_argument("--user-dir", default=None) + usr_args, _ = usr_parser.parse_known_args() + utils.import_user_module(usr_args) + + parser = argparse.ArgumentParser(allow_abbrev=False) + gen_parser_from_dataclass(parser, CommonConfig()) + + from fairseq.registry import REGISTRIES + + for registry_name, REGISTRY in REGISTRIES.items(): + parser.add_argument( + "--" + registry_name.replace("_", "-"), + default=REGISTRY["default"], + choices=REGISTRY["registry"].keys(), + ) + + # Task definitions can be found under fairseq/tasks/ + from fairseq.tasks import TASK_REGISTRY + + parser.add_argument( + "--task", + metavar="TASK", + default=default_task, + choices=TASK_REGISTRY.keys(), + help="task", + ) + # fmt: on + return parser + + +def add_preprocess_args(parser): + group = parser.add_argument_group("Preprocessing") + # fmt: off + group.add_argument("-s", "--source-lang", default=None, metavar="SRC", + help="source language") + group.add_argument("-t", "--target-lang", default=None, metavar="TARGET", + help="target language") + group.add_argument("--trainpref", metavar="FP", default=None, + help="train file prefix (also used to build dictionaries)") + group.add_argument("--validpref", metavar="FP", default=None, + help="comma separated, valid file prefixes " + "(words missing from train set are replaced with <unk>)") + group.add_argument("--testpref", metavar="FP", default=None, + help="comma separated, test file prefixes " + "(words missing from train set are replaced with <unk>)") + group.add_argument("--align-suffix", metavar="FP", default=None, + help="alignment file suffix") + group.add_argument("--destdir", metavar="DIR", default="data-bin", + help="destination dir") + group.add_argument("--thresholdtgt", metavar="N", default=0, type=int, + help="map words appearing less than threshold times to unknown") + group.add_argument("--thresholdsrc", metavar="N", default=0, type=int, + help="map words appearing less than threshold times to unknown") + group.add_argument("--tgtdict", metavar="FP", + help="reuse given target dictionary") + group.add_argument("--srcdict", metavar="FP", + help="reuse given source dictionary") + group.add_argument("--nwordstgt", metavar="N", default=-1, type=int, + help="number of target words to retain") + group.add_argument("--nwordssrc", metavar="N", default=-1, type=int, + help="number of source words to retain") + group.add_argument("--alignfile", metavar="ALIGN", default=None, + help="an alignment file (optional)") + parser.add_argument('--dataset-impl', metavar='FORMAT', default='mmap', + choices=get_available_dataset_impl(), + help='output dataset implementation') + group.add_argument("--joined-dictionary", action="store_true", + help="Generate joined dictionary") + group.add_argument("--only-source", action="store_true", + help="Only process the source language") + group.add_argument("--padding-factor", metavar="N", default=8, type=int, + help="Pad dictionary size to be multiple of N") + group.add_argument("--workers", metavar="N", default=1, type=int, + help="number of parallel workers") + group.add_argument("--dict-only", action='store_true', + help="if true, only builds a dictionary and then exits") + # fmt: on + return parser + + +def add_dataset_args(parser, train=False, gen=False): + group = parser.add_argument_group("dataset_data_loading") + gen_parser_from_dataclass(group, DatasetConfig()) + # fmt: on + return group + + +def add_distributed_training_args(parser, default_world_size=None): + group = parser.add_argument_group("distributed_training") + if default_world_size is None: + default_world_size = max(1, torch.cuda.device_count()) + gen_parser_from_dataclass( + group, DistributedTrainingConfig(distributed_world_size=default_world_size) + ) + return group + + +def add_optimization_args(parser): + group = parser.add_argument_group("optimization") + # fmt: off + gen_parser_from_dataclass(group, OptimizationConfig()) + # fmt: on + return group + + +def add_checkpoint_args(parser): + group = parser.add_argument_group("checkpoint") + # fmt: off + gen_parser_from_dataclass(group, CheckpointConfig()) + # fmt: on + return group + + +def add_common_eval_args(group): + gen_parser_from_dataclass(group, CommonEvalConfig()) + + +def add_eval_lm_args(parser): + group = parser.add_argument_group("LM Evaluation") + add_common_eval_args(group) + gen_parser_from_dataclass(group, EvalLMConfig()) + + +def add_generation_args(parser): + group = parser.add_argument_group("Generation") + add_common_eval_args(group) + gen_parser_from_dataclass(group, GenerationConfig()) + return group + + +def add_interactive_args(parser): + group = parser.add_argument_group("Interactive") + gen_parser_from_dataclass(group, InteractiveConfig()) + + +def add_model_args(parser): + group = parser.add_argument_group("Model configuration") + # fmt: off + + # Model definitions can be found under fairseq/models/ + # + # The model architecture can be specified in several ways. + # In increasing order of priority: + # 1) model defaults (lowest priority) + # 2) --arch argument + # 3) --encoder/decoder-* arguments (highest priority) + from fairseq.models import ARCH_MODEL_REGISTRY + group.add_argument('--arch', '-a', metavar='ARCH', + choices=ARCH_MODEL_REGISTRY.keys(), + help='model architecture') + # fmt: on + return group + + +def get_args( + data: Union[str, Path], + task: str = "translation", + arch: str = "transformer", + **overrides +): + parser = get_training_parser(task) + args = parse_args_and_arch(parser, [str(data), "--task", task, "--arch", arch]) + + for k, v in overrides.items(): + setattr(args, k, v) + + return args diff --git a/fairseq/pdb.py b/fairseq/pdb.py new file mode 100644 index 0000000000000000000000000000000000000000..1ba6ef0d336b30717cfdde94e1b838cfe2bfeb20 --- /dev/null +++ b/fairseq/pdb.py @@ -0,0 +1,47 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import multiprocessing +import os +import pdb +import sys + + +__all__ = ["set_trace"] + + +_stdin = [None] +_stdin_lock = multiprocessing.Lock() +try: + _stdin_fd = sys.stdin.fileno() +except Exception: + _stdin_fd = None + + +class MultiprocessingPdb(pdb.Pdb): + """A Pdb wrapper that works in a multiprocessing environment. + + Usage: `from fairseq import pdb; pdb.set_trace()` + """ + + def __init__(self): + pdb.Pdb.__init__(self, nosigint=True) + + def _cmdloop(self): + stdin_bak = sys.stdin + with _stdin_lock: + try: + if _stdin_fd is not None: + if not _stdin[0]: + _stdin[0] = os.fdopen(_stdin_fd) + sys.stdin = _stdin[0] + self.cmdloop() + finally: + sys.stdin = stdin_bak + + +def set_trace(): + pdb = MultiprocessingPdb() + pdb.set_trace(sys._getframe().f_back) diff --git a/fairseq/quantization_utils.py b/fairseq/quantization_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..11fc414c852b199b80a569bf024272535929abcc --- /dev/null +++ b/fairseq/quantization_utils.py @@ -0,0 +1,143 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging + +from fairseq.modules.quantization import pq, quantization_options, scalar +from omegaconf import DictConfig + + +logger = logging.getLogger(__name__) + + +def quantize_model_scalar(model, model_cfg: DictConfig): + quant_noise_scalar = getattr(model_cfg, "quant_noise_scalar", 0) or 0 + if quant_noise_scalar > 0: + # quantize_model edits the model in place + scalar.quantize_model_(model, p=quant_noise_scalar, bits=8, update_step=1000) + return model + + +class Quantizer(object): + def __init__(self, config_path, max_epoch, max_update): + try: + import yaml + except ImportError: + raise ImportError("Please install yaml with: pip install yaml") + + # parse config + if config_path: + with open(config_path) as config_file: + config = quantization_options.parse_config_yaml( + yaml.safe_load(config_file) + ) + else: + config = quantization_options.parse_config_yaml({}) + + self.n_centroids_config = config["n_centroids"] + self.block_sizes_config = config["block_sizes"] + self.layers_to_quantize = config["layers_to_quantize"] + + # We assume that training will run for a fixed number of epochs + # (or updates) and that we should train for equal durations + # between iterations of PQ. + num_iterations = len(self.layers_to_quantize) + if max_epoch > 0: + assert max_epoch % num_iterations == 0, ( + "for iterative PQ, --max-epoch (={}) must be evenly divisible by " + "len(layers_to_quantize) (={})".format(max_epoch, num_iterations) + ) + self.epoch_schedule = max_epoch // num_iterations + else: + self.epoch_schedule = None + if max_update > 0: + assert max_update % num_iterations == 0, ( + "for iterative PQ, --max-update (={}) must be evenly divisible by " + "len(layers_to_quantize) (={})".format(max_update, num_iterations) + ) + self.update_schedule = max_update // num_iterations + else: + self.update_schedule = None + assert (self.epoch_schedule is not None) ^ ( + self.update_schedule is not None + ), "for iterative PQ, cannot specify both --max-update and --max-epoch" + + # 0 is a special value for quantization step, which will force + # the first call to begin_epoch() to call step() + self.quantization_step = 0 + + def set_trainer(self, trainer): + self.trainer = trainer + self.size_tracker = pq.SizeTracker(self.trainer.get_model()) + + def step(self): + """Move to the next stage of quantization.""" + if self.quantization_step >= len(self.layers_to_quantize): + # Maybe we just finished the last training step or we loaded + # a checkpoint for an iterative PQ model which previously + # finished training. Either way, don't quantize again. + return + + logger.info( + "quantizing model (step={}; layers_to_quantize[step]={})".format( + self.quantization_step, self.layers_to_quantize[self.quantization_step] + ) + ) + quantized_layers = pq.quantize_model_( + self.trainer.get_model(), + self.size_tracker, + self.layers_to_quantize, + self.block_sizes_config, + self.n_centroids_config, + step=self.quantization_step, + ) + logger.info("quantized layers: {}".format(quantized_layers)) + logger.info(self.size_tracker) + + self.quantization_step += 1 + + # reintialize the Trainer since model parameters have changed + self.trainer.reinitialize() + + def begin_epoch(self, epoch): + """Called at the beginning of each epoch (epochs start at 1).""" + if ( + ( + self.epoch_schedule is not None + and epoch > 0 + and (epoch - 1) % self.epoch_schedule == 0 + ) + # we always step once in the beginning, even if using + # update-based quantization + or self.quantization_step == 0 + ): + self.step() + + def step_update(self, num_updates): + """Called at the end of each step.""" + if ( + self.update_schedule is not None + and num_updates > 0 + and num_updates % self.update_schedule == 0 + ): + self.step() + + def state_dict(self): + return { + "n_centroids_config": self.n_centroids_config, + "block_sizes_config": self.block_sizes_config, + "layers_to_quantize": self.layers_to_quantize, + "epoch_schedule": self.epoch_schedule, + "update_schedule": self.update_schedule, + "quantization_step": self.quantization_step, + } + + def load_state_dict(self, state_dict): + self.n_centroids_config = state_dict["n_centroids_config"] + self.block_sizes_config = state_dict["block_sizes_config"] + self.layers_to_quantize = state_dict["layers_to_quantize"] + self.epoch_schedule = state_dict["epoch_schedule"] + self.update_schedule = state_dict["update_schedule"] + self.quantization_step = state_dict["quantization_step"] diff --git a/fairseq/registry.py b/fairseq/registry.py new file mode 100644 index 0000000000000000000000000000000000000000..3fbaeac301855d41a5d52ff58276787e8ddebfca --- /dev/null +++ b/fairseq/registry.py @@ -0,0 +1,100 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from argparse import Namespace + +from typing import Union +from fairseq.dataclass import FairseqDataclass +from fairseq.dataclass.utils import populate_dataclass, merge_with_parent +from hydra.core.config_store import ConfigStore +from omegaconf import DictConfig + +REGISTRIES = {} + + +def setup_registry(registry_name: str, base_class=None, default=None, required=False): + assert registry_name.startswith("--") + registry_name = registry_name[2:].replace("-", "_") + + REGISTRY = {} + REGISTRY_CLASS_NAMES = set() + DATACLASS_REGISTRY = {} + + # maintain a registry of all registries + if registry_name in REGISTRIES: + return # registry already exists + REGISTRIES[registry_name] = { + "registry": REGISTRY, + "default": default, + "dataclass_registry": DATACLASS_REGISTRY, + } + + def build_x(cfg: Union[DictConfig, str, Namespace], *extra_args, **extra_kwargs): + if isinstance(cfg, DictConfig): + choice = cfg._name + + if choice and choice in DATACLASS_REGISTRY: + dc = DATACLASS_REGISTRY[choice] + cfg = merge_with_parent(dc(), cfg) + elif isinstance(cfg, str): + choice = cfg + if choice in DATACLASS_REGISTRY: + cfg = DATACLASS_REGISTRY[choice]() + else: + choice = getattr(cfg, registry_name, None) + if choice in DATACLASS_REGISTRY: + cfg = populate_dataclass(DATACLASS_REGISTRY[choice](), cfg) + + if choice is None: + if required: + raise ValueError("{} is required!".format(registry_name)) + return None + + cls = REGISTRY[choice] + if hasattr(cls, "build_" + registry_name): + builder = getattr(cls, "build_" + registry_name) + else: + builder = cls + + return builder(cfg, *extra_args, **extra_kwargs) + + def register_x(name, dataclass=None): + def register_x_cls(cls): + if name in REGISTRY: + raise ValueError( + "Cannot register duplicate {} ({})".format(registry_name, name) + ) + if cls.__name__ in REGISTRY_CLASS_NAMES: + raise ValueError( + "Cannot register {} with duplicate class name ({})".format( + registry_name, cls.__name__ + ) + ) + if base_class is not None and not issubclass(cls, base_class): + raise ValueError( + "{} must extend {}".format(cls.__name__, base_class.__name__) + ) + + if dataclass is not None and not issubclass(dataclass, FairseqDataclass): + raise ValueError( + "Dataclass {} must extend FairseqDataclass".format(dataclass) + ) + + cls.__dataclass = dataclass + if cls.__dataclass is not None: + DATACLASS_REGISTRY[name] = cls.__dataclass + + cs = ConfigStore.instance() + node = dataclass() + node._name = name + cs.store(name=name, group=registry_name, node=node, provider="fairseq") + + REGISTRY[name] = cls + + return cls + + return register_x_cls + + return build_x, register_x, REGISTRY, DATACLASS_REGISTRY diff --git a/fairseq/scoring/__init__.py b/fairseq/scoring/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..58f2f563e493327394dff1265030d18f0814b5a2 --- /dev/null +++ b/fairseq/scoring/__init__.py @@ -0,0 +1,55 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +import importlib +import os +from abc import ABC, abstractmethod + +from fairseq import registry +from omegaconf import DictConfig + + +class BaseScorer(ABC): + def __init__(self, cfg): + self.cfg = cfg + self.ref = [] + self.pred = [] + + def add_string(self, ref, pred): + self.ref.append(ref) + self.pred.append(pred) + + @abstractmethod + def score(self) -> float: + pass + + @abstractmethod + def result_string(self) -> str: + pass + + +_build_scorer, register_scorer, SCORER_REGISTRY, _ = registry.setup_registry( + "--scoring", default="bleu" +) + + +def build_scorer(choice, tgt_dict): + _choice = choice._name if isinstance(choice, DictConfig) else choice + + if _choice == "bleu": + from fairseq.scoring import bleu + + return bleu.Scorer( + bleu.BleuConfig(pad=tgt_dict.pad(), eos=tgt_dict.eos(), unk=tgt_dict.unk()) + ) + return _build_scorer(choice) + + +# automatically import any Python files in the current directory +for file in sorted(os.listdir(os.path.dirname(__file__))): + if file.endswith(".py") and not file.startswith("_"): + module = file[: file.find(".py")] + importlib.import_module("fairseq.scoring." + module) diff --git a/fairseq/scoring/bleu.py b/fairseq/scoring/bleu.py new file mode 100644 index 0000000000000000000000000000000000000000..97de5f966ec08e5a304c41358e67755c601622b7 --- /dev/null +++ b/fairseq/scoring/bleu.py @@ -0,0 +1,167 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import ctypes +import math +import sys +from dataclasses import dataclass, field + +import torch +from fairseq.dataclass import FairseqDataclass +from fairseq.scoring import BaseScorer, register_scorer +from fairseq.scoring.tokenizer import EvaluationTokenizer + + +class BleuStat(ctypes.Structure): + _fields_ = [ + ("reflen", ctypes.c_size_t), + ("predlen", ctypes.c_size_t), + ("match1", ctypes.c_size_t), + ("count1", ctypes.c_size_t), + ("match2", ctypes.c_size_t), + ("count2", ctypes.c_size_t), + ("match3", ctypes.c_size_t), + ("count3", ctypes.c_size_t), + ("match4", ctypes.c_size_t), + ("count4", ctypes.c_size_t), + ] + + +@dataclass +class SacrebleuConfig(FairseqDataclass): + sacrebleu_tokenizer: EvaluationTokenizer.ALL_TOKENIZER_TYPES = field( + default="13a", metadata={"help": "tokenizer"} + ) + sacrebleu_lowercase: bool = field( + default=False, metadata={"help": "apply lowercasing"} + ) + sacrebleu_char_level: bool = field( + default=False, metadata={"help": "evaluate at character level"} + ) + + +@register_scorer("sacrebleu", dataclass=SacrebleuConfig) +class SacrebleuScorer(BaseScorer): + def __init__(self, cfg): + super(SacrebleuScorer, self).__init__(cfg) + import sacrebleu + + self.sacrebleu = sacrebleu + self.tokenizer = EvaluationTokenizer( + tokenizer_type=cfg.sacrebleu_tokenizer, + lowercase=cfg.sacrebleu_lowercase, + character_tokenization=cfg.sacrebleu_char_level, + ) + + def add_string(self, ref, pred): + self.ref.append(self.tokenizer.tokenize(ref)) + self.pred.append(self.tokenizer.tokenize(pred)) + + def score(self, order=4): + return self.result_string(order).score + + def result_string(self, order=4): + if order != 4: + raise NotImplementedError + # tokenization and lowercasing are performed by self.tokenizer instead. + return self.sacrebleu.corpus_bleu( + self.pred, [self.ref], tokenize="none" + ).format() + + +@dataclass +class BleuConfig(FairseqDataclass): + pad: int = field(default=1, metadata={"help": "padding index"}) + eos: int = field(default=2, metadata={"help": "eos index"}) + unk: int = field(default=3, metadata={"help": "unk index"}) + + +@register_scorer("bleu", dataclass=BleuConfig) +class Scorer(object): + def __init__(self, cfg): + self.stat = BleuStat() + self.pad = cfg.pad + self.eos = cfg.eos + self.unk = cfg.unk + + try: + from fairseq import libbleu + except ImportError as e: + sys.stderr.write( + "ERROR: missing libbleu.so. run `pip install --editable .`\n" + ) + raise e + + self.C = ctypes.cdll.LoadLibrary(libbleu.__file__) + + self.reset() + + def reset(self, one_init=False): + if one_init: + self.C.bleu_one_init(ctypes.byref(self.stat)) + else: + self.C.bleu_zero_init(ctypes.byref(self.stat)) + + def add(self, ref, pred): + if not isinstance(ref, torch.IntTensor): + raise TypeError("ref must be a torch.IntTensor (got {})".format(type(ref))) + if not isinstance(pred, torch.IntTensor): + raise TypeError("pred must be a torch.IntTensor(got {})".format(type(pred))) + + # don't match unknown words + rref = ref.clone() + assert not rref.lt(0).any() + rref[rref.eq(self.unk)] = -999 + + rref = rref.contiguous().view(-1) + pred = pred.contiguous().view(-1) + + self.C.bleu_add( + ctypes.byref(self.stat), + ctypes.c_size_t(rref.size(0)), + ctypes.c_void_p(rref.data_ptr()), + ctypes.c_size_t(pred.size(0)), + ctypes.c_void_p(pred.data_ptr()), + ctypes.c_int(self.pad), + ctypes.c_int(self.eos), + ) + + def score(self, order=4): + psum = sum( + math.log(p) if p > 0 else float("-Inf") for p in self.precision()[:order] + ) + return self.brevity() * math.exp(psum / order) * 100 + + def precision(self): + def ratio(a, b): + return a / b if b > 0 else 0 + + return [ + ratio(self.stat.match1, self.stat.count1), + ratio(self.stat.match2, self.stat.count2), + ratio(self.stat.match3, self.stat.count3), + ratio(self.stat.match4, self.stat.count4), + ] + + def brevity(self): + r = self.stat.reflen / self.stat.predlen + return min(1, math.exp(1 - r)) + + def result_string(self, order=4): + assert order <= 4, "BLEU scores for order > 4 aren't supported" + fmt = "BLEU{} = {:2.2f}, {:2.1f}" + for _ in range(1, order): + fmt += "/{:2.1f}" + fmt += " (BP={:.3f}, ratio={:.3f}, syslen={}, reflen={})" + bleup = [p * 100 for p in self.precision()[:order]] + return fmt.format( + order, + self.score(order=order), + *bleup, + self.brevity(), + self.stat.predlen / self.stat.reflen, + self.stat.predlen, + self.stat.reflen + ) diff --git a/fairseq/scoring/chrf.py b/fairseq/scoring/chrf.py new file mode 100644 index 0000000000000000000000000000000000000000..0d6cb77383a44d9ac739958b79a30764f1fbf7f3 --- /dev/null +++ b/fairseq/scoring/chrf.py @@ -0,0 +1,27 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq.scoring import BaseScorer, register_scorer + + +@register_scorer("chrf") +class ChrFScorer(BaseScorer): + def __init__(self, args): + super(ChrFScorer, self).__init__(args) + import sacrebleu + + self.sacrebleu = sacrebleu + + def add_string(self, ref, pred): + self.ref.append(ref) + self.pred.append(pred) + + def score(self, order=4): + return self.result_string(order).score + + def result_string(self, order=4): + if order != 4: + raise NotImplementedError + return self.sacrebleu.corpus_chrf(self.pred, [self.ref]).format() diff --git a/fairseq/scoring/tokenizer.py b/fairseq/scoring/tokenizer.py new file mode 100644 index 0000000000000000000000000000000000000000..61cf6d4a7cc698258caad9f68f2e8559dd510eee --- /dev/null +++ b/fairseq/scoring/tokenizer.py @@ -0,0 +1,67 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import unicodedata + +from fairseq.dataclass import ChoiceEnum + + +class EvaluationTokenizer(object): + """A generic evaluation-time tokenizer, which leverages built-in tokenizers + in sacreBLEU (https://github.com/mjpost/sacrebleu). It additionally provides + lowercasing, punctuation removal and character tokenization, which are + applied after sacreBLEU tokenization. + + Args: + tokenizer_type (str): the type of sacreBLEU tokenizer to apply. + lowercase (bool): lowercase the text. + punctuation_removal (bool): remove punctuation (based on unicode + category) from text. + character_tokenization (bool): tokenize the text to characters. + """ + + SPACE = chr(32) + SPACE_ESCAPE = chr(9601) + ALL_TOKENIZER_TYPES = ChoiceEnum(["none", "13a", "intl", "zh", "ja-mecab"]) + + def __init__( + self, + tokenizer_type: str = "13a", + lowercase: bool = False, + punctuation_removal: bool = False, + character_tokenization: bool = False, + ): + from sacrebleu.tokenizers import TOKENIZERS + + assert tokenizer_type in TOKENIZERS, f"{tokenizer_type}, {TOKENIZERS}" + self.lowercase = lowercase + self.punctuation_removal = punctuation_removal + self.character_tokenization = character_tokenization + self.tokenizer = TOKENIZERS[tokenizer_type] + + @classmethod + def remove_punctuation(cls, sent: str): + """Remove punctuation based on Unicode category.""" + return cls.SPACE.join( + t + for t in sent.split(cls.SPACE) + if not all(unicodedata.category(c)[0] == "P" for c in t) + ) + + def tokenize(self, sent: str): + tokenized = self.tokenizer()(sent) + + if self.punctuation_removal: + tokenized = self.remove_punctuation(tokenized) + + if self.character_tokenization: + tokenized = self.SPACE.join( + list(tokenized.replace(self.SPACE, self.SPACE_ESCAPE)) + ) + + if self.lowercase: + tokenized = tokenized.lower() + + return tokenized diff --git a/fairseq/scoring/wer.py b/fairseq/scoring/wer.py new file mode 100644 index 0000000000000000000000000000000000000000..633dc47c247691c4c9e36cbdbab7d7cb74b38452 --- /dev/null +++ b/fairseq/scoring/wer.py @@ -0,0 +1,58 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from dataclasses import dataclass, field + +from fairseq.dataclass import FairseqDataclass +from fairseq.scoring import BaseScorer, register_scorer +from fairseq.scoring.tokenizer import EvaluationTokenizer + + +@dataclass +class WerScorerConfig(FairseqDataclass): + wer_tokenizer: EvaluationTokenizer.ALL_TOKENIZER_TYPES = field( + default="none", metadata={"help": "sacreBLEU tokenizer to use for evaluation"} + ) + wer_remove_punct: bool = field( + default=False, metadata={"help": "remove punctuation"} + ) + wer_char_level: bool = field( + default=False, metadata={"help": "evaluate at character level"} + ) + wer_lowercase: bool = field(default=False, metadata={"help": "lowercasing"}) + + +@register_scorer("wer", dataclass=WerScorerConfig) +class WerScorer(BaseScorer): + def __init__(self, cfg): + super().__init__(cfg) + self.reset() + try: + import editdistance as ed + except ImportError: + raise ImportError("Please install editdistance to use WER scorer") + self.ed = ed + self.tokenizer = EvaluationTokenizer( + tokenizer_type=self.cfg.wer_tokenizer, + lowercase=self.cfg.wer_lowercase, + punctuation_removal=self.cfg.wer_remove_punct, + character_tokenization=self.cfg.wer_char_level, + ) + + def reset(self): + self.distance = 0 + self.ref_length = 0 + + def add_string(self, ref, pred): + ref_items = self.tokenizer.tokenize(ref).split() + pred_items = self.tokenizer.tokenize(pred).split() + self.distance += self.ed.eval(ref_items, pred_items) + self.ref_length += len(ref_items) + + def result_string(self): + return f"WER: {self.score():.2f}" + + def score(self): + return 100.0 * self.distance / self.ref_length if self.ref_length > 0 else 0 diff --git a/fairseq/search.py b/fairseq/search.py new file mode 100644 index 0000000000000000000000000000000000000000..d5ea68b4ce04409c504c1d22098b7968a9ce596a --- /dev/null +++ b/fairseq/search.py @@ -0,0 +1,814 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +from typing import List, Optional + +import torch +import torch.nn as nn +from fairseq.token_generation_constraints import ( + ConstraintState, + OrderedConstraintState, + UnorderedConstraintState, +) +from torch import Tensor + + +class Search(nn.Module): + def __init__(self, tgt_dict): + super().__init__() + self.pad = tgt_dict.pad() + self.unk = tgt_dict.unk() + self.eos = tgt_dict.eos() + self.vocab_size = len(tgt_dict) + self.src_lengths = torch.tensor(-1) + self.supports_constraints = False + self.stop_on_max_len = False + + def step( + self, step, lprobs, scores, prev_output_tokens=None, original_batch_idxs=None + ): + """Take a single search step. + + Args: + step: the current search step, starting at 0 + lprobs: (bsz x input_beam_size x vocab_size) + the model's log-probabilities over the vocabulary at the current step + scores: (bsz x input_beam_size x step) + the historical model scores of each hypothesis up to this point + prev_output_tokens: (bsz x step) + the previously generated oputput tokens + original_batch_idxs: (bsz) + the tensor with the batch indices, in the range [0, bsz) + this is useful in case there has been applied a re-ordering + and we need to know the orignal indices + + Return: A tuple of (scores, indices, beams) where: + scores: (bsz x output_beam_size) + the scores of the chosen elements; output_beam_size can be + larger than input_beam_size, e.g., we may return + 2*input_beam_size to account for EOS + indices: (bsz x output_beam_size) + the indices of the chosen elements + beams: (bsz x output_beam_size) + the hypothesis ids of the chosen elements, in the range [0, input_beam_size) + """ + raise NotImplementedError + + @torch.jit.export + def set_src_lengths(self, src_lengths): + self.src_lengths = src_lengths + + @torch.jit.export + def init_constraints(self, batch_constraints: Optional[Tensor], beam_size: int): + """Initialize constraint states for constrained decoding (if supported). + + Args: + batch_constraints: (torch.Tensor, optional) + the list of constraints, in packed form + beam_size: (int) + the beam size + Returns: + *encoder_out* rearranged according to *new_order* + """ + pass + + def prune_sentences(self, batch_idxs: Tensor): + """ + Removes constraint states for completed sentences (if supported). + This is called from sequence_generator._generate() when sentences are + deleted from the batch. + + Args: + batch_idxs: Indices of *sentences* whose constraint state should be *kept*. + """ + pass + + def update_constraints(self, active_hypos: Tensor): + """ + Updates the constraint states by selecting the beam items that are retained. + This is called at each time step of sequence_generator._generate() when + the set of 2 * {beam_size} candidate hypotheses are reduced to the beam size. + + Args: + active_hypos: (batch size, beam size) + list of integers denoting, for each sentence, which beam candidate items + should be kept. + """ + pass + + +class BeamSearch(Search): + def __init__(self, tgt_dict): + super().__init__(tgt_dict) + self.constraint_states = None + + @torch.jit.export + def step( + self, + step: int, + lprobs, + scores: Optional[Tensor], + prev_output_tokens: Optional[Tensor] = None, + original_batch_idxs: Optional[Tensor] = None, + ): + bsz, beam_size, vocab_size = lprobs.size() + + if step == 0: + # at the first step all hypotheses are equally likely, so use + # only the first beam + lprobs = lprobs[:, ::beam_size, :].contiguous() + else: + # make probs contain cumulative scores for each hypothesis + assert scores is not None + lprobs = lprobs + scores[:, :, step - 1].unsqueeze(-1) + + top_prediction = torch.topk( + lprobs.view(bsz, -1), + k=min( + # Take the best 2 x beam_size predictions. We'll choose the first + # beam_size of these which don't predict eos to continue with. + beam_size * 2, + lprobs.view(bsz, -1).size(1) - 1, # -1 so we never select pad + ), + ) + scores_buf = top_prediction[0] + indices_buf = top_prediction[1] + # Project back into relative indices and beams + beams_buf = indices_buf // vocab_size + indices_buf = indices_buf.fmod(vocab_size) + + # At this point, beams_buf and indices_buf are single-dim and contain relative indices + return scores_buf, indices_buf, beams_buf + + +class PrefixConstrainedBeamSearch(Search): + def __init__(self, tgt_dict, prefix_allowed_tokens_fn): + super().__init__(tgt_dict) + self.prefix_allowed_tokens_fn = prefix_allowed_tokens_fn + self.stop_on_max_len = True + + @torch.jit.export + def apply_mask(self, x, prev_output_tokens, original_batch_idxs): + beam_size = x.shape[0] // original_batch_idxs.shape[0] + original_batch_idxs = ( + original_batch_idxs.unsqueeze(-1).repeat((1, beam_size)).flatten().tolist() + ) + + mask = torch.full_like(x, -math.inf) + for sent_i, (sent, batch_i) in enumerate( + zip(prev_output_tokens, original_batch_idxs) + ): + mask[sent_i, :, self.prefix_allowed_tokens_fn(batch_i, sent)] = 0 + + return mask + + @torch.jit.export + def step( + self, + step: int, + lprobs: Tensor, + scores: Tensor, + prev_output_tokens: Tensor, + original_batch_idxs: Tensor, + ): + bsz, beam_size, vocab_size = lprobs.size() + + lprobs += self.apply_mask( + lprobs.view(bsz * beam_size, 1, vocab_size), + prev_output_tokens, + original_batch_idxs, + ).view(bsz, beam_size, vocab_size) + + if step == 0: + # at the first step all hypotheses are equally likely, so use + # only the first beam + lprobs = lprobs[:, ::beam_size, :].contiguous() + else: + # make probs contain cumulative scores for each hypothesis + assert scores is not None + lprobs = lprobs + scores[:, :, step - 1].unsqueeze(-1) + + top_prediction = torch.topk( + lprobs.view(bsz, -1), + k=min( + # Take the best beam_size predictions. We'll choose the first + # beam_size of these which don't predict eos to continue with. + beam_size, + lprobs.view(bsz, -1).size(1) - 1, # -1 so we never select pad + ), + ) + scores_buf = top_prediction[0] + indices_buf = top_prediction[1] + beams_buf = indices_buf // vocab_size + indices_buf = indices_buf.fmod(vocab_size) + return scores_buf, indices_buf, beams_buf + + +class LexicallyConstrainedBeamSearch(Search): + """Implements lexically constrained beam search as described in + + Fast Lexically Constrained Decoding with Dynamic Beam + Allocation for Neural Machine Translation. Post & Vilar, + NAACL 2018. https://www.aclweb.org/anthology/N18-1119/ + + and + + Improved Lexically Constrained Decoding for Translation and + Monolingual Rewriting. Hu et al, NAACL + 2019. https://www.aclweb.org/anthology/N19-1090/ + + This is accomplished by maintaining, for each beam hypothesis, a + ConstraintState object (see constraints.py) that tracks which + constraints have been generated and using this information to + shape the beam for each input sentence. + """ + + def __init__(self, tgt_dict, representation): + super().__init__(tgt_dict) + self.representation = representation + self.vocab_size = len(tgt_dict) + self.num_cands = 0 + self.supports_constraints = True + + @torch.jit.export + def init_constraints(self, batch_constraints: Optional[Tensor], beam_size: int): + self.constraint_states = [] + for constraint_tensor in batch_constraints: + if self.representation == "ordered": + constraint_state = OrderedConstraintState.create(constraint_tensor) + elif self.representation == "unordered": + constraint_state = UnorderedConstraintState.create(constraint_tensor) + + self.constraint_states.append([constraint_state for i in range(beam_size)]) + + @torch.jit.export + def prune_sentences(self, batch_idxs: Tensor): + self.constraint_states = [ + self.constraint_states[i] for i in batch_idxs.tolist() + ] + + @torch.jit.export + def update_constraints(self, active_hypos: Tensor): + if self.constraint_states: + batch_size = active_hypos.size(0) + for sentid in range(batch_size): + self.constraint_states[sentid] = [ + self.constraint_states[sentid][i] for i in active_hypos[sentid] + ] + + @torch.jit.export + def step( + self, + step: int, + lprobs: Tensor, + scores: Optional[Tensor], + prev_output_tokens: Optional[Tensor] = None, + original_batch_idxs: Optional[Tensor] = None, + ): + """ + A constrained step builds a large candidates list from the following: + - the top 2 * {beam_size} items over the whole beam + - for each item in the beam + - the top {each_k} (default 1) + - all next constraints + We then compute the constrained state of each beam item, and assign + stripe codes: 0 to the best in each bank, 1 to the 2nd-best, and so + on. We then sort by (stripe, score), and truncate the list at + 2 * beam size. + + Args: + step: the decoder step + lprobs: (batch size, beam size, target vocab) + the target-vocab distributions for each item in the beam. + Retrun: A tuple of (scores, indices, beams, constraints) where: + scores: (batch, output beam size) + the scores of the chosen elements + indices: (batch, output beam size) + the target vocab indices of the chosen elements + beams: (batch, output beam size) + the 0-indexed hypothesis ids of the chosen elements + constraints: (batch, output beam size) + the new constraint states + """ + each_k = 1 + device = lprobs.device + + batch_size, beam_size, vocab_size = lprobs.size() + + self.num_cands = min( + # Just take the k-best. We'll get another k from the 1-best from each + # row, plus more from the constraints + beam_size * 2, + lprobs.view(batch_size, -1).size(1) - 1, # -1 so we never select pad + ) + + # STEP 0: Preliminary. Prevent EOS for unfinished hyps across all batch items + constraint_states = self.constraint_states + if constraint_states and step > 0: + not_finished_indices = [] + for sentno, sent_constraints in enumerate(constraint_states): + for beamno, state in enumerate(sent_constraints): + index = sentno * beam_size + beamno + if not state.finished: + not_finished_indices.append(index) + not_finished_indices = torch.tensor(not_finished_indices) + if not_finished_indices.numel() > 0: + lprobs.view(batch_size * beam_size, -1)[ + not_finished_indices, self.eos + ] = -math.inf + + if step == 0: + # at the first step all hypotheses are equally likely, so use + # only the first beam entry for each batch item + lprobs = lprobs[:, ::beam_size, :].contiguous() + else: + # make probs contain cumulative scores for each hypothesis + assert scores is not None + lprobs = lprobs + scores[:, :, step - 1].unsqueeze(-1) + + top_prediction = torch.topk( + lprobs.view(batch_size, -1), + self.num_cands, + ) + scores_buf, indices_buf = top_prediction + # Project back into relative indices and beams + beams_buf = indices_buf // vocab_size + indices_buf = indices_buf.fmod(vocab_size) + + # Short circuit if there are no constraints in this batch + if not constraint_states: + return scores_buf, indices_buf, beams_buf + + # STEP 1: get top-1 from each hypothesis across all sentences in the batch + if step > 0: + top_scores, top_indices = torch.topk( + lprobs.view(batch_size * beam_size, -1), + k=each_k, + dim=1, + ) + top_scores = top_scores.view(batch_size, -1) + top_indices = top_indices.view(batch_size, -1) + scores_buf = torch.cat((scores_buf, top_scores), dim=1) + indices_buf = torch.cat((indices_buf, top_indices), dim=1) + new_beams = torch.arange(0, beam_size, device=device).repeat(batch_size, 1) + beams_buf = torch.cat((beams_buf, new_beams), dim=1) + + # Now, process sentences in the batch one by one. + new_scores_buf = torch.zeros((batch_size, 2 * beam_size), device=device) + new_indices_buf = torch.zeros((batch_size, 2 * beam_size), device=device).long() + new_beams_buf = torch.zeros((batch_size, 2 * beam_size), device=device).long() + for sentno, states in enumerate(constraint_states): + scores, indices, beams, new_states = self.step_sentence( + step, + sentno, + lprobs[sentno], + constraint_states[sentno], + beams_buf[sentno].clone(), + indices_buf[sentno].clone(), + scores_buf[sentno].clone(), + ) + new_scores_buf[sentno] = scores + new_indices_buf[sentno] = indices + new_beams_buf[sentno] = beams + self.constraint_states[sentno] = new_states + + return new_scores_buf, new_indices_buf, new_beams_buf + + @torch.jit.export + def step_sentence( + self, + step: int, + sentno: int, + lprobs: Tensor, + constraint_states: List[List[ConstraintState]], + beams_buf: Tensor, + indices_buf: Tensor, + scores_buf: Tensor, + ): + """Does per-sentence processing. Adds all constraints for each + hypothesis to the list of candidates; then removes duplicates, + sorts, and dynamically stripes across the banks. All tensor inputs + are collapsed to those pertaining to a single input sentence. + """ + device = lprobs.device + + # STEP 2: Add all constraints for each beam item + for beamno, state in enumerate(constraint_states): + next_tokens = torch.tensor(list(state.next_tokens()), device=device).long() + if next_tokens.numel() != 0: + indices_buf = torch.cat((indices_buf, next_tokens)) + next_beams = ( + torch.tensor(beamno, device=device) + .repeat(next_tokens.size(0)) + .long() + ) + beams_buf = torch.cat((beams_buf, next_beams)) + next_values = lprobs[beamno].take(next_tokens.view(-1)) + scores_buf = torch.cat((scores_buf, next_values)) + + # At the 0th time step, there is just one beam item + if step == 0: + break + + # STEP 3: Compute the "bank" for each candidate. This is the + # number of constraints it's generated. We need this so that + # we can do round-robin allocation of the beam across these + # banks. If C is the number of constraints, we select the best + # item in bank C, then the best in bank C-1, etc, followed by + # the 2nd-best in bank C, the 2nd-best in bank C-1, etc, and so + # on, until the maximum beam size. We accomplish this by + # creating a sort key and striping across the banks. + + # Compute the new states for all candidates + cands_size = indices_buf.size(0) + constraint_states = [ + constraint_states[beams_buf[i]].advance(indices_buf[i]) + for i in range(cands_size) + ] + + banks = torch.tensor([state.bank for state in constraint_states], device=device) + + # STEP 4: Sort + num_constraint_tokens = len(state.tokens) + + # Sort by keys (bank, score) (i.e., sort banks together, and scores + # within banks). AFAIK pytorch doesn't support either stable sort or + # multi-key sorting, so we have to hack this. + MAX_SCORE = -100 + sort_key = (num_constraint_tokens - banks) * MAX_SCORE + scores_buf + sort_values, sort_indices = sort_key.sort(dim=0, descending=True) + scores_buf = scores_buf[sort_indices] + indices_buf = indices_buf[sort_indices] + beams_buf = beams_buf[sort_indices] + banks = banks[sort_indices] + + # Sort the constraints to follow suit + constraint_states = [constraint_states[i] for i in sort_indices] + + # STEP 5: Remove duplicates. The topk calls (overall and + # per-row) plus the per-row generation of constraints will + # produce duplicates. Here we remove them. + + def roll(t): + """Rolls a 1d tensor left by 1. + + [0, 1, 2, 3, 4] becomes [4, 0, 1, 2, 3] + """ + return torch.cat((t[-1].unsqueeze(0), t[0:-1]), dim=0) + + # We map candidates (beam, token_id) to a single dimension. + # This is then shifted by 1. We can then easily identify + # duplicates and create a mask that identifies unique + # extensions. + uniques_mask = beams_buf * (self.vocab_size + 1) + indices_buf + uniques_mask = roll(uniques_mask) != uniques_mask + + # Use the mask to pare down the data structures + scores_buf = torch.masked_select(scores_buf, uniques_mask) + indices_buf = torch.masked_select(indices_buf, uniques_mask) + beams_buf = torch.masked_select(beams_buf, uniques_mask) + banks = torch.masked_select(banks, uniques_mask) + i = 1 + for mask in uniques_mask[1:]: + if not mask: + constraint_states.pop(i) + i += mask + + # STEP 6: Assign IDs round-robin across banks, sort, and + # truncate. Now that the candidates are sorted by (bank, + # score) and uniqed, we dynamically allocate the {beam_size} + # beam by striping across the candidates. These stripes will + # be used as sort keys to do round-robin selection. This is + # accomplished in a single pass with offsets. Sorting by + # highest-banks (furthest-along hypotheses) first ensures + # progress through the constraints. + # + # e.g., BANKS: 3 3 3 2 2 2 2 1 1 1 0 0 + # OLD STRIPES: 0 1 2 0 1 2 3 0 1 2 0 1 + # NEW STRIPES: 0 1+4 2+8 0+1 1+5 2+9 3+11 0+2 1+6 2+10 0+3 1+7 + # = 0 5 10 1 6 11 13 2 7 12 3 8 + # + # Sorting by this then gives the following banks: + # + # 3 2 1 0 3 2 1 0 3 2 1 2 + # + # We'll take the top {beam_size} of these. + stripe_offsets = [offset * (len(banks) + 1) for offset in range(len(banks) + 1)] + stripes = torch.zeros_like(banks) + cur_bank_count = -1 + cur_bank = banks[0] + for i, bank in enumerate(banks): + if bank != cur_bank: + cur_bank_count = 0 + cur_bank = bank + else: + cur_bank_count += 1 + stripes[i] = num_constraint_tokens - bank + stripe_offsets[cur_bank_count] + + # STEP 7: Sort by the stripes values + sort_values, sort_indices = stripes.sort(dim=0) + scores_buf = scores_buf[sort_indices] + indices_buf = indices_buf[sort_indices] + beams_buf = beams_buf[sort_indices] + constraint_states = [constraint_states[i] for i in sort_indices] + + # STEP 8: Truncate to the candidates size! + scores_buf = scores_buf[: self.num_cands] + indices_buf = indices_buf[: self.num_cands] + beams_buf = beams_buf[: self.num_cands] + + return scores_buf, indices_buf, beams_buf, constraint_states + + +class LengthConstrainedBeamSearch(Search): + def __init__(self, tgt_dict, min_len_a, min_len_b, max_len_a, max_len_b): + super().__init__(tgt_dict) + self.min_len_a = min_len_a + self.min_len_b = min_len_b + self.max_len_a = max_len_a + self.max_len_b = max_len_b + self.beam = BeamSearch(tgt_dict) + self.needs_src_lengths = True + + def step( + self, + step: int, + lprobs, + scores, + prev_output_tokens: Optional[Tensor] = None, + original_batch_idxs: Optional[Tensor] = None, + ): + min_lens = self.min_len_a * self.src_lengths + self.min_len_b + max_lens = self.max_len_a * self.src_lengths + self.max_len_b + lprobs[step < min_lens, :, self.eos] = -math.inf + lprobs[step >= max_lens, :, self.eos] = 0 + return self.beam.step(step, lprobs, scores) + + +class DiverseBeamSearch(Search): + """Diverse Beam Search. + + See "Diverse Beam Search: Decoding Diverse Solutions from Neural Sequence + Models" for details. + + We only implement the Hamming Diversity penalty here, which performed best + in the original paper. + """ + + def __init__(self, tgt_dict, num_groups, diversity_strength): + super().__init__(tgt_dict) + self.num_groups = num_groups + self.diversity_strength = -diversity_strength + self.beam = BeamSearch(tgt_dict) + + @torch.jit.export + def step( + self, + step: int, + lprobs, + scores, + prev_output_tokens: Optional[Tensor] = None, + original_batch_idxs: Optional[Tensor] = None, + ): + bsz, beam_size, vocab_size = lprobs.size() + if beam_size % self.num_groups != 0: + raise ValueError( + "DiverseBeamSearch requires --beam to be divisible by the number of groups" + ) + + # initialize diversity penalty + diversity_buf = torch.zeros(lprobs[:, 0, :].size()).to(lprobs) + + scores_G, indices_G, beams_G = [], [], [] + for g in range(self.num_groups): + lprobs_g = lprobs[:, g :: self.num_groups, :] + scores_g = scores[:, g :: self.num_groups, :] if step > 0 else None + + # apply diversity penalty + if g > 0: + lprobs_g = torch.add( + lprobs_g, + other=diversity_buf.unsqueeze(1), + alpha=self.diversity_strength, + ) + else: + lprobs_g = lprobs_g.contiguous() + + scores_buf, indices_buf, beams_buf = self.beam.step( + step, lprobs_g, scores_g + ) + beams_buf.mul_(self.num_groups).add_(g) + + scores_G.append(scores_buf.clone()) + indices_G.append(indices_buf.clone()) + beams_G.append(beams_buf.clone()) + + # update diversity penalty + diversity_buf.scatter_add_( + 1, indices_buf, torch.ones(indices_buf.size()).to(diversity_buf) + ) + + # interleave results from different groups + scores_buf = torch.stack(scores_G, dim=2).view(bsz, -1) + indices_buf = torch.stack(indices_G, dim=2).view(bsz, -1) + beams_buf = torch.stack(beams_G, dim=2).view(bsz, -1) + return scores_buf, indices_buf, beams_buf + + +class Sampling(Search): + sampling_topk: int + sampling_topp: float + + def __init__(self, tgt_dict, sampling_topk=-1, sampling_topp=-1.0): + super().__init__(tgt_dict) + self.sampling_topk = sampling_topk + self.sampling_topp = sampling_topp + + def _sample_topp(self, lprobs): + """Sample among the smallest set of elements whose cumulative probability mass exceeds p. + + See `"The Curious Case of Neural Text Degeneration" + (Holtzman et al., 2019) <https://arxiv.org/abs/1904.09751>`_. + + Args: + lprobs: (bsz x input_beam_size x vocab_size) + the model's log-probabilities over the vocabulary at the current step + + Return: A tuple of (trimed_probs, truncated_indices) where: + trimed_probs: (bsz x input_beam_size x ?) + the model's probabilities over the elements selected to sample from. The + width of the third dimension is determined by top-P. + truncated_indices: (bsz x input_beam_size x ?) + the indices of the chosen elements. + """ + probs = lprobs.exp_() + + # sort the last dimension (vocab dimension) in descending order + sorted_probs, sorted_indices = probs.sort(descending=True) + + # compute a mask to indicate the words to be included in the top-P set. + cumsum_probs = sorted_probs.cumsum(dim=2) + mask = cumsum_probs.lt(self.sampling_topp) + + # note that mask was computed by 'lt'. One more word needs to be included + # so that the cumulative probability mass can exceed p. + cumsum_mask = mask.cumsum(dim=2) + last_included = cumsum_mask[:, :, -1:] + last_included.clamp_(0, mask.size()[2] - 1) + mask = mask.scatter_(2, last_included, 1) + + # truncate unnecessary dims. + max_dim = last_included.max() + truncated_mask = mask[:, :, : max_dim + 1] + truncated_probs = sorted_probs[:, :, : max_dim + 1] + truncated_indices = sorted_indices[:, :, : max_dim + 1] + + # trim the words that are not in top-P by setting their probabilities + # to 0, so that they would not be sampled later. + trim_mask = ~truncated_mask + trimed_probs = truncated_probs.masked_fill_(trim_mask, 0) + return trimed_probs, truncated_indices + + @torch.jit.export + def step( + self, + step: int, + lprobs, + scores, + prev_output_tokens: Optional[Tensor] = None, + original_batch_idxs: Optional[Tensor] = None, + ): + bsz, beam_size, vocab_size = lprobs.size() + + if step == 0: + # at the first step all hypotheses are equally likely, so use + # only the first beam + lprobs = lprobs[:, ::beam_size, :].contiguous() + + if self.sampling_topp > 0: + # only sample from the smallest set of words whose cumulative probability mass exceeds p + probs, top_indices = self._sample_topp(lprobs) + elif self.sampling_topk > 0: + # only sample from top-k candidates + lprobs, top_indices = lprobs.topk(self.sampling_topk) + probs = lprobs.exp_() + else: + probs = lprobs.exp_() + + # dummy data to be consistent with true branch for type check + top_indices = torch.empty(0).to(probs) + # sample + if step == 0: + indices_buf = torch.multinomial( + probs.view(bsz, -1), + beam_size, + replacement=True, + ).view(bsz, beam_size) + else: + indices_buf = torch.multinomial( + probs.view(bsz * beam_size, -1), + 1, + replacement=True, + ).view(bsz, beam_size) + + if step == 0: + # expand to beam size + probs = probs.expand(bsz, beam_size, -1) + + # gather scores + scores_buf = torch.gather(probs, dim=2, index=indices_buf.unsqueeze(-1)) + scores_buf = scores_buf.log_().view(bsz, -1) + + # remap indices if using top-k or top-P sampling + if self.sampling_topk > 0 or self.sampling_topp > 0: + indices_buf = torch.gather( + top_indices.expand(bsz, beam_size, -1), + dim=2, + index=indices_buf.unsqueeze(-1), + ).squeeze(2) + + if step == 0: + beams_buf = indices_buf.new_zeros(bsz, beam_size) + else: + beams_buf = torch.arange(0, beam_size).to(indices_buf).repeat(bsz, 1) + # make scores cumulative + scores_buf.add_( + torch.gather(scores[:, :, step - 1], dim=1, index=beams_buf) + ) + + return scores_buf, indices_buf, beams_buf + + +class DiverseSiblingsSearch(Search): + """ + Beam search with diverse siblings. + + See "A Simple, Fast Diverse Decoding Algorithm for Neural Generation" for details. + https://arxiv.org/abs/1611.08562 + + 1/ Calculate hypotheses for each beam + 2/ Intra-sibling ordering + 3/ Rewrite scores + 4/ Choose top K hypotheses + + if diversity_rate == 0 is equivalent to BeamSearch + """ + + def __init__(self, tgt_dict, diversity_rate): + super().__init__(tgt_dict) + self.diversity_rate = diversity_rate + self.beam = BeamSearch(tgt_dict) + + def step( + self, + step: int, + lprobs, + scores, + prev_output_tokens: Optional[Tensor] = None, + original_batch_idxs: Optional[Tensor] = None, + ): + bsz, beam_size, vocab_size = lprobs.size() + k = min( + # Take the best 2 x beam_size predictions. We'll choose the first + # beam_size of these which don't predict eos to continue with. + beam_size * 2, + lprobs.view(bsz, -1).size(1) - 1, # -1 so we never select pad + ) + s_list: List[Tensor] + i_list: List[Tensor] + s_list = [torch.empty(0).to(lprobs) for i in range(beam_size)] + i_list = [torch.LongTensor().to(device=lprobs.device) for i in range(beam_size)] + sibling_score = torch.arange(1, k + 1).to(lprobs) * self.diversity_rate + + if step == 0: + return self.beam.step(step, lprobs, scores) + lprobs.add_(scores[:, :, step - 1].unsqueeze(-1)) + + # 1/ Calculate hypotheses for each beam + for i in range(beam_size): + torch.topk(lprobs[:, i, :].view(bsz, -1), k, out=(s_list[i], i_list[i])) + i_list[i].fmod_(vocab_size) + + # 2/ Intra-sibling ordering by default from topk + 3/ Rewrite scores + s_list[i].sub_(sibling_score) + + # 4/ Choose top K hypotheses + indices = torch.stack(i_list, dim=1).view(bsz, -1) + + final_scores = torch.empty(0).to(lprobs) + final_indices = torch.LongTensor().to(device=lprobs.device) + final_beams = torch.LongTensor().to(device=lprobs.device) + (final_scores, final_indices) = torch.topk( + torch.stack(s_list, dim=1).view(bsz, -1), + k, + ) + + final_beams = final_indices // k + + for i in range(bsz): + final_indices[i] = indices[i][final_indices[i]] + + return final_scores, final_indices, final_beams diff --git a/fairseq/sequence_generator.py b/fairseq/sequence_generator.py new file mode 100644 index 0000000000000000000000000000000000000000..8a3858563ec0c3cd7f3177bcd2897d27b61dbe00 --- /dev/null +++ b/fairseq/sequence_generator.py @@ -0,0 +1,980 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +from typing import Dict, List, Optional +import sys + +import torch +import torch.nn as nn +from fairseq import search, utils +from fairseq.data import data_utils +from fairseq.models import FairseqIncrementalDecoder +from torch import Tensor +from fairseq.ngram_repeat_block import NGramRepeatBlock + + +class SequenceGenerator(nn.Module): + def __init__( + self, + models, + tgt_dict, + beam_size=1, + max_len_a=0, + max_len_b=200, + max_len=0, + min_len=1, + normalize_scores=True, + len_penalty=1.0, + unk_penalty=0.0, + temperature=1.0, + match_source_len=False, + no_repeat_ngram_size=0, + search_strategy=None, + eos=None, + symbols_to_strip_from_output=None, + lm_model=None, + lm_weight=1.0, + ): + """Generates translations of a given source sentence. + + Args: + models (List[~fairseq.models.FairseqModel]): ensemble of models, + currently support fairseq.models.TransformerModel for scripting + beam_size (int, optional): beam width (default: 1) + max_len_a/b (int, optional): generate sequences of maximum length + ax + b, where x is the source length + max_len (int, optional): the maximum length of the generated output + (not including end-of-sentence) + min_len (int, optional): the minimum length of the generated output + (not including end-of-sentence) + normalize_scores (bool, optional): normalize scores by the length + of the output (default: True) + len_penalty (float, optional): length penalty, where <1.0 favors + shorter, >1.0 favors longer sentences (default: 1.0) + unk_penalty (float, optional): unknown word penalty, where <0 + produces more unks, >0 produces fewer (default: 0.0) + temperature (float, optional): temperature, where values + >1.0 produce more uniform samples and values <1.0 produce + sharper samples (default: 1.0) + match_source_len (bool, optional): outputs should match the source + length (default: False) + """ + super().__init__() + if isinstance(models, EnsembleModel): + self.model = models + else: + self.model = EnsembleModel(models) + self.tgt_dict = tgt_dict + self.pad = tgt_dict.pad() + self.unk = tgt_dict.unk() + self.eos = tgt_dict.eos() if eos is None else eos + self.symbols_to_strip_from_output = ( + symbols_to_strip_from_output.union({self.eos}) + if symbols_to_strip_from_output is not None + else {self.eos} + ) + self.vocab_size = len(tgt_dict) + self.beam_size = beam_size + # the max beam size is the dictionary size - 1, since we never select pad + self.beam_size = min(beam_size, self.vocab_size - 1) + self.max_len_a = max_len_a + self.max_len_b = max_len_b + self.min_len = min_len + self.max_len = max_len or self.model.max_decoder_positions() + + self.normalize_scores = normalize_scores + self.len_penalty = len_penalty + self.unk_penalty = unk_penalty + self.temperature = temperature + self.match_source_len = match_source_len + + if no_repeat_ngram_size > 0: + self.repeat_ngram_blocker = NGramRepeatBlock(no_repeat_ngram_size) + else: + self.repeat_ngram_blocker = None + + assert temperature > 0, "--temperature must be greater than 0" + + self.search = ( + search.BeamSearch(tgt_dict) if search_strategy is None else search_strategy + ) + # We only need to set src_lengths in LengthConstrainedBeamSearch. + # As a module attribute, setting it would break in multithread + # settings when the model is shared. + self.should_set_src_lengths = ( + hasattr(self.search, "needs_src_lengths") and self.search.needs_src_lengths + ) + + self.model.eval() + + self.lm_model = lm_model + self.lm_weight = lm_weight + if self.lm_model is not None: + self.lm_model.eval() + + def cuda(self): + self.model.cuda() + return self + + @torch.no_grad() + def forward( + self, + sample: Dict[str, Dict[str, Tensor]], + prefix_tokens: Optional[Tensor] = None, + bos_token: Optional[int] = None, + ): + """Generate a batch of translations. + + Args: + sample (dict): batch + prefix_tokens (torch.LongTensor, optional): force decoder to begin + with these tokens + bos_token (int, optional): beginning of sentence token + (default: self.eos) + """ + return self._generate(sample, prefix_tokens, bos_token=bos_token) + + # TODO(myleott): unused, deprecate after pytorch-translate migration + def generate_batched_itr(self, data_itr, beam_size=None, cuda=False, timer=None): + """Iterate over a batched dataset and yield individual translations. + Args: + cuda (bool, optional): use GPU for generation + timer (StopwatchMeter, optional): time generations + """ + for sample in data_itr: + s = utils.move_to_cuda(sample) if cuda else sample + if "net_input" not in s: + continue + input = s["net_input"] + # model.forward normally channels prev_output_tokens into the decoder + # separately, but SequenceGenerator directly calls model.encoder + encoder_input = { + k: v for k, v in input.items() if k != "prev_output_tokens" + } + if timer is not None: + timer.start() + with torch.no_grad(): + hypos = self.generate(encoder_input) + if timer is not None: + timer.stop(sum(len(h[0]["tokens"]) for h in hypos)) + for i, id in enumerate(s["id"].data): + # remove padding + src = utils.strip_pad(input["src_tokens"].data[i, :], self.pad) + ref = ( + utils.strip_pad(s["target"].data[i, :], self.pad) + if s["target"] is not None + else None + ) + yield id, src, ref, hypos[i] + + @torch.no_grad() + def generate(self, models, sample: Dict[str, Dict[str, Tensor]], **kwargs) -> List[List[Dict[str, Tensor]]]: + """Generate translations. Match the api of other fairseq generators. + + Args: + models (List[~fairseq.models.FairseqModel]): ensemble of models + sample (dict): batch + prefix_tokens (torch.LongTensor, optional): force decoder to begin + with these tokens + constraints (torch.LongTensor, optional): force decoder to include + the list of constraints + bos_token (int, optional): beginning of sentence token + (default: self.eos) + """ + return self._generate(sample, **kwargs) + + def _generate( + self, + sample: Dict[str, Dict[str, Tensor]], + prefix_tokens: Optional[Tensor] = None, + constraints: Optional[Tensor] = None, + bos_token: Optional[int] = None, + ): + incremental_states = torch.jit.annotate( + List[Dict[str, Dict[str, Optional[Tensor]]]], + [ + torch.jit.annotate(Dict[str, Dict[str, Optional[Tensor]]], {}) + for i in range(self.model.models_size) + ], + ) + net_input = sample["net_input"] + + if "src_tokens" in net_input: + src_tokens = net_input["src_tokens"] + # length of the source text being the character length except EndOfSentence and pad + src_lengths = ( + (src_tokens.ne(self.eos) & src_tokens.ne(self.pad)).long().sum(dim=1) + ) + elif "source" in net_input: + src_tokens = net_input["source"] + src_lengths = ( + net_input["padding_mask"].size(-1) - net_input["padding_mask"].sum(-1) + if net_input["padding_mask"] is not None + else torch.tensor(src_tokens.size(-1)).to(src_tokens) + ) + elif "features" in net_input: + src_tokens = net_input["features"] + src_lengths = ( + net_input["padding_mask"].size(-1) - net_input["padding_mask"].sum(-1) + if net_input["padding_mask"] is not None + else torch.tensor(src_tokens.size(-1)).to(src_tokens) + ) + else: + raise Exception("expected src_tokens or source in net input. input keys: " + str(net_input.keys())) + + # bsz: total number of sentences in beam + # Note that src_tokens may have more than 2 dimensions (i.e. audio features) + bsz, src_len = src_tokens.size()[:2] + beam_size = self.beam_size + + if constraints is not None and not self.search.supports_constraints: + raise NotImplementedError( + "Target-side constraints were provided, but search method doesn't support them" + ) + + # Initialize constraints, when active + self.search.init_constraints(constraints, beam_size) + + max_len: int = -1 + if self.match_source_len: + max_len = src_lengths.max().item() + else: + max_len = min( + int(self.max_len_a * src_len + self.max_len_b), + self.max_len - 1, + ) + assert ( + self.min_len <= max_len + ), "min_len cannot be larger than max_len, please adjust these!" + # compute the encoder output for each beam + encoder_outs = self.model.forward_encoder(net_input) + + # placeholder of indices for bsz * beam_size to hold tokens and accumulative scores + new_order = torch.arange(bsz).view(-1, 1).repeat(1, beam_size).view(-1) + new_order = new_order.to(src_tokens.device).long() + encoder_outs = self.model.reorder_encoder_out(encoder_outs, new_order) + # ensure encoder_outs is a List. + assert encoder_outs is not None + + # initialize buffers + scores = ( + torch.zeros(bsz * beam_size, max_len + 1).to(src_tokens).float() + ) # +1 for eos; pad is never chosen for scoring + tokens = ( + torch.zeros(bsz * beam_size, max_len + 2) + .to(src_tokens) + .long() + .fill_(self.pad) + ) # +2 for eos and pad + tokens[:, 0] = self.eos if bos_token is None else bos_token + attn: Optional[Tensor] = None + + # A list that indicates candidates that should be ignored. + # For example, suppose we're sampling and have already finalized 2/5 + # samples. Then cands_to_ignore would mark 2 positions as being ignored, + # so that we only finalize the remaining 3 samples. + cands_to_ignore = ( + torch.zeros(bsz, beam_size).to(src_tokens).eq(-1) + ) # forward and backward-compatible False mask + + # list of completed sentences + finalized = torch.jit.annotate( + List[List[Dict[str, Tensor]]], + [torch.jit.annotate(List[Dict[str, Tensor]], []) for i in range(bsz)], + ) # contains lists of dictionaries of infomation about the hypothesis being finalized at each step + + # a boolean array indicating if the sentence at the index is finished or not + finished = [False for i in range(bsz)] + num_remaining_sent = bsz # number of sentences remaining + + # number of candidate hypos per step + cand_size = 2 * beam_size # 2 x beam size in case half are EOS + + # offset arrays for converting between different indexing schemes + bbsz_offsets = ( + (torch.arange(0, bsz) * beam_size) + .unsqueeze(1) + .type_as(tokens) + .to(src_tokens.device) + ) + cand_offsets = torch.arange(0, cand_size).type_as(tokens).to(src_tokens.device) + + reorder_state: Optional[Tensor] = None + batch_idxs: Optional[Tensor] = None + + original_batch_idxs: Optional[Tensor] = None + if "id" in sample and isinstance(sample["id"], Tensor): + original_batch_idxs = sample["id"] + else: + original_batch_idxs = torch.arange(0, bsz).type_as(tokens) + + for step in range(max_len + 1): # one extra step for EOS marker + # reorder decoder internal states based on the prev choice of beams + if reorder_state is not None: + if batch_idxs is not None: + # update beam indices to take into account removed sentences + corr = batch_idxs - torch.arange(batch_idxs.numel()).type_as( + batch_idxs + ) + reorder_state.view(-1, beam_size).add_( + corr.unsqueeze(-1) * beam_size + ) + original_batch_idxs = original_batch_idxs[batch_idxs] + self.model.reorder_incremental_state(incremental_states, reorder_state) + encoder_outs = self.model.reorder_encoder_out( + encoder_outs, reorder_state + ) + + lprobs, avg_attn_scores = self.model.forward_decoder( + tokens[:, : step + 1], + encoder_outs, + incremental_states, + self.temperature, + ) + + if self.lm_model is not None: + lm_out = self.lm_model(tokens[:, : step + 1]) + probs = self.lm_model.get_normalized_probs( + lm_out, log_probs=True, sample=None + ) + probs = probs[:, -1, :] * self.lm_weight + lprobs += probs + + lprobs[lprobs != lprobs] = torch.tensor(-math.inf).to(lprobs) + + lprobs[:, self.pad] = -math.inf # never select pad + lprobs[:, self.unk] -= self.unk_penalty # apply unk penalty + + # handle max length constraint + if step >= max_len: + lprobs[:, : self.eos] = -math.inf + lprobs[:, self.eos + 1 :] = -math.inf + + # handle prefix tokens (possibly with different lengths) + if ( + prefix_tokens is not None + and step < prefix_tokens.size(1) + and step < max_len + ): + lprobs, tokens, scores = self._prefix_tokens( + step, lprobs, scores, tokens, prefix_tokens, beam_size + ) + elif step < self.min_len: + # minimum length constraint (does not apply if using prefix_tokens) + lprobs[:, self.eos] = -math.inf + + # Record attention scores, only support avg_attn_scores is a Tensor + if avg_attn_scores is not None: + if attn is None: + attn = torch.empty( + bsz * beam_size, avg_attn_scores.size(1), max_len + 2 + ).to(scores) + attn[:, :, step + 1].copy_(avg_attn_scores) + + scores = scores.type_as(lprobs) + eos_bbsz_idx = torch.empty(0).to( + tokens + ) # indices of hypothesis ending with eos (finished sentences) + eos_scores = torch.empty(0).to( + scores + ) # scores of hypothesis ending with eos (finished sentences) + + if self.should_set_src_lengths: + self.search.set_src_lengths(src_lengths) + + if self.repeat_ngram_blocker is not None: + lprobs = self.repeat_ngram_blocker(tokens, lprobs, bsz, beam_size, step) + + # Shape: (batch, cand_size) + cand_scores, cand_indices, cand_beams = self.search.step( + step, + lprobs.view(bsz, -1, self.vocab_size), + scores.view(bsz, beam_size, -1)[:, :, :step], + tokens[:, : step + 1], + original_batch_idxs, + ) + + # cand_bbsz_idx contains beam indices for the top candidate + # hypotheses, with a range of values: [0, bsz*beam_size), + # and dimensions: [bsz, cand_size] + cand_bbsz_idx = cand_beams.add(bbsz_offsets) + + # finalize hypotheses that end in eos + # Shape of eos_mask: (batch size, beam size) + eos_mask = cand_indices.eq(self.eos) & cand_scores.ne(-math.inf) + eos_mask[:, :beam_size][cands_to_ignore] = torch.tensor(0).to(eos_mask) + + # only consider eos when it's among the top beam_size indices + # Now we know what beam item(s) to finish + # Shape: 1d list of absolute-numbered + eos_bbsz_idx = torch.masked_select( + cand_bbsz_idx[:, :beam_size], mask=eos_mask[:, :beam_size] + ) + + finalized_sents: List[int] = [] + if eos_bbsz_idx.numel() > 0: + eos_scores = torch.masked_select( + cand_scores[:, :beam_size], mask=eos_mask[:, :beam_size] + ) + + finalized_sents = self.finalize_hypos( + step, + eos_bbsz_idx, + eos_scores, + tokens, + scores, + finalized, + finished, + beam_size, + attn, + src_lengths, + max_len, + ) + num_remaining_sent -= len(finalized_sents) + + assert num_remaining_sent >= 0 + if num_remaining_sent == 0: + break + if self.search.stop_on_max_len and step >= max_len: + break + assert step < max_len, f"{step} < {max_len}" + + # Remove finalized sentences (ones for which {beam_size} + # finished hypotheses have been generated) from the batch. + if len(finalized_sents) > 0: + new_bsz = bsz - len(finalized_sents) + + # construct batch_idxs which holds indices of batches to keep for the next pass + batch_mask = torch.ones( + bsz, dtype=torch.bool, device=cand_indices.device + ) + batch_mask[finalized_sents] = False + # TODO replace `nonzero(as_tuple=False)` after TorchScript supports it + batch_idxs = torch.arange( + bsz, device=cand_indices.device + ).masked_select(batch_mask) + + # Choose the subset of the hypothesized constraints that will continue + self.search.prune_sentences(batch_idxs) + + eos_mask = eos_mask[batch_idxs] + cand_beams = cand_beams[batch_idxs] + bbsz_offsets.resize_(new_bsz, 1) + cand_bbsz_idx = cand_beams.add(bbsz_offsets) + cand_scores = cand_scores[batch_idxs] + cand_indices = cand_indices[batch_idxs] + + if prefix_tokens is not None: + prefix_tokens = prefix_tokens[batch_idxs] + src_lengths = src_lengths[batch_idxs] + cands_to_ignore = cands_to_ignore[batch_idxs] + + scores = scores.view(bsz, -1)[batch_idxs].view(new_bsz * beam_size, -1) + tokens = tokens.view(bsz, -1)[batch_idxs].view(new_bsz * beam_size, -1) + if attn is not None: + attn = attn.view(bsz, -1)[batch_idxs].view( + new_bsz * beam_size, attn.size(1), -1 + ) + bsz = new_bsz + else: + batch_idxs = None + + # Set active_mask so that values > cand_size indicate eos hypos + # and values < cand_size indicate candidate active hypos. + # After, the min values per row are the top candidate active hypos + + # Rewrite the operator since the element wise or is not supported in torchscript. + + eos_mask[:, :beam_size] = ~((~cands_to_ignore) & (~eos_mask[:, :beam_size])) + active_mask = torch.add( + eos_mask.type_as(cand_offsets) * cand_size, + cand_offsets[: eos_mask.size(1)], + ) + + # get the top beam_size active hypotheses, which are just + # the hypos with the smallest values in active_mask. + # {active_hypos} indicates which {beam_size} hypotheses + # from the list of {2 * beam_size} candidates were + # selected. Shapes: (batch size, beam size) + new_cands_to_ignore, active_hypos = torch.topk( + active_mask, k=beam_size, dim=1, largest=False + ) + + # update cands_to_ignore to ignore any finalized hypos. + cands_to_ignore = new_cands_to_ignore.ge(cand_size)[:, :beam_size] + # Make sure there is at least one active item for each sentence in the batch. + assert (~cands_to_ignore).any(dim=1).all() + + # update cands_to_ignore to ignore any finalized hypos + + # {active_bbsz_idx} denotes which beam number is continued for each new hypothesis (a beam + # can be selected more than once). + active_bbsz_idx = torch.gather(cand_bbsz_idx, dim=1, index=active_hypos) + active_scores = torch.gather(cand_scores, dim=1, index=active_hypos) + + active_bbsz_idx = active_bbsz_idx.view(-1) + active_scores = active_scores.view(-1) + + # copy tokens and scores for active hypotheses + + # Set the tokens for each beam (can select the same row more than once) + tokens[:, : step + 1] = torch.index_select( + tokens[:, : step + 1], dim=0, index=active_bbsz_idx + ) + # Select the next token for each of them + tokens.view(bsz, beam_size, -1)[:, :, step + 1] = torch.gather( + cand_indices, dim=1, index=active_hypos + ) + if step > 0: + scores[:, :step] = torch.index_select( + scores[:, :step], dim=0, index=active_bbsz_idx + ) + scores.view(bsz, beam_size, -1)[:, :, step] = torch.gather( + cand_scores, dim=1, index=active_hypos + ) + + # Update constraints based on which candidates were selected for the next beam + self.search.update_constraints(active_hypos) + + # copy attention for active hypotheses + if attn is not None: + attn[:, :, : step + 2] = torch.index_select( + attn[:, :, : step + 2], dim=0, index=active_bbsz_idx + ) + + # reorder incremental state in decoder + reorder_state = active_bbsz_idx + + # sort by score descending + for sent in range(len(finalized)): + scores = torch.tensor( + [float(elem["score"].item()) for elem in finalized[sent]] + ) + _, sorted_scores_indices = torch.sort(scores, descending=True) + finalized[sent] = [finalized[sent][ssi] for ssi in sorted_scores_indices] + finalized[sent] = torch.jit.annotate( + List[Dict[str, Tensor]], finalized[sent] + ) + return finalized + + def _prefix_tokens( + self, step: int, lprobs, scores, tokens, prefix_tokens, beam_size: int + ): + """Handle prefix tokens""" + prefix_toks = prefix_tokens[:, step].unsqueeze(-1).repeat(1, beam_size).view(-1) + prefix_lprobs = lprobs.gather(-1, prefix_toks.unsqueeze(-1)) + prefix_mask = prefix_toks.ne(self.pad) + lprobs[prefix_mask] = torch.tensor(-math.inf).to(lprobs) + lprobs[prefix_mask] = lprobs[prefix_mask].scatter( + -1, prefix_toks[prefix_mask].unsqueeze(-1), prefix_lprobs[prefix_mask] + ) + # if prefix includes eos, then we should make sure tokens and + # scores are the same across all beams + eos_mask = prefix_toks.eq(self.eos) + if eos_mask.any(): + # validate that the first beam matches the prefix + first_beam = tokens[eos_mask].view(-1, beam_size, tokens.size(-1))[ + :, 0, 1 : step + 1 + ] + eos_mask_batch_dim = eos_mask.view(-1, beam_size)[:, 0] + target_prefix = prefix_tokens[eos_mask_batch_dim][:, :step] + assert (first_beam == target_prefix).all() + + # copy tokens, scores and lprobs from the first beam to all beams + tokens = self.replicate_first_beam(tokens, eos_mask_batch_dim, beam_size) + scores = self.replicate_first_beam(scores, eos_mask_batch_dim, beam_size) + lprobs = self.replicate_first_beam(lprobs, eos_mask_batch_dim, beam_size) + return lprobs, tokens, scores + + def replicate_first_beam(self, tensor, mask, beam_size: int): + tensor = tensor.view(-1, beam_size, tensor.size(-1)) + tensor[mask] = tensor[mask][:, :1, :] + return tensor.view(-1, tensor.size(-1)) + + def finalize_hypos( + self, + step: int, + bbsz_idx, + eos_scores, + tokens, + scores, + finalized: List[List[Dict[str, Tensor]]], + finished: List[bool], + beam_size: int, + attn: Optional[Tensor], + src_lengths, + max_len: int, + ): + """Finalize hypothesis, store finalized information in `finalized`, and change `finished` accordingly. + A sentence is finalized when {beam_size} finished items have been collected for it. + + Returns number of sentences (not beam items) being finalized. + These will be removed from the batch and not processed further. + Args: + bbsz_idx (Tensor): + """ + assert bbsz_idx.numel() == eos_scores.numel() + + # clone relevant token and attention tensors. + # tokens is (batch * beam, max_len). So the index_select + # gets the newly EOS rows, then selects cols 1..{step + 2} + tokens_clone = tokens.index_select(0, bbsz_idx)[ + :, 1 : step + 2 + ] # skip the first index, which is EOS + + tokens_clone[:, step] = self.eos + attn_clone = ( + attn.index_select(0, bbsz_idx)[:, :, 1 : step + 2] + if attn is not None + else None + ) + + # compute scores per token position + pos_scores = scores.index_select(0, bbsz_idx)[:, : step + 1] + pos_scores[:, step] = eos_scores + # convert from cumulative to per-position scores + pos_scores[:, 1:] = pos_scores[:, 1:] - pos_scores[:, :-1] + + # normalize sentence-level scores + if self.normalize_scores: + eos_scores /= (step + 1) ** self.len_penalty + + # cum_unfin records which sentences in the batch are finished. + # It helps match indexing between (a) the original sentences + # in the batch and (b) the current, possibly-reduced set of + # sentences. + cum_unfin: List[int] = [] + prev = 0 + for f in finished: + if f: + prev += 1 + else: + cum_unfin.append(prev) + + # The keys here are of the form "{sent}_{unfin_idx}", where + # "unfin_idx" is the index in the current (possibly reduced) + # list of sentences, and "sent" is the index in the original, + # unreduced batch + # set() is not supported in script export + sents_seen: Dict[str, Optional[Tensor]] = {} + + # For every finished beam item + for i in range(bbsz_idx.size()[0]): + idx = bbsz_idx[i] + score = eos_scores[i] + # sentence index in the current (possibly reduced) batch + unfin_idx = idx // beam_size + # sentence index in the original (unreduced) batch + sent = unfin_idx + cum_unfin[unfin_idx] + # Cannot create dict for key type '(int, int)' in torchscript. + # The workaround is to cast int to string + seen = str(sent.item()) + "_" + str(unfin_idx.item()) + if seen not in sents_seen: + sents_seen[seen] = None + + if self.match_source_len and step > src_lengths[unfin_idx]: + score = torch.tensor(-math.inf).to(score) + + # An input sentence (among those in a batch) is finished when + # beam_size hypotheses have been collected for it + if len(finalized[sent]) < beam_size: + if attn_clone is not None: + # remove padding tokens from attn scores + hypo_attn = attn_clone[i] + else: + hypo_attn = torch.empty(0) + + finalized[sent].append( + { + "tokens": tokens_clone[i], + "score": score, + "attention": hypo_attn, # src_len x tgt_len + "alignment": torch.empty(0), + "positional_scores": pos_scores[i], + } + ) + + newly_finished: List[int] = [] + + for seen in sents_seen.keys(): + # check termination conditions for this sentence + sent: int = int(float(seen.split("_")[0])) + unfin_idx: int = int(float(seen.split("_")[1])) + + if not finished[sent] and self.is_finished( + step, unfin_idx, max_len, len(finalized[sent]), beam_size + ): + finished[sent] = True + newly_finished.append(unfin_idx) + + return newly_finished + + def is_finished( + self, + step: int, + unfin_idx: int, + max_len: int, + finalized_sent_len: int, + beam_size: int, + ): + """ + Check whether decoding for a sentence is finished, which + occurs when the list of finalized sentences has reached the + beam size, or when we reach the maximum length. + """ + assert finalized_sent_len <= beam_size + if finalized_sent_len == beam_size or step == max_len: + return True + return False + + +class EnsembleModel(nn.Module): + """A wrapper around an ensemble of models.""" + + def __init__(self, models): + super().__init__() + self.models_size = len(models) + # method '__len__' is not supported in ModuleList for torch script + self.single_model = models[0] + self.models = nn.ModuleList(models) + + self.has_incremental: bool = False + if all( + hasattr(m, "decoder") and isinstance(m.decoder, FairseqIncrementalDecoder) + for m in models + ): + self.has_incremental = True + + def forward(self): + pass + + def has_encoder(self): + return hasattr(self.single_model, "encoder") + + def has_incremental_states(self): + return self.has_incremental + + def max_decoder_positions(self): + return min([m.max_decoder_positions() for m in self.models if hasattr(m, "max_decoder_positions")] + [sys.maxsize]) + + @torch.jit.export + def forward_encoder(self, net_input: Dict[str, Tensor]): + if not self.has_encoder(): + return None + return [model.encoder.forward_torchscript(net_input) for model in self.models] + + @torch.jit.export + def forward_decoder( + self, + tokens, + encoder_outs: List[Dict[str, List[Tensor]]], + incremental_states: List[Dict[str, Dict[str, Optional[Tensor]]]], + temperature: float = 1.0, + ): + log_probs = [] + avg_attn: Optional[Tensor] = None + encoder_out: Optional[Dict[str, List[Tensor]]] = None + for i, model in enumerate(self.models): + if self.has_encoder(): + encoder_out = encoder_outs[i] + # decode each model + if self.has_incremental_states(): + decoder_out = model.decoder.forward( + tokens, + encoder_out=encoder_out, + incremental_state=incremental_states[i], + ) + else: + if hasattr(model, "decoder"): + decoder_out = model.decoder.forward(tokens, encoder_out=encoder_out) + else: + decoder_out = model.forward(tokens) + + attn: Optional[Tensor] = None + decoder_len = len(decoder_out) + if decoder_len > 1 and decoder_out[1] is not None: + if isinstance(decoder_out[1], Tensor): + attn = decoder_out[1] + else: + attn_holder = decoder_out[1]["attn"] + if isinstance(attn_holder, Tensor): + attn = attn_holder + elif attn_holder is not None: + attn = attn_holder[0] + if attn is not None: + attn = attn[:, -1, :] + + decoder_out_tuple = ( + decoder_out[0][:, -1:, :].div_(temperature), + None if decoder_len <= 1 else decoder_out[1], + ) + probs = model.get_normalized_probs( + decoder_out_tuple, log_probs=True, sample=None + ) + probs = probs[:, -1, :] + if self.models_size == 1: + return probs, attn + + log_probs.append(probs) + if attn is not None: + if avg_attn is None: + avg_attn = attn + else: + avg_attn.add_(attn) + + avg_probs = torch.logsumexp(torch.stack(log_probs, dim=0), dim=0) - math.log( + self.models_size + ) + + if avg_attn is not None: + avg_attn.div_(self.models_size) + return avg_probs, avg_attn + + @torch.jit.export + def reorder_encoder_out( + self, encoder_outs: Optional[List[Dict[str, List[Tensor]]]], new_order + ): + """ + Reorder encoder output according to *new_order*. + + Args: + encoder_out: output from the ``forward()`` method + new_order (LongTensor): desired order + + Returns: + *encoder_out* rearranged according to *new_order* + """ + new_outs: List[Dict[str, List[Tensor]]] = [] + if not self.has_encoder(): + return new_outs + for i, model in enumerate(self.models): + assert encoder_outs is not None + new_outs.append( + model.encoder.reorder_encoder_out(encoder_outs[i], new_order) + ) + return new_outs + + @torch.jit.export + def reorder_incremental_state( + self, + incremental_states: List[Dict[str, Dict[str, Optional[Tensor]]]], + new_order, + ): + if not self.has_incremental_states(): + return + for i, model in enumerate(self.models): + model.decoder.reorder_incremental_state_scripting( + incremental_states[i], new_order + ) + + +class SequenceGeneratorWithAlignment(SequenceGenerator): + def __init__( + self, models, tgt_dict, left_pad_target=False, print_alignment="hard", **kwargs + ): + """Generates translations of a given source sentence. + + Produces alignments following "Jointly Learning to Align and + Translate with Transformer Models" (Garg et al., EMNLP 2019). + + Args: + left_pad_target (bool, optional): Whether or not the + hypothesis should be left padded or not when they are + teacher forced for generating alignments. + """ + super().__init__(EnsembleModelWithAlignment(models), tgt_dict, **kwargs) + self.left_pad_target = left_pad_target + + if print_alignment == "hard": + self.extract_alignment = utils.extract_hard_alignment + elif print_alignment == "soft": + self.extract_alignment = utils.extract_soft_alignment + + @torch.no_grad() + def generate(self, models, sample, **kwargs): + finalized = super()._generate(sample, **kwargs) + + src_tokens = sample["net_input"]["src_tokens"] + bsz = src_tokens.shape[0] + beam_size = self.beam_size + ( + src_tokens, + src_lengths, + prev_output_tokens, + tgt_tokens, + ) = self._prepare_batch_for_alignment(sample, finalized) + if any(getattr(m, "full_context_alignment", False) for m in self.model.models): + attn = self.model.forward_align(src_tokens, src_lengths, prev_output_tokens) + else: + attn = [ + finalized[i // beam_size][i % beam_size]["attention"].transpose(1, 0) + for i in range(bsz * beam_size) + ] + + if src_tokens.device != "cpu": + src_tokens = src_tokens.to("cpu") + tgt_tokens = tgt_tokens.to("cpu") + attn = [i.to("cpu") for i in attn] + + # Process the attn matrix to extract hard alignments. + for i in range(bsz * beam_size): + alignment = self.extract_alignment( + attn[i], src_tokens[i], tgt_tokens[i], self.pad, self.eos + ) + finalized[i // beam_size][i % beam_size]["alignment"] = alignment + return finalized + + def _prepare_batch_for_alignment(self, sample, hypothesis): + src_tokens = sample["net_input"]["src_tokens"] + bsz = src_tokens.shape[0] + src_tokens = ( + src_tokens[:, None, :] + .expand(-1, self.beam_size, -1) + .contiguous() + .view(bsz * self.beam_size, -1) + ) + src_lengths = sample["net_input"]["src_lengths"] + src_lengths = ( + src_lengths[:, None] + .expand(-1, self.beam_size) + .contiguous() + .view(bsz * self.beam_size) + ) + prev_output_tokens = data_utils.collate_tokens( + [beam["tokens"] for example in hypothesis for beam in example], + self.pad, + self.eos, + self.left_pad_target, + move_eos_to_beginning=True, + ) + tgt_tokens = data_utils.collate_tokens( + [beam["tokens"] for example in hypothesis for beam in example], + self.pad, + self.eos, + self.left_pad_target, + move_eos_to_beginning=False, + ) + return src_tokens, src_lengths, prev_output_tokens, tgt_tokens + + +class EnsembleModelWithAlignment(EnsembleModel): + """A wrapper around an ensemble of models.""" + + def __init__(self, models): + super().__init__(models) + + def forward_align(self, src_tokens, src_lengths, prev_output_tokens): + avg_attn = None + for model in self.models: + decoder_out = model(src_tokens, src_lengths, prev_output_tokens) + attn = decoder_out[1]["attn"][0] + if avg_attn is None: + avg_attn = attn + else: + avg_attn.add_(attn) + if len(self.models) > 1: + avg_attn.div_(len(self.models)) + return avg_attn diff --git a/fairseq/sequence_scorer.py b/fairseq/sequence_scorer.py new file mode 100644 index 0000000000000000000000000000000000000000..411d4df4445ef8dd3f1907ad56f9de6943d1fed8 --- /dev/null +++ b/fairseq/sequence_scorer.py @@ -0,0 +1,153 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import sys + +import torch +from fairseq import utils + + +class SequenceScorer(object): + """Scores the target for a given source sentence.""" + + def __init__( + self, + tgt_dict, + softmax_batch=None, + compute_alignment=False, + eos=None, + symbols_to_strip_from_output=None, + ): + self.pad = tgt_dict.pad() + self.eos = tgt_dict.eos() if eos is None else eos + self.softmax_batch = softmax_batch or sys.maxsize + assert self.softmax_batch > 0 + self.compute_alignment = compute_alignment + self.symbols_to_strip_from_output = ( + symbols_to_strip_from_output.union({self.eos}) + if symbols_to_strip_from_output is not None + else {self.eos} + ) + + @torch.no_grad() + def generate(self, models, sample, **kwargs): + """Score a batch of translations.""" + net_input = sample["net_input"] + + def batch_for_softmax(dec_out, target): + # assumes decoder_out[0] is the only thing needed (may not be correct for future models!) + first, rest = dec_out[0], dec_out[1:] + bsz, tsz, dim = first.shape + if bsz * tsz < self.softmax_batch: + yield dec_out, target, True + else: + flat = first.contiguous().view(1, -1, dim) + flat_tgt = target.contiguous().view(flat.shape[:-1]) + s = 0 + while s < flat.size(1): + e = s + self.softmax_batch + yield (flat[:, s:e],) + rest, flat_tgt[:, s:e], False + s = e + + def gather_target_probs(probs, target): + probs = probs.gather( + dim=2, + index=target.unsqueeze(-1), + ) + return probs + + orig_target = sample["target"] + + # compute scores for each model in the ensemble + avg_probs = None + avg_attn = None + for model in models: + model.eval() + decoder_out = model(**net_input) + attn = decoder_out[1] if len(decoder_out) > 1 else None + if type(attn) is dict: + attn = attn.get("attn", None) + + batched = batch_for_softmax(decoder_out, orig_target) + probs, idx = None, 0 + for bd, tgt, is_single in batched: + sample["target"] = tgt + curr_prob = model.get_normalized_probs( + bd, log_probs=len(models) == 1, sample=sample + ).data + if is_single: + probs = gather_target_probs(curr_prob, orig_target) + else: + if probs is None: + probs = curr_prob.new(orig_target.numel()) + step = curr_prob.size(0) * curr_prob.size(1) + end = step + idx + tgt_probs = gather_target_probs( + curr_prob.view(tgt.shape + (curr_prob.size(-1),)), tgt + ) + probs[idx:end] = tgt_probs.view(-1) + idx = end + sample["target"] = orig_target + + probs = probs.view(sample["target"].shape) + + if avg_probs is None: + avg_probs = probs + else: + avg_probs.add_(probs) + if attn is not None: + if torch.is_tensor(attn): + attn = attn.data + else: + attn = attn[0] + if avg_attn is None: + avg_attn = attn + else: + avg_attn.add_(attn) + if len(models) > 1: + avg_probs.div_(len(models)) + avg_probs.log_() + if avg_attn is not None: + avg_attn.div_(len(models)) + + bsz = avg_probs.size(0) + hypos = [] + start_idxs = sample["start_indices"] if "start_indices" in sample else [0] * bsz + for i in range(bsz): + # remove padding from ref + ref = ( + utils.strip_pad(sample["target"][i, start_idxs[i] :], self.pad) + if sample["target"] is not None + else None + ) + tgt_len = ref.numel() + avg_probs_i = avg_probs[i][start_idxs[i] : start_idxs[i] + tgt_len] + score_i = avg_probs_i.sum() / tgt_len + if avg_attn is not None: + avg_attn_i = avg_attn[i] + if self.compute_alignment: + alignment = utils.extract_hard_alignment( + avg_attn_i, + sample["net_input"]["src_tokens"][i], + sample["target"][i], + self.pad, + self.eos, + ) + else: + alignment = None + else: + avg_attn_i = alignment = None + hypos.append( + [ + { + "tokens": ref, + "score": score_i, + "attention": avg_attn_i, + "alignment": alignment, + "positional_scores": avg_probs_i, + } + ] + ) + return hypos diff --git a/fairseq/tasks/__init__.py b/fairseq/tasks/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..79dde74057f40a368590cbf0ca0d290f1787a264 --- /dev/null +++ b/fairseq/tasks/__init__.py @@ -0,0 +1,136 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +"""isort:skip_file""" + +import argparse +import importlib +import os + +from fairseq.dataclass import FairseqDataclass +from fairseq.dataclass.utils import merge_with_parent, populate_dataclass +from hydra.core.config_store import ConfigStore + +from .fairseq_task import FairseqTask, LegacyFairseqTask # noqa + + +# register dataclass +TASK_DATACLASS_REGISTRY = {} +TASK_REGISTRY = {} +TASK_CLASS_NAMES = set() + + +def setup_task(cfg: FairseqDataclass, **kwargs): + task = None + task_name = getattr(cfg, "task", None) + + if isinstance(task_name, str): + # legacy tasks + task = TASK_REGISTRY[task_name] + if task_name in TASK_DATACLASS_REGISTRY: + dc = TASK_DATACLASS_REGISTRY[task_name] + cfg = populate_dataclass(dc(), cfg) + else: + task_name = getattr(cfg, "_name", None) + + if task_name and task_name in TASK_DATACLASS_REGISTRY: + dc = TASK_DATACLASS_REGISTRY[task_name] + cfg = merge_with_parent(dc(), cfg) + task = TASK_REGISTRY[task_name] + + assert ( + task is not None + ), f"Could not infer task type from {cfg}. Available tasks: {TASK_REGISTRY.keys()}" + + return task.setup_task(cfg, **kwargs) + + +def register_task(name, dataclass=None): + """ + New tasks can be added to fairseq with the + :func:`~fairseq.tasks.register_task` function decorator. + + For example:: + + @register_task('classification') + class ClassificationTask(FairseqTask): + (...) + + .. note:: + + All Tasks must implement the :class:`~fairseq.tasks.FairseqTask` + interface. + + Args: + name (str): the name of the task + """ + + def register_task_cls(cls): + if name in TASK_REGISTRY: + raise ValueError("Cannot register duplicate task ({})".format(name)) + if not issubclass(cls, FairseqTask): + raise ValueError( + "Task ({}: {}) must extend FairseqTask".format(name, cls.__name__) + ) + if cls.__name__ in TASK_CLASS_NAMES: + raise ValueError( + "Cannot register task with duplicate class name ({})".format( + cls.__name__ + ) + ) + TASK_REGISTRY[name] = cls + TASK_CLASS_NAMES.add(cls.__name__) + + if dataclass is not None and not issubclass(dataclass, FairseqDataclass): + raise ValueError( + "Dataclass {} must extend FairseqDataclass".format(dataclass) + ) + + cls.__dataclass = dataclass + if dataclass is not None: + TASK_DATACLASS_REGISTRY[name] = dataclass + + cs = ConfigStore.instance() + node = dataclass() + node._name = name + cs.store(name=name, group="task", node=node, provider="fairseq") + + return cls + + return register_task_cls + + +def get_task(name): + return TASK_REGISTRY[name] + + +def import_tasks(tasks_dir, namespace): + for file in os.listdir(tasks_dir): + path = os.path.join(tasks_dir, file) + if ( + not file.startswith("_") + and not file.startswith(".") + and (file.endswith(".py") or os.path.isdir(path)) + ): + task_name = file[: file.find(".py")] if file.endswith(".py") else file + importlib.import_module(namespace + "." + task_name) + + # expose `task_parser` for sphinx + if task_name in TASK_REGISTRY: + parser = argparse.ArgumentParser(add_help=False) + group_task = parser.add_argument_group("Task name") + # fmt: off + group_task.add_argument('--task', metavar=task_name, + help='Enable this task with: ``--task=' + task_name + '``') + # fmt: on + group_args = parser.add_argument_group( + "Additional command-line arguments" + ) + TASK_REGISTRY[task_name].add_args(group_args) + globals()[task_name + "_parser"] = parser + + +# automatically import any Python files in the tasks/ directory +tasks_dir = os.path.dirname(__file__) +import_tasks(tasks_dir, "fairseq.tasks") diff --git a/fairseq/tasks/audio_pretraining.py b/fairseq/tasks/audio_pretraining.py new file mode 100644 index 0000000000000000000000000000000000000000..c642ff5226e5f98332b5b2ae90716842c2addacc --- /dev/null +++ b/fairseq/tasks/audio_pretraining.py @@ -0,0 +1,381 @@ +# Copyright (c) 2017-present, Facebook, Inc. +# All rights reserved. +# +# This source code is licensed under the license found in the LICENSE file in +# the root directory of this source tree. An additional grant of patent rights +# can be found in the PATENTS file in the same directory. + +import logging +import os +import sys +import torch + +from argparse import Namespace +from dataclasses import dataclass, field +from typing import Optional, Any +from omegaconf import MISSING, II, OmegaConf + +from fairseq.data import ( + AddTargetDataset, + BinarizedAudioDataset, + Dictionary, + FileAudioDataset, + encoders, +) +from fairseq.dataclass import FairseqDataclass +from fairseq.dataclass.configs import GenerationConfig + +from . import FairseqTask, register_task +from .. import utils +from ..logging import metrics + + +logger = logging.getLogger(__name__) + + +class LabelEncoder(object): + def __init__(self, dictionary): + self.dictionary = dictionary + + def __call__(self, label): + return self.dictionary.encode_line( + label, append_eos=False, add_if_not_exist=False + ) + + +@dataclass +class InferredW2vConfig: + # The following are needed to precompute mask and mask channel indices + # before model's forward. + mask_length: Optional[int] = II("model.mask_length") + mask_prob: Optional[float] = II("model.mask_prob") + mask_selection: Optional[str] = II("model.mask_selection") + mask_other: Optional[float] = II("model.mask_other") + no_mask_overlap: Optional[bool] = II("model.no_mask_overlap") + mask_min_space: Optional[int] = II("model.mask_min_space") + mask_channel_length: Optional[int] = II("model.mask_channel_length") + mask_channel_prob: Optional[float] = II("model.mask_channel_prob") + mask_channel_selection: Optional[str] = II("model.mask_channel_selection") + mask_channel_other: Optional[float] = II("model.mask_channel_other") + no_mask_channel_overlap: Optional[bool] = II("model.no_mask_channel_overlap") + mask_channel_min_space: Optional[int] = II("model.mask_channel_min_space") + + conv_feature_layers: Optional[str] = II("model.conv_feature_layers") + encoder_embed_dim: Optional[int] = II("model.encoder_embed_dim") + + +@dataclass +class AudioPretrainingConfig(FairseqDataclass): + data: str = field(default=MISSING, metadata={"help": "path to data directory"}) + labels: Optional[str] = field( + default=None, + metadata={"help": "extension of the label file to load, used for fine-tuning"}, + ) + binarized_dataset: bool = field( + default=False, + metadata={ + "help": "if true, loads binarized dataset (useful for very large datasets). " + "See examples/wav2vec/scripts/binarize_manifest.sh" + }, + ) + sample_rate: int = field( + default=16_000, + metadata={ + "help": "target sample rate. audio files will be up/down sampled to this rate" + }, + ) + normalize: bool = field( + default=False, + metadata={"help": "if set, normalizes input to have 0 mean and unit variance"}, + ) + enable_padding: bool = field( + default=False, metadata={"help": "pad shorter samples instead of cropping"} + ) + max_sample_size: Optional[int] = field( + default=None, metadata={"help": "max sample size to crop to for batching"} + ) + min_sample_size: Optional[int] = field( + default=None, metadata={"help": "min sample size to skip small examples"} + ) + + # Options for reporting WER metrics during validation. Only applicable to + # Seq2Seq models during fine-tuning + eval_wer: bool = field( + default=False, metadata={"help": "compute WER for Seq2Seq models"} + ) + eval_wer_config: GenerationConfig = field( + default_factory=lambda: GenerationConfig(), + metadata={"help": "beam search config for evaluating wer during training"}, + ) + eval_wer_tokenizer: Any = field( + default=None, + metadata={"help": "tokenizer config for evaluating wer during training"}, + ) + eval_wer_post_process: str = field( + default="letter", + metadata={ + "help": "remove BPE tokens before scoring (can be sentencepiece, letter, and more)" + }, + ) + autoregressive: bool = field( + default=False, + metadata={ + "help": "required for autoregressive decoders (like seq2seq models); " + "adds 'prev_output_tokens' to input and appends eos to target" + }, + ) + num_batch_buckets: int = field( + default=0, + metadata={"help": "number of buckets"}, + ) + precompute_mask_indices: bool = field( + default=False, + metadata={ + "help": "flag to compute mask indices in data preparation.", + }, + ) + + inferred_w2v_config: Optional[InferredW2vConfig] = field( + default=None, + metadata={ + "help": "wav2vec 2.0 masking arguments used to pre-compute masks (required for TPU)", + }, + ) + + tpu: bool = II("common.tpu") + + +@register_task("audio_pretraining", dataclass=AudioPretrainingConfig) +class AudioPretrainingTask(FairseqTask): + """ """ + + cfg: AudioPretrainingConfig + + def __init__( + self, + cfg: AudioPretrainingConfig, + ): + super().__init__(cfg) + if cfg.eval_wer: + assert cfg.labels is not None, "eval_wer can only be set during fine-tuning" + self.blank_symbol = "<s>" + + self.state.add_factory("target_dictionary", self.load_target_dictionary) + + @classmethod + def setup_task(cls, cfg: AudioPretrainingConfig, **kwargs): + """Setup the task (e.g., load dictionaries). + + Args: + cfg (AudioPretrainingConfig): configuration of this task + """ + + return cls(cfg) + + def load_target_dictionary(self): + if self.cfg.labels: + dict_path = os.path.join(self.cfg.data, f"dict.{self.cfg.labels}.txt") + return Dictionary.load(dict_path) + return None + + def _get_mask_precompute_kwargs(self, cfg): + if self.cfg.precompute_mask_indices or self.cfg.tpu: + assert ( + cfg.inferred_w2v_config is not None + ), "inferred_w2v_config must be set" + return OmegaConf.to_container( + cfg.inferred_w2v_config, resolve=True, enum_to_str=True + ) + else: + return {} + + def load_dataset(self, split: str, task_cfg: FairseqDataclass = None, **kwargs): + data_path = self.cfg.data + task_cfg = task_cfg or self.cfg + + # upgrade old task + if isinstance(task_cfg, Namespace): + if not hasattr(task_cfg, "autoregressive"): + task_cfg.autoregressive = not task_cfg.criterion == "ctc" + + if getattr(task_cfg, "binarized_dataset", False): + self.datasets[split] = BinarizedAudioDataset( + data_path, + split=split, + sample_rate=task_cfg.get("sample_rate", self.cfg.sample_rate), + max_sample_size=self.cfg.max_sample_size, + min_sample_size=self.cfg.min_sample_size, + pad=task_cfg.labels is not None or task_cfg.enable_padding, + normalize=task_cfg.normalize, + num_buckets=self.cfg.num_batch_buckets or int(self.cfg.tpu), + compute_mask_indices=(self.cfg.precompute_mask_indices or self.cfg.tpu), + **self._get_mask_precompute_kwargs(task_cfg), + ) + else: + manifest_path = os.path.join(data_path, "{}.tsv".format(split)) + + self.datasets[split] = FileAudioDataset( + manifest_path=manifest_path, + sample_rate=task_cfg.get("sample_rate", self.cfg.sample_rate), + max_sample_size=self.cfg.max_sample_size, + min_sample_size=self.cfg.min_sample_size, + pad=task_cfg.labels is not None or task_cfg.enable_padding, + normalize=task_cfg.normalize, + num_buckets=self.cfg.num_batch_buckets or int(self.cfg.tpu), + compute_mask_indices=(self.cfg.precompute_mask_indices or self.cfg.tpu), + **self._get_mask_precompute_kwargs(task_cfg), + ) + + if self.cfg.tpu and task_cfg["mask_channel_prob"] == 0.0: + logger.info( + "Pretraining on TPUs may suffer convergence " + "issues when training with `mask_channel_prob` value of " + "0. You may want to set this to a low value close to 0." + ) + + if task_cfg.labels: + label_path = os.path.join(data_path, f"{split}.{task_cfg.labels}") + skipped_indices = getattr(self.datasets[split], "skipped_indices", set()) + with open(label_path, "r") as f: + labels = [line for i, line in enumerate(f) if i not in skipped_indices] + + assert len(labels) == len(self.datasets[split]), ( + f"labels length ({len(labels)}) and dataset length " + f"({len(self.datasets[split])}) do not match" + ) + + process_label = LabelEncoder(self.target_dictionary) + + self.datasets[split] = AddTargetDataset( + self.datasets[split], + labels, + pad=self.target_dictionary.pad(), + eos=self.target_dictionary.eos(), + batch_targets=True, + process_label=process_label, + add_to_input=task_cfg.get("autoregressive", False), + ) + + @property + def source_dictionary(self): + return None + + @property + def target_dictionary(self): + """Return the :class:`~fairseq.data.Dictionary` for the language + model.""" + return self.state.target_dictionary + + def max_positions(self): + """Maximum input length supported by the encoder.""" + return (sys.maxsize, sys.maxsize) + + def filter_indices_by_size( + self, + indices, + dataset, + max_positions=None, + ignore_invalid_inputs=False, + ): + # we do not need to filter by size in this task as dataloaders take care of this + return indices + + def valid_step(self, sample, model, criterion): + loss, sample_size, logging_output = super().valid_step(sample, model, criterion) + if self.cfg.eval_wer and self.cfg.autoregressive: + metrics = self._inference_with_wer(self.sequence_generator, sample, model) + logging_output["_num_char_errors"] = metrics["num_char_errors"] + logging_output["_num_chars"] = metrics["num_chars"] + logging_output["_num_word_errors"] = metrics["num_word_errors"] + logging_output["_num_words"] = metrics["num_words"] + return loss, sample_size, logging_output + + def build_model(self, model_cfg: FairseqDataclass): + model = super().build_model(model_cfg) + + if self.cfg.eval_wer and self.cfg.autoregressive: + self.sequence_generator = self.build_generator( + [model], + self.cfg.eval_wer_config, + ) + if self.cfg.eval_wer_tokenizer: + self.tokenizer = encoders.build_tokenizer(self.cfg.eval_wer_tokenizer) + else: + self.tokenizer = None + + actualized_cfg = getattr(model, "cfg", None) + if actualized_cfg is not None: + if "w2v_args" in actualized_cfg: + model_cfg.w2v_args = actualized_cfg.w2v_args + + return model + + def _inference_with_wer(self, generator, sample, model): + import editdistance + + def decode(toks): + s = self.target_dictionary.string( + toks.int().cpu(), + self.cfg.eval_wer_post_process, + escape_unk=True, + ) + if self.tokenizer: + s = self.tokenizer.decode(s) + return s + + num_word_errors, num_char_errors = 0, 0 + num_chars, num_words = 0, 0 + gen_out = self.inference_step(generator, [model], sample, None) + for i in range(len(gen_out)): + hyp = decode(gen_out[i][0]["tokens"]) + ref = decode( + utils.strip_pad(sample["target"][i], self.target_dictionary.pad()), + ) + num_char_errors += editdistance.eval(hyp, ref) + num_chars += len(ref) + hyp_words = hyp.split() + ref_words = ref.split() + num_word_errors += editdistance.eval(hyp_words, ref_words) + num_words += len(ref_words) + + return { + "num_char_errors": num_char_errors, + "num_chars": num_chars, + "num_word_errors": num_word_errors, + "num_words": num_words, + } + + def reduce_metrics(self, logging_outputs, criterion): + super().reduce_metrics(logging_outputs, criterion) + + zero = torch.scalar_tensor(0.0) + num_char_errors = sum( + log.get("_num_char_errors", zero) for log in logging_outputs + ) + num_chars = sum(log.get("_num_chars", zero) for log in logging_outputs) + num_word_errors = sum( + log.get("_num_word_errors", zero) for log in logging_outputs + ) + num_words = sum(log.get("_num_words", zero) for log in logging_outputs) + metrics.log_scalar("_num_char_errors", num_char_errors) + metrics.log_scalar("_num_chars", num_chars) + metrics.log_scalar("_num_word_errors", num_word_errors) + metrics.log_scalar("_num_words", num_words) + if num_chars > 0: + metrics.log_derived( + "uer", + lambda meters: meters["_num_char_errors"].sum + * 100.0 + / meters["_num_chars"].sum + if meters["_num_chars"].sum > 0 + else float("nan"), + ) + if num_words > 0: + metrics.log_derived( + "wer", + lambda meters: meters["_num_word_errors"].sum + * 100.0 + / meters["_num_words"].sum + if meters["_num_words"].sum > 0 + else float("nan"), + ) diff --git a/fairseq/tasks/cross_lingual_lm.py b/fairseq/tasks/cross_lingual_lm.py new file mode 100644 index 0000000000000000000000000000000000000000..8f8fe7e2de181e41bd0e6a2bf96948ee78de5ae8 --- /dev/null +++ b/fairseq/tasks/cross_lingual_lm.py @@ -0,0 +1,191 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import itertools +import logging +import os +from collections import OrderedDict + +import numpy as np +from fairseq import tokenizer, utils +from fairseq.data import ConcatDataset, Dictionary, TokenBlockDataset, data_utils +from fairseq.data.legacy.masked_lm_dataset import MaskedLMDataset +from fairseq.data.legacy.masked_lm_dictionary import MaskedLMDictionary +from fairseq.data.multi_corpus_sampled_dataset import MultiCorpusSampledDataset +from fairseq.tasks import LegacyFairseqTask, register_task + + +logger = logging.getLogger(__name__) + + +@register_task("cross_lingual_lm") +class CrossLingualLMTask(LegacyFairseqTask): + """ + Task for training cross-lingual language models. + + For more details look at: https://arxiv.org/pdf/1901.07291.pdf + + Args: + dictionary (Dictionary): the dictionary for the input of the task + """ + + @staticmethod + def add_args(parser): + """Add task-specific arguments to the parser.""" + parser.add_argument( + "data", + help="colon separated path to data directories list, \ + will be iterated upon during epochs in round-robin manner", + ) + parser.add_argument( + "--tokens-per-sample", + default=512, + type=int, + help="max number of total tokens over all segments" " per sample", + ) + parser.add_argument( + "--monolingual-langs", + default="en", + type=str, + help="comma separated list of languages for which we" + " want to train XLM on", + ) + parser.add_argument( + "--shuffle", + action="store_true", + help="shuffle each monolingual dataset while" " training", + ) + + def __init__(self, args, dictionary): + super().__init__(args) + self.dictionary = dictionary + self.seed = args.seed + self.distributed_world_size = args.distributed_world_size + self.langs2id = self._lang_to_id(args.monolingual_langs) + + def _lang_to_id(self, languages: str): + """ + Build a map from languages to ids. These ids are used as segment labels + for cross-lingual LM training. + """ + lang2id = {} + langs = [l.strip() for l in languages.split(",")] + for id, lang in enumerate(langs): + lang2id[lang] = id + return lang2id + + @classmethod + def load_dictionary(cls, filename): + return MaskedLMDictionary.load(filename) + + @classmethod + def build_dictionary( + cls, filenames, workers=1, threshold=-1, nwords=-1, padding_factor=8 + ): + d = MaskedLMDictionary() + for filename in filenames: + Dictionary.add_file_to_dictionary( + filename, d, tokenizer.tokenize_line, workers + ) + d.finalize(threshold=threshold, nwords=nwords, padding_factor=padding_factor) + return d + + @property + def target_dictionary(self): + return self.dictionary + + @classmethod + def setup_task(cls, args, **kwargs): + """Setup the task.""" + dictionary = MaskedLMDictionary.load(os.path.join(args.data, "dict.txt")) + logger.info("dictionary: {} types".format(len(dictionary))) + return cls(args, dictionary) + + def _load_single_lang_dataset(self, split, epoch): + loaded_datasets = [] + + paths = utils.split_paths(self.args.data) + assert len(paths) > 0 + data_path = paths[(epoch - 1) % len(paths)] + + for k in itertools.count(): + split_k = split + (str(k) if k > 0 else "") + path = os.path.join(data_path, split_k) + + ds = data_utils.load_indexed_dataset( + path, self.dictionary, self.args.dataset_impl + ) + if ds is None: + if k > 0: + break + else: + raise FileNotFoundError( + "Dataset not found: {} ({})".format(split, data_path) + ) + + # Since we append each block with the classification_token, + # we need to effectively create blocks of length + # tokens_per_sample-1 + loaded_datasets.append( + TokenBlockDataset( + ds, + ds.sizes, + self.args.tokens_per_sample - 1, + pad=self.dictionary.pad(), + eos=self.dictionary.eos(), + ) + ) + + logger.info( + "{} {} {} examples".format(data_path, split_k, len(loaded_datasets[-1])) + ) + + if len(loaded_datasets) == 1: + dataset = loaded_datasets[0] + sizes = dataset.sizes + else: + dataset = ConcatDataset(loaded_datasets) + sizes = np.concatenate([ds.sizes for ds in loaded_datasets]) + + return dataset, sizes + + def load_dataset(self, split, epoch=1, combine=False, **kwargs): + """Load a given dataset split. + + Args: + split (str): name of the split (e.g., train, valid, test) + """ + dataset_map = OrderedDict() + + for lang in self.langs2id.keys(): + # Datasets are expected to be in "split.lang" format (Eg: train.en) + language_split = "{}.{}".format(split, lang) + + block_dataset, sizes = self._load_single_lang_dataset( + split=language_split, epoch=epoch + ) + + dataset_map[lang] = MaskedLMDataset( + dataset=block_dataset, + sizes=sizes, + vocab=self.dictionary, + pad_idx=self.dictionary.pad(), + mask_idx=self.dictionary.mask(), + classif_token_idx=self.dictionary.eos(), + sep_token_idx=self.dictionary.eos(), + shuffle=getattr(self.args, "shuffle", False), + has_pairs=False, + segment_id=self.langs2id[lang], + seed=self.seed, + ) + + self.datasets[split] = MultiCorpusSampledDataset(dataset_map) + logger.info( + "{} {} {} examples".format( + utils.split_paths(self.args.data)[epoch - 1], + split, + len(self.datasets[split]), + ) + ) diff --git a/fairseq/tasks/denoising.py b/fairseq/tasks/denoising.py new file mode 100644 index 0000000000000000000000000000000000000000..cbf01e14dfad17ee8ab0ae1ca67c2458b84559cb --- /dev/null +++ b/fairseq/tasks/denoising.py @@ -0,0 +1,274 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import os + +from fairseq import utils +from fairseq.data import ( + AppendTokenDataset, + DenoisingDataset, + Dictionary, + IdDataset, + NestedDictionaryDataset, + NumelDataset, + PadDataset, + PrependTokenDataset, + StripTokenDataset, + TokenBlockDataset, + data_utils, +) +from fairseq.data.encoders.utils import get_whole_word_mask +from fairseq.data.shorten_dataset import maybe_shorten_dataset +from fairseq.tasks import LegacyFairseqTask, register_task +import numpy as np + + +logger = logging.getLogger(__name__) + + +@register_task("denoising") +class DenoisingTask(LegacyFairseqTask): + """ + Denoising task for applying sequence to sequence denoising. (ie. BART) + """ + + @staticmethod + def add_args(parser): + """Add task-specific arguments to the parser.""" + parser.add_argument("data", help="path to data directory") + parser.add_argument( + "--tokens-per-sample", + default=512, + type=int, + help="max number of total tokens over all segments" + " per sample for dataset", + ) + parser.add_argument( + "--sample-break-mode", + default="complete_doc", + type=str, + help="mode for breaking sentence", + ) + parser.add_argument( + "--mask", + default=0.0, + type=float, + help="fraction of words/subwords that will be masked", + ) + parser.add_argument( + "--mask-random", + default=0.0, + type=float, + help="instead of using [MASK], use random token this often", + ) + parser.add_argument( + "--insert", + default=0.0, + type=float, + help="insert this percentage of additional random tokens", + ) + parser.add_argument( + "--permute", + default=0.0, + type=float, + help="take this proportion of subwords and permute them", + ) + parser.add_argument( + "--rotate", + default=0.5, + type=float, + help="rotate this proportion of inputs", + ) + parser.add_argument( + "--poisson-lambda", + default=3.0, + type=float, + help="randomly shuffle sentences for this proportion of inputs", + ) + parser.add_argument( + "--permute-sentences", + default=0.0, + type=float, + help="shuffle this proportion of sentences in all inputs", + ) + parser.add_argument( + "--mask-length", + default="subword", + type=str, + choices=["subword", "word", "span-poisson"], + help="mask length to choose", + ) + parser.add_argument( + "--replace-length", + default=-1, + type=int, + help="when masking N tokens, replace with 0, 1, or N tokens (use -1 for N)", + ) + parser.add_argument( + "--max-source-positions", + default=1024, + type=int, + metavar="N", + help="max number of tokens in the source sequence", + ) + parser.add_argument( + "--max-target-positions", + default=1024, + type=int, + metavar="N", + help="max number of tokens in the target sequence", + ) + + parser.add_argument( + "--shorten-method", + default="none", + choices=["none", "truncate", "random_crop"], + help="if not none, shorten sequences that exceed --tokens-per-sample", + ) + parser.add_argument( + "--shorten-data-split-list", + default="", + help="comma-separated list of dataset splits to apply shortening to, " + 'e.g., "train,valid" (default: all dataset splits)', + ) + + + def __init__(self, args, dictionary): + super().__init__(args) + self.dictionary = dictionary + self.seed = args.seed + + # add mask token + self.mask_idx = self.dictionary.add_symbol("<mask>") + + @classmethod + def setup_task(cls, args, **kwargs): + """Setup the task.""" + dictionary = Dictionary.load(os.path.join(args.data, "dict.txt")) + logger.info("dictionary: {} types".format(len(dictionary))) + if not hasattr(args, "shuffle_instance"): + args.shuffle_instance = False + return cls(args, dictionary) + + def load_dataset(self, split, epoch=1, combine=False, **kwargs): + """Load a given dataset split. + + Args: + split (str): name of the split (e.g., train, valid, test) + """ + paths = utils.split_paths(self.args.data) + assert len(paths) > 0 + data_path = paths[(epoch - 1) % len(paths)] + split_path = os.path.join(data_path, split) + + dataset = data_utils.load_indexed_dataset( + split_path, + self.dictionary, + self.args.dataset_impl, + combine=combine, + ) + if dataset is None: + raise FileNotFoundError( + "Dataset not found: {} ({})".format(split, split_path) + ) + + dataset = StripTokenDataset(dataset, self.dictionary.eos()) + + dataset = maybe_shorten_dataset( + dataset, + split, + self.args.shorten_data_split_list, + self.args.shorten_method, + self.args.tokens_per_sample, + self.args.seed, + ) + + # create continuous blocks of tokens + dataset = TokenBlockDataset( + dataset, + dataset.sizes, + self.args.tokens_per_sample - 2, # one less for <s> and one for </s> + pad=self.dictionary.pad(), + eos=self.dictionary.eos(), + break_mode=self.args.sample_break_mode, + document_sep_len=0, + ) + + # prepend beginning-of-sentence token (<s>, equiv. to [CLS] in BERT) + dataset = PrependTokenDataset(dataset, self.source_dictionary.bos()) + dataset = AppendTokenDataset(dataset, self.source_dictionary.eos()) + + mask_whole_words = ( + get_whole_word_mask(self.args, self.source_dictionary) + if self.args.mask_length != "subword" + else None + ) + + self.datasets[split] = DenoisingDataset( + dataset, + dataset.sizes, + self.dictionary, + self.mask_idx, + mask_whole_words, + shuffle=self.args.shuffle_instance, + seed=self.seed, + args=self.args, + ) + logger.info( + "Split: {0}, Loaded {1} samples of denoising_dataset".format( + split, + len(self.datasets[split]), + ) + ) + + def build_dataset_for_inference(self, src_tokens, src_lengths, **kwargs): + """ + Generate batches for inference. We assume that the input begins with a + bos symbol (`<s>`) and ends with an eos symbol (`</s>`). + """ + pad = self.source_dictionary.pad() + eos = self.source_dictionary.eos() + src_dataset = TokenBlockDataset( + src_tokens, + src_lengths, + block_size=self.args.tokens_per_sample - 2, # for <s> and </s> + pad=pad, + eos=eos, + break_mode=self.args.sample_break_mode, + document_sep_len=0, + ) + prev_output_tokens = PrependTokenDataset( + StripTokenDataset(src_dataset, eos), eos + ) + src_dataset = PadDataset(src_dataset, pad_idx=pad, left_pad=False) + return NestedDictionaryDataset( + { + "id": IdDataset(), + "net_input": { + "src_tokens": src_dataset, + "src_lengths": NumelDataset(src_dataset, reduce=False), + "prev_output_tokens": PadDataset( + prev_output_tokens, pad_idx=pad, left_pad=False + ), + }, + "target": src_dataset, + }, + sizes=[np.array(src_lengths)], + ) + + def max_positions(self): + """Return the max sentence length allowed by the task.""" + return (self.args.max_source_positions, self.args.max_target_positions) + + @property + def source_dictionary(self): + """Return the source :class:`~fairseq.data.Dictionary`.""" + return self.dictionary + + @property + def target_dictionary(self): + """Return the target :class:`~fairseq.data.Dictionary`.""" + return self.dictionary diff --git a/fairseq/tasks/fairseq_task.py b/fairseq/tasks/fairseq_task.py new file mode 100644 index 0000000000000000000000000000000000000000..fbec9bb2a557e97cb921b705846bde482d85f169 --- /dev/null +++ b/fairseq/tasks/fairseq_task.py @@ -0,0 +1,677 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import os +import warnings +from argparse import Namespace +from typing import Any, Callable, Dict, List + +import torch +from fairseq import metrics, search, tokenizer, utils +from fairseq.data import Dictionary, FairseqDataset, data_utils, encoders, iterators +from fairseq.dataclass import FairseqDataclass +from fairseq.dataclass.utils import gen_parser_from_dataclass +from fairseq.optim.amp_optimizer import AMPOptimizer +from omegaconf import DictConfig + + +logger = logging.getLogger(__name__) + + +class StatefulContainer(object): + + _state: Dict[str, Any] = dict() + _factories: Dict[str, Callable[[], Any]] = dict() + + def add_factory(self, name, factory: Callable[[], Any]): + self._factories[name] = factory + + def merge_state_dict(self, state_dict: Dict[str, Any]): + self._state.update(state_dict) + + @property + def state_dict(self) -> Dict[str, Any]: + return self._state + + def __getattr__(self, name): + if name not in self._state and name in self._factories: + self._state[name] = self._factories[name]() + + if name in self._state: + return self._state[name] + + raise AttributeError(f"Task state has no factory for attribute {name}") + + +class FairseqTask(object): + """ + Tasks store dictionaries and provide helpers for loading/iterating over + Datasets, initializing the Model/Criterion and calculating the loss. + + Tasks have limited statefulness. In particular, state that needs to be + saved to/loaded from checkpoints needs to be stored in the `self.state` + :class:`StatefulContainer` object. For example:: + + self.state.add_factory("dictionary", self.load_dictionary) + print(self.state.dictionary) # calls self.load_dictionary() + + This is necessary so that when loading checkpoints, we can properly + recreate the task state after initializing the task instance. + """ + + @classmethod + def add_args(cls, parser): + """Add task-specific arguments to the parser.""" + dc = getattr(cls, "__dataclass", None) + if dc is not None: + gen_parser_from_dataclass(parser, dc()) + + @staticmethod + def logging_outputs_can_be_summed(criterion) -> bool: + """ + Whether the logging outputs returned by `train_step` and `valid_step` can + be summed across workers prior to calling `aggregate_logging_outputs`. + Setting this to True will improves distributed training speed. + """ + return criterion.logging_outputs_can_be_summed() + + cfg: FairseqDataclass + datasets: Dict[str, FairseqDataset] + dataset_to_epoch_iter: Dict[FairseqDataset, Any] + state: StatefulContainer = None + + def __init__(self, cfg: FairseqDataclass, **kwargs): + self.cfg = cfg + self.datasets = dict() + self.dataset_to_epoch_iter = dict() + self.state = StatefulContainer() + + @classmethod + def load_dictionary(cls, filename): + """Load the dictionary from the filename + + Args: + filename (str): the filename + """ + return Dictionary.load(filename) + + @classmethod + def build_dictionary( + cls, filenames, workers=1, threshold=-1, nwords=-1, padding_factor=8 + ): + """Build the dictionary + + Args: + filenames (list): list of filenames + workers (int): number of concurrent workers + threshold (int): defines the minimum word count + nwords (int): defines the total number of words in the final dictionary, + including special symbols + padding_factor (int): can be used to pad the dictionary size to be a + multiple of 8, which is important on some hardware (e.g., Nvidia + Tensor Cores). + """ + d = Dictionary() + for filename in filenames: + Dictionary.add_file_to_dictionary( + filename, d, tokenizer.tokenize_line, workers + ) + d.finalize(threshold=threshold, nwords=nwords, padding_factor=padding_factor) + return d + + @classmethod + def setup_task(cls, cfg: DictConfig, **kwargs): + """Setup the task (e.g., load dictionaries). + + Args: + cfg (omegaconf.DictConfig): parsed command-line arguments + """ + return cls(cfg, **kwargs) + + def has_sharded_data(self, split): + return os.pathsep in getattr(self.cfg, "data", "") + + def load_dataset( + self, + split: str, + combine: bool = False, + task_cfg: FairseqDataclass = None, + **kwargs + ): + """Load a given dataset split. + + Args: + split (str): name of the split (e.g., train, valid, test) + combine (bool): combines a split segmented into pieces into one dataset + task_cfg (FairseqDataclass): optional task configuration stored in the checkpoint that can be used + to load datasets + """ + raise NotImplementedError + + def dataset(self, split): + """ + Return a loaded dataset split. + + Args: + split (str): name of the split (e.g., train, valid, test) + + Returns: + a :class:`~fairseq.data.FairseqDataset` corresponding to *split* + """ + from fairseq.data import FairseqDataset + + if split not in self.datasets: + raise KeyError("Dataset not loaded: " + split) + if not isinstance(self.datasets[split], FairseqDataset): + raise TypeError("Datasets are expected to be of type FairseqDataset") + return self.datasets[split] + + def filter_indices_by_size( + self, indices, dataset, max_positions=None, ignore_invalid_inputs=False + ): + """ + Filter examples that are too large + + Args: + indices (np.array): original array of sample indices + dataset (~fairseq.data.FairseqDataset): dataset to batch + max_positions (optional): max sentence length supported by the + model (default: None). + ignore_invalid_inputs (bool, optional): don't raise Exception for + sentences that are too long (default: False). + Returns: + np.array: array of filtered sample indices + """ + indices, ignored = dataset.filter_indices_by_size(indices, max_positions) + if len(ignored) > 0: + if not ignore_invalid_inputs: + raise Exception( + ( + "Size of sample #{} is invalid (={}) since max_positions={}, " + "skip this example with --skip-invalid-size-inputs-valid-test" + ).format(ignored[0], dataset.size(ignored[0]), max_positions) + ) + logger.warning( + ( + "{:,} samples have invalid sizes and will be skipped, " + "max_positions={}, first few sample ids={}" + ).format(len(ignored), max_positions, ignored[:10]) + ) + return indices + + def can_reuse_epoch_itr(self, dataset): + # We can reuse the epoch iterator across epochs as long as the dataset + # hasn't disabled it. We default to ``False`` here, although in practice + # this will be ``True`` for most datasets that inherit from + # ``FairseqDataset`` due to the base implementation there. + return getattr(dataset, "can_reuse_epoch_itr_across_epochs", False) + + def get_batch_iterator( + self, + dataset, + max_tokens=None, + max_sentences=None, + max_positions=None, + ignore_invalid_inputs=False, + required_batch_size_multiple=1, + seed=1, + num_shards=1, + shard_id=0, + num_workers=0, + epoch=1, + data_buffer_size=0, + disable_iterator_cache=False, + ): + """ + Get an iterator that yields batches of data from the given dataset. + + Args: + dataset (~fairseq.data.FairseqDataset): dataset to batch + max_tokens (int, optional): max number of tokens in each batch + (default: None). + max_sentences (int, optional): max number of sentences in each + batch (default: None). + max_positions (optional): max sentence length supported by the + model (default: None). + ignore_invalid_inputs (bool, optional): don't raise Exception for + sentences that are too long (default: False). + required_batch_size_multiple (int, optional): require batch size to + be a multiple of N (default: 1). + seed (int, optional): seed for random number generator for + reproducibility (default: 1). + num_shards (int, optional): shard the data iterator into N + shards (default: 1). + shard_id (int, optional): which shard of the data iterator to + return (default: 0). + num_workers (int, optional): how many subprocesses to use for data + loading. 0 means the data will be loaded in the main process + (default: 0). + epoch (int, optional): the epoch to start the iterator from + (default: 1). + data_buffer_size (int, optional): number of batches to + preload (default: 0). + disable_iterator_cache (bool, optional): don't cache the + EpochBatchIterator (ignores `FairseqTask::can_reuse_epoch_itr`) + (default: False). + Returns: + ~fairseq.iterators.EpochBatchIterator: a batched iterator over the + given dataset split + """ + can_reuse_epoch_itr = not disable_iterator_cache and self.can_reuse_epoch_itr( + dataset + ) + if can_reuse_epoch_itr and dataset in self.dataset_to_epoch_iter: + logger.debug("reusing EpochBatchIterator for epoch {}".format(epoch)) + return self.dataset_to_epoch_iter[dataset] + + assert isinstance(dataset, FairseqDataset) + + # initialize the dataset with the correct starting epoch + dataset.set_epoch(epoch) + + # get indices ordered by example size + with data_utils.numpy_seed(seed): + indices = dataset.ordered_indices() + + # filter examples that are too large + if max_positions is not None: + indices = self.filter_indices_by_size( + indices, dataset, max_positions, ignore_invalid_inputs + ) + + # create mini-batches with given size constraints + batch_sampler = dataset.batch_by_size( + indices, + max_tokens=max_tokens, + max_sentences=max_sentences, + required_batch_size_multiple=required_batch_size_multiple, + ) + + # return a reusable, sharded iterator + epoch_iter = iterators.EpochBatchIterator( + dataset=dataset, + collate_fn=dataset.collater, + batch_sampler=batch_sampler, + seed=seed, + num_shards=num_shards, + shard_id=shard_id, + num_workers=num_workers, + epoch=epoch, + buffer_size=data_buffer_size, + ) + + if can_reuse_epoch_itr: + self.dataset_to_epoch_iter[dataset] = epoch_iter + + return epoch_iter + + def build_model(self, cfg: FairseqDataclass): + """ + Build the :class:`~fairseq.models.BaseFairseqModel` instance for this + task. + + Args: + cfg (FairseqDataclass): configuration object + + Returns: + a :class:`~fairseq.models.BaseFairseqModel` instance + """ + from fairseq import models, quantization_utils + + model = models.build_model(cfg, self) + model = quantization_utils.quantize_model_scalar(model, cfg) + return model + + def build_criterion(self, cfg: DictConfig): + """ + Build the :class:`~fairseq.criterions.FairseqCriterion` instance for + this task. + + Args: + cfg (omegaconf.DictConfig): configration object + + Returns: + a :class:`~fairseq.criterions.FairseqCriterion` instance + """ + from fairseq import criterions + + return criterions.build_criterion(cfg, self) + + def build_generator( + self, models, args, seq_gen_cls=None, extra_gen_cls_kwargs=None, prefix_allowed_tokens_fn=None, + ): + """ + Build a :class:`~fairseq.SequenceGenerator` instance for this + task. + + Args: + models (List[~fairseq.models.FairseqModel]): ensemble of models + args (fairseq.dataclass.configs.GenerationConfig): + configuration object (dataclass) for generation + extra_gen_cls_kwargs (Dict[str, Any]): extra options to pass + through to SequenceGenerator + prefix_allowed_tokens_fn (Callable[[int, torch.Tensor], List[int]]): + If provided, this function constrains the beam search to + allowed tokens only at each step. The provided function + should take 2 arguments: the batch ID (`batch_id: int`) + and a unidimensional tensor of token ids (`inputs_ids: + torch.Tensor`). It has to return a `List[int]` with the + allowed tokens for the next generation step conditioned + on the previously generated tokens (`inputs_ids`) and + the batch ID (`batch_id`). This argument is useful for + constrained generation conditioned on the prefix, as + described in "Autoregressive Entity Retrieval" + (https://arxiv.org/abs/2010.00904) and + https://github.com/facebookresearch/GENRE. + """ + if getattr(args, "score_reference", False): + from fairseq.sequence_scorer import SequenceScorer + + return SequenceScorer( + self.target_dictionary, + compute_alignment=getattr(args, "print_alignment", False), + ) + + from fairseq.sequence_generator import ( + SequenceGenerator, + SequenceGeneratorWithAlignment, + ) + try: + from fairseq.fb_sequence_generator import FBSequenceGenerator + except ModuleNotFoundError: + pass + + # Choose search strategy. Defaults to Beam Search. + sampling = getattr(args, "sampling", False) + sampling_topk = getattr(args, "sampling_topk", -1) + sampling_topp = getattr(args, "sampling_topp", -1.0) + diverse_beam_groups = getattr(args, "diverse_beam_groups", -1) + diverse_beam_strength = getattr(args, "diverse_beam_strength", 0.5) + match_source_len = getattr(args, "match_source_len", False) + diversity_rate = getattr(args, "diversity_rate", -1) + constrained = getattr(args, "constraints", False) + if prefix_allowed_tokens_fn is None: + prefix_allowed_tokens_fn = getattr(args, "prefix_allowed_tokens_fn", None) + if ( + sum( + int(cond) + for cond in [ + sampling, + diverse_beam_groups > 0, + match_source_len, + diversity_rate > 0, + ] + ) + > 1 + ): + raise ValueError("Provided Search parameters are mutually exclusive.") + assert sampling_topk < 0 or sampling, "--sampling-topk requires --sampling" + assert sampling_topp < 0 or sampling, "--sampling-topp requires --sampling" + + if sampling: + search_strategy = search.Sampling( + self.target_dictionary, sampling_topk, sampling_topp + ) + elif diverse_beam_groups > 0: + search_strategy = search.DiverseBeamSearch( + self.target_dictionary, diverse_beam_groups, diverse_beam_strength + ) + elif match_source_len: + # this is useful for tagging applications where the output + # length should match the input length, so we hardcode the + # length constraints for simplicity + search_strategy = search.LengthConstrainedBeamSearch( + self.target_dictionary, + min_len_a=1, + min_len_b=0, + max_len_a=1, + max_len_b=0, + ) + elif diversity_rate > -1: + search_strategy = search.DiverseSiblingsSearch( + self.target_dictionary, diversity_rate + ) + elif constrained: + search_strategy = search.LexicallyConstrainedBeamSearch( + self.target_dictionary, args.constraints + ) + elif prefix_allowed_tokens_fn: + search_strategy = search.PrefixConstrainedBeamSearch( + self.target_dictionary, prefix_allowed_tokens_fn + ) + else: + search_strategy = search.BeamSearch(self.target_dictionary) + + extra_gen_cls_kwargs = extra_gen_cls_kwargs or {} + if seq_gen_cls is None: + if getattr(args, "print_alignment", False): + seq_gen_cls = SequenceGeneratorWithAlignment + extra_gen_cls_kwargs["print_alignment"] = args.print_alignment + elif getattr(args, "fb_seq_gen", False): + seq_gen_cls = FBSequenceGenerator + else: + seq_gen_cls = SequenceGenerator + + return seq_gen_cls( + models, + self.target_dictionary, + beam_size=getattr(args, "beam", 5), + max_len_a=getattr(args, "max_len_a", 0), + max_len_b=getattr(args, "max_len_b", 200), + min_len=getattr(args, "min_len", 1), + normalize_scores=(not getattr(args, "unnormalized", False)), + len_penalty=getattr(args, "lenpen", 1), + unk_penalty=getattr(args, "unkpen", 0), + temperature=getattr(args, "temperature", 1.0), + match_source_len=getattr(args, "match_source_len", False), + no_repeat_ngram_size=getattr(args, "no_repeat_ngram_size", 0), + search_strategy=search_strategy, + **extra_gen_cls_kwargs, + ) + + def train_step( + self, sample, model, criterion, optimizer, update_num, ignore_grad=False + ): + """ + Do forward and backward, and return the loss as computed by *criterion* + for the given *model* and *sample*. + + Args: + sample (dict): the mini-batch. The format is defined by the + :class:`~fairseq.data.FairseqDataset`. + model (~fairseq.models.BaseFairseqModel): the model + criterion (~fairseq.criterions.FairseqCriterion): the criterion + optimizer (~fairseq.optim.FairseqOptimizer): the optimizer + update_num (int): the current update + ignore_grad (bool): multiply loss by 0 if this is set to True + + Returns: + tuple: + - the loss + - the sample size, which is used as the denominator for the + gradient + - logging outputs to display while training + """ + model.train() + model.set_num_updates(update_num) + with torch.autograd.profiler.record_function("forward"): + with torch.cuda.amp.autocast(enabled=(isinstance(optimizer, AMPOptimizer))): + loss, sample_size, logging_output = criterion(model, sample) + if ignore_grad: + loss *= 0 + with torch.autograd.profiler.record_function("backward"): + optimizer.backward(loss) + return loss, sample_size, logging_output + + def valid_step(self, sample, model, criterion): + model.eval() + with torch.no_grad(): + loss, sample_size, logging_output = criterion(model, sample) + return loss, sample_size, logging_output + + def optimizer_step(self, optimizer, model, update_num): + optimizer.step() + + def build_dataset_for_inference( + self, src_tokens: List[torch.Tensor], src_lengths: List[int], **kwargs + ) -> torch.utils.data.Dataset: + raise NotImplementedError + + def inference_step( + self, generator, models, sample, prefix_tokens=None, constraints=None + ): + with torch.no_grad(): + return generator.generate( + models, sample, prefix_tokens=prefix_tokens, constraints=constraints + ) + + def begin_epoch(self, epoch, model): + """Hook function called before the start of each epoch.""" + pass + + def begin_valid_epoch(self, epoch, model): + """Hook function called before the start of each validation epoch.""" + pass + + def aggregate_logging_outputs(self, logging_outputs, criterion): + """[deprecated] Aggregate logging outputs from data parallel training.""" + utils.deprecation_warning( + "The aggregate_logging_outputs API is deprecated. " + "Please use the reduce_metrics API instead." + ) + with metrics.aggregate() as agg: + self.reduce_metrics(logging_outputs, criterion) + return agg.get_smoothed_values() + + def reduce_metrics(self, logging_outputs, criterion): + """Aggregate logging outputs from data parallel training.""" + # backward compatibility for tasks that override aggregate_logging_outputs + base_func = FairseqTask.aggregate_logging_outputs + self_func = getattr(self, "aggregate_logging_outputs").__func__ + if self_func is not base_func: + utils.deprecation_warning( + "Tasks should implement the reduce_metrics API. " + "Falling back to deprecated aggregate_logging_outputs API." + ) + agg_logging_outputs = self.aggregate_logging_outputs( + logging_outputs, criterion + ) + for k, v in agg_logging_outputs.items(): + metrics.log_scalar(k, v) + return + + if not any("ntokens" in log for log in logging_outputs): + warnings.warn( + "ntokens not found in Criterion logging outputs, cannot log wpb or wps" + ) + else: + ntokens = sum(log.get("ntokens", 0) for log in logging_outputs) + metrics.log_scalar("wpb", ntokens, priority=180, round=1) + metrics.log_speed("wps", ntokens, priority=90, round=1) + + if not any("nsentences" in log for log in logging_outputs): + warnings.warn( + "nsentences not found in Criterion logging outputs, cannot log bsz" + ) + else: + nsentences = sum(log.get("nsentences", 0) for log in logging_outputs) + metrics.log_scalar("bsz", nsentences, priority=190, round=1) + + criterion.__class__.reduce_metrics(logging_outputs) + + def state_dict(self): + if self.state is not None: + return self.state.state_dict + return {} + + def load_state_dict(self, state_dict: Dict[str, Any]): + if self.state is not None: + self.state.merge_state_dict(state_dict) + + def max_positions(self): + """Return the max input length allowed by the task.""" + return None + + @property + def source_dictionary(self): + """Return the source :class:`~fairseq.data.Dictionary` (if applicable + for this task).""" + raise NotImplementedError + + @property + def target_dictionary(self): + """Return the target :class:`~fairseq.data.Dictionary` (if applicable + for this task).""" + raise NotImplementedError + + def build_tokenizer(self, args): + """Build the pre-tokenizer for this task.""" + return encoders.build_tokenizer(args) + + def build_bpe(self, args): + """Build the tokenizer for this task.""" + return encoders.build_bpe(args) + + def get_interactive_tokens_and_lengths(self, lines, encode_fn): + tokens = [ + self.source_dictionary.encode_line( + encode_fn(src_str), add_if_not_exist=False + ).long() + for src_str in lines + ] + lengths = [t.numel() for t in tokens] + return tokens, lengths + + +class LegacyFairseqTask(FairseqTask): + def __init__(self, args: Namespace): + self.args = args + self.datasets = {} + self.dataset_to_epoch_iter = {} + + @classmethod + def setup_task(cls, args: Namespace, **kwargs): + """Setup the task (e.g., load dictionaries). + + Args: + args (argparse.Namespace): parsed command-line arguments + """ + return cls(args, **kwargs) + + def has_sharded_data(self, split): + return os.pathsep in getattr(self.args, "data", "") + + def build_model(self, args: Namespace): + """ + Build the :class:`~fairseq.models.BaseFairseqModel` instance for this + task. + + Args: + args (argparse.Namespace): parsed command-line arguments + + Returns: + a :class:`~fairseq.models.BaseFairseqModel` instance + """ + from fairseq import models, quantization_utils + + model = models.build_model(args, self) + model = quantization_utils.quantize_model_scalar(model, args) + return model + + def build_criterion(self, args: Namespace): + """ + Build the :class:`~fairseq.criterions.FairseqCriterion` instance for + this task. + + Args: + args (argparse.Namespace): parsed command-line arguments + + Returns: + a :class:`~fairseq.criterions.FairseqCriterion` instance + """ + from fairseq import criterions + + return criterions.build_criterion(args, self) diff --git a/fairseq/tasks/hubert_pretraining.py b/fairseq/tasks/hubert_pretraining.py new file mode 100644 index 0000000000000000000000000000000000000000..a63f2f6ef8ff4953e6ca424bc4413fe55695b273 --- /dev/null +++ b/fairseq/tasks/hubert_pretraining.py @@ -0,0 +1,193 @@ +# Copyright (c) 2017-present, Facebook, Inc. +# All rights reserved. +# +# This source code is licensed under the license found in the LICENSE file in +# the root directory of this source tree. An additional grant of patent rights +# can be found in the PATENTS file in the same directory. + +import logging +import os +import sys +from typing import Dict, List, Optional, Tuple + +import numpy as np + +from dataclasses import dataclass, field +from fairseq.data import Dictionary, HubertDataset +from fairseq.dataclass.configs import FairseqDataclass +from fairseq.tasks import register_task +from fairseq.tasks.fairseq_task import FairseqTask +from omegaconf import MISSING + +logger = logging.getLogger(__name__) + + +class LabelEncoder(object): + def __init__(self, dictionary: Dictionary) -> None: + self.dictionary = dictionary + + def __call__(self, label: str) -> List[str]: + return self.dictionary.encode_line( + label, append_eos=False, add_if_not_exist=False, + ) + + +@dataclass +class HubertPretrainingConfig(FairseqDataclass): + data: str = field( + default=MISSING, metadata={"help": "path to data directory"} + ) + fine_tuning: bool = field( + default=False, metadata={"help": "set to true if fine-tuning Hubert"} + ) + labels: List[str] = field( + default_factory=lambda: ["ltr"], + metadata={ + "help": ( + "extension of the label files to load, frame-level labels for" + " pre-training, and sequence-level label for fine-tuning" + ) + }, + ) + label_dir: Optional[str] = field( + default=None, + metadata={ + "help": "if set, looks for labels in this directory instead", + }, + ) + label_rate: int = field( + default=-1, + metadata={"help": "label frame rate. -1 for sequence label"}, + ) + sample_rate: int = field( + default=16_000, + metadata={ + "help": "target sample rate. audio files will be up/down " + "sampled to this rate" + }, + ) + normalize: bool = field( + default=False, + metadata={ + "help": "if set, normalizes input to have 0 mean and unit variance" + }, + ) + enable_padding: bool = field( + default=False, + metadata={"help": "pad shorter samples instead of cropping"}, + ) + max_sample_size: Optional[int] = field( + default=None, + metadata={"help": "max sample size to crop to for batching"}, + ) + min_sample_size: Optional[int] = field( + default=None, + metadata={"help": "min sample size to crop to for batching"}, + ) + single_target: Optional[bool] = field( + default=False, + metadata={ + "help": "if set, AddTargetDatasets outputs same keys " + "as AddTargetDataset" + }, + ) + random_crop: Optional[bool] = field( + default=True, + metadata={"help": "always crop from the beginning if false"}, + ) + pad_audio: Optional[bool] = field( + default=False, + metadata={"help": "pad audio to the longest one in the batch if true"}, + ) + + +@register_task("hubert_pretraining", dataclass=HubertPretrainingConfig) +class HubertPretrainingTask(FairseqTask): + + cfg: HubertPretrainingConfig + + def __init__( + self, + cfg: HubertPretrainingConfig, + ) -> None: + super().__init__(cfg) + + logger.info(f"current directory is {os.getcwd()}") + logger.info(f"HubertPretrainingTask Config {cfg}") + + self.cfg = cfg + self.fine_tuning = cfg.fine_tuning + + if cfg.fine_tuning: + self.state.add_factory("target_dictionary", lambda: self.load_dictionaries) + else: + self.state.add_factory("dictionaries", lambda: self.load_dictionaries) + + self._source_dictionary = None + + self.blank_symbol = "<s>" + + @property + def source_dictionary(self) -> Optional[Dictionary]: + return self._source_dictionary + + @property + def target_dictionary(self) -> Optional[Dictionary]: + return self.state.target_dictionary + + @property + def dictionaries(self) -> List[Dictionary]: + return self.state.dictionaries + + @classmethod + def setup_task( + cls, cfg: HubertPretrainingConfig, **kwargs + ) -> "HubertPretrainingTask": + return cls(cfg) + + def load_dictionaries(self): + label_dir = self.cfg.data if self.cfg.label_dir is None else self.cfg.label_dir + dictionaries = [Dictionary.load(f"{label_dir}/dict.{label}.txt") for label in self.cfg.labels] + return dictionaries[0] if self.cfg.fine_tuning else dictionaries + + def get_label_dir(self) -> str: + if self.cfg.label_dir is None: + return self.cfg.data + return self.cfg.label_dir + + def load_dataset(self, split: str, **kwargs) -> None: + manifest = f"{self.cfg.data}/{split}.tsv" + dicts = [self.target_dictionary] if self.cfg.fine_tuning else self.dictionaries + pad_list = [dict.pad() for dict in dicts] + eos_list = [dict.eos() for dict in dicts] + procs = [LabelEncoder(dict) for dict in dicts] + paths = [ + f"{self.get_label_dir()}/{split}.{l}" for l in self.cfg.labels + ] + + # hubert v1: pad_audio=True, random_crop=False; + self.datasets[split] = HubertDataset( + manifest, + sample_rate=self.cfg.sample_rate, + label_paths=paths, + label_rates=self.cfg.label_rate, + pad_list=pad_list, + eos_list=eos_list, + label_processors=procs, + max_keep_sample_size=None, + min_keep_sample_size=self.cfg.min_sample_size, + max_sample_size=self.cfg.max_sample_size, + pad_audio=self.cfg.pad_audio, + normalize=self.cfg.normalize, + store_labels=False, + random_crop=self.cfg.random_crop, + single_target=self.cfg.single_target, + ) + + def max_positions(self) -> Tuple[int, int]: + return (sys.maxsize, sys.maxsize) + + def filter_indices_by_size( + self, indices: np.array, *args, **kwargs + ) -> np.array: + return indices diff --git a/fairseq/tasks/language_modeling.py b/fairseq/tasks/language_modeling.py new file mode 100644 index 0000000000000000000000000000000000000000..4b76a51c61d71c4358de07bdd4eb3f93894737a8 --- /dev/null +++ b/fairseq/tasks/language_modeling.py @@ -0,0 +1,379 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import os +from dataclasses import dataclass, field +from typing import Optional + +import numpy as np +import torch +from fairseq import utils +from fairseq.data import ( + AppendTokenDataset, + Dictionary, + IdDataset, + LMContextWindowDataset, + MonolingualDataset, + NestedDictionaryDataset, + NumelDataset, + PadDataset, + PrependTokenDataset, + StripTokenDataset, + TokenBlockDataset, + TruncatedDictionary, + data_utils, +) +from fairseq.data.indexed_dataset import get_available_dataset_impl +from fairseq.data.shorten_dataset import maybe_shorten_dataset +from fairseq.dataclass import ChoiceEnum, FairseqDataclass +from fairseq.tasks import LegacyFairseqTask, register_task +from omegaconf import II + + +SAMPLE_BREAK_MODE_CHOICES = ChoiceEnum(["none", "complete", "complete_doc", "eos"]) +SHORTEN_METHOD_CHOICES = ChoiceEnum(["none", "truncate", "random_crop"]) +logger = logging.getLogger(__name__) + + +@dataclass +class LanguageModelingConfig(FairseqDataclass): + data: Optional[str] = field( + default=None, metadata={"help": "path to data directory"} + ) + sample_break_mode: SAMPLE_BREAK_MODE_CHOICES = field( + default="none", + metadata={ + "help": 'If omitted or "none", fills each sample with tokens-per-sample ' + 'tokens. If set to "complete", splits samples only at the end ' + "of sentence, but may include multiple sentences per sample. " + '"complete_doc" is similar but respects doc boundaries. ' + 'If set to "eos", includes only one sentence per sample.' + }, + ) + tokens_per_sample: int = field( + default=1024, + metadata={"help": "max number of tokens per sample for LM dataset"}, + ) + output_dictionary_size: int = field( + default=-1, metadata={"help": "limit the size of output dictionary"} + ) + self_target: bool = field(default=False, metadata={"help": "include self target"}) + future_target: bool = field( + default=False, metadata={"help": "include future target"} + ) + past_target: bool = field(default=False, metadata={"help": "include past target"}) + add_bos_token: bool = field( + default=False, metadata={"help": "prepend beginning of sentence token (<s>)"} + ) + max_target_positions: Optional[int] = field( + default=None, metadata={"help": "max number of tokens in the target sequence"} + ) + shorten_method: SHORTEN_METHOD_CHOICES = field( + default="none", + metadata={ + "help": "if not none, shorten sequences that exceed --tokens-per-sample" + }, + ) + shorten_data_split_list: str = field( + default="", + metadata={ + "help": "comma-separated list of dataset splits to apply shortening to, " + 'e.g., "train,valid" (default: all dataset splits)' + }, + ) + pad_to_fixed_length: Optional[bool] = field( + default=False, metadata={"help": "pad to fixed length"}, + ) + pad_to_fixed_bsz: Optional[bool] = field( + default=False, metadata={"help": "boolean to pad to fixed batch size"}, + ) + + # TODO common vars below add to parent + seed: int = II("common.seed") + batch_size: Optional[int] = II("dataset.batch_size") + batch_size_valid: Optional[int] = II("dataset.batch_size_valid") + dataset_impl: Optional[ChoiceEnum(get_available_dataset_impl())] = II( + "dataset.dataset_impl" + ) + data_buffer_size: int = II("dataset.data_buffer_size") + tpu: bool = II("common.tpu") + use_plasma_view: bool = II("common.use_plasma_view") + plasma_path: str = II("common.plasma_path") + + +@register_task("language_modeling", dataclass=LanguageModelingConfig) +class LanguageModelingTask(LegacyFairseqTask): + """ + Train a language model. + + Args: + dictionary (~fairseq.data.Dictionary): the dictionary for the input of + the language model + output_dictionary (~fairseq.data.Dictionary): the dictionary for the + output of the language model. In most cases it will be the same as + *dictionary*, but could possibly be a more limited version of the + dictionary (if ``--output-dictionary-size`` is used). + targets (List[str]): list of the target types that the language model + should predict. Can be one of "self", "future", and "past". + Defaults to "future". + + .. note:: + + The language modeling task is compatible with :mod:`fairseq-train`, + :mod:`fairseq-generate`, :mod:`fairseq-interactive` and + :mod:`fairseq-eval-lm`. + + The language modeling task provides the following additional command-line + arguments: + + .. argparse:: + :ref: fairseq.tasks.language_modeling_parser + :prog: + """ + + def __init__(self, args, dictionary, output_dictionary=None, targets=None): + super().__init__(args) + self.dictionary = dictionary + self.output_dictionary = output_dictionary or dictionary + + if targets is None: + targets = ["future"] + self.targets = targets + + @classmethod + def setup_dictionary(cls, args, **kwargs): + dictionary = None + output_dictionary = None + if args.data: + paths = utils.split_paths(args.data) + assert len(paths) > 0 + dictionary = Dictionary.load(os.path.join(paths[0], "dict.txt")) + logger.info("dictionary: {} types".format(len(dictionary))) + output_dictionary = dictionary + if args.output_dictionary_size >= 0: + output_dictionary = TruncatedDictionary( + dictionary, args.output_dictionary_size + ) + return (dictionary, output_dictionary) + + @classmethod + def setup_task(cls, args, **kwargs): + """Setup the task (e.g., load dictionaries). + + Args: + args (argparse.Namespace): parsed command-line arguments + """ + dictionary, output_dictionary = cls.setup_dictionary(args, **kwargs) + + # upgrade old checkpoints + if getattr(args, "exclude_self_target", False): + args.self_target = False + + targets = [] + if getattr(args, "self_target", False): + targets.append("self") + if getattr(args, "future_target", False): + targets.append("future") + if getattr(args, "past_target", False): + targets.append("past") + if len(targets) == 0: + # standard language modeling + targets = ["future"] + + return cls(args, dictionary, output_dictionary, targets=targets) + + def build_model(self, args): + model = super().build_model(args) + for target in self.targets: + if target not in model.supported_targets: + raise ValueError( + "Unsupported language modeling target: {}".format(target) + ) + + return model + + def load_dataset( + self, split: str, epoch=1, combine=False, **kwargs + ) -> MonolingualDataset: + """Load a given dataset split. + + Args: + split (str): name of the split (e.g., train, valid, valid1, test) + """ + paths = utils.split_paths(self.args.data) + assert len(paths) > 0 + + data_path = paths[(epoch - 1) % len(paths)] + split_path = os.path.join(data_path, split) + + # each process has its own copy of the raw data (likely to be an np.memmap) + dataset = data_utils.load_indexed_dataset( + split_path, self.dictionary, self.args.dataset_impl, combine=combine + ) + if dataset is None: + raise FileNotFoundError(f"Dataset not found: {split} ({split_path})") + + dataset = maybe_shorten_dataset( + dataset, + split, + self.args.shorten_data_split_list, + self.args.shorten_method, + self.args.tokens_per_sample, + self.args.seed, + ) + dataset = TokenBlockDataset( + dataset, + dataset.sizes, + self.args.tokens_per_sample, + pad=self.dictionary.pad(), + eos=self.dictionary.eos(), + break_mode=self.args.sample_break_mode, + include_targets=True, + use_plasma_view=self.args.use_plasma_view, + split_path=split_path, + plasma_path=self.args.plasma_path, + ) + + add_eos_for_other_targets = ( + self.args.sample_break_mode is not None + and self.args.sample_break_mode != "none" + ) + fixed_pad_length = None + if self.args.pad_to_fixed_length: + fixed_pad_length = self.args.tokens_per_sample + + pad_to_bsz = None + if self.args.pad_to_fixed_bsz: + pad_to_bsz = self.args.batch_size_valid if 'valid' in split else self.args.batch_size + + self.datasets[split] = MonolingualDataset( + dataset=dataset, + sizes=dataset.sizes, + src_vocab=self.dictionary, + tgt_vocab=self.output_dictionary, + add_eos_for_other_targets=add_eos_for_other_targets, + shuffle=True, + targets=self.targets, + add_bos_token=self.args.add_bos_token, + fixed_pad_length=fixed_pad_length, + pad_to_bsz=pad_to_bsz, + ) + + def build_dataset_for_inference(self, src_tokens, src_lengths, **kwargs): + """ + Generate batches for inference. We prepend an eos token to src_tokens + (or bos if `--add-bos-token` is set) and we append a <pad> to target. + This is convenient both for generation with a prefix and LM scoring. + """ + dataset = StripTokenDataset( + TokenBlockDataset( + src_tokens, + src_lengths, + block_size=None, # ignored for "eos" break mode + pad=self.source_dictionary.pad(), + eos=self.source_dictionary.eos(), + break_mode="eos", + ), + # remove eos from (end of) target sequence + self.source_dictionary.eos(), + ) + src_dataset = PrependTokenDataset( + dataset, + token=( + self.source_dictionary.bos() + if getattr(self.args, "add_bos_token", False) + else self.source_dictionary.eos() + ), + ) + tgt_dataset = AppendTokenDataset(dataset, token=self.source_dictionary.pad()) + return NestedDictionaryDataset( + { + "id": IdDataset(), + "net_input": { + "src_tokens": PadDataset( + src_dataset, + pad_idx=self.source_dictionary.pad(), + left_pad=False, + ), + "src_lengths": NumelDataset(src_dataset, reduce=False), + }, + "target": PadDataset( + tgt_dataset, pad_idx=self.source_dictionary.pad(), left_pad=False + ), + }, + sizes=[np.array(src_lengths)], + ) + + def inference_step( + self, generator, models, sample, prefix_tokens=None, constraints=None + ): + with torch.no_grad(): + # Generation will always be conditioned on bos_token + if getattr(self.args, "add_bos_token", False): + bos_token = self.source_dictionary.bos() + else: + bos_token = self.source_dictionary.eos() + + if constraints is not None: + raise NotImplementedError( + "Constrained decoding with the language_modeling task is not supported" + ) + + # SequenceGenerator doesn't use src_tokens directly, we need to + # pass the `prefix_tokens` argument instead + if prefix_tokens is None and sample["net_input"]["src_tokens"].nelement(): + prefix_tokens = sample["net_input"]["src_tokens"] + if prefix_tokens[:, 0].eq(bos_token).all(): + prefix_tokens = prefix_tokens[:, 1:] + + return generator.generate( + models, sample, prefix_tokens=prefix_tokens, bos_token=bos_token + ) + + def eval_lm_dataloader( + self, + dataset, + max_tokens: Optional[int] = 36000, + batch_size: Optional[int] = None, + max_positions: Optional[int] = None, + num_shards: int = 1, + shard_id: int = 0, + num_workers: int = 1, + data_buffer_size: int = 10, + # ensures that every evaluated token has access to a context of at least + # this size, if possible + context_window: int = 0, + ): + if context_window > 0: + dataset = LMContextWindowDataset( + dataset=dataset, + tokens_per_sample=self.args.tokens_per_sample, + context_window=context_window, + pad_idx=self.source_dictionary.pad(), + ) + return self.get_batch_iterator( + dataset=dataset, + max_tokens=max_tokens, + max_sentences=batch_size, + max_positions=max_positions, + ignore_invalid_inputs=True, + num_shards=num_shards, + shard_id=shard_id, + num_workers=num_workers, + data_buffer_size=data_buffer_size, + ).next_epoch_itr(shuffle=False) + + @property + def source_dictionary(self): + """Return the :class:`~fairseq.data.Dictionary` for the language + model.""" + return self.dictionary + + @property + def target_dictionary(self): + """Return the :class:`~fairseq.data.Dictionary` for the language + model.""" + return self.output_dictionary diff --git a/fairseq/tasks/legacy_masked_lm.py b/fairseq/tasks/legacy_masked_lm.py new file mode 100644 index 0000000000000000000000000000000000000000..975497654926b64fff6c4960f54c4e6932e7fce1 --- /dev/null +++ b/fairseq/tasks/legacy_masked_lm.py @@ -0,0 +1,152 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import itertools +import logging +import os + +import numpy as np +from fairseq import tokenizer, utils +from fairseq.data import ConcatDataset, Dictionary, data_utils, indexed_dataset +from fairseq.data.legacy.block_pair_dataset import BlockPairDataset +from fairseq.data.legacy.masked_lm_dataset import MaskedLMDataset +from fairseq.data.legacy.masked_lm_dictionary import BertDictionary +from fairseq.tasks import LegacyFairseqTask, register_task + + +logger = logging.getLogger(__name__) + + +@register_task("legacy_masked_lm") +class LegacyMaskedLMTask(LegacyFairseqTask): + """ + Task for training Masked LM (BERT) model. + Args: + dictionary (Dictionary): the dictionary for the input of the task + """ + + @staticmethod + def add_args(parser): + """Add task-specific arguments to the parser.""" + parser.add_argument( + "data", + help="colon separated path to data directories list, \ + will be iterated upon during epochs in round-robin manner", + ) + parser.add_argument( + "--tokens-per-sample", + default=512, + type=int, + help="max number of total tokens over all segments" + " per sample for BERT dataset", + ) + parser.add_argument( + "--break-mode", default="doc", type=str, help="mode for breaking sentence" + ) + parser.add_argument("--shuffle-dataset", action="store_true", default=False) + + def __init__(self, args, dictionary): + super().__init__(args) + self.dictionary = dictionary + self.seed = args.seed + + @classmethod + def load_dictionary(cls, filename): + return BertDictionary.load(filename) + + @classmethod + def build_dictionary( + cls, filenames, workers=1, threshold=-1, nwords=-1, padding_factor=8 + ): + d = BertDictionary() + for filename in filenames: + Dictionary.add_file_to_dictionary( + filename, d, tokenizer.tokenize_line, workers + ) + d.finalize(threshold=threshold, nwords=nwords, padding_factor=padding_factor) + return d + + @property + def target_dictionary(self): + return self.dictionary + + @classmethod + def setup_task(cls, args, **kwargs): + """Setup the task.""" + paths = utils.split_paths(args.data) + assert len(paths) > 0 + dictionary = BertDictionary.load(os.path.join(paths[0], "dict.txt")) + logger.info("dictionary: {} types".format(len(dictionary))) + + return cls(args, dictionary) + + def load_dataset(self, split, epoch=1, combine=False): + """Load a given dataset split. + + Args: + split (str): name of the split (e.g., train, valid, test) + """ + loaded_datasets = [] + + paths = utils.split_paths(self.args.data) + assert len(paths) > 0 + data_path = paths[(epoch - 1) % len(paths)] + logger.info("data_path", data_path) + + for k in itertools.count(): + split_k = split + (str(k) if k > 0 else "") + path = os.path.join(data_path, split_k) + ds = indexed_dataset.make_dataset( + path, + impl=self.args.dataset_impl, + fix_lua_indexing=True, + dictionary=self.dictionary, + ) + + if ds is None: + if k > 0: + break + else: + raise FileNotFoundError( + "Dataset not found: {} ({})".format(split, data_path) + ) + + with data_utils.numpy_seed(self.seed + k): + loaded_datasets.append( + BlockPairDataset( + ds, + self.dictionary, + ds.sizes, + self.args.tokens_per_sample, + break_mode=self.args.break_mode, + doc_break_size=1, + ) + ) + + logger.info( + "{} {} {} examples".format(data_path, split_k, len(loaded_datasets[-1])) + ) + + if not combine: + break + + if len(loaded_datasets) == 1: + dataset = loaded_datasets[0] + sizes = dataset.sizes + else: + dataset = ConcatDataset(loaded_datasets) + sizes = np.concatenate([ds.sizes for ds in loaded_datasets]) + + self.datasets[split] = MaskedLMDataset( + dataset=dataset, + sizes=sizes, + vocab=self.dictionary, + pad_idx=self.dictionary.pad(), + mask_idx=self.dictionary.mask(), + classif_token_idx=self.dictionary.cls(), + sep_token_idx=self.dictionary.sep(), + shuffle=self.args.shuffle_dataset, + seed=self.seed, + ) diff --git a/fairseq/tasks/masked_lm.py b/fairseq/tasks/masked_lm.py new file mode 100644 index 0000000000000000000000000000000000000000..fd2ea6ade15e94f963045db8cc1a20d4a3d5c7c1 --- /dev/null +++ b/fairseq/tasks/masked_lm.py @@ -0,0 +1,258 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import os + +import numpy as np +from fairseq import utils +from fairseq.data import ( + Dictionary, + IdDataset, + MaskTokensDataset, + NestedDictionaryDataset, + NumelDataset, + NumSamplesDataset, + PrependTokenDataset, + RightPadDataset, + SortDataset, + TokenBlockDataset, + data_utils, +) +from fairseq.data.encoders.utils import get_whole_word_mask +from fairseq.data.shorten_dataset import maybe_shorten_dataset +from fairseq.tasks import LegacyFairseqTask, register_task + + +logger = logging.getLogger(__name__) + + +@register_task("masked_lm") +class MaskedLMTask(LegacyFairseqTask): + """Task for training masked language models (e.g., BERT, RoBERTa).""" + + @staticmethod + def add_args(parser): + """Add task-specific arguments to the parser.""" + parser.add_argument( + "data", + help="colon separated path to data directories list, \ + will be iterated upon during epochs in round-robin manner", + ) + parser.add_argument( + "--sample-break-mode", + default="complete", + choices=["none", "complete", "complete_doc", "eos"], + help='If omitted or "none", fills each sample with tokens-per-sample ' + 'tokens. If set to "complete", splits samples only at the end ' + "of sentence, but may include multiple sentences per sample. " + '"complete_doc" is similar but respects doc boundaries. ' + 'If set to "eos", includes only one sentence per sample.', + ) + parser.add_argument( + "--tokens-per-sample", + default=512, + type=int, + help="max number of total tokens over all segments " + "per sample for BERT dataset", + ) + parser.add_argument( + "--mask-prob", + default=0.15, + type=float, + help="probability of replacing a token with mask", + ) + parser.add_argument( + "--leave-unmasked-prob", + default=0.1, + type=float, + help="probability that a masked token is unmasked", + ) + parser.add_argument( + "--random-token-prob", + default=0.1, + type=float, + help="probability of replacing a token with a random token", + ) + parser.add_argument( + "--freq-weighted-replacement", + default=False, + action="store_true", + help="sample random replacement words based on word frequencies", + ) + parser.add_argument( + "--mask-whole-words", + default=False, + action="store_true", + help="mask whole words; you may also want to set --bpe", + ) + parser.add_argument( + "--mask-multiple-length", + default=1, + type=int, + help="repeat the mask indices multiple times", + ) + parser.add_argument( + "--mask-stdev", default=0.0, type=float, help="stdev of the mask length" + ) + parser.add_argument( + "--shorten-method", + default="none", + choices=["none", "truncate", "random_crop"], + help="if not none, shorten sequences that exceed --tokens-per-sample", + ) + parser.add_argument( + "--shorten-data-split-list", + default="", + help="comma-separated list of dataset splits to apply shortening to, " + 'e.g., "train,valid" (default: all dataset splits)', + ) + + def __init__(self, args, dictionary): + super().__init__(args) + self.dictionary = dictionary + self.seed = args.seed + + # add mask token + self.mask_idx = dictionary.add_symbol("<mask>") + + @classmethod + def setup_task(cls, args, **kwargs): + paths = utils.split_paths(args.data) + assert len(paths) > 0 + dictionary = Dictionary.load(os.path.join(paths[0], "dict.txt")) + logger.info("dictionary: {} types".format(len(dictionary))) + return cls(args, dictionary) + + def load_dataset(self, split, epoch=1, combine=False, **kwargs): + """Load a given dataset split. + + Args: + split (str): name of the split (e.g., train, valid, test) + """ + paths = utils.split_paths(self.args.data) + assert len(paths) > 0 + data_path = paths[(epoch - 1) % len(paths)] + split_path = os.path.join(data_path, split) + + dataset = data_utils.load_indexed_dataset( + split_path, + self.source_dictionary, + self.args.dataset_impl, + combine=combine, + ) + if dataset is None: + raise FileNotFoundError( + "Dataset not found: {} ({})".format(split, split_path) + ) + + dataset = maybe_shorten_dataset( + dataset, + split, + self.args.shorten_data_split_list, + self.args.shorten_method, + self.args.tokens_per_sample, + self.args.seed, + ) + + # create continuous blocks of tokens + dataset = TokenBlockDataset( + dataset, + dataset.sizes, + self.args.tokens_per_sample - 1, # one less for <s> + pad=self.source_dictionary.pad(), + eos=self.source_dictionary.eos(), + break_mode=self.args.sample_break_mode, + ) + logger.info("loaded {} blocks from: {}".format(len(dataset), split_path)) + + # prepend beginning-of-sentence token (<s>, equiv. to [CLS] in BERT) + dataset = PrependTokenDataset(dataset, self.source_dictionary.bos()) + + # create masked input and targets + mask_whole_words = ( + get_whole_word_mask(self.args, self.source_dictionary) + if self.args.mask_whole_words + else None + ) + + src_dataset, tgt_dataset = MaskTokensDataset.apply_mask( + dataset, + self.source_dictionary, + pad_idx=self.source_dictionary.pad(), + mask_idx=self.mask_idx, + seed=self.args.seed, + mask_prob=self.args.mask_prob, + leave_unmasked_prob=self.args.leave_unmasked_prob, + random_token_prob=self.args.random_token_prob, + freq_weighted_replacement=self.args.freq_weighted_replacement, + mask_whole_words=mask_whole_words, + mask_multiple_length=self.args.mask_multiple_length, + mask_stdev=self.args.mask_stdev, + ) + + with data_utils.numpy_seed(self.args.seed): + shuffle = np.random.permutation(len(src_dataset)) + + self.datasets[split] = SortDataset( + NestedDictionaryDataset( + { + "id": IdDataset(), + "net_input": { + "src_tokens": RightPadDataset( + src_dataset, + pad_idx=self.source_dictionary.pad(), + ), + "src_lengths": NumelDataset(src_dataset, reduce=False), + }, + "target": RightPadDataset( + tgt_dataset, + pad_idx=self.source_dictionary.pad(), + ), + "nsentences": NumSamplesDataset(), + "ntokens": NumelDataset(src_dataset, reduce=True), + }, + sizes=[src_dataset.sizes], + ), + sort_order=[ + shuffle, + src_dataset.sizes, + ], + ) + + def build_dataset_for_inference(self, src_tokens, src_lengths, sort=True): + src_dataset = RightPadDataset( + TokenBlockDataset( + src_tokens, + src_lengths, + self.args.tokens_per_sample - 1, # one less for <s> + pad=self.source_dictionary.pad(), + eos=self.source_dictionary.eos(), + break_mode="eos", + ), + pad_idx=self.source_dictionary.pad(), + ) + src_dataset = PrependTokenDataset(src_dataset, self.source_dictionary.bos()) + src_dataset = NestedDictionaryDataset( + { + "id": IdDataset(), + "net_input": { + "src_tokens": src_dataset, + "src_lengths": NumelDataset(src_dataset, reduce=False), + }, + }, + sizes=src_lengths, + ) + if sort: + src_dataset = SortDataset(src_dataset, sort_order=[src_lengths]) + return src_dataset + + @property + def source_dictionary(self): + return self.dictionary + + @property + def target_dictionary(self): + return self.dictionary diff --git a/fairseq/tasks/multilingual_denoising.py b/fairseq/tasks/multilingual_denoising.py new file mode 100644 index 0000000000000000000000000000000000000000..d1c914917feb5165aad7482cd1377f5f65b21635 --- /dev/null +++ b/fairseq/tasks/multilingual_denoising.py @@ -0,0 +1,254 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import os + +import numpy as np +from fairseq.data import ( + AppendTokenDataset, + ConcatDataset, + DenoisingDataset, + Dictionary, + PrependTokenDataset, + ResamplingDataset, + SortDataset, + TokenBlockDataset, + data_utils, +) +from fairseq.data.encoders.utils import get_whole_word_mask +from fairseq.tasks import register_task + +from .denoising import DenoisingTask + + +logger = logging.getLogger(__name__) + + +@register_task("multilingual_denoising") +class MultilingualDenoisingTask(DenoisingTask): + @staticmethod + def add_args(parser): + DenoisingTask.add_args(parser) + parser.add_argument( + "--multilang-sampling-alpha", + type=float, + default=1.0, + help="smoothing alpha for sample ratios across multiple datasets", + ) + parser.add_argument("--add-lang-token", default=False, action="store_true") + parser.add_argument( + "--langs", type=str, help="language ids we are considering", default=None + ) + parser.add_argument( + "--no-whole-word-mask-langs", + type=str, + default="", + metavar="N", + help="languages without spacing between words dont support whole word masking", + ) + + @classmethod + def setup_task(cls, args, **kwargs): + """Setup the task.""" + paths = args.data.split(":") + assert len(paths) > 0 + dictionary = Dictionary.load(os.path.join(paths[0], "dict.txt")) + + data_path = paths[0] + if args.langs is None: + languages = sorted( + [ + name + for name in os.listdir(data_path) + if os.path.isdir(os.path.join(data_path, name)) + ] + ) + else: + languages = args.langs.split(",") + + if args.add_lang_token: + for lang in languages: + dictionary.add_symbol("[{}]".format(lang)) + + logger.info("dictionary: {} types".format(len(dictionary))) + if not hasattr(args, "shuffle_instance"): + args.shuffle_instance = False + return cls(args, dictionary) + + def __init__(self, args, dictionary): + super().__init__(args, dictionary) + self.dictionary = dictionary + self.seed = args.seed + + # add mask token + self.mask_idx = self.dictionary.add_symbol("<mask>") + self.langs = args.langs + self.args = args + + def _get_sample_prob(self, dataset_lens): + """ + Get smoothed sampling porbability by languages. This helps low resource + languages by upsampling them. + """ + prob = dataset_lens / dataset_lens.sum() + smoothed_prob = prob ** self.args.multilang_sampling_alpha + smoothed_prob = smoothed_prob / smoothed_prob.sum() + return smoothed_prob + + def load_dataset(self, split, epoch=1, combine=False, **kwargs): + """Load a given dataset split. + + Args: + split (str): name of the split (e.g., train, valid, test) + """ + paths = self.args.data.split(":") + assert len(paths) > 0 + data_path = paths[(epoch - 1) % len(paths)] + split_path = os.path.join(data_path, split) + + if self.langs is None: + languages = sorted( + [ + name + for name in os.listdir(data_path) + if os.path.isdir(os.path.join(data_path, name)) + ] + ) + else: + languages = self.langs.split(",") + for name in languages: + p = os.path.join(data_path, name) + assert os.path.exists(p), "data not found: {}".format(p) + + logger.info("Training on {0} languages: {1}".format(len(languages), languages)) + logger.info( + "Language to id mapping: ", {lang: id for id, lang in enumerate(languages)} + ) + + mask_whole_words = get_whole_word_mask(self.args, self.dictionary) + language_without_segmentations = self.args.no_whole_word_mask_langs.split(",") + lang_datasets = [] + for language in languages: + split_path = os.path.join(data_path, language, split) + + dataset = data_utils.load_indexed_dataset( + split_path, + self.source_dictionary, + self.args.dataset_impl, + combine=combine, + ) + if dataset is None: + raise FileNotFoundError( + "Dataset not found: {} ({})".format(split, split_path) + ) + + end_token = ( + self.source_dictionary.index("[{}]".format(language)) + if self.args.add_lang_token + else self.source_dictionary.eos() + ) + + # create continuous blocks of tokens + dataset = TokenBlockDataset( + dataset, + dataset.sizes, + self.args.tokens_per_sample - 2, # one less for <s> + pad=self.source_dictionary.pad(), + eos=end_token, + break_mode=self.args.sample_break_mode, + ) + logger.info("loaded {} blocks from: {}".format(len(dataset), split_path)) + + # prepend beginning-of-sentence token (<s>, equiv. to [CLS] in BERT) + dataset = PrependTokenDataset(dataset, self.source_dictionary.bos()) + dataset = AppendTokenDataset(dataset, end_token) + + lang_mask_whole_words = ( + mask_whole_words + if language not in language_without_segmentations + else None + ) + lang_dataset = DenoisingDataset( + dataset, + dataset.sizes, + self.dictionary, + self.mask_idx, + lang_mask_whole_words, + shuffle=self.args.shuffle_instance, + seed=self.seed, + args=self.args, + eos=None + if not self.args.add_lang_token + else self.source_dictionary.index("[{}]".format(language)), + ) + lang_datasets.append(lang_dataset) + + dataset_lengths = np.array( + [len(d) for d in lang_datasets], + dtype=float, + ) + logger.info( + "loaded total {} blocks for all languages".format( + int(dataset_lengths.sum()), + ) + ) + if split == self.args.train_subset: + # For train subset, additionally up or down sample languages. + sample_probs = self._get_sample_prob(dataset_lengths) + logger.info( + "Sample probability by language: {}".format( + { + lang: "{0:.4f}".format(sample_probs[id]) + for id, lang in enumerate(languages) + } + ) + ) + size_ratio = (sample_probs * dataset_lengths.sum()) / dataset_lengths + logger.info( + "Up/Down Sampling ratio by language: {}".format( + { + lang: "{0:.2f}".format(size_ratio[id]) + for id, lang in enumerate(languages) + } + ) + ) + + resampled_lang_datasets = [ + ResamplingDataset( + lang_datasets[i], + size_ratio=size_ratio[i], + seed=self.args.seed, + epoch=epoch, + replace=size_ratio[i] >= 1.0, + ) + for i, d in enumerate(lang_datasets) + ] + dataset = ConcatDataset( + resampled_lang_datasets, + ) + else: + dataset = ConcatDataset(lang_datasets) + lang_splits = [split] + for lang_id, lang_dataset in enumerate(lang_datasets): + split_name = split + "_" + languages[lang_id] + lang_splits.append(split_name) + self.datasets[split_name] = lang_dataset + + if split in self.args.valid_subset: + self.args.valid_subset = self.args.valid_subset.replace( + split, ",".join(lang_splits) + ) + + with data_utils.numpy_seed(self.args.seed + epoch): + shuffle = np.random.permutation(len(dataset)) + + self.datasets[split] = SortDataset( + dataset, + sort_order=[ + shuffle, + dataset.sizes, + ], + ) diff --git a/fairseq/tasks/multilingual_masked_lm.py b/fairseq/tasks/multilingual_masked_lm.py new file mode 100644 index 0000000000000000000000000000000000000000..9e6ce4b8a2f77ed889a6e1451321a8e3ac21dc67 --- /dev/null +++ b/fairseq/tasks/multilingual_masked_lm.py @@ -0,0 +1,338 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import os + +import numpy as np +import torch +from fairseq import utils +from fairseq.data import ( + ConcatDataset, + Dictionary, + IdDataset, + MaskTokensDataset, + NestedDictionaryDataset, + NumelDataset, + NumSamplesDataset, + PadDataset, + PrependTokenDataset, + RawLabelDataset, + ResamplingDataset, + SortDataset, + TokenBlockDataset, + data_utils, + encoders, +) +from fairseq.tasks import LegacyFairseqTask, register_task + + +logger = logging.getLogger(__name__) + + +@register_task("multilingual_masked_lm") +class MultiLingualMaskedLMTask(LegacyFairseqTask): + """Task for training masked language models (e.g., BERT, RoBERTa).""" + + @staticmethod + def add_args(parser): + """Add task-specific arguments to the parser.""" + parser.add_argument( + "data", + help="colon separated path to data directories list, \ + will be iterated upon during epochs in round-robin manner", + ) + parser.add_argument( + "--sample-break-mode", + default="complete", + choices=["none", "complete", "complete_doc", "eos"], + help='If omitted or "none", fills each sample with tokens-per-sample ' + 'tokens. If set to "complete", splits samples only at the end ' + "of sentence, but may include multiple sentences per sample. " + '"complete_doc" is similar but respects doc boundaries. ' + 'If set to "eos", includes only one sentence per sample.', + ) + parser.add_argument( + "--tokens-per-sample", + default=512, + type=int, + help="max number of total tokens over all segments " + "per sample for BERT dataset", + ) + parser.add_argument( + "--mask-prob", + default=0.15, + type=float, + help="probability of replacing a token with mask", + ) + parser.add_argument( + "--leave-unmasked-prob", + default=0.1, + type=float, + help="probability that a masked token is unmasked", + ) + parser.add_argument( + "--random-token-prob", + default=0.1, + type=float, + help="probability of replacing a token with a random token", + ) + parser.add_argument( + "--freq-weighted-replacement", + action="store_true", + help="sample random replacement words based on word frequencies", + ) + parser.add_argument( + "--mask-whole-words", + default=False, + action="store_true", + help="mask whole words; you may also want to set --bpe", + ) + parser.add_argument( + "--multilang-sampling-alpha", + type=float, + default=1.0, + help="smoothing alpha for sample rations across multiple datasets", + ) + + def __init__(self, args, dictionary): + super().__init__(args) + self.dictionary = dictionary + self.seed = args.seed + + # add mask token + self.mask_idx = dictionary.add_symbol("<mask>") + + @classmethod + def setup_task(cls, args, **kwargs): + paths = utils.split_paths(args.data) + assert len(paths) > 0 + dictionary = Dictionary.load(os.path.join(paths[0], "dict.txt")) + logger.info("dictionary: {} types".format(len(dictionary))) + return cls(args, dictionary) + + def _get_whole_word_mask(self): + # create masked input and targets + if self.args.mask_whole_words: + bpe = encoders.build_bpe(self.args) + if bpe is not None: + + def is_beginning_of_word(i): + if i < self.source_dictionary.nspecial: + # special elements are always considered beginnings + return True + tok = self.source_dictionary[i] + if tok.startswith("madeupword"): + return True + try: + return bpe.is_beginning_of_word(tok) + except ValueError: + return True + + mask_whole_words = torch.ByteTensor( + list(map(is_beginning_of_word, range(len(self.source_dictionary)))) + ) + else: + mask_whole_words = None + return mask_whole_words + + def _get_sample_prob(self, dataset_lens): + """ + Get smoothed sampling porbability by languages. This helps low resource + languages by upsampling them. + """ + prob = dataset_lens / dataset_lens.sum() + smoothed_prob = prob ** self.args.multilang_sampling_alpha + smoothed_prob = smoothed_prob / smoothed_prob.sum() + return smoothed_prob + + def load_dataset(self, split, epoch=1, combine=False, **kwargs): + """Load a given dataset split. + + Args: + split (str): name of the split (e.g., train, valid, test) + """ + paths = utils.split_paths(self.args.data) + assert len(paths) > 0 + data_path = paths[(epoch - 1) % len(paths)] + + languages = sorted( + name + for name in os.listdir(data_path) + if os.path.isdir(os.path.join(data_path, name)) + ) + + logger.info("Training on {0} languages: {1}".format(len(languages), languages)) + logger.info( + "Language to id mapping: ", {lang: id for id, lang in enumerate(languages)} + ) + + mask_whole_words = self._get_whole_word_mask() + lang_datasets = [] + for lang_id, language in enumerate(languages): + split_path = os.path.join(data_path, language, split) + + dataset = data_utils.load_indexed_dataset( + split_path, + self.source_dictionary, + self.args.dataset_impl, + combine=combine, + ) + if dataset is None: + raise FileNotFoundError( + "Dataset not found: {} ({})".format(split, split_path) + ) + + # create continuous blocks of tokens + dataset = TokenBlockDataset( + dataset, + dataset.sizes, + self.args.tokens_per_sample - 1, # one less for <s> + pad=self.source_dictionary.pad(), + eos=self.source_dictionary.eos(), + break_mode=self.args.sample_break_mode, + ) + logger.info("loaded {} blocks from: {}".format(len(dataset), split_path)) + + # prepend beginning-of-sentence token (<s>, equiv. to [CLS] in BERT) + dataset = PrependTokenDataset(dataset, self.source_dictionary.bos()) + + src_dataset, tgt_dataset = MaskTokensDataset.apply_mask( + dataset, + self.source_dictionary, + pad_idx=self.source_dictionary.pad(), + mask_idx=self.mask_idx, + seed=self.args.seed, + mask_prob=self.args.mask_prob, + leave_unmasked_prob=self.args.leave_unmasked_prob, + random_token_prob=self.args.random_token_prob, + freq_weighted_replacement=self.args.freq_weighted_replacement, + mask_whole_words=mask_whole_words, + ) + + lang_dataset = NestedDictionaryDataset( + { + "net_input": { + "src_tokens": PadDataset( + src_dataset, + pad_idx=self.source_dictionary.pad(), + left_pad=False, + ), + "src_lengths": NumelDataset(src_dataset, reduce=False), + }, + "target": PadDataset( + tgt_dataset, + pad_idx=self.source_dictionary.pad(), + left_pad=False, + ), + "nsentences": NumSamplesDataset(), + "ntokens": NumelDataset(src_dataset, reduce=True), + "lang_id": RawLabelDataset([lang_id] * src_dataset.sizes.shape[0]), + }, + sizes=[src_dataset.sizes], + ) + lang_datasets.append(lang_dataset) + + dataset_lengths = np.array( + [len(d) for d in lang_datasets], + dtype=float, + ) + logger.info( + "loaded total {} blocks for all languages".format( + dataset_lengths.sum(), + ) + ) + if split == self.args.train_subset: + # For train subset, additionally up or down sample languages. + sample_probs = self._get_sample_prob(dataset_lengths) + logger.info( + "Sample probability by language: ", + { + lang: "{0:.4f}".format(sample_probs[id]) + for id, lang in enumerate(languages) + }, + ) + size_ratio = (sample_probs * dataset_lengths.sum()) / dataset_lengths + logger.info( + "Up/Down Sampling ratio by language: ", + { + lang: "{0:.2f}".format(size_ratio[id]) + for id, lang in enumerate(languages) + }, + ) + + resampled_lang_datasets = [ + ResamplingDataset( + lang_datasets[i], + size_ratio=size_ratio[i], + seed=self.args.seed, + epoch=epoch, + replace=size_ratio[i] >= 1.0, + ) + for i, d in enumerate(lang_datasets) + ] + dataset = ConcatDataset(resampled_lang_datasets) + else: + dataset = ConcatDataset(lang_datasets) + lang_splits = [split] + for lang_id, lang_dataset in enumerate(lang_datasets): + split_name = split + "_" + languages[lang_id] + lang_splits.append(split_name) + self.datasets[split_name] = lang_dataset + + # [TODO]: This is hacky for now to print validation ppl for each + # language individually. Maybe need task API changes to allow it + # in more generic ways. + if split in self.args.valid_subset: + self.args.valid_subset = self.args.valid_subset.replace( + split, ",".join(lang_splits) + ) + + with data_utils.numpy_seed(self.args.seed + epoch): + shuffle = np.random.permutation(len(dataset)) + + self.datasets[split] = SortDataset( + dataset, + sort_order=[ + shuffle, + dataset.sizes, + ], + ) + + def build_dataset_for_inference(self, src_tokens, src_lengths, sort=True): + src_dataset = PadDataset( + TokenBlockDataset( + src_tokens, + src_lengths, + self.args.tokens_per_sample - 1, # one less for <s> + pad=self.source_dictionary.pad(), + eos=self.source_dictionary.eos(), + break_mode="eos", + ), + pad_idx=self.source_dictionary.pad(), + left_pad=False, + ) + src_dataset = PrependTokenDataset(src_dataset, self.source_dictionary.bos()) + src_dataset = NestedDictionaryDataset( + { + "id": IdDataset(), + "net_input": { + "src_tokens": src_dataset, + "src_lengths": NumelDataset(src_dataset, reduce=False), + }, + }, + sizes=src_lengths, + ) + if sort: + src_dataset = SortDataset(src_dataset, sort_order=[src_lengths]) + return src_dataset + + @property + def source_dictionary(self): + return self.dictionary + + @property + def target_dictionary(self): + return self.dictionary diff --git a/fairseq/tasks/multilingual_translation.py b/fairseq/tasks/multilingual_translation.py new file mode 100644 index 0000000000000000000000000000000000000000..26e0b529d5f2902bd80c8207a001ae28af393291 --- /dev/null +++ b/fairseq/tasks/multilingual_translation.py @@ -0,0 +1,457 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import contextlib +import logging +import os +from collections import OrderedDict + +import torch +from fairseq import metrics, options, utils +from fairseq.data import ( + Dictionary, + LanguagePairDataset, + RoundRobinZipDatasets, + TransformEosLangPairDataset, +) +from fairseq.models import FairseqMultiModel +from fairseq.tasks.translation import load_langpair_dataset + +from . import LegacyFairseqTask, register_task + + +logger = logging.getLogger(__name__) + + +def _lang_token(lang: str): + return "__{}__".format(lang) + + +def _lang_token_index(dic: Dictionary, lang: str): + """Return language token index.""" + idx = dic.index(_lang_token(lang)) + assert idx != dic.unk_index, "cannot find language token for lang {}".format(lang) + return idx + + +@register_task("multilingual_translation") +class MultilingualTranslationTask(LegacyFairseqTask): + """A task for training multiple translation models simultaneously. + + We iterate round-robin over batches from multiple language pairs, ordered + according to the `--lang-pairs` argument. + + The training loop is roughly: + + for i in range(len(epoch)): + for lang_pair in args.lang_pairs: + batch = next_batch_for_lang_pair(lang_pair) + loss = criterion(model_for_lang_pair(lang_pair), batch) + loss.backward() + optimizer.step() + + In practice, `next_batch_for_lang_pair` is abstracted in a FairseqDataset + (e.g., `RoundRobinZipDatasets`) and `model_for_lang_pair` is a model that + implements the `FairseqMultiModel` interface. + + During inference it is required to specify a single `--source-lang` and + `--target-lang`, which indicates the inference langauge direction. + `--lang-pairs`, `--encoder-langtok`, `--decoder-langtok` have to be set to + the same value as training. + """ + + @staticmethod + def add_args(parser): + """Add task-specific arguments to the parser.""" + # fmt: off + parser.add_argument('data', metavar='DIR', help='path to data directory') + parser.add_argument('--lang-pairs', default=None, metavar='PAIRS', + help='comma-separated list of language pairs (in training order): en-de,en-fr,de-fr') + parser.add_argument('-s', '--source-lang', default=None, metavar='SRC', + help='source language (only needed for inference)') + parser.add_argument('-t', '--target-lang', default=None, metavar='TARGET', + help='target language (only needed for inference)') + parser.add_argument('--left-pad-source', default='True', type=str, metavar='BOOL', + help='pad the source on the left (default: True)') + parser.add_argument('--left-pad-target', default='False', type=str, metavar='BOOL', + help='pad the target on the left (default: False)') + parser.add_argument('--max-source-positions', default=1024, type=int, metavar='N', + help='max number of tokens in the source sequence') + parser.add_argument('--max-target-positions', default=1024, type=int, metavar='N', + help='max number of tokens in the target sequence') + parser.add_argument('--upsample-primary', default=1, type=int, + help='amount to upsample primary dataset') + parser.add_argument('--encoder-langtok', default=None, type=str, choices=['src', 'tgt'], + metavar='SRCTGT', + help='replace beginning-of-sentence in source sentence with source or target ' + 'language token. (src/tgt)') + parser.add_argument('--decoder-langtok', action='store_true', + help='replace beginning-of-sentence in target sentence with target language token') + # fmt: on + + def __init__(self, args, dicts, training): + super().__init__(args) + self.dicts = dicts + self.training = training + if training: + self.lang_pairs = args.lang_pairs + else: + self.lang_pairs = ["{}-{}".format(args.source_lang, args.target_lang)] + # eval_lang_pairs for multilingual translation is usually all of the + # lang_pairs. However for other multitask settings or when we want to + # optimize for certain languages we want to use a different subset. Thus + # the eval_lang_pairs class variable is provided for classes that extend + # this class. + self.eval_lang_pairs = self.lang_pairs + # model_lang_pairs will be used to build encoder-decoder model pairs in + # models.build_model(). This allows multitask type of sub-class can + # build models other than the input lang_pairs + self.model_lang_pairs = self.lang_pairs + self.langs = list(dicts.keys()) + + @classmethod + def setup_task(cls, args, **kwargs): + dicts, training = cls.prepare(args, **kwargs) + return cls(args, dicts, training) + + @classmethod + def update_args(cls, args): + args.left_pad_source = utils.eval_bool(args.left_pad_source) + args.left_pad_target = utils.eval_bool(args.left_pad_target) + + if args.lang_pairs is None: + raise ValueError( + "--lang-pairs is required. List all the language pairs in the training objective." + ) + if isinstance(args.lang_pairs, str): + args.lang_pairs = args.lang_pairs.split(",") + + @classmethod + def prepare(cls, args, **kargs): + cls.update_args(args) + sorted_langs = sorted( + list({x for lang_pair in args.lang_pairs for x in lang_pair.split("-")}) + ) + if args.source_lang is not None or args.target_lang is not None: + training = False + else: + training = True + + # load dictionaries + dicts = OrderedDict() + for lang in sorted_langs: + paths = utils.split_paths(args.data) + assert len(paths) > 0 + dicts[lang] = cls.load_dictionary( + os.path.join(paths[0], "dict.{}.txt".format(lang)) + ) + if len(dicts) > 0: + assert dicts[lang].pad() == dicts[sorted_langs[0]].pad() + assert dicts[lang].eos() == dicts[sorted_langs[0]].eos() + assert dicts[lang].unk() == dicts[sorted_langs[0]].unk() + if args.encoder_langtok is not None or args.decoder_langtok: + for lang_to_add in sorted_langs: + dicts[lang].add_symbol(_lang_token(lang_to_add)) + logger.info("[{}] dictionary: {} types".format(lang, len(dicts[lang]))) + return dicts, training + + def get_encoder_langtok(self, src_lang, tgt_lang): + if self.args.encoder_langtok is None: + return self.dicts[src_lang].eos() + if self.args.encoder_langtok == "src": + return _lang_token_index(self.dicts[src_lang], src_lang) + else: + return _lang_token_index(self.dicts[src_lang], tgt_lang) + + def get_decoder_langtok(self, tgt_lang): + if not self.args.decoder_langtok: + return self.dicts[tgt_lang].eos() + return _lang_token_index(self.dicts[tgt_lang], tgt_lang) + + def alter_dataset_langtok( + self, + lang_pair_dataset, + src_eos=None, + src_lang=None, + tgt_eos=None, + tgt_lang=None, + ): + if self.args.encoder_langtok is None and not self.args.decoder_langtok: + return lang_pair_dataset + + new_src_eos = None + if ( + self.args.encoder_langtok is not None + and src_eos is not None + and src_lang is not None + and tgt_lang is not None + ): + new_src_eos = self.get_encoder_langtok(src_lang, tgt_lang) + else: + src_eos = None + + new_tgt_bos = None + if self.args.decoder_langtok and tgt_eos is not None and tgt_lang is not None: + new_tgt_bos = self.get_decoder_langtok(tgt_lang) + else: + tgt_eos = None + + return TransformEosLangPairDataset( + lang_pair_dataset, + src_eos=src_eos, + new_src_eos=new_src_eos, + tgt_bos=tgt_eos, + new_tgt_bos=new_tgt_bos, + ) + + def load_dataset(self, split, epoch=1, **kwargs): + """Load a dataset split.""" + paths = utils.split_paths(self.args.data) + assert len(paths) > 0 + data_path = paths[(epoch - 1) % len(paths)] + + def language_pair_dataset(lang_pair): + src, tgt = lang_pair.split("-") + langpair_dataset = load_langpair_dataset( + data_path, + split, + src, + self.dicts[src], + tgt, + self.dicts[tgt], + combine=True, + dataset_impl=self.args.dataset_impl, + upsample_primary=self.args.upsample_primary, + left_pad_source=self.args.left_pad_source, + left_pad_target=self.args.left_pad_target, + max_source_positions=self.args.max_source_positions, + max_target_positions=self.args.max_target_positions, + ) + return self.alter_dataset_langtok( + langpair_dataset, + src_eos=self.dicts[src].eos(), + src_lang=src, + tgt_eos=self.dicts[tgt].eos(), + tgt_lang=tgt, + ) + + self.datasets[split] = RoundRobinZipDatasets( + OrderedDict( + [ + (lang_pair, language_pair_dataset(lang_pair)) + for lang_pair in self.lang_pairs + ] + ), + eval_key=None + if self.training + else "%s-%s" % (self.args.source_lang, self.args.target_lang), + ) + + def build_dataset_for_inference(self, src_tokens, src_lengths, constraints=None): + if constraints is not None: + raise NotImplementedError( + "Constrained decoding with the multilingual_translation task is not supported" + ) + + lang_pair = "%s-%s" % (self.args.source_lang, self.args.target_lang) + return RoundRobinZipDatasets( + OrderedDict( + [ + ( + lang_pair, + self.alter_dataset_langtok( + LanguagePairDataset( + src_tokens, src_lengths, self.source_dictionary + ), + src_eos=self.source_dictionary.eos(), + src_lang=self.args.source_lang, + tgt_eos=self.target_dictionary.eos(), + tgt_lang=self.args.target_lang, + ), + ) + ] + ), + eval_key=lang_pair, + ) + + def build_model(self, args): + def check_args(): + messages = [] + if ( + len(set(self.args.lang_pairs).symmetric_difference(args.lang_pairs)) + != 0 + ): + messages.append( + "--lang-pairs should include all the language pairs {}.".format( + args.lang_pairs + ) + ) + if self.args.encoder_langtok != args.encoder_langtok: + messages.append( + "--encoder-langtok should be {}.".format(args.encoder_langtok) + ) + if self.args.decoder_langtok != args.decoder_langtok: + messages.append( + "--decoder-langtok should {} be set.".format( + "" if args.decoder_langtok else "not" + ) + ) + + if len(messages) > 0: + raise ValueError(" ".join(messages)) + + # Update args -> the fact that the constructor here + # changes the args object doesn't mean you get the same one here + self.update_args(args) + + # Check if task args are consistant with model args + check_args() + + from fairseq import models + + model = models.build_model(args, self) + if not isinstance(model, FairseqMultiModel): + raise ValueError( + "MultilingualTranslationTask requires a FairseqMultiModel architecture" + ) + return model + + def _per_lang_pair_train_loss( + self, lang_pair, model, update_num, criterion, sample, optimizer, ignore_grad + ): + loss, sample_size, logging_output = criterion( + model.models[lang_pair], sample[lang_pair] + ) + if ignore_grad: + loss *= 0 + optimizer.backward(loss) + return loss, sample_size, logging_output + + def train_step( + self, sample, model, criterion, optimizer, update_num, ignore_grad=False + ): + model.train() + from collections import defaultdict + + agg_loss, agg_sample_size, agg_logging_output = 0.0, 0.0, defaultdict(float) + curr_lang_pairs = [ + lang_pair + for lang_pair in self.model_lang_pairs + if sample[lang_pair] is not None and len(sample[lang_pair]) != 0 + ] + + for idx, lang_pair in enumerate(curr_lang_pairs): + + def maybe_no_sync(): + if ( + self.args.distributed_world_size > 1 + and hasattr(model, "no_sync") + and idx < len(curr_lang_pairs) - 1 + ): + return model.no_sync() + else: + return contextlib.ExitStack() # dummy contextmanager + + with maybe_no_sync(): + loss, sample_size, logging_output = self._per_lang_pair_train_loss( + lang_pair, + model, + update_num, + criterion, + sample, + optimizer, + ignore_grad, + ) + agg_loss += loss.detach().item() + # TODO make summing of the sample sizes configurable + agg_sample_size += sample_size + for k in logging_output: + agg_logging_output[k] += logging_output[k] + agg_logging_output[f"{lang_pair}:{k}"] += logging_output[k] + return agg_loss, agg_sample_size, agg_logging_output + + def _per_lang_pair_valid_loss(self, lang_pair, model, criterion, sample): + return criterion(model.models[lang_pair], sample[lang_pair]) + + def valid_step(self, sample, model, criterion): + model.eval() + with torch.no_grad(): + from collections import defaultdict + + agg_loss, agg_sample_size, agg_logging_output = 0.0, 0.0, defaultdict(float) + for lang_pair in self.eval_lang_pairs: + if ( + lang_pair not in sample + or sample[lang_pair] is None + or len(sample[lang_pair]) == 0 + ): + continue + loss, sample_size, logging_output = self._per_lang_pair_valid_loss( + lang_pair, model, criterion, sample + ) + agg_loss += loss.data.item() + # TODO make summing of the sample sizes configurable + agg_sample_size += sample_size + for k in logging_output: + agg_logging_output[k] += logging_output[k] + agg_logging_output[f"{lang_pair}:{k}"] += logging_output[k] + return agg_loss, agg_sample_size, agg_logging_output + + def inference_step( + self, generator, models, sample, prefix_tokens=None, constraints=None + ): + with torch.no_grad(): + if self.args.decoder_langtok: + bos_token = _lang_token_index( + self.target_dictionary, self.args.target_lang + ) + else: + bos_token = self.target_dictionary.eos() + return generator.generate( + models, + sample, + prefix_tokens=prefix_tokens, + constraints=constraints, + bos_token=bos_token, + ) + + def reduce_metrics(self, logging_outputs, criterion): + with metrics.aggregate(): + # pass 'sample_size', 'nsentences', 'ntokens' stats to fairseq_task + super().reduce_metrics(logging_outputs, criterion) + for k in ["sample_size", "nsentences", "ntokens"]: + metrics.log_scalar(k, sum(l[k] for l in logging_outputs)) + + @property + def source_dictionary(self): + if self.training: + return next(iter(self.dicts.values())) + else: + return self.dicts[self.args.source_lang] + + @property + def target_dictionary(self): + if self.training: + return next(iter(self.dicts.values())) + else: + return self.dicts[self.args.target_lang] + + def max_positions(self): + """Return the max sentence length allowed by the task.""" + if len(self.datasets.values()) == 0: + return { + "%s-%s" + % (self.args.source_lang, self.args.target_lang): ( + self.args.max_source_positions, + self.args.max_target_positions, + ) + } + return OrderedDict( + [ + (key, (self.args.max_source_positions, self.args.max_target_positions)) + for split in self.datasets.keys() + for key in self.datasets[split].datasets.keys() + ] + ) diff --git a/fairseq/tasks/online_backtranslation.py b/fairseq/tasks/online_backtranslation.py new file mode 100644 index 0000000000000000000000000000000000000000..2545624cd4ad9a7ec684aca798dca339feeff58b --- /dev/null +++ b/fairseq/tasks/online_backtranslation.py @@ -0,0 +1,677 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import contextlib +import json +import logging +import math +import os +from argparse import Namespace +from collections import OrderedDict, defaultdict +from pathlib import Path +from typing import Dict, Sequence, Tuple + +import numpy as np +import torch +import torch.nn as nn +import torch.nn.functional as F + +import fairseq +from fairseq import metrics, options, utils +from fairseq.data import ( + FairseqDataset, + LanguagePairDataset, + NoisingDataset, + PrependTokenDataset, + RoundRobinZipDatasets, + TransformEosLangPairDataset, + data_utils, + encoders, +) +from fairseq.sequence_generator import SequenceGenerator +from fairseq.tasks import register_task +from fairseq.tasks.translation import TranslationTask, load_langpair_dataset + +logger = logging.getLogger(__name__) + + +class PiecewiseLinearFn: + """Piecewise linear function. Can be configured with a string.""" + + def __init__(self, pieces: Sequence[Tuple[int, float]]): + assert pieces == sorted( + pieces + ), f"PiecewiseLinearFn configuration should be sorted, received: {pieces}" + + self.pieces = pieces + + def __call__(self, x: int) -> float: + for i, (x_a, y_a) in enumerate(self.pieces[:-1]): + x_b, y_b = self.pieces[i + 1] + if x_a <= x <= x_b: + return y_a + (x - x_a) * (y_b - y_a) / (x_b - x_a) + + return self.pieces[-1][1] + + @staticmethod + def from_string(configuration: str) -> "PiecewiseLinearFn": + """ + Parse the configuration of lambda coefficient (for scheduling). + x = "3" # lambda will be a constant equal to x + x = "0:1,1000:0" # lambda will start from 1 and linearly decrease + # to 0 during the first 1000 iterations + x = "0:0,1000:0,2000:1" # lambda will be equal to 0 for the first 1000 + # iterations, then will linearly increase to 1 until iteration 2000 + """ + if isinstance(configuration, float): + return PiecewiseLinearFn([(0, configuration)]) + + try: + parts = configuration.split(",") + if len(parts) == 1: + v = float(configuration) + return PiecewiseLinearFn([(0, v)]) + + split = [s.split(":") for s in parts] + pieces = [(int(t), float(v)) for t, v in split] + return PiecewiseLinearFn(pieces) + except Exception: + raise ValueError( + f"Invalid PiecewiseLinearFn configuration: {configuration!r}" + ) + + @staticmethod + def one() -> "PiecewiseLinearFn": + return PiecewiseLinearFn([(0, 1.0)]) + + +@register_task("online_backtranslation") +class OnlineBackTranslationTask(TranslationTask): + @staticmethod + def add_args(parser): + """Add task-specific arguments to the parser.""" + # fmt: off + # Generic translation args + parser.add_argument('data', help='colon separated path to data directories list, \ + will be iterated upon during epochs in round-robin manner; \ + however, valid and test data are always in the first directory to \ + avoid the need for repeating them in all directories') + parser.add_argument('--mono-langs', metavar='MONO_LANGS', + help='monolingual languages for training') + parser.add_argument('--valid-lang-pairs', default=None, metavar='VALID_LANG_PAIRS', + help='language pairs for validation') + parser.add_argument('--load-alignments', action='store_true', + help='load the binarized alignments') + parser.add_argument('--left-pad-source', default='False', type=str, metavar='BOOL', + help='pad the source on the left') + parser.add_argument('--left-pad-target', default='False', type=str, metavar='BOOL', + help='pad the target on the left') + parser.add_argument('--upsample-primary', default=1, type=int, + help='amount to upsample primary dataset') + parser.add_argument('--max-source-positions', default=1024, type=int, metavar='N', + help='max number of tokens in the source sequence') + parser.add_argument('--max-target-positions', default=1024, type=int, metavar='N', + help='max number of tokens in the target sequence') + parser.add_argument('--truncate-source', action='store_true', default=False, + help='truncate source to max-source-positions') + parser.add_argument('--num-batch-buckets', default=0, type=int, metavar='N', + help='if >0, then bucket source and target lengths into N ' + 'buckets and pad accordingly; this is useful on TPUs ' + 'to minimize the number of compilations') + + # Denoising args + parser.add_argument('--max-word-shuffle-distance', default=3.0, type=float, metavar='N', + help='maximum word shuffle distance for denoising autoencoding data generation') + parser.add_argument('--word-dropout-prob', default=0.1, type=float, metavar='N', + help='word dropout probability for denoising autoencoding data generation') + parser.add_argument('--word-blanking-prob', default=0.2, type=float, metavar='N', + help='word blanking probability for denoising autoencoding data generation') + + # Backtranslation args + parser.add_argument('--lambda-bt', default="1.0", type=str, metavar='N', + help='back-translation weight') + parser.add_argument('--lambda-dae', default="1.0", type=str, metavar='N', + help='denoising auto-encoder weight') + + # Evaluation args + parser.add_argument('--generate-one-by-one', action='store_true', + help='generate one sentence at a time for backtranslation') + + parser.add_argument('--eval-bleu', action='store_true', + help='evaluation with BLEU scores') + parser.add_argument('--eval-bleu-detok', type=str, default="space", + help='detokenize before computing BLEU (e.g., "moses"); ' + 'required if using --eval-bleu; use "space" to ' + 'disable detokenization; see fairseq.data.encoders ' + 'for other options') + parser.add_argument('--eval-bleu-detok-args', type=str, metavar='JSON', + help='args for building the tokenizer, if needed') + parser.add_argument('--eval-tokenized-bleu', action='store_true', default=False, + help='compute tokenized BLEU instead of sacrebleu') + parser.add_argument('--eval-bleu-remove-bpe', nargs='?', const='@@ ', default=None, + help='remove BPE before computing BLEU') + parser.add_argument('--eval-bleu-args', type=str, metavar='JSON', + help='generation args for BLUE scoring, ' + 'e.g., \'{"beam": 4, "lenpen": 0.6}\'') + parser.add_argument('--eval-bleu-print-samples', action='store_true', + help='print sample generations during validation') + # fmt: on + + def __init__(self, args, common_dict, mono_langs, valid_lang_pairs): + super().__init__(args, common_dict, common_dict) + self.common_dict = common_dict + self.mono_langs = mono_langs + self.valid_lang_pairs = valid_lang_pairs + + self.SHOW_SAMPLES_INTERVAL = 1000 + # Start by showing samples + self._show_samples_ctr = self.SHOW_SAMPLES_INTERVAL + self.SHOW_SAMPLES_NUMBER = 5 + self.lambda_bt = PiecewiseLinearFn.from_string(args.lambda_bt) + self.lambda_dae = PiecewiseLinearFn.from_string(args.lambda_dae) + + self.args = args + self.data = utils.split_paths(self.args.data) + if len(self.data) == 1: + shards = list(Path(self.data[0]).glob("shard*")) + if len(shards) > 0: + # keep this as strings, since it can also be a manifold path + old_data = self.data + self.data = [str(shard) for shard in shards] + logging.warning(f"Expanded data directory {old_data} to {self.data}") + + @classmethod + def setup_task(cls, args, **kwargs): + """Setup the task (e.g., load dictionaries). + + Args: + args (argparse.Namespace): parsed command-line arguments + """ + args.left_pad_source = options.eval_bool(args.left_pad_source) + args.left_pad_target = options.eval_bool(args.left_pad_target) + + paths = utils.split_paths(args.data) + assert len(paths) > 0 + assert args.mono_langs is not None + + mono_langs = args.mono_langs.split(",") + valid_lang_pairs = args.valid_lang_pairs.split(",") + + # load dictionary + dict_path = os.path.join(paths[0], "dict.txt") + common_dict = cls.load_dictionary(dict_path) + + return cls(args, common_dict, mono_langs, valid_lang_pairs) + + def load_dataset(self, split, epoch=1, combine=False, **kwargs) -> FairseqDataset: + """Load a given dataset split. + + Args: + split (str): name of the split (e.g., train, valid, test) + """ + if split == "train": + data_path = self.data[(epoch - 1) % len(self.data)] + dataset = self.load_train_dataset(data_path) + else: + # valid/test should always be the same. + dataset = self.load_translation_dataset(split, self.data[0]) + + self.datasets[split] = dataset + return dataset + + def load_train_dataset(self, data_path: str) -> FairseqDataset: + """The training dataset is made of backtranslation dataset and denoising dataset.""" + data = [] + for lang in self.mono_langs: + train_path = os.path.join(data_path, lang, "train") + # TODO: could we do the BT using denoise sample ? + # this would half the data loading work + data.append((f"{lang}-BT", self.load_bt_dataset(train_path, lang))) + data.append( + (f"{lang}-DENOISE", self.load_denoise_dataset(train_path, lang)) + ) + + return RoundRobinZipDatasets(OrderedDict(data)) + + def _langpair_dataset( + self, src: FairseqDataset, tgt: FairseqDataset + ) -> LanguagePairDataset: + return LanguagePairDataset( + src, + src.sizes, + self.dictionary, + tgt=tgt, + tgt_sizes=tgt.sizes, + tgt_dict=self.dictionary, + left_pad_source=self.args.left_pad_source, + left_pad_target=self.args.left_pad_target, + # TODO: should we shuffle ? we are already sorting batch by sizes so ? + # shuffle=True, + ) + + def _prepend_lang_bos_to_target( + self, dataset: LanguagePairDataset, lang: str + ) -> LanguagePairDataset: + bos = _lang_token_index(self.dictionary, lang) + return TransformEosLangPairDataset( + dataset, + src_eos=self.dictionary.eos(), + new_src_eos=self.dictionary.eos(), + tgt_bos=self.dictionary.eos(), + new_tgt_bos=bos, + ) + + def load_bt_dataset(self, data_path: str, lang: str) -> FairseqDataset: + """The BT dataset is generated with (tgt, tgt) pairs. + The actual translation to a (generated_src, tgt) pair + is done on the fly during training. + """ + mono_dataset = data_utils.load_indexed_dataset( + data_path, self.common_dict, self.args.dataset_impl + ) + assert mono_dataset is not None, f"No dataset found for {lang}" + + mono_dataset_src = PrependTokenDataset( + mono_dataset, _lang_token_index(self.dictionary, lang) + ) + + mono_dataset_bt = self._langpair_dataset(mono_dataset_src, mono_dataset) + logger.info( + f"mono_lang = {lang} " + f"lang token index = {_lang_token_index(self.dictionary, lang)} " + f"lang token = {_lang_token(lang)}" + ) + + mono_dataset_bt = self._prepend_lang_bos_to_target(mono_dataset_bt, lang) + return mono_dataset_bt + + def load_denoise_dataset(self, data_path: str, lang: str) -> FairseqDataset: + """Classic denoising dataset""" + dataset = data_utils.load_indexed_dataset( + data_path, self.common_dict, self.args.dataset_impl + ) + noisy_dataset = NoisingDataset( + dataset, + self.dictionary, + seed=1, + max_word_shuffle_distance=self.args.max_word_shuffle_distance, + word_dropout_prob=self.args.word_dropout_prob, + word_blanking_prob=self.args.word_blanking_prob, + ) + noisy_dataset = PrependTokenDataset( + noisy_dataset, _lang_token_index(self.dictionary, lang) + ) + + clean_dataset = data_utils.load_indexed_dataset( + data_path, self.common_dict, self.args.dataset_impl + ) + denoising_dataset = self._langpair_dataset(noisy_dataset, clean_dataset) + denoising_dataset = self._prepend_lang_bos_to_target(denoising_dataset, lang) + return denoising_dataset + + def load_translation_dataset( + self, split: str, data_path: str, combine: bool = False + ): + # only judging with one language pair for the moment, + # since ConcatDataset doesn't work as expected + assert len(self.valid_lang_pairs) == 1, "For now..." + valid_lang_pair = self.valid_lang_pairs[0] + src, tgt = valid_lang_pair.split("-") + + # use the same function than TranslationTask + src_tgt_dt = load_langpair_dataset( + data_path, + split, + src, + self.common_dict, + tgt, + self.common_dict, + combine=combine, + dataset_impl=self.args.dataset_impl, + upsample_primary=self.args.upsample_primary, + left_pad_source=self.args.left_pad_source, + left_pad_target=self.args.left_pad_target, + max_source_positions=self.args.max_source_positions, + max_target_positions=self.args.max_target_positions, + load_alignments=self.args.load_alignments, + truncate_source=self.args.truncate_source, + num_buckets=self.args.num_batch_buckets, + shuffle=(split != "test"), + prepend_bos_src=_lang_token_index(self.dictionary, src), + ) + + src_tgt_eos_dt = self._prepend_lang_bos_to_target(src_tgt_dt, tgt) + src_tgt_eos_dt.args = self.args + return src_tgt_eos_dt + + def build_dataset_for_inference(self, src_tokens, src_lengths, constraints=None): + raise NotImplementedError + + def build_model(self, args): + # torch.autograd.set_detect_anomaly(True) + model = super().build_model(args) + + add_secial_tokens_to_dict_and_model(self.common_dict, model, self.mono_langs) + + self.sequence_generators = {} + for mono_lang in self.mono_langs: + self.sequence_generators[mono_lang] = SequenceGenerator( + [model], + tgt_dict=self.dictionary, + beam_size=1, + max_len_a=1.3, + max_len_b=5, + min_len=5, + # keep 1 to be able to prepend bos + max_len=model.max_decoder_positions() - 1, + ) + + if getattr(args, "eval_bleu", False): + assert getattr(args, "eval_bleu_detok", None) is not None, ( + "--eval-bleu-detok is required if using --eval-bleu; " + "try --eval-bleu-detok=moses (or --eval-bleu-detok=space " + "to disable detokenization, e.g., when using sentencepiece)" + ) + detok_args = json.loads(getattr(args, "eval_bleu_detok_args", "{}") or "{}") + self.tokenizer = encoders.build_tokenizer( + Namespace( + tokenizer=getattr(args, "eval_bleu_detok", None), **detok_args + ) + ) + + gen_args = json.loads(getattr(args, "eval_bleu_args", "{}") or "{}") + self.bleu_sequence_generator = self.build_generator( + [model], Namespace(**gen_args) + ) + + return model + + def max_positions(self): + """Return the max sentence length allowed by the task.""" + return (self.args.max_source_positions, self.args.max_target_positions) + + @property + def dictionary(self): + """Return the source :class:`~fairseq.data.Dictionary`.""" + return self.common_dict + + def display_samples_once_in_a_while(self, smp, mono_lang, other_lang): + self._show_samples_ctr += 1 + if self._show_samples_ctr < self.SHOW_SAMPLES_INTERVAL: + return + self._show_samples_ctr = 0 + + ln = smp["net_input"]["src_tokens"].shape[0] + + logger.info( + f"(r:{self.args.distributed_rank}) : " + f"{other_lang} ---> {mono_lang} " + f"({other_lang} was generated by back-translation.) {ln} samples" + ) + + for i in range(min(ln, self.SHOW_SAMPLES_NUMBER)): + src_tokens = smp["net_input"]["src_tokens"][i] + tgt_tokens = smp["target"][i] + + src_str = self.dictionary.string(src_tokens, "sentencepiece") + tgt_str = self.dictionary.string(tgt_tokens, "sentencepiece") + logger.info( + f"\n{i}\t\t[{other_lang} generated] {src_str}\n" + f"\t\t[{mono_lang} original ] {tgt_str}\n" + f"\t\t[ src tokens] {src_tokens}\n" + ) + + def backtranslate_sample(self, smp, orig_lang, other_lang) -> None: + """ + * WARNING: smp is modified in place. + * At the start of this function, `smp` has the same input and target: + |--------------------------------------------------------| + | smp['net_input']['src_tokens'] | smp['target'] | + | (from data) __en__ hello world | __en__ hello world | + |--------------------------------------------------------| + + * We call generator.generate(smp, bos_token = token("ro")), + and copy the result as input + * At the end, `smp` has the translation to other language. + |--------------------------------------------------------| + | smp['net_input']['src_tokens'] | smp['target'] | + | (generated) __ro__ salut lume | __en__ hello world | + |--------------------------------------------------------| + + """ + bos_token = _lang_token_index(self.dictionary, other_lang) + generated = self.sequence_generators[orig_lang].generate( + models=[], sample=smp, bos_token=bos_token + ) + + max_lngth = max([gn[0]["tokens"].size(0) for gn in generated]) + net_input = smp["net_input"] + n_src_tokens = torch.empty( + size=(len(generated), max_lngth + 1), dtype=net_input["src_tokens"].dtype + ) + n_src_lengths = torch.empty( + len(generated), dtype=net_input["src_lengths"].dtype + ) + + for i, gn in enumerate(generated): + tokens = gn[0]["tokens"] + tokens_size = tokens.size(0) + padding_needed = max_lngth - tokens_size + tokens = torch.cat([tokens.new([bos_token]), tokens]) + tokens = F.pad(tokens, (0, padding_needed), value=self.dictionary.pad()) + n_src_tokens[i] = tokens + n_src_lengths[i] = tokens_size + 1 + + device = net_input["src_tokens"].device + # This seems to be important + del net_input["src_tokens"] + del net_input["src_lengths"] + net_input["src_tokens"] = n_src_tokens.to(device) + net_input["src_lengths"] = n_src_lengths.to(device) + + def generate(self, smp, model): + model.eval() + orig_lang = ( + self.dictionary[smp["net_input"]["src_tokens"][0][0]] + .replace(" ", "") + .replace("_", "") + ) + bos_token = smp["net_input"]["prev_output_tokens"][0][0] + with torch.no_grad(): + generated = self.sequence_generators[orig_lang].generate( + models=[model], sample=smp, bos_token=bos_token + ) + return generated + + def get_other_lang(self, lang): + # TODO: allow more complex mapping + if lang != self.mono_langs[0]: + return self.mono_langs[0] + if len(self.mono_langs) == 2: + return self.mono_langs[1] + return self.mono_langs[np.random.randint(1, len(self.mono_langs))] + + def train_step( + self, sample, model, criterion, optimizer, update_num, ignore_grad=False + ): + + model.train() + model.set_num_updates(update_num) + + agg_loss, agg_sample_size = 0.0, 0.0 + agg_logging_output: Dict[str, float] = defaultdict(float) + + dataset_keys = self.datasets["train"].datasets.keys() + + weights = { + "BT": self.lambda_bt(update_num), + "DENOISE": self.lambda_dae(update_num), + } + log_keys = {"BT": "bt_", "DENOISE": "dae_"} + + for dataset_key in dataset_keys: + smp = sample[dataset_key] + mono_lang, task_subtype = dataset_key.split("-") + if weights[task_subtype] == 0: + continue + + if task_subtype == "BT": + with torch.autograd.profiler.record_function("backtranslation"): + model.eval() + # TODO: Could we translate to several language at once ? + # this would allow to share encoder_out and maximize GPU usage. + other_lang = self.get_other_lang(mono_lang) + self.backtranslate_sample(smp, mono_lang, other_lang) + self.display_samples_once_in_a_while(smp, mono_lang, other_lang) + model.train() + + # Like in FairseqTask.train_step + with torch.autograd.profiler.record_function("forward"): + loss, sample_size, logging_output = criterion(model, smp) + loss *= weights[task_subtype] + if ignore_grad: + loss *= 0 + with torch.autograd.profiler.record_function("backward"): + optimizer.backward(loss) + + agg_loss += loss.item() + agg_sample_size += sample_size + for k in logging_output: + agg_logging_output[log_keys[task_subtype] + k] += logging_output[k] + agg_logging_output[k] += logging_output[k] + + return agg_loss, agg_sample_size, agg_logging_output + + def get_bos_token_from_sample(self, sample): + net_input = sample["net_input"] + source_lang_token_id = torch.unique(net_input["src_tokens"][:, 0]).item() + source_lang_token = self.dictionary[source_lang_token_id].replace("_", "") + target_lang_token_id = _lang_token_index( + self.dictionary, self.get_other_lang(source_lang_token) + ) + + return target_lang_token_id + + def reduce_metrics(self, logging_outputs, criterion): + super().reduce_metrics(logging_outputs, criterion) + bt_sample_size = sum(x.get("bt_sample_size", 0) for x in logging_outputs) + if bt_sample_size: + bt_loss_sum = sum(x.get("bt_loss", 0) for x in logging_outputs) + bt_loss_sum *= 1 / bt_sample_size / math.log(2) + metrics.log_scalar("bt_loss", bt_loss_sum, bt_sample_size, round=3) + + bt_nll_loss_sum = sum(x.get("bt_nll_loss", 0) for x in logging_outputs) + bt_ntokens = sum(x.get("bt_ntokens", 0) for x in logging_outputs) + bt_nll_loss_sum *= 1 / bt_ntokens / math.log(2) + metrics.log_scalar("bt_nll_loss", bt_nll_loss_sum, bt_ntokens, round=3) + metrics.log_derived( + "bt_ppl", lambda meters: utils.get_perplexity(meters["bt_nll_loss"].avg) + ) + + dae_sample_size = sum(x.get("dae_sample_size", 0) for x in logging_outputs) + if dae_sample_size: + dae_loss_sum = sum(x.get("dae_loss", 0) for x in logging_outputs) + dae_loss_sum *= 1 / dae_sample_size / math.log(2) + metrics.log_scalar("dae_loss", dae_loss_sum, dae_sample_size, round=3) + + dae_nll_loss_sum = sum(x.get("dae_nll_loss", 0) for x in logging_outputs) + dae_ntokens = sum(x.get("dae_ntokens", 0) for x in logging_outputs) + dae_nll_loss_sum *= 1 / dae_ntokens / math.log(2) + metrics.log_scalar("dae_nll_loss", dae_nll_loss_sum, dae_ntokens, round=3) + metrics.log_derived( + "dae_ppl", + lambda meters: utils.get_perplexity(meters["dae_nll_loss"].avg), + ) + + +@torch.no_grad() +def extend_embedding( + emb: nn.Module, new_vocab_size: int, copy_from_token_id: int +) -> None: + old_emb_data = emb.weight.data + (old_vocab_size, dim) = old_emb_data.shape + assert new_vocab_size >= old_vocab_size + + if new_vocab_size > old_vocab_size: + emb.weight.data = torch.zeros((new_vocab_size, dim)) + emb.weight.data[:old_vocab_size, :] = old_emb_data + # initialize new embeddings + emb.weight.data[old_vocab_size:, :] = old_emb_data[copy_from_token_id] + if hasattr(emb, "num_embeddings"): + emb.num_embeddings = new_vocab_size + if hasattr(emb, "out_features"): + emb.out_features = new_vocab_size + + if getattr(emb, "bias", None) is None: + return + + # Fix the bias. + # Bias shape can be different from the previous vocab size + # if the weight matrix was shared and alread extended but not the bias. + (old_vocab_size,) = emb.bias.shape + assert new_vocab_size >= old_vocab_size + if new_vocab_size > old_vocab_size: + old_bias = emb.bias.data + new_bias = torch.zeros( + (new_vocab_size,), dtype=old_bias.dtype, device=old_bias.device + ) + new_bias[:old_vocab_size] = old_bias + emb.bias.data = new_bias + + +def add_secial_tokens_to_dict_and_model( + dictionary: "fairseq.data.Dictionary", + model: nn.Module, + mono_langs: Sequence[str], +) -> None: + embs = model.encoder.embed_tokens + vocab_size, embedding_dim = embs.weight.shape + + # The model may or may not have a '<mask>' embedding yet + assert ( + len(dictionary) <= vocab_size <= len(dictionary) + 1 + ), f"Dictionary len ({len(dictionary)}) doesn't match embs shape ({embs.weight.shape})" + # TODO: we should reuse the pretrained model dict which already has <mask> + dictionary.add_symbol("<mask>") + + for lang in mono_langs: + lang_token = _lang_token(lang) + dictionary.add_symbol(lang_token) + logger.info( + f"dictionary: {len(dictionary)} -> {vocab_size} tokens " + f"after adding {len(mono_langs)} lang tokens." + ) + + if len(dictionary) <= vocab_size: + return + + extend_embedding(embs, len(dictionary), dictionary.bos()) + dec_embs = model.decoder.embed_tokens + extend_embedding(dec_embs, len(dictionary), dictionary.bos()) + lm_head = model.decoder.output_projection + extend_embedding(lm_head, len(dictionary), dictionary.bos()) + assert lm_head.weight.shape == (len(dictionary), embedding_dim) + + +def _lang_token(lang: str) -> str: + return f"__{lang}__" + + +def _lang_token_index(dictionary, lang: str) -> int: + return dictionary.index(_lang_token(lang)) + + +@contextlib.contextmanager +def assert_weights_have_changed(model: nn.Module): + def checksum(model: nn.Module) -> float: + return sum(p.sum().item() for p in model.parameters()) + + initial_checksum = checksum(model) + yield model + final_checksum = checksum(model) + logger.info( + f"initial_checksum={initial_checksum} -> final_checksum={final_checksum}" + ) + assert initial_checksum != final_checksum, "Model hasn't changed !" diff --git a/fairseq/tasks/semisupervised_translation.py b/fairseq/tasks/semisupervised_translation.py new file mode 100644 index 0000000000000000000000000000000000000000..b2f9bf9a733d94e50b588e4316b4a02e1c8bcf51 --- /dev/null +++ b/fairseq/tasks/semisupervised_translation.py @@ -0,0 +1,485 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import os +from collections import OrderedDict + +from fairseq import utils +from fairseq.data import ( + BacktranslationDataset, + IndexedCachedDataset, + IndexedDataset, + IndexedRawTextDataset, + LanguagePairDataset, + NoisingDataset, + RoundRobinZipDatasets, + data_utils, + indexed_dataset, +) +from fairseq.models import FairseqMultiModel +from fairseq.sequence_generator import SequenceGenerator + +from . import register_task +from .multilingual_translation import MultilingualTranslationTask + + +logger = logging.getLogger(__name__) + + +def _get_bt_dataset_key(lang_pair): + return "bt:" + lang_pair + + +def _get_denoising_dataset_key(lang_pair): + return "denoising:" + lang_pair + + +# ported from UnsupervisedMT +def parse_lambda_config(x): + """ + Parse the configuration of lambda coefficient (for scheduling). + x = "3" # lambda will be a constant equal to x + x = "0:1,1000:0" # lambda will start from 1 and linearly decrease + # to 0 during the first 1000 iterations + x = "0:0,1000:0,2000:1" # lambda will be equal to 0 for the first 1000 + # iterations, then will linearly increase to 1 until iteration 2000 + """ + split = x.split(",") + if len(split) == 1: + return float(x), None + else: + split = [s.split(os.pathsep) for s in split] + assert all(len(s) == 2 for s in split) + assert all(k.isdigit() for k, _ in split) + assert all( + int(split[i][0]) < int(split[i + 1][0]) for i in range(len(split) - 1) + ) + return float(split[0][1]), [(int(k), float(v)) for k, v in split] + + +@register_task("semisupervised_translation") +class SemisupervisedTranslationTask(MultilingualTranslationTask): + """A task for training multiple translation models simultaneously. + + We iterate round-robin over batches from multiple language pairs, ordered + according to the `--lang-pairs` argument. + + The training loop is roughly: + + for i in range(len(epoch)): + for lang_pair in args.lang_pairs: + batch = next_batch_for_lang_pair(lang_pair) + loss = criterion(model_for_lang_pair(lang_pair), batch) + loss.backward() + optimizer.step() + + In practice, `next_batch_for_lang_pair` is abstracted in a FairseqDataset + (e.g., `RoundRobinZipDatasets`) and `model_for_lang_pair` is a model that + implements the `FairseqMultiModel` interface. + + During inference it is required to specify a single `--source-lang` and + `--target-lang`, instead of `--lang-pairs`. + """ + + @staticmethod + def add_args(parser): + """Add task-specific arguments to the parser.""" + # fmt: off + MultilingualTranslationTask.add_args(parser) + parser.add_argument('--lambda-parallel-config', default="1.0", type=str, metavar='CONFIG', + help='cross-entropy reconstruction coefficient (parallel data). ' + 'use fixed weight during training if set to floating point number. ' + 'use piecewise linear function over number of updates to schedule the ' + 'weight with the format: w0:step0,w1:step1,...') + parser.add_argument('--lambda-denoising-config', default="0.0", type=str, metavar='CONFIG', + help='Cross-entropy reconstruction coefficient (denoising autoencoding)' + 'use fixed weight during training if set to floating point number. ' + 'use piecewise linear function over number of updates to schedule the ' + 'weight with the format: w0:step0,w1:step1,...') + parser.add_argument('--lambda-otf-bt-config', default="0.0", type=str, metavar='CONFIG', + help='cross-entropy reconstruction coefficient (on-the-fly back-translation parallel data)' + 'use fixed weight during training if set to floating point number. ' + 'use piecewise linear function over number of updates to schedule the ' + 'weight with the format: w0:step0,w1:step1,...') + parser.add_argument('--bt-max-len-a', default=1.1, type=float, metavar='N', + help='generate back-translated sequences of maximum length ax + b, where x is the ' + 'source length') + parser.add_argument('--bt-max-len-b', default=10.0, type=float, metavar='N', + help='generate back-translated sequences of maximum length ax + b, where x is the ' + 'source length') + parser.add_argument('--bt-beam-size', default=1, type=int, metavar='N', + help='beam size used in beam search of online back-translation') + parser.add_argument('--max-word-shuffle-distance', default=3.0, type=float, metavar='N', + help='maximum word shuffle distance for denoising autoencoding data generation') + parser.add_argument('--word-dropout-prob', default=0.1, type=float, metavar='N', + help='word dropout probability for denoising autoencoding data generation') + parser.add_argument('--word-blanking-prob', default=0.2, type=float, metavar='N', + help='word blanking probability for denoising autoencoding data generation') + # fmt: on + + def __init__(self, args, dicts, training): + super().__init__(args, dicts, training) + self.lambda_parallel, self.lambda_parallel_steps = parse_lambda_config( + args.lambda_parallel_config + ) + self.lambda_otf_bt, self.lambda_otf_bt_steps = parse_lambda_config( + args.lambda_otf_bt_config + ) + self.lambda_denoising, self.lambda_denoising_steps = parse_lambda_config( + args.lambda_denoising_config + ) + if self.lambda_denoising > 0.0 or self.lambda_denoising_steps is not None: + denoising_lang_pairs = [ + "%s-%s" % (tgt, tgt) + for tgt in {lang_pair.split("-")[1] for lang_pair in args.lang_pairs} + ] + self.model_lang_pairs = self.model_lang_pairs + denoising_lang_pairs + self.backtranslate_datasets = {} + self.backtranslators = {} + + @classmethod + def setup_task(cls, args, **kwargs): + dicts, training = MultilingualTranslationTask.prepare(args, **kwargs) + return cls(args, dicts, training) + + def load_dataset(self, split, epoch=1, **kwargs): + """Load a dataset split.""" + paths = utils.split_paths(self.args.data) + assert len(paths) > 0 + data_path = paths[(epoch - 1) % len(paths)] + + def split_exists(split, src, tgt, lang): + if src is not None: + filename = os.path.join( + data_path, "{}.{}-{}.{}".format(split, src, tgt, lang) + ) + else: + filename = os.path.join( + data_path, "{}.{}-None.{}".format(split, src, tgt) + ) + return indexed_dataset.dataset_exists(filename, impl=self.args.dataset_impl) + + def load_indexed_dataset(path, dictionary): + return data_utils.load_indexed_dataset( + path, dictionary, self.args.dataset_impl + ) + + # load parallel datasets + src_datasets, tgt_datasets = {}, {} + if ( + self.lambda_parallel > 0.0 + or self.lambda_parallel_steps is not None + or not split.startswith("train") + ): + for lang_pair in self.lang_pairs: + src, tgt = lang_pair.split("-") + if split_exists(split, src, tgt, src): + prefix = os.path.join( + data_path, "{}.{}-{}.".format(split, src, tgt) + ) + elif split_exists(split, tgt, src, src): + prefix = os.path.join( + data_path, "{}.{}-{}.".format(split, tgt, src) + ) + else: + continue + src_datasets[lang_pair] = load_indexed_dataset( + prefix + src, self.dicts[src] + ) + tgt_datasets[lang_pair] = load_indexed_dataset( + prefix + tgt, self.dicts[tgt] + ) + logger.info( + "parallel-{} {} {} examples".format( + data_path, split, len(src_datasets[lang_pair]) + ) + ) + if len(src_datasets) == 0: + raise FileNotFoundError( + "Dataset not found: {} ({})".format(split, data_path) + ) + + # back translation datasets + backtranslate_datasets = {} + if ( + self.lambda_otf_bt > 0.0 or self.lambda_otf_bt_steps is not None + ) and split.startswith("train"): + for lang_pair in self.lang_pairs: + src, tgt = lang_pair.split("-") + if not split_exists(split, tgt, None, tgt): + raise FileNotFoundError( + "Dataset not found: backtranslation {} ({})".format( + split, data_path + ) + ) + filename = os.path.join( + data_path, "{}.{}-None.{}".format(split, tgt, tgt) + ) + dataset = load_indexed_dataset(filename, self.dicts[tgt]) + lang_pair_dataset_tgt = LanguagePairDataset( + dataset, + dataset.sizes, + self.dicts[tgt], + left_pad_source=self.args.left_pad_source, + left_pad_target=self.args.left_pad_target, + ) + lang_pair_dataset = LanguagePairDataset( + dataset, + dataset.sizes, + src_dict=self.dicts[src], + tgt=dataset, + tgt_sizes=dataset.sizes, + tgt_dict=self.dicts[tgt], + left_pad_source=self.args.left_pad_source, + left_pad_target=self.args.left_pad_target, + ) + backtranslate_datasets[lang_pair] = BacktranslationDataset( + tgt_dataset=self.alter_dataset_langtok( + lang_pair_dataset_tgt, + src_eos=self.dicts[tgt].eos(), + src_lang=tgt, + tgt_lang=src, + ), + backtranslation_fn=self.backtranslators[lang_pair], + src_dict=self.dicts[src], + tgt_dict=self.dicts[tgt], + output_collater=self.alter_dataset_langtok( + lang_pair_dataset=lang_pair_dataset, + src_eos=self.dicts[src].eos(), + src_lang=src, + tgt_eos=self.dicts[tgt].eos(), + tgt_lang=tgt, + ).collater, + ) + logger.info( + "backtranslate-{}: {} {} {} examples".format( + tgt, + data_path, + split, + len(backtranslate_datasets[lang_pair]), + ) + ) + self.backtranslate_datasets[lang_pair] = backtranslate_datasets[ + lang_pair + ] + + # denoising autoencoder + noising_datasets = {} + if ( + self.lambda_denoising > 0.0 or self.lambda_denoising_steps is not None + ) and split.startswith("train"): + for lang_pair in self.lang_pairs: + _, tgt = lang_pair.split("-") + if not split_exists(split, tgt, None, tgt): + continue + filename = os.path.join( + data_path, "{}.{}-None.{}".format(split, tgt, tgt) + ) + tgt_dataset1 = load_indexed_dataset(filename, self.dicts[tgt]) + tgt_dataset2 = load_indexed_dataset(filename, self.dicts[tgt]) + noising_dataset = NoisingDataset( + tgt_dataset1, + self.dicts[tgt], + seed=1, + max_word_shuffle_distance=self.args.max_word_shuffle_distance, + word_dropout_prob=self.args.word_dropout_prob, + word_blanking_prob=self.args.word_blanking_prob, + ) + noising_datasets[lang_pair] = self.alter_dataset_langtok( + LanguagePairDataset( + noising_dataset, + tgt_dataset1.sizes, + self.dicts[tgt], + tgt_dataset2, + tgt_dataset2.sizes, + self.dicts[tgt], + left_pad_source=self.args.left_pad_source, + left_pad_target=self.args.left_pad_target, + ), + src_eos=self.dicts[tgt].eos(), + src_lang=tgt, + tgt_eos=self.dicts[tgt].eos(), + tgt_lang=tgt, + ) + logger.info( + "denoising-{}: {} {} {} examples".format( + tgt, + data_path, + split, + len(noising_datasets[lang_pair]), + ) + ) + + def language_pair_dataset(lang_pair): + src, tgt = lang_pair.split("-") + src_dataset, tgt_dataset = src_datasets[lang_pair], tgt_datasets[lang_pair] + return self.alter_dataset_langtok( + LanguagePairDataset( + src_dataset, + src_dataset.sizes, + self.dicts[src], + tgt_dataset, + tgt_dataset.sizes, + self.dicts[tgt], + left_pad_source=self.args.left_pad_source, + left_pad_target=self.args.left_pad_target, + ), + self.dicts[src].eos(), + src, + self.dicts[tgt].eos(), + tgt, + ) + + self.datasets[split] = RoundRobinZipDatasets( + OrderedDict( + [ + (lang_pair, language_pair_dataset(lang_pair)) + for lang_pair in src_datasets.keys() + ] + + [ + (_get_bt_dataset_key(lang_pair), dataset) + for lang_pair, dataset in backtranslate_datasets.items() + ] + + [ + (_get_denoising_dataset_key(lang_pair), dataset) + for lang_pair, dataset in noising_datasets.items() + ] + ), + eval_key=None + if self.training + else "%s-%s" % (self.args.source_lang, self.args.target_lang), + ) + + def build_model(self, args): + from fairseq import models + + model = models.build_model(args, self) + if not isinstance(model, FairseqMultiModel): + raise ValueError( + "SemisupervisedTranslationTask requires a FairseqMultiModel architecture" + ) + + # create SequenceGenerator for each model that has backtranslation dependency on it + self.sequence_generators = {} + if ( + self.lambda_otf_bt > 0.0 or self.lambda_otf_bt_steps is not None + ) and self.training: + for lang_pair in self.lang_pairs: + src, tgt = lang_pair.split("-") + key = "{}-{}".format(tgt, src) + self.sequence_generators[key] = SequenceGenerator( + [model.models[key]], + tgt_dict=self.dicts[src], + beam_size=args.bt_beam_size, + max_len_a=args.bt_max_len_a, + max_len_b=args.bt_max_len_b, + ) + decoder_lang_tok_idx = self.get_decoder_langtok(src) + + def backtranslate_fn( + sample, + model=model.models[key], + bos_token=decoder_lang_tok_idx, + sequence_generator=self.sequence_generators[key], + ): + return sequence_generator.generate( + [model], + sample, + bos_token=bos_token, + ) + + self.backtranslators[lang_pair] = backtranslate_fn + + return model + + def train_step( + self, sample, model, criterion, optimizer, update_num, ignore_grad=False + ): + model.train() + + if update_num > 0: + self.update_step(update_num) + + agg_loss, agg_sample_size, agg_logging_output = 0.0, 0.0, {} + + def forward_backward(model, samples, logging_output_key, weight): + nonlocal agg_loss, agg_sample_size, agg_logging_output + if samples is None or len(samples) == 0: + return + loss, sample_size, logging_output = criterion(model, samples) + if ignore_grad: + loss *= 0 + else: + loss *= weight + optimizer.backward(loss) + agg_loss += loss.detach().item() + # TODO make summing of the sample sizes configurable + agg_sample_size += sample_size + for k in logging_output: + agg_logging_output[k] += logging_output[k] + agg_logging_output[logging_output_key] += logging_output[k] + + if self.lambda_parallel > 0.0: + for lang_pair in self.lang_pairs: + forward_backward( + model.models[lang_pair], + sample[lang_pair], + lang_pair, + self.lambda_parallel, + ) + + if self.lambda_otf_bt > 0.0: + for lang_pair in self.lang_pairs: + sample_key = _get_bt_dataset_key(lang_pair) + forward_backward( + model.models[lang_pair], + sample[sample_key], + sample_key, + self.lambda_otf_bt, + ) + + if self.lambda_denoising > 0.0: + for lang_pair in self.lang_pairs: + _, tgt = lang_pair.split("-") + sample_key = _get_denoising_dataset_key(lang_pair) + forward_backward( + model.models["{0}-{0}".format(tgt)], + sample[sample_key], + sample_key, + self.lambda_denoising, + ) + + return agg_loss, agg_sample_size, agg_logging_output + + def update_step(self, num_updates): + def lambda_step_func(config, n_iter): + """ + Update a lambda value according to its schedule configuration. + """ + ranges = [ + i + for i in range(len(config) - 1) + if config[i][0] <= n_iter < config[i + 1][0] + ] + if len(ranges) == 0: + assert n_iter >= config[-1][0] + return config[-1][1] + assert len(ranges) == 1 + i = ranges[0] + x_a, y_a = config[i] + x_b, y_b = config[i + 1] + return y_a + (n_iter - x_a) * float(y_b - y_a) / float(x_b - x_a) + + if self.lambda_parallel_steps is not None: + self.lambda_parallel = lambda_step_func( + self.lambda_parallel_steps, num_updates + ) + if self.lambda_denoising_steps is not None: + self.lambda_denoising = lambda_step_func( + self.lambda_denoising_steps, num_updates + ) + if self.lambda_otf_bt_steps is not None: + self.lambda_otf_bt = lambda_step_func(self.lambda_otf_bt_steps, num_updates) diff --git a/fairseq/tasks/sentence_prediction.py b/fairseq/tasks/sentence_prediction.py new file mode 100644 index 0000000000000000000000000000000000000000..6732728de981da7174eae32ecaf4c47901d65399 --- /dev/null +++ b/fairseq/tasks/sentence_prediction.py @@ -0,0 +1,286 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import os + +import numpy as np +from fairseq import utils +from fairseq.data import ( + ConcatSentencesDataset, + Dictionary, + IdDataset, + NestedDictionaryDataset, + NumelDataset, + NumSamplesDataset, + OffsetTokensDataset, + PrependTokenDataset, + RawLabelDataset, + RightPadDataset, + RollDataset, + SortDataset, + StripTokenDataset, + data_utils, +) +from fairseq.data.shorten_dataset import maybe_shorten_dataset +from fairseq.tasks import LegacyFairseqTask, register_task + + +logger = logging.getLogger(__name__) + + +@register_task("sentence_prediction") +class SentencePredictionTask(LegacyFairseqTask): + """ + Sentence (or sentence pair) prediction (classification or regression) task. + + Args: + dictionary (Dictionary): the dictionary for the input of the task + """ + + @staticmethod + def add_args(parser): + """Add task-specific arguments to the parser.""" + parser.add_argument("data", metavar="FILE", help="file prefix for data") + parser.add_argument( + "--num-classes", + type=int, + default=-1, + help="number of classes or regression targets", + ) + parser.add_argument( + "--init-token", + type=int, + default=None, + help="add token at the beginning of each batch item", + ) + parser.add_argument( + "--separator-token", + type=int, + default=None, + help="add separator token between inputs", + ) + parser.add_argument("--regression-target", action="store_true", default=False) + parser.add_argument("--no-shuffle", action="store_true", default=False) + parser.add_argument( + "--shorten-method", + default="none", + choices=["none", "truncate", "random_crop"], + help="if not none, shorten sequences that exceed --tokens-per-sample", + ) + parser.add_argument( + "--shorten-data-split-list", + default="", + help="comma-separated list of dataset splits to apply shortening to, " + 'e.g., "train,valid" (default: all dataset splits)', + ) + parser.add_argument( + "--add-prev-output-tokens", + action="store_true", + default=False, + help="add prev_output_tokens to sample, used for encoder-decoder arch", + ) + + def __init__(self, args, data_dictionary, label_dictionary): + super().__init__(args) + self.dictionary = data_dictionary + self._label_dictionary = label_dictionary + if not hasattr(args, "max_positions"): + self._max_positions = ( + args.max_source_positions, + args.max_target_positions, + ) + else: + self._max_positions = args.max_positions + args.tokens_per_sample = self._max_positions + + @classmethod + def load_dictionary(cls, args, filename, source=True): + """Load the dictionary from the filename + + Args: + filename (str): the filename + """ + dictionary = Dictionary.load(filename) + dictionary.add_symbol("<mask>") + return dictionary + + @classmethod + def setup_task(cls, args, **kwargs): + assert args.num_classes > 0, "Must set --num-classes" + + # load data dictionary + data_dict = cls.load_dictionary( + args, + os.path.join(args.data, "input0", "dict.txt"), + source=True, + ) + logger.info("[input] dictionary: {} types".format(len(data_dict))) + + # load label dictionary + if not args.regression_target: + label_dict = cls.load_dictionary( + args, + os.path.join(args.data, "label", "dict.txt"), + source=False, + ) + logger.info("[label] dictionary: {} types".format(len(label_dict))) + else: + label_dict = data_dict + return cls(args, data_dict, label_dict) + + def load_dataset(self, split, combine=False, **kwargs): + """Load a given dataset split (e.g., train, valid, test).""" + + def get_path(key, split): + return os.path.join(self.args.data, key, split) + + def make_dataset(key, dictionary): + split_path = get_path(key, split) + + try: + dataset = data_utils.load_indexed_dataset( + split_path, + dictionary, + self.args.dataset_impl, + combine=combine, + ) + except Exception as e: + if "StorageException: [404] Path not found" in str(e): + logger.warning(f"dataset {e} not found") + dataset = None + else: + raise e + return dataset + + input0 = make_dataset("input0", self.source_dictionary) + assert input0 is not None, "could not find dataset: {}".format( + get_path("input0", split) + ) + input1 = make_dataset("input1", self.source_dictionary) + + if self.args.init_token is not None: + input0 = PrependTokenDataset(input0, self.args.init_token) + + if input1 is None: + src_tokens = input0 + else: + if self.args.separator_token is not None: + input1 = PrependTokenDataset(input1, self.args.separator_token) + + src_tokens = ConcatSentencesDataset(input0, input1) + + with data_utils.numpy_seed(self.args.seed): + shuffle = np.random.permutation(len(src_tokens)) + + src_tokens = maybe_shorten_dataset( + src_tokens, + split, + self.args.shorten_data_split_list, + self.args.shorten_method, + self.max_positions(), + self.args.seed, + ) + + dataset = { + "id": IdDataset(), + "net_input": { + "src_tokens": RightPadDataset( + src_tokens, + pad_idx=self.source_dictionary.pad(), + ), + "src_lengths": NumelDataset(src_tokens, reduce=False), + }, + "nsentences": NumSamplesDataset(), + "ntokens": NumelDataset(src_tokens, reduce=True), + } + + if self.args.add_prev_output_tokens: + prev_tokens_dataset = RightPadDataset( + RollDataset(src_tokens, 1), + pad_idx=self.dictionary.pad(), + ) + dataset["net_input"].update( + prev_output_tokens=prev_tokens_dataset, + ) + + if not self.args.regression_target: + label_dataset = make_dataset("label", self.label_dictionary) + if label_dataset is not None: + dataset.update( + target=OffsetTokensDataset( + StripTokenDataset( + label_dataset, + id_to_strip=self.label_dictionary.eos(), + ), + offset=-self.label_dictionary.nspecial, + ) + ) + else: + label_path = "{0}.label".format(get_path("label", split)) + if os.path.exists(label_path): + + def parse_regression_target(i, line): + values = line.split() + assert ( + len(values) == self.args.num_classes + ), f'expected num_classes={self.args.num_classes} regression target values on line {i}, found: "{line}"' + return [float(x) for x in values] + + with open(label_path) as h: + dataset.update( + target=RawLabelDataset( + [ + parse_regression_target(i, line.strip()) + for i, line in enumerate(h.readlines()) + ] + ) + ) + + nested_dataset = NestedDictionaryDataset( + dataset, + sizes=[src_tokens.sizes], + ) + + if self.args.no_shuffle: + dataset = nested_dataset + else: + dataset = SortDataset( + nested_dataset, + # shuffle + sort_order=[shuffle], + ) + + logger.info("Loaded {0} with #samples: {1}".format(split, len(dataset))) + + self.datasets[split] = dataset + return self.datasets[split] + + def build_model(self, args): + from fairseq import models + + model = models.build_model(args, self) + + model.register_classification_head( + getattr(args, "classification_head_name", "sentence_classification_head"), + num_classes=self.args.num_classes, + ) + + return model + + def max_positions(self): + return self._max_positions + + @property + def source_dictionary(self): + return self.dictionary + + @property + def target_dictionary(self): + return self.dictionary + + @property + def label_dictionary(self): + return self._label_dictionary diff --git a/fairseq/tasks/sentence_ranking.py b/fairseq/tasks/sentence_ranking.py new file mode 100644 index 0000000000000000000000000000000000000000..bed44f34e5f8e506b6ae7ba30ddaa661bf4a7522 --- /dev/null +++ b/fairseq/tasks/sentence_ranking.py @@ -0,0 +1,219 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import os + +import numpy as np +from fairseq import utils +from fairseq.data import ( + ConcatSentencesDataset, + Dictionary, + IdDataset, + NestedDictionaryDataset, + NumelDataset, + NumSamplesDataset, + PrependTokenDataset, + RawLabelDataset, + RightPadDataset, + SortDataset, + TruncateDataset, + data_utils, +) +from fairseq.data.shorten_dataset import maybe_shorten_dataset +from fairseq.tasks import LegacyFairseqTask, register_task + + +logger = logging.getLogger(__name__) + + +@register_task("sentence_ranking") +class SentenceRankingTask(LegacyFairseqTask): + """ + Ranking task on multiple sentences. + + Args: + dictionary (Dictionary): the dictionary for the input of the task + """ + + @staticmethod + def add_args(parser): + """Add task-specific arguments to the parser.""" + parser.add_argument("data", metavar="FILE", help="file prefix for data") + parser.add_argument( + "--num-classes", type=int, help="number of sentences to be ranked" + ) + parser.add_argument( + "--init-token", + type=int, + help="add token at the beginning of each batch item", + ) + parser.add_argument( + "--separator-token", type=int, help="add separator token between inputs" + ) + parser.add_argument("--no-shuffle", action="store_true") + parser.add_argument( + "--shorten-method", + default="none", + choices=["none", "truncate", "random_crop"], + help="if not none, shorten sequences that exceed --tokens-per-sample", + ) + parser.add_argument( + "--shorten-data-split-list", + default="", + help="comma-separated list of dataset splits to apply shortening to, " + 'e.g., "train,valid" (default: all dataset splits)', + ) + parser.add_argument( + "--max-option-length", type=int, help="max length for each option" + ) + + def __init__(self, args, dictionary): + super().__init__(args) + self.dictionary = dictionary + + @classmethod + def load_dictionary(cls, args, filename, source=True): + """Load the dictionary from the filename + + Args: + filename (str): the filename + """ + dictionary = Dictionary.load(filename) + dictionary.add_symbol("<mask>") + return dictionary + + @classmethod + def setup_task(cls, args, **kwargs): + assert ( + args.criterion == "sentence_ranking" + ), "Must set --criterion=sentence_ranking" + + # load data dictionary + data_dict = cls.load_dictionary( + args, + os.path.join(args.data, "input0", "dict.txt"), + source=True, + ) + logger.info("[input] dictionary: {} types".format(len(data_dict))) + return SentenceRankingTask(args, data_dict) + + def load_dataset(self, split, combine=False, **kwargs): + """Load a given dataset split (e.g., train, valid, test).""" + + def get_path(type, split): + return os.path.join(self.args.data, type, split) + + def make_dataset(type, dictionary): + split_path = get_path(type, split) + + dataset = data_utils.load_indexed_dataset( + split_path, + self.source_dictionary, + self.args.dataset_impl, + combine=combine, + ) + return dataset + + input0 = make_dataset("input0", self.source_dictionary) + input_options = [ + make_dataset("input{idx}".format(idx=idx + 1), self.source_dictionary) + for idx in range(self.args.num_classes) + ] + + if self.args.separator_token is not None: + input0 = PrependTokenDataset(input0, self.args.separator_token) + + src_tokens = [] + for input_option in input_options: + if self.args.init_token is not None: + input_option = PrependTokenDataset(input_option, self.args.init_token) + if self.args.max_option_length is not None: + input_option = TruncateDataset( + input_option, self.args.max_option_length + ) + src_token = ConcatSentencesDataset(input_option, input0) + src_token = maybe_shorten_dataset( + src_token, + split, + self.args.shorten_data_split_list, + self.args.shorten_method, + self.args.max_positions, + self.args.seed, + ) + src_tokens.append(src_token) + + with data_utils.numpy_seed(self.args.seed): + shuffle = np.random.permutation(len(src_tokens[0])) + + dataset = { + "id": IdDataset(), + "nsentences": NumSamplesDataset(), + "ntokens": NumelDataset(src_tokens[0], reduce=True), + } + + for src_token_idx in range(len(src_tokens)): + dataset.update( + { + "net_input{idx}".format(idx=src_token_idx + 1): { + "src_tokens": RightPadDataset( + src_tokens[src_token_idx], + pad_idx=self.source_dictionary.pad(), + ), + "src_lengths": NumelDataset( + src_tokens[src_token_idx], reduce=False + ), + } + } + ) + + label_path = "{}.label".format(get_path("label", split)) + if os.path.exists(label_path): + with open(label_path) as h: + dataset.update( + target=RawLabelDataset([int(x.strip()) for x in h.readlines()]) + ) + + nested_dataset = NestedDictionaryDataset( + dataset, + sizes=[np.maximum.reduce([src_token.sizes for src_token in src_tokens])], + ) + + if self.args.no_shuffle: + dataset = nested_dataset + else: + dataset = SortDataset( + nested_dataset, + # shuffle + sort_order=[shuffle], + ) + + logger.info("Loaded {0} with #samples: {1}".format(split, len(dataset))) + + self.datasets[split] = dataset + return self.datasets[split] + + def build_model(self, args): + from fairseq import models + + model = models.build_model(args, self) + + model.register_classification_head( + getattr(args, "ranking_head_name", "sentence_classification_head"), + num_classes=1, + ) + + return model + + def max_positions(self): + return self.args.max_positions + + @property + def source_dictionary(self): + return self.dictionary + + @property + def target_dictionary(self): + return self.dictionary diff --git a/fairseq/tasks/simultaneous_translation.py b/fairseq/tasks/simultaneous_translation.py new file mode 100644 index 0000000000000000000000000000000000000000..11c7dc1ea966a54f8915ef164377e40f90e851a1 --- /dev/null +++ b/fairseq/tasks/simultaneous_translation.py @@ -0,0 +1,42 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +from fairseq.tasks import register_task +from fairseq.tasks.speech_to_text import SpeechToTextTask +from fairseq.tasks.translation import ( + TranslationTask, TranslationConfig +) + +try: + import examples.simultaneous_translation # noqa + import_successful = True +except BaseException: + import_successful = False + + +logger = logging.getLogger(__name__) + + +def check_import(flag): + if not flag: + raise ImportError( + "'examples.simultaneous_translation' is not correctly imported. " + "Please considering `pip install -e $FAIRSEQ_DIR`." + ) + + +@register_task("simul_speech_to_text") +class SimulSpeechToTextTask(SpeechToTextTask): + def __init__(self, args, tgt_dict): + check_import(import_successful) + super().__init__(args, tgt_dict) + + +@register_task("simul_text_to_text", dataclass=TranslationConfig) +class SimulTextToTextTask(TranslationTask): + def __init__(self, cfg, src_dict, tgt_dict): + check_import(import_successful) + super().__init__(cfg, src_dict, tgt_dict) diff --git a/fairseq/tasks/speech_to_text.py b/fairseq/tasks/speech_to_text.py new file mode 100644 index 0000000000000000000000000000000000000000..8bdf21564367d3647d582c72a6c3c9924760933e --- /dev/null +++ b/fairseq/tasks/speech_to_text.py @@ -0,0 +1,149 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import os.path as op +from argparse import Namespace + +from fairseq.data import Dictionary, encoders +from fairseq.data.audio.speech_to_text_dataset import ( + S2TDataConfig, + SpeechToTextDataset, + SpeechToTextDatasetCreator, + get_features_or_waveform +) +from fairseq.tasks import LegacyFairseqTask, register_task + + +logger = logging.getLogger(__name__) + + +@register_task("speech_to_text") +class SpeechToTextTask(LegacyFairseqTask): + @staticmethod + def add_args(parser): + parser.add_argument("data", help="manifest root path") + parser.add_argument( + "--config-yaml", + type=str, + default="config.yaml", + help="Configuration YAML filename (under manifest root)", + ) + parser.add_argument( + "--max-source-positions", + default=6000, + type=int, + metavar="N", + help="max number of tokens in the source sequence", + ) + parser.add_argument( + "--max-target-positions", + default=1024, + type=int, + metavar="N", + help="max number of tokens in the target sequence", + ) + + def __init__(self, args, tgt_dict): + super().__init__(args) + self.tgt_dict = tgt_dict + self.data_cfg = S2TDataConfig(op.join(args.data, args.config_yaml)) + + @classmethod + def setup_task(cls, args, **kwargs): + data_cfg = S2TDataConfig(op.join(args.data, args.config_yaml)) + dict_path = op.join(args.data, data_cfg.vocab_filename) + if not op.isfile(dict_path): + raise FileNotFoundError(f"Dict not found: {dict_path}") + tgt_dict = Dictionary.load(dict_path) + logger.info( + f"dictionary size ({data_cfg.vocab_filename}): " f"{len(tgt_dict):,}" + ) + + if getattr(args, "train_subset", None) is not None: + if not all(s.startswith("train") for s in args.train_subset.split(",")): + raise ValueError('Train splits should be named like "train*".') + return cls(args, tgt_dict) + + def build_criterion(self, args): + from fairseq import criterions + + if self.data_cfg.prepend_tgt_lang_tag and args.ignore_prefix_size != 1: + raise ValueError( + 'Please set "--ignore-prefix-size 1" since ' + "target language ID token is prepended as BOS." + ) + return criterions.build_criterion(args, self) + + def load_dataset(self, split, epoch=1, combine=False, **kwargs): + is_train_split = split.startswith("train") + pre_tokenizer = self.build_tokenizer(self.args) + bpe_tokenizer = self.build_bpe(self.args) + self.datasets[split] = SpeechToTextDatasetCreator.from_tsv( + self.args.data, + self.data_cfg, + split, + self.tgt_dict, + pre_tokenizer, + bpe_tokenizer, + is_train_split=is_train_split, + epoch=epoch, + seed=self.args.seed, + ) + + @property + def target_dictionary(self): + return self.tgt_dict + + @property + def source_dictionary(self): + return None + + def max_positions(self): + return self.args.max_source_positions, self.args.max_target_positions + + def build_model(self, args): + args.input_feat_per_channel = self.data_cfg.input_feat_per_channel + args.input_channels = self.data_cfg.input_channels + return super(SpeechToTextTask, self).build_model(args) + + def build_generator( + self, + models, + args, + seq_gen_cls=None, + extra_gen_cls_kwargs=None, + ): + if self.data_cfg.prepend_tgt_lang_tag and args.prefix_size != 1: + raise ValueError( + 'Please set "--prefix-size 1" since ' + "target language ID token is prepended as BOS." + ) + lang_token_ids = { + i + for s, i in self.tgt_dict.indices.items() + if SpeechToTextDataset.is_lang_tag(s) + } + extra_gen_cls_kwargs = {"symbols_to_strip_from_output": lang_token_ids} + return super().build_generator( + models, args, seq_gen_cls=None, extra_gen_cls_kwargs=extra_gen_cls_kwargs + ) + + def build_tokenizer(self, args): + logger.info(f"pre-tokenizer: {self.data_cfg.pre_tokenizer}") + return encoders.build_tokenizer(Namespace(**self.data_cfg.pre_tokenizer)) + + def build_bpe(self, args): + logger.info(f"tokenizer: {self.data_cfg.bpe_tokenizer}") + return encoders.build_bpe(Namespace(**self.data_cfg.bpe_tokenizer)) + + def get_interactive_tokens_and_lengths(self, lines, encode_fn): + n_frames = [get_features_or_waveform(p).shape[0] for p in lines] + return lines, n_frames + + def build_dataset_for_inference(self, src_tokens, src_lengths, **kwargs): + return SpeechToTextDataset( + "interactive", False, self.data_cfg, src_tokens, src_lengths + ) diff --git a/fairseq/tasks/translation.py b/fairseq/tasks/translation.py new file mode 100644 index 0000000000000000000000000000000000000000..ea80fa2e73a0ee0e6d22d1b880e9a57877c48742 --- /dev/null +++ b/fairseq/tasks/translation.py @@ -0,0 +1,487 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from dataclasses import dataclass, field +import itertools +import json +import logging +import os +from typing import Optional +from argparse import Namespace +from omegaconf import II + +import numpy as np +from fairseq import metrics, utils +from fairseq.data import ( + AppendTokenDataset, + ConcatDataset, + LanguagePairDataset, + PrependTokenDataset, + StripTokenDataset, + TruncateDataset, + data_utils, + encoders, + indexed_dataset, +) +from fairseq.data.indexed_dataset import get_available_dataset_impl +from fairseq.dataclass import ChoiceEnum, FairseqDataclass +from fairseq.tasks import FairseqTask, register_task + + +EVAL_BLEU_ORDER = 4 + + +logger = logging.getLogger(__name__) + + +def load_langpair_dataset( + data_path, + split, + src, + src_dict, + tgt, + tgt_dict, + combine, + dataset_impl, + upsample_primary, + left_pad_source, + left_pad_target, + max_source_positions, + max_target_positions, + prepend_bos=False, + load_alignments=False, + truncate_source=False, + append_source_id=False, + num_buckets=0, + shuffle=True, + pad_to_multiple=1, + prepend_bos_src=None, +): + def split_exists(split, src, tgt, lang, data_path): + filename = os.path.join(data_path, "{}.{}-{}.{}".format(split, src, tgt, lang)) + return indexed_dataset.dataset_exists(filename, impl=dataset_impl) + + src_datasets = [] + tgt_datasets = [] + + for k in itertools.count(): + split_k = split + (str(k) if k > 0 else "") + + # infer langcode + if split_exists(split_k, src, tgt, src, data_path): + prefix = os.path.join(data_path, "{}.{}-{}.".format(split_k, src, tgt)) + elif split_exists(split_k, tgt, src, src, data_path): + prefix = os.path.join(data_path, "{}.{}-{}.".format(split_k, tgt, src)) + else: + if k > 0: + break + else: + raise FileNotFoundError( + "Dataset not found: {} ({})".format(split, data_path) + ) + + src_dataset = data_utils.load_indexed_dataset( + prefix + src, src_dict, dataset_impl + ) + if truncate_source: + src_dataset = AppendTokenDataset( + TruncateDataset( + StripTokenDataset(src_dataset, src_dict.eos()), + max_source_positions - 1, + ), + src_dict.eos(), + ) + src_datasets.append(src_dataset) + + tgt_dataset = data_utils.load_indexed_dataset( + prefix + tgt, tgt_dict, dataset_impl + ) + if tgt_dataset is not None: + tgt_datasets.append(tgt_dataset) + + logger.info( + "{} {} {}-{} {} examples".format( + data_path, split_k, src, tgt, len(src_datasets[-1]) + ) + ) + + if not combine: + break + + assert len(src_datasets) == len(tgt_datasets) or len(tgt_datasets) == 0 + + if len(src_datasets) == 1: + src_dataset = src_datasets[0] + tgt_dataset = tgt_datasets[0] if len(tgt_datasets) > 0 else None + else: + sample_ratios = [1] * len(src_datasets) + sample_ratios[0] = upsample_primary + src_dataset = ConcatDataset(src_datasets, sample_ratios) + if len(tgt_datasets) > 0: + tgt_dataset = ConcatDataset(tgt_datasets, sample_ratios) + else: + tgt_dataset = None + + if prepend_bos: + assert hasattr(src_dict, "bos_index") and hasattr(tgt_dict, "bos_index") + src_dataset = PrependTokenDataset(src_dataset, src_dict.bos()) + if tgt_dataset is not None: + tgt_dataset = PrependTokenDataset(tgt_dataset, tgt_dict.bos()) + elif prepend_bos_src is not None: + logger.info(f"prepending src bos: {prepend_bos_src}") + src_dataset = PrependTokenDataset(src_dataset, prepend_bos_src) + + eos = None + if append_source_id: + src_dataset = AppendTokenDataset( + src_dataset, src_dict.index("[{}]".format(src)) + ) + if tgt_dataset is not None: + tgt_dataset = AppendTokenDataset( + tgt_dataset, tgt_dict.index("[{}]".format(tgt)) + ) + eos = tgt_dict.index("[{}]".format(tgt)) + + align_dataset = None + if load_alignments: + align_path = os.path.join(data_path, "{}.align.{}-{}".format(split, src, tgt)) + if indexed_dataset.dataset_exists(align_path, impl=dataset_impl): + align_dataset = data_utils.load_indexed_dataset( + align_path, None, dataset_impl + ) + + tgt_dataset_sizes = tgt_dataset.sizes if tgt_dataset is not None else None + return LanguagePairDataset( + src_dataset, + src_dataset.sizes, + src_dict, + tgt_dataset, + tgt_dataset_sizes, + tgt_dict, + left_pad_source=left_pad_source, + left_pad_target=left_pad_target, + align_dataset=align_dataset, + eos=eos, + num_buckets=num_buckets, + shuffle=shuffle, + pad_to_multiple=pad_to_multiple, + ) + + +@dataclass +class TranslationConfig(FairseqDataclass): + data: Optional[str] = field( + default=None, + metadata={ + "help": "colon separated path to data directories list, will be iterated upon during epochs " + "in round-robin manner; however, valid and test data are always in the first directory " + "to avoid the need for repeating them in all directories" + }, + ) + source_lang: Optional[str] = field( + default=None, + metadata={ + "help": "source language", + "argparse_alias": "-s", + }, + ) + target_lang: Optional[str] = field( + default=None, + metadata={ + "help": "target language", + "argparse_alias": "-t", + }, + ) + load_alignments: bool = field( + default=False, metadata={"help": "load the binarized alignments"} + ) + left_pad_source: bool = field( + default=True, metadata={"help": "pad the source on the left"} + ) + left_pad_target: bool = field( + default=False, metadata={"help": "pad the target on the left"} + ) + max_source_positions: int = field( + default=1024, metadata={"help": "max number of tokens in the source sequence"} + ) + max_target_positions: int = field( + default=1024, metadata={"help": "max number of tokens in the target sequence"} + ) + upsample_primary: int = field( + default=-1, metadata={"help": "the amount of upsample primary dataset"} + ) + truncate_source: bool = field( + default=False, metadata={"help": "truncate source to max-source-positions"} + ) + num_batch_buckets: int = field( + default=0, + metadata={ + "help": "if >0, then bucket source and target lengths into " + "N buckets and pad accordingly; this is useful on TPUs to minimize the number of compilations" + }, + ) + train_subset: str = II("dataset.train_subset") + dataset_impl: Optional[ChoiceEnum(get_available_dataset_impl())] = II( + "dataset.dataset_impl" + ) + required_seq_len_multiple: int = II("dataset.required_seq_len_multiple") + + # options for reporting BLEU during validation + eval_bleu: bool = field( + default=False, metadata={"help": "evaluation with BLEU scores"} + ) + eval_bleu_args: Optional[str] = field( + default="{}", + metadata={ + "help": 'generation args for BLUE scoring, e.g., \'{"beam": 4, "lenpen": 0.6}\', as JSON string' + }, + ) + eval_bleu_detok: str = field( + default="space", + metadata={ + "help": "detokenize before computing BLEU (e.g., 'moses'); required if using --eval-bleu; " + "use 'space' to disable detokenization; see fairseq.data.encoders for other options" + }, + ) + eval_bleu_detok_args: Optional[str] = field( + default="{}", + metadata={"help": "args for building the tokenizer, if needed, as JSON string"}, + ) + eval_tokenized_bleu: bool = field( + default=False, metadata={"help": "compute tokenized BLEU instead of sacrebleu"} + ) + eval_bleu_remove_bpe: Optional[str] = field( + default=None, + metadata={ + "help": "remove BPE before computing BLEU", + "argparse_const": "@@ ", + }, + ) + eval_bleu_print_samples: bool = field( + default=False, metadata={"help": "print sample generations during validation"} + ) + + +@register_task("translation", dataclass=TranslationConfig) +class TranslationTask(FairseqTask): + """ + Translate from one (source) language to another (target) language. + + Args: + src_dict (~fairseq.data.Dictionary): dictionary for the source language + tgt_dict (~fairseq.data.Dictionary): dictionary for the target language + + .. note:: + + The translation task is compatible with :mod:`fairseq-train`, + :mod:`fairseq-generate` and :mod:`fairseq-interactive`. + """ + + cfg: TranslationConfig + + def __init__(self, cfg: TranslationConfig, src_dict, tgt_dict): + super().__init__(cfg) + self.src_dict = src_dict + self.tgt_dict = tgt_dict + + @classmethod + def setup_task(cls, cfg: TranslationConfig, **kwargs): + """Setup the task (e.g., load dictionaries). + + Args: + args (argparse.Namespace): parsed command-line arguments + """ + + paths = utils.split_paths(cfg.data) + assert len(paths) > 0 + # find language pair automatically + if cfg.source_lang is None or cfg.target_lang is None: + cfg.source_lang, cfg.target_lang = data_utils.infer_language_pair(paths[0]) + if cfg.source_lang is None or cfg.target_lang is None: + raise Exception( + "Could not infer language pair, please provide it explicitly" + ) + + # load dictionaries + src_dict = cls.load_dictionary( + os.path.join(paths[0], "dict.{}.txt".format(cfg.source_lang)) + ) + tgt_dict = cls.load_dictionary( + os.path.join(paths[0], "dict.{}.txt".format(cfg.target_lang)) + ) + assert src_dict.pad() == tgt_dict.pad() + assert src_dict.eos() == tgt_dict.eos() + assert src_dict.unk() == tgt_dict.unk() + logger.info("[{}] dictionary: {} types".format(cfg.source_lang, len(src_dict))) + logger.info("[{}] dictionary: {} types".format(cfg.target_lang, len(tgt_dict))) + + return cls(cfg, src_dict, tgt_dict) + + def load_dataset(self, split, epoch=1, combine=False, **kwargs): + """Load a given dataset split. + + Args: + split (str): name of the split (e.g., train, valid, test) + """ + paths = utils.split_paths(self.cfg.data) + assert len(paths) > 0 + if split != self.cfg.train_subset: + # if not training data set, use the first shard for valid and test + paths = paths[:1] + data_path = paths[(epoch - 1) % len(paths)] + + # infer langcode + src, tgt = self.cfg.source_lang, self.cfg.target_lang + + self.datasets[split] = load_langpair_dataset( + data_path, + split, + src, + self.src_dict, + tgt, + self.tgt_dict, + combine=combine, + dataset_impl=self.cfg.dataset_impl, + upsample_primary=self.cfg.upsample_primary, + left_pad_source=self.cfg.left_pad_source, + left_pad_target=self.cfg.left_pad_target, + max_source_positions=self.cfg.max_source_positions, + max_target_positions=self.cfg.max_target_positions, + load_alignments=self.cfg.load_alignments, + truncate_source=self.cfg.truncate_source, + num_buckets=self.cfg.num_batch_buckets, + shuffle=(split != "test"), + pad_to_multiple=self.cfg.required_seq_len_multiple, + ) + + def build_dataset_for_inference(self, src_tokens, src_lengths, constraints=None): + return LanguagePairDataset( + src_tokens, + src_lengths, + self.source_dictionary, + tgt_dict=self.target_dictionary, + constraints=constraints, + ) + + def build_model(self, cfg): + model = super().build_model(cfg) + if self.cfg.eval_bleu: + detok_args = json.loads(self.cfg.eval_bleu_detok_args) + self.tokenizer = encoders.build_tokenizer( + Namespace(tokenizer=self.cfg.eval_bleu_detok, **detok_args) + ) + + gen_args = json.loads(self.cfg.eval_bleu_args) + self.sequence_generator = self.build_generator( + [model], Namespace(**gen_args) + ) + return model + + def valid_step(self, sample, model, criterion): + loss, sample_size, logging_output = super().valid_step(sample, model, criterion) + if self.cfg.eval_bleu: + bleu = self._inference_with_bleu(self.sequence_generator, sample, model) + logging_output["_bleu_sys_len"] = bleu.sys_len + logging_output["_bleu_ref_len"] = bleu.ref_len + # we split counts into separate entries so that they can be + # summed efficiently across workers using fast-stat-sync + assert len(bleu.counts) == EVAL_BLEU_ORDER + for i in range(EVAL_BLEU_ORDER): + logging_output["_bleu_counts_" + str(i)] = bleu.counts[i] + logging_output["_bleu_totals_" + str(i)] = bleu.totals[i] + return loss, sample_size, logging_output + + def reduce_metrics(self, logging_outputs, criterion): + super().reduce_metrics(logging_outputs, criterion) + if self.cfg.eval_bleu: + + def sum_logs(key): + import torch + result = sum(log.get(key, 0) for log in logging_outputs) + if torch.is_tensor(result): + result = result.cpu() + return result + + counts, totals = [], [] + for i in range(EVAL_BLEU_ORDER): + counts.append(sum_logs("_bleu_counts_" + str(i))) + totals.append(sum_logs("_bleu_totals_" + str(i))) + + if max(totals) > 0: + # log counts as numpy arrays -- log_scalar will sum them correctly + metrics.log_scalar("_bleu_counts", np.array(counts)) + metrics.log_scalar("_bleu_totals", np.array(totals)) + metrics.log_scalar("_bleu_sys_len", sum_logs("_bleu_sys_len")) + metrics.log_scalar("_bleu_ref_len", sum_logs("_bleu_ref_len")) + + def compute_bleu(meters): + import inspect + import sacrebleu + + fn_sig = inspect.getfullargspec(sacrebleu.compute_bleu)[0] + if "smooth_method" in fn_sig: + smooth = {"smooth_method": "exp"} + else: + smooth = {"smooth": "exp"} + bleu = sacrebleu.compute_bleu( + correct=meters["_bleu_counts"].sum, + total=meters["_bleu_totals"].sum, + sys_len=meters["_bleu_sys_len"].sum, + ref_len=meters["_bleu_ref_len"].sum, + **smooth + ) + return round(bleu.score, 2) + + metrics.log_derived("bleu", compute_bleu) + + def max_positions(self): + """Return the max sentence length allowed by the task.""" + return (self.cfg.max_source_positions, self.cfg.max_target_positions) + + @property + def source_dictionary(self): + """Return the source :class:`~fairseq.data.Dictionary`.""" + return self.src_dict + + @property + def target_dictionary(self): + """Return the target :class:`~fairseq.data.Dictionary`.""" + return self.tgt_dict + + def _inference_with_bleu(self, generator, sample, model): + import sacrebleu + + def decode(toks, escape_unk=False): + s = self.tgt_dict.string( + toks.int().cpu(), + self.cfg.eval_bleu_remove_bpe, + # The default unknown string in fairseq is `<unk>`, but + # this is tokenized by sacrebleu as `< unk >`, inflating + # BLEU scores. Instead, we use a somewhat more verbose + # alternative that is unlikely to appear in the real + # reference, but doesn't get split into multiple tokens. + unk_string=("UNKNOWNTOKENINREF" if escape_unk else "UNKNOWNTOKENINHYP"), + ) + if self.tokenizer: + s = self.tokenizer.decode(s) + return s + + gen_out = self.inference_step(generator, [model], sample, prefix_tokens=None) + hyps, refs = [], [] + for i in range(len(gen_out)): + hyps.append(decode(gen_out[i][0]["tokens"])) + refs.append( + decode( + utils.strip_pad(sample["target"][i], self.tgt_dict.pad()), + escape_unk=True, # don't count <unk> as matches to the hypo + ) + ) + if self.cfg.eval_bleu_print_samples: + logger.info("example hypothesis: " + hyps[0]) + logger.info("example reference: " + refs[0]) + if self.cfg.eval_tokenized_bleu: + return sacrebleu.corpus_bleu(hyps, [refs], tokenize="none") + else: + return sacrebleu.corpus_bleu(hyps, [refs]) diff --git a/fairseq/tasks/translation_from_pretrained_bart.py b/fairseq/tasks/translation_from_pretrained_bart.py new file mode 100644 index 0000000000000000000000000000000000000000..0fd7a5b29f0e34699b5d5ef7574bc39b8c6052c9 --- /dev/null +++ b/fairseq/tasks/translation_from_pretrained_bart.py @@ -0,0 +1,132 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +from fairseq import utils +from fairseq.data import LanguagePairDataset + +from . import register_task +from .translation import TranslationTask, load_langpair_dataset + + +@register_task("translation_from_pretrained_bart") +class TranslationFromPretrainedBARTTask(TranslationTask): + """ + Translate from source language to target language with a model initialized with a multilingual pretrain. + + Args: + src_dict (~fairseq.data.Dictionary): dictionary for the source language + tgt_dict (~fairseq.data.Dictionary): dictionary for the target language + + .. note:: + + The translation task is compatible with :mod:`fairseq-train`, + :mod:`fairseq-generate` and :mod:`fairseq-interactive`. + + The translation task provides the following additional command-line + arguments: + + .. argparse:: + :ref: fairseq.tasks.translation_parser + :prog: + """ + + @staticmethod + def add_args(parser): + """Add task-specific arguments to the parser.""" + # fmt: off + TranslationTask.add_args(parser) + parser.add_argument('--langs', type=str, metavar='LANG', + help='comma-separated list of monolingual language, ' + 'for example, "en,de,fr". These should match the ' + 'langs from pretraining (and be in the same order). ' + 'You should always add all pretraining language idx ' + 'during finetuning.') + parser.add_argument('--prepend-bos', action='store_true', + help='prepend bos token to each sentence, which matches ' + 'mBART pretraining') + # fmt: on + + def __init__(self, args, src_dict, tgt_dict): + super().__init__(args, src_dict, tgt_dict) + self.langs = args.langs.split(",") + for d in [src_dict, tgt_dict]: + for l in self.langs: + d.add_symbol("[{}]".format(l)) + d.add_symbol("<mask>") + + def load_dataset(self, split, epoch=1, combine=False, **kwargs): + """Load a given dataset split. + + Args: + split (str): name of the split (e.g., train, valid, test) + """ + paths = utils.split_paths(self.args.data) + assert len(paths) > 0 + data_path = paths[(epoch - 1) % len(paths)] + + # infer langcode + src, tgt = self.args.source_lang, self.args.target_lang + + self.datasets[split] = load_langpair_dataset( + data_path, + split, + src, + self.src_dict, + tgt, + self.tgt_dict, + combine=combine, + dataset_impl=self.args.dataset_impl, + upsample_primary=self.args.upsample_primary, + left_pad_source=self.args.left_pad_source, + left_pad_target=self.args.left_pad_target, + max_source_positions=getattr(self.args, "max_source_positions", 1024), + max_target_positions=getattr(self.args, "max_target_positions", 1024), + load_alignments=self.args.load_alignments, + prepend_bos=getattr(self.args, "prepend_bos", False), + append_source_id=True, + ) + + def build_generator(self, models, args, **unused): + if getattr(args, "score_reference", False): + from fairseq.sequence_scorer import SequenceScorer + + return SequenceScorer( + self.target_dictionary, + eos=self.tgt_dict.index("[{}]".format(self.args.target_lang)), + ) + else: + from fairseq.sequence_generator import SequenceGenerator + + return SequenceGenerator( + models, + self.target_dictionary, + beam_size=getattr(args, "beam", 5), + max_len_a=getattr(args, "max_len_a", 0), + max_len_b=getattr(args, "max_len_b", 200), + min_len=getattr(args, "min_len", 1), + normalize_scores=(not getattr(args, "unnormalized", False)), + len_penalty=getattr(args, "lenpen", 1), + unk_penalty=getattr(args, "unkpen", 0), + temperature=getattr(args, "temperature", 1.0), + match_source_len=getattr(args, "match_source_len", False), + no_repeat_ngram_size=getattr(args, "no_repeat_ngram_size", 0), + eos=self.tgt_dict.index("[{}]".format(self.args.target_lang)), + ) + + def build_dataset_for_inference(self, src_tokens, src_lengths, constraints=None): + src_lang_id = self.source_dictionary.index("[{}]".format(self.args.source_lang)) + source_tokens = [] + for s_t in src_tokens: + s_t = torch.cat([s_t, s_t.new(1).fill_(src_lang_id)]) + source_tokens.append(s_t) + dataset = LanguagePairDataset( + source_tokens, + src_lengths, + self.source_dictionary, + tgt_dict=self.target_dictionary, + constraints=constraints, + ) + return dataset diff --git a/fairseq/tasks/translation_from_pretrained_xlm.py b/fairseq/tasks/translation_from_pretrained_xlm.py new file mode 100644 index 0000000000000000000000000000000000000000..a05f2891524a8b23482e206c1742c3b816b77afb --- /dev/null +++ b/fairseq/tasks/translation_from_pretrained_xlm.py @@ -0,0 +1,39 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from dataclasses import dataclass +from fairseq.data.legacy.masked_lm_dictionary import MaskedLMDictionary +from fairseq.tasks.translation import TranslationConfig, TranslationTask + +from . import register_task + + +@dataclass +class TranslationFromPretrainedXLMConfig(TranslationConfig): + pass + + +@register_task( + "translation_from_pretrained_xlm", dataclass=TranslationFromPretrainedXLMConfig +) +class TranslationFromPretrainedXLMTask(TranslationTask): + """ + Same as TranslationTask except use the MaskedLMDictionary class so that + we can load data that was binarized with the MaskedLMDictionary class. + + This task should be used for the entire training pipeline when we want to + train an NMT model from a pretrained XLM checkpoint: binarizing NMT data, + training NMT with the pretrained XLM checkpoint, and subsequent evaluation + of that trained model. + """ + + @classmethod + def load_dictionary(cls, filename): + """Load the masked LM dictionary from the filename + + Args: + filename (str): the filename + """ + return MaskedLMDictionary.load(filename) diff --git a/fairseq/tasks/translation_lev.py b/fairseq/tasks/translation_lev.py new file mode 100644 index 0000000000000000000000000000000000000000..041279305dc4978f6a3a4178c5ec4c72c5fb2b5c --- /dev/null +++ b/fairseq/tasks/translation_lev.py @@ -0,0 +1,191 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from dataclasses import dataclass, field +import torch +from fairseq import utils +from fairseq.data import LanguagePairDataset +from fairseq.dataclass import ChoiceEnum +from fairseq.tasks import register_task +from fairseq.tasks.translation import TranslationConfig, TranslationTask, load_langpair_dataset +from fairseq.utils import new_arange + + +NOISE_CHOICES = ChoiceEnum(["random_delete", "random_mask", "no_noise", "full_mask"]) + +@dataclass +class TranslationLevenshteinConfig(TranslationConfig): + noise: NOISE_CHOICES = field( + default="random_delete", + metadata={ + "help": "type of noise" + }, + ) + +@register_task("translation_lev", dataclass=TranslationLevenshteinConfig) +class TranslationLevenshteinTask(TranslationTask): + """ + Translation (Sequence Generation) task for Levenshtein Transformer + See `"Levenshtein Transformer" <https://arxiv.org/abs/1905.11006>`_. + """ + + cfg: TranslationLevenshteinConfig + + def load_dataset(self, split, epoch=1, combine=False, **kwargs): + """Load a given dataset split. + + Args: + split (str): name of the split (e.g., train, valid, test) + """ + paths = utils.split_paths(self.cfg.data) + assert len(paths) > 0 + data_path = paths[(epoch - 1) % len(paths)] + + # infer langcode + src, tgt = self.cfg.source_lang, self.cfg.target_lang + + self.datasets[split] = load_langpair_dataset( + data_path, + split, + src, + self.src_dict, + tgt, + self.tgt_dict, + combine=combine, + dataset_impl=self.cfg.dataset_impl, + upsample_primary=self.cfg.upsample_primary, + left_pad_source=self.cfg.left_pad_source, + left_pad_target=self.cfg.left_pad_target, + max_source_positions=self.cfg.max_source_positions, + max_target_positions=self.cfg.max_target_positions, + prepend_bos=True, + ) + + def inject_noise(self, target_tokens): + def _random_delete(target_tokens): + pad = self.tgt_dict.pad() + bos = self.tgt_dict.bos() + eos = self.tgt_dict.eos() + + max_len = target_tokens.size(1) + target_mask = target_tokens.eq(pad) + target_score = target_tokens.clone().float().uniform_() + target_score.masked_fill_( + target_tokens.eq(bos) | target_tokens.eq(eos), 0.0 + ) + target_score.masked_fill_(target_mask, 1) + target_score, target_rank = target_score.sort(1) + target_length = target_mask.size(1) - target_mask.float().sum( + 1, keepdim=True + ) + + # do not delete <bos> and <eos> (we assign 0 score for them) + target_cutoff = ( + 2 + + ( + (target_length - 2) + * target_score.new_zeros(target_score.size(0), 1).uniform_() + ).long() + ) + target_cutoff = target_score.sort(1)[1] >= target_cutoff + + prev_target_tokens = ( + target_tokens.gather(1, target_rank) + .masked_fill_(target_cutoff, pad) + .gather(1, target_rank.masked_fill_(target_cutoff, max_len).sort(1)[1]) + ) + prev_target_tokens = prev_target_tokens[ + :, : prev_target_tokens.ne(pad).sum(1).max() + ] + + return prev_target_tokens + + def _random_mask(target_tokens): + pad = self.tgt_dict.pad() + bos = self.tgt_dict.bos() + eos = self.tgt_dict.eos() + unk = self.tgt_dict.unk() + + target_masks = ( + target_tokens.ne(pad) & target_tokens.ne(bos) & target_tokens.ne(eos) + ) + target_score = target_tokens.clone().float().uniform_() + target_score.masked_fill_(~target_masks, 2.0) + target_length = target_masks.sum(1).float() + target_length = target_length * target_length.clone().uniform_() + target_length = target_length + 1 # make sure to mask at least one token. + + _, target_rank = target_score.sort(1) + target_cutoff = new_arange(target_rank) < target_length[:, None].long() + prev_target_tokens = target_tokens.masked_fill( + target_cutoff.scatter(1, target_rank, target_cutoff), unk + ) + return prev_target_tokens + + def _full_mask(target_tokens): + pad = self.tgt_dict.pad() + bos = self.tgt_dict.bos() + eos = self.tgt_dict.eos() + unk = self.tgt_dict.unk() + + target_mask = ( + target_tokens.eq(bos) | target_tokens.eq(eos) | target_tokens.eq(pad) + ) + return target_tokens.masked_fill(~target_mask, unk) + + if self.cfg.noise == "random_delete": + return _random_delete(target_tokens) + elif self.cfg.noise == "random_mask": + return _random_mask(target_tokens) + elif self.cfg.noise == "full_mask": + return _full_mask(target_tokens) + elif self.cfg.noise == "no_noise": + return target_tokens + else: + raise NotImplementedError + + def build_generator(self, models, args, **unused): + # add models input to match the API for SequenceGenerator + from fairseq.iterative_refinement_generator import IterativeRefinementGenerator + + return IterativeRefinementGenerator( + self.target_dictionary, + eos_penalty=getattr(args, "iter_decode_eos_penalty", 0.0), + max_iter=getattr(args, "iter_decode_max_iter", 10), + beam_size=getattr(args, "iter_decode_with_beam", 1), + reranking=getattr(args, "iter_decode_with_external_reranker", False), + decoding_format=getattr(args, "decoding_format", None), + adaptive=not getattr(args, "iter_decode_force_max_iter", False), + retain_history=getattr(args, "retain_iter_history", False), + ) + + def build_dataset_for_inference(self, src_tokens, src_lengths, constraints=None): + if constraints is not None: + # Though see Susanto et al. (ACL 2020): https://www.aclweb.org/anthology/2020.acl-main.325/ + raise NotImplementedError( + "Constrained decoding with the translation_lev task is not supported" + ) + + return LanguagePairDataset( + src_tokens, src_lengths, self.source_dictionary, append_bos=True + ) + + def train_step( + self, sample, model, criterion, optimizer, update_num, ignore_grad=False + ): + model.train() + sample["prev_target"] = self.inject_noise(sample["target"]) + loss, sample_size, logging_output = criterion(model, sample) + if ignore_grad: + loss *= 0 + optimizer.backward(loss) + return loss, sample_size, logging_output + + def valid_step(self, sample, model, criterion): + model.eval() + with torch.no_grad(): + sample["prev_target"] = self.inject_noise(sample["target"]) + loss, sample_size, logging_output = criterion(model, sample) + return loss, sample_size, logging_output diff --git a/fairseq/tasks/translation_multi_simple_epoch.py b/fairseq/tasks/translation_multi_simple_epoch.py new file mode 100644 index 0000000000000000000000000000000000000000..6f36e5b93e98497de31969d203ae04dbb4bd9306 --- /dev/null +++ b/fairseq/tasks/translation_multi_simple_epoch.py @@ -0,0 +1,430 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import datetime +import logging +import time + +import torch +from fairseq.data import ( + FairseqDataset, + LanguagePairDataset, + ListDataset, + data_utils, + iterators, +) +from fairseq.data.multilingual.multilingual_data_manager import ( + MultilingualDatasetManager, +) +from fairseq.data.multilingual.sampling_method import SamplingMethod +from fairseq.tasks import LegacyFairseqTask, register_task +from fairseq.utils import FileContentsAction + + +### +def get_time_gap(s, e): + return ( + datetime.datetime.fromtimestamp(e) - datetime.datetime.fromtimestamp(s) + ).__str__() + + +### + + +logger = logging.getLogger(__name__) + + +@register_task("translation_multi_simple_epoch") +class TranslationMultiSimpleEpochTask(LegacyFairseqTask): + """ + Translate from one (source) language to another (target) language. + + Args: + langs (List[str]): a list of languages that are being supported + dicts (Dict[str, fairseq.data.Dictionary]): mapping from supported languages to their dictionaries + training (bool): whether the task should be configured for training or not + + .. note:: + + The translation task is compatible with :mod:`fairseq-train`, + :mod:`fairseq-generate` and :mod:`fairseq-interactive`. + + The translation task provides the following additional command-line + arguments: + + .. argparse:: + :ref: fairseq.tasks.translation_parser + :prog: + """ + + @staticmethod + def add_args(parser): + """Add task-specific arguments to the parser.""" + # fmt: off + parser.add_argument('-s', '--source-lang', default=None, metavar='SRC', + help='inference source language') + parser.add_argument('-t', '--target-lang', default=None, metavar='TARGET', + help='inference target language') + parser.add_argument('--lang-pairs', default=None, metavar='PAIRS', + help='comma-separated list of language pairs (in training order): en-de,en-fr,de-fr', + action=FileContentsAction) + parser.add_argument('--keep-inference-langtok', action='store_true', + help='keep language tokens in inference output (e.g. for analysis or debugging)') + + SamplingMethod.add_arguments(parser) + MultilingualDatasetManager.add_args(parser) + # fmt: on + + def __init__(self, args, langs, dicts, training): + super().__init__(args) + self.langs = langs + self.dicts = dicts + self.training = training + if training: + self.lang_pairs = args.lang_pairs + else: + self.lang_pairs = ["{}-{}".format(args.source_lang, args.target_lang)] + # eval_lang_pairs for multilingual translation is usually all of the + # lang_pairs. However for other multitask settings or when we want to + # optimize for certain languages we want to use a different subset. Thus + # the eval_lang_pairs class variable is provided for classes that extend + # this class. + self.eval_lang_pairs = self.lang_pairs + # model_lang_pairs will be used to build encoder-decoder model pairs in + # models.build_model(). This allows multitask type of sub-class can + # build models other than the input lang_pairs + self.model_lang_pairs = self.lang_pairs + self.source_langs = [d.split("-")[0] for d in self.lang_pairs] + self.target_langs = [d.split("-")[1] for d in self.lang_pairs] + self.check_dicts(self.dicts, self.source_langs, self.target_langs) + + self.sampling_method = SamplingMethod.build_sampler(args, self) + self.data_manager = MultilingualDatasetManager.setup_data_manager( + args, self.lang_pairs, langs, dicts, self.sampling_method + ) + + def check_dicts(self, dicts, source_langs, target_langs): + if self.args.source_dict is not None or self.args.target_dict is not None: + # no need to check whether the source side and target side are sharing dictionaries + return + src_dict = dicts[source_langs[0]] + tgt_dict = dicts[target_langs[0]] + for src_lang in source_langs: + assert ( + src_dict == dicts[src_lang] + ), "Diffrent dictionary are specified for different source languages; " + "TranslationMultiSimpleEpochTask only supports one shared dictionary across all source languages" + for tgt_lang in target_langs: + assert ( + tgt_dict == dicts[tgt_lang] + ), "Diffrent dictionary are specified for different target languages; " + "TranslationMultiSimpleEpochTask only supports one shared dictionary across all target languages" + + @classmethod + def setup_task(cls, args, **kwargs): + langs, dicts, training = MultilingualDatasetManager.prepare( + cls.load_dictionary, args, **kwargs + ) + return cls(args, langs, dicts, training) + + def has_sharded_data(self, split): + return self.data_manager.has_sharded_data(split) + + def load_dataset(self, split, epoch=1, combine=False, **kwargs): + """Load a given dataset split. + + Args: + split (str): name of the split (e.g., train, valid, test) + """ + if split in self.datasets: + dataset = self.datasets[split] + if self.has_sharded_data(split): + if self.args.virtual_epoch_size is not None: + if dataset.load_next_shard: + shard_epoch = dataset.shard_epoch + else: + # no need to load next shard so skip loading + # also this avoid always loading from beginning of the data + return + else: + shard_epoch = epoch + else: + # estimate the shard epoch from virtual data size and virtual epoch size + shard_epoch = self.data_manager.estimate_global_pass_epoch(epoch) + logger.info(f"loading data for {split} epoch={epoch}/{shard_epoch}") + logger.info(f"mem usage: {data_utils.get_mem_usage()}") + if split in self.datasets: + del self.datasets[split] + logger.info("old dataset deleted manually") + logger.info(f"mem usage: {data_utils.get_mem_usage()}") + self.datasets[split] = self.data_manager.load_dataset( + split, + self.training, + epoch=epoch, + combine=combine, + shard_epoch=shard_epoch, + **kwargs, + ) + + def build_dataset_for_inference(self, src_tokens, src_lengths, constraints=None): + if constraints is not None: + raise NotImplementedError( + "Constrained decoding with the multilingual_translation task is not supported" + ) + + src_data = ListDataset(src_tokens, src_lengths) + dataset = LanguagePairDataset(src_data, src_lengths, self.source_dictionary) + src_langtok_spec, tgt_langtok_spec = self.args.langtoks["main"] + if self.args.lang_tok_replacing_bos_eos: + dataset = self.data_manager.alter_dataset_langtok( + dataset, + src_eos=self.source_dictionary.eos(), + src_lang=self.args.source_lang, + tgt_eos=self.target_dictionary.eos(), + tgt_lang=self.args.target_lang, + src_langtok_spec=src_langtok_spec, + tgt_langtok_spec=tgt_langtok_spec, + ) + else: + dataset.src = self.data_manager.src_dataset_tranform_func( + self.args.source_lang, + self.args.target_lang, + dataset=dataset.src, + spec=src_langtok_spec, + ) + return dataset + + def build_generator( + self, + models, + args, + seq_gen_cls=None, + extra_gen_cls_kwargs=None, + ): + if not getattr(args, "keep_inference_langtok", False): + _, tgt_langtok_spec = self.args.langtoks["main"] + if tgt_langtok_spec: + tgt_lang_tok = self.data_manager.get_decoder_langtok( + self.args.target_lang, tgt_langtok_spec + ) + extra_gen_cls_kwargs = extra_gen_cls_kwargs or {} + extra_gen_cls_kwargs["symbols_to_strip_from_output"] = {tgt_lang_tok} + + return super().build_generator( + models, args, seq_gen_cls=None, extra_gen_cls_kwargs=extra_gen_cls_kwargs + ) + + def build_model(self, args): + return super().build_model(args) + + def valid_step(self, sample, model, criterion): + loss, sample_size, logging_output = super().valid_step(sample, model, criterion) + return loss, sample_size, logging_output + + def inference_step( + self, generator, models, sample, prefix_tokens=None, constraints=None + ): + with torch.no_grad(): + _, tgt_langtok_spec = self.args.langtoks["main"] + if not self.args.lang_tok_replacing_bos_eos: + if prefix_tokens is None and tgt_langtok_spec: + tgt_lang_tok = self.data_manager.get_decoder_langtok( + self.args.target_lang, tgt_langtok_spec + ) + src_tokens = sample["net_input"]["src_tokens"] + bsz = src_tokens.size(0) + prefix_tokens = ( + torch.LongTensor([[tgt_lang_tok]]).expand(bsz, 1).to(src_tokens) + ) + return generator.generate( + models, + sample, + prefix_tokens=prefix_tokens, + constraints=constraints, + ) + else: + return generator.generate( + models, + sample, + prefix_tokens=prefix_tokens, + bos_token=self.data_manager.get_decoder_langtok( + self.args.target_lang, tgt_langtok_spec + ) + if tgt_langtok_spec + else self.target_dictionary.eos(), + ) + + def reduce_metrics(self, logging_outputs, criterion): + super().reduce_metrics(logging_outputs, criterion) + + def max_positions(self): + """Return the max sentence length allowed by the task.""" + return (self.args.max_source_positions, self.args.max_target_positions) + + @property + def source_dictionary(self): + return self.data_manager.get_source_dictionary(self.source_langs[0]) + + @property + def target_dictionary(self): + return self.data_manager.get_target_dictionary(self.target_langs[0]) + + def create_batch_sampler_func( + self, + max_positions, + ignore_invalid_inputs, + max_tokens, + max_sentences, + required_batch_size_multiple=1, + seed=1, + ): + def construct_batch_sampler(dataset, epoch): + splits = [ + s for s, _ in self.datasets.items() if self.datasets[s] == dataset + ] + split = splits[0] if len(splits) > 0 else None + # NEW implementation + if epoch is not None: + # initialize the dataset with the correct starting epoch + dataset.set_epoch(epoch) + + # get indices ordered by example size + start_time = time.time() + logger.info(f"start batch sampler: mem usage: {data_utils.get_mem_usage()}") + + with data_utils.numpy_seed(seed): + indices = dataset.ordered_indices() + logger.info( + f"[{split}] @batch_sampler order indices time: {get_time_gap(start_time, time.time())}" + ) + logger.info(f"mem usage: {data_utils.get_mem_usage()}") + + # filter examples that are too large + if max_positions is not None: + my_time = time.time() + indices = self.filter_indices_by_size( + indices, dataset, max_positions, ignore_invalid_inputs + ) + logger.info( + f"[{split}] @batch_sampler filter_by_size time: {get_time_gap(my_time, time.time())}" + ) + logger.info(f"mem usage: {data_utils.get_mem_usage()}") + + # create mini-batches with given size constraints + my_time = time.time() + batch_sampler = dataset.batch_by_size( + indices, + max_tokens=max_tokens, + max_sentences=max_sentences, + required_batch_size_multiple=required_batch_size_multiple, + ) + + logger.info( + f"[{split}] @batch_sampler batch_by_size time: {get_time_gap(my_time, time.time())}" + ) + logger.info( + f"[{split}] per epoch batch_sampler set-up time: {get_time_gap(start_time, time.time())}" + ) + logger.info(f"mem usage: {data_utils.get_mem_usage()}") + + return batch_sampler + + return construct_batch_sampler + + # we need to override get_batch_iterator because we want to reset the epoch iterator each time + def get_batch_iterator( + self, + dataset, + max_tokens=None, + max_sentences=None, + max_positions=None, + ignore_invalid_inputs=False, + required_batch_size_multiple=1, + seed=1, + num_shards=1, + shard_id=0, + num_workers=0, + epoch=1, + data_buffer_size=0, + disable_iterator_cache=False, + ): + """ + Get an iterator that yields batches of data from the given dataset. + + Args: + dataset (~fairseq.data.FairseqDataset): dataset to batch + max_tokens (int, optional): max number of tokens in each batch + (default: None). + max_sentences (int, optional): max number of sentences in each + batch (default: None). + max_positions (optional): max sentence length supported by the + model (default: None). + ignore_invalid_inputs (bool, optional): don't raise Exception for + sentences that are too long (default: False). + required_batch_size_multiple (int, optional): require batch size to + be a multiple of N (default: 1). + seed (int, optional): seed for random number generator for + reproducibility (default: 1). + num_shards (int, optional): shard the data iterator into N + shards (default: 1). + shard_id (int, optional): which shard of the data iterator to + return (default: 0). + num_workers (int, optional): how many subprocesses to use for data + loading. 0 means the data will be loaded in the main process + (default: 0). + epoch (int, optional): the epoch to start the iterator from + (default: 0). + data_buffer_size (int, optional): number of batches to + preload (default: 0). + disable_iterator_cache (bool, optional): don't cache the + EpochBatchIterator (ignores `FairseqTask::can_reuse_epoch_itr`) + (default: False). + Returns: + ~fairseq.iterators.EpochBatchIterator: a batched iterator over the + given dataset split + """ + # initialize the dataset with the correct starting epoch + assert isinstance(dataset, FairseqDataset) + if dataset in self.dataset_to_epoch_iter: + return self.dataset_to_epoch_iter[dataset] + if self.args.sampling_method == "RoundRobin": + batch_iter = super().get_batch_iterator( + dataset, + max_tokens=max_tokens, + max_sentences=max_sentences, + max_positions=max_positions, + ignore_invalid_inputs=ignore_invalid_inputs, + required_batch_size_multiple=required_batch_size_multiple, + seed=seed, + num_shards=num_shards, + shard_id=shard_id, + num_workers=num_workers, + epoch=epoch, + data_buffer_size=data_buffer_size, + disable_iterator_cache=disable_iterator_cache, + ) + self.dataset_to_epoch_iter[dataset] = batch_iter + return batch_iter + + construct_batch_sampler = self.create_batch_sampler_func( + max_positions, + ignore_invalid_inputs, + max_tokens, + max_sentences, + required_batch_size_multiple=required_batch_size_multiple, + seed=seed, + ) + + epoch_iter = iterators.EpochBatchIterator( + dataset=dataset, + collate_fn=dataset.collater, + batch_sampler=construct_batch_sampler, + seed=seed, + num_shards=num_shards, + shard_id=shard_id, + num_workers=num_workers, + epoch=epoch, + ) + return epoch_iter diff --git a/fairseq/token_generation_constraints.py b/fairseq/token_generation_constraints.py new file mode 100644 index 0000000000000000000000000000000000000000..e708dc51bcb0ffb7b411496239c74d5e6f3c2448 --- /dev/null +++ b/fairseq/token_generation_constraints.py @@ -0,0 +1,506 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +"""Implements tracking of constraints for a beam item. + +A list of constraints is given as a list of one or more token +sequences, each of length at least one token. For example, for an input sentence + +> Die maschinelle Übersetzung ist schwer zu kontrollieren. + +We could have the constraints: +* to influence +* hard + +There are two implementations: +* OrderedConstraintState: Tracks progress through an ordered list of multitoken constraints. +* UnorderedConstraintState: Tracks progress through an unordered list of multitoken constraints. + +The difference is that in the first, the constraints are assumed to be +in order; the algorithm will permit zero or more tokens between them. +In the second, the constraints are not ordered, so many orderings will +be explored. + +The same sequence can be present any number of times, and will appear +that many times in the output. +""" + +from collections import Counter +from typing import List, Optional, Set, Tuple + +import torch + + +class ConstraintState: + def __init__(self): + pass + + +def pack_constraints(batch_constraints: List[List[torch.Tensor]]) -> torch.Tensor: + """Takes a list of list of constraints in tensor form (a list of + tensor constraints for each sentence) and transforms it into a + packed Tensor. For example, here is a batch of size 3 with 3, 0, + and 1 constraints: + + [ [ [3 1 2], [3], [4 5 6 7], ] + [], + [ [1 8 9 10 1 4 11 12], ] + ] + + Its corresponding packed structure is: + + [ [ 3 3 1 2 0 3 0 4 5 6 7 0], + [ 0 0 0 0 0 0 0 0 0 0 0 0], + [ 1 1 8 9 10 1 4 11 12 0 0 0] ] + + The packed tensor has shape (batch size, maxlen), where + maxlen is defined below. Each row contains concatenated + constraint tokens for that sentence, with 0 appended after + each constraint. The first item in each row is the number + of constraints for that sentence. So maxlen is the maximum + of + + (number of constraints) + (sum length of constraints) + 1. + + across all sentences in the batch. + """ + # The maximum word length of concatenated constraints for any sentence + max_constraints_len = 1 + for sentence_constraints in batch_constraints: + if len(sentence_constraints): + # number of constraints, plus sum of constrain lens, plus a zero after each + constraints_len = ( + 1 + + sum([c.size(0) for c in sentence_constraints]) + + len(sentence_constraints) + ) + max_constraints_len = max(max_constraints_len, constraints_len) + + batch_size = len(batch_constraints) + constraints_tensor = torch.zeros((batch_size, max_constraints_len)).long() + for i, sentence_constraints in enumerate(batch_constraints): + constraints_tensor[i, 0] = len(sentence_constraints) + offset = 1 + for j, constraint in enumerate(sentence_constraints): + this_len = constraint.size(0) + constraints_tensor[i, offset : offset + this_len] = constraint + offset += this_len + 1 + + return constraints_tensor.long() + + +def unpack_constraints(constraint_tensor: torch.Tensor) -> List[torch.Tensor]: + """ + Transforms *one row* of a packed constraint tensor (e.g., for one + sentence in the batch) into a list of constraint tensors. + """ + constraint_list = [] + num_constraints = constraint_tensor[0] + constraints = constraint_tensor.tolist() + offset = 1 + for i in range(num_constraints): + where = constraints.index(0, offset) + constraint_list.append(constraint_tensor[offset:where]) + offset = where + 1 + + return constraint_list + + +class ConstraintNode: + """ + Represents a node in a trie managing unordered constraints. + """ + + def __init__(self, token: int = None, parent=None): + # The token associate with this node (None for the root) + self.token = int(token) if token is not None else None + # The parent (None at the root) + self.parent = parent + # Whether this node is a completed constraint + self.terminal = 0 + # List of child nodes + self.children = {} + + # The cumulative number of constraints from this point in the + # trie forward + self.num_constraints = 0 + + @property + def id(self): + return self.token + + def __str__(self): + term = self.terminal != 0 + return f"[{self.token}].{term}#{self.num_constraints}" + + def __getitem__(self, key: int): + return self.children.get(key, None) + + def next_tokens(self) -> Set[int]: + """The set of child labels.""" + return set(self.children.keys()) + + @staticmethod + def create(constraints: List[List[int]]): + root = ConstraintNode() + for sequence in constraints: + root.add_sequence(sequence) + + return root + + @staticmethod + def print_graph(node: "ConstraintNode"): + if len(node.children) == 0: + return str(node) + else: + s = f"({node}" + for child in node.children.values(): + s += " " + ConstraintNode.print_graph(child) + s += ")" + return s + + def token_counts(self) -> Counter: + """Returns a counter of the number of times each token is used + in a constraint. + """ + token_counts = Counter() + kids = list(self.children.values()) + while len(kids) > 0: + kid = kids.pop() + token_counts[kid.id] += kid.num_constraints + kids += list(kid.children.values()) + + return token_counts + + def tokens(self) -> Set[int]: + """Returns the set of tokens in constraints.""" + return set(self.token_counts().keys()) + + def add_sequence(self, sequence: List[int]): + """Adds a constraint, represented as a list of integers, to + the trie.""" + assert len(sequence) > 0 + + token = int(sequence[0]) + if token not in self.children: + self.children[token] = ConstraintNode(token, parent=self) + + node = self.children[token] + if len(sequence) == 1: + node.terminal += 1 + node.num_constraints += 1 + parent = node.parent + while parent is not None: + parent.num_constraints += 1 + parent = parent.parent + else: + node.add_sequence(sequence[1:]) + + +class UnorderedConstraintState(ConstraintState): + """ + Records progress through the set of constraints for each item in the beam + using a trie. + """ + + def __init__(self, node: ConstraintNode, copy_from: "ConstraintState" = None): + self.node = node + + if copy_from is None: + # The root node + self.root = node + # The set of states in the graph that have been completed + self.completed = Counter() + # The... + self.generated = Counter() + # The list of tokens we need to generate + self.needed_tokens = self.root.tokens() + else: + self.completed = Counter(copy_from.completed) + self.generated = Counter(copy_from.generated) + self.root = copy_from.root + + # Mark the node as generated + if self.node != self.root: + self.generated[node] += 1 + + @staticmethod + def create(constraint_tensor: torch.Tensor): + constraint_list = unpack_constraints(constraint_tensor) + constraint_trie_root = ConstraintNode.create(constraint_list) + return UnorderedConstraintState(constraint_trie_root) + + def __str__(self): + gen_str = ",".join([str(node) for node in self.generated]) + return f"{self.name}/{self.bank}({gen_str})x{self.num_completed}" + + def __copy__(self): + copied_state = UnorderedConstraintState(self.node, copy_from=self) + return copied_state + + def copy(self): + return self.__copy__() + + @property + def name(self): + if self.node.id is None: + return "ROOT" + else: + return str(self.node.id) + + @property + def is_root(self): + return self.node == self.root + + @property + def bank(self): + return sum(self.generated.values()) + + @property + def num_completed(self): + """The number of constraints (not constraint tokens) that are completed. + In addition to the already-completed states, we need to account for the + current state, which might get marked as completed when another token + is generated. + """ + in_final = self.node.terminal and self.completed[self.node] < self.node.terminal + return sum(self.completed.values()) + in_final + + @property + def finished(self): + return self.root.num_constraints - self.num_completed == 0 + + @property + def token_counts(self): + return self.root.token_counts() + + @property + def tokens(self): + return self.root.tokens() + + @property + def num_constraint_tokens(self): + return sum(self.token_counts.values()) + + def next_tokens(self) -> Set[int]: + """Returns the list of tokens that could come next. + These are (a) all tokens extending the root state and, for + non-root states, additionally all tokens extending the current + state.""" + + if self.node != self.root: + return self.root.next_tokens().union(self.node.next_tokens()) + else: + return self.root.next_tokens() + + def advance(self, token: int): + """Reads in a token and advances the state. Here's how it works. + + We can advance to the next state if: + - there is a matching child + - its path isn't blocked + + A path is blocked when all constraints that are descendants of + that node have already been generated, in the current state. + + If we are not able to advance from the current state, we "fall + off the graph" and return to the root state. There, we again + try to advance, checking the same criteria. + + In any case, when falling off the graph, we need to do some + bookkeeping. We: + - check whether any constraints were met (all prefixes of + current state) + - if one is found, mark it as completed + - adjust visited nodes accordingly + """ + token = int(token) + + next_state = None + child = self.node[token] + if child is not None and self.generated[child] < child.num_constraints: + next_state = UnorderedConstraintState(child, copy_from=self) + + def rewind(): + """If we're mid-trie and an "illegal" token is chosen next, we need + to reset our state to the root state. However, along the way, we need + to check whether a prefix of the current trie state represents a state + we could mark as completed. + """ + node = self.node + while node != self.root: + if node.terminal and self.completed[node] < node.terminal: + next_state.completed[node] += 1 + return + + next_state.generated[node] -= 1 + node = node.parent + + # Fall off the graph, check the root + if next_state is None and token in self.root.next_tokens(): + child = self.root[token] + # We can only traverse this edge if it's not saturated + if self.generated[child] < child.num_constraints: + next_state = UnorderedConstraintState(child, copy_from=self) + else: + next_state = UnorderedConstraintState(self.root, copy_from=self) + + # Rewind + rewind() + + elif next_state is None: + next_state = UnorderedConstraintState(self.root, copy_from=self) + # Rewind + rewind() + + return next_state + + +class ConstraintSequence: + def __init__(self, sequences: List[List[int]]): + """Represents a set of possibly multitoken constraints by + concatenating them and internally recording the end points. + """ + self.sequences = [] + self.endpoints = [] + self.num_tokens = 0 + self.tokens = set() + for sequence in sequences: + for token in sequence: + self.tokens.add(token) + self.num_tokens += len(sequence) + self.endpoints += [False for x in range(len(sequence) - 1)] + [True] + self.sequences += sequence + + def __getitem__(self, key: int): + return self.sequences[key] + + def __len__(self): + return len(self.sequences) + + def __str__(self): + return str(self.sequences) + + +class OrderedConstraintState(ConstraintState): + """ + Records progress through the set of linear nonbranching constraints with gaps. + """ + + def __init__(self, sequence: ConstraintSequence, state: int = -1): + self.sequence = sequence + self.state = state + + @staticmethod + def create(constraint_tensor: torch.Tensor): + constraint_list = unpack_constraints(constraint_tensor) + return OrderedConstraintState(ConstraintSequence(constraint_list), -1) + + def __str__(self): + return f"{self.state}/{self.bank}x{self.num_completed}" + + def __copy__(self): + return OrderedConstraintState(self.sequence, self.state) + + def copy(self): + return self.__copy__() + + @property + def num_completed(self): + if self.state == -1: + return 0 + count = len( + list(filter(lambda x: x, self.sequence.endpoints[0 : self.state + 1])) + ) + return count + + @property + def is_root(self): + return self.state == -1 + + @property + def name(self): + if self.state == -1: + return "ROOT" + else: + return str(self.sequence[self.state]) + + @property + def bank(self) -> int: + return self.state + 1 + + @property + def finished(self): + return self.state + 1 == len(self.sequence) + + @property + def token_counts(self): + return self.sequence.token_counts() + + @property + def tokens(self): + return self.sequence.tokens + + @property + def num_constraint_tokens(self): + return sum(self.token_counts.values()) + + def next_tokens(self) -> Set[int]: + """Returns the list of tokens that could come next. + These are (a) all tokens extending the root state and, for + non-root states, additionally all tokens extending the current + state.""" + + tokens = set() + if self.state > 0: + tokens.add(self.sequence[0]) + if not self.finished: + tokens.add(self.sequence[self.state + 1]) + return tokens + + def advance(self, token: int): + """Reads in a token and advances the state. Here's how it works. + + We can advance to the next state if: + - there is a matching child + - its path isn't blocked + + A path is blocked when all constraints that are descendants of + that node have already been generated, in the current state. + + If we are not able to advance from the current state, we "fall + off the graph" and return to the root state. There, we again + try to advance, checking the same criteria. + + In any case, when falling off the graph, we need to do some + bookkeeping. We: + - check whether any constraints were met (all prefixes of + current state) + - if one is found, mark it as completed + - adjust visited nodes accordingly + """ + token = int(token) + # print(f"{self} ADVANCE({token}) {self.sequence} -> ", end="") + + if self.finished: + # Accept anything + next_state = self.copy() + + elif self.sequence[self.state + 1] == token: + # Advance to the next token + next_state = OrderedConstraintState(self.sequence, self.state + 1) + + elif self.sequence.endpoints[self.state]: + # Accept anything between constraints (*) + next_state = self.copy() + + elif token == self.sequence[0]: + # Start over having generated the first token + next_state = OrderedConstraintState(self.sequence, 0) + else: + # Start over from the root + next_state = OrderedConstraintState(self.sequence, -1) + + return next_state diff --git a/fairseq/tokenizer.py b/fairseq/tokenizer.py new file mode 100644 index 0000000000000000000000000000000000000000..42131f7b1d334020c3b48a6e44d4139f7c62ad28 --- /dev/null +++ b/fairseq/tokenizer.py @@ -0,0 +1,15 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import re + + +SPACE_NORMALIZER = re.compile(r"\s+") + + +def tokenize_line(line): + line = SPACE_NORMALIZER.sub(" ", line) + line = line.strip() + return line.split() diff --git a/fairseq/trainer.py b/fairseq/trainer.py new file mode 100644 index 0000000000000000000000000000000000000000..1deb14326f90dea246b9a1a8d3b97b95c5472a5e --- /dev/null +++ b/fairseq/trainer.py @@ -0,0 +1,1439 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +""" +Train a network across multiple GPUs. +""" + +import contextlib +import logging +import sys +import time +from argparse import Namespace +from itertools import chain +from typing import Any, Dict, List + +import torch +from fairseq import checkpoint_utils, models, optim, utils +from fairseq.dataclass.configs import FairseqConfig +from fairseq.dataclass.utils import convert_namespace_to_omegaconf +from fairseq.distributed import utils as distributed_utils +from fairseq.file_io import PathManager +from fairseq.logging import meters, metrics +from fairseq.nan_detector import NanDetector +from fairseq.optim import lr_scheduler +from omegaconf import OmegaConf + +logger = logging.getLogger(__name__) + + +class Trainer(object): + """Main class for data parallel training. + + This class supports synchronous distributed data parallel training, + where multiple workers each have a full model replica and gradients + are accumulated across workers before each update. We use + :class:`~torch.nn.parallel.DistributedDataParallel` to handle + communication of the gradients across workers. + """ + + def __init__(self, cfg: FairseqConfig, task, model, criterion, quantizer=None): + + if isinstance(cfg, Namespace): + logger.warning( + "argparse.Namespace configuration is deprecated! Automatically converting to OmegaConf" + ) + cfg = convert_namespace_to_omegaconf(cfg) + + self.cfg = cfg + self.task = task + + # catalog shared parameters + shared_params = _catalog_shared_params(model) + self.tpu = cfg.common.tpu + self.cuda = torch.cuda.is_available() and not cfg.common.cpu and not self.tpu + if self.cuda: + self.device = torch.device("cuda") + elif self.tpu: + self.device = utils.get_tpu_device() + else: + self.device = torch.device("cpu") + + if self.cfg.distributed_training.ddp_backend == "fully_sharded": + if self.cfg.common.bf16: + raise ValueError( + "FullyShardedDataParallel is not compatible with --bf16 or " + "--memory-efficient-bf16" + ) + if self.cfg.distributed_training.zero_sharding != "none": + raise ValueError( + "FullyShardedDataParallel is not compatible with --zero-sharding " + "option (it's already built in)" + ) + else: + if ( + hasattr(self.cfg.distributed_training, "cpu_offload") + and self.cfg.distributed_training.cpu_offload + ): + raise ValueError("--cpu-offload requires --ddp-backend=fully_sharded") + + # copy model and criterion to current device/dtype + self._criterion = criterion + self._model = model + if cfg.distributed_training.ddp_backend != "fully_sharded": + if cfg.common.fp16: + assert not cfg.common.amp, "Cannot use fp16 and AMP together" + self._criterion = self._criterion.half() + self._model = self._model.half() + elif cfg.common.bf16: + self._criterion = self._criterion.to(dtype=torch.bfloat16) + self._model = self._model.to(dtype=torch.bfloat16) + elif cfg.common.amp: + self._amp_retries = 0 + if ( + not cfg.distributed_training.pipeline_model_parallel + # the DistributedFairseqModel wrapper will handle moving to device, + # so only handle cases which don't use the wrapper + and not self.use_distributed_wrapper + ): + self._criterion = self._criterion.to(device=self.device) + self._model = self._model.to(device=self.device) + self.pipeline_model_parallel = cfg.distributed_training.pipeline_model_parallel + self.last_device = None + if self.cuda and self.pipeline_model_parallel: + self.last_device = torch.device( + cfg.distributed_training.pipeline_devices[-1] + ) + + # check that shared parameters are preserved after device transfer + for shared_param in shared_params: + ref = _get_module_by_path(self._model, shared_param[0]) + for path in shared_param[1:]: + logger.info( + "detected shared parameter: {} <- {}".format(shared_param[0], path) + ) + _set_module_by_path(self._model, path, ref) + + self._dummy_batch = None # indicates we don't have a dummy batch at first + self._lr_scheduler = None + self._num_updates = 0 + self._num_xla_compiles = 0 # for TPUs + self._optim_history = None + self._optimizer = None + self._warn_once = set() + self._wrapped_criterion = None + self._wrapped_model = None + + # TODO(myleott): support tpu + if self.cuda and self.data_parallel_world_size > 1: + self._grad_norm_buf = torch.cuda.DoubleTensor(self.data_parallel_world_size) + else: + self._grad_norm_buf = None + + self.quantizer = quantizer + if self.quantizer is not None: + self.quantizer.set_trainer(self) + + # get detailed cuda environment + if self.cuda: + self.cuda_env = utils.CudaEnvironment() + if self.data_parallel_world_size > 1: + self.cuda_env_arr = distributed_utils.all_gather_list( + self.cuda_env, group=distributed_utils.get_global_group() + ) + else: + self.cuda_env_arr = [self.cuda_env] + if self.data_parallel_rank == 0: + utils.CudaEnvironment.pretty_print_cuda_env_list(self.cuda_env_arr) + else: + self.cuda_env = None + self.cuda_env_arr = None + + metrics.log_start_time("wall", priority=790, round=0) + + self._start_time = time.time() + self._previous_training_time = 0 + self._cumulative_training_time = None + + def reinitialize(self): + """Reinitialize the Trainer, typically after model params change.""" + self._lr_scheduler = None + self._optimizer = None + self._wrapped_criterion = None + self._wrapped_model = None + + @property + def data_parallel_world_size(self): + if self.cfg.distributed_training.distributed_world_size == 1: + return 1 + return distributed_utils.get_data_parallel_world_size() + + @property + def data_parallel_process_group(self): + return distributed_utils.get_data_parallel_group() + + @property + def data_parallel_rank(self): + if self.cfg.distributed_training.distributed_world_size == 1: + return 0 + return distributed_utils.get_data_parallel_rank() + + @property + def is_data_parallel_master(self): + # NOTE: this returns true for all model parallel replicas with data + # parallel rank 0 + return self.data_parallel_rank == 0 + + @property + def use_distributed_wrapper(self) -> bool: + return ( + self.data_parallel_world_size > 1 and not self.cfg.optimization.use_bmuf + ) or ( + self.cfg.distributed_training.ddp_backend == "fully_sharded" + and self.cfg.distributed_training.cpu_offload + ) + + @property + def should_save_checkpoint_on_current_rank(self) -> bool: + """Indicates whether to save checkpoints on the current DDP rank.""" + if ( + self.cfg.distributed_training.ddp_backend == "fully_sharded" + and self.cfg.distributed_training.use_sharded_state + ) or getattr(self.cfg.model, "base_layers", 0) > 0: + return True + else: + return self.is_data_parallel_master + + @property + def always_call_state_dict_during_save_checkpoint(self) -> bool: + if ( + self.cfg.distributed_training.ddp_backend == "fully_sharded" + and not self.cfg.distributed_training.use_sharded_state + ): + # FSDP calls communication collective when consolidating checkpoints + return True + else: + return False + + @property + def checkpoint_suffix(self) -> str: + """Suffix to add to the checkpoint file name.""" + if ( + self.cfg.distributed_training.ddp_backend == "fully_sharded" + and self.cfg.distributed_training.use_sharded_state + ): + return self.cfg.checkpoint.checkpoint_suffix + "-shard{0}".format( + self.data_parallel_rank + ) + else: + return self.cfg.checkpoint.checkpoint_suffix or "" + + @property + def criterion(self): + if self._wrapped_criterion is None: + if utils.has_parameters(self._criterion) and self.use_distributed_wrapper: + self._wrapped_criterion = models.DistributedFairseqModel( + self.cfg.distributed_training, + self._criterion, + process_group=self.data_parallel_process_group, + device=self.device, + ) + else: + self._wrapped_criterion = self._criterion + return self._wrapped_criterion + + @property + def model(self): + if self._wrapped_model is None: + if self.use_distributed_wrapper: + self._wrapped_model = models.DistributedFairseqModel( + self.cfg.distributed_training, + self._model, + process_group=self.data_parallel_process_group, + device=self.device, + ) + else: + self._wrapped_model = self._model + return self._wrapped_model + + @property + def optimizer(self): + if self._optimizer is None: + self._build_optimizer() + return self._optimizer + + @property + def lr_scheduler(self): + if self._lr_scheduler is None: + self._build_optimizer() # this will initialize self._lr_scheduler + return self._lr_scheduler + + def _build_optimizer(self): + params = list( + filter( + lambda p: p.requires_grad, + chain(self.model.parameters(), self.criterion.parameters()), + ) + ) + + if ( + self.cfg.distributed_training.ddp_backend == "fully_sharded" + and self.cfg.common.fp16 + ): + # FullyShardedDataParallel always uses MemoryEfficientFP16 wrapper, + # mostly for the grad scaling. But if we don't have the + # --memory-efficient-fp16 flag set, then we're effectively doing + # regular --fp16 and can allow the use of optimizers that would + # otherwise be unsupported by MemoryEfficientFP16Optimizer. + allow_unsupported = not self.cfg.common.memory_efficient_fp16 + self._optimizer = optim.MemoryEfficientFP16Optimizer.build_optimizer( + self.cfg, params, allow_unsupported=allow_unsupported + ) + elif self.cfg.common.fp16 or self.cfg.common.bf16 or self.cfg.common.amp: + if self.cuda and torch.cuda.get_device_capability(0)[0] < 7: + logger.info( + "NOTE: your device does NOT support faster training with --fp16 or --amp, " + "please switch to FP32 which is likely to be faster" + ) + if ( + self.cfg.common.memory_efficient_fp16 + or self.cfg.common.memory_efficient_bf16 + ): + self._optimizer = optim.MemoryEfficientFP16Optimizer.build_optimizer( + self.cfg, params + ) + elif self.cfg.common.amp: + self._optimizer = optim.AMPOptimizer.build_optimizer(self.cfg, params) + else: + self._optimizer = optim.FP16Optimizer.build_optimizer(self.cfg, params) + else: + if self.cuda and torch.cuda.get_device_capability(0)[0] >= 7: + logger.info("NOTE: your device may support faster training with --fp16 or --amp") + self._optimizer = optim.build_optimizer(self.cfg.optimizer, params) + + if self.cfg.distributed_training.ddp_backend == "fully_sharded": + assert ( + not self.cfg.optimization.use_bmuf + ), "--ddp-backend=fully_sharded is not compatible with BMUF" + assert self._optimizer.supports_flat_params, ( + "--ddp-backend=fully_sharded is only compatible with pointwise " + "optimizers (e.g., Adam, AdamW, Adadelta, Adamax, SGD, etc.). " + "However, the sharding will result in slightly different results when " + "using non-pointwise optimizers (e.g., Adagrad, Adafactor, LAMB)" + ) + + if self.cfg.optimization.use_bmuf: + self._optimizer = optim.FairseqBMUF( + self.cfg.bmuf, + self._optimizer, + ) + + if self.cfg.distributed_training.zero_sharding == "os": + if ( + self.cfg.common.fp16 + and not self.cfg.common.memory_efficient_fp16 + and not self.cfg.common.memory_efficient_bf16 + ) and not self.cfg.common.fp16_no_flatten_grads: + raise ValueError( + "ZeRO is incomptabile with fp16 and flattened grads. " + "Please use --fp16-no-flatten-grads" + ) + else: + optim.shard_(self._optimizer, self.data_parallel_process_group) + + # We should initialize the learning rate scheduler immediately after + # building the optimizer, so that the initial learning rate is set. + self._lr_scheduler = lr_scheduler.build_lr_scheduler( + self.cfg.lr_scheduler, + self.optimizer, + ) + self._lr_scheduler.step_update(0) + + def consolidate_optimizer(self): + """For OSS, we need to consolidate the state dict.""" + if self.cfg.checkpoint.no_save_optimizer_state: + return + self._gathered_optim_state = None + if hasattr(self.optimizer.optimizer, "consolidate_state_dict"): + self.optimizer.optimizer.consolidate_state_dict() + + elif ( + self.cfg.distributed_training.ddp_backend == "fully_sharded" + and not self.model.use_sharded_state + ): + st = self.model.gather_full_optim_state_dict( + self.optimizer + ) # only returns on rank 0 + self._gathered_optim_state = st + + def state_dict(self): + state_dict = { + "args": None, # legacy + "cfg": ( + OmegaConf.to_container(self.cfg, resolve=True, enum_to_str=True) + if OmegaConf.is_config(self.cfg) + else self.cfg + ), + "model": self.model.state_dict(), + "criterion": ( + self.criterion.state_dict() + if utils.has_parameters(self.criterion) + else None + ), + "optimizer_history": (self._optim_history or []) + + [ + { + "criterion_name": self.get_criterion().__class__.__name__, + "optimizer_name": self.optimizer.__class__.__name__, + "lr_scheduler_state": self.lr_scheduler.state_dict(), + "num_updates": self.get_num_updates(), + } + ], + "task_state": self.task.state_dict() if self.task is not None else {}, + "extra_state": { + "metrics": metrics.state_dict(), + "previous_training_time": self.cumulative_training_time(), + }, + } + if not self.cfg.checkpoint.no_save_optimizer_state: + if self._gathered_optim_state is not None: + state_dict["last_optimizer_state"] = self._gathered_optim_state + self._gathered_optim_state = None + else: + state_dict["last_optimizer_state"] = self.optimizer.state_dict() + if self.cfg.distributed_training.ddp_backend == "fully_sharded": + # save meta data for recombining checkpoint upon loading + state_dict["fsdp_metadata"] = self.model.local_metadata_dict() + return state_dict + + def save_checkpoint(self, filename, extra_state): + """Save all training state in a checkpoint file.""" + logger.info(f"Saving checkpoint to {filename}") + # call state_dict on all ranks in case it needs internal communication + state_dict = utils.move_to_cpu(self.state_dict()) + state_dict["extra_state"].update(extra_state) + if self.should_save_checkpoint_on_current_rank: + checkpoint_utils.torch_persistent_save( + state_dict, + filename, + async_write=self.cfg.checkpoint.write_checkpoints_asynchronously, + ) + logger.info(f"Finished saving checkpoint to {filename}") + + def load_checkpoint( + self, + filename, + reset_optimizer=False, + reset_lr_scheduler=False, + optimizer_overrides=None, + reset_meters=False, + ): + """ + Load all training state from a checkpoint file. + rank = 0 will load the checkpoint, and then broadcast it to all + other ranks. + """ + extra_state, self._optim_history, last_optim_state = None, [], None + + logger.info(f"Preparing to load checkpoint {filename}") + is_distributed = self.data_parallel_world_size > 1 + bexists = PathManager.isfile(filename) + if bexists: + load_on_all_ranks = ( + self.cfg.checkpoint.load_checkpoint_on_all_dp_ranks + # TPUs don't support broadcast yet, so load checkpoints + # on every worker for now + or self.tpu + # FSDP requires loading checkpoint shards on all ranks + or ( + self.cfg.distributed_training.ddp_backend == "fully_sharded" + and self.cfg.distributed_training.use_sharded_state + ) + or getattr(self.cfg.model, "base_layers", 0) > 0 + ) + + if load_on_all_ranks or self.data_parallel_rank == 0: + state = checkpoint_utils.load_checkpoint_to_cpu( + filename, load_on_all_ranks=load_on_all_ranks + ) + last_optim_state = state.get("last_optimizer_state", None) + + # If doing zero_sharding, do not broadcast global optimizer + # state. Later we will broadcast sharded states to each rank + # to avoid memory from exploding. + if ( + not load_on_all_ranks + and self.cfg.distributed_training.zero_sharding == "os" + and "last_optimizer_state" in state + and is_distributed + ): + state["last_optimizer_state"] = "SHARDED" + else: + last_optim_state = None + state = None + + if is_distributed and not load_on_all_ranks: + state = distributed_utils.broadcast_object( + state, + src_rank=0, + group=self.data_parallel_process_group, + dist_device=self.device, + ) + if self.data_parallel_rank > 0: + last_optim_state = state.get("last_optimizer_state", None) + + # load model parameters + try: + self.model.load_state_dict( + state["model"], strict=True, model_cfg=self.cfg.model + ) + # save memory for later steps + del state["model"] + if utils.has_parameters(self.get_criterion()): + self.get_criterion().load_state_dict( + state["criterion"], strict=True + ) + del state["criterion"] + + except Exception: + raise Exception( + "Cannot load model parameters from checkpoint {}; " + "please ensure that the architectures match.".format(filename) + ) + extra_state = state["extra_state"] + self._optim_history = state["optimizer_history"] + + if last_optim_state is not None and not reset_optimizer: + # rebuild optimizer after loading model, since params may have changed + self._build_optimizer() + + # only reload optimizer and lr_scheduler if they match + last_optim = self._optim_history[-1] + assert ( + last_optim["criterion_name"] == self.get_criterion().__class__.__name__ + ), f"Criterion does not match; please reset the optimizer (--reset-optimizer). {last_optim['criterion_name']} vs {self.get_criterion().__class__.__name__}" + assert ( + last_optim["optimizer_name"] == self.optimizer.__class__.__name__ + ), f"Optimizer does not match; please reset the optimizer (--reset-optimizer). {last_optim['optimizer_name']} vs {self.optimizer.__class__.__name__}" + + if not reset_lr_scheduler: + self.lr_scheduler.load_state_dict(last_optim["lr_scheduler_state"]) + + if ( + self.cfg.distributed_training.ddp_backend == "fully_sharded" + and not self.model.use_sharded_state + ): + # if use_sharded_state, the last_optim_state is already sharded, skip this + last_optim_state = self.model.get_shard_from_optim_state_dict( + last_optim_state + ) + elif not load_on_all_ranks and is_distributed: + last_optim_state = self.optimizer.broadcast_global_state_dict( + last_optim_state + ) + + self.optimizer.load_state_dict(last_optim_state, optimizer_overrides) + + self.set_num_updates(last_optim["num_updates"]) + + if extra_state is not None: + itr_state = extra_state["train_iterator"] + epoch = itr_state["epoch"] + + if "previous_training_time" in extra_state: + self._previous_training_time = extra_state["previous_training_time"] + self._start_time = time.time() + + self.lr_step(epoch) + + if ( + itr_state.get("version", 1) >= 2 + and itr_state["iterations_in_epoch"] == 0 + ): + # reset meters at start of epoch + reset_meters = True + + if "metrics" in extra_state and not reset_meters: + metrics.load_state_dict(extra_state["metrics"]) + + # reset TimeMeters, since their start times don't make sense anymore + for meter in metrics.get_meters("default"): + if isinstance(meter, meters.TimeMeter): + meter.reset() + + logger.info( + "Loaded checkpoint {} (epoch {} @ {} updates)".format( + filename, epoch, self.get_num_updates() + ) + ) + + else: + logger.info("No existing checkpoint found {}".format(filename)) + + return extra_state + + def get_train_iterator( + self, + epoch, + combine=True, + load_dataset=True, + data_selector=None, + shard_batch_itr=True, + disable_iterator_cache=False, + ): + """Return an EpochBatchIterator over the training set for a given epoch.""" + if load_dataset: + logger.info("loading train data for epoch {}".format(epoch)) + self.task.load_dataset( + self.cfg.dataset.train_subset, + epoch=epoch, + combine=combine, + data_selector=data_selector, + tpu=self.tpu, + ) + batch_iterator = self.task.get_batch_iterator( + dataset=self.task.dataset(self.cfg.dataset.train_subset), + max_tokens=self.cfg.dataset.max_tokens, + max_sentences=self.cfg.dataset.batch_size, + max_positions=utils.resolve_max_positions( + self.task.max_positions(), + self.model.max_positions(), + self.cfg.dataset.max_tokens, + ), + ignore_invalid_inputs=True, + required_batch_size_multiple=self.cfg.dataset.required_batch_size_multiple, + seed=self.cfg.common.seed, + num_shards=self.data_parallel_world_size if shard_batch_itr else 1, + shard_id=self.data_parallel_rank if shard_batch_itr else 0, + num_workers=self.cfg.dataset.num_workers, + epoch=epoch, + data_buffer_size=self.cfg.dataset.data_buffer_size, + disable_iterator_cache=disable_iterator_cache, + ) + self.reset_dummy_batch(batch_iterator.first_batch) + return batch_iterator + + def get_valid_iterator( + self, + subset, + disable_iterator_cache=False, + ): + """Return an EpochBatchIterator over given validation subset for a given epoch.""" + batch_iterator = self.task.get_batch_iterator( + dataset=self.task.dataset(subset), + max_tokens=self.cfg.dataset.max_tokens_valid, + max_sentences=self.cfg.dataset.batch_size_valid, + max_positions=utils.resolve_max_positions( + self.task.max_positions(), + self.model.max_positions(), + ), + ignore_invalid_inputs=self.cfg.dataset.skip_invalid_size_inputs_valid_test, + required_batch_size_multiple=self.cfg.dataset.required_batch_size_multiple, + seed=self.cfg.common.seed, + num_shards=self.data_parallel_world_size, + shard_id=self.data_parallel_rank, + num_workers=self.cfg.dataset.num_workers, + # always pass a fixed "epoch" to keep validation data consistent + # across training epochs + epoch=1, + data_buffer_size=self.cfg.dataset.data_buffer_size, + disable_iterator_cache=disable_iterator_cache, + ) + self.reset_dummy_batch(batch_iterator.first_batch) + return batch_iterator + + def begin_epoch(self, epoch): + """Called at the beginning of each epoch.""" + logger.info("begin training epoch {}".format(epoch)) + + self.lr_step_begin_epoch(epoch) + + if self.quantizer is not None: + self.quantizer.begin_epoch(epoch) + + # task specific setup per epoch + self.task.begin_epoch(epoch, self.get_model()) + + if self.tpu: + import torch_xla.core.xla_model as xm + + xm.rendezvous("begin_epoch") # wait for all workers + xm.mark_step() + + def begin_valid_epoch(self, epoch): + """Called at the beginning of each validation epoch.""" + + # task specific setup per validation epoch + self.task.begin_valid_epoch(epoch, self.get_model()) + + def reset_dummy_batch(self, batch): + self._dummy_batch = batch + + @metrics.aggregate("train") + def train_step(self, samples, raise_oom=False): + """Do forward, backward and parameter update.""" + self._set_seed() + self.model.train() + self.criterion.train() + self.zero_grad() + + metrics.log_start_time("train_wall", priority=800, round=0) + + # forward and backward pass + logging_outputs, sample_size, ooms = [], 0, 0 + for i, sample in enumerate(samples): # delayed update loop + sample, is_dummy_batch = self._prepare_sample(sample) + + def maybe_no_sync(): + """ + Whenever *samples* contains more than one mini-batch, we + want to accumulate gradients locally and only call + all-reduce in the last backwards pass. + """ + if ( + self.data_parallel_world_size > 1 + and hasattr(self.model, "no_sync") + and i < len(samples) - 1 + ): + return self.model.no_sync() + else: + return contextlib.ExitStack() # dummy contextmanager + + try: + with maybe_no_sync(): + # forward and backward + loss, sample_size_i, logging_output = self.task.train_step( + sample=sample, + model=self.model, + criterion=self.criterion, + optimizer=self.optimizer, + update_num=self.get_num_updates(), + ignore_grad=is_dummy_batch, + ) + del loss + + logging_outputs.append(logging_output) + sample_size += sample_size_i + + # emptying the CUDA cache after the first step can + # reduce the chance of OOM + if self.cuda and self.get_num_updates() == 0: + torch.cuda.empty_cache() + except RuntimeError as e: + if "out of memory" in str(e): + self._log_oom(e) + if raise_oom: + raise e + logger.warning( + "attempting to recover from OOM in forward/backward pass" + ) + ooms += 1 + self.zero_grad() + if self.cuda: + torch.cuda.empty_cache() + if self.cfg.distributed_training.distributed_world_size == 1: + return None + else: + raise e + + if self.tpu and i < len(samples) - 1: + # tpu-comment: every XLA operation before marking step is + # appended to the IR graph, and processing too many batches + # before marking step can lead to OOM errors. + # To handle gradient accumulation use case, we explicitly + # mark step here for every forward pass without a backward pass + self._xla_markstep_and_send_to_cpu() + + if is_dummy_batch: + if torch.is_tensor(sample_size): + sample_size.zero_() + else: + sample_size *= 0.0 + + if torch.is_tensor(sample_size): + sample_size = sample_size.float() + else: + sample_size = float(sample_size) + + # gather logging outputs from all replicas + if self._sync_stats(): + train_time = self._local_cumulative_training_time() + logging_outputs, ( + sample_size, + ooms, + total_train_time, + ) = self._aggregate_logging_outputs( + logging_outputs, sample_size, ooms, train_time, ignore=is_dummy_batch + ) + self._cumulative_training_time = ( + total_train_time / self.data_parallel_world_size + ) + + overflow = False + try: + with torch.autograd.profiler.record_function("reduce-grads"): + # reduce gradients across workers + self.optimizer.all_reduce_grads(self.model) + if utils.has_parameters(self.criterion): + self.optimizer.all_reduce_grads(self.criterion) + + with torch.autograd.profiler.record_function("multiply-grads"): + # multiply gradients by (data_parallel_size / sample_size) since + # DDP normalizes by the number of data parallel workers for + # improved fp16 precision. + # Thus we get (sum_of_gradients / sample_size) at the end. + # In case of fp16, this step also undoes loss scaling. + # (Debugging note: Some optimizers perform this scaling on the + # fly, so inspecting model.parameters() or optimizer.params may + # still show the original, unscaled gradients.) + numer = ( + self.data_parallel_world_size + if not self.cfg.optimization.use_bmuf or self._sync_stats() + else 1 + ) + self.optimizer.multiply_grads(numer / (sample_size or 1.0)) + # Note: (sample_size or 1.0) handles the case of a zero gradient, in a + # way that avoids CPU/device transfers in case sample_size is a GPU or + # TPU object. The assumption is that the gradient itself is also 0. + + with torch.autograd.profiler.record_function("clip-grads"): + # clip grads + grad_norm = self.clip_grad_norm(self.cfg.optimization.clip_norm) + + # check that grad norms are consistent across workers + # on tpu check tensor is slow + if not self.tpu: + if ( + not self.cfg.optimization.use_bmuf + and self.cfg.distributed_training.ddp_backend != "slow_mo" + ): + self._check_grad_norms(grad_norm) + if not torch.isfinite(grad_norm).all(): + # in case of AMP, if gradients are Nan/Inf then + # optimizer step is still required + if self.cfg.common.amp: + overflow = True + else: + # check local gradnorm single GPU case, trigger NanDetector + raise FloatingPointError("gradients are Nan/Inf") + + with torch.autograd.profiler.record_function("optimizer"): + # take an optimization step + self.task.optimizer_step( + self.optimizer, model=self.model, update_num=self.get_num_updates() + ) + if self.cfg.common.amp and overflow: + if self._amp_retries == self.cfg.common.amp_batch_retries: + logger.info("AMP: skipping this batch.") + self._amp_retries = 0 + else: + self._amp_retries += 1 + return self.train_step(samples, raise_oom) # recursion to feed in same batch + + except FloatingPointError: + # re-run the forward and backward pass with hooks attached to print + # out where it fails + self.zero_grad() + with NanDetector(self.get_model()): + for _, sample in enumerate(samples): + sample, _ = self._prepare_sample(sample) + self.task.train_step( + sample, + self.model, + self.criterion, + self.optimizer, + self.get_num_updates(), + ignore_grad=False, + ) + raise + except OverflowError as e: + overflow = True + logger.info( + f"NOTE: gradient overflow detected, ignoring gradient, {str(e)}" + ) + grad_norm = torch.tensor(0.0).cuda() + self.zero_grad() + except RuntimeError as e: + if "out of memory" in str(e): + self._log_oom(e) + logger.error("OOM during optimization, irrecoverable") + raise e + + # Some distributed wrappers (e.g., SlowMo) need access to the optimizer + # after the step + if hasattr(self.model, "perform_additional_optimizer_actions"): + if hasattr(self.optimizer, "fp32_params"): + self.model.perform_additional_optimizer_actions( + self.optimizer.optimizer, self.optimizer.fp32_params + ) + else: + self.model.perform_additional_optimizer_actions( + self.optimizer.optimizer + ) + + logging_output = None + if not overflow or self.cfg.distributed_training.ddp_backend == "slow_mo": + self.set_num_updates(self.get_num_updates() + 1) + + if self.tpu: + import torch_xla.core.xla_model as xm + + # mark step on TPUs + self._xla_markstep_and_send_to_cpu() + + # only log stats every log_interval steps + # this causes wps to be misreported when log_interval > 1 + logging_output = {} + if self.get_num_updates() % self.cfg.common.log_interval == 0: + # log memory usage + mem_info = xm.get_memory_info(self.device) + gb_free = mem_info["kb_free"] / 1024 / 1024 + gb_total = mem_info["kb_total"] / 1024 / 1024 + metrics.log_scalar( + "gb_free", gb_free, priority=1500, round=1, weight=0 + ) + metrics.log_scalar( + "gb_total", gb_total, priority=1600, round=1, weight=0 + ) + logging_outputs = self._xla_markstep_and_send_to_cpu( + logging_outputs + ) + logging_output = self._reduce_and_log_stats( + logging_outputs, sample_size, grad_norm + ) + + # log whenever there's an XLA compilation, since these + # slow down training and may indicate opportunities for + # optimization + self._check_xla_compilation() + else: + if self.cuda and self.cuda_env is not None: + # log minimum free memory over the iteration + gb_used = torch.cuda.max_memory_allocated() / 1024 / 1024 / 1024 + torch.cuda.reset_peak_memory_stats() + gb_free = self.cuda_env.total_memory_in_GB - gb_used + metrics.log_scalar( + "gb_free", gb_free, priority=1500, round=1, weight=0 + ) + + # log stats + logging_output = self._reduce_and_log_stats( + logging_outputs, sample_size, grad_norm + ) + + # clear CUDA cache to reduce memory fragmentation + if ( + self.cuda + and self.cfg.common.empty_cache_freq > 0 + and ( + (self.get_num_updates() + self.cfg.common.empty_cache_freq - 1) + % self.cfg.common.empty_cache_freq + ) + == 0 + ): + torch.cuda.empty_cache() + + if self.cfg.common.fp16 or self.cfg.common.amp: + metrics.log_scalar( + "loss_scale", + ( + self.optimizer.scaler.loss_scale + if self.cfg.common.fp16 + else self.optimizer.scaler.get_scale() + ), + priority=700, + round=4, + weight=0, + ) + + metrics.log_stop_time("train_wall") + return logging_output + + @metrics.aggregate("valid") + def valid_step(self, sample, raise_oom=False): + """Do forward pass in evaluation mode.""" + if self.tpu: + import torch_xla.core.xla_model as xm + + xm.rendezvous("valid_step") # wait for all workers + + with torch.no_grad(): + self.model.eval() + self.criterion.eval() + + sample, is_dummy_batch = self._prepare_sample(sample) + + try: + _loss, sample_size, logging_output = self.task.valid_step( + sample, self.model, self.criterion + ) + except RuntimeError as e: + if "out of memory" in str(e): + self._log_oom(e) + if not raise_oom: + logger.warning( + "ran out of memory in validation step, retrying batch" + ) + for p in self.model.parameters(): + if p.grad is not None: + p.grad = None # free some memory + if self.cuda: + torch.cuda.empty_cache() + return self.valid_step(sample, raise_oom=True) + raise e + + logging_outputs = [logging_output] + if is_dummy_batch: + if torch.is_tensor(sample_size): + sample_size.zero_() + else: + sample_size *= 0.0 + + # gather logging outputs from all replicas + if self.data_parallel_world_size > 1: + logging_outputs, (sample_size,) = self._aggregate_logging_outputs( + logging_outputs, + sample_size, + ignore=is_dummy_batch, + ) + + # log validation stats + if self.tpu: + logging_outputs = self._xla_markstep_and_send_to_cpu(logging_outputs) + logging_output = self._reduce_and_log_stats(logging_outputs, sample_size) + + return logging_output + + def zero_grad(self): + self.optimizer.zero_grad() + + def lr_step_begin_epoch(self, epoch): + """Adjust the learning rate at the beginning of the epoch.""" + self.lr_scheduler.step_begin_epoch(epoch) + # prefer updating the LR based on the number of steps + return self.lr_step_update() + + def lr_step(self, epoch, val_loss=None): + """Adjust the learning rate at the end of the epoch.""" + self.lr_scheduler.step(epoch, val_loss) + # prefer updating the LR based on the number of steps + return self.lr_step_update() + + def lr_step_update(self): + """Update the learning rate after each update.""" + new_lr = self.lr_scheduler.step_update(self.get_num_updates()) + if isinstance(new_lr, dict): + for k, v in new_lr.items(): + metrics.log_scalar(f"lr_{k}", v, weight=0, priority=300) + new_lr = new_lr.get("default", next(iter(new_lr.values()))) + else: + metrics.log_scalar("lr", new_lr, weight=0, priority=300) + return new_lr + + def get_lr(self): + """Get the current learning rate.""" + return self.optimizer.get_lr() + + def get_model(self): + """Get the (non-wrapped) model instance.""" + return self._model + + def get_criterion(self): + """Get the (non-wrapped) criterion instance.""" + return self._criterion + + def get_meter(self, name): + """[deprecated] Get a specific meter by name.""" + from fairseq import meters + + if "get_meter" not in self._warn_once: + self._warn_once.add("get_meter") + utils.deprecation_warning( + "Trainer.get_meter is deprecated. Please use fairseq.metrics instead." + ) + + train_meters = metrics.get_meters("train") + if train_meters is None: + train_meters = {} + + if name == "train_loss" and "loss" in train_meters: + return train_meters["loss"] + elif name == "train_nll_loss": + # support for legacy train.py, which assumed this meter is + # always initialized + m = train_meters.get("nll_loss", None) + return m or meters.AverageMeter() + elif name == "wall": + # support for legacy train.py, which assumed this meter is + # always initialized + m = metrics.get_meter("default", "wall") + return m or meters.TimeMeter() + elif name == "wps": + m = metrics.get_meter("train", "wps") + return m or meters.TimeMeter() + elif name in {"valid_loss", "valid_nll_loss"}: + # support for legacy train.py, which assumed these meters + # are always initialized + k = name[len("valid_") :] + m = metrics.get_meter("valid", k) + return m or meters.AverageMeter() + elif name == "oom": + return meters.AverageMeter() + elif name in train_meters: + return train_meters[name] + return None + + def get_num_updates(self): + """Get the number of parameters updates.""" + return self._num_updates + + def set_num_updates(self, num_updates): + """Set the number of parameters updates.""" + self._num_updates = num_updates + self.lr_step_update() + if self.quantizer: + self.quantizer.step_update(self._num_updates) + metrics.log_scalar("num_updates", self._num_updates, weight=0, priority=200) + + def clip_grad_norm(self, clip_norm): + def agg_norm_fn(total_norm): + total_norm = total_norm.cuda().float() ** 2 + total_norm = distributed_utils.all_reduce( + total_norm, group=self.data_parallel_process_group + ) + return total_norm ** 0.5 + + should_agg_norm = ( + self.cfg.distributed_training.ddp_backend == "fully_sharded" + and ( + self.data_parallel_process_group is not None + or torch.distributed.is_initialized() + ) + ) + return self.optimizer.clip_grad_norm( + clip_norm, aggregate_norm_fn=agg_norm_fn if should_agg_norm else None + ) + + def cumulative_training_time(self): + if self._cumulative_training_time is None: + # single GPU + return self._local_cumulative_training_time() + else: + return self._cumulative_training_time + + def _local_cumulative_training_time(self): + """Aggregate training time in seconds.""" + return time.time() - self._start_time + self._previous_training_time + + def _fp_convert_sample(self, sample): + def apply_half(t): + if t.dtype is torch.float32: + return t.to(dtype=torch.half) + return t + + def apply_bfloat16(t): + if t.dtype is torch.float32: + return t.to(dtype=torch.bfloat16) + return t + + if self.cfg.common.fp16: + sample = utils.apply_to_sample(apply_half, sample) + + if self.cfg.common.bf16: + sample = utils.apply_to_sample(apply_bfloat16, sample) + + return sample + + def _prepare_sample(self, sample, is_dummy=False): + if sample == "DUMMY": + raise Exception( + "Trying to use an uninitialized 'dummy' batch. This usually indicates " + "that the total number of batches is smaller than the number of " + "participating GPUs. Try reducing the batch size or using fewer GPUs." + ) + + if sample is None or len(sample) == 0: + assert ( + self._dummy_batch is not None and len(self._dummy_batch) > 0 + ), "Invalid dummy batch: {}".format(self._dummy_batch) + sample, _ = self._prepare_sample(self._dummy_batch, is_dummy=True) + return sample, True + + # Given that PCIe/NVLink bandwidth is significantly smaller than DRAM bandwidth + # it makes sense to do the format conversion on the CPU and then transfer + # a smaller buffer to the device. This also saves GPU memory capacity. + + if self.cfg.common.on_cpu_convert_precision: + sample = self._fp_convert_sample(sample) + + if self.cuda: + if self.pipeline_model_parallel: + if 'target' in sample: + sample['target'] = utils.move_to_cuda(sample['target'], device=self.last_device) + else: + sample = utils.move_to_cuda(sample) + elif self.tpu and is_dummy: + # the dummy batch may not be on the appropriate device + sample = utils.move_to_cuda(sample, device=self.device) + + if not self.cfg.common.on_cpu_convert_precision: + sample = self._fp_convert_sample(sample) + + if self._dummy_batch == "DUMMY": + self._dummy_batch = sample + + return sample, False + + def _set_seed(self): + # Set seed based on args.seed and the update number so that we get + # reproducible results when resuming from checkpoints + seed = self.cfg.common.seed + self.get_num_updates() + utils.set_torch_seed(seed) + + def _sync_stats(self): + # Return True if it's using multiple GPUs and DDP or multiple GPUs with + # BMUF and it's a bmuf sync with warmup iterations completed before. + if self.data_parallel_world_size == 1: + return False + elif self.cfg.optimization.use_bmuf: + return ( + self.get_num_updates() + 1 + ) % self.cfg.bmuf.global_sync_iter == 0 and ( + self.get_num_updates() + 1 + ) > self.cfg.bmuf.warmup_iterations + else: + return True + + def _log_oom(self, exc): + msg = "OOM: Ran out of memory with exception: {}".format(exc) + logger.warning(msg) + if torch.cuda.is_available() and hasattr(torch.cuda, "memory_summary"): + for device_idx in range(torch.cuda.device_count()): + logger.warning(torch.cuda.memory_summary(device=device_idx)) + sys.stderr.flush() + + def _aggregate_logging_outputs( + self, + logging_outputs: List[Dict[str, Any]], + *extra_stats_to_sum, + ignore=False, + ): + if self.task.__class__.logging_outputs_can_be_summed(self.get_criterion()): + return self._fast_stat_sync_sum( + logging_outputs, *extra_stats_to_sum, ignore=ignore + ) + else: + return self._all_gather_list_sync( + logging_outputs, *extra_stats_to_sum, ignore=ignore + ) + + def _all_gather_list_sync( + self, + logging_outputs: List[Dict[str, Any]], + *extra_stats_to_sum, + ignore=False, + ): + """ + Sync logging outputs across workers. all_gather_list_sync is + suitable when logging outputs are complex types. + """ + if self.tpu: + raise NotImplementedError + if ignore: + logging_outputs = [] + results = list( + zip( + *distributed_utils.all_gather_list( + [logging_outputs] + list(extra_stats_to_sum), + max_size=getattr(self.cfg.common, "all_gather_list_size", 16384), + group=self.data_parallel_process_group, + ) + ) + ) + logging_outputs, extra_stats_to_sum = results[0], results[1:] + logging_outputs = list(chain.from_iterable(logging_outputs)) + extra_stats_to_sum = [sum(s) for s in extra_stats_to_sum] + return logging_outputs, extra_stats_to_sum + + def _fast_stat_sync_sum( + self, + logging_outputs: List[Dict[str, Any]], + *extra_stats_to_sum, + ignore=False, + ): + """ + Sync logging outputs across workers. fast_stat_sync_sum is + faster than all_gather_list_sync, but is only suitable when + logging outputs are scalars and can be summed. Note that + *logging_outputs* cannot contain any nested dicts/lists. + """ + data = {} + for i, stat in enumerate(extra_stats_to_sum): + data["extra_stats_" + str(i)] = stat + if len(logging_outputs) > 0: + log_keys = list(logging_outputs[0].keys()) + for k in log_keys: + if not ignore: + v = sum(log[k] for log in logging_outputs if k in log) + else: + v = logging_outputs[0][k] + v = torch.zeros_like(v) if torch.is_tensor(v) else 0 + data["logging_outputs_" + k] = v + else: + log_keys = None + + data = distributed_utils.all_reduce_dict( + data, device=self.device, group=self.data_parallel_process_group + ) + + extra_stats_to_sum = [ + data["extra_stats_" + str(i)] for i in range(len(extra_stats_to_sum)) + ] + if log_keys is not None: + logging_outputs = [{k: data["logging_outputs_" + k] for k in log_keys}] + else: + logging_outputs = [] + return logging_outputs, extra_stats_to_sum + + def _check_grad_norms(self, grad_norm): + """Check that grad norms are consistent across workers.""" + if self._grad_norm_buf is not None: + self._grad_norm_buf.zero_() + self._grad_norm_buf[self.data_parallel_rank] = grad_norm + distributed_utils.all_reduce( + self._grad_norm_buf, group=self.data_parallel_process_group + ) + + def is_consistent(tensor): + max_abs_diff = torch.max(torch.abs(tensor - tensor[0])) + return ( + (torch.isfinite(tensor).all() + and (max_abs_diff / (tensor[0] + 1e-6) < 1e-6).all()) + or + (self.cfg.common.amp and not torch.isfinite(tensor).all()) + # in case of amp non-finite grads are fine + ) + + if not is_consistent(self._grad_norm_buf): + pretty_detail = "\n".join( + "rank {:3d} = {:.8f}".format(r, n) + for r, n in enumerate(self._grad_norm_buf.tolist()) + ) + error_detail = "grad_norm across the workers:\n{}\n".format( + pretty_detail + ) + # use FloatingPointError to trigger NanDetector + raise FloatingPointError( + "Fatal error: gradients are inconsistent between workers. " + "Try --ddp-backend=legacy_ddp. " + "Or are you mixing up different generation of GPUs in training?" + + "\n" + + "-" * 80 + + "\n{}\n".format(error_detail) + + "-" * 80 + ) + + def _reduce_and_log_stats(self, logging_outputs, sample_size, grad_norm=None): + if grad_norm is not None and ( + not torch.is_tensor(grad_norm) or torch.isfinite(grad_norm) + ): + metrics.log_speed("ups", 1.0, priority=100, round=2) + metrics.log_scalar("gnorm", grad_norm, priority=400, round=3) + if self.cfg.optimization.clip_norm > 0: + metrics.log_scalar( + "clip", + torch.where( + grad_norm > self.cfg.optimization.clip_norm, + grad_norm.new_tensor(100), + grad_norm.new_tensor(0), + ), + priority=500, + round=1, + ) + + with metrics.aggregate() as agg: + if logging_outputs is not None: + self.task.reduce_metrics(logging_outputs, self.get_criterion()) + del logging_outputs + + # extra warning for criterions that don't properly log a loss value + if "loss" not in agg: + if "loss" not in self._warn_once: + self._warn_once.add("loss") + logger.warning( + "Criterion.reduce_metrics did not log a 'loss' value, " + "which may break some functionality" + ) + metrics.log_scalar("loss", -1) + + # support legacy interface + if self.tpu: + logging_output = {} + else: + logging_output = agg.get_smoothed_values() + logging_output["sample_size"] = sample_size + for key_to_delete in ["ppl", "wps", "wpb", "bsz"]: + if key_to_delete in logging_output: + del logging_output[key_to_delete] + return logging_output + + def _check_xla_compilation(self): + import torch_xla.debug.metrics as met + + compile_stats = met.metric_data("CompileTime") + if compile_stats is None: + return + num_xla_compiles = compile_stats[0] + if num_xla_compiles > self._num_xla_compiles: + logger.warning( + "XLA compilation detected on device #{}; too many of these can lead " + "to slow training, but we expect a few in the beginning".format( + self.cfg.distributed_training.distributed_rank + ) + ) + self._num_xla_compiles = num_xla_compiles + + def _xla_markstep_and_send_to_cpu(self, data=None): + import torch_xla.core.xla_model as xm + + xm.mark_step() + if data is not None: + from fairseq.utils import xla_device_to_cpu + + return xla_device_to_cpu(data) + + +def _catalog_shared_params(module, memo=None, prefix=""): + if memo is None: + first_call = True + memo = {} + else: + first_call = False + for name, param in module._parameters.items(): + param_prefix = prefix + ("." if prefix else "") + name + if param not in memo: + memo[param] = [] + memo[param].append(param_prefix) + for name, m in module._modules.items(): + if m is None: + continue + submodule_prefix = prefix + ("." if prefix else "") + name + _catalog_shared_params(m, memo, submodule_prefix) + if first_call: + return [x for x in memo.values() if len(x) > 1] + + +def _get_module_by_path(module, path): + path = path.split(".") + for name in path: + module = getattr(module, name) + return module + + +def _set_module_by_path(module, path, value): + path = path.split(".") + for name in path[:-1]: + module = getattr(module, name) + setattr(module, path[-1], value) diff --git a/fairseq/utils.py b/fairseq/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..4fe95b9e8b2b277cd545e12d5980561492b70783 --- /dev/null +++ b/fairseq/utils.py @@ -0,0 +1,807 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import contextlib +import copy +import importlib +import logging +import os +import sys +import warnings +from itertools import accumulate +from typing import Callable, Dict, List, Optional + +import torch +import torch.nn.functional as F +from fairseq.modules.multihead_attention import MultiheadAttention +from torch import Tensor + + +try: + from amp_C import multi_tensor_l2norm + + multi_tensor_l2norm_available = True +except ImportError: + multi_tensor_l2norm_available = False + +try: + import torch_xla.core.xla_model as xm +except ImportError: + xm = None + + +logger = logging.getLogger(__name__) + + +MANIFOLD_PATH_SEP = "|" + + +class FileContentsAction(argparse.Action): + def __init__(self, option_strings, dest, nargs=None, **kwargs): + if nargs is not None: + raise ValueError("nargs not allowed") + super(FileContentsAction, self).__init__(option_strings, dest, **kwargs) + + def __call__(self, parser, namespace, values, option_string=None): + from fairseq.file_io import PathManager + + if PathManager.isfile(values): + with PathManager.open(values) as f: + argument = f.read().strip() + else: + argument = values + setattr(namespace, self.dest, argument) + + +def split_paths(paths: str, separator=os.pathsep) -> List[str]: + return ( + paths.split(separator) if "://" not in paths else paths.split(MANIFOLD_PATH_SEP) + ) + + +def load_ensemble_for_inference(filenames, task, model_arg_overrides=None): + from fairseq import checkpoint_utils + + deprecation_warning( + "utils.load_ensemble_for_inference is deprecated. " + "Please use checkpoint_utils.load_model_ensemble instead." + ) + return checkpoint_utils.load_model_ensemble( + filenames, arg_overrides=model_arg_overrides, task=task + ) + + +def apply_to_sample(f, sample): + if hasattr(sample, "__len__") and len(sample) == 0: + return {} + + def _apply(x): + if torch.is_tensor(x): + return f(x) + elif isinstance(x, dict): + return {key: _apply(value) for key, value in x.items()} + elif isinstance(x, list): + return [_apply(x) for x in x] + elif isinstance(x, tuple): + return tuple(_apply(x) for x in x) + elif isinstance(x, set): + return {_apply(x) for x in x} + else: + return x + + return _apply(sample) + + +def move_to_cuda(sample, device=None): + device = device or torch.cuda.current_device() + + def _move_to_cuda(tensor): + # non_blocking is ignored if tensor is not pinned, so we can always set + # to True (see github.com/PyTorchLightning/pytorch-lightning/issues/620) + return tensor.to(device=device, non_blocking=True) + + return apply_to_sample(_move_to_cuda, sample) + + +def move_to_cpu(sample): + def _move_to_cpu(tensor): + # PyTorch has poor support for half tensors (float16) on CPU. + # Move any such tensors to float32. + if tensor.dtype in {torch.bfloat16, torch.float16}: + tensor = tensor.to(dtype=torch.float32) + return tensor.cpu() + + return apply_to_sample(_move_to_cpu, sample) + + +def move_to_tpu(sample): + + import torch_xla.core.xla_model as xm + + device = xm.xla_device() + + def _move_to_tpu(tensor): + return tensor.to(device) + + return apply_to_sample(_move_to_tpu, sample) + + +def get_incremental_state( + module: MultiheadAttention, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]], + key: str, +) -> Optional[Dict[str, Optional[Tensor]]]: + """Helper for getting incremental state for an nn.Module.""" + return module.get_incremental_state(incremental_state, key) + + +def set_incremental_state( + module: MultiheadAttention, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]], + key: str, + value: Dict[str, Optional[Tensor]], +) -> Optional[Dict[str, Dict[str, Optional[Tensor]]]]: + """Helper for setting incremental state for an nn.Module.""" + if incremental_state is not None: + result = module.set_incremental_state(incremental_state, key, value) + if result is not None: + incremental_state = result + return incremental_state + + +def load_align_dict(replace_unk): + if replace_unk is None: + align_dict = None + elif isinstance(replace_unk, str) and len(replace_unk) > 0: + # Load alignment dictionary for unknown word replacement if it was passed as an argument. + align_dict = {} + with open(replace_unk, "r") as f: + for line in f: + cols = line.split() + align_dict[cols[0]] = cols[1] + else: + # No alignment dictionary provided but we still want to perform unknown word replacement by copying the + # original source word. + align_dict = {} + return align_dict + + +def print_embed_overlap(embed_dict, vocab_dict): + embed_keys = set(embed_dict.keys()) + vocab_keys = set(vocab_dict.symbols) + overlap = len(embed_keys & vocab_keys) + logger.info("found {}/{} types in embedding file".format(overlap, len(vocab_dict))) + + +def parse_embedding(embed_path): + """Parse embedding text file into a dictionary of word and embedding tensors. + + The first line can have vocabulary size and dimension. The following lines + should contain word and embedding separated by spaces. + + Example: + 2 5 + the -0.0230 -0.0264 0.0287 0.0171 0.1403 + at -0.0395 -0.1286 0.0275 0.0254 -0.0932 + """ + embed_dict = {} + with open(embed_path) as f_embed: + next(f_embed) # skip header + for line in f_embed: + pieces = line.rstrip().split(" ") + embed_dict[pieces[0]] = torch.Tensor( + [float(weight) for weight in pieces[1:]] + ) + return embed_dict + + +def load_embedding(embed_dict, vocab, embedding): + for idx in range(len(vocab)): + token = vocab[idx] + if token in embed_dict: + embedding.weight.data[idx] = embed_dict[token] + return embedding + + +def replace_unk(hypo_str, src_str, alignment, align_dict, unk): + from fairseq import tokenizer + + # Tokens are strings here + hypo_tokens = tokenizer.tokenize_line(hypo_str) + # TODO: Very rare cases where the replacement is '<eos>' should be handled gracefully + src_tokens = tokenizer.tokenize_line(src_str) + ["<eos>"] + for i, ht in enumerate(hypo_tokens): + if ht == unk: + src_token = src_tokens[alignment[i]] + # Either take the corresponding value in the aligned dictionary or just copy the original value. + hypo_tokens[i] = align_dict.get(src_token, src_token) + return " ".join(hypo_tokens) + + +def post_process_prediction( + hypo_tokens, + src_str, + alignment, + align_dict, + tgt_dict, + remove_bpe=None, + extra_symbols_to_ignore=None, +): + hypo_str = tgt_dict.string( + hypo_tokens, remove_bpe, extra_symbols_to_ignore=extra_symbols_to_ignore + ) + if align_dict is not None: + hypo_str = replace_unk( + hypo_str, src_str, alignment, align_dict, tgt_dict.unk_string() + ) + if align_dict is not None or remove_bpe is not None: + # Convert back to tokens for evaluating with unk replacement or without BPE + # Note that the dictionary can be modified inside the method. + hypo_tokens = tgt_dict.encode_line(hypo_str, add_if_not_exist=True) + return hypo_tokens, hypo_str, alignment + + +def make_positions(tensor, padding_idx: int, onnx_trace: bool = False): + """Replace non-padding symbols with their position numbers. + + Position numbers begin at padding_idx+1. Padding symbols are ignored. + """ + # The series of casts and type-conversions here are carefully + # balanced to both work with ONNX export and XLA. In particular XLA + # prefers ints, cumsum defaults to output longs, and ONNX doesn't know + # how to handle the dtype kwarg in cumsum. + mask = tensor.ne(padding_idx).int() + return (torch.cumsum(mask, dim=1).type_as(mask) * mask).long() + padding_idx + + +def strip_pad(tensor, pad): + return tensor[tensor.ne(pad)] + + +def buffered_arange(max): + if not hasattr(buffered_arange, "buf"): + buffered_arange.buf = torch.LongTensor() + if max > buffered_arange.buf.numel(): + buffered_arange.buf.resize_(max) + torch.arange(max, out=buffered_arange.buf) + return buffered_arange.buf[:max] + + +def convert_padding_direction( + src_tokens, padding_idx, right_to_left: bool = False, left_to_right: bool = False +): + assert right_to_left ^ left_to_right + pad_mask = src_tokens.eq(padding_idx) + if not pad_mask.any(): + # no padding, return early + return src_tokens + if left_to_right and not pad_mask[:, 0].any(): + # already right padded + return src_tokens + if right_to_left and not pad_mask[:, -1].any(): + # already left padded + return src_tokens + max_len = src_tokens.size(1) + buffered = torch.empty(0).long() + if max_len > 0: + torch.arange(max_len, out=buffered) + range = buffered.type_as(src_tokens).expand_as(src_tokens) + num_pads = pad_mask.long().sum(dim=1, keepdim=True) + if right_to_left: + index = torch.remainder(range - num_pads, max_len) + else: + index = torch.remainder(range + num_pads, max_len) + return src_tokens.gather(1, index) + + +def item(tensor): + # tpu-comment: making this a no-op for xla devices. + if torch.is_tensor(tensor) and tensor.device.type == "xla": + return tensor.detach() + if hasattr(tensor, "item"): + return tensor.item() + if hasattr(tensor, "__getitem__"): + return tensor[0] + return tensor + + +def multi_tensor_total_norm(grads, chunk_size=2048 * 32) -> torch.Tensor: + per_device_grads = {} + norms = [] + for grad in grads: + device = grad.device + cur_device_grads = per_device_grads.get(device) + if cur_device_grads is None: + cur_device_grads = [] + per_device_grads[device] = cur_device_grads + cur_device_grads.append(grad) + for device in per_device_grads.keys(): + cur_device_grads = per_device_grads[device] + if device.type == "cuda": + # TODO(msb) return has_inf + has_inf = torch.zeros((1, 1), dtype=torch.int, device=device) + with torch.cuda.device(device): + norm = multi_tensor_l2norm( + chunk_size, has_inf, [cur_device_grads], False + ) + norms.append(norm[0].to(torch.cuda.current_device())) + else: + norms += [torch.norm(g, p=2, dtype=torch.float32) for g in cur_device_grads] + total_norm = torch.norm(torch.stack(norms)) + return total_norm + + +@torch.no_grad() +def clip_grad_norm_(params, max_norm, aggregate_norm_fn=None) -> torch.Tensor: + def grad_exists(p): + return p is not None and getattr(p, "grad", None) is not None + + if isinstance(params, torch.Tensor): + params = [params] + params = list(params) + grads = [ + p.grad.detach() for p in params if grad_exists(p) and not hasattr(p, "expert") + ] + expert_grads = [ + p.grad.detach() for p in params if grad_exists(p) and hasattr(p, "expert") + ] + + if len(grads) == 0: + if len(params) > 0: + return params[0].new_tensor(0.0) + else: + return torch.tensor(0.0) + + if len(grads) == 1: + total_norm = torch.norm(grads[0], p=2, dtype=torch.float32) + else: + if multi_tensor_l2norm_available: + total_norm = multi_tensor_total_norm(grads) + else: + if torch.cuda.is_available(): + warnings.warn( + "amp_C fused kernels unavailable, disabling multi_tensor_l2norm; " + "you may get better performance by installing NVIDIA's apex library" + ) + device = torch.cuda.current_device() + elif grads[0].device.type == "xla": + device = grads[0].device + else: + device = torch.device("cpu") + total_norm = torch.norm( + torch.stack( + [torch.norm(g, p=2, dtype=torch.float32).to(device) for g in grads] + ) + ) + + if aggregate_norm_fn is not None: + total_norm = aggregate_norm_fn(total_norm) + + if max_norm > 0: + max_norm = float(max_norm) + clip_coef = (max_norm / (total_norm + 1e-6)).clamp_(max=1) + for g in grads + expert_grads: + g.mul_(clip_coef) + return total_norm + + +def fill_with_neg_inf(t): + """FP16-compatible function that fills a tensor with -inf.""" + return t.float().fill_(float("-inf")).type_as(t) + + +def _match_types(arg1, arg2): + """Convert the numerical argument to the same type as the other argument""" + + def upgrade(arg_number, arg_structure): + if isinstance(arg_structure, tuple): + return tuple([arg_number] * len(arg_structure)) + elif isinstance(arg_structure, dict): + arg = copy.deepcopy(arg_structure) + for k in arg: + arg[k] = upgrade(arg_number, arg_structure[k]) + return arg + else: + return arg_number + + if isinstance(arg1, float) or isinstance(arg1, int): + return upgrade(arg1, arg2), arg2 + elif isinstance(arg2, float) or isinstance(arg2, int): + return arg1, upgrade(arg2, arg1) + + return arg1, arg2 + + +def resolve_max_positions(*args): + """Resolve max position constraints from multiple sources.""" + + def map_value_update(d1, d2): + updated_value = copy.deepcopy(d1) + for key in d2: + if key not in updated_value: + updated_value[key] = d2[key] + else: + updated_value[key] = min(d1[key], d2[key]) + return updated_value + + def nullsafe_min(l): + minim = None + for item in l: + if minim is None: + minim = item + elif item is not None and item < minim: + minim = item + return minim + + max_positions = None + for arg in args: + if max_positions is None: + max_positions = arg + elif arg is not None: + max_positions, arg = _match_types(max_positions, arg) + if isinstance(arg, float) or isinstance(arg, int): + max_positions = min(max_positions, arg) + elif isinstance(arg, dict): + max_positions = map_value_update(max_positions, arg) + else: + max_positions = tuple(map(nullsafe_min, zip(max_positions, arg))) + + return max_positions + + +def import_user_module(args): + module_path = getattr(args, "user_dir", None) + if module_path is not None: + module_path = os.path.abspath(args.user_dir) + if not os.path.exists(module_path) and not os.path.isfile( + os.path.dirname(module_path) + ): + fairseq_rel_path = os.path.join(os.path.dirname(__file__), args.user_dir) + if os.path.exists(fairseq_rel_path): + module_path = fairseq_rel_path + else: + fairseq_rel_path = os.path.join( + os.path.dirname(__file__), "..", args.user_dir + ) + if os.path.exists(fairseq_rel_path): + module_path = fairseq_rel_path + else: + raise FileNotFoundError(module_path) + + # ensure that user modules are only imported once + import_user_module.memo = getattr(import_user_module, "memo", set()) + if module_path not in import_user_module.memo: + import_user_module.memo.add(module_path) + + module_parent, module_name = os.path.split(module_path) + if module_name not in sys.modules: + sys.path.insert(0, module_parent) + importlib.import_module(module_name) + + tasks_path = os.path.join(module_path, "tasks") + if os.path.exists(tasks_path): + from fairseq.tasks import import_tasks + + import_tasks(tasks_path, f"{module_name}.tasks") + + models_path = os.path.join(module_path, "models") + if os.path.exists(models_path): + from fairseq.models import import_models + + import_models(models_path, f"{module_name}.models") + else: + raise ImportError( + "Failed to import --user-dir={} because the corresponding module name " + "({}) is not globally unique. Please rename the directory to " + "something unique and try again.".format(module_path, module_name) + ) + + +def softmax(x, dim: int, onnx_trace: bool = False): + if onnx_trace: + return F.softmax(x.float(), dim=dim) + else: + return F.softmax(x, dim=dim, dtype=torch.float32) + + +def log_softmax(x, dim: int, onnx_trace: bool = False): + if onnx_trace: + return F.log_softmax(x.float(), dim=dim) + else: + return F.log_softmax(x, dim=dim, dtype=torch.float32) + + +def get_perplexity(loss, round=2, base=2): + from fairseq.logging.meters import safe_round + + if loss is None: + return 0.0 + try: + return safe_round(base ** loss, round) + except OverflowError: + return float("inf") + + +def deprecation_warning(message, stacklevel=3): + # don't use DeprecationWarning, since it's ignored by default + warnings.warn(message, stacklevel=stacklevel) + + +def get_activation_fn(activation: str) -> Callable: + """Returns the activation function corresponding to `activation`""" + from fairseq.modules import gelu, gelu_accurate + + if activation == "relu": + return F.relu + elif activation == "gelu": + return gelu + elif activation == "gelu_fast": + deprecation_warning( + "--activation-fn=gelu_fast has been renamed to gelu_accurate" + ) + return gelu_accurate + elif activation == "gelu_accurate": + return gelu_accurate + elif activation == "tanh": + return torch.tanh + elif activation == "linear": + return lambda x: x + else: + raise RuntimeError("--activation-fn {} not supported".format(activation)) + + +def get_available_activation_fns() -> List: + return [ + "relu", + "gelu", + "gelu_fast", # deprecated + "gelu_accurate", + "tanh", + "linear", + ] + + +@contextlib.contextmanager +def model_eval(model): + is_training = model.training + model.eval() + yield + model.train(is_training) + + +def has_parameters(module): + try: + next(module.parameters()) + return True + except StopIteration: + return False + + +def get_rng_state(): + state = {"torch_rng_state": torch.get_rng_state()} + if xm is not None: + state["xla_rng_state"] = xm.get_rng_state() + if torch.cuda.is_available(): + state["cuda_rng_state"] = torch.cuda.get_rng_state() + return state + + +def set_rng_state(state): + torch.set_rng_state(state["torch_rng_state"]) + if xm is not None: + xm.set_rng_state(state["xla_rng_state"]) + if torch.cuda.is_available(): + torch.cuda.set_rng_state(state["cuda_rng_state"]) + + +class set_torch_seed(object): + def __init__(self, seed): + assert isinstance(seed, int) + self.rng_state = get_rng_state() + + torch.manual_seed(seed) + if xm is not None: + xm.set_rng_state(seed) + if torch.cuda.is_available(): + torch.cuda.manual_seed(seed) + + def __enter__(self): + return self + + def __exit__(self, *exc): + set_rng_state(self.rng_state) + + +def parse_alignment(line): + """ + Parses a single line from the alingment file. + + Args: + line (str): String containing the alignment of the format: + <src_idx_1>-<tgt_idx_1> <src_idx_2>-<tgt_idx_2> .. + <src_idx_m>-<tgt_idx_m>. All indices are 0 indexed. + + Returns: + torch.IntTensor: packed alignments of shape (2 * m). + """ + alignments = line.strip().split() + parsed_alignment = torch.IntTensor(2 * len(alignments)) + for idx, alignment in enumerate(alignments): + src_idx, tgt_idx = alignment.split("-") + parsed_alignment[2 * idx] = int(src_idx) + parsed_alignment[2 * idx + 1] = int(tgt_idx) + return parsed_alignment + + +def get_token_to_word_mapping(tokens, exclude_list): + n = len(tokens) + word_start = [int(token not in exclude_list) for token in tokens] + word_idx = list(accumulate(word_start)) + token_to_word = {i: word_idx[i] for i in range(n)} + return token_to_word + + +def extract_hard_alignment(attn, src_sent, tgt_sent, pad, eos): + tgt_valid = ( + ((tgt_sent != pad) & (tgt_sent != eos)).nonzero(as_tuple=False).squeeze(dim=-1) + ) + src_invalid = ( + ((src_sent == pad) | (src_sent == eos)).nonzero(as_tuple=False).squeeze(dim=-1) + ) + src_token_to_word = get_token_to_word_mapping(src_sent, [eos, pad]) + tgt_token_to_word = get_token_to_word_mapping(tgt_sent, [eos, pad]) + alignment = [] + if len(tgt_valid) != 0 and len(src_invalid) < len(src_sent): + attn_valid = attn[tgt_valid] + attn_valid[:, src_invalid] = float("-inf") + _, src_indices = attn_valid.max(dim=1) + for tgt_idx, src_idx in zip(tgt_valid, src_indices): + alignment.append( + ( + src_token_to_word[src_idx.item()] - 1, + tgt_token_to_word[tgt_idx.item()] - 1, + ) + ) + return alignment + + +def extract_soft_alignment(attn, src_sent, tgt_sent, pad, eos): + tgt_valid = ((tgt_sent != pad)).nonzero(as_tuple=False) + src_valid = ((src_sent != pad)).nonzero(as_tuple=False).squeeze(dim=-1) + alignment = [] + if len(tgt_valid) != 0 and len(src_valid) != 0: + attn_valid = attn[tgt_valid, src_valid] + alignment = [ + ["{:.6f}".format(p) for p in src_probs.tolist()] for src_probs in attn_valid + ] + return alignment + + +def new_arange(x, *size): + """ + Return a Tensor of `size` filled with a range function on the device of x. + If size is empty, using the size of the variable x. + """ + if len(size) == 0: + size = x.size() + return torch.arange(size[-1], device=x.device).expand(*size).contiguous() + + +def get_tpu_device(): + return xm.xla_device() + + +def tpu_data_loader(itr): + import torch_xla.core.xla_model as xm + import torch_xla.distributed.parallel_loader as pl + from fairseq.data import iterators + + xm.rendezvous("tpu_data_loader") # wait for all workers + xm.mark_step() + device = xm.xla_device() + return iterators.CountingIterator( + pl.ParallelLoader(itr, [device]).per_device_loader(device), + start=getattr(itr, "n", 0), + total=len(itr), + ) + + +def is_xla_tensor(tensor): + return torch.is_tensor(tensor) and tensor.device.type == "xla" + + +def index_put(tensor, indices, value): + if is_xla_tensor(tensor): + for _ in range(indices.dim(), tensor.dim()): + indices = indices.unsqueeze(-1) + if indices.size(-1) < tensor.size(-1): + indices = indices.expand_as(tensor) + tensor = torch.mul(tensor, ~indices) + torch.mul(value, indices) + else: + tensor[indices] = value + return tensor + + +def xla_device_to_cpu(dat): + import torch_xla.core.xla_model as xm + + return xm._maybe_convert_to_cpu(dat) + + +class CudaEnvironment(object): + def __init__(self): + cur_device = torch.cuda.current_device() + prop = torch.cuda.get_device_properties("cuda:{}".format(cur_device)) + self.name = prop.name + self.major = prop.major + self.minor = prop.minor + self.total_memory_in_GB = prop.total_memory / 1024 / 1024 / 1024 + + @staticmethod + def pretty_print_cuda_env_list(cuda_env_list): + """ + Given a list of CudaEnviorments, pretty print them + """ + num_workers = len(cuda_env_list) + center = "CUDA enviroments for all {} workers".format(num_workers) + banner_len = 40 - len(center) // 2 + first_line = "*" * banner_len + center + "*" * banner_len + logger.info(first_line) + for r, env in enumerate(cuda_env_list): + logger.info( + "rank {:3d}: ".format(r) + + "capabilities = {:2d}.{:<2d} ; ".format(env.major, env.minor) + + "total memory = {:.3f} GB ; ".format(env.total_memory_in_GB) + + "name = {:40s}".format(env.name) + ) + logger.info(first_line) + + +def csv_str_list(x): + return x.split(",") + + +def eval_str_list(x, type=float): + if x is None: + return None + if isinstance(x, str): + x = eval(x) + try: + return list(map(type, x)) + except TypeError: + return [type(x)] + + +def eval_str_dict(x, type=dict): + if x is None: + return None + if isinstance(x, str): + x = eval(x) + return x + + +def eval_bool(x, default=False): + if x is None: + return default + try: + return bool(eval(x)) + except TypeError: + return default + + +def reset_logging(): + root = logging.getLogger() + for handler in root.handlers: + root.removeHandler(handler) + root.setLevel(os.environ.get("LOGLEVEL", "INFO").upper()) + handler = logging.StreamHandler(sys.stdout) + handler.setFormatter( + logging.Formatter( + fmt="%(asctime)s | %(levelname)s | %(name)s | %(message)s", + datefmt="%Y-%m-%d %H:%M:%S", + ) + ) + root.addHandler(handler) diff --git a/fairseq/version.txt b/fairseq/version.txt new file mode 100644 index 0000000000000000000000000000000000000000..41432f00d9ce57fadd55cc7dd27b391ddf5ca0b9 --- /dev/null +++ b/fairseq/version.txt @@ -0,0 +1 @@ +1.0.0a0 diff --git a/fairseq_cli/__init__.py b/fairseq_cli/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/fairseq_cli/eval_lm.py b/fairseq_cli/eval_lm.py new file mode 100644 index 0000000000000000000000000000000000000000..ab6e77029ef738291efd190b1cfe2435dd403dea --- /dev/null +++ b/fairseq_cli/eval_lm.py @@ -0,0 +1,347 @@ +#!/usr/bin/env python3 -u +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +""" +Evaluate the perplexity of a trained language model. +""" + +import logging +import math +import os +import sys +from argparse import Namespace +from typing import Iterable, List, Optional + +import torch +import fairseq +from fairseq import checkpoint_utils, distributed_utils, options, tasks, utils +from fairseq.dataclass.utils import convert_namespace_to_omegaconf +from fairseq.logging import progress_bar +from fairseq.logging.meters import StopwatchMeter +from fairseq.sequence_scorer import SequenceScorer +from omegaconf import DictConfig + + +logging.basicConfig( + format="%(asctime)s | %(levelname)s | %(name)s | %(message)s", + datefmt="%Y-%m-%d %H:%M:%S", + level=os.environ.get("LOGLEVEL", "INFO").upper(), + stream=sys.stdout, +) +logger = logging.getLogger("fairseq_cli.eval_lm") + + +def eval_lm( + models: List[fairseq.models.FairseqModel], + source_dictionary: fairseq.data.Dictionary, + batch_iterator: Iterable, + post_process: Optional[str] = None, + output_word_probs: bool = False, + output_word_stats: bool = False, + target_dictionary: Optional[fairseq.data.Dictionary] = None, + softmax_batch: int = 0, + remove_bos_token: bool = False, + device: Optional[torch.device] = None, +): + """ + Args: + models (List[~fairseq.models.FairseqModel]): list of models to + evaluate. Models are essentially `nn.Module` instances, but + must be compatible with fairseq's `SequenceScorer`. + source_dictionary (~fairseq.data.Dictionary): dictionary for + applying any relevant post processing or outputing word + probs/stats. + batch_iterator (Iterable): yield batches of data + post_process (Optional[str]): post-process text by removing BPE, + letter segmentation, etc. Valid options can be found in + fairseq.data.utils.post_process, although not all options + are implemented here. + output_word_probs (Optional[bool]): output words and their + predicted log probabilities + output_word_stats (Optional[bool]): output word statistics such + as word count and average probability + target_dictionary (Optional[~fairseq.data.Dictionary]): output + dictionary (defaults to *source_dictionary*) + softmax_batch (Optional[bool]): if BxT is more than this, will + batch the softmax over vocab to this amount of tokens, in + order to fit into GPU memory + remove_bos_token (Optional[bool]): if True, confirm that the + first token is the beginning-of-sentence symbol (according + to the relevant dictionary) and remove it from the output + device (Optional[torch.device]): device to use for evaluation + (defaults to device of first model parameter) + """ + if target_dictionary is None: + target_dictionary = source_dictionary + if device is None: + device = next(models[0].parameters()).device + + gen_timer = StopwatchMeter() + scorer = SequenceScorer(target_dictionary, softmax_batch) + + score_sum = 0.0 + count = 0 + + if post_process is not None: + if post_process in {"subword_nmt", "@@ "}: + bpe_cont = post_process.rstrip() + bpe_toks = { + i + for i in range(len(source_dictionary)) + if source_dictionary[i].endswith(bpe_cont) + } + else: + raise NotImplementedError( + "--post-process={post_process} is not implemented" + ) + bpe_len = len(bpe_cont) + else: + bpe_toks = None + bpe_len = 0 + + word_stats = dict() + + for sample in batch_iterator: + if "net_input" not in sample: + continue + + sample = utils.move_to_cuda(sample, device=device) + + gen_timer.start() + hypos = scorer.generate(models, sample) + gen_timer.stop(sample["ntokens"]) + + for i, hypos_i in enumerate(hypos): + hypo = hypos_i[0] + sample_id = sample["id"][i] + + tokens = hypo["tokens"] + tgt_len = tokens.numel() + pos_scores = hypo["positional_scores"].float() + + if remove_bos_token: + assert hypo["tokens"][0].item() == target_dictionary.bos() + tokens = tokens[1:] + pos_scores = pos_scores[1:] + + skipped_toks = 0 + if bpe_toks is not None: + for i in range(tgt_len - 1): + if tokens[i].item() in bpe_toks: + skipped_toks += 1 + pos_scores[i + 1] += pos_scores[i] + pos_scores[i] = 0 + + inf_scores = pos_scores.eq(float("inf")) | pos_scores.eq(float("-inf")) + if inf_scores.any(): + logger.info( + "skipping tokens with inf scores:", + target_dictionary.string(tokens[inf_scores.nonzero()]), + ) + pos_scores = pos_scores[(~inf_scores).nonzero()] + score_sum += pos_scores.sum().cpu() + count += pos_scores.numel() - skipped_toks + + if output_word_probs or output_word_stats: + w = "" + word_prob = [] + is_bpe = False + for i in range(len(tokens)): + w_ind = tokens[i].item() + w += source_dictionary[w_ind] + if bpe_toks is not None and w_ind in bpe_toks: + w = w[:-bpe_len] + is_bpe = True + else: + word_prob.append((w, pos_scores[i].item())) + + next_prob = None + ind = i + 1 + while ind < len(tokens): + if pos_scores[ind].item() != 0: + next_prob = pos_scores[ind] + break + ind += 1 + + word_stats.setdefault(w, WordStat(w, is_bpe)).add( + pos_scores[i].item(), next_prob + ) + is_bpe = False + w = "" + if output_word_probs: + logger.info( + str(int(sample_id)) + + " " + + ( + "\t".join( + "{} [{:2f}]".format(x[0], x[1]) for x in word_prob + ) + ) + ) + + avg_nll_loss = ( + -score_sum / count / math.log(2) if count > 0 else 0 + ) # convert to base 2 + logger.info( + "Evaluated {:,} tokens in {:.1f}s ({:.2f} tokens/s)".format( + gen_timer.n, gen_timer.sum, 1.0 / gen_timer.avg if gen_timer.avg > 0 else 0 + ) + ) + + if output_word_stats: + for ws in sorted(word_stats.values(), key=lambda x: x.count, reverse=True): + logger.info(ws) + + return { + "loss": avg_nll_loss, + "perplexity": 2 ** avg_nll_loss, + } + + +class WordStat(object): + def __init__(self, word, is_bpe): + self.word = word + self.is_bpe = is_bpe + self.log_prob = 0 + self.next_word_prob = 0 + self.count = 0 + self.missing_next_words = 0 + + def add(self, log_prob, next_word_prob): + """increments counters for the sum of log probs of current word and next + word (given context ending at current word). Since the next word might be at the end of the example, + or it might be not counted because it is not an ending subword unit, + also keeps track of how many of those we have seen""" + if next_word_prob is not None: + self.next_word_prob += next_word_prob + else: + self.missing_next_words += 1 + self.log_prob += log_prob + self.count += 1 + + def __str__(self): + return "{}\t{}\t{}\t{}\t{}\t{}".format( + self.word, + self.count, + self.log_prob, + self.is_bpe, + self.next_word_prob, + self.count - self.missing_next_words, + ) + + +def main(cfg: DictConfig, **unused_kwargs): + if isinstance(cfg, Namespace): + cfg = convert_namespace_to_omegaconf(cfg) + + utils.import_user_module(cfg.common) + + logger.info(cfg) + + if cfg.eval_lm.context_window > 0: + # reduce tokens per sample by the required context window size + cfg.task.tokens_per_sample -= cfg.eval_lm.context_window + + # Initialize the task using the current *cfg* + task = tasks.setup_task(cfg.task) + + # Load ensemble + logger.info("loading model(s) from {}".format(cfg.common_eval.path)) + models, model_args, task = checkpoint_utils.load_model_ensemble_and_task( + [cfg.common_eval.path], + arg_overrides=eval(cfg.common_eval.model_overrides), + suffix=cfg.checkpoint.checkpoint_suffix, + strict=(cfg.checkpoint.checkpoint_shard_count == 1), + num_shards=cfg.checkpoint.checkpoint_shard_count, + task=task, + ) + + use_fp16 = cfg.common.fp16 + use_cuda = torch.cuda.is_available() and not cfg.common.cpu + if use_cuda: + torch.cuda.set_device(cfg.distributed_training.device_id) + + # Optimize ensemble for generation and set the source and dest dicts on the model + # (required by scorer) + for model in models: + if use_fp16: + model.half() + if use_cuda and not cfg.distributed_training.pipeline_model_parallel: + model.cuda() + model.prepare_for_inference_(cfg) + + assert len(models) > 0 + + logger.info( + "num. model params: {:,}".format(sum(p.numel() for p in models[0].parameters())) + ) + + # Load dataset splits + task.load_dataset(cfg.dataset.gen_subset) + dataset = task.dataset(cfg.dataset.gen_subset) + logger.info( + "{} {} {:,} examples".format( + cfg.task.data, cfg.dataset.gen_subset, len(dataset) + ) + ) + + itr = task.eval_lm_dataloader( + dataset=dataset, + max_tokens=cfg.dataset.max_tokens or 36000, + batch_size=cfg.dataset.batch_size, + max_positions=utils.resolve_max_positions( + *[model.max_positions() for model in models] + ), + num_shards=max( + cfg.dataset.num_shards, + cfg.distributed_training.distributed_world_size, + ), + shard_id=max( + cfg.dataset.shard_id, + cfg.distributed_training.distributed_rank, + ), + num_workers=cfg.dataset.num_workers, + data_buffer_size=cfg.dataset.data_buffer_size, + context_window=cfg.eval_lm.context_window, + ) + + itr = progress_bar.progress_bar( + itr, + log_format=cfg.common.log_format, + log_interval=cfg.common.log_interval, + default_log_format=("tqdm" if not cfg.common.no_progress_bar else "simple"), + ) + + results = eval_lm( + models=models, + source_dictionary=task.source_dictionary, + batch_iterator=itr, + post_process=cfg.common_eval.post_process, + output_word_probs=cfg.eval_lm.output_word_probs, + output_word_stats=cfg.eval_lm.output_word_stats, + target_dictionary=task.target_dictionary, + softmax_batch=cfg.eval_lm.softmax_batch, + remove_bos_token=getattr(cfg.task, "add_bos_token", False), + ) + + logger.info( + "Loss (base 2): {:.4f}, Perplexity: {:.2f}".format( + results["loss"], results["perplexity"] + ) + ) + + return results + + +def cli_main(): + parser = options.get_eval_lm_parser() + args = options.parse_args_and_arch(parser) + + distributed_utils.call_main(convert_namespace_to_omegaconf(args), main) + + +if __name__ == "__main__": + cli_main() diff --git a/fairseq_cli/generate.py b/fairseq_cli/generate.py new file mode 100644 index 0000000000000000000000000000000000000000..7bd582b25670a937921755ad98af12b107ad9fc7 --- /dev/null +++ b/fairseq_cli/generate.py @@ -0,0 +1,408 @@ +#!/usr/bin/env python3 -u +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +Translate pre-processed data with a trained model. +""" + +import ast +import logging +import math +import os +import sys +from argparse import Namespace +from itertools import chain + +import numpy as np +import torch +from fairseq import checkpoint_utils, options, scoring, tasks, utils +from fairseq.dataclass.utils import convert_namespace_to_omegaconf +from fairseq.logging import progress_bar +from fairseq.logging.meters import StopwatchMeter, TimeMeter +from omegaconf import DictConfig + + +def main(cfg: DictConfig): + + if isinstance(cfg, Namespace): + cfg = convert_namespace_to_omegaconf(cfg) + + assert cfg.common_eval.path is not None, "--path required for generation!" + assert ( + not cfg.generation.sampling or cfg.generation.nbest == cfg.generation.beam + ), "--sampling requires --nbest to be equal to --beam" + assert ( + cfg.generation.replace_unk is None or cfg.dataset.dataset_impl == "raw" + ), "--replace-unk requires a raw text dataset (--dataset-impl=raw)" + + if cfg.common_eval.results_path is not None: + os.makedirs(cfg.common_eval.results_path, exist_ok=True) + output_path = os.path.join( + cfg.common_eval.results_path, + "generate-{}.txt".format(cfg.dataset.gen_subset), + ) + with open(output_path, "w", buffering=1, encoding="utf-8") as h: + return _main(cfg, h) + else: + return _main(cfg, sys.stdout) + + +def get_symbols_to_strip_from_output(generator): + if hasattr(generator, "symbols_to_strip_from_output"): + return generator.symbols_to_strip_from_output + else: + return {generator.eos} + + +def _main(cfg: DictConfig, output_file): + logging.basicConfig( + format="%(asctime)s | %(levelname)s | %(name)s | %(message)s", + datefmt="%Y-%m-%d %H:%M:%S", + level=os.environ.get("LOGLEVEL", "INFO").upper(), + stream=output_file, + ) + logger = logging.getLogger("fairseq_cli.generate") + + utils.import_user_module(cfg.common) + + if cfg.dataset.max_tokens is None and cfg.dataset.batch_size is None: + cfg.dataset.max_tokens = 12000 + logger.info(cfg) + + # Fix seed for stochastic decoding + if cfg.common.seed is not None and not cfg.generation.no_seed_provided: + np.random.seed(cfg.common.seed) + utils.set_torch_seed(cfg.common.seed) + + use_cuda = torch.cuda.is_available() and not cfg.common.cpu + + # Load dataset splits + task = tasks.setup_task(cfg.task) + + + # Set dictionaries + try: + src_dict = getattr(task, "source_dictionary", None) + except NotImplementedError: + src_dict = None + tgt_dict = task.target_dictionary + + overrides = ast.literal_eval(cfg.common_eval.model_overrides) + + # Load ensemble + logger.info("loading model(s) from {}".format(cfg.common_eval.path)) + models, saved_cfg = checkpoint_utils.load_model_ensemble( + utils.split_paths(cfg.common_eval.path), + arg_overrides=overrides, + task=task, + suffix=cfg.checkpoint.checkpoint_suffix, + strict=(cfg.checkpoint.checkpoint_shard_count == 1), + num_shards=cfg.checkpoint.checkpoint_shard_count, + ) + + # loading the dataset should happen after the checkpoint has been loaded so we can give it the saved task config + task.load_dataset(cfg.dataset.gen_subset, task_cfg=saved_cfg.task) + + if cfg.generation.lm_path is not None: + overrides["data"] = cfg.task.data + + try: + lms, _ = checkpoint_utils.load_model_ensemble( + [cfg.generation.lm_path], arg_overrides=overrides, task=None + ) + except: + logger.warning( + f"Failed to load language model! Please make sure that the language model dict is the same " + f"as target dict and is located in the data dir ({cfg.task.data})" + ) + raise + + assert len(lms) == 1 + else: + lms = [None] + + # Optimize ensemble for generation + for model in chain(models, lms): + if model is None: + continue + if cfg.common.fp16: + model.half() + if use_cuda and not cfg.distributed_training.pipeline_model_parallel: + model.cuda() + model.prepare_for_inference_(cfg) + + # Load alignment dictionary for unknown word replacement + # (None if no unknown word replacement, empty if no path to align dictionary) + align_dict = utils.load_align_dict(cfg.generation.replace_unk) + + # Load dataset (possibly sharded) + itr = task.get_batch_iterator( + dataset=task.dataset(cfg.dataset.gen_subset), + max_tokens=cfg.dataset.max_tokens, + max_sentences=cfg.dataset.batch_size, + max_positions=utils.resolve_max_positions( + task.max_positions(), *[m.max_positions() for m in models] + ), + ignore_invalid_inputs=cfg.dataset.skip_invalid_size_inputs_valid_test, + required_batch_size_multiple=cfg.dataset.required_batch_size_multiple, + seed=cfg.common.seed, + num_shards=cfg.distributed_training.distributed_world_size, + shard_id=cfg.distributed_training.distributed_rank, + num_workers=cfg.dataset.num_workers, + data_buffer_size=cfg.dataset.data_buffer_size, + ).next_epoch_itr(shuffle=False) + progress = progress_bar.progress_bar( + itr, + log_format=cfg.common.log_format, + log_interval=cfg.common.log_interval, + default_log_format=("tqdm" if not cfg.common.no_progress_bar else "simple"), + ) + + # Initialize generator + gen_timer = StopwatchMeter() + + extra_gen_cls_kwargs = {"lm_model": lms[0], "lm_weight": cfg.generation.lm_weight} + generator = task.build_generator( + models, cfg.generation, extra_gen_cls_kwargs=extra_gen_cls_kwargs + ) + + # Handle tokenization and BPE + tokenizer = task.build_tokenizer(cfg.tokenizer) + bpe = task.build_bpe(cfg.bpe) + + def decode_fn(x): + if bpe is not None: + x = bpe.decode(x) + if tokenizer is not None: + x = tokenizer.decode(x) + return x + + scorer = scoring.build_scorer(cfg.scoring, tgt_dict) + + num_sentences = 0 + has_target = True + wps_meter = TimeMeter() + for sample in progress: + sample = utils.move_to_cuda(sample) if use_cuda else sample + if "net_input" not in sample: + continue + + prefix_tokens = None + if cfg.generation.prefix_size > 0: + prefix_tokens = sample["target"][:, : cfg.generation.prefix_size] + + constraints = None + if "constraints" in sample: + constraints = sample["constraints"] + + gen_timer.start() + hypos = task.inference_step( + generator, + models, + sample, + prefix_tokens=prefix_tokens, + constraints=constraints, + ) + num_generated_tokens = sum(len(h[0]["tokens"]) for h in hypos) + gen_timer.stop(num_generated_tokens) + + for i, sample_id in enumerate(sample["id"].tolist()): + has_target = sample["target"] is not None + + # Remove padding + if "src_tokens" in sample["net_input"]: + src_tokens = utils.strip_pad( + sample["net_input"]["src_tokens"][i, :], tgt_dict.pad() + ) + else: + src_tokens = None + + target_tokens = None + if has_target: + target_tokens = ( + utils.strip_pad(sample["target"][i, :], tgt_dict.pad()).int().cpu() + ) + + # Either retrieve the original sentences or regenerate them from tokens. + if align_dict is not None: + src_str = task.dataset(cfg.dataset.gen_subset).src.get_original_text( + sample_id + ) + target_str = task.dataset(cfg.dataset.gen_subset).tgt.get_original_text( + sample_id + ) + else: + if src_dict is not None: + src_str = src_dict.string(src_tokens, cfg.common_eval.post_process) + else: + src_str = "" + if has_target: + target_str = tgt_dict.string( + target_tokens, + cfg.common_eval.post_process, + escape_unk=True, + extra_symbols_to_ignore=get_symbols_to_strip_from_output( + generator + ), + ) + + src_str = decode_fn(src_str) + if has_target: + target_str = decode_fn(target_str) + + if not cfg.common_eval.quiet: + if src_dict is not None: + print("S-{}\t{}".format(sample_id, src_str), file=output_file) + if has_target: + print("T-{}\t{}".format(sample_id, target_str), file=output_file) + + # Process top predictions + for j, hypo in enumerate(hypos[i][: cfg.generation.nbest]): + hypo_tokens, hypo_str, alignment = utils.post_process_prediction( + hypo_tokens=hypo["tokens"].int().cpu(), + src_str=src_str, + alignment=hypo["alignment"], + align_dict=align_dict, + tgt_dict=tgt_dict, + remove_bpe=cfg.common_eval.post_process, + extra_symbols_to_ignore=get_symbols_to_strip_from_output(generator), + ) + detok_hypo_str = decode_fn(hypo_str) + if not cfg.common_eval.quiet: + score = hypo["score"] / math.log(2) # convert to base 2 + # original hypothesis (after tokenization and BPE) + print( + "H-{}\t{}\t{}".format(sample_id, score, hypo_str), + file=output_file, + ) + # detokenized hypothesis + print( + "D-{}\t{}\t{}".format(sample_id, score, detok_hypo_str), + file=output_file, + ) + print( + "P-{}\t{}".format( + sample_id, + " ".join( + map( + lambda x: "{:.4f}".format(x), + # convert from base e to base 2 + hypo["positional_scores"] + .div_(math.log(2)) + .tolist(), + ) + ), + ), + file=output_file, + ) + + if cfg.generation.print_alignment == "hard": + print( + "A-{}\t{}".format( + sample_id, + " ".join( + [ + "{}-{}".format(src_idx, tgt_idx) + for src_idx, tgt_idx in alignment + ] + ), + ), + file=output_file, + ) + if cfg.generation.print_alignment == "soft": + print( + "A-{}\t{}".format( + sample_id, + " ".join( + [ + ",".join(src_probs) + for src_probs in alignment + ] + ), + ), + file=output_file, + ) + + if cfg.generation.print_step: + print( + "I-{}\t{}".format(sample_id, hypo["steps"]), + file=output_file, + ) + + if cfg.generation.retain_iter_history: + for step, h in enumerate(hypo["history"]): + _, h_str, _ = utils.post_process_prediction( + hypo_tokens=h["tokens"].int().cpu(), + src_str=src_str, + alignment=None, + align_dict=None, + tgt_dict=tgt_dict, + remove_bpe=None, + ) + print( + "E-{}_{}\t{}".format(sample_id, step, h_str), + file=output_file, + ) + + # Score only the top hypothesis + if has_target and j == 0: + if align_dict is not None or cfg.common_eval.post_process is not None: + # Convert back to tokens for evaluation with unk replacement and/or without BPE + target_tokens = tgt_dict.encode_line( + target_str, add_if_not_exist=True + ) + hypo_tokens = tgt_dict.encode_line( + detok_hypo_str, add_if_not_exist=True + ) + if hasattr(scorer, "add_string"): + scorer.add_string(target_str, detok_hypo_str) + else: + scorer.add(target_tokens, hypo_tokens) + + wps_meter.update(num_generated_tokens) + progress.log({"wps": round(wps_meter.avg)}) + num_sentences += ( + sample["nsentences"] if "nsentences" in sample else sample["id"].numel() + ) + + logger.info("NOTE: hypothesis and token scores are output in base 2") + logger.info( + "Translated {:,} sentences ({:,} tokens) in {:.1f}s ({:.2f} sentences/s, {:.2f} tokens/s)".format( + num_sentences, + gen_timer.n, + gen_timer.sum, + num_sentences / gen_timer.sum, + 1.0 / gen_timer.avg, + ) + ) + if has_target: + if cfg.bpe and not cfg.generation.sacrebleu: + if cfg.common_eval.post_process: + logger.warning( + "BLEU score is being computed by splitting detokenized string on spaces, this is probably not what you want. Use --sacrebleu for standard 13a BLEU tokenization" + ) + else: + logger.warning( + "If you are using BPE on the target side, the BLEU score is computed on BPE tokens, not on proper words. Use --sacrebleu for standard 13a BLEU tokenization" + ) + # use print to be consistent with other main outputs: S-, H-, T-, D- and so on + print( + "Generate {} with beam={}: {}".format( + cfg.dataset.gen_subset, cfg.generation.beam, scorer.result_string() + ), + file=output_file, + ) + + return scorer + + +def cli_main(): + parser = options.get_generation_parser() + args = options.parse_args_and_arch(parser) + main(args) + + +if __name__ == "__main__": + cli_main() diff --git a/fairseq_cli/hydra_train.py b/fairseq_cli/hydra_train.py new file mode 100644 index 0000000000000000000000000000000000000000..9de01084ba01a77a2297a3eace652a9be9b50380 --- /dev/null +++ b/fairseq_cli/hydra_train.py @@ -0,0 +1,80 @@ +#!/usr/bin/env python3 -u +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import os + +from fairseq.dataclass.initialize import add_defaults, hydra_init +from fairseq_cli.train import main as pre_main +from fairseq import distributed_utils, metrics +from fairseq.dataclass.configs import FairseqConfig +from fairseq.utils import reset_logging + +import hydra +from hydra.core.hydra_config import HydraConfig +import torch +from omegaconf import OmegaConf, open_dict + + +logger = logging.getLogger("fairseq_cli.hydra_train") + + +@hydra.main(config_path=os.path.join("..", "fairseq", "config"), config_name="config") +def hydra_main(cfg: FairseqConfig) -> float: + add_defaults(cfg) + + if cfg.common.reset_logging: + reset_logging() # Hydra hijacks logging, fix that + else: + with open_dict(cfg): + # make hydra logging work with ddp (see # see https://github.com/facebookresearch/hydra/issues/1126) + cfg.job_logging_cfg = OmegaConf.to_container(HydraConfig.get().job_logging, resolve=True) + + cfg = OmegaConf.create(OmegaConf.to_container(cfg, resolve=True, enum_to_str=True)) + OmegaConf.set_struct(cfg, True) + + try: + if cfg.common.profile: + with torch.cuda.profiler.profile(): + with torch.autograd.profiler.emit_nvtx(): + distributed_utils.call_main(cfg, pre_main) + else: + distributed_utils.call_main(cfg, pre_main) + except BaseException as e: + if not cfg.common.suppress_crashes: + raise + else: + logger.error("Crashed! " + str(e)) + + # get best val and return - useful for sweepers + try: + best_val = metrics.get_smoothed_value( + "valid", cfg.checkpoint.best_checkpoint_metric + ) + except: + best_val = None + + if best_val is None: + best_val = float("inf") + + return best_val + + +def cli_main(): + try: + from hydra._internal.utils import get_args + + cfg_name = get_args().config_name or "config" + except: + logger.warning("Failed to get config name from hydra args") + cfg_name = "config" + + hydra_init(cfg_name) + hydra_main() + + +if __name__ == "__main__": + cli_main() diff --git a/fairseq_cli/interactive.py b/fairseq_cli/interactive.py new file mode 100644 index 0000000000000000000000000000000000000000..cadef2821a74a3b2f051c792d835129bf775714f --- /dev/null +++ b/fairseq_cli/interactive.py @@ -0,0 +1,316 @@ +#!/usr/bin/env python3 -u +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +Translate raw text with a trained model. Batches data on-the-fly. +""" + +import ast +import fileinput +import logging +import math +import os +import sys +import time +from argparse import Namespace +from collections import namedtuple + +import numpy as np +import torch +from fairseq import checkpoint_utils, distributed_utils, options, tasks, utils +from fairseq.dataclass.configs import FairseqConfig +from fairseq.dataclass.utils import convert_namespace_to_omegaconf +from fairseq.token_generation_constraints import pack_constraints, unpack_constraints +from fairseq_cli.generate import get_symbols_to_strip_from_output + + +logging.basicConfig( + format="%(asctime)s | %(levelname)s | %(name)s | %(message)s", + datefmt="%Y-%m-%d %H:%M:%S", + level=os.environ.get("LOGLEVEL", "INFO").upper(), + stream=sys.stdout, +) +logger = logging.getLogger("fairseq_cli.interactive") + + +Batch = namedtuple("Batch", "ids src_tokens src_lengths constraints") +Translation = namedtuple("Translation", "src_str hypos pos_scores alignments") + + +def buffered_read(input, buffer_size): + buffer = [] + with fileinput.input(files=[input], openhook=fileinput.hook_encoded("utf-8")) as h: + for src_str in h: + buffer.append(src_str.strip()) + if len(buffer) >= buffer_size: + yield buffer + buffer = [] + + if len(buffer) > 0: + yield buffer + + +def make_batches(lines, cfg, task, max_positions, encode_fn): + def encode_fn_target(x): + return encode_fn(x) + + if cfg.generation.constraints: + # Strip (tab-delimited) contraints, if present, from input lines, + # store them in batch_constraints + batch_constraints = [list() for _ in lines] + for i, line in enumerate(lines): + if "\t" in line: + lines[i], *batch_constraints[i] = line.split("\t") + + # Convert each List[str] to List[Tensor] + for i, constraint_list in enumerate(batch_constraints): + batch_constraints[i] = [ + task.target_dictionary.encode_line( + encode_fn_target(constraint), + append_eos=False, + add_if_not_exist=False, + ) + for constraint in constraint_list + ] + + if cfg.generation.constraints: + constraints_tensor = pack_constraints(batch_constraints) + else: + constraints_tensor = None + + tokens, lengths = task.get_interactive_tokens_and_lengths(lines, encode_fn) + + itr = task.get_batch_iterator( + dataset=task.build_dataset_for_inference( + tokens, lengths, constraints=constraints_tensor + ), + max_tokens=cfg.dataset.max_tokens, + max_sentences=cfg.dataset.batch_size, + max_positions=max_positions, + ignore_invalid_inputs=cfg.dataset.skip_invalid_size_inputs_valid_test, + ).next_epoch_itr(shuffle=False) + for batch in itr: + ids = batch["id"] + src_tokens = batch["net_input"]["src_tokens"] + src_lengths = batch["net_input"]["src_lengths"] + constraints = batch.get("constraints", None) + + yield Batch( + ids=ids, + src_tokens=src_tokens, + src_lengths=src_lengths, + constraints=constraints, + ) + + +def main(cfg: FairseqConfig): + if isinstance(cfg, Namespace): + cfg = convert_namespace_to_omegaconf(cfg) + + start_time = time.time() + total_translate_time = 0 + + utils.import_user_module(cfg.common) + + if cfg.interactive.buffer_size < 1: + cfg.interactive.buffer_size = 1 + if cfg.dataset.max_tokens is None and cfg.dataset.batch_size is None: + cfg.dataset.batch_size = 1 + + assert ( + not cfg.generation.sampling or cfg.generation.nbest == cfg.generation.beam + ), "--sampling requires --nbest to be equal to --beam" + assert ( + not cfg.dataset.batch_size + or cfg.dataset.batch_size <= cfg.interactive.buffer_size + ), "--batch-size cannot be larger than --buffer-size" + + logger.info(cfg) + + # Fix seed for stochastic decoding + if cfg.common.seed is not None and not cfg.generation.no_seed_provided: + np.random.seed(cfg.common.seed) + utils.set_torch_seed(cfg.common.seed) + + use_cuda = torch.cuda.is_available() and not cfg.common.cpu + + # Setup task, e.g., translation + task = tasks.setup_task(cfg.task) + + # Load ensemble + overrides = ast.literal_eval(cfg.common_eval.model_overrides) + logger.info("loading model(s) from {}".format(cfg.common_eval.path)) + models, _model_args = checkpoint_utils.load_model_ensemble( + utils.split_paths(cfg.common_eval.path), + arg_overrides=overrides, + task=task, + suffix=cfg.checkpoint.checkpoint_suffix, + strict=(cfg.checkpoint.checkpoint_shard_count == 1), + num_shards=cfg.checkpoint.checkpoint_shard_count, + ) + + # Set dictionaries + src_dict = task.source_dictionary + tgt_dict = task.target_dictionary + + # Optimize ensemble for generation + for model in models: + if model is None: + continue + if cfg.common.fp16: + model.half() + if use_cuda and not cfg.distributed_training.pipeline_model_parallel: + model.cuda() + model.prepare_for_inference_(cfg) + + # Initialize generator + generator = task.build_generator(models, cfg.generation) + + # Handle tokenization and BPE + tokenizer = task.build_tokenizer(cfg.tokenizer) + bpe = task.build_bpe(cfg.bpe) + + def encode_fn(x): + if tokenizer is not None: + x = tokenizer.encode(x) + if bpe is not None: + x = bpe.encode(x) + return x + + def decode_fn(x): + if bpe is not None: + x = bpe.decode(x) + if tokenizer is not None: + x = tokenizer.decode(x) + return x + + # Load alignment dictionary for unknown word replacement + # (None if no unknown word replacement, empty if no path to align dictionary) + align_dict = utils.load_align_dict(cfg.generation.replace_unk) + + max_positions = utils.resolve_max_positions( + task.max_positions(), *[model.max_positions() for model in models] + ) + + if cfg.generation.constraints: + logger.warning( + "NOTE: Constrained decoding currently assumes a shared subword vocabulary." + ) + + if cfg.interactive.buffer_size > 1: + logger.info("Sentence buffer size: %s", cfg.interactive.buffer_size) + logger.info("NOTE: hypothesis and token scores are output in base 2") + logger.info("Type the input sentence and press return:") + start_id = 0 + for inputs in buffered_read(cfg.interactive.input, cfg.interactive.buffer_size): + results = [] + for batch in make_batches(inputs, cfg, task, max_positions, encode_fn): + bsz = batch.src_tokens.size(0) + src_tokens = batch.src_tokens + src_lengths = batch.src_lengths + constraints = batch.constraints + if use_cuda: + src_tokens = src_tokens.cuda() + src_lengths = src_lengths.cuda() + if constraints is not None: + constraints = constraints.cuda() + + sample = { + "net_input": { + "src_tokens": src_tokens, + "src_lengths": src_lengths, + }, + } + translate_start_time = time.time() + translations = task.inference_step( + generator, models, sample, constraints=constraints + ) + translate_time = time.time() - translate_start_time + total_translate_time += translate_time + list_constraints = [[] for _ in range(bsz)] + if cfg.generation.constraints: + list_constraints = [unpack_constraints(c) for c in constraints] + for i, (id, hypos) in enumerate(zip(batch.ids.tolist(), translations)): + src_tokens_i = utils.strip_pad(src_tokens[i], tgt_dict.pad()) + constraints = list_constraints[i] + results.append( + ( + start_id + id, + src_tokens_i, + hypos, + { + "constraints": constraints, + "time": translate_time / len(translations), + }, + ) + ) + + # sort output to match input order + for id_, src_tokens, hypos, info in sorted(results, key=lambda x: x[0]): + src_str = '' + if src_dict is not None: + src_str = src_dict.string(src_tokens, cfg.common_eval.post_process) + print("S-{}\t{}".format(id_, src_str)) + print("W-{}\t{:.3f}\tseconds".format(id_, info["time"])) + for constraint in info["constraints"]: + print( + "C-{}\t{}".format( + id_, tgt_dict.string(constraint, cfg.common_eval.post_process) + ) + ) + + # Process top predictions + for hypo in hypos[: min(len(hypos), cfg.generation.nbest)]: + hypo_tokens, hypo_str, alignment = utils.post_process_prediction( + hypo_tokens=hypo["tokens"].int().cpu(), + src_str=src_str, + alignment=hypo["alignment"], + align_dict=align_dict, + tgt_dict=tgt_dict, + remove_bpe=cfg.common_eval.post_process, + extra_symbols_to_ignore=get_symbols_to_strip_from_output(generator), + ) + detok_hypo_str = decode_fn(hypo_str) + score = hypo["score"] / math.log(2) # convert to base 2 + # original hypothesis (after tokenization and BPE) + print("H-{}\t{}\t{}".format(id_, score, hypo_str)) + # detokenized hypothesis + print("D-{}\t{}\t{}".format(id_, score, detok_hypo_str)) + print( + "P-{}\t{}".format( + id_, + " ".join( + map( + lambda x: "{:.4f}".format(x), + # convert from base e to base 2 + hypo["positional_scores"].div_(math.log(2)).tolist(), + ) + ), + ) + ) + if cfg.generation.print_alignment: + alignment_str = " ".join( + ["{}-{}".format(src, tgt) for src, tgt in alignment] + ) + print("A-{}\t{}".format(id_, alignment_str)) + + # update running id_ counter + start_id += len(inputs) + + logger.info( + "Total time: {:.3f} seconds; translation time: {:.3f}".format( + time.time() - start_time, total_translate_time + ) + ) + + +def cli_main(): + parser = options.get_interactive_generation_parser() + args = options.parse_args_and_arch(parser) + distributed_utils.call_main(convert_namespace_to_omegaconf(args), main) + + +if __name__ == "__main__": + cli_main() diff --git a/fairseq_cli/preprocess.py b/fairseq_cli/preprocess.py new file mode 100644 index 0000000000000000000000000000000000000000..b788900d30af6f0bcc1a3e807a8bb249e70b7a43 --- /dev/null +++ b/fairseq_cli/preprocess.py @@ -0,0 +1,401 @@ +#!/usr/bin/env python3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +Data pre-processing: build vocabularies and binarize training data. +""" + +import logging +import os +import shutil +import sys +from collections import Counter +from itertools import zip_longest +from multiprocessing import Pool + +from fairseq import options, tasks, utils +from fairseq.binarizer import Binarizer +from fairseq.data import indexed_dataset + + +logging.basicConfig( + format="%(asctime)s | %(levelname)s | %(name)s | %(message)s", + datefmt="%Y-%m-%d %H:%M:%S", + level=os.environ.get("LOGLEVEL", "INFO").upper(), + stream=sys.stdout, +) +logger = logging.getLogger("fairseq_cli.preprocess") + + +def main(args): + utils.import_user_module(args) + + os.makedirs(args.destdir, exist_ok=True) + + logger.addHandler( + logging.FileHandler( + filename=os.path.join(args.destdir, "preprocess.log"), + ) + ) + logger.info(args) + + task = tasks.get_task(args.task) + + def train_path(lang): + return "{}{}".format(args.trainpref, ("." + lang) if lang else "") + + def file_name(prefix, lang): + fname = prefix + if lang is not None: + fname += ".{lang}".format(lang=lang) + return fname + + def dest_path(prefix, lang): + return os.path.join(args.destdir, file_name(prefix, lang)) + + def dict_path(lang): + return dest_path("dict", lang) + ".txt" + + def build_dictionary(filenames, src=False, tgt=False): + assert src ^ tgt + return task.build_dictionary( + filenames, + workers=args.workers, + threshold=args.thresholdsrc if src else args.thresholdtgt, + nwords=args.nwordssrc if src else args.nwordstgt, + padding_factor=args.padding_factor, + ) + + target = not args.only_source + + if not args.srcdict and os.path.exists(dict_path(args.source_lang)): + raise FileExistsError(dict_path(args.source_lang)) + if target and not args.tgtdict and os.path.exists(dict_path(args.target_lang)): + raise FileExistsError(dict_path(args.target_lang)) + + if args.joined_dictionary: + assert ( + not args.srcdict or not args.tgtdict + ), "cannot use both --srcdict and --tgtdict with --joined-dictionary" + + if args.srcdict: + src_dict = task.load_dictionary(args.srcdict) + elif args.tgtdict: + src_dict = task.load_dictionary(args.tgtdict) + else: + assert ( + args.trainpref + ), "--trainpref must be set if --srcdict is not specified" + src_dict = build_dictionary( + {train_path(lang) for lang in [args.source_lang, args.target_lang]}, + src=True, + ) + tgt_dict = src_dict + else: + if args.srcdict: + src_dict = task.load_dictionary(args.srcdict) + else: + assert ( + args.trainpref + ), "--trainpref must be set if --srcdict is not specified" + src_dict = build_dictionary([train_path(args.source_lang)], src=True) + + if target: + if args.tgtdict: + tgt_dict = task.load_dictionary(args.tgtdict) + else: + assert ( + args.trainpref + ), "--trainpref must be set if --tgtdict is not specified" + tgt_dict = build_dictionary([train_path(args.target_lang)], tgt=True) + else: + tgt_dict = None + + src_dict.save(dict_path(args.source_lang)) + if target and tgt_dict is not None: + tgt_dict.save(dict_path(args.target_lang)) + + if args.dict_only: + return + + def make_binary_dataset(vocab, input_prefix, output_prefix, lang, num_workers): + logger.info("[{}] Dictionary: {} types".format(lang, len(vocab))) + n_seq_tok = [0, 0] + replaced = Counter() + + def merge_result(worker_result): + replaced.update(worker_result["replaced"]) + n_seq_tok[0] += worker_result["nseq"] + n_seq_tok[1] += worker_result["ntok"] + + input_file = "{}{}".format( + input_prefix, ("." + lang) if lang is not None else "" + ) + offsets = Binarizer.find_offsets(input_file, num_workers) + pool = None + if num_workers > 1: + pool = Pool(processes=num_workers - 1) + for worker_id in range(1, num_workers): + prefix = "{}{}".format(output_prefix, worker_id) + pool.apply_async( + binarize, + ( + args, + input_file, + vocab, + prefix, + lang, + offsets[worker_id], + offsets[worker_id + 1], + ), + callback=merge_result, + ) + pool.close() + + ds = indexed_dataset.make_builder( + dataset_dest_file(args, output_prefix, lang, "bin"), + impl=args.dataset_impl, + vocab_size=len(vocab), + ) + merge_result( + Binarizer.binarize( + input_file, vocab, lambda t: ds.add_item(t), offset=0, end=offsets[1] + ) + ) + if num_workers > 1: + pool.join() + for worker_id in range(1, num_workers): + prefix = "{}{}".format(output_prefix, worker_id) + temp_file_path = dataset_dest_prefix(args, prefix, lang) + ds.merge_file_(temp_file_path) + os.remove(indexed_dataset.data_file_path(temp_file_path)) + os.remove(indexed_dataset.index_file_path(temp_file_path)) + + ds.finalize(dataset_dest_file(args, output_prefix, lang, "idx")) + + logger.info( + "[{}] {}: {} sents, {} tokens, {:.3}% replaced by {}".format( + lang, + input_file, + n_seq_tok[0], + n_seq_tok[1], + 100 * sum(replaced.values()) / n_seq_tok[1], + vocab.unk_word, + ) + ) + + def make_binary_alignment_dataset(input_prefix, output_prefix, num_workers): + nseq = [0] + + def merge_result(worker_result): + nseq[0] += worker_result["nseq"] + + input_file = input_prefix + offsets = Binarizer.find_offsets(input_file, num_workers) + pool = None + if num_workers > 1: + pool = Pool(processes=num_workers - 1) + for worker_id in range(1, num_workers): + prefix = "{}{}".format(output_prefix, worker_id) + pool.apply_async( + binarize_alignments, + ( + args, + input_file, + utils.parse_alignment, + prefix, + offsets[worker_id], + offsets[worker_id + 1], + ), + callback=merge_result, + ) + pool.close() + + ds = indexed_dataset.make_builder( + dataset_dest_file(args, output_prefix, None, "bin"), impl=args.dataset_impl + ) + + merge_result( + Binarizer.binarize_alignments( + input_file, + utils.parse_alignment, + lambda t: ds.add_item(t), + offset=0, + end=offsets[1], + ) + ) + if num_workers > 1: + pool.join() + for worker_id in range(1, num_workers): + prefix = "{}{}".format(output_prefix, worker_id) + temp_file_path = dataset_dest_prefix(args, prefix, None) + ds.merge_file_(temp_file_path) + os.remove(indexed_dataset.data_file_path(temp_file_path)) + os.remove(indexed_dataset.index_file_path(temp_file_path)) + + ds.finalize(dataset_dest_file(args, output_prefix, None, "idx")) + + logger.info("[alignments] {}: parsed {} alignments".format(input_file, nseq[0])) + + def make_dataset(vocab, input_prefix, output_prefix, lang, num_workers=1): + if args.dataset_impl == "raw": + # Copy original text file to destination folder + output_text_file = dest_path( + output_prefix + ".{}-{}".format(args.source_lang, args.target_lang), + lang, + ) + shutil.copyfile(file_name(input_prefix, lang), output_text_file) + else: + make_binary_dataset(vocab, input_prefix, output_prefix, lang, num_workers) + + def make_all(lang, vocab): + if args.trainpref: + make_dataset(vocab, args.trainpref, "train", lang, num_workers=args.workers) + if args.validpref: + for k, validpref in enumerate(args.validpref.split(",")): + outprefix = "valid{}".format(k) if k > 0 else "valid" + make_dataset( + vocab, validpref, outprefix, lang, num_workers=args.workers + ) + if args.testpref: + for k, testpref in enumerate(args.testpref.split(",")): + outprefix = "test{}".format(k) if k > 0 else "test" + make_dataset(vocab, testpref, outprefix, lang, num_workers=args.workers) + + def make_all_alignments(): + if args.trainpref and os.path.exists(args.trainpref + "." + args.align_suffix): + make_binary_alignment_dataset( + args.trainpref + "." + args.align_suffix, + "train.align", + num_workers=args.workers, + ) + if args.validpref and os.path.exists(args.validpref + "." + args.align_suffix): + make_binary_alignment_dataset( + args.validpref + "." + args.align_suffix, + "valid.align", + num_workers=args.workers, + ) + if args.testpref and os.path.exists(args.testpref + "." + args.align_suffix): + make_binary_alignment_dataset( + args.testpref + "." + args.align_suffix, + "test.align", + num_workers=args.workers, + ) + + make_all(args.source_lang, src_dict) + if target: + make_all(args.target_lang, tgt_dict) + if args.align_suffix: + make_all_alignments() + + logger.info("Wrote preprocessed data to {}".format(args.destdir)) + + if args.alignfile: + assert args.trainpref, "--trainpref must be set if --alignfile is specified" + src_file_name = train_path(args.source_lang) + tgt_file_name = train_path(args.target_lang) + freq_map = {} + with open(args.alignfile, "r", encoding="utf-8") as align_file: + with open(src_file_name, "r", encoding="utf-8") as src_file: + with open(tgt_file_name, "r", encoding="utf-8") as tgt_file: + for a, s, t in zip_longest(align_file, src_file, tgt_file): + si = src_dict.encode_line(s, add_if_not_exist=False) + ti = tgt_dict.encode_line(t, add_if_not_exist=False) + ai = list(map(lambda x: tuple(x.split("-")), a.split())) + for sai, tai in ai: + srcidx = si[int(sai)] + tgtidx = ti[int(tai)] + if srcidx != src_dict.unk() and tgtidx != tgt_dict.unk(): + assert srcidx != src_dict.pad() + assert srcidx != src_dict.eos() + assert tgtidx != tgt_dict.pad() + assert tgtidx != tgt_dict.eos() + + if srcidx not in freq_map: + freq_map[srcidx] = {} + if tgtidx not in freq_map[srcidx]: + freq_map[srcidx][tgtidx] = 1 + else: + freq_map[srcidx][tgtidx] += 1 + + align_dict = {} + for srcidx in freq_map.keys(): + align_dict[srcidx] = max(freq_map[srcidx], key=freq_map[srcidx].get) + + with open( + os.path.join( + args.destdir, + "alignment.{}-{}.txt".format(args.source_lang, args.target_lang), + ), + "w", + encoding="utf-8", + ) as f: + for k, v in align_dict.items(): + print("{} {}".format(src_dict[k], tgt_dict[v]), file=f) + + +def binarize(args, filename, vocab, output_prefix, lang, offset, end, append_eos=True): + ds = indexed_dataset.make_builder( + dataset_dest_file(args, output_prefix, lang, "bin"), + impl=args.dataset_impl, + vocab_size=len(vocab), + ) + + def consumer(tensor): + ds.add_item(tensor) + + res = Binarizer.binarize( + filename, vocab, consumer, append_eos=append_eos, offset=offset, end=end + ) + ds.finalize(dataset_dest_file(args, output_prefix, lang, "idx")) + return res + + +def binarize_alignments(args, filename, parse_alignment, output_prefix, offset, end): + ds = indexed_dataset.make_builder( + dataset_dest_file(args, output_prefix, None, "bin"), + impl=args.dataset_impl, + vocab_size=None, + ) + + def consumer(tensor): + ds.add_item(tensor) + + res = Binarizer.binarize_alignments( + filename, parse_alignment, consumer, offset=offset, end=end + ) + ds.finalize(dataset_dest_file(args, output_prefix, None, "idx")) + return res + + +def dataset_dest_prefix(args, output_prefix, lang): + base = "{}/{}".format(args.destdir, output_prefix) + if lang is not None: + lang_part = ".{}-{}.{}".format(args.source_lang, args.target_lang, lang) + elif args.only_source: + lang_part = "" + else: + lang_part = ".{}-{}".format(args.source_lang, args.target_lang) + + return "{}{}".format(base, lang_part) + + +def dataset_dest_file(args, output_prefix, lang, extension): + base = dataset_dest_prefix(args, output_prefix, lang) + return "{}.{}".format(base, extension) + + +def get_offsets(input_file, num_workers): + return Binarizer.find_offsets(input_file, num_workers) + + +def cli_main(): + parser = options.get_preprocessing_parser() + args = parser.parse_args() + main(args) + + +if __name__ == "__main__": + cli_main() diff --git a/fairseq_cli/score.py b/fairseq_cli/score.py new file mode 100644 index 0000000000000000000000000000000000000000..0b207be959d55f6a56d8c5eb7db3dbe0c1ac977e --- /dev/null +++ b/fairseq_cli/score.py @@ -0,0 +1,102 @@ +#!/usr/bin/env python3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +BLEU scoring of generated translations against reference translations. +""" + +import argparse +import os +import sys + +from fairseq.data import dictionary +from fairseq.scoring import bleu + + +def get_parser(): + parser = argparse.ArgumentParser( + description="Command-line script for BLEU scoring." + ) + # fmt: off + parser.add_argument('-s', '--sys', default='-', help='system output') + parser.add_argument('-r', '--ref', required=True, help='references') + parser.add_argument('-o', '--order', default=4, metavar='N', + type=int, help='consider ngrams up to this order') + parser.add_argument('--ignore-case', action='store_true', + help='case-insensitive scoring') + parser.add_argument('--sacrebleu', action='store_true', + help='score with sacrebleu') + parser.add_argument('--sentence-bleu', action='store_true', + help='report sentence-level BLEUs (i.e., with +1 smoothing)') + # fmt: on + return parser + + +def cli_main(): + parser = get_parser() + args = parser.parse_args() + print(args) + + assert args.sys == "-" or os.path.exists( + args.sys + ), "System output file {} does not exist".format(args.sys) + assert os.path.exists(args.ref), "Reference file {} does not exist".format(args.ref) + + dict = dictionary.Dictionary() + + def readlines(fd): + for line in fd.readlines(): + if args.ignore_case: + yield line.lower() + else: + yield line + + if args.sacrebleu: + import sacrebleu + + def score(fdsys): + with open(args.ref) as fdref: + print(sacrebleu.corpus_bleu(fdsys, [fdref]).format()) + + elif args.sentence_bleu: + + def score(fdsys): + with open(args.ref) as fdref: + scorer = bleu.Scorer(dict.pad(), dict.eos(), dict.unk()) + for i, (sys_tok, ref_tok) in enumerate( + zip(readlines(fdsys), readlines(fdref)) + ): + scorer.reset(one_init=True) + sys_tok = dict.encode_line(sys_tok) + ref_tok = dict.encode_line(ref_tok) + scorer.add(ref_tok, sys_tok) + print(i, scorer.result_string(args.order)) + + else: + + def score(fdsys): + with open(args.ref) as fdref: + scorer = bleu.Scorer( + bleu.BleuConfig( + pad=dict.pad(), + eos=dict.eos(), + unk=dict.unk(), + ) + ) + for sys_tok, ref_tok in zip(readlines(fdsys), readlines(fdref)): + sys_tok = dict.encode_line(sys_tok) + ref_tok = dict.encode_line(ref_tok) + scorer.add(ref_tok, sys_tok) + print(scorer.result_string(args.order)) + + if args.sys == "-": + score(sys.stdin) + else: + with open(args.sys, "r") as f: + score(f) + + +if __name__ == "__main__": + cli_main() diff --git a/fairseq_cli/train.py b/fairseq_cli/train.py new file mode 100644 index 0000000000000000000000000000000000000000..83475873138c5d1bac288c234afb6b4a1a7882d7 --- /dev/null +++ b/fairseq_cli/train.py @@ -0,0 +1,514 @@ +#!/usr/bin/env python3 -u +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +Train a new model on one or across multiple GPUs. +""" + +import argparse +import logging +import math +import os +import sys +from typing import Dict, Optional, Any, List, Tuple, Callable + +# We need to setup root logger before importing any fairseq libraries. +logging.basicConfig( + format="%(asctime)s | %(levelname)s | %(name)s | %(message)s", + datefmt="%Y-%m-%d %H:%M:%S", + level=os.environ.get("LOGLEVEL", "INFO").upper(), + stream=sys.stdout, +) +logger = logging.getLogger("fairseq_cli.train") + +import numpy as np +import torch +from fairseq import ( + checkpoint_utils, + options, + quantization_utils, + tasks, + utils, +) +from fairseq.data import iterators, data_utils +from fairseq.data.plasma_utils import PlasmaStore +from fairseq.dataclass.configs import FairseqConfig +from fairseq.dataclass.utils import convert_namespace_to_omegaconf +from fairseq.distributed import fsdp_enable_wrap, fsdp_wrap, utils as distributed_utils +from fairseq.file_io import PathManager +from fairseq.logging import meters, metrics, progress_bar +from fairseq.model_parallel.megatron_trainer import MegatronTrainer +from fairseq.trainer import Trainer +from omegaconf import DictConfig, OmegaConf + + + + +def main(cfg: FairseqConfig) -> None: + if isinstance(cfg, argparse.Namespace): + cfg = convert_namespace_to_omegaconf(cfg) + + utils.import_user_module(cfg.common) + + if distributed_utils.is_master(cfg.distributed_training) and "job_logging_cfg" in cfg: + # make hydra logging work with ddp (see # see https://github.com/facebookresearch/hydra/issues/1126) + logging.config.dictConfig(OmegaConf.to_container(cfg.job_logging_cfg)) + + assert ( + cfg.dataset.max_tokens is not None or cfg.dataset.batch_size is not None + ), "Must specify batch size either with --max-tokens or --batch-size" + metrics.reset() + + if cfg.common.log_file is not None: + handler = logging.FileHandler(filename=cfg.common.log_file) + logger.addHandler(handler) + + np.random.seed(cfg.common.seed) + utils.set_torch_seed(cfg.common.seed) + + if distributed_utils.is_master(cfg.distributed_training): + checkpoint_utils.verify_checkpoint_directory(cfg.checkpoint.save_dir) + + # Print args + logger.info(cfg) + + if cfg.checkpoint.write_checkpoints_asynchronously: + try: + import iopath # noqa: F401 + except ImportError: + logging.exception( + "Asynchronous checkpoint writing is specified but iopath is " + "not installed: `pip install iopath`" + ) + return + + # Setup task, e.g., translation, language modeling, etc. + task = tasks.setup_task(cfg.task) + + assert cfg.criterion, "Please specify criterion to train a model" + + # Build model and criterion + if cfg.distributed_training.ddp_backend == "fully_sharded": + with fsdp_enable_wrap(cfg.distributed_training): + model = fsdp_wrap(task.build_model(cfg.model)) + else: + model = task.build_model(cfg.model) + criterion = task.build_criterion(cfg.criterion) + logger.info(model) + logger.info("task: {}".format(task.__class__.__name__)) + logger.info("model: {}".format(model.__class__.__name__)) + logger.info("criterion: {}".format(criterion.__class__.__name__)) + logger.info( + "num. shared model params: {:,} (num. trained: {:,})".format( + sum(p.numel() for p in model.parameters() if not getattr(p, "expert", False)), + sum(p.numel() for p in model.parameters() if not getattr(p, "expert", False) and p.requires_grad) + ) + ) + + logger.info( + "num. expert model params: {} (num. trained: {})".format( + sum(p.numel() for p in model.parameters() if getattr(p, "expert", False)), + sum(p.numel() for p in model.parameters() if getattr(p, "expert", False) and p.requires_grad), + ) + ) + + # Load valid dataset (we load training data below, based on the latest checkpoint) + # We load the valid dataset AFTER building the model + data_utils.raise_if_valid_subsets_unintentionally_ignored(cfg) + if cfg.dataset.combine_valid_subsets: + task.load_dataset("valid", combine=True, epoch=1) + else: + for valid_sub_split in cfg.dataset.valid_subset.split(","): + task.load_dataset(valid_sub_split, combine=False, epoch=1) + + # (optionally) Configure quantization + if cfg.common.quantization_config_path is not None: + quantizer = quantization_utils.Quantizer( + config_path=cfg.common.quantization_config_path, + max_epoch=cfg.optimization.max_epoch, + max_update=cfg.optimization.max_update, + ) + else: + quantizer = None + + # Build trainer + if cfg.common.model_parallel_size == 1: + trainer = Trainer(cfg, task, model, criterion, quantizer) + else: + trainer = MegatronTrainer(cfg, task, model, criterion) + logger.info( + "training on {} devices (GPUs/TPUs)".format( + cfg.distributed_training.distributed_world_size + ) + ) + logger.info( + "max tokens per device = {} and max sentences per device = {}".format( + cfg.dataset.max_tokens, + cfg.dataset.batch_size, + ) + ) + + # Load the latest checkpoint if one is available and restore the + # corresponding train iterator + extra_state, epoch_itr = checkpoint_utils.load_checkpoint( + cfg.checkpoint, + trainer, + # don't cache epoch iterators for sharded datasets + disable_iterator_cache=task.has_sharded_data("train"), + ) + if cfg.common.tpu: + import torch_xla.core.xla_model as xm + xm.rendezvous("load_checkpoint") # wait for all workers + + max_epoch = cfg.optimization.max_epoch or math.inf + lr = trainer.get_lr() + + train_meter = meters.StopwatchMeter() + train_meter.start() + while epoch_itr.next_epoch_idx <= max_epoch: + if lr <= cfg.optimization.stop_min_lr: + logger.info( + f"stopping training because current learning rate ({lr}) is smaller " + "than or equal to minimum learning rate " + f"(--stop-min-lr={cfg.optimization.stop_min_lr})" + ) + break + + # train for one epoch + valid_losses, should_stop = train(cfg, trainer, task, epoch_itr) + if should_stop: + break + + # only use first validation loss to update the learning rate + lr = trainer.lr_step(epoch_itr.epoch, valid_losses[0]) + + epoch_itr = trainer.get_train_iterator( + epoch_itr.next_epoch_idx, + # sharded data: get train iterator for next epoch + load_dataset=task.has_sharded_data("train"), + # don't cache epoch iterators for sharded datasets + disable_iterator_cache=task.has_sharded_data("train"), + ) + train_meter.stop() + logger.info("done training in {:.1f} seconds".format(train_meter.sum)) + + # ioPath implementation to wait for all asynchronous file writes to complete. + if cfg.checkpoint.write_checkpoints_asynchronously: + logger.info( + "ioPath PathManager waiting for all asynchronous checkpoint " + "writes to finish." + ) + PathManager.async_close() + logger.info("ioPath PathManager finished waiting.") + + +def should_stop_early(cfg: DictConfig, valid_loss: float) -> bool: + # skip check if no validation was done in the current epoch + if valid_loss is None: + return False + if cfg.checkpoint.patience <= 0: + return False + + def is_better(a, b): + return a > b if cfg.checkpoint.maximize_best_checkpoint_metric else a < b + + prev_best = getattr(should_stop_early, "best", None) + if prev_best is None or is_better(valid_loss, prev_best): + should_stop_early.best = valid_loss + should_stop_early.num_runs = 0 + return False + else: + should_stop_early.num_runs += 1 + if should_stop_early.num_runs >= cfg.checkpoint.patience: + logger.info( + "early stop since valid performance hasn't improved for last {} runs".format( + cfg.checkpoint.patience + ) + ) + return True + else: + return False + + +@metrics.aggregate("train") +def train( + cfg: DictConfig, trainer: Trainer, task: tasks.FairseqTask, epoch_itr +) -> Tuple[List[Optional[float]], bool]: + """Train the model for one epoch and return validation losses.""" + # Initialize data iterator + itr = epoch_itr.next_epoch_itr( + fix_batches_to_gpus=cfg.distributed_training.fix_batches_to_gpus, + shuffle=(epoch_itr.next_epoch_idx > cfg.dataset.curriculum), + ) + update_freq = ( + cfg.optimization.update_freq[epoch_itr.epoch - 1] + if epoch_itr.epoch <= len(cfg.optimization.update_freq) + else cfg.optimization.update_freq[-1] + ) + itr = iterators.GroupedIterator(itr, update_freq) + if cfg.common.tpu: + itr = utils.tpu_data_loader(itr) + progress = progress_bar.progress_bar( + itr, + log_format=cfg.common.log_format, + log_file=cfg.common.log_file, + log_interval=cfg.common.log_interval, + epoch=epoch_itr.epoch, + tensorboard_logdir=( + cfg.common.tensorboard_logdir + if distributed_utils.is_master(cfg.distributed_training) + else None + ), + default_log_format=("tqdm" if not cfg.common.no_progress_bar else "simple"), + wandb_project=( + cfg.common.wandb_project + if distributed_utils.is_master(cfg.distributed_training) + else None + ), + wandb_run_name=os.environ.get( + "WANDB_NAME", os.path.basename(cfg.checkpoint.save_dir) + ), + azureml_logging=( + cfg.common.azureml_logging + if distributed_utils.is_master(cfg.distributed_training) + else False + ), + ) + progress.update_config(_flatten_config(cfg)) + + trainer.begin_epoch(epoch_itr.epoch) + + valid_subsets = cfg.dataset.valid_subset.split(",") + should_stop = False + num_updates = trainer.get_num_updates() + logger.info("Start iterating over samples") + for i, samples in enumerate(progress): + with metrics.aggregate("train_inner"), torch.autograd.profiler.record_function( + "train_step-%d" % i + ): + log_output = trainer.train_step(samples) + + if log_output is not None: # not OOM, overflow, ... + # log mid-epoch stats + num_updates = trainer.get_num_updates() + if num_updates % cfg.common.log_interval == 0: + stats = get_training_stats(metrics.get_smoothed_values("train_inner")) + progress.log(stats, tag="train_inner", step=num_updates) + + # reset mid-epoch stats after each log interval + # the end-of-epoch stats will still be preserved + metrics.reset_meters("train_inner") + + end_of_epoch = not itr.has_next() + valid_losses, should_stop = validate_and_save( + cfg, trainer, task, epoch_itr, valid_subsets, end_of_epoch + ) + + if should_stop: + break + + # log end-of-epoch stats + logger.info("end of epoch {} (average epoch stats below)".format(epoch_itr.epoch)) + stats = get_training_stats(metrics.get_smoothed_values("train")) + progress.print(stats, tag="train", step=num_updates) + + # reset epoch-level meters + metrics.reset_meters("train") + return valid_losses, should_stop + + +def _flatten_config(cfg: DictConfig): + config = OmegaConf.to_container(cfg) + # remove any legacy Namespaces and replace with a single "args" + namespace = None + for k, v in list(config.items()): + if isinstance(v, argparse.Namespace): + namespace = v + del config[k] + if namespace is not None: + config["args"] = vars(namespace) + return config + + +def validate_and_save( + cfg: DictConfig, + trainer: Trainer, + task: tasks.FairseqTask, + epoch_itr, + valid_subsets: List[str], + end_of_epoch: bool, +) -> Tuple[List[Optional[float]], bool]: + num_updates = trainer.get_num_updates() + max_update = cfg.optimization.max_update or math.inf + + # Stopping conditions (and an additional one based on validation loss later + # on) + should_stop = False + if num_updates >= max_update: + should_stop = True + logger.info( + f"Stopping training due to " + f"num_updates: {num_updates} >= max_update: {max_update}" + ) + + training_time_hours = trainer.cumulative_training_time() / (60 * 60) + if ( + cfg.optimization.stop_time_hours > 0 + and training_time_hours > cfg.optimization.stop_time_hours + ): + should_stop = True + logger.info( + f"Stopping training due to " + f"cumulative_training_time: {training_time_hours} > " + f"stop_time_hours: {cfg.optimization.stop_time_hours} hour(s)" + ) + + do_save = ( + (end_of_epoch and epoch_itr.epoch % cfg.checkpoint.save_interval == 0) + or should_stop + or ( + cfg.checkpoint.save_interval_updates > 0 + and num_updates > 0 + and num_updates % cfg.checkpoint.save_interval_updates == 0 + and num_updates >= cfg.dataset.validate_after_updates + ) + ) + do_validate = ( + (not end_of_epoch and do_save) # validate during mid-epoch saves + or (end_of_epoch and epoch_itr.epoch % cfg.dataset.validate_interval == 0) + or should_stop + or ( + cfg.dataset.validate_interval_updates > 0 + and num_updates > 0 + and num_updates % cfg.dataset.validate_interval_updates == 0 + ) + ) and not cfg.dataset.disable_validation and num_updates >= cfg.dataset.validate_after_updates + + # Validate + valid_losses = [None] + if do_validate: + valid_losses = validate(cfg, trainer, task, epoch_itr, valid_subsets) + + should_stop |= should_stop_early(cfg, valid_losses[0]) + + # Save checkpoint + if do_save or should_stop: + checkpoint_utils.save_checkpoint( + cfg.checkpoint, trainer, epoch_itr, valid_losses[0] + ) + + return valid_losses, should_stop + + +def get_training_stats(stats: Dict[str, Any]) -> Dict[str, Any]: + stats["wall"] = round(metrics.get_meter("default", "wall").elapsed_time, 0) + return stats + + +def validate( + cfg: DictConfig, + trainer: Trainer, + task: tasks.FairseqTask, + epoch_itr, + subsets: List[str], +) -> List[Optional[float]]: + """Evaluate the model on the validation set(s) and return the losses.""" + + if cfg.dataset.fixed_validation_seed is not None: + # set fixed seed for every validation + utils.set_torch_seed(cfg.dataset.fixed_validation_seed) + + trainer.begin_valid_epoch(epoch_itr.epoch) + valid_losses = [] + for subset in subsets: + logger.info('begin validation on "{}" subset'.format(subset)) + + # Initialize data iterator + itr = trainer.get_valid_iterator(subset).next_epoch_itr( + shuffle=False, set_dataset_epoch=False # use a fixed valid set + ) + if cfg.common.tpu: + itr = utils.tpu_data_loader(itr) + progress = progress_bar.progress_bar( + itr, + log_format=cfg.common.log_format, + log_interval=cfg.common.log_interval, + epoch=epoch_itr.epoch, + prefix=f"valid on '{subset}' subset", + tensorboard_logdir=( + cfg.common.tensorboard_logdir + if distributed_utils.is_master(cfg.distributed_training) + else None + ), + default_log_format=("tqdm" if not cfg.common.no_progress_bar else "simple"), + wandb_project=( + cfg.common.wandb_project + if distributed_utils.is_master(cfg.distributed_training) + else None + ), + wandb_run_name=os.environ.get( + "WANDB_NAME", os.path.basename(cfg.checkpoint.save_dir) + ), + ) + + # create a new root metrics aggregator so validation metrics + # don't pollute other aggregators (e.g., train meters) + with metrics.aggregate(new_root=True) as agg: + for i, sample in enumerate(progress): + if cfg.dataset.max_valid_steps is not None and i > cfg.dataset.max_valid_steps: + break + trainer.valid_step(sample) + + # log validation stats + stats = get_valid_stats(cfg, trainer, agg.get_smoothed_values()) + + if hasattr(task, "post_validate"): + task.post_validate(trainer.get_model(), stats, agg) + + progress.print(stats, tag=subset, step=trainer.get_num_updates()) + + valid_losses.append(stats[cfg.checkpoint.best_checkpoint_metric]) + return valid_losses + + +def get_valid_stats( + cfg: DictConfig, trainer: Trainer, stats: Dict[str, Any] +) -> Dict[str, Any]: + stats["num_updates"] = trainer.get_num_updates() + if hasattr(checkpoint_utils.save_checkpoint, "best"): + key = "best_{0}".format(cfg.checkpoint.best_checkpoint_metric) + best_function = max if cfg.checkpoint.maximize_best_checkpoint_metric else min + stats[key] = best_function( + checkpoint_utils.save_checkpoint.best, + stats[cfg.checkpoint.best_checkpoint_metric], + ) + return stats + + +def cli_main( + modify_parser: Optional[Callable[[argparse.ArgumentParser], None]] = None +) -> None: + parser = options.get_training_parser() + args = options.parse_args_and_arch(parser, modify_parser=modify_parser) + + cfg = convert_namespace_to_omegaconf(args) + + if cfg.common.use_plasma_view: + server = PlasmaStore(path=cfg.common.plasma_path) + logger.info(f"Started plasma server pid {server.server.pid} {cfg.common.plasma_path}") + + if args.profile: + with torch.cuda.profiler.profile(): + with torch.autograd.profiler.emit_nvtx(): + distributed_utils.call_main(cfg, main) + else: + distributed_utils.call_main(cfg, main) + + # if cfg.common.use_plasma_view: + # server.server.kill() + + +if __name__ == "__main__": + cli_main() diff --git a/fairseq_cli/validate.py b/fairseq_cli/validate.py new file mode 100644 index 0000000000000000000000000000000000000000..22b93e9a6a1e1fbcff67075019177110905270f2 --- /dev/null +++ b/fairseq_cli/validate.py @@ -0,0 +1,155 @@ +#!/usr/bin/env python3 -u +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import os +import sys +from argparse import Namespace +from itertools import chain + +import torch +from fairseq import checkpoint_utils, distributed_utils, options, utils +from fairseq.dataclass.utils import convert_namespace_to_omegaconf +from fairseq.logging import metrics, progress_bar +from fairseq.utils import reset_logging +from omegaconf import DictConfig + + +logging.basicConfig( + format="%(asctime)s | %(levelname)s | %(name)s | %(message)s", + datefmt="%Y-%m-%d %H:%M:%S", + level=os.environ.get("LOGLEVEL", "INFO").upper(), + stream=sys.stdout, +) +logger = logging.getLogger("fairseq_cli.validate") + + +def main(cfg: DictConfig, override_args=None): + if isinstance(cfg, Namespace): + cfg = convert_namespace_to_omegaconf(cfg) + + utils.import_user_module(cfg.common) + + reset_logging() + + assert ( + cfg.dataset.max_tokens is not None or cfg.dataset.batch_size is not None + ), "Must specify batch size either with --max-tokens or --batch-size" + + use_fp16 = cfg.common.fp16 + use_cuda = torch.cuda.is_available() and not cfg.common.cpu + + if use_cuda: + torch.cuda.set_device(cfg.distributed_training.device_id) + + if cfg.distributed_training.distributed_world_size > 1: + data_parallel_world_size = distributed_utils.get_data_parallel_world_size() + data_parallel_rank = distributed_utils.get_data_parallel_rank() + else: + data_parallel_world_size = 1 + data_parallel_rank = 0 + + if override_args is not None: + overrides = vars(override_args) + overrides.update(eval(getattr(override_args, "model_overrides", "{}"))) + else: + overrides = None + + # Load ensemble + logger.info("loading model(s) from {}".format(cfg.common_eval.path)) + models, saved_cfg, task = checkpoint_utils.load_model_ensemble_and_task( + [cfg.common_eval.path], + arg_overrides=overrides, + suffix=cfg.checkpoint.checkpoint_suffix, + ) + model = models[0] + + # Move models to GPU + for model in models: + model.eval() + if use_fp16: + model.half() + if use_cuda: + model.cuda() + + # Print args + logger.info(saved_cfg) + + # Build criterion + criterion = task.build_criterion(saved_cfg.criterion) + criterion.eval() + + for subset in cfg.dataset.valid_subset.split(","): + try: + task.load_dataset(subset, combine=False, epoch=1, task_cfg=saved_cfg.task) + dataset = task.dataset(subset) + except KeyError: + raise Exception("Cannot find dataset: " + subset) + + # Initialize data iterator + itr = task.get_batch_iterator( + dataset=dataset, + max_tokens=cfg.dataset.max_tokens, + max_sentences=cfg.dataset.batch_size, + max_positions=utils.resolve_max_positions( + task.max_positions(), + *[m.max_positions() for m in models], + ), + ignore_invalid_inputs=cfg.dataset.skip_invalid_size_inputs_valid_test, + required_batch_size_multiple=cfg.dataset.required_batch_size_multiple, + seed=cfg.common.seed, + num_shards=data_parallel_world_size, + shard_id=data_parallel_rank, + num_workers=cfg.dataset.num_workers, + data_buffer_size=cfg.dataset.data_buffer_size, + ).next_epoch_itr(shuffle=False) + progress = progress_bar.progress_bar( + itr, + log_format=cfg.common.log_format, + log_interval=cfg.common.log_interval, + prefix=f"valid on '{subset}' subset", + default_log_format=("tqdm" if not cfg.common.no_progress_bar else "simple"), + ) + + log_outputs = [] + for i, sample in enumerate(progress): + sample = utils.move_to_cuda(sample) if use_cuda else sample + _loss, _sample_size, log_output = task.valid_step(sample, model, criterion) + progress.log(log_output, step=i) + log_outputs.append(log_output) + + if data_parallel_world_size > 1: + log_outputs = distributed_utils.all_gather_list( + log_outputs, + max_size=cfg.common.all_gather_list_size, + group=distributed_utils.get_data_parallel_group(), + ) + log_outputs = list(chain.from_iterable(log_outputs)) + + with metrics.aggregate() as agg: + task.reduce_metrics(log_outputs, criterion) + log_output = agg.get_smoothed_values() + + progress.print(log_output, tag=subset, step=i) + + +def cli_main(): + parser = options.get_validation_parser() + args = options.parse_args_and_arch(parser) + + # only override args that are explicitly given on the command line + override_parser = options.get_validation_parser() + override_args = options.parse_args_and_arch( + override_parser, suppress_defaults=True + ) + + distributed_utils.call_main( + convert_namespace_to_omegaconf(args), main, override_args=override_args + ) + + +if __name__ == "__main__": + cli_main() diff --git a/gradiodemo.py b/gradiodemo.py new file mode 100644 index 0000000000000000000000000000000000000000..0389d28745d31ff348ac958f5a39fbcd315192d1 --- /dev/null +++ b/gradiodemo.py @@ -0,0 +1,14 @@ +import gradio as gr + + + +description = "HuBERT: Self-Supervised Speech Representation Learning. To use it, simply add your audio or click one of the examples to load them. Read more at the links below." +article = "<p style='text-align: center'><a href='https://arxiv.org/abs/2106.07447'>HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units</a> | <a href='https://github.com/pytorch/fairseq/tree/master/examples/hubert'>Github Repo</a></p>" + +gr.Interface.load("huggingface/facebook/hubert-large-ls960-ft", + description=description, + article=article, + examples=[ + ["./audio1.mp3"], + ["./audio2.mp3"] +]).launch() diff --git a/hubconf.py b/hubconf.py new file mode 100644 index 0000000000000000000000000000000000000000..5949e274edd02e86cb323331211641ce0d0b9b93 --- /dev/null +++ b/hubconf.py @@ -0,0 +1,73 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +"""isort:skip_file""" + +import functools +import importlib + + +dependencies = [ + "dataclasses", + "hydra", + "numpy", + "omegaconf", + "regex", + "requests", + "torch", +] + + +# Check for required dependencies and raise a RuntimeError if any are missing. +missing_deps = [] +for dep in dependencies: + try: + importlib.import_module(dep) + except ImportError: + # Hack: the hydra package is provided under the "hydra-core" name in + # pypi. We don't want the user mistakenly calling `pip install hydra` + # since that will install an unrelated package. + if dep == "hydra": + dep = "hydra-core" + missing_deps.append(dep) +if len(missing_deps) > 0: + raise RuntimeError("Missing dependencies: {}".format(", ".join(missing_deps))) + + +# only do fairseq imports after checking for dependencies +from fairseq.hub_utils import ( # noqa; noqa + BPEHubInterface as bpe, + TokenizerHubInterface as tokenizer, +) +from fairseq.models import MODEL_REGISTRY # noqa + + +# torch.hub doesn't build Cython components, so if they are not found then try +# to build them here +try: + import fairseq.data.token_block_utils_fast # noqa +except ImportError: + try: + import cython # noqa + import os + from setuptools import sandbox + + sandbox.run_setup( + os.path.join(os.path.dirname(__file__), "setup.py"), + ["build_ext", "--inplace"], + ) + except ImportError: + print( + "Unable to build Cython components. Please make sure Cython is " + "installed if the torch.hub model you are loading depends on it." + ) + + +# automatically expose models defined in FairseqModel::hub_models +for _model_type, _cls in MODEL_REGISTRY.items(): + for model_name in _cls.hub_models().keys(): + globals()[model_name] = functools.partial( + _cls.from_pretrained, + model_name, + ) diff --git a/pyproject.toml b/pyproject.toml new file mode 100644 index 0000000000000000000000000000000000000000..6d1b4c5b6fb56a63069147e3a1de922ce71a45d8 --- /dev/null +++ b/pyproject.toml @@ -0,0 +1,3 @@ +[build-system] +requires = ["setuptools", "wheel", "cython"] +build-backend = "setuptools.build_meta" diff --git a/requirements.txt b/requirements.txt new file mode 100644 index 0000000000000000000000000000000000000000..d1cdd97e46d09f57f1c454483c7f4ee6bd11f489 --- /dev/null +++ b/requirements.txt @@ -0,0 +1,2 @@ +gradio +torch diff --git a/scripts/__init__.py b/scripts/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/scripts/average_checkpoints.py b/scripts/average_checkpoints.py new file mode 100644 index 0000000000000000000000000000000000000000..c512f802bce6b3395cc42a0e4eb39181e9f8c873 --- /dev/null +++ b/scripts/average_checkpoints.py @@ -0,0 +1,158 @@ +#!/usr/bin/env python3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import collections +import os +import re + +import torch +from fairseq.file_io import PathManager + + +def average_checkpoints(inputs): + """Loads checkpoints from inputs and returns a model with averaged weights. + + Args: + inputs: An iterable of string paths of checkpoints to load from. + + Returns: + A dict of string keys mapping to various values. The 'model' key + from the returned dict should correspond to an OrderedDict mapping + string parameter names to torch Tensors. + """ + params_dict = collections.OrderedDict() + params_keys = None + new_state = None + num_models = len(inputs) + + for fpath in inputs: + with PathManager.open(fpath, "rb") as f: + state = torch.load( + f, + map_location=( + lambda s, _: torch.serialization.default_restore_location(s, "cpu") + ), + ) + # Copies over the settings from the first checkpoint + if new_state is None: + new_state = state + + model_params = state["model"] + + model_params_keys = list(model_params.keys()) + if params_keys is None: + params_keys = model_params_keys + elif params_keys != model_params_keys: + raise KeyError( + "For checkpoint {}, expected list of params: {}, " + "but found: {}".format(f, params_keys, model_params_keys) + ) + + for k in params_keys: + p = model_params[k] + if isinstance(p, torch.HalfTensor): + p = p.float() + if k not in params_dict: + params_dict[k] = p.clone() + # NOTE: clone() is needed in case of p is a shared parameter + else: + params_dict[k] += p + + averaged_params = collections.OrderedDict() + for k, v in params_dict.items(): + averaged_params[k] = v + if averaged_params[k].is_floating_point(): + averaged_params[k].div_(num_models) + else: + averaged_params[k] //= num_models + new_state["model"] = averaged_params + return new_state + + +def last_n_checkpoints(paths, n, update_based, upper_bound=None): + assert len(paths) == 1 + path = paths[0] + if update_based: + pt_regexp = re.compile(r"checkpoint_\d+_(\d+)\.pt") + else: + pt_regexp = re.compile(r"checkpoint(\d+)\.pt") + files = PathManager.ls(path) + + entries = [] + for f in files: + m = pt_regexp.fullmatch(f) + if m is not None: + sort_key = int(m.group(1)) + if upper_bound is None or sort_key <= upper_bound: + entries.append((sort_key, m.group(0))) + if len(entries) < n: + raise Exception( + "Found {} checkpoint files but need at least {}", len(entries), n + ) + return [os.path.join(path, x[1]) for x in sorted(entries, reverse=True)[:n]] + + +def main(): + parser = argparse.ArgumentParser( + description="Tool to average the params of input checkpoints to " + "produce a new checkpoint", + ) + # fmt: off + parser.add_argument('--inputs', required=True, nargs='+', + help='Input checkpoint file paths.') + parser.add_argument('--output', required=True, metavar='FILE', + help='Write the new checkpoint containing the averaged weights to this path.') + num_group = parser.add_mutually_exclusive_group() + num_group.add_argument('--num-epoch-checkpoints', type=int, + help='if set, will try to find checkpoints with names checkpoint_xx.pt in the path specified by input, ' + 'and average last this many of them.') + num_group.add_argument('--num-update-checkpoints', type=int, + help='if set, will try to find checkpoints with names checkpoint_ee_xx.pt in the path specified by input, ' + 'and average last this many of them.') + parser.add_argument('--checkpoint-upper-bound', type=int, + help='when using --num-epoch-checkpoints, this will set an upper bound on which epoch to use, ' + 'when using --num-update-checkpoints, this will set an upper bound on which update to use' + 'e.g., with --num-epoch-checkpoints=10 --checkpoint-upper-bound=50, checkpoints 41-50 would be averaged.' + 'e.g., with --num-update-checkpoints=10 --checkpoint-upper-bound=50000, checkpoints 40500-50000 would be averaged assuming --save-interval-updates 500' + ) + # fmt: on + args = parser.parse_args() + print(args) + + num = None + is_update_based = False + if args.num_update_checkpoints is not None: + num = args.num_update_checkpoints + is_update_based = True + elif args.num_epoch_checkpoints is not None: + num = args.num_epoch_checkpoints + + assert args.checkpoint_upper_bound is None or ( + args.num_epoch_checkpoints is not None + or args.num_update_checkpoints is not None + ), "--checkpoint-upper-bound requires --num-epoch-checkpoints or --num-update-checkpoints" + assert ( + args.num_epoch_checkpoints is None or args.num_update_checkpoints is None + ), "Cannot combine --num-epoch-checkpoints and --num-update-checkpoints" + + if num is not None: + args.inputs = last_n_checkpoints( + args.inputs, + num, + is_update_based, + upper_bound=args.checkpoint_upper_bound, + ) + print("averaging checkpoints: ", args.inputs) + + new_state = average_checkpoints(args.inputs) + with PathManager.open(args.output, "wb") as f: + torch.save(new_state, f) + print("Finished writing averaged checkpoint to {}".format(args.output)) + + +if __name__ == "__main__": + main() diff --git a/scripts/build_sym_alignment.py b/scripts/build_sym_alignment.py new file mode 100644 index 0000000000000000000000000000000000000000..0ca5c18f7bd4b0fbf58b203793506ca395466129 --- /dev/null +++ b/scripts/build_sym_alignment.py @@ -0,0 +1,97 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +Use this script in order to build symmetric alignments for your translation +dataset. +This script depends on fast_align and mosesdecoder tools. You will need to +build those before running the script. +fast_align: + github: http://github.com/clab/fast_align + instructions: follow the instructions in README.md +mosesdecoder: + github: http://github.com/moses-smt/mosesdecoder + instructions: http://www.statmt.org/moses/?n=Development.GetStarted +The script produces the following files under --output_dir: + text.joined - concatenation of lines from the source_file and the + target_file. + align.forward - forward pass of fast_align. + align.backward - backward pass of fast_align. + aligned.sym_heuristic - symmetrized alignment. +""" + +import argparse +import os +from itertools import zip_longest + + +def main(): + parser = argparse.ArgumentParser(description="symmetric alignment builer") + # fmt: off + parser.add_argument('--fast_align_dir', + help='path to fast_align build directory') + parser.add_argument('--mosesdecoder_dir', + help='path to mosesdecoder root directory') + parser.add_argument('--sym_heuristic', + help='heuristic to use for symmetrization', + default='grow-diag-final-and') + parser.add_argument('--source_file', + help='path to a file with sentences ' + 'in the source language') + parser.add_argument('--target_file', + help='path to a file with sentences ' + 'in the target language') + parser.add_argument('--output_dir', + help='output directory') + # fmt: on + args = parser.parse_args() + + fast_align_bin = os.path.join(args.fast_align_dir, "fast_align") + symal_bin = os.path.join(args.mosesdecoder_dir, "bin", "symal") + sym_fast_align_bin = os.path.join( + args.mosesdecoder_dir, "scripts", "ems", "support", "symmetrize-fast-align.perl" + ) + + # create joined file + joined_file = os.path.join(args.output_dir, "text.joined") + with open(args.source_file, "r", encoding="utf-8") as src, open( + args.target_file, "r", encoding="utf-8" + ) as tgt: + with open(joined_file, "w", encoding="utf-8") as joined: + for s, t in zip_longest(src, tgt): + print("{} ||| {}".format(s.strip(), t.strip()), file=joined) + + bwd_align_file = os.path.join(args.output_dir, "align.backward") + + # run forward alignment + fwd_align_file = os.path.join(args.output_dir, "align.forward") + fwd_fast_align_cmd = "{FASTALIGN} -i {JOINED} -d -o -v > {FWD}".format( + FASTALIGN=fast_align_bin, JOINED=joined_file, FWD=fwd_align_file + ) + assert os.system(fwd_fast_align_cmd) == 0 + + # run backward alignment + bwd_align_file = os.path.join(args.output_dir, "align.backward") + bwd_fast_align_cmd = "{FASTALIGN} -i {JOINED} -d -o -v -r > {BWD}".format( + FASTALIGN=fast_align_bin, JOINED=joined_file, BWD=bwd_align_file + ) + assert os.system(bwd_fast_align_cmd) == 0 + + # run symmetrization + sym_out_file = os.path.join(args.output_dir, "aligned") + sym_cmd = "{SYMFASTALIGN} {FWD} {BWD} {SRC} {TGT} {OUT} {HEURISTIC} {SYMAL}".format( + SYMFASTALIGN=sym_fast_align_bin, + FWD=fwd_align_file, + BWD=bwd_align_file, + SRC=args.source_file, + TGT=args.target_file, + OUT=sym_out_file, + HEURISTIC=args.sym_heuristic, + SYMAL=symal_bin, + ) + assert os.system(sym_cmd) == 0 + + +if __name__ == "__main__": + main() diff --git a/scripts/compare_namespaces.py b/scripts/compare_namespaces.py new file mode 100644 index 0000000000000000000000000000000000000000..bc24db624f8db36f546c263ba3a806dae6d466bf --- /dev/null +++ b/scripts/compare_namespaces.py @@ -0,0 +1,46 @@ +#!/usr/bin/env python +"""Helper script to compare two argparse.Namespace objects.""" + +from argparse import Namespace # noqa + + +def main(): + + ns1 = eval(input("Namespace 1: ")) + ns2 = eval(input("Namespace 2: ")) + + def keys(ns): + ks = set() + for k in dir(ns): + if not k.startswith("_"): + ks.add(k) + return ks + + k1 = keys(ns1) + k2 = keys(ns2) + + def print_keys(ks, ns1, ns2=None): + for k in ks: + if ns2 is None: + print("{}\t{}".format(k, getattr(ns1, k, None))) + else: + print( + "{}\t{}\t{}".format(k, getattr(ns1, k, None), getattr(ns2, k, None)) + ) + + print("Keys unique to namespace 1:") + print_keys(k1 - k2, ns1) + print() + + print("Keys unique to namespace 2:") + print_keys(k2 - k1, ns2) + print() + + print("Overlapping keys with different values:") + ks = [k for k in k1 & k2 if getattr(ns1, k, "None") != getattr(ns2, k, "None")] + print_keys(ks, ns1, ns2) + print() + + +if __name__ == "__main__": + main() diff --git a/scripts/compound_split_bleu.sh b/scripts/compound_split_bleu.sh new file mode 100644 index 0000000000000000000000000000000000000000..1972fddcebff9a43a70bcf14c287175c68f60e3f --- /dev/null +++ b/scripts/compound_split_bleu.sh @@ -0,0 +1,20 @@ +#!/bin/bash + +if [ $# -ne 1 ]; then + echo "usage: $0 GENERATE_PY_OUTPUT" + exit 1 +fi + +GEN=$1 + +SYS=$GEN.sys +REF=$GEN.ref + +if [ $(tail -n 1 $GEN | grep BLEU | wc -l) -ne 1 ]; then + echo "not done generating" + exit +fi + +grep ^H $GEN | awk -F '\t' '{print $NF}' | perl -ple 's{(\S)-(\S)}{$1 ##AT##-##AT## $2}g' > $SYS +grep ^T $GEN | cut -f2- | perl -ple 's{(\S)-(\S)}{$1 ##AT##-##AT## $2}g' > $REF +fairseq-score --sys $SYS --ref $REF diff --git a/scripts/constraints/extract.py b/scripts/constraints/extract.py new file mode 100755 index 0000000000000000000000000000000000000000..f6155d0a0538aadb46bf612256b6b949728de69e --- /dev/null +++ b/scripts/constraints/extract.py @@ -0,0 +1,92 @@ +#!/usr/bin/env python3 +# +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +"""Extracts random constraints from reference files.""" + +import argparse +import random +import sys + +from sacrebleu import extract_ngrams + + +def get_phrase(words, index, length): + assert index < len(words) - length + 1 + phr = " ".join(words[index : index + length]) + for i in range(index, index + length): + words.pop(index) + return phr + + +def main(args): + + if args.seed: + random.seed(args.seed) + + for line in sys.stdin: + constraints = [] + + def add_constraint(constraint): + constraints.append(constraint) + + source = line.rstrip() + if "\t" in line: + source, target = line.split("\t") + if args.add_sos: + target = f"<s> {target}" + if args.add_eos: + target = f"{target} </s>" + + if len(target.split()) >= args.len: + words = [target] + + num = args.number + + choices = {} + for i in range(num): + if len(words) == 0: + break + segmentno = random.choice(range(len(words))) + segment = words.pop(segmentno) + tokens = segment.split() + phrase_index = random.choice(range(len(tokens))) + choice = " ".join( + tokens[phrase_index : min(len(tokens), phrase_index + args.len)] + ) + for j in range( + phrase_index, min(len(tokens), phrase_index + args.len) + ): + tokens.pop(phrase_index) + if phrase_index > 0: + words.append(" ".join(tokens[0:phrase_index])) + if phrase_index + 1 < len(tokens): + words.append(" ".join(tokens[phrase_index:])) + choices[target.find(choice)] = choice + + # mask out with spaces + target = target.replace(choice, " " * len(choice), 1) + + for key in sorted(choices.keys()): + add_constraint(choices[key]) + + print(source, *constraints, sep="\t") + + +if __name__ == "__main__": + parser = argparse.ArgumentParser() + parser.add_argument("--number", "-n", type=int, default=1, help="number of phrases") + parser.add_argument("--len", "-l", type=int, default=1, help="phrase length") + parser.add_argument( + "--add-sos", default=False, action="store_true", help="add <s> token" + ) + parser.add_argument( + "--add-eos", default=False, action="store_true", help="add </s> token" + ) + parser.add_argument("--seed", "-s", default=0, type=int) + args = parser.parse_args() + + main(args) diff --git a/scripts/constraints/validate.py b/scripts/constraints/validate.py new file mode 100755 index 0000000000000000000000000000000000000000..d531ad9f39b1df42c98fe8f26ad61fe53a9ac0c5 --- /dev/null +++ b/scripts/constraints/validate.py @@ -0,0 +1,34 @@ +#!/usr/bin/env python3 +# +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import sys + + +"""Reads in a fairseq output file, and verifies that the constraints +(C- lines) are present in the output (the first H- line). Assumes that +constraints are listed prior to the first hypothesis. +""" + +constraints = [] +found = 0 +total = 0 +for line in sys.stdin: + if line.startswith("C-"): + constraints.append(line.rstrip().split("\t")[1]) + elif line.startswith("H-"): + text = line.split("\t")[2] + + for constraint in constraints: + total += 1 + if constraint in text: + found += 1 + else: + print(f"No {constraint} in {text}", file=sys.stderr) + + constraints = [] + +print(f"Found {found} / {total} = {100 * found / total:.1f}%") diff --git a/scripts/convert_dictionary.lua b/scripts/convert_dictionary.lua new file mode 100644 index 0000000000000000000000000000000000000000..14ee8c997f642c8ff196617c2dcd0584037a60c4 --- /dev/null +++ b/scripts/convert_dictionary.lua @@ -0,0 +1,34 @@ +-- Copyright (c) Facebook, Inc. and its affiliates. +-- +-- This source code is licensed under the MIT license found in the +-- LICENSE file in the root directory of this source tree. +-- +-- Usage: convert_dictionary.lua <dict.th7> +require 'fairseq' +require 'torch' +require 'paths' + +if #arg < 1 then + print('usage: convert_dictionary.lua <dict.th7>') + os.exit(1) +end +if not paths.filep(arg[1]) then + print('error: file does not exit: ' .. arg[1]) + os.exit(1) +end + +dict = torch.load(arg[1]) +dst = paths.basename(arg[1]):gsub('.th7', '.txt') +assert(dst:match('.txt$')) + +f = io.open(dst, 'w') +for idx, symbol in ipairs(dict.index_to_symbol) do + if idx > dict.cutoff then + break + end + f:write(symbol) + f:write(' ') + f:write(dict.index_to_freq[idx]) + f:write('\n') +end +f:close() diff --git a/scripts/convert_model.lua b/scripts/convert_model.lua new file mode 100644 index 0000000000000000000000000000000000000000..61b92139294fb90a25989ebd2ee52a765fb278a2 --- /dev/null +++ b/scripts/convert_model.lua @@ -0,0 +1,108 @@ +-- Copyright (c) Facebook, Inc. and its affiliates. +-- +-- This source code is licensed under the MIT license found in the +-- LICENSE file in the root directory of this source tree. +-- +-- Usage: convert_model.lua <model_epoch1.th7> +require 'torch' +local fairseq = require 'fairseq' + +model = torch.load(arg[1]) + +function find_weight_norm(container, module) + for _, wn in ipairs(container:listModules()) do + if torch.type(wn) == 'nn.WeightNorm' and wn.modules[1] == module then + return wn + end + end +end + +function push_state(dict, key, module) + if torch.type(module) == 'nn.Linear' then + local wn = find_weight_norm(model.module, module) + assert(wn) + dict[key .. '.weight_v'] = wn.v:float() + dict[key .. '.weight_g'] = wn.g:float() + elseif torch.type(module) == 'nn.TemporalConvolutionTBC' then + local wn = find_weight_norm(model.module, module) + assert(wn) + local v = wn.v:float():view(wn.viewOut):transpose(2, 3) + dict[key .. '.weight_v'] = v + dict[key .. '.weight_g'] = wn.g:float():view(module.weight:size(3), 1, 1) + else + dict[key .. '.weight'] = module.weight:float() + end + if module.bias then + dict[key .. '.bias'] = module.bias:float() + end +end + +encoder_dict = {} +decoder_dict = {} +combined_dict = {} + +function encoder_state(encoder) + luts = encoder:findModules('nn.LookupTable') + push_state(encoder_dict, 'embed_tokens', luts[1]) + push_state(encoder_dict, 'embed_positions', luts[2]) + + fcs = encoder:findModules('nn.Linear') + assert(#fcs >= 2) + local nInputPlane = fcs[1].weight:size(1) + push_state(encoder_dict, 'fc1', table.remove(fcs, 1)) + push_state(encoder_dict, 'fc2', table.remove(fcs, #fcs)) + + for i, module in ipairs(encoder:findModules('nn.TemporalConvolutionTBC')) do + push_state(encoder_dict, 'convolutions.' .. tostring(i - 1), module) + if nInputPlane ~= module.weight:size(3) / 2 then + push_state(encoder_dict, 'projections.' .. tostring(i - 1), table.remove(fcs, 1)) + end + nInputPlane = module.weight:size(3) / 2 + end + assert(#fcs == 0) +end + +function decoder_state(decoder) + luts = decoder:findModules('nn.LookupTable') + push_state(decoder_dict, 'embed_tokens', luts[1]) + push_state(decoder_dict, 'embed_positions', luts[2]) + + fcs = decoder:findModules('nn.Linear') + local nInputPlane = fcs[1].weight:size(1) + push_state(decoder_dict, 'fc1', table.remove(fcs, 1)) + push_state(decoder_dict, 'fc2', fcs[#fcs - 1]) + push_state(decoder_dict, 'fc3', fcs[#fcs]) + + table.remove(fcs, #fcs) + table.remove(fcs, #fcs) + + for i, module in ipairs(decoder:findModules('nn.TemporalConvolutionTBC')) do + if nInputPlane ~= module.weight:size(3) / 2 then + push_state(decoder_dict, 'projections.' .. tostring(i - 1), table.remove(fcs, 1)) + end + nInputPlane = module.weight:size(3) / 2 + + local prefix = 'attention.' .. tostring(i - 1) + push_state(decoder_dict, prefix .. '.in_projection', table.remove(fcs, 1)) + push_state(decoder_dict, prefix .. '.out_projection', table.remove(fcs, 1)) + push_state(decoder_dict, 'convolutions.' .. tostring(i - 1), module) + end + assert(#fcs == 0) +end + + +_encoder = model.module.modules[2] +_decoder = model.module.modules[3] + +encoder_state(_encoder) +decoder_state(_decoder) + +for k, v in pairs(encoder_dict) do + combined_dict['encoder.' .. k] = v +end +for k, v in pairs(decoder_dict) do + combined_dict['decoder.' .. k] = v +end + + +torch.save('state_dict.t7', combined_dict) diff --git a/scripts/count_docs.py b/scripts/count_docs.py new file mode 100644 index 0000000000000000000000000000000000000000..58d85af85e91377a34dbd01f7674436152fd08e8 --- /dev/null +++ b/scripts/count_docs.py @@ -0,0 +1,58 @@ +#!/usr/bin/env python3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +Count the number of documents and average number of lines and tokens per +document in a large file. Documents should be separated by a single empty line. +""" + +import argparse +import gzip +import sys + +import numpy as np + + +def main(): + parser = argparse.ArgumentParser() + parser.add_argument("input") + parser.add_argument("--gzip", action="store_true") + args = parser.parse_args() + + def gopen(): + if args.gzip: + return gzip.open(args.input, "r") + else: + return open(args.input, "r", encoding="utf-8") + + num_lines = [] + num_toks = [] + with gopen() as h: + num_docs = 1 + num_lines_in_doc = 0 + num_toks_in_doc = 0 + for i, line in enumerate(h): + if len(line.strip()) == 0: # empty line indicates new document + num_docs += 1 + num_lines.append(num_lines_in_doc) + num_toks.append(num_toks_in_doc) + num_lines_in_doc = 0 + num_toks_in_doc = 0 + else: + num_lines_in_doc += 1 + num_toks_in_doc += len(line.rstrip().split()) + if i % 1000000 == 0: + print(i, file=sys.stderr, end="", flush=True) + elif i % 100000 == 0: + print(".", file=sys.stderr, end="", flush=True) + print(file=sys.stderr, flush=True) + + print("found {} docs".format(num_docs)) + print("average num lines per doc: {}".format(np.mean(num_lines))) + print("average num toks per doc: {}".format(np.mean(num_toks))) + + +if __name__ == "__main__": + main() diff --git a/scripts/read_binarized.py b/scripts/read_binarized.py new file mode 100644 index 0000000000000000000000000000000000000000..a414095d03fb022a6753e816fc8bfd80e11db24d --- /dev/null +++ b/scripts/read_binarized.py @@ -0,0 +1,48 @@ +#!/usr/bin/env python3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse + +from fairseq.data import Dictionary, data_utils, indexed_dataset + + +def get_parser(): + parser = argparse.ArgumentParser( + description="writes text from binarized file to stdout" + ) + # fmt: off + parser.add_argument('--dataset-impl', help='dataset implementation', + choices=indexed_dataset.get_available_dataset_impl()) + parser.add_argument('--dict', metavar='FP', help='dictionary containing known words', default=None) + parser.add_argument('--input', metavar='FP', required=True, help='binarized file to read') + # fmt: on + + return parser + + +def main(): + parser = get_parser() + args = parser.parse_args() + + dictionary = Dictionary.load(args.dict) if args.dict is not None else None + dataset = data_utils.load_indexed_dataset( + args.input, + dictionary, + dataset_impl=args.dataset_impl, + default="lazy", + ) + + for tensor_line in dataset: + if dictionary is None: + line = " ".join([str(int(x)) for x in tensor_line]) + else: + line = dictionary.string(tensor_line) + + print(line) + + +if __name__ == "__main__": + main() diff --git a/scripts/rm_pt.py b/scripts/rm_pt.py new file mode 100644 index 0000000000000000000000000000000000000000..6cd063d21f0610fa7c42c2cfb2ee8af7c9c78677 --- /dev/null +++ b/scripts/rm_pt.py @@ -0,0 +1,141 @@ +#!/usr/bin/env python3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import os +import re +import shutil +import sys + + +pt_regexp = re.compile(r"checkpoint(\d+|_\d+_\d+|_[a-z]+)\.pt") +pt_regexp_epoch_based = re.compile(r"checkpoint(\d+)\.pt") +pt_regexp_update_based = re.compile(r"checkpoint_\d+_(\d+)\.pt") + + +def parse_checkpoints(files): + entries = [] + for f in files: + m = pt_regexp_epoch_based.fullmatch(f) + if m is not None: + entries.append((int(m.group(1)), m.group(0))) + else: + m = pt_regexp_update_based.fullmatch(f) + if m is not None: + entries.append((int(m.group(1)), m.group(0))) + return entries + + +def last_n_checkpoints(files, n): + entries = parse_checkpoints(files) + return [x[1] for x in sorted(entries, reverse=True)[:n]] + + +def every_n_checkpoints(files, n): + entries = parse_checkpoints(files) + return [x[1] for x in sorted(sorted(entries)[::-n])] + + +def main(): + parser = argparse.ArgumentParser( + description=( + "Recursively delete checkpoint files from `root_dir`, " + "but preserve checkpoint_best.pt and checkpoint_last.pt" + ) + ) + parser.add_argument("root_dirs", nargs="*") + parser.add_argument( + "--save-last", type=int, default=0, help="number of last checkpoints to save" + ) + parser.add_argument( + "--save-every", type=int, default=0, help="interval of checkpoints to save" + ) + parser.add_argument( + "--preserve-test", + action="store_true", + help="preserve checkpoints in dirs that start with test_ prefix (default: delete them)", + ) + parser.add_argument( + "--delete-best", action="store_true", help="delete checkpoint_best.pt" + ) + parser.add_argument( + "--delete-last", action="store_true", help="delete checkpoint_last.pt" + ) + parser.add_argument( + "--no-dereference", action="store_true", help="don't dereference symlinks" + ) + args = parser.parse_args() + + files_to_desymlink = [] + files_to_preserve = [] + files_to_delete = [] + for root_dir in args.root_dirs: + for root, _subdirs, files in os.walk(root_dir): + if args.save_last > 0: + to_save = last_n_checkpoints(files, args.save_last) + else: + to_save = [] + if args.save_every > 0: + to_save += every_n_checkpoints(files, args.save_every) + for file in files: + if not pt_regexp.fullmatch(file): + continue + full_path = os.path.join(root, file) + if ( + not os.path.basename(root).startswith("test_") or args.preserve_test + ) and ( + (file == "checkpoint_last.pt" and not args.delete_last) + or (file == "checkpoint_best.pt" and not args.delete_best) + or file in to_save + ): + if os.path.islink(full_path) and not args.no_dereference: + files_to_desymlink.append(full_path) + else: + files_to_preserve.append(full_path) + else: + files_to_delete.append(full_path) + + if len(files_to_desymlink) == 0 and len(files_to_delete) == 0: + print("Nothing to do.") + sys.exit(0) + + files_to_desymlink = sorted(files_to_desymlink) + files_to_preserve = sorted(files_to_preserve) + files_to_delete = sorted(files_to_delete) + + print("Operations to perform (in order):") + if len(files_to_desymlink) > 0: + for file in files_to_desymlink: + print(" - preserve (and dereference symlink): " + file) + if len(files_to_preserve) > 0: + for file in files_to_preserve: + print(" - preserve: " + file) + if len(files_to_delete) > 0: + for file in files_to_delete: + print(" - delete: " + file) + while True: + resp = input("Continue? (Y/N): ") + if resp.strip().lower() == "y": + break + elif resp.strip().lower() == "n": + sys.exit(0) + + print("Executing...") + if len(files_to_desymlink) > 0: + for file in files_to_desymlink: + realpath = os.path.realpath(file) + print("rm " + file) + os.remove(file) + print("cp {} {}".format(realpath, file)) + shutil.copyfile(realpath, file) + if len(files_to_delete) > 0: + for file in files_to_delete: + print("rm " + file) + os.remove(file) + + +if __name__ == "__main__": + main() diff --git a/scripts/sacrebleu.sh b/scripts/sacrebleu.sh new file mode 100644 index 0000000000000000000000000000000000000000..c10bf2b76ea032deabab6f5c9d8a3e1e884f1642 --- /dev/null +++ b/scripts/sacrebleu.sh @@ -0,0 +1,27 @@ +#!/bin/bash + +if [ $# -ne 4 ]; then + echo "usage: $0 TESTSET SRCLANG TGTLANG GEN" + exit 1 +fi + +TESTSET=$1 +SRCLANG=$2 +TGTLANG=$3 + +GEN=$4 + +if ! command -v sacremoses &> /dev/null +then + echo "sacremoses could not be found, please install with: pip install sacremoses" + exit +fi + +grep ^H $GEN \ +| sed 's/^H\-//' \ +| sort -n -k 1 \ +| cut -f 3 \ +| sacremoses detokenize \ +> $GEN.sorted.detok + +sacrebleu --test-set $TESTSET --language-pair "${SRCLANG}-${TGTLANG}" < $GEN.sorted.detok diff --git a/scripts/shard_docs.py b/scripts/shard_docs.py new file mode 100644 index 0000000000000000000000000000000000000000..97232c3c845ee01dc5ab627388934cc0f9588280 --- /dev/null +++ b/scripts/shard_docs.py @@ -0,0 +1,54 @@ +#!/usr/bin/env python3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +Split a large file into shards while respecting document boundaries. Documents +should be separated by a single empty line. +""" + +import argparse +import contextlib + + +def main(): + parser = argparse.ArgumentParser() + parser.add_argument("input") + parser.add_argument("--num-shards", type=int) + args = parser.parse_args() + + assert args.num_shards is not None and args.num_shards > 1 + + with open(args.input, "r", encoding="utf-8") as h: + with contextlib.ExitStack() as stack: + outputs = [ + stack.enter_context( + open(args.input + ".shard" + str(i), "w", encoding="utf-8") + ) + for i in range(args.num_shards) + ] + + doc = [] + first_doc = [True] * args.num_shards + + def output_doc(i): + if not first_doc[i]: + outputs[i].write("\n") + first_doc[i] = False + for line in doc: + outputs[i].write(line) + doc.clear() + + num_docs = 0 + for line in h: + if line.strip() == "": # empty line indicates new document + output_doc(num_docs % args.num_shards) + num_docs += 1 + else: + doc.append(line) + output_doc(num_docs % args.num_shards) + + +if __name__ == "__main__": + main() diff --git a/scripts/split_train_valid_docs.py b/scripts/split_train_valid_docs.py new file mode 100644 index 0000000000000000000000000000000000000000..ff159785284a13b44626b207d84430c592acaf8f --- /dev/null +++ b/scripts/split_train_valid_docs.py @@ -0,0 +1,86 @@ +#!/usr/bin/env python3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +Split a large file into a train and valid set while respecting document +boundaries. Documents should be separated by a single empty line. +""" + +import argparse +import random +import sys + + +def main(): + parser = argparse.ArgumentParser() + parser.add_argument("input") + parser.add_argument("sample_output", help="train output file") + parser.add_argument("remainder_output", help="valid output file") + parser.add_argument("-k", type=int, help="remainder size") + parser.add_argument( + "--lines", action="store_true", help="split lines instead of docs" + ) + args = parser.parse_args() + + assert args.k is not None + + sample = [] + remainder = [] + num_docs = [0] + + def update_sample(doc): + if len(sample) < args.k: + sample.append(doc.copy()) + else: + i = num_docs[0] + j = random.randrange(i + 1) + if j < args.k: + remainder.append(sample[j]) + sample[j] = doc.copy() + else: + remainder.append(doc.copy()) + num_docs[0] += 1 + doc.clear() + + with open(args.input, "r", encoding="utf-8") as h: + doc = [] + for i, line in enumerate(h): + if line.strip() == "": # empty line indicates new document + update_sample(doc) + else: + doc.append(line) + if args.lines: + update_sample(doc) + if i % 1000000 == 0: + print(i, file=sys.stderr, end="", flush=True) + elif i % 100000 == 0: + print(".", file=sys.stderr, end="", flush=True) + if len(doc) > 0: + update_sample(doc) + print(file=sys.stderr, flush=True) + + assert len(sample) == args.k + + with open(args.sample_output, "w", encoding="utf-8") as out: + first = True + for doc in sample: + if not first and not args.lines: + out.write("\n") + first = False + for line in doc: + out.write(line) + + with open(args.remainder_output, "w", encoding="utf-8") as out: + first = True + for doc in remainder: + if not first and not args.lines: + out.write("\n") + first = False + for line in doc: + out.write(line) + + +if __name__ == "__main__": + main() diff --git a/scripts/spm_decode.py b/scripts/spm_decode.py new file mode 100644 index 0000000000000000000000000000000000000000..1c18b1d2a7d7628b7aeb6fdb6c4ab5a096e9edf8 --- /dev/null +++ b/scripts/spm_decode.py @@ -0,0 +1,53 @@ +#!/usr/bin/env python +# Copyright (c) Facebook, Inc. and its affiliates. +# All rights reserved. +# +# This source code is licensed under the license found in the +# LICENSE file in the root directory of this source tree. + +from __future__ import absolute_import, division, print_function, unicode_literals + +import argparse + +import sentencepiece as spm + + +def main(): + parser = argparse.ArgumentParser() + parser.add_argument( + "--model", required=True, help="sentencepiece model to use for decoding" + ) + parser.add_argument("--input", required=True, help="input file to decode") + parser.add_argument("--input_format", choices=["piece", "id"], default="piece") + args = parser.parse_args() + + sp = spm.SentencePieceProcessor() + sp.Load(args.model) + + if args.input_format == "piece": + + def decode(l): + return "".join(sp.DecodePieces(l)) + + elif args.input_format == "id": + + def decode(l): + return "".join(sp.DecodeIds(l)) + + else: + raise NotImplementedError + + def tok2int(tok): + # remap reference-side <unk> (represented as <<unk>>) to 0 + return int(tok) if tok != "<<unk>>" else 0 + + with open(args.input, "r", encoding="utf-8") as h: + for line in h: + if args.input_format == "id": + print(decode(list(map(tok2int, line.rstrip().split())))) + elif args.input_format == "piece": + print(decode(line.rstrip().split())) + + +if __name__ == "__main__": + main() diff --git a/scripts/spm_encode.py b/scripts/spm_encode.py new file mode 100644 index 0000000000000000000000000000000000000000..83facfb3b184aff8b9cc3f0c82dd53668c63e57b --- /dev/null +++ b/scripts/spm_encode.py @@ -0,0 +1,119 @@ +#!/usr/bin/env python +# Copyright (c) Facebook, Inc. and its affiliates. +# All rights reserved. +# +# This source code is licensed under the license found in the +# LICENSE file in the root directory of this source tree. + +from __future__ import absolute_import, division, print_function, unicode_literals + +import argparse +import contextlib +import sys + +import sentencepiece as spm + + +def main(): + parser = argparse.ArgumentParser() + parser.add_argument( + "--model", required=True, help="sentencepiece model to use for encoding" + ) + parser.add_argument( + "--inputs", nargs="+", default=["-"], help="input files to filter/encode" + ) + parser.add_argument( + "--outputs", nargs="+", default=["-"], help="path to save encoded outputs" + ) + parser.add_argument("--output_format", choices=["piece", "id"], default="piece") + parser.add_argument( + "--min-len", + type=int, + metavar="N", + help="filter sentence pairs with fewer than N tokens", + ) + parser.add_argument( + "--max-len", + type=int, + metavar="N", + help="filter sentence pairs with more than N tokens", + ) + args = parser.parse_args() + + assert len(args.inputs) == len( + args.outputs + ), "number of input and output paths should match" + + sp = spm.SentencePieceProcessor() + sp.Load(args.model) + + if args.output_format == "piece": + + def encode(l): + return sp.EncodeAsPieces(l) + + elif args.output_format == "id": + + def encode(l): + return list(map(str, sp.EncodeAsIds(l))) + + else: + raise NotImplementedError + + if args.min_len is not None or args.max_len is not None: + + def valid(line): + return (args.min_len is None or len(line) >= args.min_len) and ( + args.max_len is None or len(line) <= args.max_len + ) + + else: + + def valid(lines): + return True + + with contextlib.ExitStack() as stack: + inputs = [ + stack.enter_context(open(input, "r", encoding="utf-8")) + if input != "-" + else sys.stdin + for input in args.inputs + ] + outputs = [ + stack.enter_context(open(output, "w", encoding="utf-8")) + if output != "-" + else sys.stdout + for output in args.outputs + ] + + stats = { + "num_empty": 0, + "num_filtered": 0, + } + + def encode_line(line): + line = line.strip() + if len(line) > 0: + line = encode(line) + if valid(line): + return line + else: + stats["num_filtered"] += 1 + else: + stats["num_empty"] += 1 + return None + + for i, lines in enumerate(zip(*inputs), start=1): + enc_lines = list(map(encode_line, lines)) + if not any(enc_line is None for enc_line in enc_lines): + for enc_line, output_h in zip(enc_lines, outputs): + print(" ".join(enc_line), file=output_h) + if i % 10000 == 0: + print("processed {} lines".format(i), file=sys.stderr) + + print("skipped {} empty lines".format(stats["num_empty"]), file=sys.stderr) + print("filtered {} lines".format(stats["num_filtered"]), file=sys.stderr) + + +if __name__ == "__main__": + main() diff --git a/scripts/spm_train.py b/scripts/spm_train.py new file mode 100644 index 0000000000000000000000000000000000000000..9db668fd4166a860198784990de68ea26157995d --- /dev/null +++ b/scripts/spm_train.py @@ -0,0 +1,16 @@ +#!/usr/bin/env python +# Copyright (c) Facebook, Inc. and its affiliates. +# All rights reserved. +# +# This source code is licensed under the license found in the +# LICENSE file in the root directory of this source tree. + +from __future__ import absolute_import, division, print_function, unicode_literals + +import sys + +import sentencepiece as spm + + +if __name__ == "__main__": + spm.SentencePieceTrainer.Train(" ".join(sys.argv[1:])) diff --git a/scripts/test_fsdp.sh b/scripts/test_fsdp.sh new file mode 100755 index 0000000000000000000000000000000000000000..1f428a035e4474427ded991f8e8307ea59f61f69 --- /dev/null +++ b/scripts/test_fsdp.sh @@ -0,0 +1,24 @@ +#!/usr/bin/env bash +rm -rf fsdp_dummy +mkdir -p fsdp_dummy +CUDA_VISIBLE_DEVICES=0,1,2,3 fairseq-train /private/home/sshleifer/data-bin/stories_mmap \ + --ddp-backend fully_sharded --fp16 --fp16-init-scale 4 \ + --cpu-offload --checkpoint-activations \ + --task language_modeling --tokens-per-sample 256 --batch-size 8 \ + --arch transformer_lm_gpt2_tiny \ + --optimizer cpu_adam --adam-betas "(0.9,0.98)" \ + --lr 0.0001 --lr-scheduler polynomial_decay --warmup-updates 5 --total-num-update 10 \ + --max-update 5 --log-format json --log-interval 1 \ + --save-interval-updates 5 --save-dir fsdp_dummy --disable-validation \ + --restore-file x.pt "$@" + +# Now we try to load the checkpoint +CUDA_VISIBLE_DEVICES=0,1 fairseq-train /private/home/sshleifer/data-bin/stories_mmap \ + --ddp-backend fully_sharded --fp16 --fp16-init-scale 4 \ + --cpu-offload --checkpoint-activations \ + --task language_modeling --tokens-per-sample 256 --batch-size 8 \ + --arch transformer_lm_gpt2_tiny \ + --optimizer cpu_adam --adam-betas "(0.9,0.98)" \ + --lr 0.0001 --lr-scheduler polynomial_decay --warmup-updates 5 --total-num-update 10 \ + --max-update 2 --log-format json --log-interval 1 \ + --save-interval-updates 2 --save-dir fsdp_dummy diff --git a/setup.py b/setup.py new file mode 100644 index 0000000000000000000000000000000000000000..51e555229c6111616362583731b181125e489ad7 --- /dev/null +++ b/setup.py @@ -0,0 +1,271 @@ +#!/usr/bin/env python3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import os +import subprocess +import sys +from setuptools import setup, find_packages, Extension + +from setuptools import Extension, find_packages, setup + + +if sys.version_info < (3, 6): + sys.exit("Sorry, Python >= 3.6 is required for fairseq.") + + +def write_version_py(): + with open(os.path.join("fairseq", "version.txt")) as f: + version = f.read().strip() + + # append latest commit hash to version string + try: + sha = ( + subprocess.check_output(["git", "rev-parse", "HEAD"]) + .decode("ascii") + .strip() + ) + version += "+" + sha[:7] + except Exception: + pass + + # write version info to fairseq/version.py + with open(os.path.join("fairseq", "version.py"), "w") as f: + f.write('__version__ = "{}"\n'.format(version)) + return version + + +version = write_version_py() + + +with open("README.md") as f: + readme = f.read() + + +if sys.platform == "darwin": + extra_compile_args = ["-stdlib=libc++", "-O3"] +else: + extra_compile_args = ["-std=c++11", "-O3"] + + +class NumpyExtension(Extension): + """Source: https://stackoverflow.com/a/54128391""" + + def __init__(self, *args, **kwargs): + self.__include_dirs = [] + super().__init__(*args, **kwargs) + + @property + def include_dirs(self): + import numpy + + return self.__include_dirs + [numpy.get_include()] + + @include_dirs.setter + def include_dirs(self, dirs): + self.__include_dirs = dirs + + +extensions = [ + Extension( + "fairseq.libbleu", + sources=[ + "fairseq/clib/libbleu/libbleu.cpp", + "fairseq/clib/libbleu/module.cpp", + ], + extra_compile_args=extra_compile_args, + ), + NumpyExtension( + "fairseq.data.data_utils_fast", + sources=["fairseq/data/data_utils_fast.pyx"], + language="c++", + extra_compile_args=extra_compile_args, + ), + NumpyExtension( + "fairseq.data.token_block_utils_fast", + sources=["fairseq/data/token_block_utils_fast.pyx"], + language="c++", + extra_compile_args=extra_compile_args, + ), +] + + +cmdclass = {} + + +try: + # torch is not available when generating docs + from torch.utils import cpp_extension + + extensions.extend( + [ + cpp_extension.CppExtension( + "fairseq.libbase", + sources=[ + "fairseq/clib/libbase/balanced_assignment.cpp", + ], + ) + ] + ) + + extensions.extend( + [ + cpp_extension.CppExtension( + "fairseq.libnat", + sources=[ + "fairseq/clib/libnat/edit_dist.cpp", + ], + ) + ] + ) + if "CUDA_HOME" in os.environ: + extensions.extend( + [ + cpp_extension.CppExtension( + "fairseq.libnat_cuda", + sources=[ + "fairseq/clib/libnat_cuda/edit_dist.cu", + "fairseq/clib/libnat_cuda/binding.cpp", + ], + ), + cpp_extension.CppExtension( + "fairseq.ngram_repeat_block_cuda", + sources=[ + "fairseq/clib/cuda/ngram_repeat_block_cuda.cpp", + "fairseq/clib/cuda/ngram_repeat_block_cuda_kernel.cu", + ], + ), + ] + ) + cmdclass["build_ext"] = cpp_extension.BuildExtension + +except ImportError: + pass + + +if "READTHEDOCS" in os.environ: + # don't build extensions when generating docs + extensions = [] + if "build_ext" in cmdclass: + del cmdclass["build_ext"] + + # use CPU build of PyTorch + dependency_links = [ + "https://download.pytorch.org/whl/cpu/torch-1.7.0%2Bcpu-cp36-cp36m-linux_x86_64.whl" + ] +else: + dependency_links = [] + + +if "clean" in sys.argv[1:]: + # Source: https://bit.ly/2NLVsgE + print("deleting Cython files...") + import subprocess + + subprocess.run( + ["rm -f fairseq/*.so fairseq/**/*.so fairseq/*.pyd fairseq/**/*.pyd"], + shell=True, + ) + + +extra_packages = [] +if os.path.exists(os.path.join("fairseq", "model_parallel", "megatron", "mpu")): + extra_packages.append("fairseq.model_parallel.megatron.mpu") + + +def do_setup(package_data): + setup( + name="fairseq", + version=version, + description="Facebook AI Research Sequence-to-Sequence Toolkit", + url="https://github.com/pytorch/fairseq", + classifiers=[ + "Intended Audience :: Science/Research", + "License :: OSI Approved :: MIT License", + "Programming Language :: Python :: 3.6", + "Programming Language :: Python :: 3.7", + "Programming Language :: Python :: 3.8", + "Topic :: Scientific/Engineering :: Artificial Intelligence", + ], + long_description=readme, + long_description_content_type="text/markdown", + setup_requires=[ + "cython", + 'numpy<1.20.0; python_version<"3.7"', + 'numpy; python_version>="3.7"', + "setuptools>=18.0", + ], + install_requires=[ + "cffi", + "cython", + 'dataclasses; python_version<"3.7"', + "hydra-core<1.1", + "omegaconf<2.1", + 'numpy<1.20.0; python_version<"3.7"', + 'numpy; python_version>="3.7"', + "regex", + "sacrebleu>=1.4.12", + "torch", + "tqdm", + ], + dependency_links=dependency_links, + packages=find_packages( + exclude=[ + "examples", + "examples.*", + "scripts", + "scripts.*", + "tests", + "tests.*", + ] + ) + + extra_packages, + package_data=package_data, + ext_modules=extensions, + test_suite="tests", + entry_points={ + "console_scripts": [ + "fairseq-eval-lm = fairseq_cli.eval_lm:cli_main", + "fairseq-generate = fairseq_cli.generate:cli_main", + "fairseq-hydra-train = fairseq_cli.hydra_train:cli_main", + "fairseq-interactive = fairseq_cli.interactive:cli_main", + "fairseq-preprocess = fairseq_cli.preprocess:cli_main", + "fairseq-score = fairseq_cli.score:cli_main", + "fairseq-train = fairseq_cli.train:cli_main", + "fairseq-validate = fairseq_cli.validate:cli_main", + ], + }, + cmdclass=cmdclass, + zip_safe=False, + ) + + +def get_files(path, relative_to="fairseq"): + all_files = [] + for root, _dirs, files in os.walk(path, followlinks=True): + root = os.path.relpath(root, relative_to) + for file in files: + if file.endswith(".pyc"): + continue + all_files.append(os.path.join(root, file)) + return all_files + + +if __name__ == "__main__": + try: + # symlink examples into fairseq package so package_data accepts them + fairseq_examples = os.path.join("fairseq", "examples") + if "build_ext" not in sys.argv[1:] and not os.path.exists(fairseq_examples): + os.symlink(os.path.join("..", "examples"), fairseq_examples) + + package_data = { + "fairseq": ( + get_files(fairseq_examples) + get_files(os.path.join("fairseq", "config")) + ) + } + do_setup(package_data) + finally: + if "build_ext" not in sys.argv[1:] and os.path.islink(fairseq_examples): + os.unlink(fairseq_examples) diff --git a/tests/__init__.py b/tests/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/tests/distributed/__init__.py b/tests/distributed/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/tests/distributed/test_bmuf.py b/tests/distributed/test_bmuf.py new file mode 100644 index 0000000000000000000000000000000000000000..8b7cadb094d49587b6b82432248459fdcf42457e --- /dev/null +++ b/tests/distributed/test_bmuf.py @@ -0,0 +1,207 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import functools +import random +import unittest +from multiprocessing import Manager + +import torch +import torch.nn as nn +from fairseq import optim +from fairseq.distributed import utils as distributed_utils +from omegaconf import OmegaConf + + +class Model(nn.Module): + def __init__(self, input_size, output_size): + super(Model, self).__init__() + self.fc = nn.Linear(input_size, output_size) + + def forward(self, input): + output = self.fc(input) + return output + + +def setup_model_loss_criterion(cfg, args, rank, is_cuda): + """ + setup model, criterion and optimizer based on input args + """ + args.distributed_rank = rank + cfg.distributed_training.distributed_rank = args.distributed_rank + if cfg.distributed_training.distributed_world_size > 1: + distributed_utils.distributed_init(cfg) + torch.manual_seed(1) + model = Model(args.input_size, args.nb_classes) + loss_fn = nn.CrossEntropyLoss() + if is_cuda: + model = model.cuda() + loss_fn = loss_fn.cuda() + + optimizer = optim.sgd.SGD(args, model.parameters()) + optimizer = optim.FairseqBMUF( + cfg=cfg.bmuf, + optimizer=optimizer + ) + + return model, loss_fn, optimizer + + +def train_step(input, target, model, loss_fn, optimizer, **unused): + """Do forward, backward and parameter update.""" + model.train() + output = model(input) + loss = loss_fn(output, target) + optimizer.backward(loss) + optimizer.step() + + +def single_gpu_training(cfg, args, rank, iterations, shared_results): + + is_cuda = torch.cuda.is_available() + if is_cuda: + torch.cuda.set_device(rank) + + model, loss_fn, optimizer = setup_model_loss_criterion(cfg, args, rank, is_cuda) + + for _ in range(iterations): + input = torch.randn(1, args.input_size) + target = torch.empty(args.batch_size, dtype=torch.long).random_(args.nb_classes) + + if is_cuda: + input = input.cuda() + target = target.cuda() + train_step(input, target, model, loss_fn, optimizer) + + results = [] + for param in model.parameters(): + if len(results) == 0: + results = param.flatten().cpu().data + else: + results = torch.cat((results, param.flatten().cpu().data), 0) + + shared_results[rank] = results + + +def setup_args(): + args = argparse.Namespace() + args.global_sync_iter = 20 + args.block_momentum = 0.875 + args.block_lr = 0.5 + args.input_size = 5 + args.nb_classes = 2 + args.batch_size = 1 + args.lr = [1e-3] + args.momentum = 0 + args.weight_decay = 0 + args.warmup_iterations = 0 + args.use_nbm = True + args.average_sync = True + args.global_sync_iter = 1 + args.model_parallel_size = 1 + args.distributed_backend = "gloo" + + args.distributed_world_size = 2 + port = random.randint(10000, 20000) + args.distributed_init_method = "tcp://localhost:{port}".format(port=port) + args.distributed_init_host = "localhost" + args.distributed_port = port + 1 + args.local_world_size = args.distributed_world_size + + cfg = OmegaConf.create() + cfg.optimization = OmegaConf.create() + cfg.common = OmegaConf.create() + cfg.distributed_training = OmegaConf.create() + cfg.dataset = OmegaConf.create() + cfg.bmuf = OmegaConf.create() + cfg.optimizer = OmegaConf.create() + + cfg.bmuf.global_sync_iter = args.global_sync_iter + cfg.bmuf.block_momentum = args.block_momentum + cfg.bmuf.block_lr = args.block_lr + cfg.dataset.batch_size = args.batch_size + cfg.optimization.lr = args.lr + cfg.optimizer.momentum = args.momentum + cfg.optimizer.weight_decay = args.weight_decay + cfg.bmuf.warmup_iterations = args.warmup_iterations + cfg.bmuf.use_nbm = args.use_nbm + cfg.bmuf.average_sync = args.average_sync + cfg.common.model_parallel_size = args.model_parallel_size + cfg.distributed_training.distributed_backend = args.distributed_backend + cfg.distributed_training.distributed_world_size = args.distributed_world_size + cfg.bmuf.distributed_world_size = args.distributed_world_size + cfg.distributed_training.distributed_init_method = args.distributed_init_method + cfg.distributed_training.distributed_port = args.distributed_port + + return cfg, args + + +@unittest.skipIf(torch.cuda.device_count() < 2, "test requires 2 GPUs") +class TestBMUF(unittest.TestCase): + def bmuf_process(self, cfg, args, iterations): + processes = [] + results = Manager().dict() + torch.multiprocessing.spawn( + fn=functools.partial(single_gpu_training, cfg, args), + args=(iterations, results), + nprocs=args.distributed_world_size, + join=True, + ) + return results + + def test_bmuf_sync(self): + # Train model for 1 iteration and do bmuf sync without doing warmup + cfg, args = setup_args() + iterations = 1 + results = self.bmuf_process(cfg, args, iterations) + # Make sure params in both machines are same + assert len(results) == 2 + self.assertAlmostEqual(results[0], results[1]) + + def test_warmup_sync(self): + # Train model for 20 iteration and do warmup sync without doing bmuf sync + cfg, args = setup_args() + args.warmup_iterations = 20 + cfg.bmuf.warmup_iterations = args.warmup_iterations + iterations = 20 + results = self.bmuf_process(cfg, args, iterations) + # Make sure params in both machines are same + assert len(results) == 2 + self.assertAlmostEqual(results[0], results[1]) + + def test_warmup_sync_bmuf_sync(self): + # Train model for 25 iteration and do warmup sync after 20 iteration + # and bmuf sync after 25 iteration + cfg, args = setup_args() + args.warmup_iterations = 20 + args.global_sync_iter = 5 + cfg.bmuf.warmup_iterations = args.warmup_iterations + cfg.bmuf.global_sync_iter = args.global_sync_iter + iterations = 25 + results = self.bmuf_process(cfg, args, iterations) + # Make sure params in both machines are same + assert len(results) == 2 + self.assertAlmostEqual(results[0], results[1]) + + def test_single_gpu_bmuf(self): + # Train model for 5 iterations and use GPU 1 + cfg, args = setup_args() + args.distributed_world_size = 1 + args.warmup_iterations = 5 + cfg.distributed_training.distributed_world_size = args.distributed_world_size + cfg.bmuf.distributed_world_size = args.distributed_world_size + cfg.bmuf.warmup_iterations = args.warmup_iterations + iterations = 20 + results = self.bmuf_process(cfg, args, iterations) + assert len(results) == 1 + + def assertAlmostEqual(self, t1, t2): + self.assertEqual(t1.size(), t2.size(), "size mismatch") + self.assertLess((t1 - t2).abs().max(), 1e-4) + + +if __name__ == "__main__": + unittest.main() diff --git a/tests/distributed/test_distributed_timeout_wrapper.py b/tests/distributed/test_distributed_timeout_wrapper.py new file mode 100644 index 0000000000000000000000000000000000000000..27908b9d3f7d6d880351e2a12effb12f9bc27971 --- /dev/null +++ b/tests/distributed/test_distributed_timeout_wrapper.py @@ -0,0 +1,54 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import signal +import time +import unittest + +import torch +from torch import nn + +from fairseq.distributed import DistributedTimeoutWrapper + + +class ModuleWithDelay(nn.Module): + + def __init__(self, delay): + super().__init__() + self.delay = delay + + def forward(self, x): + time.sleep(self.delay) + return x + + +class TestDistributedTimeoutWrapper(unittest.TestCase): + + def setUp(self): + logging.disable(logging.CRITICAL) + + def tearDown(self): + logging.disable(logging.NOTSET) + + def test_no_timeout(self): + module = DistributedTimeoutWrapper(ModuleWithDelay(1), 0, signal.SIGINT) + module(torch.rand(5)) + module.stop_timeout() + + def test_timeout_safe(self): + module = DistributedTimeoutWrapper(ModuleWithDelay(1), 10, signal.SIGINT) + module(torch.rand(5)) + module.stop_timeout() + + def test_timeout_killed(self): + with self.assertRaises(KeyboardInterrupt): + module = DistributedTimeoutWrapper(ModuleWithDelay(5), 1, signal.SIGINT) + module(torch.rand(5)) + module.stop_timeout() + + +if __name__ == "__main__": + unittest.main() diff --git a/tests/distributed/test_module_proxy_wrapper.py b/tests/distributed/test_module_proxy_wrapper.py new file mode 100644 index 0000000000000000000000000000000000000000..2803a044cdcc12e0a348f40d06ce89c571d307ed --- /dev/null +++ b/tests/distributed/test_module_proxy_wrapper.py @@ -0,0 +1,75 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import unittest + +import torch +from torch import nn + +from fairseq.distributed import ModuleProxyWrapper + +from .utils import objects_are_equal + + +class MockDDPWrapper(nn.Module): + """A simple wrapper with an interface similar to DistributedDataParallel.""" + + def __init__(self, module): + super().__init__() + self.module = module + + def forward(self, x): + return self.module(x) + + +class Model(nn.Module): + def __init__(self): + super().__init__() + self.linear = nn.Linear(5, 10) + self.xyz = "hello" + + def forward(self, x): + return self.linear(x) + + def get_xyz(self): + return self.xyz + + +class TestModuleProxyWrapper(unittest.TestCase): + + def _get_module(self): + module = Model() + wrapped_module = MockDDPWrapper(module) + wrapped_module = ModuleProxyWrapper(wrapped_module) + return wrapped_module, module + + def test_getattr_forwarding(self): + wrapped_module, module = self._get_module() + assert module.xyz == "hello" + assert module.get_xyz() == "hello" + assert wrapped_module.xyz == "hello" + + wrapped_module.xyz = "world" + assert wrapped_module.xyz == "world" + assert module.get_xyz() == "hello" + + def test_state_dict(self): + wrapped_module, module = self._get_module() + assert objects_are_equal(wrapped_module.state_dict(), module.state_dict()) + + def test_load_state_dict(self): + wrapped_module, module = self._get_module() + wrapped_module.load_state_dict(module.state_dict()) + input = torch.rand(4, 5) + torch.testing.assert_allclose(wrapped_module(input), module(input)) + + def test_forward(self): + wrapped_module, module = self._get_module() + input = torch.rand(4, 5) + torch.testing.assert_allclose(wrapped_module(input), module(input)) + + +if __name__ == "__main__": + unittest.main() diff --git a/tests/distributed/test_utils.py b/tests/distributed/test_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..30f995b67acd39af5816d2eb412d6b4df7f44f8c --- /dev/null +++ b/tests/distributed/test_utils.py @@ -0,0 +1,124 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import functools +import sys +import unittest + +import torch + +from fairseq.distributed import utils as dist_utils + +from .utils import objects_are_equal, spawn_and_init + + +class DistributedTest(unittest.TestCase): + def setUp(self): + if not torch.cuda.is_available(): + raise unittest.SkipTest("CUDA not available, skipping test") + if sys.platform == "win32": + raise unittest.SkipTest("NCCL doesn't support Windows, skipping test") + if torch.cuda.device_count() < 2: + raise unittest.SkipTest("distributed tests require 2+ GPUs, skipping") + + +class TestBroadcastObject(DistributedTest): + def test_str(self): + spawn_and_init( + functools.partial( + TestBroadcastObject._test_broadcast_object, "hello world" + ), + world_size=2, + ) + + def test_tensor(self): + spawn_and_init( + functools.partial( + TestBroadcastObject._test_broadcast_object, + torch.rand(5), + ), + world_size=2, + ) + + def test_complex(self): + spawn_and_init( + functools.partial( + TestBroadcastObject._test_broadcast_object, + { + "a": "1", + "b": [2, torch.rand(2, 3), 3], + "c": (torch.rand(2, 3), 4), + "d": {5, torch.rand(5)}, + "e": torch.rand(5), + "f": torch.rand(5).int().cuda(), + }, + ), + world_size=2, + ) + + @staticmethod + def _test_broadcast_object(ref_obj, rank, group): + obj = dist_utils.broadcast_object( + ref_obj if rank == 0 else None, src_rank=0, group=group + ) + assert objects_are_equal(ref_obj, obj) + + +class TestAllGatherList(DistributedTest): + def test_str_equality(self): + spawn_and_init( + functools.partial( + TestAllGatherList._test_all_gather_list_equality, + "hello world", + ), + world_size=2, + ) + + def test_tensor_equality(self): + spawn_and_init( + functools.partial( + TestAllGatherList._test_all_gather_list_equality, + torch.rand(5), + ), + world_size=2, + ) + + def test_complex_equality(self): + spawn_and_init( + functools.partial( + TestAllGatherList._test_all_gather_list_equality, + { + "a": "1", + "b": [2, torch.rand(2, 3), 3], + "c": (torch.rand(2, 3), 4), + "d": {5, torch.rand(5)}, + "e": torch.rand(5), + "f": torch.rand(5).int(), + }, + ), + world_size=2, + ) + + @staticmethod + def _test_all_gather_list_equality(ref_obj, rank, group): + objs = dist_utils.all_gather_list(ref_obj, group) + for obj in objs: + assert objects_are_equal(ref_obj, obj) + + def test_rank_tensor(self): + spawn_and_init( + TestAllGatherList._test_all_gather_list_rank_tensor, world_size=2 + ) + + @staticmethod + def _test_all_gather_list_rank_tensor(rank, group): + obj = torch.tensor([rank]) + objs = dist_utils.all_gather_list(obj, group) + for i, obj in enumerate(objs): + assert obj.item() == i + + +if __name__ == "__main__": + unittest.main() diff --git a/tests/distributed/utils.py b/tests/distributed/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..c8040392a8e27eb4c3a74032c702643a91d11a3e --- /dev/null +++ b/tests/distributed/utils.py @@ -0,0 +1,62 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import functools +import tempfile + +import torch + + +def spawn_and_init(fn, world_size, args=None): + if args is None: + args = () + with tempfile.NamedTemporaryFile(delete=False) as tmp_file: + torch.multiprocessing.spawn( + fn=functools.partial(init_and_run, fn, args), + args=(world_size, tmp_file.name,), + nprocs=world_size, + join=True, + ) + + +def distributed_init(rank, world_size, tmp_file): + torch.distributed.init_process_group( + backend="nccl", + init_method="file://{}".format(tmp_file), + world_size=world_size, + rank=rank, + ) + torch.cuda.set_device(rank) + + +def init_and_run(fn, args, rank, world_size, tmp_file): + distributed_init(rank, world_size, tmp_file) + group = torch.distributed.new_group() + fn(rank, group, *args) + + +def objects_are_equal(a, b) -> bool: + if type(a) is not type(b): + return False + if isinstance(a, dict): + if set(a.keys()) != set(b.keys()): + return False + for k in a.keys(): + if not objects_are_equal(a[k], b[k]): + return False + return True + elif isinstance(a, (list, tuple, set)): + if len(a) != len(b): + return False + return all(objects_are_equal(x, y) for x, y in zip(a, b)) + elif torch.is_tensor(a): + return ( + a.size() == b.size() + and a.dtype == b.dtype + and a.device == b.device + and torch.all(a == b) + ) + else: + return a == b diff --git a/tests/gpu/__init__.py b/tests/gpu/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/tests/gpu/test_binaries_gpu.py b/tests/gpu/test_binaries_gpu.py new file mode 100644 index 0000000000000000000000000000000000000000..de8c2426134089035c6e0e5da223647bab6f3dba --- /dev/null +++ b/tests/gpu/test_binaries_gpu.py @@ -0,0 +1,449 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import contextlib +import logging +import json +import os +import tempfile +import unittest +from io import StringIO + +import torch +from fairseq import options +from fairseq_cli import train +from tests.utils import ( + create_dummy_data, + generate_main, + preprocess_lm_data, + preprocess_translation_data, + train_translation_model, +) + + +@unittest.skipIf(not torch.cuda.is_available(), "test requires a GPU") +class TestTranslationGPU(unittest.TestCase): + def setUp(self): + logging.disable(logging.CRITICAL) + + def tearDown(self): + logging.disable(logging.NOTSET) + + def test_fp16_multigpu(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_fp16") as data_dir: + log = os.path.join(data_dir, "train.log") + create_dummy_data(data_dir) + preprocess_translation_data(data_dir) + train_translation_model( + data_dir, + "fconv_iwslt_de_en", + ["--fp16", "--log-file", log], + world_size=min(torch.cuda.device_count(), 2), + ) + generate_main(data_dir) + assert os.path.exists(log) + + @staticmethod + def parse_logs(logfile): + logs = [] + for ln in open(logfile, "r").readlines(): + try: + logs.append(json.loads(ln)) + except json.JSONDecodeError: + continue + return logs + + def test_resume_training_fsdp(self): + self._test_resume_training(["--ddp-backend", "fully_sharded"]) + + def test_resume_training_fsdp_sharded_state(self): + self._test_resume_training(["--ddp-backend", "fully_sharded", "--use-sharded-state"]) + + def test_resume_training_noc10d(self): + self._test_resume_training([]) + + def _test_resume_training(self, extra_clargs, arch="fconv_iwslt_de_en"): + flags = [ + "--fp16", + "--log-format", + "json", + "--max-update", + "10", + "--save-interval-updates", + "2", + "--log-interval", + "1", + ] + extra_clargs + world_size = min(torch.cuda.device_count(), 2) + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_fp16") as data_dir: + log = os.path.join(data_dir, "train.log") + create_dummy_data(data_dir) + preprocess_translation_data(data_dir) + train_translation_model( + data_dir, arch, flags + ["--log-file", log], world_size=world_size, + ) + log2 = os.path.join(data_dir, "resume.log") + restore_file = os.path.join(data_dir, "checkpoint_1_2.pt") + train_translation_model( + data_dir, + arch, + flags + ["--log-file", log2, "--restore-file", restore_file], + world_size=world_size, + ) + + l1 = self.parse_logs(log) + l2 = self.parse_logs(log2) + assert int(l2[0]["num_updates"]) == 3, f"{l1}\n\n {l2}" + for k in [ + "train_loss", + "train_num_updates", + "train_ppl", + "train_gnorm", + ]: + from_scratch, resumed = l1[-1][k], l2[-1][k] + assert ( + from_scratch == resumed + ), f"difference at {k} {from_scratch} != {resumed}" + + def test_memory_efficient_fp16(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_memory_efficient_fp16") as data_dir: + create_dummy_data(data_dir) + preprocess_translation_data(data_dir) + train_translation_model( + data_dir, "fconv_iwslt_de_en", ["--memory-efficient-fp16"] + ) + generate_main(data_dir) + + def test_transformer_fp16(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_transformer") as data_dir: + create_dummy_data(data_dir) + preprocess_translation_data(data_dir) + train_translation_model( + data_dir, + "transformer_iwslt_de_en", + [ + "--encoder-layers", + "2", + "--decoder-layers", + "2", + "--encoder-embed-dim", + "64", + "--decoder-embed-dim", + "64", + "--fp16", + ], + run_validation=True, + ) + generate_main(data_dir) + + @unittest.skipIf(not torch.cuda.is_available(), "test requires a GPU") + def test_amp(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_amp") as data_dir: + create_dummy_data(data_dir) + preprocess_translation_data(data_dir) + train_translation_model(data_dir, "fconv_iwslt_de_en", ["--amp"]) + generate_main(data_dir) + + @unittest.skipIf(not torch.cuda.is_available(), "test requires a GPU") + def test_transformer_amp(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_transformer") as data_dir: + create_dummy_data(data_dir) + preprocess_translation_data(data_dir) + train_translation_model( + data_dir, + "transformer_iwslt_de_en", + [ + "--encoder-layers", + "2", + "--decoder-layers", + "2", + "--encoder-embed-dim", + "64", + "--decoder-embed-dim", + "64", + "--amp", + ], + run_validation=True, + ) + generate_main(data_dir) + + @unittest.skipIf(not torch.cuda.is_available(), "test requires a GPU") + def test_levenshtein_transformer(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory( + "test_levenshtein_transformer" + ) as data_dir: + create_dummy_data(data_dir) + preprocess_translation_data(data_dir, ["--joined-dictionary"]) + train_translation_model( + data_dir, + "levenshtein_transformer", + [ + "--apply-bert-init", + "--early-exit", + "6,6,6", + "--criterion", + "nat_loss", + ], + task="translation_lev", + ) + gen_config = [ + "--task", + "translation_lev", + "--iter-decode-max-iter", + "9", + "--iter-decode-eos-penalty", + "0", + "--print-step", + ] + # non-ensemble generation + generate_main(data_dir, gen_config) + # ensemble generation + generate_main( + data_dir, + gen_config, + path=os.pathsep.join( + [ + os.path.join(data_dir, "checkpoint_last.pt"), + os.path.join(data_dir, "checkpoint_last.pt"), + ] + ), + ) + + def test_fsdp_checkpoint_generate(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_fsdp_sharded") as data_dir: + log = os.path.join(data_dir, "train.log") + create_dummy_data(data_dir) + preprocess_translation_data(data_dir) + world_size = min(torch.cuda.device_count(), 2) + train_translation_model( + data_dir, + "fconv_iwslt_de_en", + ["--log-file", log, "--ddp-backend", "fully_sharded"], + world_size=world_size, + ) + generate_main(data_dir) + assert os.path.exists(log) + + def test_fsdp_sharded_checkpoint_generate(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_fsdp_sharded") as data_dir: + log = os.path.join(data_dir, "train.log") + create_dummy_data(data_dir) + preprocess_translation_data(data_dir) + world_size = min(torch.cuda.device_count(), 2) + train_translation_model( + data_dir, + "fconv_iwslt_de_en", + ["--log-file", log, "--ddp-backend", "fully_sharded", "--use-sharded-state"], + world_size=world_size, + ) + generate_main(data_dir, ["--checkpoint-shard-count", str(world_size)]) + assert os.path.exists(log) + + +def _quantize_language_model(data_dir, arch, extra_flags=None, run_validation=False): + train_parser = options.get_training_parser() + train_args = options.parse_args_and_arch( + train_parser, + [ + "--task", + "language_modeling", + data_dir, + "--arch", + arch, + "--optimizer", + "adam", + "--lr", + "0.0001", + "--criterion", + "adaptive_loss", + "--adaptive-softmax-cutoff", + "5,10,15", + "--max-tokens", + "500", + "--tokens-per-sample", + "500", + "--save-dir", + data_dir, + "--max-epoch", + "1", + "--no-progress-bar", + "--distributed-world-size", + "1", + "--ddp-backend", + "no_c10d", + "--num-workers", + "0", + ] + + (extra_flags or []), + ) + train.main(train_args) + + # try scalar quantization + scalar_quant_train_parser = options.get_training_parser() + scalar_quant_train_args = options.parse_args_and_arch( + scalar_quant_train_parser, + [ + "--task", + "language_modeling", + data_dir, + "--arch", + arch, + "--optimizer", + "adam", + "--lr", + "0.0001", + "--criterion", + "adaptive_loss", + "--adaptive-softmax-cutoff", + "5,10,15", + "--max-tokens", + "500", + "--tokens-per-sample", + "500", + "--save-dir", + data_dir, + "--max-update", + "3", + "--no-progress-bar", + "--distributed-world-size", + "1", + "--ddp-backend", + "no_c10d", + "--num-workers", + "0", + "--quant-noise-scalar", + "0.5", + ] + + (extra_flags or []), + ) + train.main(scalar_quant_train_args) + + # try iterative PQ quantization + quantize_parser = options.get_training_parser() + quantize_args = options.parse_args_and_arch( + quantize_parser, + [ + "--task", + "language_modeling", + data_dir, + "--arch", + arch, + "--optimizer", + "adam", + "--lr", + "0.0001", + "--criterion", + "adaptive_loss", + "--adaptive-softmax-cutoff", + "5,10,15", + "--max-tokens", + "50", + "--tokens-per-sample", + "50", + "--max-update", + "6", + "--no-progress-bar", + "--distributed-world-size", + "1", + "--ddp-backend", + "no_c10d", + "--num-workers", + "0", + "--restore-file", + os.path.join(data_dir, "checkpoint_last.pt"), + "--reset-optimizer", + "--quantization-config-path", + os.path.join( + os.path.dirname(__file__), "transformer_quantization_config.yaml" + ), + ] + + (extra_flags or []), + ) + train.main(quantize_args) + + +@unittest.skipIf(not torch.cuda.is_available(), "test requires a GPU") +class TestQuantization(unittest.TestCase): + def setUp(self): + logging.disable(logging.CRITICAL) + + def tearDown(self): + logging.disable(logging.NOTSET) + + def test_quantization(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_quantization") as data_dir: + create_dummy_data(data_dir) + preprocess_lm_data(data_dir) + # tests both scalar and iterative PQ quantization + _quantize_language_model(data_dir, "transformer_lm") + + +@unittest.skipIf(not torch.cuda.is_available(), "test requires a GPU") +class TestOptimizersGPU(unittest.TestCase): + def setUp(self): + logging.disable(logging.CRITICAL) + + def tearDown(self): + logging.disable(logging.NOTSET) + + def test_flat_grads(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_flat_grads") as data_dir: + # Use just a bit of data and tiny model to keep this test runtime reasonable + create_dummy_data(data_dir, num_examples=10, maxlen=5) + preprocess_translation_data(data_dir) + with self.assertRaises(RuntimeError): + # adafactor isn't compatible with flat grads, which + # are used by default with --fp16 + train_translation_model( + data_dir, + "lstm", + [ + "--required-batch-size-multiple", + "1", + "--encoder-layers", + "1", + "--encoder-hidden-size", + "32", + "--decoder-layers", + "1", + "--optimizer", + "adafactor", + "--fp16", + ], + ) + # but it should pass once we set --fp16-no-flatten-grads + train_translation_model( + data_dir, + "lstm", + [ + "--required-batch-size-multiple", + "1", + "--encoder-layers", + "1", + "--encoder-hidden-size", + "32", + "--decoder-layers", + "1", + "--optimizer", + "adafactor", + "--fp16", + "--fp16-no-flatten-grads", + ], + ) + + +if __name__ == "__main__": + unittest.main() diff --git a/tests/gpu/transformer_quantization_config.yaml b/tests/gpu/transformer_quantization_config.yaml new file mode 100644 index 0000000000000000000000000000000000000000..de31d8116ced675b81eb74119642217d768e7736 --- /dev/null +++ b/tests/gpu/transformer_quantization_config.yaml @@ -0,0 +1,28 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +# This file defines example configuration arguments for quantizing +# a transformer model with product quantization + +n_centroids: + Linear: + key: in_features + value: {"*": 8} + Embedding: + key: embedding_dim + value: {"*": 8} + +block_sizes: + Linear: + key: fuzzy_name + value: {fc: 8, attn: 4, emb: 4} + Embedding: + key: fuzzy_name + value: {emb: 8} + +layers_to_quantize: + - decoder\\.layers\\.\d+\\.fc[12] + - decoder\\.embed_tokens\\.embeddings\\.[012]\\.[01] + - decoder\\.layers\\.\d+\\.self_attn\\.(k_proj|v_proj|q_proj|out_proj) diff --git a/tests/speech_recognition/__init__.py b/tests/speech_recognition/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/tests/speech_recognition/asr_test_base.py b/tests/speech_recognition/asr_test_base.py new file mode 100644 index 0000000000000000000000000000000000000000..8c5d414e7bf17ee02f280d024fa5d07e28b79d6b --- /dev/null +++ b/tests/speech_recognition/asr_test_base.py @@ -0,0 +1,557 @@ +#!/usr/bin/env python3 + +import argparse +import os +import unittest +from inspect import currentframe, getframeinfo + +import numpy as np +import torch +from examples.speech_recognition.data.data_utils import lengths_to_encoder_padding_mask +from fairseq.data import data_utils as fairseq_data_utils +from fairseq.data.dictionary import Dictionary +from fairseq.models import ( + BaseFairseqModel, + FairseqDecoder, + FairseqEncoder, + FairseqEncoderDecoderModel, + FairseqEncoderModel, + FairseqModel, +) +from fairseq.tasks.fairseq_task import LegacyFairseqTask + + +DEFAULT_TEST_VOCAB_SIZE = 100 + + +# /////////////////////////////////////////////////////////////////////////// +# utility function to setup dummy dict/task/input +# /////////////////////////////////////////////////////////////////////////// + + +def get_dummy_dictionary(vocab_size=DEFAULT_TEST_VOCAB_SIZE): + dummy_dict = Dictionary() + # add dummy symbol to satisfy vocab size + for id, _ in enumerate(range(vocab_size)): + dummy_dict.add_symbol("{}".format(id), 1000) + return dummy_dict + + +class DummyTask(LegacyFairseqTask): + def __init__(self, args): + super().__init__(args) + self.dictionary = get_dummy_dictionary() + if getattr(self.args, "ctc", False): + self.dictionary.add_symbol("<ctc_blank>") + self.tgt_dict = self.dictionary + + @property + def target_dictionary(self): + return self.dictionary + + +def get_dummy_task_and_parser(): + """ + to build a fariseq model, we need some dummy parse and task. This function + is used to create dummy task and parser to faciliate model/criterion test + + Note: we use FbSpeechRecognitionTask as the dummy task. You may want + to use other task by providing another function + """ + parser = argparse.ArgumentParser( + description="test_dummy_s2s_task", argument_default=argparse.SUPPRESS + ) + DummyTask.add_args(parser) + args = parser.parse_args([]) + task = DummyTask.setup_task(args) + return task, parser + + +def get_dummy_input(T=100, D=80, B=5, K=100): + forward_input = {} + # T max sequence length + # D feature vector dimension + # B batch size + # K target dimension size + feature = torch.randn(B, T, D) + # this (B, T, D) layout is just a convention, you can override it by + # write your own _prepare_forward_input function + src_lengths = torch.from_numpy( + np.random.randint(low=1, high=T, size=B, dtype=np.int64) + ) + src_lengths[0] = T # make sure the maximum length matches + prev_output_tokens = [] + for b in range(B): + token_length = np.random.randint(low=1, high=src_lengths[b].item() + 1) + tokens = np.random.randint(low=0, high=K, size=token_length, dtype=np.int64) + prev_output_tokens.append(torch.from_numpy(tokens)) + + prev_output_tokens = fairseq_data_utils.collate_tokens( + prev_output_tokens, + pad_idx=1, + eos_idx=2, + left_pad=False, + move_eos_to_beginning=False, + ) + src_lengths, sorted_order = src_lengths.sort(descending=True) + forward_input["src_tokens"] = feature.index_select(0, sorted_order) + forward_input["src_lengths"] = src_lengths + forward_input["prev_output_tokens"] = prev_output_tokens + + return forward_input + + +def get_dummy_encoder_output(encoder_out_shape=(100, 80, 5)): + """ + This only provides an example to generate dummy encoder output + """ + (T, B, D) = encoder_out_shape + encoder_out = {} + + encoder_out["encoder_out"] = torch.from_numpy( + np.random.randn(*encoder_out_shape).astype(np.float32) + ) + seq_lengths = torch.from_numpy(np.random.randint(low=1, high=T, size=B)) + # some dummy mask + encoder_out["encoder_padding_mask"] = torch.arange(T).view(1, T).expand( + B, -1 + ) >= seq_lengths.view(B, 1).expand(-1, T) + encoder_out["encoder_padding_mask"].t_() + + # encoer_padding_mask is (T, B) tensor, with (t, b)-th element indicate + # whether encoder_out[t, b] is valid (=0) or not (=1) + return encoder_out + + +def _current_postion_info(): + cf = currentframe() + frameinfo = " (at {}:{})".format( + os.path.basename(getframeinfo(cf).filename), cf.f_back.f_lineno + ) + return frameinfo + + +def check_encoder_output(encoder_output, batch_size=None): + """we expect encoder_output to be a dict with the following + key/value pairs: + - encoder_out: a Torch.Tensor + - encoder_padding_mask: a binary Torch.Tensor + """ + if not isinstance(encoder_output, dict): + msg = ( + "FairseqEncoderModel.forward(...) must be a dict" + _current_postion_info() + ) + return False, msg + + if "encoder_out" not in encoder_output: + msg = ( + "FairseqEncoderModel.forward(...) must contain encoder_out" + + _current_postion_info() + ) + return False, msg + + if "encoder_padding_mask" not in encoder_output: + msg = ( + "FairseqEncoderModel.forward(...) must contain encoder_padding_mask" + + _current_postion_info() + ) + return False, msg + + if not isinstance(encoder_output["encoder_out"], torch.Tensor): + msg = "encoder_out must be a torch.Tensor" + _current_postion_info() + return False, msg + + if encoder_output["encoder_out"].dtype != torch.float32: + msg = "encoder_out must have float32 dtype" + _current_postion_info() + return False, msg + + mask = encoder_output["encoder_padding_mask"] + if mask is not None: + if not isinstance(mask, torch.Tensor): + msg = ( + "encoder_padding_mask must be a torch.Tensor" + _current_postion_info() + ) + return False, msg + if mask.dtype != torch.uint8 and ( + not hasattr(torch, "bool") or mask.dtype != torch.bool + ): + msg = ( + "encoder_padding_mask must have dtype of uint8" + + _current_postion_info() + ) + return False, msg + + if mask.dim() != 2: + msg = ( + "we expect encoder_padding_mask to be a 2-d tensor, in shape (T, B)" + + _current_postion_info() + ) + return False, msg + + if batch_size is not None and mask.size(1) != batch_size: + msg = ( + "we expect encoder_padding_mask to be a 2-d tensor, with size(1)" + + " being the batch size" + + _current_postion_info() + ) + return False, msg + return True, None + + +def check_decoder_output(decoder_output): + """we expect output from a decoder is a tuple with the following constraint: + - the first element is a torch.Tensor + - the second element can be anything (reserved for future use) + """ + if not isinstance(decoder_output, tuple): + msg = "FariseqDecoder output must be a tuple" + _current_postion_info() + return False, msg + + if len(decoder_output) != 2: + msg = "FairseqDecoder output must be 2-elem tuple" + _current_postion_info() + return False, msg + + if not isinstance(decoder_output[0], torch.Tensor): + msg = ( + "FariseqDecoder output[0] must be a torch.Tensor" + _current_postion_info() + ) + return False, msg + + return True, None + + +# /////////////////////////////////////////////////////////////////////////// +# Base Test class +# /////////////////////////////////////////////////////////////////////////// + + +class TestBaseFairseqModelBase(unittest.TestCase): + """ + This class is used to facilitate writing unittest for any class derived from + `BaseFairseqModel`. + """ + + @classmethod + def setUpClass(cls): + if cls is TestBaseFairseqModelBase: + raise unittest.SkipTest("Skipping test case in base") + super().setUpClass() + + def setUpModel(self, model): + self.assertTrue(isinstance(model, BaseFairseqModel)) + self.model = model + + def setupInput(self): + pass + + def setUp(self): + self.model = None + self.forward_input = None + pass + + +class TestFairseqEncoderDecoderModelBase(TestBaseFairseqModelBase): + """ + base code to test FairseqEncoderDecoderModel (formally known as + `FairseqModel`) must be derived from this base class + """ + + @classmethod + def setUpClass(cls): + if cls is TestFairseqEncoderDecoderModelBase: + raise unittest.SkipTest("Skipping test case in base") + super().setUpClass() + + def setUpModel(self, model_cls, extra_args_setters=None): + self.assertTrue( + issubclass(model_cls, (FairseqEncoderDecoderModel, FairseqModel)), + msg="This class only tests for FairseqModel subclasses", + ) + + task, parser = get_dummy_task_and_parser() + model_cls.add_args(parser) + + args = parser.parse_args([]) + + if extra_args_setters is not None: + for args_setter in extra_args_setters: + args_setter(args) + model = model_cls.build_model(args, task) + self.model = model + + def setUpInput(self, input=None): + self.forward_input = get_dummy_input() if input is None else input + + def setUp(self): + super().setUp() + + def test_forward(self): + if self.model and self.forward_input: + forward_output = self.model.forward(**self.forward_input) + # for FairseqEncoderDecoderModel, forward returns a tuple of two + # elements, the first one is a Torch.Tensor + succ, msg = check_decoder_output(forward_output) + if not succ: + self.assertTrue(succ, msg=msg) + self.forward_output = forward_output + + def test_get_normalized_probs(self): + if self.model and self.forward_input: + forward_output = self.model.forward(**self.forward_input) + logprob = self.model.get_normalized_probs(forward_output, log_probs=True) + prob = self.model.get_normalized_probs(forward_output, log_probs=False) + + # in order for different models/criterion to play with each other + # we need to know whether the logprob or prob output is batch_first + # or not. We assume an additional attribute will be attached to logprob + # or prob. If you find your code failed here, simply override + # FairseqModel.get_normalized_probs, see example at + # https://fburl.com/batch_first_example + self.assertTrue(hasattr(logprob, "batch_first")) + self.assertTrue(hasattr(prob, "batch_first")) + + self.assertTrue(torch.is_tensor(logprob)) + self.assertTrue(torch.is_tensor(prob)) + + +class TestFairseqEncoderModelBase(TestBaseFairseqModelBase): + """ + base class to test FairseqEncoderModel + """ + + @classmethod + def setUpClass(cls): + if cls is TestFairseqEncoderModelBase: + raise unittest.SkipTest("Skipping test case in base") + super().setUpClass() + + def setUpModel(self, model_cls, extra_args_setters=None): + self.assertTrue( + issubclass(model_cls, FairseqEncoderModel), + msg="This class is only used for testing FairseqEncoderModel", + ) + task, parser = get_dummy_task_and_parser() + model_cls.add_args(parser) + args = parser.parse_args([]) + if extra_args_setters is not None: + for args_setter in extra_args_setters: + args_setter(args) + + model = model_cls.build_model(args, task) + self.model = model + + def setUpInput(self, input=None): + self.forward_input = get_dummy_input() if input is None else input + # get_dummy_input() is originally for s2s, here we delete extra dict + # items, so it can be used for EncoderModel / Encoder as well + self.forward_input.pop("prev_output_tokens", None) + + def setUp(self): + super().setUp() + + def test_forward(self): + if self.forward_input and self.model: + bsz = self.forward_input["src_tokens"].size(0) + forward_output = self.model.forward(**self.forward_input) + + # we expect forward_output to be a dict with the following + # key/value pairs: + # - encoder_out: a Torch.Tensor + # - encoder_padding_mask: a binary Torch.Tensor + succ, msg = check_encoder_output(forward_output, batch_size=bsz) + if not succ: + self.assertTrue(succ, msg=msg) + self.forward_output = forward_output + + def test_get_normalized_probs(self): + if self.model and self.forward_input: + forward_output = self.model.forward(**self.forward_input) + logprob = self.model.get_normalized_probs(forward_output, log_probs=True) + prob = self.model.get_normalized_probs(forward_output, log_probs=False) + + # in order for different models/criterion to play with each other + # we need to know whether the logprob or prob output is batch_first + # or not. We assume an additional attribute will be attached to logprob + # or prob. If you find your code failed here, simply override + # FairseqModel.get_normalized_probs, see example at + # https://fburl.com/batch_first_example + self.assertTrue(hasattr(logprob, "batch_first")) + self.assertTrue(hasattr(prob, "batch_first")) + + self.assertTrue(torch.is_tensor(logprob)) + self.assertTrue(torch.is_tensor(prob)) + + +class TestFairseqEncoderBase(unittest.TestCase): + """ + base class to test FairseqEncoder + """ + + @classmethod + def setUpClass(cls): + if cls is TestFairseqEncoderBase: + raise unittest.SkipTest("Skipping test case in base") + super().setUpClass() + + def setUpEncoder(self, encoder): + self.assertTrue( + isinstance(encoder, FairseqEncoder), + msg="This class is only used for test FairseqEncoder", + ) + self.encoder = encoder + + def setUpInput(self, input=None): + self.forward_input = get_dummy_input() if input is None else input + # get_dummy_input() is originally for s2s, here we delete extra dict + # items, so it can be used for EncoderModel / Encoder as well + self.forward_input.pop("prev_output_tokens", None) + + def setUp(self): + self.encoder = None + self.forward_input = None + + def test_forward(self): + if self.encoder and self.forward_input: + bsz = self.forward_input["src_tokens"].size(0) + + forward_output = self.encoder.forward(**self.forward_input) + succ, msg = check_encoder_output(forward_output, batch_size=bsz) + if not succ: + self.assertTrue(succ, msg=msg) + self.forward_output = forward_output + + +class TestFairseqDecoderBase(unittest.TestCase): + """ + base class to test FairseqDecoder + """ + + @classmethod + def setUpClass(cls): + if cls is TestFairseqDecoderBase: + raise unittest.SkipTest("Skipping test case in base") + super().setUpClass() + + def setUpDecoder(self, decoder): + self.assertTrue( + isinstance(decoder, FairseqDecoder), + msg="This class is only used for test FairseqDecoder", + ) + self.decoder = decoder + + def setUpInput(self, input=None): + self.forward_input = get_dummy_encoder_output() if input is None else input + + def setUpPrevOutputTokens(self, tokens=None): + if tokens is None: + self.encoder_input = get_dummy_input() + self.prev_output_tokens = self.encoder_input["prev_output_tokens"] + else: + self.prev_output_tokens = tokens + + def setUp(self): + self.decoder = None + self.forward_input = None + self.prev_output_tokens = None + + def test_forward(self): + if ( + self.decoder is not None + and self.forward_input is not None + and self.prev_output_tokens is not None + ): + forward_output = self.decoder.forward( + prev_output_tokens=self.prev_output_tokens, + encoder_out=self.forward_input, + ) + succ, msg = check_decoder_output(forward_output) + if not succ: + self.assertTrue(succ, msg=msg) + self.forward_input = forward_output + + +class DummyEncoderModel(FairseqEncoderModel): + def __init__(self, encoder): + super().__init__(encoder) + + @classmethod + def build_model(cls, args, task): + return cls(DummyEncoder()) + + def get_logits(self, net_output): + # Inverse of sigmoid to use with BinaryCrossEntropyWithLogitsCriterion as + # F.binary_cross_entropy_with_logits combines sigmoid and CE + return torch.log( + torch.div(net_output["encoder_out"], 1 - net_output["encoder_out"]) + ) + + def get_normalized_probs(self, net_output, log_probs, sample=None): + lprobs = super().get_normalized_probs(net_output, log_probs, sample=sample) + lprobs.batch_first = True + return lprobs + + +class DummyEncoder(FairseqEncoder): + def __init__(self): + super().__init__(None) + + def forward(self, src_tokens, src_lengths): + mask, max_len = lengths_to_encoder_padding_mask(src_lengths) + return {"encoder_out": src_tokens, "encoder_padding_mask": mask} + + +class CrossEntropyCriterionTestBase(unittest.TestCase): + @classmethod + def setUpClass(cls): + if cls is CrossEntropyCriterionTestBase: + raise unittest.SkipTest("Skipping base class test case") + super().setUpClass() + + def setUpArgs(self): + args = argparse.Namespace() + args.sentence_avg = False + args.threshold = 0.1 # to use with BinaryCrossEntropyWithLogitsCriterion + return args + + def setUp(self): + args = self.setUpArgs() + self.model = DummyEncoderModel(encoder=DummyEncoder()) + self.criterion = self.criterion_cls.build_criterion(args, task=DummyTask(args)) + + def get_src_tokens(self, correct_prediction, aggregate): + """ + correct_prediction: True if the net_output (src_tokens) should + predict the correct target + aggregate: True if the criterion expects net_output (src_tokens) + aggregated across time axis + """ + predicted_idx = 0 if correct_prediction else 1 + if aggregate: + src_tokens = torch.zeros((2, 2), dtype=torch.float) + for b in range(2): + src_tokens[b][predicted_idx] = 1.0 + else: + src_tokens = torch.zeros((2, 10, 2), dtype=torch.float) + for b in range(2): + for t in range(10): + src_tokens[b][t][predicted_idx] = 1.0 + return src_tokens + + def get_target(self, soft_target): + if soft_target: + target = torch.zeros((2, 2), dtype=torch.float) + for b in range(2): + target[b][0] = 1.0 + else: + target = torch.zeros((2, 10), dtype=torch.long) + return target + + def get_test_sample(self, correct, soft_target, aggregate): + src_tokens = self.get_src_tokens(correct, aggregate) + target = self.get_target(soft_target) + L = src_tokens.size(1) + return { + "net_input": {"src_tokens": src_tokens, "src_lengths": torch.tensor([L])}, + "target": target, + "ntokens": src_tokens.size(0) * src_tokens.size(1), + } diff --git a/tests/speech_recognition/test_collaters.py b/tests/speech_recognition/test_collaters.py new file mode 100644 index 0000000000000000000000000000000000000000..6a5029a48faea2426d7a0277655a2c7c08c1d16c --- /dev/null +++ b/tests/speech_recognition/test_collaters.py @@ -0,0 +1,58 @@ +#!/usr/bin/env python3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import unittest + +import numpy as np +import torch +from examples.speech_recognition.data.collaters import Seq2SeqCollater + + +class TestSeq2SeqCollator(unittest.TestCase): + def test_collate(self): + + eos_idx = 1 + pad_idx = 0 + collater = Seq2SeqCollater( + feature_index=0, label_index=1, pad_index=pad_idx, eos_index=eos_idx + ) + + # 2 frames in the first sample and 3 frames in the second one + frames1 = np.array([[7, 8], [9, 10]]) + frames2 = np.array([[1, 2], [3, 4], [5, 6]]) + target1 = np.array([4, 2, 3, eos_idx]) + target2 = np.array([3, 2, eos_idx]) + sample1 = {"id": 0, "data": [frames1, target1]} + sample2 = {"id": 1, "data": [frames2, target2]} + batch = collater.collate([sample1, sample2]) + + # collate sort inputs by frame's length before creating the batch + self.assertTensorEqual(batch["id"], torch.tensor([1, 0])) + self.assertEqual(batch["ntokens"], 7) + self.assertTensorEqual( + batch["net_input"]["src_tokens"], + torch.tensor( + [[[1, 2], [3, 4], [5, 6]], [[7, 8], [9, 10], [pad_idx, pad_idx]]] + ), + ) + self.assertTensorEqual( + batch["net_input"]["prev_output_tokens"], + torch.tensor([[eos_idx, 3, 2, pad_idx], [eos_idx, 4, 2, 3]]), + ) + self.assertTensorEqual(batch["net_input"]["src_lengths"], torch.tensor([3, 2])) + self.assertTensorEqual( + batch["target"], + torch.tensor([[3, 2, eos_idx, pad_idx], [4, 2, 3, eos_idx]]), + ) + self.assertEqual(batch["nsentences"], 2) + + def assertTensorEqual(self, t1, t2): + self.assertEqual(t1.size(), t2.size(), "size mismatch") + self.assertEqual(t1.ne(t2).long().sum(), 0) + + +if __name__ == "__main__": + unittest.main() diff --git a/tests/speech_recognition/test_cross_entropy.py b/tests/speech_recognition/test_cross_entropy.py new file mode 100644 index 0000000000000000000000000000000000000000..b05400ed95e22762c3e3e5e8fd3ebfa6caf1e325 --- /dev/null +++ b/tests/speech_recognition/test_cross_entropy.py @@ -0,0 +1,37 @@ +#!/usr/bin/env python3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from examples.speech_recognition.criterions.cross_entropy_acc import ( + CrossEntropyWithAccCriterion, +) + +from .asr_test_base import CrossEntropyCriterionTestBase + + +class CrossEntropyWithAccCriterionTest(CrossEntropyCriterionTestBase): + def setUp(self): + self.criterion_cls = CrossEntropyWithAccCriterion + super().setUp() + + def test_cross_entropy_all_correct(self): + sample = self.get_test_sample(correct=True, soft_target=False, aggregate=False) + loss, sample_size, logging_output = self.criterion( + self.model, sample, "sum", log_probs=True + ) + assert logging_output["correct"] == 20 + assert logging_output["total"] == 20 + assert logging_output["sample_size"] == 20 + assert logging_output["ntokens"] == 20 + + def test_cross_entropy_all_wrong(self): + sample = self.get_test_sample(correct=False, soft_target=False, aggregate=False) + loss, sample_size, logging_output = self.criterion( + self.model, sample, "sum", log_probs=True + ) + assert logging_output["correct"] == 0 + assert logging_output["total"] == 20 + assert logging_output["sample_size"] == 20 + assert logging_output["ntokens"] == 20 diff --git a/tests/speech_recognition/test_data_utils.py b/tests/speech_recognition/test_data_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..a72e0b66948da1349d87eafdef4c4004dd535c96 --- /dev/null +++ b/tests/speech_recognition/test_data_utils.py @@ -0,0 +1,62 @@ +#!/usr/bin/env python3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +import unittest + +import torch +from examples.speech_recognition.data import data_utils + + +class DataUtilsTest(unittest.TestCase): + def test_normalization(self): + sample_len1 = torch.tensor( + [ + [ + -0.7661, + -1.3889, + -2.0972, + -0.9134, + -0.7071, + -0.9765, + -0.8700, + -0.8283, + 0.7512, + 1.3211, + 2.1532, + 2.1174, + 1.2800, + 1.2633, + 1.6147, + 1.6322, + 2.0723, + 3.1522, + 3.2852, + 2.2309, + 2.5569, + 2.2183, + 2.2862, + 1.5886, + 0.8773, + 0.8725, + 1.2662, + 0.9899, + 1.1069, + 1.3926, + 1.2795, + 1.1199, + 1.1477, + 1.2687, + 1.3843, + 1.1903, + 0.8355, + 1.1367, + 1.2639, + 1.4707, + ] + ] + ) + out = data_utils.apply_mv_norm(sample_len1) + assert not torch.isnan(out).any() + assert (out == sample_len1).all() diff --git a/tests/speech_recognition/test_vggtransformer.py b/tests/speech_recognition/test_vggtransformer.py new file mode 100644 index 0000000000000000000000000000000000000000..4dc73b8c7379970dc0bcc16fcb088a64a1bd7e3b --- /dev/null +++ b/tests/speech_recognition/test_vggtransformer.py @@ -0,0 +1,135 @@ +#!/usr/bin/env python3 + +# import models/encoder/decoder to be tested +from examples.speech_recognition.models.vggtransformer import ( + TransformerDecoder, + VGGTransformerEncoder, + VGGTransformerModel, + vggtransformer_1, + vggtransformer_2, + vggtransformer_base, +) + +# import base test class +from .asr_test_base import ( + DEFAULT_TEST_VOCAB_SIZE, + TestFairseqDecoderBase, + TestFairseqEncoderBase, + TestFairseqEncoderDecoderModelBase, + get_dummy_dictionary, + get_dummy_encoder_output, + get_dummy_input, +) + + +class VGGTransformerModelTest_mid(TestFairseqEncoderDecoderModelBase): + def setUp(self): + def override_config(args): + """ + vggtrasformer_1 use 14 layers of transformer, + for testing purpose, it is too expensive. For fast turn-around + test, reduce the number of layers to 3. + """ + args.transformer_enc_config = ( + "((1024, 16, 4096, True, 0.15, 0.15, 0.15),) * 3" + ) + + super().setUp() + extra_args_setter = [vggtransformer_1, override_config] + + self.setUpModel(VGGTransformerModel, extra_args_setter) + self.setUpInput(get_dummy_input(T=50, D=80, B=5, K=DEFAULT_TEST_VOCAB_SIZE)) + + +class VGGTransformerModelTest_big(TestFairseqEncoderDecoderModelBase): + def setUp(self): + def override_config(args): + """ + vggtrasformer_2 use 16 layers of transformer, + for testing purpose, it is too expensive. For fast turn-around + test, reduce the number of layers to 3. + """ + args.transformer_enc_config = ( + "((1024, 16, 4096, True, 0.15, 0.15, 0.15),) * 3" + ) + + super().setUp() + extra_args_setter = [vggtransformer_2, override_config] + + self.setUpModel(VGGTransformerModel, extra_args_setter) + self.setUpInput(get_dummy_input(T=50, D=80, B=5, K=DEFAULT_TEST_VOCAB_SIZE)) + + +class VGGTransformerModelTest_base(TestFairseqEncoderDecoderModelBase): + def setUp(self): + def override_config(args): + """ + vggtrasformer_base use 12 layers of transformer, + for testing purpose, it is too expensive. For fast turn-around + test, reduce the number of layers to 3. + """ + args.transformer_enc_config = ( + "((512, 8, 2048, True, 0.15, 0.15, 0.15),) * 3" + ) + + super().setUp() + extra_args_setter = [vggtransformer_base, override_config] + + self.setUpModel(VGGTransformerModel, extra_args_setter) + self.setUpInput(get_dummy_input(T=50, D=80, B=5, K=DEFAULT_TEST_VOCAB_SIZE)) + + +class VGGTransformerEncoderTest(TestFairseqEncoderBase): + def setUp(self): + super().setUp() + + self.setUpInput(get_dummy_input(T=50, D=80, B=5)) + + def test_forward(self): + print("1. test standard vggtransformer") + self.setUpEncoder(VGGTransformerEncoder(input_feat_per_channel=80)) + super().test_forward() + print("2. test vggtransformer with limited right context") + self.setUpEncoder( + VGGTransformerEncoder( + input_feat_per_channel=80, transformer_context=(-1, 5) + ) + ) + super().test_forward() + print("3. test vggtransformer with limited left context") + self.setUpEncoder( + VGGTransformerEncoder( + input_feat_per_channel=80, transformer_context=(5, -1) + ) + ) + super().test_forward() + print("4. test vggtransformer with limited right context and sampling") + self.setUpEncoder( + VGGTransformerEncoder( + input_feat_per_channel=80, + transformer_context=(-1, 12), + transformer_sampling=(2, 2), + ) + ) + super().test_forward() + print("5. test vggtransformer with windowed context and sampling") + self.setUpEncoder( + VGGTransformerEncoder( + input_feat_per_channel=80, + transformer_context=(12, 12), + transformer_sampling=(2, 2), + ) + ) + + +class TransformerDecoderTest(TestFairseqDecoderBase): + def setUp(self): + super().setUp() + + dict = get_dummy_dictionary(vocab_size=DEFAULT_TEST_VOCAB_SIZE) + decoder = TransformerDecoder(dict) + dummy_encoder_output = get_dummy_encoder_output(encoder_out_shape=(50, 5, 256)) + + self.setUpDecoder(decoder) + self.setUpInput(dummy_encoder_output) + self.setUpPrevOutputTokens() diff --git a/tests/test_activation_checkpointing.py b/tests/test_activation_checkpointing.py new file mode 100644 index 0000000000000000000000000000000000000000..647a9572886f8aff09a4aadc0b21e1d5817ff38e --- /dev/null +++ b/tests/test_activation_checkpointing.py @@ -0,0 +1,79 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import unittest + +import torch +import torch.nn as nn +from fairseq.modules.checkpoint_activations import checkpoint_wrapper +from torch.utils.checkpoint import checkpoint + + +class Model(nn.Module): + def __init__( + self, use_pytorch_checkpoint=False, use_fairseq_checkpoint=False, **kwargs + ): + super().__init__() + torch.manual_seed(0) + self.use_pytorch_checkpoint = use_pytorch_checkpoint + self.ffn = nn.Sequential( + nn.Linear(32, 128), + # add a Dropout layer to test RNG save/restore + nn.Dropout(p=0.5), + nn.Linear(128, 32), + ) + if use_fairseq_checkpoint: + self.ffn = checkpoint_wrapper(self.ffn, **kwargs) + self.out = nn.Linear(32, 1) + + def forward(self, x): + if self.use_pytorch_checkpoint: + x = checkpoint(self.ffn, x) + else: + x = self.ffn(x) + return self.out(x) + + +class TestActivationCheckpointing(unittest.TestCase): + def _test_checkpoint_wrapper(self, device, log_memory_usage=False): + def get_loss_and_gnorm(model): + torch.manual_seed(1) + input = torch.rand(2, 16, 32).requires_grad_(True).to(device) + model.zero_grad() + loss = model(input).sum() + loss.backward() + gnorm = torch.norm( + torch.stack([torch.norm(p.grad.detach()) for p in model.parameters()]) + ) + return {"loss": loss, "gnorm": gnorm} + + model = Model().to(device) + no_cpt = get_loss_and_gnorm(model) + + model = Model(use_pytorch_checkpoint=True).to(device) + pyt_cpt = get_loss_and_gnorm(model) + torch.testing.assert_allclose(no_cpt["loss"], pyt_cpt["loss"]) + torch.testing.assert_allclose(no_cpt["gnorm"], pyt_cpt["gnorm"]) + + model = Model(use_fairseq_checkpoint=True).to(device) + fairseq_cpt = get_loss_and_gnorm(model) + torch.testing.assert_allclose(no_cpt["loss"], fairseq_cpt["loss"]) + torch.testing.assert_allclose(no_cpt["gnorm"], fairseq_cpt["gnorm"]) + + model = Model(use_fairseq_checkpoint=True, offload_to_cpu=True).to(device) + fairseq_cpt_offload = get_loss_and_gnorm(model) + torch.testing.assert_allclose(no_cpt["loss"], fairseq_cpt_offload["loss"]) + torch.testing.assert_allclose(no_cpt["gnorm"], fairseq_cpt_offload["gnorm"]) + + def test_checkpoint_wrapper_cpu(self): + self._test_checkpoint_wrapper(device=torch.device("cpu")) + + @unittest.skipIf(not torch.cuda.is_available(), "test requires a GPU") + def test_checkpoint_wrapper_cuda(self): + self._test_checkpoint_wrapper(device=torch.device("cuda")) + + +if __name__ == "__main__": + unittest.main() diff --git a/tests/test_amp_optimizer.py b/tests/test_amp_optimizer.py new file mode 100644 index 0000000000000000000000000000000000000000..3a785e1830e91b7e090e841d428fe4ea61f3a65c --- /dev/null +++ b/tests/test_amp_optimizer.py @@ -0,0 +1,78 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import copy +import unittest + +import torch +from torch.cuda.amp import autocast, GradScaler +from fairseq.optim import build_optimizer + + +@unittest.skipIf(not torch.cuda.is_available(), "test requires a GPU") +class TestGradientScalingAMP(unittest.TestCase): + def setUp(self): + self.x = torch.tensor([2.0]).cuda().half() + weight = 3.0 + bias = 5.0 + self.error = 1.0 + self.target = torch.tensor([self.x * weight + bias + self.error]).cuda() + self.loss_fn = torch.nn.L1Loss() + + self.model = torch.nn.Linear(1, 1) + self.model.weight.data = torch.tensor([[weight]]) + self.model.bias.data = torch.tensor([bias]) + self.model.cuda() + self.params = list(self.model.parameters()) + + self.namespace_dls = argparse.Namespace( + optimizer="adam", + lr=[0.1], + adam_betas="(0.9, 0.999)", + adam_eps=1e-8, + weight_decay=0.0, + threshold_loss_scale=1, + min_loss_scale=1e-4, + ) + self.scaler = GradScaler( + init_scale=1, + growth_interval=1, + ) + + def run_iter(self, model, params, optimizer): + optimizer.zero_grad() + with autocast(): + y = model(self.x) + loss = self.loss_fn(y, self.target) + self.scaler.scale(loss).backward() + self.assertEqual(loss, torch.tensor(1.0, device="cuda:0", dtype=torch.float16)) + + self.scaler.unscale_(optimizer) + grad_norm = optimizer.clip_grad_norm(0) + self.assertAlmostEqual(grad_norm.item(), 2.2361, 4) + + self.scaler.step(optimizer) + self.scaler.update() + self.assertEqual( + model.weight, + torch.tensor( + [[3.1]], device="cuda:0", requires_grad=True + ), + ) + self.assertEqual( + model.bias, + torch.tensor( + [5.1], device="cuda:0", requires_grad=True + ), + ) + self.assertEqual(self.scaler.get_scale(), 2.0) + + def test_automatic_mixed_precision(self): + model = copy.deepcopy(self.model) + params = list(model.parameters()) + optimizer = build_optimizer(self.namespace_dls, params) + + self.run_iter(model, params, optimizer) diff --git a/tests/test_average_checkpoints.py b/tests/test_average_checkpoints.py new file mode 100644 index 0000000000000000000000000000000000000000..f348b56b869372d8434fe03f13324d78e9093fa2 --- /dev/null +++ b/tests/test_average_checkpoints.py @@ -0,0 +1,134 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import collections +import os +import shutil +import tempfile +import unittest + +import numpy as np +import torch +from scripts.average_checkpoints import average_checkpoints +from torch import nn + + +class ModelWithSharedParameter(nn.Module): + def __init__(self): + super(ModelWithSharedParameter, self).__init__() + self.embedding = nn.Embedding(1000, 200) + self.FC1 = nn.Linear(200, 200) + self.FC2 = nn.Linear(200, 200) + # tie weight in FC2 to FC1 + self.FC2.weight = nn.Parameter(self.FC1.weight) + self.FC2.bias = nn.Parameter(self.FC1.bias) + + self.relu = nn.ReLU() + + def forward(self, input): + return self.FC2(self.ReLU(self.FC1(input))) + self.FC1(input) + + +class TestAverageCheckpoints(unittest.TestCase): + def test_average_checkpoints(self): + params_0 = collections.OrderedDict( + [ + ("a", torch.DoubleTensor([100.0])), + ("b", torch.FloatTensor([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]])), + ("c", torch.IntTensor([7, 8, 9])), + ] + ) + params_1 = collections.OrderedDict( + [ + ("a", torch.DoubleTensor([1.0])), + ("b", torch.FloatTensor([[1.0, 1.0, 1.0], [1.0, 1.0, 1.0]])), + ("c", torch.IntTensor([2, 2, 2])), + ] + ) + params_avg = collections.OrderedDict( + [ + ("a", torch.DoubleTensor([50.5])), + ("b", torch.FloatTensor([[1.0, 1.5, 2.0], [2.5, 3.0, 3.5]])), + # We expect truncation for integer division + ("c", torch.IntTensor([4, 5, 5])), + ] + ) + + fd_0, path_0 = tempfile.mkstemp() + fd_1, path_1 = tempfile.mkstemp() + torch.save(collections.OrderedDict([("model", params_0)]), path_0) + torch.save(collections.OrderedDict([("model", params_1)]), path_1) + + output = average_checkpoints([path_0, path_1])["model"] + + os.close(fd_0) + os.remove(path_0) + os.close(fd_1) + os.remove(path_1) + + for (k_expected, v_expected), (k_out, v_out) in zip( + params_avg.items(), output.items() + ): + self.assertEqual( + k_expected, + k_out, + "Key mismatch - expected {} but found {}. " + "(Expected list of keys: {} vs actual list of keys: {})".format( + k_expected, k_out, params_avg.keys(), output.keys() + ), + ) + np.testing.assert_allclose( + v_expected.numpy(), + v_out.numpy(), + err_msg="Tensor value mismatch for key {}".format(k_expected), + ) + + def test_average_checkpoints_with_shared_parameters(self): + def _construct_model_with_shared_parameters(path, value): + m = ModelWithSharedParameter() + nn.init.constant_(m.FC1.weight, value) + torch.save({"model": m.state_dict()}, path) + return m + + tmpdir = tempfile.mkdtemp() + paths = [] + path = os.path.join(tmpdir, "m1.pt") + m1 = _construct_model_with_shared_parameters(path, 1.0) + paths.append(path) + + path = os.path.join(tmpdir, "m2.pt") + m2 = _construct_model_with_shared_parameters(path, 2.0) + paths.append(path) + + path = os.path.join(tmpdir, "m3.pt") + m3 = _construct_model_with_shared_parameters(path, 3.0) + paths.append(path) + + new_model = average_checkpoints(paths) + self.assertTrue( + torch.equal( + new_model["model"]["embedding.weight"], + (m1.embedding.weight + m2.embedding.weight + m3.embedding.weight) / 3.0, + ) + ) + + self.assertTrue( + torch.equal( + new_model["model"]["FC1.weight"], + (m1.FC1.weight + m2.FC1.weight + m3.FC1.weight) / 3.0, + ) + ) + + self.assertTrue( + torch.equal( + new_model["model"]["FC2.weight"], + (m1.FC2.weight + m2.FC2.weight + m3.FC2.weight) / 3.0, + ) + ) + shutil.rmtree(tmpdir) + + +if __name__ == "__main__": + unittest.main() diff --git a/tests/test_backtranslation_dataset.py b/tests/test_backtranslation_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..dffc3b49387dfdc046ea23d7db179377040b7cbc --- /dev/null +++ b/tests/test_backtranslation_dataset.py @@ -0,0 +1,123 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import unittest + +import tests.utils as test_utils +import torch +from fairseq.data import ( + BacktranslationDataset, + LanguagePairDataset, + TransformEosDataset, +) +from fairseq.sequence_generator import SequenceGenerator + + +class TestBacktranslationDataset(unittest.TestCase): + def setUp(self): + ( + self.tgt_dict, + self.w1, + self.w2, + self.src_tokens, + self.src_lengths, + self.model, + ) = test_utils.sequence_generator_setup() + + dummy_src_samples = self.src_tokens + + self.tgt_dataset = test_utils.TestDataset(data=dummy_src_samples) + self.cuda = torch.cuda.is_available() + + def _backtranslation_dataset_helper( + self, + remove_eos_from_input_src, + remove_eos_from_output_src, + ): + tgt_dataset = LanguagePairDataset( + src=self.tgt_dataset, + src_sizes=self.tgt_dataset.sizes, + src_dict=self.tgt_dict, + tgt=None, + tgt_sizes=None, + tgt_dict=None, + ) + + generator = SequenceGenerator( + [self.model], + tgt_dict=self.tgt_dict, + max_len_a=0, + max_len_b=200, + beam_size=2, + unk_penalty=0, + ) + + backtranslation_dataset = BacktranslationDataset( + tgt_dataset=TransformEosDataset( + dataset=tgt_dataset, + eos=self.tgt_dict.eos(), + # remove eos from the input src + remove_eos_from_src=remove_eos_from_input_src, + ), + src_dict=self.tgt_dict, + backtranslation_fn=( + lambda sample: generator.generate([self.model], sample) + ), + output_collater=TransformEosDataset( + dataset=tgt_dataset, + eos=self.tgt_dict.eos(), + # if we remove eos from the input src, then we need to add it + # back to the output tgt + append_eos_to_tgt=remove_eos_from_input_src, + remove_eos_from_src=remove_eos_from_output_src, + ).collater, + cuda=self.cuda, + ) + dataloader = torch.utils.data.DataLoader( + backtranslation_dataset, + batch_size=2, + collate_fn=backtranslation_dataset.collater, + ) + backtranslation_batch_result = next(iter(dataloader)) + + eos, pad, w1, w2 = self.tgt_dict.eos(), self.tgt_dict.pad(), self.w1, self.w2 + + # Note that we sort by src_lengths and add left padding, so actually + # ids will look like: [1, 0] + expected_src = torch.LongTensor([[w1, w2, w1, eos], [pad, pad, w1, eos]]) + if remove_eos_from_output_src: + expected_src = expected_src[:, :-1] + expected_tgt = torch.LongTensor([[w1, w2, eos], [w1, w2, eos]]) + generated_src = backtranslation_batch_result["net_input"]["src_tokens"] + tgt_tokens = backtranslation_batch_result["target"] + + self.assertTensorEqual(expected_src, generated_src) + self.assertTensorEqual(expected_tgt, tgt_tokens) + + def test_backtranslation_dataset_no_eos_in_output_src(self): + self._backtranslation_dataset_helper( + remove_eos_from_input_src=False, + remove_eos_from_output_src=True, + ) + + def test_backtranslation_dataset_with_eos_in_output_src(self): + self._backtranslation_dataset_helper( + remove_eos_from_input_src=False, + remove_eos_from_output_src=False, + ) + + def test_backtranslation_dataset_no_eos_in_input_src(self): + self._backtranslation_dataset_helper( + remove_eos_from_input_src=True, + remove_eos_from_output_src=False, + ) + + def assertTensorEqual(self, t1, t2): + self.assertEqual(t1.size(), t2.size(), "size mismatch") + self.assertEqual(t1.ne(t2).long().sum(), 0) + + +if __name__ == "__main__": + unittest.main() diff --git a/tests/test_binaries.py b/tests/test_binaries.py new file mode 100644 index 0000000000000000000000000000000000000000..4e207742625427f108f78bcd24d487a081b6ccf7 --- /dev/null +++ b/tests/test_binaries.py @@ -0,0 +1,1874 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import contextlib +import logging +import json +import os +import random +import sys +import tempfile +import unittest +from io import StringIO +from typing import List, Dict +import torch +from fairseq import options +from fairseq_cli import eval_lm, train +from tests.utils import ( + create_dummy_data, + generate_main, + preprocess_lm_data, + preprocess_summarization_data, + preprocess_translation_data, + create_laser_data_and_config_json, + train_translation_model, + train_language_model, +) + + +try: + import transformers # noqa + + has_hf_transformers = True +except ImportError: + has_hf_transformers = False + + +class TestTranslation(unittest.TestCase): + def setUp(self): + logging.disable(logging.CRITICAL) + + def tearDown(self): + logging.disable(logging.NOTSET) + + def test_fconv(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_fconv") as data_dir: + create_dummy_data(data_dir) + preprocess_translation_data(data_dir) + train_translation_model(data_dir, "fconv_iwslt_de_en") + generate_main(data_dir) + + def test_raw(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_fconv_raw") as data_dir: + create_dummy_data(data_dir) + preprocess_translation_data(data_dir, ["--dataset-impl", "raw"]) + train_translation_model( + data_dir, "fconv_iwslt_de_en", ["--dataset-impl", "raw"] + ) + generate_main(data_dir, ["--dataset-impl", "raw"]) + + def test_update_freq(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_update_freq") as data_dir: + create_dummy_data(data_dir) + preprocess_translation_data(data_dir) + train_translation_model( + data_dir, "fconv_iwslt_de_en", ["--update-freq", "3"] + ) + generate_main(data_dir) + + def test_max_positions(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_max_positions") as data_dir: + create_dummy_data(data_dir) + preprocess_translation_data(data_dir) + with self.assertRaises(Exception) as context: + train_translation_model( + data_dir, + "fconv_iwslt_de_en", + ["--max-target-positions", "5"], + ) + self.assertTrue( + "skip this example with --skip-invalid-size-inputs-valid-test" + in str(context.exception) + ) + train_translation_model( + data_dir, + "fconv_iwslt_de_en", + [ + "--max-target-positions", + "5", + "--skip-invalid-size-inputs-valid-test", + ], + ) + with self.assertRaises(Exception) as context: + generate_main(data_dir) + generate_main(data_dir, ["--skip-invalid-size-inputs-valid-test"]) + + def test_generation(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_sampling") as data_dir: + create_dummy_data(data_dir) + preprocess_translation_data(data_dir) + train_translation_model(data_dir, "fconv_iwslt_de_en") + generate_main( + data_dir, + [ + "--sampling", + "--temperature", + "2", + "--beam", + "2", + "--nbest", + "2", + ], + ) + generate_main( + data_dir, + [ + "--sampling", + "--sampling-topk", + "3", + "--beam", + "2", + "--nbest", + "2", + ], + ) + generate_main( + data_dir, + [ + "--sampling", + "--sampling-topp", + "0.2", + "--beam", + "2", + "--nbest", + "2", + ], + ) + generate_main( + data_dir, + [ + "--diversity-rate", + "0.5", + "--beam", + "6", + ], + ) + with self.assertRaises(ValueError): + generate_main( + data_dir, + [ + "--diverse-beam-groups", + "4", + "--match-source-len", + ], + ) + generate_main(data_dir, ["--prefix-size", "2"]) + generate_main(data_dir, ["--retain-dropout"]) + + def test_eval_bleu(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_eval_bleu") as data_dir: + create_dummy_data(data_dir) + preprocess_translation_data(data_dir) + train_translation_model( + data_dir, + "fconv_iwslt_de_en", + [ + "--eval-bleu", + "--eval-bleu-print-samples", + "--eval-bleu-remove-bpe", + "--eval-bleu-detok", + "space", + "--eval-bleu-args", + '{"beam": 4, "min_len": 10}', + ], + ) + + def test_lstm(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_lstm") as data_dir: + create_dummy_data(data_dir) + preprocess_translation_data(data_dir) + train_translation_model( + data_dir, + "lstm_wiseman_iwslt_de_en", + [ + "--encoder-layers", + "2", + "--decoder-layers", + "2", + "--encoder-embed-dim", + "8", + "--decoder-embed-dim", + "8", + "--decoder-out-embed-dim", + "8", + ], + ) + generate_main(data_dir) + + def test_lstm_bidirectional(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_lstm_bidirectional") as data_dir: + create_dummy_data(data_dir) + preprocess_translation_data(data_dir) + train_translation_model( + data_dir, + "lstm", + [ + "--encoder-layers", + "2", + "--encoder-bidirectional", + "--encoder-hidden-size", + "16", + "--encoder-embed-dim", + "8", + "--decoder-embed-dim", + "8", + "--decoder-out-embed-dim", + "8", + "--decoder-layers", + "2", + ], + ) + generate_main(data_dir) + + def test_transformer(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_transformer") as data_dir: + create_dummy_data(data_dir) + preprocess_translation_data(data_dir) + train_translation_model( + data_dir, + "transformer_iwslt_de_en", + [ + "--encoder-layers", + "2", + "--decoder-layers", + "2", + "--encoder-embed-dim", + "8", + "--decoder-embed-dim", + "8", + ], + run_validation=True, + ) + generate_main(data_dir) + + def test_multilingual_transformer(self): + # test with all combinations of encoder/decoder lang tokens + encoder_langtok_flags = [ + [], + ["--encoder-langtok", "src"], + ["--encoder-langtok", "tgt"], + ] + decoder_langtok_flags = [[], ["--decoder-langtok"]] + with contextlib.redirect_stdout(StringIO()): + for i in range(len(encoder_langtok_flags)): + for j in range(len(decoder_langtok_flags)): + enc_ltok_flag = encoder_langtok_flags[i] + dec_ltok_flag = decoder_langtok_flags[j] + with tempfile.TemporaryDirectory( + f"test_multilingual_transformer_{i}_{j}" + ) as data_dir: + create_dummy_data(data_dir) + preprocess_translation_data(data_dir) + train_translation_model( + data_dir, + arch="multilingual_transformer", + task="multilingual_translation", + extra_flags=[ + "--encoder-layers", + "2", + "--decoder-layers", + "2", + "--encoder-embed-dim", + "8", + "--decoder-embed-dim", + "8", + ] + + enc_ltok_flag + + dec_ltok_flag, + lang_flags=["--lang-pairs", "in-out,out-in"], + run_validation=True, + extra_valid_flags=enc_ltok_flag + dec_ltok_flag, + ) + generate_main( + data_dir, + extra_flags=[ + "--task", + "multilingual_translation", + "--lang-pairs", + "in-out,out-in", + "--source-lang", + "in", + "--target-lang", + "out", + ] + + enc_ltok_flag + + dec_ltok_flag, + ) + + @unittest.skipIf( + sys.platform.lower() == "darwin", "skip latent depth test on MacOS" + ) + def test_multilingual_translation_latent_depth(self): + # test with latent depth in encoder, decoder, or both + encoder_latent_layer = [[], ["--encoder-latent-layer"]] + decoder_latent_layer = [[], ["--decoder-latent-layer"]] + with contextlib.redirect_stdout(StringIO()): + for i in range(len(encoder_latent_layer)): + for j in range(len(decoder_latent_layer)): + if i == 0 and j == 0: + continue + enc_ll_flag = encoder_latent_layer[i] + dec_ll_flag = decoder_latent_layer[j] + with tempfile.TemporaryDirectory( + f"test_multilingual_translation_latent_depth_{i}_{j}" + ) as data_dir: + create_dummy_data(data_dir) + preprocess_translation_data( + data_dir, extra_flags=["--joined-dictionary"] + ) + train_translation_model( + data_dir, + arch="latent_multilingual_transformer", + task="multilingual_translation_latent_depth", + extra_flags=[ + "--user-dir", + "examples/latent_depth/latent_depth_src", + "--encoder-layers", + "2", + "--decoder-layers", + "2", + "--encoder-embed-dim", + "8", + "--decoder-embed-dim", + "8", + "--share-encoders", + "--share-decoders", + "--sparsity-weight", + "0.1", + ] + + enc_ll_flag + + dec_ll_flag, + lang_flags=["--lang-pairs", "in-out,out-in"], + run_validation=True, + extra_valid_flags=[ + "--user-dir", + "examples/latent_depth/latent_depth_src", + ] + + enc_ll_flag + + dec_ll_flag, + ) + generate_main( + data_dir, + extra_flags=[ + "--user-dir", + "examples/latent_depth/latent_depth_src", + "--task", + "multilingual_translation_latent_depth", + "--lang-pairs", + "in-out,out-in", + "--source-lang", + "in", + "--target-lang", + "out", + ] + + enc_ll_flag + + dec_ll_flag, + ) + + def test_translation_multi_simple_epoch(self): + # test with all combinations of encoder/decoder lang tokens + encoder_langtok_flags = [ + [], + ["--encoder-langtok", "src"], + ["--encoder-langtok", "tgt"], + ] + decoder_langtok_flags = [[], ["--decoder-langtok"]] + with contextlib.redirect_stdout(StringIO()): + for i in range(len(encoder_langtok_flags)): + for j in range(len(decoder_langtok_flags)): + enc_ltok_flag = encoder_langtok_flags[i] + dec_ltok_flag = decoder_langtok_flags[j] + with tempfile.TemporaryDirectory( + f"test_translation_multi_simple_epoch_{i}_{j}" + ) as data_dir: + create_dummy_data(data_dir) + preprocess_translation_data( + data_dir, extra_flags=["--joined-dictionary"] + ) + train_translation_model( + data_dir, + arch="transformer", + task="translation_multi_simple_epoch", + extra_flags=[ + "--encoder-layers", + "2", + "--decoder-layers", + "2", + "--encoder-embed-dim", + "8", + "--decoder-embed-dim", + "8", + "--sampling-method", + "temperature", + "--sampling-temperature", + "1.5", + "--virtual-epoch-size", + "1000", + ] + + enc_ltok_flag + + dec_ltok_flag, + lang_flags=["--lang-pairs", "in-out,out-in"], + run_validation=True, + extra_valid_flags=enc_ltok_flag + dec_ltok_flag, + ) + generate_main( + data_dir, + extra_flags=[ + "--task", + "translation_multi_simple_epoch", + "--lang-pairs", + "in-out,out-in", + "--source-lang", + "in", + "--target-lang", + "out", + ] + + enc_ltok_flag + + dec_ltok_flag, + ) + + def test_translation_multi_simple_epoch_no_vepoch(self): + # test with all combinations of encoder/decoder lang tokens + with contextlib.redirect_stdout(StringIO()): + enc_ltok_flag = ["--encoder-langtok", "src"] + dec_ltok_flag = ["--decoder-langtok"] + with tempfile.TemporaryDirectory( + "test_translation_multi_simple_epoch_dict" + ) as data_dir: + create_dummy_data(data_dir) + preprocess_translation_data(data_dir, extra_flags=[]) + train_translation_model( + data_dir, + arch="transformer", + task="translation_multi_simple_epoch", + extra_flags=[ + "--encoder-layers", + "2", + "--decoder-layers", + "2", + "--encoder-embed-dim", + "8", + "--decoder-embed-dim", + "8", + "--sampling-method", + "temperature", + "--sampling-temperature", + "1.5", + ] + + enc_ltok_flag + + dec_ltok_flag, + lang_flags=["--lang-pairs", "in-out"], + run_validation=True, + extra_valid_flags=enc_ltok_flag + dec_ltok_flag, + ) + generate_main( + data_dir, + extra_flags=[ + "--task", + "translation_multi_simple_epoch", + "--lang-pairs", + "in-out", + "--source-lang", + "in", + "--target-lang", + "out", + ] + + enc_ltok_flag + + dec_ltok_flag, + ) + + def test_translation_multi_simple_epoch_dicts(self): + # test with all combinations of encoder/decoder lang tokens + with contextlib.redirect_stdout(StringIO()): + enc_ltok_flag = ["--encoder-langtok", "src"] + dec_ltok_flag = ["--decoder-langtok"] + with tempfile.TemporaryDirectory( + "test_translation_multi_simple_epoch_dict" + ) as data_dir: + create_dummy_data(data_dir) + preprocess_translation_data(data_dir, extra_flags=[]) + train_translation_model( + data_dir, + arch="transformer", + task="translation_multi_simple_epoch", + extra_flags=[ + "--encoder-layers", + "2", + "--decoder-layers", + "2", + "--encoder-embed-dim", + "8", + "--decoder-embed-dim", + "8", + "--sampling-method", + "temperature", + "--sampling-temperature", + "1.5", + "--virtual-epoch-size", + "1000", + ] + + enc_ltok_flag + + dec_ltok_flag, + lang_flags=["--lang-pairs", "in-out"], + run_validation=True, + extra_valid_flags=enc_ltok_flag + dec_ltok_flag, + ) + generate_main( + data_dir, + extra_flags=[ + "--task", + "translation_multi_simple_epoch", + "--lang-pairs", + "in-out", + "--source-lang", + "in", + "--target-lang", + "out", + ] + + enc_ltok_flag + + dec_ltok_flag, + ) + + def test_translation_multi_simple_epoch_src_tgt_dict_spec(self): + # test the specification of explicit --src-dict and --tgt-dict + with contextlib.redirect_stdout(StringIO()): + enc_ltok_flag = ["--encoder-langtok", "src"] + dec_ltok_flag = ["--decoder-langtok"] + with tempfile.TemporaryDirectory( + "test_translation_multi_simple_epoch_dict" + ) as data_dir: + create_dummy_data(data_dir) + preprocess_translation_data(data_dir, extra_flags=[]) + train_translation_model( + data_dir, + arch="transformer", + task="translation_multi_simple_epoch", + extra_flags=[ + "--source-dict", + f"{data_dir}/dict.in.txt", + "--target-dict", + f"{data_dir}/dict.out.txt", + "--encoder-layers", + "2", + "--decoder-layers", + "2", + "--encoder-embed-dim", + "8", + "--decoder-embed-dim", + "8", + "--sampling-method", + "temperature", + "--sampling-temperature", + "1.5", + "--virtual-epoch-size", + "1000", + ] + + enc_ltok_flag + + dec_ltok_flag, + lang_flags=["--lang-pairs", "in-out"], + run_validation=True, + extra_valid_flags=enc_ltok_flag + dec_ltok_flag, + ) + generate_main( + data_dir, + extra_flags=[ + "--task", + "translation_multi_simple_epoch", + "--lang-pairs", + "in-out", + "--source-lang", + "in", + "--target-lang", + "out", + ] + + enc_ltok_flag + + dec_ltok_flag, + ) + + def test_transformer_cross_self_attention(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory( + "test_transformer_cross_self_attention" + ) as data_dir: + create_dummy_data(data_dir) + preprocess_translation_data(data_dir) + train_translation_model( + data_dir, + "transformer_iwslt_de_en", + [ + "--encoder-layers", + "2", + "--decoder-layers", + "2", + "--encoder-embed-dim", + "8", + "--decoder-embed-dim", + "8", + "--decoder-embed-dim", + "8", + "--no-cross-attention", + "--cross-self-attention", + ], + run_validation=True, + ) + generate_main(data_dir, extra_flags=[]) + + def test_transformer_pointer_generator(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory( + "test_transformer_pointer_generator" + ) as data_dir: + create_dummy_data(data_dir) + preprocess_summarization_data(data_dir) + train_translation_model( + data_dir, + "transformer_pointer_generator", + extra_flags=[ + "--user-dir", + "examples/pointer_generator/pointer_generator_src", + "--encoder-layers", + "2", + "--decoder-layers", + "2", + "--encoder-embed-dim", + "8", + "--decoder-embed-dim", + "8", + "--alignment-layer", + "-1", + "--alignment-heads", + "1", + "--source-position-markers", + "0", + ], + run_validation=True, + extra_valid_flags=[ + "--user-dir", + "examples/pointer_generator/pointer_generator_src", + ], + ) + generate_main( + data_dir, + extra_flags=[ + "--user-dir", + "examples/pointer_generator/pointer_generator_src", + ], + ) + + def test_lightconv(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_lightconv") as data_dir: + create_dummy_data(data_dir) + preprocess_translation_data(data_dir) + train_translation_model( + data_dir, + "lightconv_iwslt_de_en", + [ + "--encoder-conv-type", + "lightweight", + "--decoder-conv-type", + "lightweight", + "--encoder-embed-dim", + "8", + "--decoder-embed-dim", + "8", + ], + ) + generate_main(data_dir) + + def test_dynamicconv(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_dynamicconv") as data_dir: + create_dummy_data(data_dir) + preprocess_translation_data(data_dir) + train_translation_model( + data_dir, + "lightconv_iwslt_de_en", + [ + "--encoder-conv-type", + "dynamic", + "--decoder-conv-type", + "dynamic", + "--encoder-embed-dim", + "8", + "--decoder-embed-dim", + "8", + ], + ) + generate_main(data_dir) + + def test_cmlm_transformer(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_cmlm_transformer") as data_dir: + create_dummy_data(data_dir) + preprocess_translation_data(data_dir, ["--joined-dictionary"]) + train_translation_model( + data_dir, + "cmlm_transformer", + [ + "--apply-bert-init", + "--criterion", + "nat_loss", + "--noise", + "full_mask", + "--pred-length-offset", + "--length-loss-factor", + "0.1", + ], + task="translation_lev", + ) + generate_main( + data_dir, + [ + "--task", + "translation_lev", + "--iter-decode-max-iter", + "9", + "--iter-decode-eos-penalty", + "0", + "--print-step", + ], + ) + + def test_nonautoregressive_transformer(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory( + "test_nonautoregressive_transformer" + ) as data_dir: + create_dummy_data(data_dir) + preprocess_translation_data(data_dir, ["--joined-dictionary"]) + train_translation_model( + data_dir, + "nonautoregressive_transformer", + [ + "--apply-bert-init", + "--src-embedding-copy", + "--criterion", + "nat_loss", + "--noise", + "full_mask", + "--pred-length-offset", + "--length-loss-factor", + "0.1", + ], + task="translation_lev", + ) + generate_main( + data_dir, + [ + "--task", + "translation_lev", + "--iter-decode-max-iter", + "0", + "--iter-decode-eos-penalty", + "0", + "--print-step", + ], + ) + + # def test_nat_crf_transformer(self): + # with contextlib.redirect_stdout(StringIO()): + # with tempfile.TemporaryDirectory('test_nat_crf_transformer') as data_dir: + # create_dummy_data(data_dir) + # preprocess_translation_data(data_dir, ['--joined-dictionary']) + # train_translation_model(data_dir, 'nacrf_transformer', [ + # '--apply-bert-init', '--criterion', + # 'nat_loss', '--noise', 'full_mask', '--pred-length-offset', + # '--length-loss-factor', '0.1', + # '--word-ins-loss-factor', '0.5', + # '--crf-lowrank-approx', '1', + # '--crf-beam-approx', '1' + # ], task='translation_lev') + # generate_main(data_dir, [ + # '--task', 'translation_lev', + # '--iter-decode-max-iter', '0', + # '--iter-decode-eos-penalty', '0', + # '--print-step', + # ]) + + def test_iterative_nonautoregressive_transformer(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory( + "test_iterative_nonautoregressive_transformer" + ) as data_dir: + create_dummy_data(data_dir) + preprocess_translation_data(data_dir, ["--joined-dictionary"]) + train_translation_model( + data_dir, + "iterative_nonautoregressive_transformer", + [ + "--apply-bert-init", + "--src-embedding-copy", + "--criterion", + "nat_loss", + "--noise", + "full_mask", + "--stochastic-approx", + "--dae-ratio", + "0.5", + "--train-step", + "3", + ], + task="translation_lev", + ) + generate_main( + data_dir, + [ + "--task", + "translation_lev", + "--iter-decode-max-iter", + "9", + "--iter-decode-eos-penalty", + "0", + "--print-step", + ], + ) + + def test_insertion_transformer(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_insertion_transformer") as data_dir: + create_dummy_data(data_dir) + preprocess_translation_data(data_dir, ["--joined-dictionary"]) + train_translation_model( + data_dir, + "insertion_transformer", + [ + "--apply-bert-init", + "--criterion", + "nat_loss", + "--noise", + "random_mask", + ], + task="translation_lev", + ) + generate_main( + data_dir, + [ + "--task", + "translation_lev", + "--iter-decode-max-iter", + "9", + "--iter-decode-eos-penalty", + "0", + "--print-step", + ], + ) + + def test_mixture_of_experts(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_moe") as data_dir: + create_dummy_data(data_dir) + preprocess_translation_data(data_dir) + train_translation_model( + data_dir, + "transformer_iwslt_de_en", + [ + "--task", + "translation_moe", + "--user-dir", + "examples/translation_moe/translation_moe_src", + "--method", + "hMoElp", + "--mean-pool-gating-network", + "--num-experts", + "3", + "--encoder-layers", + "2", + "--decoder-layers", + "2", + "--encoder-embed-dim", + "8", + "--decoder-embed-dim", + "8", + ], + ) + generate_main( + data_dir, + [ + "--task", + "translation_moe", + "--user-dir", + "examples/translation_moe/translation_moe_src", + "--method", + "hMoElp", + "--mean-pool-gating-network", + "--num-experts", + "3", + "--gen-expert", + "0", + ], + ) + + def test_alignment(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_alignment") as data_dir: + create_dummy_data(data_dir, alignment=True) + preprocess_translation_data(data_dir, ["--align-suffix", "align"]) + train_translation_model( + data_dir, + "transformer_align", + [ + "--encoder-layers", + "2", + "--decoder-layers", + "2", + "--encoder-embed-dim", + "8", + "--decoder-embed-dim", + "8", + "--load-alignments", + "--alignment-layer", + "1", + "--criterion", + "label_smoothed_cross_entropy_with_alignment", + ], + run_validation=True, + ) + generate_main(data_dir) + + def test_laser_lstm(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_laser_lstm") as data_dir: + laser_config_file = create_laser_data_and_config_json(data_dir) + train_translation_model( + laser_config_file.name, + "laser_lstm", + [ + "--user-dir", + "examples/laser/laser_src", + "--weighting-alpha", + "0.3", + "--encoder-bidirectional", + "--encoder-hidden-size", + "512", + "--encoder-layers", + "5", + "--decoder-layers", + "1", + "--encoder-embed-dim", + "320", + "--decoder-embed-dim", + "320", + "--decoder-lang-embed-dim", + "32", + "--save-dir", + data_dir, + "--disable-validation", + ], + task="laser", + lang_flags=[], + ) + + def test_laser_transformer(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_laser_transformer") as data_dir: + laser_config_file = create_laser_data_and_config_json(data_dir) + train_translation_model( + laser_config_file.name, + "laser_transformer", + [ + "--user-dir", + "examples/laser/laser_src", + "--weighting-alpha", + "0.3", + "--encoder-embed-dim", + "320", + "--decoder-embed-dim", + "320", + "--decoder-lang-embed-dim", + "32", + "--save-dir", + data_dir, + "--disable-validation", + ], + task="laser", + lang_flags=[], + ) + + def test_alignment_full_context(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_alignment") as data_dir: + create_dummy_data(data_dir, alignment=True) + preprocess_translation_data(data_dir, ["--align-suffix", "align"]) + train_translation_model( + data_dir, + "transformer_align", + [ + "--encoder-layers", + "2", + "--decoder-layers", + "2", + "--encoder-embed-dim", + "8", + "--decoder-embed-dim", + "8", + "--load-alignments", + "--alignment-layer", + "1", + "--criterion", + "label_smoothed_cross_entropy_with_alignment", + "--full-context-alignment", + ], + run_validation=True, + ) + generate_main(data_dir) + + def test_transformer_layerdrop(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_transformer_layerdrop") as data_dir: + create_dummy_data(data_dir) + preprocess_translation_data(data_dir) + train_translation_model( + data_dir, + "transformer_iwslt_de_en", + [ + "--encoder-layers", + "3", + "--decoder-layers", + "3", + "--encoder-embed-dim", + "8", + "--decoder-embed-dim", + "8", + "--encoder-layerdrop", + "0.01", + "--decoder-layerdrop", + "0.01", + ], + ) + generate_main(data_dir) + generate_main( + data_dir, + [ + "--model-overrides", + "{'encoder_layers_to_keep':'0,2','decoder_layers_to_keep':'1'}", + ], + ) + + +class TestStories(unittest.TestCase): + def setUp(self): + logging.disable(logging.CRITICAL) + + def tearDown(self): + logging.disable(logging.NOTSET) + + def test_fconv_self_att_wp(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_fconv_self_att_wp") as data_dir: + create_dummy_data(data_dir) + preprocess_translation_data(data_dir) + config = [ + "--encoder-layers", + "[(128, 3)] * 2", + "--decoder-layers", + "[(128, 3)] * 2", + "--decoder-attention", + "True", + "--encoder-attention", + "False", + "--gated-attention", + "True", + "--self-attention", + "True", + "--project-input", + "True", + "--encoder-embed-dim", + "8", + "--decoder-embed-dim", + "8", + "--decoder-out-embed-dim", + "8", + "--multihead-self-attention-nheads", + "2", + ] + train_translation_model(data_dir, "fconv_self_att_wp", config) + generate_main(data_dir) + + # fusion model + os.rename( + os.path.join(data_dir, "checkpoint_last.pt"), + os.path.join(data_dir, "pretrained.pt"), + ) + config.extend( + [ + "--pretrained", + "True", + "--pretrained-checkpoint", + os.path.join(data_dir, "pretrained.pt"), + "--save-dir", + os.path.join(data_dir, "fusion_model"), + ] + ) + train_translation_model(data_dir, "fconv_self_att_wp", config) + + +class TestLanguageModeling(unittest.TestCase): + def setUp(self): + logging.disable(logging.CRITICAL) + + def tearDown(self): + logging.disable(logging.NOTSET) + + def test_fconv_lm(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_fconv_lm") as data_dir: + create_dummy_data(data_dir) + preprocess_lm_data(data_dir) + train_language_model( + data_dir, + "fconv_lm", + [ + "--decoder-layers", + "[(850, 3)] * 2 + [(1024,4)]", + "--decoder-embed-dim", + "280", + "--optimizer", + "nag", + "--lr", + "0.1", + ], + ) + eval_lm_main(data_dir) + generate_main( + data_dir, + [ + "--task", + "language_modeling", + "--sample-break-mode", + "eos", + "--tokens-per-sample", + "500", + ], + ) + + def test_transformer_lm(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_transformer_lm") as data_dir: + create_dummy_data(data_dir) + preprocess_lm_data(data_dir) + train_language_model( + data_dir, + "transformer_lm", + ["--add-bos-token", '--nval', '1'], + run_validation=True, + ) + eval_lm_main(data_dir) + eval_lm_main(data_dir, extra_flags=["--context-window", "25"]) + generate_main( + data_dir, + [ + "--task", + "language_modeling", + "--sample-break-mode", + "eos", + "--tokens-per-sample", + "500", + ], + ) + + def test_transformer_lm_with_adaptive_softmax(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory( + "test_transformer_lm_with_adaptive_softmax" + ) as data_dir: + create_dummy_data(data_dir) + preprocess_lm_data(data_dir) + train_language_model( + data_dir, + "transformer_lm", + [ + "--add-bos-token", + "--criterion", + "adaptive_loss", + "--adaptive-softmax-cutoff", + "5,10,15", + ], + run_validation=True, + ) + eval_lm_main(data_dir) + generate_main( + data_dir, + [ + "--task", + "language_modeling", + "--sample-break-mode", + "eos", + "--tokens-per-sample", + "500", + ], + ) + + def test_lightconv_lm(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_lightconv_lm") as data_dir: + create_dummy_data(data_dir) + preprocess_lm_data(data_dir) + train_language_model( + data_dir, + "lightconv_lm", + ["--add-bos-token"], + run_validation=True, + ) + eval_lm_main(data_dir) + generate_main( + data_dir, + [ + "--task", + "language_modeling", + "--sample-break-mode", + "eos", + "--tokens-per-sample", + "500", + ], + ) + + def test_lstm_lm(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_lstm_lm") as data_dir: + create_dummy_data(data_dir) + preprocess_lm_data(data_dir) + train_language_model( + data_dir, + "lstm_lm", + ["--add-bos-token"], + run_validation=True, + ) + eval_lm_main(data_dir) + generate_main( + data_dir, + [ + "--task", + "language_modeling", + "--sample-break-mode", + "eos", + "--tokens-per-sample", + "500", + ], + ) + + def test_lstm_lm_residuals(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_lstm_lm_residuals") as data_dir: + create_dummy_data(data_dir) + preprocess_lm_data(data_dir) + train_language_model( + data_dir, + "lstm_lm", + ["--add-bos-token", "--residuals"], + run_validation=True, + ) + eval_lm_main(data_dir) + generate_main( + data_dir, + [ + "--task", + "language_modeling", + "--sample-break-mode", + "eos", + "--tokens-per-sample", + "500", + ], + ) + + @unittest.skipIf(not has_hf_transformers, "skip test if transformers is missing") + def test_transformer_xl_bptt_lm(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_transformer_xl_bptt_lm") as data_dir: + create_dummy_data(data_dir) + preprocess_lm_data(data_dir) + task_flags = [ + "--user-dir", + "examples/truncated_bptt", + "--task", + "truncated_bptt_lm", + "--batch-size", + "2", + "--tokens-per-sample", + "50", + ] + train_language_model( + data_dir=data_dir, + arch="transformer_xl", + extra_flags=task_flags + + [ + "--n-layer", + "2", + ], + task="truncated_bptt_lm", + run_validation=True, + extra_valid_flags=task_flags, + ) + eval_lm_main(data_dir, extra_flags=task_flags) + # Train with activation offloading + train_language_model( + data_dir=data_dir, + arch="transformer_xl", + extra_flags=task_flags + + [ + "--n-layer", + "2", + "--offload-activations", + ], + task="truncated_bptt_lm", + run_validation=True, + extra_valid_flags=task_flags, + ) + + +class TestMaskedLanguageModel(unittest.TestCase): + def setUp(self): + logging.disable(logging.CRITICAL) + + def tearDown(self): + logging.disable(logging.NOTSET) + + def test_legacy_masked_lm(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_legacy_mlm") as data_dir: + create_dummy_data(data_dir) + preprocess_lm_data(data_dir) + train_legacy_masked_language_model(data_dir, "masked_lm") + + def test_roberta_masked_lm(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_roberta_mlm") as data_dir: + create_dummy_data(data_dir) + preprocess_lm_data(data_dir) + train_masked_lm( + data_dir, "roberta_base", extra_flags=["--encoder-layers", "2"] + ) + + def test_roberta_sentence_prediction(self): + num_classes = 3 + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_roberta_head") as data_dir: + create_dummy_roberta_head_data(data_dir, num_classes=num_classes) + preprocess_lm_data(os.path.join(data_dir, "input0")) + preprocess_lm_data(os.path.join(data_dir, "label")) + train_roberta_head(data_dir, "roberta_base", num_classes=num_classes) + + def test_roberta_regression_single(self): + num_classes = 1 + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory( + "test_roberta_regression_single" + ) as data_dir: + create_dummy_roberta_head_data( + data_dir, num_classes=num_classes, regression=True + ) + preprocess_lm_data(os.path.join(data_dir, "input0")) + train_roberta_head( + data_dir, + "roberta_base", + num_classes=num_classes, + extra_flags=["--regression-target"], + ) + + def test_roberta_regression_multiple(self): + num_classes = 3 + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory( + "test_roberta_regression_multiple" + ) as data_dir: + create_dummy_roberta_head_data( + data_dir, num_classes=num_classes, regression=True + ) + preprocess_lm_data(os.path.join(data_dir, "input0")) + train_roberta_head( + data_dir, + "roberta_base", + num_classes=num_classes, + extra_flags=["--regression-target"], + ) + + def test_linformer_roberta_masked_lm(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_linformer_roberta_mlm") as data_dir: + create_dummy_data(data_dir) + preprocess_lm_data(data_dir) + train_masked_lm( + data_dir, + "linformer_roberta_base", + extra_flags=[ + "--user-dir", + "examples/linformer/linformer_src", + "--encoder-layers", + "2", + ], + ) + + def test_linformer_roberta_sentence_prediction(self): + num_classes = 3 + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_linformer_roberta_head") as data_dir: + create_dummy_roberta_head_data(data_dir, num_classes=num_classes) + preprocess_lm_data(os.path.join(data_dir, "input0")) + preprocess_lm_data(os.path.join(data_dir, "label")) + train_roberta_head( + data_dir, + "linformer_roberta_base", + num_classes=num_classes, + extra_flags=["--user-dir", "examples/linformer/linformer_src"], + ) + + def test_linformer_roberta_regression_single(self): + num_classes = 1 + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory( + "test_linformer_roberta_regression_single" + ) as data_dir: + create_dummy_roberta_head_data( + data_dir, num_classes=num_classes, regression=True + ) + preprocess_lm_data(os.path.join(data_dir, "input0")) + train_roberta_head( + data_dir, + "linformer_roberta_base", + num_classes=num_classes, + extra_flags=[ + "--regression-target", + "--user-dir", + "examples/linformer/linformer_src", + ], + ) + + def test_linformer_roberta_regression_multiple(self): + num_classes = 3 + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory( + "test_linformer_roberta_regression_multiple" + ) as data_dir: + create_dummy_roberta_head_data( + data_dir, num_classes=num_classes, regression=True + ) + preprocess_lm_data(os.path.join(data_dir, "input0")) + train_roberta_head( + data_dir, + "linformer_roberta_base", + num_classes=num_classes, + extra_flags=[ + "--regression-target", + "--user-dir", + "examples/linformer/linformer_src", + ], + ) + + def _test_pretrained_masked_lm_for_translation(self, learned_pos_emb, encoder_only): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_mlm") as data_dir: + create_dummy_data(data_dir) + preprocess_lm_data(data_dir) + train_legacy_masked_language_model( + data_dir, + arch="masked_lm", + extra_args=("--encoder-learned-pos",) if learned_pos_emb else (), + ) + with tempfile.TemporaryDirectory( + "test_mlm_translation" + ) as translation_dir: + create_dummy_data(translation_dir) + preprocess_translation_data( + translation_dir, extra_flags=["--joined-dictionary"] + ) + # Train transformer with data_dir/checkpoint_last.pt + train_translation_model( + translation_dir, + arch="transformer_from_pretrained_xlm", + extra_flags=[ + "--decoder-layers", + "1", + "--decoder-embed-dim", + "32", + "--decoder-attention-heads", + "1", + "--decoder-ffn-embed-dim", + "32", + "--encoder-layers", + "1", + "--encoder-embed-dim", + "32", + "--encoder-attention-heads", + "1", + "--encoder-ffn-embed-dim", + "32", + "--pretrained-xlm-checkpoint", + "{}/checkpoint_last.pt".format(data_dir), + "--activation-fn", + "gelu", + "--max-source-positions", + "500", + "--max-target-positions", + "500", + ] + + ( + ["--encoder-learned-pos", "--decoder-learned-pos"] + if learned_pos_emb + else [] + ) + + (["--init-encoder-only"] if encoder_only else []), + task="translation_from_pretrained_xlm", + ) + + def test_pretrained_masked_lm_for_translation_learned_pos_emb(self): + self._test_pretrained_masked_lm_for_translation(True, False) + + def test_pretrained_masked_lm_for_translation_sinusoidal_pos_emb(self): + self._test_pretrained_masked_lm_for_translation(False, False) + + def test_pretrained_masked_lm_for_translation_encoder_only(self): + self._test_pretrained_masked_lm_for_translation(True, True) + + def test_r4f_roberta(self): + num_classes = 3 + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_r4f_roberta_head") as data_dir: + create_dummy_roberta_head_data(data_dir, num_classes=num_classes) + preprocess_lm_data(os.path.join(data_dir, "input0")) + preprocess_lm_data(os.path.join(data_dir, "label")) + train_roberta_head( + data_dir, + "roberta_base", + num_classes=num_classes, + extra_flags=[ + "--user-dir", + "examples/rxf/rxf_src", + "--criterion", + "sentence_prediction_r3f", + "--spectral-norm-classification-head", + ], + ) + + +def train_legacy_masked_language_model(data_dir, arch, extra_args=()): + train_parser = options.get_training_parser() + # TODO: langs should be in and out right? + train_args = options.parse_args_and_arch( + train_parser, + [ + "--task", + "cross_lingual_lm", + data_dir, + "--arch", + arch, + # Optimizer args + "--optimizer", + "adam", + "--lr-scheduler", + "reduce_lr_on_plateau", + "--lr-shrink", + "0.5", + "--lr", + "0.0001", + "--stop-min-lr", + "1e-09", + # dropout, attention args + "--dropout", + "0.1", + "--attention-dropout", + "0.1", + # MLM args + "--criterion", + "legacy_masked_lm_loss", + "--masked-lm-only", + "--monolingual-langs", + "in,out", + "--num-segment", + "5", + # Transformer args: use a small transformer model for fast training + "--encoder-layers", + "1", + "--encoder-embed-dim", + "32", + "--encoder-attention-heads", + "1", + "--encoder-ffn-embed-dim", + "32", + # Other training args + "--max-tokens", + "500", + "--tokens-per-sample", + "500", + "--save-dir", + data_dir, + "--max-epoch", + "1", + "--no-progress-bar", + "--distributed-world-size", + "1", + "--dataset-impl", + "raw", + "--num-workers", + "0", + ] + + list(extra_args), + ) + train.main(train_args) + + +class TestOptimizers(unittest.TestCase): + def setUp(self): + logging.disable(logging.CRITICAL) + + def tearDown(self): + logging.disable(logging.NOTSET) + + def test_optimizers(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_optimizers") as data_dir: + # Use just a bit of data and tiny model to keep this test runtime reasonable + create_dummy_data(data_dir, num_examples=10, maxlen=5) + preprocess_translation_data(data_dir) + optimizers = ["adafactor", "adam", "nag", "adagrad", "sgd", "adadelta"] + last_checkpoint = os.path.join(data_dir, "checkpoint_last.pt") + for optimizer in optimizers: + if os.path.exists(last_checkpoint): + os.remove(last_checkpoint) + train_translation_model( + data_dir, + "lstm", + [ + "--required-batch-size-multiple", + "1", + "--encoder-layers", + "1", + "--encoder-hidden-size", + "32", + "--decoder-layers", + "1", + "--optimizer", + optimizer, + ], + ) + generate_main(data_dir) + + +def read_last_log_entry( + logs: List[logging.LogRecord], logger_name: str +) -> Dict[str, float]: + for x in reversed(logs): + if x.name == logger_name: + return json.loads(x.message) + raise ValueError(f"No entries from {logger_name} found in captured logs") + + +class TestActivationCheckpointing(unittest.TestCase): + base_flags = [ + "--encoder-layers", + "2", + "--decoder-layers", + "2", + "--encoder-embed-dim", + "8", + "--decoder-embed-dim", + "8", + "--restore-file", + "x.pt", + "--log-format", + "json", + "--log-interval", + "1", + "--max-update", + "2", + ] + + def _train(self, data_dir, extra_flags): + with self.assertLogs() as logs: + train_translation_model( + data_dir, + "transformer_iwslt_de_en", + self.base_flags + extra_flags, + run_validation=True, + extra_valid_flags=["--log-format", "json"], + ) + return logs.records + + def test_activation_offloading_does_not_change_metrics(self): + """Neither ----checkpoint-activations nor --offload-activations should change loss""" + with tempfile.TemporaryDirectory("test_transformer_with_act_cpt") as data_dir: + + with self.assertLogs(): + create_dummy_data(data_dir, num_examples=20) + preprocess_translation_data(data_dir) + offload_logs = self._train(data_dir, ["--offload-activations"]) + baseline_logs = self._train(data_dir, []) + + assert len(baseline_logs) == len(offload_logs) + + baseline_valid_stats = read_last_log_entry(baseline_logs, "valid") + offload_valid_stats = read_last_log_entry(offload_logs, "valid") + baseline_train_stats = read_last_log_entry(baseline_logs, "train") + offload_train_stats = read_last_log_entry(offload_logs, "train") + + assert ( + baseline_train_stats["train_loss"] == offload_train_stats["train_loss"] + ) + assert ( + baseline_valid_stats["valid_loss"] == offload_valid_stats["valid_loss"] + ) + + def test_activation_checkpointing_does_not_change_metrics(self): + """--checkpoint-activations should not change loss""" + + with tempfile.TemporaryDirectory("test_transformer_with_act_cpt") as data_dir: + with self.assertLogs(): + create_dummy_data(data_dir, num_examples=20) + preprocess_translation_data(data_dir) + ckpt_logs = self._train(data_dir, ["--checkpoint-activations"]) + baseline_logs = self._train(data_dir, []) + assert len(baseline_logs) == len(ckpt_logs) + + baseline_train_stats = read_last_log_entry(baseline_logs, "train") + ckpt_train_stats = read_last_log_entry(ckpt_logs, "train") + assert baseline_train_stats["train_loss"] == ckpt_train_stats["train_loss"] + + baseline_valid_stats = read_last_log_entry(baseline_logs, "valid") + ckpt_valid_stats = read_last_log_entry(ckpt_logs, "valid") + assert baseline_valid_stats["valid_loss"] == ckpt_valid_stats["valid_loss"] + + +def create_dummy_roberta_head_data( + data_dir, num_examples=100, maxlen=10, num_classes=2, regression=False +): + input_dir = "input0" + + def _create_dummy_data(filename): + random_data = torch.rand(num_examples * maxlen) + input_data = 97 + torch.floor(26 * random_data).int() + if regression: + output_data = torch.rand((num_examples, num_classes)) + else: + output_data = 1 + torch.floor(num_classes * torch.rand(num_examples)).int() + with open(os.path.join(data_dir, input_dir, filename + ".out"), "w") as f_in: + label_filename = filename + ".label" if regression else filename + ".out" + with open(os.path.join(data_dir, "label", label_filename), "w") as f_out: + offset = 0 + for i in range(num_examples): + # write example input + ex_len = random.randint(1, maxlen) + ex_str = " ".join(map(chr, input_data[offset : offset + ex_len])) + print(ex_str, file=f_in) + # write example label + if regression: + class_str = " ".join(map(str, output_data[i].numpy())) + print(class_str, file=f_out) + else: + class_str = "class{}".format(output_data[i]) + print(class_str, file=f_out) + offset += ex_len + + os.mkdir(os.path.join(data_dir, input_dir)) + os.mkdir(os.path.join(data_dir, "label")) + _create_dummy_data("train") + _create_dummy_data("valid") + _create_dummy_data("test") + + +def train_masked_lm(data_dir, arch, extra_flags=None): + train_parser = options.get_training_parser() + train_args = options.parse_args_and_arch( + train_parser, + [ + "--task", + "masked_lm", + data_dir, + "--arch", + arch, + "--optimizer", + "adam", + "--lr", + "0.0001", + "--criterion", + "masked_lm", + "--batch-size", + "500", + "--save-dir", + data_dir, + "--max-epoch", + "1", + "--no-progress-bar", + "--distributed-world-size", + "1", + "--ddp-backend", + "no_c10d", + "--num-workers", + "0", + ] + + (extra_flags or []), + ) + train.main(train_args) + + +def train_roberta_head(data_dir, arch, num_classes=2, extra_flags=None): + train_parser = options.get_training_parser() + train_args = options.parse_args_and_arch( + train_parser, + [ + "--task", + "sentence_prediction", + data_dir, + "--arch", + arch, + "--encoder-layers", + "2", + "--num-classes", + str(num_classes), + "--optimizer", + "adam", + "--lr", + "0.0001", + "--criterion", + "sentence_prediction", + "--max-tokens", + "500", + "--max-positions", + "500", + "--batch-size", + "500", + "--save-dir", + data_dir, + "--max-epoch", + "1", + "--no-progress-bar", + "--distributed-world-size", + "1", + "--ddp-backend", + "no_c10d", + "--num-workers", + "0", + ] + + (extra_flags or []), + ) + train.main(train_args) + + +def eval_lm_main(data_dir, extra_flags=None): + eval_lm_parser = options.get_eval_lm_parser() + eval_lm_args = options.parse_args_and_arch( + eval_lm_parser, + [ + data_dir, + "--path", + os.path.join(data_dir, "checkpoint_last.pt"), + "--no-progress-bar", + "--num-workers", + "0", + ] + + (extra_flags or []), + ) + eval_lm.main(eval_lm_args) + + +if __name__ == "__main__": + unittest.main() diff --git a/tests/test_character_token_embedder.py b/tests/test_character_token_embedder.py new file mode 100644 index 0000000000000000000000000000000000000000..24940ebd21a0e4465ca6052409353a3179e9cf6d --- /dev/null +++ b/tests/test_character_token_embedder.py @@ -0,0 +1,48 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import unittest + +import torch +from fairseq.data import Dictionary +from fairseq.modules import CharacterTokenEmbedder + + +class TestCharacterTokenEmbedder(unittest.TestCase): + def test_character_token_embedder(self): + vocab = Dictionary() + vocab.add_symbol("hello") + vocab.add_symbol("there") + + embedder = CharacterTokenEmbedder( + vocab, [(2, 16), (4, 32), (8, 64), (16, 2)], 64, 5, 2 + ) + + test_sents = [["hello", "unk", "there"], ["there"], ["hello", "there"]] + max_len = max(len(s) for s in test_sents) + input = torch.LongTensor(len(test_sents), max_len + 2).fill_(vocab.pad()) + for i in range(len(test_sents)): + input[i][0] = vocab.eos() + for j in range(len(test_sents[i])): + input[i][j + 1] = vocab.index(test_sents[i][j]) + input[i][j + 2] = vocab.eos() + embs = embedder(input) + + assert embs.size() == (len(test_sents), max_len + 2, 5) + self.assertAlmostEqual(embs[0][0], embs[1][0]) + self.assertAlmostEqual(embs[0][0], embs[0][-1]) + self.assertAlmostEqual(embs[0][1], embs[2][1]) + self.assertAlmostEqual(embs[0][3], embs[1][1]) + + embs.sum().backward() + assert embedder.char_embeddings.weight.grad is not None + + def assertAlmostEqual(self, t1, t2): + self.assertEqual(t1.size(), t2.size(), "size mismatch") + self.assertLess((t1 - t2).abs().max(), 1e-6) + + +if __name__ == "__main__": + unittest.main() diff --git a/tests/test_checkpoint_utils.py b/tests/test_checkpoint_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..0f28222633a68943497616507ce412ead76864d6 --- /dev/null +++ b/tests/test_checkpoint_utils.py @@ -0,0 +1,106 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import contextlib +import logging +import os +import tempfile +import unittest +from io import StringIO +from unittest.mock import patch + +from fairseq import checkpoint_utils +from omegaconf import OmegaConf + +from tests.utils import ( + create_dummy_data, + preprocess_translation_data, + train_translation_model, +) + + +class TestCheckpointUtils(unittest.TestCase): + def setUp(self): + logging.disable(logging.CRITICAL) + + def tearDown(self): + logging.disable(logging.NOTSET) + + @contextlib.contextmanager + def _train_transformer(self, seed, extra_args=None): + if extra_args is None: + extra_args = [] + with tempfile.TemporaryDirectory(f"_train_transformer_seed{seed}") as data_dir: + create_dummy_data(data_dir) + preprocess_translation_data(data_dir) + train_translation_model( + data_dir, + "transformer_iwslt_de_en", + [ + "--encoder-layers", + "3", + "--decoder-layers", + "3", + "--encoder-embed-dim", + "8", + "--decoder-embed-dim", + "8", + "--seed", + str(seed), + ] + + extra_args, + ) + yield os.path.join(data_dir, "checkpoint_last.pt") + + def test_load_model_ensemble_and_task(self): + # with contextlib.redirect_stdout(StringIO()): + with self._train_transformer(seed=123) as model1: + with self._train_transformer(seed=456) as model2: + ensemble, cfg, task = checkpoint_utils.load_model_ensemble_and_task( + filenames=[model1, model2] + ) + self.assertEqual(len(ensemble), 2) + + # after Transformer has been migrated to Hydra, this will probably + # become cfg.common.seed + self.assertEqual(ensemble[0].args.seed, 123) + self.assertEqual(ensemble[1].args.seed, 456) + + # the task from the first model should be returned + self.assertTrue("seed123" in task.cfg.data) + + # last cfg is saved + self.assertEqual(cfg.common.seed, 456) + + def test_prune_state_dict(self): + with contextlib.redirect_stdout(StringIO()): + extra_args = ["--encoder-layerdrop", "0.01", "--decoder-layerdrop", "0.01"] + with self._train_transformer(seed=1, extra_args=extra_args) as model: + ensemble, cfg, task = checkpoint_utils.load_model_ensemble_and_task( + filenames=[model], + arg_overrides={ + "encoder_layers_to_keep": "0,2", + "decoder_layers_to_keep": "1", + }, + ) + self.assertEqual(len(ensemble), 1) + self.assertEqual(len(ensemble[0].encoder.layers), 2) + self.assertEqual(len(ensemble[0].decoder.layers), 1) + + def test_torch_persistent_save_async(self): + state_dict = {} + filename = "async_checkpoint.pt" + + with patch(f"{checkpoint_utils.__name__}.PathManager.opena") as mock_opena: + with patch(f"{checkpoint_utils.__name__}._torch_persistent_save") as mock_save: + checkpoint_utils.torch_persistent_save( + state_dict, filename, async_write=True + ) + mock_opena.assert_called_with(filename, "wb") + mock_save.assert_called() + + +if __name__ == "__main__": + unittest.main() diff --git a/tests/test_concat_dataset.py b/tests/test_concat_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d94aeffd481a2e107eb5747e41d76435b3f3dc8a --- /dev/null +++ b/tests/test_concat_dataset.py @@ -0,0 +1,58 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import unittest + +import torch +from fairseq.data import LanguagePairDataset, TokenBlockDataset +from fairseq.data.concat_dataset import ConcatDataset +from tests.test_train import mock_dict + + +class TestConcatDataset(unittest.TestCase): + def setUp(self): + d = mock_dict() + tokens_1 = torch.LongTensor([1]).view(1, -1) + tokens_ds1 = TokenBlockDataset( + tokens_1, + sizes=[tokens_1.size(-1)], + block_size=1, + pad=0, + eos=1, + include_targets=False, + ) + self.dataset_1 = LanguagePairDataset( + tokens_ds1, tokens_ds1.sizes, d, shuffle=False + ) + tokens_2 = torch.LongTensor([2]).view(1, -1) + tokens_ds2 = TokenBlockDataset( + tokens_2, + sizes=[tokens_2.size(-1)], + block_size=1, + pad=0, + eos=1, + include_targets=False, + ) + self.dataset_2 = LanguagePairDataset( + tokens_ds2, tokens_ds2.sizes, d, shuffle=False + ) + + def test_concat_dataset_basics(self): + d = ConcatDataset([self.dataset_1, self.dataset_2]) + assert len(d) == 2 + assert d[0]["source"][0] == 1 + assert d[1]["source"][0] == 2 + + d = ConcatDataset([self.dataset_1, self.dataset_2], sample_ratios=[1, 2]) + assert len(d) == 3 + assert d[0]["source"][0] == 1 + assert d[1]["source"][0] == 2 + assert d[2]["source"][0] == 2 + + d = ConcatDataset([self.dataset_1, self.dataset_2], sample_ratios=[2, 1]) + assert len(d) == 3 + assert d[0]["source"][0] == 1 + assert d[1]["source"][0] == 1 + assert d[2]["source"][0] == 2 diff --git a/tests/test_constraints.py b/tests/test_constraints.py new file mode 100755 index 0000000000000000000000000000000000000000..1c37f7e1fb26d8ea5349fedd3a60f566d09cf598 --- /dev/null +++ b/tests/test_constraints.py @@ -0,0 +1,269 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import sys +import unittest + +import torch +from fairseq.token_generation_constraints import * + + +def tensorize(constraints: List[List[int]]) -> torch.Tensor: + return [torch.tensor(x) for x in constraints] + + +class TestHelperRoutines(unittest.TestCase): + def setUp(self): + self.examples = [ + ([[]], torch.tensor([[0]])), + ([[], []], torch.tensor([[0], [0]])), + ([[torch.tensor([1, 2])], []], torch.tensor([[1, 1, 2, 0], [0, 0, 0, 0]])), + ( + [ + [ + torch.tensor([3, 1, 2]), + torch.tensor([3]), + torch.tensor([4, 5, 6, 7]), + ], + [], + [torch.tensor([1, 8, 9, 10, 1, 4, 11, 12])], + ], + torch.tensor( + [ + [3, 3, 1, 2, 0, 3, 0, 4, 5, 6, 7, 0], + [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], + [1, 1, 8, 9, 10, 1, 4, 11, 12, 0, 0, 0], + ] + ), + ), + ] + + def test_packing(self): + """Ensures the list of lists of tensors gets packed correctly.""" + for batch_constraints, expected_tensor in self.examples: + packed = pack_constraints(batch_constraints) + assert torch.equal(packed, expected_tensor) + + +class TestUnorderedConstraintState(unittest.TestCase): + def setUp(self): + # Tuples of (contraint set, expected printed graph, token counts per node) + self.examples = [ + ( + tensorize([[1, 2, 3], [1, 3], [1, 4], [4, 5, 6, 7], [1], [4, 5]]), + "([None].False#6 ([1].True#4 ([2].False#1 [3].True#1) [3].True#1 [4].True#1) ([4].False#2 ([5].True#2 ([6].False#1 [7].True#1))))", + {1: 4, 2: 1, 3: 2, 4: 3, 5: 2, 6: 1, 7: 1}, + ), + ([], "[None].False#0", {}), + (tensorize([[0]]), "([None].False#1 [0].True#1)", {0: 1}), + ( + tensorize([[100000, 1, 2, 3, 4, 5]]), + "([None].False#1 ([100000].False#1 ([1].False#1 ([2].False#1 ([3].False#1 ([4].False#1 [5].True#1))))))", + {100000: 1, 1: 1, 2: 1, 3: 1, 4: 1, 5: 1}, + ), + ( + tensorize([[1, 2], [1, 2]]), + "([None].False#2 ([1].False#2 [2].True#2))", + {1: 2, 2: 2}, + ), + ( + tensorize([[1, 2], [3, 4]]), + "([None].False#2 ([1].False#1 [2].True#1) ([3].False#1 [4].True#1))", + {1: 1, 2: 1, 3: 1, 4: 1}, + ), + ] + + self.sequences = [ + ( + self.examples[0][0], + [], + {"bank": 0, "num_completed": 0, "finished": False, "is_root": True}, + ), + ( + self.examples[0][0], + [1, 2], + {"bank": 2, "num_completed": 0, "finished": False, "is_root": False}, + ), + ( + self.examples[0][0], + [1, 2, 94], + {"bank": 1, "num_completed": 1, "finished": False, "is_root": True}, + ), + ( + self.examples[0][0], + [1, 3, 999, 1, 4], + {"bank": 4, "num_completed": 2, "finished": False, "is_root": False}, + ), + ( + self.examples[0][0], + [1, 3, 999, 1, 4, 999], + {"bank": 4, "num_completed": 2, "finished": False, "is_root": True}, + ), + ( + self.examples[0][0], + [4, 5, 6, 8], + {"bank": 2, "num_completed": 1, "finished": False, "is_root": True}, + ), + ( + self.examples[0][0], + # Tricky, because in last three, goes down [1->4] branch, could miss [1] and [4->5] + # [[1, 2, 3], [1, 3], [1, 4], [4, 5, 6, 7], [1], [4, 5]], + [1, 2, 3, 1, 3, 1, 4, 4, 5, 6, 7, 1, 4, 5], + {"bank": 14, "num_completed": 6, "finished": True, "is_root": False}, + ), + ( + self.examples[0][0], + [1, 2, 3, 999, 1, 3, 1, 4, 4, 5, 6, 7, 1, 4, 5, 117], + {"bank": 14, "num_completed": 6, "finished": True, "is_root": True}, + ), + ( + tensorize([[1], [2, 3]]), + # Should not be able to get credit for entering 1 a second time + [1, 1], + {"bank": 1, "num_completed": 1, "finished": False, "is_root": True}, + ), + ( + self.examples[4][0], + [1, 2, 1, 2], + {"bank": 4, "num_completed": 2, "finished": True, "is_root": False}, + ), + ( + self.examples[4][0], + [1, 2, 1, 2, 1], + {"bank": 4, "num_completed": 2, "finished": True, "is_root": True}, + ), + ( + self.examples[5][0], + [1, 2, 3, 4, 5], + {"bank": 4, "num_completed": 2, "finished": True, "is_root": True}, + ), + ] + + def test_graphs(self): + """ + Test whether unordered graph systems are created correctly. + """ + for example in self.examples: + constraints, expected, gold_counts = example + c = ConstraintNode.create(constraints) + assert ( + ConstraintNode.print_graph(c) == expected + ), f"got {ConstraintNode.print_graph(c)}, expected {expected}" + assert ( + c.token_counts() == gold_counts + ), f"{c} got {c.token_counts()} wanted {gold_counts}" + + def test_next_tokens(self): + """ + Tests that the set of next tokens is correct. + """ + for example in self.examples: + constraints, expected, gold_counts = example + root = ConstraintNode.create(constraints) + + root_tokens = set(root.children.keys()) + for sequence in constraints: + state = UnorderedConstraintState(root) + for token in sequence: + all_tokens = root_tokens.union(state.node.children.keys()) + assert ( + all_tokens == state.next_tokens() + ), f"ALL {all_tokens} NEXT {state.next_tokens()}" + state = state.advance(token) + + def test_sequences(self): + for constraints, tokens, expected in self.sequences: + state = UnorderedConstraintState.create(pack_constraints([constraints])[0]) + for token in tokens: + state = state.advance(token) + result = {} + for attr in expected.keys(): + result[attr] = getattr(state, attr) + + assert ( + result == expected + ), f"TEST({tokens}) GOT: {result} WANTED: {expected}" + + +class TestOrderedConstraintState(unittest.TestCase): + def setUp(self): + self.sequences = [ + ( + tensorize([[1, 2, 3], [1, 3], [1, 4], [4, 5, 6, 7], [1], [4, 5]]), + [], + {"bank": 0, "num_completed": 0, "finished": False, "is_root": True}, + ), + ( + tensorize([[1, 2, 3], [1, 3], [1, 4], [4, 5, 6, 7], [1], [4, 5]]), + [1, 2], + {"bank": 2, "num_completed": 0, "finished": False, "is_root": False}, + ), + ( + tensorize([[1, 2, 3], [1, 3], [1, 4], [4, 5, 6, 7], [1], [4, 5]]), + [1, 2, 94], + {"bank": 0, "num_completed": 0, "finished": False, "is_root": True}, + ), + ( + tensorize([[1, 2, 3], [1, 3], [1, 4], [4, 5, 6, 7], [1], [4, 5]]), + [1, 3, 999, 1, 4], + {"bank": 0, "num_completed": 0, "finished": False, "is_root": True}, + ), + ( + tensorize([[1, 2, 3], [1, 3], [1, 4], [4, 5, 6, 7], [1], [4, 5]]), + [1, 2, 3, 999, 999], + {"bank": 3, "num_completed": 1, "finished": False, "is_root": False}, + ), + ( + tensorize([[1, 2, 3], [1, 3], [1, 4], [4, 5, 6, 7], [1], [4, 5]]), + [1, 2, 3, 77, 1, 3, 1], + {"bank": 6, "num_completed": 2, "finished": False, "is_root": False}, + ), + ( + tensorize([[1, 2, 3], [1, 3], [1, 4], [4, 5, 6, 7], [1], [4, 5]]), + [1, 2, 3, 1, 3, 1, 4, 4, 5, 6, 7, 1, 4, 5], + {"bank": 14, "num_completed": 6, "finished": True, "is_root": False}, + ), + ( + tensorize([[1, 2, 3], [1, 3], [1, 4], [4, 5, 6, 7], [1], [4, 5]]), + [1, 2, 999, 1, 2, 3, 999, 1, 3, 1, 4, 4, 5, 6, 7, 1, 4, 5, 117], + {"bank": 14, "num_completed": 6, "finished": True, "is_root": False}, + ), + ( + tensorize([[1], [2, 3]]), + [1, 1], + {"bank": 1, "num_completed": 1, "finished": False, "is_root": False}, + ), + ( + tensorize([[1, 2], [1, 2]]), + [1, 2, 1, 2], + {"bank": 4, "num_completed": 2, "finished": True, "is_root": False}, + ), + ( + tensorize([[1, 2], [1, 2]]), + [1, 2, 1, 2, 1], + {"bank": 4, "num_completed": 2, "finished": True, "is_root": False}, + ), + ( + tensorize([[1, 2], [3, 4]]), + [1, 2, 3, 4, 5], + {"bank": 4, "num_completed": 2, "finished": True, "is_root": False}, + ), + ] + + def test_sequences(self): + for i, (constraints, tokens, expected) in enumerate(self.sequences): + state = OrderedConstraintState.create(pack_constraints([constraints])[0]) + for token in tokens: + state = state.advance(token) + result = {} + for attr in expected.keys(): + result[attr] = getattr(state, attr) + assert ( + result == expected + ), f"TEST({tokens}) GOT: {result} WANTED: {expected}" + + +if __name__ == "__main__": + unittest.main() diff --git a/tests/test_convtbc.py b/tests/test_convtbc.py new file mode 100644 index 0000000000000000000000000000000000000000..3a3c9b91e70f597ab77b9b01459cc429db5d7956 --- /dev/null +++ b/tests/test_convtbc.py @@ -0,0 +1,54 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import unittest + +import torch +import torch.nn as nn +from fairseq.modules import ConvTBC + + +class TestConvTBC(unittest.TestCase): + def test_convtbc(self): + # ksz, in_channels, out_channels + conv_tbc = ConvTBC(4, 5, kernel_size=3, padding=1) + # out_channels, in_channels, ksz + conv1d = nn.Conv1d(4, 5, kernel_size=3, padding=1) + + conv_tbc.weight.data.copy_(conv1d.weight.data.transpose(0, 2)) + conv_tbc.bias.data.copy_(conv1d.bias.data) + + input_tbc = torch.randn(7, 2, 4, requires_grad=True) + input1d = input_tbc.data.transpose(0, 1).transpose(1, 2) + input1d.requires_grad = True + + output_tbc = conv_tbc(input_tbc) + output1d = conv1d(input1d) + + self.assertAlmostEqual( + output_tbc.data.transpose(0, 1).transpose(1, 2), output1d.data + ) + + grad_tbc = torch.randn(output_tbc.size()) + grad1d = grad_tbc.transpose(0, 1).transpose(1, 2).contiguous() + + output_tbc.backward(grad_tbc) + output1d.backward(grad1d) + + self.assertAlmostEqual( + conv_tbc.weight.grad.data.transpose(0, 2), conv1d.weight.grad.data + ) + self.assertAlmostEqual(conv_tbc.bias.grad.data, conv1d.bias.grad.data) + self.assertAlmostEqual( + input_tbc.grad.data.transpose(0, 1).transpose(1, 2), input1d.grad.data + ) + + def assertAlmostEqual(self, t1, t2): + self.assertEqual(t1.size(), t2.size(), "size mismatch") + self.assertLess((t1 - t2).abs().max(), 1e-4) + + +if __name__ == "__main__": + unittest.main() diff --git a/tests/test_data_utils.py b/tests/test_data_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..2acfc8dc184015ad762db154dd9929f4c4043093 --- /dev/null +++ b/tests/test_data_utils.py @@ -0,0 +1,136 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import unittest + +import numpy as np +from fairseq.data.data_utils_fast import batch_by_size_fn +from fairseq.data.data_utils_fast import batch_by_size_vec + + +class TestBatchBySize(unittest.TestCase): + @classmethod + def batch_by_size_baseline( + cls, + indices, + num_tokens_vec, + max_tokens, + max_sentences, + bsz_mult, + ): + """Simple, reliable and slow implementation of batch by size """ + batches = [] + start = 0 + while start < len(indices): + for end in range(start + 1, len(indices) + 1): + max_val = max(num_tokens_vec[pos] for pos in range(start, end)) + sent_count = end - start + num_tokens = max_val * sent_count + overflow = num_tokens > max_tokens > 0 or sent_count > max_sentences > 0 + terminate = overflow or end == len(indices) + if overflow: + sent_count -= 1 + if terminate: + if sent_count > bsz_mult: + sent_count = sent_count - sent_count % bsz_mult + batches.append(indices[start : start + sent_count]) + start = start + sent_count + break + return batches + + @classmethod + def _get_error_message( + cls, max_sentences, max_tokens, bsz_mult, num_tokens_vec, validation, results + ): + return f"""Reference batch_by_size implementation should produce + same output as the baseline method. + Params: + max_sentences={max_sentences}, + max_tokens={max_tokens}, + bsz_mult={bsz_mult}, + num_tokens_vec={num_tokens_vec}, + expected_batches={validation}, + returned_batches={results}""" + + def _compare_results( + self, + indices_len, + batch_by_size_impl, + max_sentences, + max_tokens, + bsz_mult, + num_tokens_vec, + ): + indices = np.array(list(range(indices_len))) + validation = self.batch_by_size_baseline( + indices, + num_tokens_vec, + max_tokens=max_tokens, + max_sentences=max_sentences, + bsz_mult=bsz_mult, + ) + results = batch_by_size_impl( + indices, + num_tokens_vec, + max_tokens=max_tokens, + max_sentences=max_sentences, + bsz_mult=bsz_mult, + ) + error_msg = self._get_error_message( + max_sentences, max_tokens, bsz_mult, num_tokens_vec, validation, results + ) + self.assertEqual(len(validation), len(results), error_msg) + for first, second in zip(validation, results): + self.assertTrue(np.array_equal(first, second), error_msg) + + def _run_compare_with_baseline_sweep(self, batch_by_size_impl): + """Compare reference batch_by_size implementation with batch_by_size_baseline + across a dense grid of hyperparam values""" + MAX_MAX_TOKENS = 10 + NUM_TOKENS_VECS_COUNT = 5 + for indices_len in [10, 11]: # try odd and even len of indices + for max_sentences in range(0, indices_len + 2): + for max_tokens in range(0, MAX_MAX_TOKENS): + for bsz_mult in range(1, max(MAX_MAX_TOKENS, indices_len) + 2): + for _ in range(NUM_TOKENS_VECS_COUNT): + num_tokens_vec = np.random.randint( + 0, max_tokens + 1, size=indices_len + ) + self._compare_results( + indices_len, + batch_by_size_impl, + max_sentences, + max_tokens, + bsz_mult, + num_tokens_vec, + ) + + +class TestBatchBySizeVec(TestBatchBySize): + def test_compare_with_baseline(self): + self._run_compare_with_baseline_sweep(batch_by_size_vec) + + +class TestBatchBySizeFn(TestBatchBySize): + def test_compare_with_baseline(self): + def batch_by_size_fn_wrapper( + indices, + num_tokens_vec, + max_tokens, + max_sentences, + bsz_mult, + ): + def num_tokens_fn(idx): + return num_tokens_vec[idx] + + return batch_by_size_fn( + indices, num_tokens_fn, max_tokens, max_sentences, bsz_mult + ) + + self._run_compare_with_baseline_sweep(batch_by_size_fn_wrapper) + + +if __name__ == "__main__": + unittest.main() diff --git a/tests/test_dataset.py b/tests/test_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..a3e3970028bc4b0259153e403951e1735bb0cd3e --- /dev/null +++ b/tests/test_dataset.py @@ -0,0 +1,66 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import unittest +from typing import Sequence + +from fairseq.data import LanguagePairDataset, ListDataset, RoundRobinZipDatasets +from tests.test_train import mock_dict + + +def lang_pair_dataset(lengths: Sequence[int]) -> LanguagePairDataset: + tokens = [[i] * l for i, l in enumerate(lengths)] + return LanguagePairDataset(ListDataset(tokens), lengths, mock_dict()) + + +def sample(id: int, length: int): + return {"id": id, "source": [id] * length, "target": None} + + +class TestDataset(unittest.TestCase): + def setUp(self): + logging.disable(logging.CRITICAL) + + def tearDown(self): + logging.disable(logging.NOTSET) + + def test_round_robin_zip_datasets(self): + long_dataset = lang_pair_dataset([10, 9, 8, 11]) + short_dataset = lang_pair_dataset([11, 9]) + + dataset = RoundRobinZipDatasets({"a": long_dataset, "b": short_dataset}) + # Dataset is now sorted by sentence length + dataset.ordered_indices() + assert dataset.longest_dataset is long_dataset + self.assertEqual(dict(dataset[0]), {"a": sample(2, 8), "b": sample(1, 9)}) + # The item 2 of dataset 'a' is with item (2 % 2 = 0) of dataset 'b' + self.assertEqual(dict(dataset[2]), {"a": sample(0, 10), "b": sample(1, 9)}) + + def test_round_robin_zip_datasets_filtered(self): + long_dataset = lang_pair_dataset([10, 20, 8, 11, 1000, 7, 12]) + short_dataset = lang_pair_dataset([11, 20, 9, 1000]) + + dataset = RoundRobinZipDatasets({"a": long_dataset, "b": short_dataset}) + # Dataset is now sorted by sentence length + idx = dataset.ordered_indices() + idx, _ = dataset.filter_indices_by_size(idx, {"a": 19, "b": 900}) + self.assertEqual(list(idx), [0, 1, 2, 3, 4]) + self.assertEqual(dict(dataset[0]), {"a": sample(5, 7), "b": sample(2, 9)}) + self.assertEqual(dict(dataset[2]), {"a": sample(0, 10), "b": sample(1, 20)}) + self.assertEqual(dict(dataset[4]), {"a": sample(6, 12), "b": sample(0, 11)}) + + def test_round_robin_zip_datasets_filtered_with_tuple(self): + long_dataset = lang_pair_dataset([10, 20, 8, 11, 1000, 7, 12]) + short_dataset = lang_pair_dataset([11, 20, 9, 1000]) + + dataset = RoundRobinZipDatasets({"a": long_dataset, "b": short_dataset}) + # Dataset is now sorted by sentence length + idx = dataset.ordered_indices() + idx, _ = dataset.filter_indices_by_size(idx, 19) + self.assertEqual(list(idx), [0, 1, 2, 3, 4]) + self.assertEqual(dict(dataset[0]), {"a": sample(5, 7), "b": sample(2, 9)}) + self.assertEqual(dict(dataset[2]), {"a": sample(0, 10), "b": sample(2, 9)}) + self.assertEqual(dict(dataset[4]), {"a": sample(6, 12), "b": sample(2, 9)}) diff --git a/tests/test_dictionary.py b/tests/test_dictionary.py new file mode 100644 index 0000000000000000000000000000000000000000..81ce102f4f555822e36298034cdeb3d1c0650255 --- /dev/null +++ b/tests/test_dictionary.py @@ -0,0 +1,116 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import io +import tempfile +import unittest + +import torch +from fairseq.data import Dictionary + + +class TestDictionary(unittest.TestCase): + def test_finalize(self): + txt = [ + "A B C D", + "B C D", + "C D", + "D", + ] + ref_ids1 = list( + map( + torch.IntTensor, + [ + [4, 5, 6, 7, 2], + [5, 6, 7, 2], + [6, 7, 2], + [7, 2], + ], + ) + ) + ref_ids2 = list( + map( + torch.IntTensor, + [ + [7, 6, 5, 4, 2], + [6, 5, 4, 2], + [5, 4, 2], + [4, 2], + ], + ) + ) + + # build dictionary + d = Dictionary() + for line in txt: + d.encode_line(line, add_if_not_exist=True) + + def get_ids(dictionary): + ids = [] + for line in txt: + ids.append(dictionary.encode_line(line, add_if_not_exist=False)) + return ids + + def assertMatch(ids, ref_ids): + for toks, ref_toks in zip(ids, ref_ids): + self.assertEqual(toks.size(), ref_toks.size()) + self.assertEqual(0, (toks != ref_toks).sum().item()) + + ids = get_ids(d) + assertMatch(ids, ref_ids1) + + # check finalized dictionary + d.finalize() + finalized_ids = get_ids(d) + assertMatch(finalized_ids, ref_ids2) + + # write to disk and reload + with tempfile.NamedTemporaryFile(mode="w") as tmp_dict: + d.save(tmp_dict.name) + d = Dictionary.load(tmp_dict.name) + reload_ids = get_ids(d) + assertMatch(reload_ids, ref_ids2) + assertMatch(finalized_ids, reload_ids) + + def test_overwrite(self): + # for example, Camembert overwrites <unk>, <s> and </s> + dict_file = io.StringIO( + "<unk> 999 #fairseq:overwrite\n" + "<s> 999 #fairseq:overwrite\n" + "</s> 999 #fairseq:overwrite\n" + ", 999\n" + "▁de 999\n" + ) + d = Dictionary() + d.add_from_file(dict_file) + self.assertEqual(d.index("<pad>"), 1) + self.assertEqual(d.index("foo"), 3) + self.assertEqual(d.index("<unk>"), 4) + self.assertEqual(d.index("<s>"), 5) + self.assertEqual(d.index("</s>"), 6) + self.assertEqual(d.index(","), 7) + self.assertEqual(d.index("▁de"), 8) + + def test_no_overwrite(self): + # for example, Camembert overwrites <unk>, <s> and </s> + dict_file = io.StringIO( + "<unk> 999\n" "<s> 999\n" "</s> 999\n" ", 999\n" "▁de 999\n" + ) + d = Dictionary() + with self.assertRaisesRegex(RuntimeError, "Duplicate"): + d.add_from_file(dict_file) + + def test_space(self): + # for example, character models treat space as a symbol + dict_file = io.StringIO(" 999\n" "a 999\n" "b 999\n") + d = Dictionary() + d.add_from_file(dict_file) + self.assertEqual(d.index(" "), 4) + self.assertEqual(d.index("a"), 5) + self.assertEqual(d.index("b"), 6) + + +if __name__ == "__main__": + unittest.main() diff --git a/tests/test_export.py b/tests/test_export.py new file mode 100644 index 0000000000000000000000000000000000000000..b380697b9aff8799f90c1e0819e408826ecf2932 --- /dev/null +++ b/tests/test_export.py @@ -0,0 +1,121 @@ +#!/usr/bin/env python3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import tempfile +import unittest + +import torch +from fairseq.data.dictionary import Dictionary +from fairseq.models.transformer import TransformerModel +from fairseq.modules import multihead_attention, sinusoidal_positional_embedding +from fairseq.tasks.fairseq_task import LegacyFairseqTask + + +DEFAULT_TEST_VOCAB_SIZE = 100 + + +class DummyTask(LegacyFairseqTask): + def __init__(self, args): + super().__init__(args) + self.dictionary = get_dummy_dictionary() + if getattr(self.args, "ctc", False): + self.dictionary.add_symbol("<ctc_blank>") + self.src_dict = self.dictionary + self.tgt_dict = self.dictionary + + @property + def source_dictionary(self): + return self.src_dict + + @property + def target_dictionary(self): + return self.dictionary + + +def get_dummy_dictionary(vocab_size=DEFAULT_TEST_VOCAB_SIZE): + dummy_dict = Dictionary() + # add dummy symbol to satisfy vocab size + for id, _ in enumerate(range(vocab_size)): + dummy_dict.add_symbol("{}".format(id), 1000) + return dummy_dict + + +def get_dummy_task_and_parser(): + """ + Return a dummy task and argument parser, which can be used to + create a model/criterion. + """ + parser = argparse.ArgumentParser( + description="test_dummy_s2s_task", argument_default=argparse.SUPPRESS + ) + DummyTask.add_args(parser) + args = parser.parse_args([]) + task = DummyTask.setup_task(args) + return task, parser + + +def _test_save_and_load(scripted_module): + with tempfile.NamedTemporaryFile() as f: + scripted_module.save(f.name) + torch.jit.load(f.name) + + +class TestExportModels(unittest.TestCase): + def test_export_multihead_attention(self): + module = multihead_attention.MultiheadAttention(embed_dim=8, num_heads=2) + scripted = torch.jit.script(module) + _test_save_and_load(scripted) + + def test_incremental_state_multihead_attention(self): + module1 = multihead_attention.MultiheadAttention(embed_dim=8, num_heads=2) + module1 = torch.jit.script(module1) + module2 = multihead_attention.MultiheadAttention(embed_dim=8, num_heads=2) + module2 = torch.jit.script(module2) + + state = {} + state = module1.set_incremental_state(state, "key", {"a": torch.tensor([1])}) + state = module2.set_incremental_state(state, "key", {"a": torch.tensor([2])}) + v1 = module1.get_incremental_state(state, "key")["a"] + v2 = module2.get_incremental_state(state, "key")["a"] + + self.assertEqual(v1, 1) + self.assertEqual(v2, 2) + + def test_positional_embedding(self): + module = sinusoidal_positional_embedding.SinusoidalPositionalEmbedding( + embedding_dim=8, padding_idx=1 + ) + scripted = torch.jit.script(module) + _test_save_and_load(scripted) + + @unittest.skipIf( + torch.__version__ < "1.6.0", "Targeting OSS scriptability for the 1.6 release" + ) + def test_export_transformer(self): + task, parser = get_dummy_task_and_parser() + TransformerModel.add_args(parser) + args = parser.parse_args([]) + model = TransformerModel.build_model(args, task) + scripted = torch.jit.script(model) + _test_save_and_load(scripted) + + @unittest.skipIf( + torch.__version__ < "1.6.0", "Targeting OSS scriptability for the 1.6 release" + ) + def test_export_transformer_no_token_pos_emb(self): + task, parser = get_dummy_task_and_parser() + TransformerModel.add_args(parser) + args = parser.parse_args([]) + args.no_token_positional_embeddings = True + model = TransformerModel.build_model(args, task) + scripted = torch.jit.script(model) + _test_save_and_load(scripted) + + + +if __name__ == "__main__": + unittest.main() diff --git a/tests/test_file_io.py b/tests/test_file_io.py new file mode 100644 index 0000000000000000000000000000000000000000..425812bf1672489093941e5fa09f9da3171559ee --- /dev/null +++ b/tests/test_file_io.py @@ -0,0 +1,58 @@ +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import os +import shutil +import sys +import tempfile +import unittest +from typing import Optional +from unittest.mock import MagicMock + + +class TestFileIO(unittest.TestCase): + + _tmpdir: Optional[str] = None + _tmpfile: Optional[str] = None + _tmpfile_contents = "Hello, World" + + @classmethod + def setUpClass(cls) -> None: + cls._tmpdir = tempfile.mkdtemp() + with open(os.path.join(cls._tmpdir, "test.txt"), "w") as f: + cls._tmpfile = f.name + f.write(cls._tmpfile_contents) + f.flush() + + @classmethod + def tearDownClass(cls) -> None: + # Cleanup temp working dir. + if cls._tmpdir is not None: + shutil.rmtree(cls._tmpdir) # type: ignore + + def test_file_io(self): + from fairseq.file_io import PathManager + + with PathManager.open(os.path.join(self._tmpdir, "test.txt"), "r") as f: + s = f.read() + self.assertEqual(s, self._tmpfile_contents) + + def test_file_io_oss(self): + # Mock iopath to simulate oss environment. + sys.modules["iopath"] = MagicMock() + from fairseq.file_io import PathManager + + with PathManager.open(os.path.join(self._tmpdir, "test.txt"), "r") as f: + s = f.read() + self.assertEqual(s, self._tmpfile_contents) + + def test_file_io_async(self): + # ioPath `PathManager` is initialized after the first `opena` call. + try: + from fairseq.file_io import IOPathManager, PathManager + _asyncfile = os.path.join(self._tmpdir, "async.txt") + f = PathManager.opena(_asyncfile, "wb") + f.close() + + finally: + self.assertTrue(PathManager.async_close()) diff --git a/tests/test_fp16_optimizer.py b/tests/test_fp16_optimizer.py new file mode 100644 index 0000000000000000000000000000000000000000..ce4f1c055ce68b8e3933636fae66cca73c5e9d18 --- /dev/null +++ b/tests/test_fp16_optimizer.py @@ -0,0 +1,112 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import copy +import logging +import unittest + +import torch +from fairseq.optim.fp16_optimizer import FP16Optimizer, MemoryEfficientFP16Optimizer +from omegaconf import OmegaConf + + +@unittest.skipIf(not torch.cuda.is_available(), "test requires a GPU") +class TestGradientScaling(unittest.TestCase): + def setUp(self): + self.x = torch.tensor([2.0]).cuda().half() + weight = 3.0 + bias = 5.0 + self.error = 1.0 + self.target = torch.tensor([self.x * weight + bias + self.error]).cuda().half() + self.loss_fn = torch.nn.L1Loss() + + self.model = torch.nn.Linear(1, 1) + self.model.weight.data = torch.tensor([[weight]]) + self.model.bias.data = torch.tensor([bias]) + self.model.cuda().half() + self.params = list(self.model.parameters()) + + self.cfg_dls = OmegaConf.create( + { + "optimization": { + "lr": [0.1], + }, + "optimizer": { + "_name": "adam", + "lr": [0.1], + "adam_betas": "(0.9, 0.999)", + "adam_eps": 1e-8, + "weight_decay": 0.0, + }, + "common": { + "fp16_init_scale": 1, + "fp16_scale_window": 1, + "fp16_scale_tolerance": 1, + "threshold_loss_scale": 1, + "min_loss_scale": 1e-4, + "tpu": False, + }, + } + ) + logging.disable(logging.CRITICAL) + + def tearDown(self): + logging.disable(logging.NOTSET) + + def run_iter(self, model, params, optimizer): + optimizer.zero_grad() + y = model(self.x) + loss = self.loss_fn(y, self.target) + optimizer.backward(loss) + self.assertEqual(loss, torch.tensor(1.0, device="cuda:0", dtype=torch.float16)) + + grad_norm = optimizer.clip_grad_norm(0) + self.assertAlmostEqual(grad_norm.item(), 2.2361, 4) + + optimizer.step() + self.assertEqual( + model.weight, + torch.tensor( + [[3.0996]], device="cuda:0", dtype=torch.float16, requires_grad=True + ), + ) + self.assertEqual( + model.bias, + torch.tensor( + [5.1016], device="cuda:0", dtype=torch.float16, requires_grad=True + ), + ) + self.assertEqual(optimizer.scaler.loss_scale, 2.0) + + def test_mixed_precision(self): + model = copy.deepcopy(self.model) + params = list(model.parameters()) + optimizer = FP16Optimizer.build_optimizer(self.cfg_dls, params) + + self.run_iter(model, params, optimizer) + self.assertTrue( + all( + torch.all( + fp32_params.eq( + torch.tensor( + [3.1000, 5.1000], device="cuda:0", requires_grad=True + ) + ) + ) + for fp32_params in optimizer.fp32_params.values() + ) + ) + + def test_memory_efficient(self): + model = copy.deepcopy(self.model) + params = list(model.parameters()) + optimizer = MemoryEfficientFP16Optimizer.build_optimizer(self.cfg_dls, params) + + self.run_iter(model, params, optimizer) + + +if __name__ == "__main__": + unittest.main() diff --git a/tests/test_inference_dropout.py b/tests/test_inference_dropout.py new file mode 100644 index 0000000000000000000000000000000000000000..353ac674780a9795492c75aa0a7bc0677b07a9c9 --- /dev/null +++ b/tests/test_inference_dropout.py @@ -0,0 +1,70 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import unittest + +from fairseq.dataclass.utils import convert_namespace_to_omegaconf +from fairseq.models.transformer import TransformerModel +from tests.test_sequence_generator import get_dummy_task_and_parser + + +class TestInferenceDropout(unittest.TestCase): + def setUp(self): + self.task, self.parser = get_dummy_task_and_parser() + TransformerModel.add_args(self.parser) + self.args = self.parser.parse_args([]) + self.args.encoder_layers = 2 + self.args.decoder_layers = 1 + logging.disable(logging.CRITICAL) + + def tearDown(self): + logging.disable(logging.NOTSET) + + def test_sets_inference_dropout_to_true(self): + self.args.retain_dropout = True + self.transformer_model = TransformerModel.build_model(self.args, self.task) + cfg = convert_namespace_to_omegaconf(self.args) + self.transformer_model.prepare_for_inference_(cfg) + assert self.transformer_model.encoder.dropout_module.apply_during_inference + assert self.transformer_model.decoder.dropout_module.apply_during_inference + for layer in self.transformer_model.encoder.layers: + assert layer.dropout_module.apply_during_inference + + def test_inference_dropout_false_by_default(self): + self.transformer_model = TransformerModel.build_model(self.args, self.task) + cfg = convert_namespace_to_omegaconf(self.args) + self.transformer_model.prepare_for_inference_(cfg) + assert not self.transformer_model.encoder.dropout_module.apply_during_inference + assert not self.transformer_model.decoder.dropout_module.apply_during_inference + for layer in self.transformer_model.encoder.layers: + assert not layer.dropout_module.apply_during_inference + for layer in self.transformer_model.decoder.layers: + assert not layer.dropout_module.apply_during_inference + + def test_applies_training_mode(self): + self.transformer_model = TransformerModel.build_model(self.args, self.task) + assert self.transformer_model.encoder.dropout_module.training + for layer in self.transformer_model.encoder.layers: + assert layer.dropout_module.training + + self.transformer_model.eval() + assert not self.transformer_model.decoder.dropout_module.training + for layer in self.transformer_model.encoder.layers: + assert not layer.dropout_module.training + + def test_retain_modules(self): + self.args.retain_dropout = True + self.args.retain_dropout_modules = [ + "TransformerEncoder", + "TransformerEncoderLayer", + ] + self.transformer_model = TransformerModel.build_model(self.args, self.task) + cfg = convert_namespace_to_omegaconf(self.args) + self.transformer_model.prepare_for_inference_(cfg) + assert self.transformer_model.encoder.dropout_module.apply_during_inference + assert not self.transformer_model.decoder.dropout_module.apply_during_inference + for layer in self.transformer_model.decoder.layers: + assert not layer.dropout_module.apply_during_inference diff --git a/tests/test_iopath.py b/tests/test_iopath.py new file mode 100644 index 0000000000000000000000000000000000000000..908261a6619806f7ef9b5dd1beb5d6817b249a6e --- /dev/null +++ b/tests/test_iopath.py @@ -0,0 +1,29 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import unittest +from unittest import mock + + +class TestIOPath(unittest.TestCase): + + def test_no_iopath(self): + from .test_reproducibility import TestReproducibility + + with mock.patch.dict("sys.modules", {"iopath": None}): + # reuse reproducibility tests, which are e2e tests that should cover + # most checkpoint related functionality + TestReproducibility._test_reproducibility(self, "test_reproducibility") + + def test_no_supports_rename(self): + from .test_reproducibility import TestReproducibility + + with mock.patch("fairseq.file_io.PathManager.supports_rename") as mock_fn: + mock_fn.return_value = False + TestReproducibility._test_reproducibility(self, "test_reproducibility") + + +if __name__ == "__main__": + unittest.main() diff --git a/tests/test_iterators.py b/tests/test_iterators.py new file mode 100644 index 0000000000000000000000000000000000000000..7b3dd4848553357e5e8326ed3a31cf5d68ceea94 --- /dev/null +++ b/tests/test_iterators.py @@ -0,0 +1,137 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import unittest + +from fairseq.data import iterators + + +class TestIterators(unittest.TestCase): + def test_counting_iterator_index(self, ref=None, itr=None): + # Test the indexing functionality of CountingIterator + if ref is None: + assert itr is None + ref = list(range(10)) + itr = iterators.CountingIterator(ref) + else: + assert len(ref) == 10 + assert itr is not None + + self.assertTrue(itr.has_next()) + self.assertEqual(itr.n, 0) + self.assertEqual(next(itr), ref[0]) + self.assertEqual(itr.n, 1) + self.assertEqual(next(itr), ref[1]) + self.assertEqual(itr.n, 2) + itr.skip(3) + self.assertEqual(itr.n, 5) + self.assertEqual(next(itr), ref[5]) + itr.skip(2) + self.assertEqual(itr.n, 8) + self.assertEqual(list(itr), [ref[8], ref[9]]) + self.assertFalse(itr.has_next()) + + def test_counting_iterator_length_mismatch(self): + ref = list(range(10)) + # When the underlying iterable is longer than the CountingIterator, + # the remaining items in the iterable should be ignored + itr = iterators.CountingIterator(ref, total=8) + self.assertEqual(list(itr), ref[:8]) + # When the underlying iterable is shorter than the CountingIterator, + # raise an IndexError when the underlying iterable is exhausted + itr = iterators.CountingIterator(ref, total=12) + self.assertRaises(IndexError, list, itr) + + def test_counting_iterator_take(self): + # Test the "take" method of CountingIterator + ref = list(range(10)) + itr = iterators.CountingIterator(ref) + itr.take(5) + self.assertEqual(len(itr), len(list(iter(itr)))) + self.assertEqual(len(itr), 5) + + itr = iterators.CountingIterator(ref) + itr.take(5) + self.assertEqual(next(itr), ref[0]) + self.assertEqual(next(itr), ref[1]) + itr.skip(2) + self.assertEqual(next(itr), ref[4]) + self.assertFalse(itr.has_next()) + + def test_grouped_iterator(self): + # test correctness + x = list(range(10)) + itr = iterators.GroupedIterator(x, 1) + self.assertEqual(list(itr), [[0], [1], [2], [3], [4], [5], [6], [7], [8], [9]]) + itr = iterators.GroupedIterator(x, 4) + self.assertEqual(list(itr), [[0, 1, 2, 3], [4, 5, 6, 7], [8, 9]]) + itr = iterators.GroupedIterator(x, 5) + self.assertEqual(list(itr), [[0, 1, 2, 3, 4], [5, 6, 7, 8, 9]]) + + # test the GroupIterator also works correctly as a CountingIterator + x = list(range(30)) + ref = list(iterators.GroupedIterator(x, 3)) + itr = iterators.GroupedIterator(x, 3) + self.test_counting_iterator_index(ref, itr) + + def test_sharded_iterator(self): + # test correctness + x = list(range(10)) + itr = iterators.ShardedIterator(x, num_shards=1, shard_id=0) + self.assertEqual(list(itr), x) + itr = iterators.ShardedIterator(x, num_shards=2, shard_id=0) + self.assertEqual(list(itr), [0, 2, 4, 6, 8]) + itr = iterators.ShardedIterator(x, num_shards=2, shard_id=1) + self.assertEqual(list(itr), [1, 3, 5, 7, 9]) + itr = iterators.ShardedIterator(x, num_shards=3, shard_id=0) + self.assertEqual(list(itr), [0, 3, 6, 9]) + itr = iterators.ShardedIterator(x, num_shards=3, shard_id=1) + self.assertEqual(list(itr), [1, 4, 7, None]) + itr = iterators.ShardedIterator(x, num_shards=3, shard_id=2) + self.assertEqual(list(itr), [2, 5, 8, None]) + + # test CountingIterator functionality + x = list(range(30)) + ref = list(iterators.ShardedIterator(x, num_shards=3, shard_id=0)) + itr = iterators.ShardedIterator(x, num_shards=3, shard_id=0) + self.test_counting_iterator_index(ref, itr) + + def test_counting_iterator_buffered_iterator_take(self): + ref = list(range(10)) + buffered_itr = iterators.BufferedIterator(2, ref) + itr = iterators.CountingIterator(buffered_itr) + itr.take(5) + self.assertEqual(len(itr), len(list(iter(itr)))) + self.assertEqual(len(itr), 5) + + buffered_itr = iterators.BufferedIterator(2, ref) + itr = iterators.CountingIterator(buffered_itr) + itr.take(5) + self.assertEqual(len(buffered_itr), 5) + self.assertEqual(len(list(iter(buffered_itr))), 5) + + buffered_itr = iterators.BufferedIterator(2, ref) + itr = iterators.CountingIterator(buffered_itr) + itr.take(5) + self.assertEqual(next(itr), ref[0]) + self.assertEqual(next(itr), ref[1]) + itr.skip(2) + self.assertEqual(next(itr), ref[4]) + self.assertFalse(itr.has_next()) + self.assertRaises(StopIteration, next, buffered_itr) + + ref = list(range(4, 10)) + buffered_itr = iterators.BufferedIterator(2, ref) + itr = iterators.CountingIterator(buffered_itr, start=4) + itr.take(5) + self.assertEqual(len(itr), 5) + self.assertEqual(len(buffered_itr), 1) + self.assertEqual(next(itr), ref[0]) + self.assertFalse(itr.has_next()) + self.assertRaises(StopIteration, next, buffered_itr) + + +if __name__ == "__main__": + unittest.main() diff --git a/tests/test_label_smoothing.py b/tests/test_label_smoothing.py new file mode 100644 index 0000000000000000000000000000000000000000..04c0f974ac80f7606327f868e948712c3c18f1d0 --- /dev/null +++ b/tests/test_label_smoothing.py @@ -0,0 +1,123 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import copy +import unittest + +import tests.utils as test_utils +import torch +from fairseq.criterions.cross_entropy import CrossEntropyCriterion +from fairseq.criterions.label_smoothed_cross_entropy import ( + LabelSmoothedCrossEntropyCriterion, +) + + +class TestLabelSmoothing(unittest.TestCase): + def setUp(self): + # build dictionary + self.d = test_utils.dummy_dictionary(3) + vocab = len(self.d) + self.assertEqual(vocab, 4 + 3) # 4 special + 3 tokens + self.assertEqual(self.d.pad(), 1) + self.assertEqual(self.d.eos(), 2) + self.assertEqual(self.d.unk(), 3) + pad, eos, unk, w1, w2, w3 = 1, 2, 3, 4, 5, 6 # noqa: F841 + + # build dataset + self.data = [ + # the first batch item has padding + { + "source": torch.LongTensor([w1, eos]), + "target": torch.LongTensor([w1, eos]), + }, + { + "source": torch.LongTensor([w1, eos]), + "target": torch.LongTensor([w1, w1, eos]), + }, + ] + self.sample = next(test_utils.dummy_dataloader(self.data)) + + # build model + self.args = argparse.Namespace() + self.args.sentence_avg = False + self.args.report_accuracy = False + self.args.probs = ( + torch.FloatTensor( + [ + # pad eos unk w1 w2 w3 + [0.05, 0.05, 0.1, 0.05, 0.3, 0.4, 0.05], + [0.05, 0.10, 0.2, 0.05, 0.2, 0.3, 0.10], + [0.05, 0.15, 0.3, 0.05, 0.1, 0.2, 0.15], + ] + ) + .unsqueeze(0) + .expand(2, 3, 7) + ) # add batch dimension + self.task = test_utils.TestTranslationTask.setup_task(self.args, self.d, self.d) + self.model = self.task.build_model(self.args) + + def test_nll_loss(self): + self.args.label_smoothing = 0.1 + nll_crit = CrossEntropyCriterion.build_criterion(self.args, self.task) + smooth_crit = LabelSmoothedCrossEntropyCriterion.build_criterion( + self.args, self.task + ) + nll_loss, nll_sample_size, nll_logging_output = nll_crit( + self.model, self.sample + ) + smooth_loss, smooth_sample_size, smooth_logging_output = smooth_crit( + self.model, self.sample + ) + self.assertLess(abs(nll_loss - nll_logging_output["loss"]), 1e-6) + self.assertLess(abs(nll_loss - smooth_logging_output["nll_loss"]), 1e-6) + + def test_padding(self): + self.args.label_smoothing = 0.1 + crit = LabelSmoothedCrossEntropyCriterion.build_criterion(self.args, self.task) + loss, _, logging_output = crit(self.model, self.sample) + + def get_one_no_padding(idx): + # create a new sample with just a single batch item so that there's + # no padding + sample1 = next(test_utils.dummy_dataloader([self.data[idx]])) + args1 = copy.copy(self.args) + args1.probs = args1.probs[idx, :, :].unsqueeze(0) + model1 = self.task.build_model(args1) + loss1, _, _ = crit(model1, sample1) + return loss1 + + loss1 = get_one_no_padding(0) + loss2 = get_one_no_padding(1) + self.assertAlmostEqual(loss, loss1 + loss2) + + def test_reduction(self): + self.args.label_smoothing = 0.1 + crit = LabelSmoothedCrossEntropyCriterion.build_criterion(self.args, self.task) + loss, _, logging_output = crit(self.model, self.sample, reduce=True) + unreduced_loss, _, _ = crit(self.model, self.sample, reduce=False) + self.assertAlmostEqual(loss, unreduced_loss.sum()) + + def test_zero_eps(self): + self.args.label_smoothing = 0.0 + nll_crit = CrossEntropyCriterion.build_criterion(self.args, self.task) + smooth_crit = LabelSmoothedCrossEntropyCriterion.build_criterion( + self.args, self.task + ) + nll_loss, nll_sample_size, nll_logging_output = nll_crit( + self.model, self.sample + ) + smooth_loss, smooth_sample_size, smooth_logging_output = smooth_crit( + self.model, self.sample + ) + self.assertAlmostEqual(nll_loss, smooth_loss) + + def assertAlmostEqual(self, t1, t2): + self.assertEqual(t1.size(), t2.size(), "size mismatch") + self.assertLess((t1 - t2).abs().max(), 1e-6) + + +if __name__ == "__main__": + unittest.main() diff --git a/tests/test_lm_context_window.py b/tests/test_lm_context_window.py new file mode 100644 index 0000000000000000000000000000000000000000..7415e86abdf8ddc2d797092bf98f7a1331e038d6 --- /dev/null +++ b/tests/test_lm_context_window.py @@ -0,0 +1,51 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import unittest + +import torch +from fairseq.data import MonolingualDataset +from fairseq.tasks.language_modeling import LanguageModelingTask, LanguageModelingConfig +from tests import utils as test_utils + + +class TestLMContextWindow(unittest.TestCase): + + def test_eval_dataloader(self): + dictionary = test_utils.dummy_dictionary(10) + assert len(dictionary) == 14 # 4 extra special symbols + assert dictionary.pad() == 1 + + dataset = test_utils.TestDataset([ + torch.tensor([4, 5, 6, 7], dtype=torch.long), + torch.tensor([8, 9, 10, 11], dtype=torch.long), + torch.tensor([12, 13], dtype=torch.long), + ]) + dataset = MonolingualDataset(dataset, sizes=[4, 4, 2], src_vocab=dictionary) + + config = LanguageModelingConfig(tokens_per_sample=4) + task = LanguageModelingTask(config, dictionary) + + eval_dataloader = task.eval_lm_dataloader( + dataset=dataset, + batch_size=1, + context_window=2, + ) + + batch = next(eval_dataloader) + assert batch["net_input"]["src_tokens"][0].tolist() == [4, 5, 6, 7, 1, 1] + assert batch["target"][0].tolist() == [4, 5, 6, 7, 1, 1] + + batch = next(eval_dataloader) + assert batch["net_input"]["src_tokens"][0].tolist() == [6, 7, 8, 9, 10, 11] + assert batch["target"][0].tolist() == [1, 1, 8, 9, 10, 11] + + batch = next(eval_dataloader) + assert batch["net_input"]["src_tokens"][0].tolist() == [10, 11, 12, 13] + assert batch["target"][0].tolist() == [1, 1, 12, 13] + + +if __name__ == "__main__": + unittest.main() diff --git a/tests/test_lstm_jitable.py b/tests/test_lstm_jitable.py new file mode 100644 index 0000000000000000000000000000000000000000..38f79d17931c32447e96c0fbae2630ac397e1804 --- /dev/null +++ b/tests/test_lstm_jitable.py @@ -0,0 +1,115 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import tempfile +import unittest + +import torch +from fairseq.data.dictionary import Dictionary +from fairseq.models.lstm import LSTMModel +from fairseq.tasks.fairseq_task import LegacyFairseqTask + + +DEFAULT_TEST_VOCAB_SIZE = 100 + + +class DummyTask(LegacyFairseqTask): + def __init__(self, args): + super().__init__(args) + self.dictionary = get_dummy_dictionary() + if getattr(self.args, "ctc", False): + self.dictionary.add_symbol("<ctc_blank>") + self.src_dict = self.dictionary + self.tgt_dict = self.dictionary + + @property + def source_dictionary(self): + return self.src_dict + + @property + def target_dictionary(self): + return self.dictionary + + +def get_dummy_dictionary(vocab_size=DEFAULT_TEST_VOCAB_SIZE): + dummy_dict = Dictionary() + # add dummy symbol to satisfy vocab size + for id, _ in enumerate(range(vocab_size)): + dummy_dict.add_symbol("{}".format(id), 1000) + return dummy_dict + + +def get_dummy_task_and_parser(): + """ + to build a fariseq model, we need some dummy parse and task. This function + is used to create dummy task and parser to faciliate model/criterion test + + Note: we use FbSpeechRecognitionTask as the dummy task. You may want + to use other task by providing another function + """ + parser = argparse.ArgumentParser( + description="test_dummy_s2s_task", argument_default=argparse.SUPPRESS + ) + DummyTask.add_args(parser) + args = parser.parse_args([]) + task = DummyTask.setup_task(args) + return task, parser + + +class TestJitLSTMModel(unittest.TestCase): + def _test_save_and_load(self, scripted_module): + with tempfile.NamedTemporaryFile() as f: + scripted_module.save(f.name) + torch.jit.load(f.name) + + def assertTensorEqual(self, t1, t2): + t1 = t1[~torch.isnan(t1)] # can cause size mismatch errors if there are NaNs + t2 = t2[~torch.isnan(t2)] + self.assertEqual(t1.size(), t2.size(), "size mismatch") + self.assertEqual(t1.ne(t2).long().sum(), 0) + + def test_jit_and_export_lstm(self): + task, parser = get_dummy_task_and_parser() + LSTMModel.add_args(parser) + args = parser.parse_args([]) + args.criterion = "" + model = LSTMModel.build_model(args, task) + scripted_model = torch.jit.script(model) + self._test_save_and_load(scripted_model) + + def test_assert_jit_vs_nonjit_(self): + task, parser = get_dummy_task_and_parser() + LSTMModel.add_args(parser) + args = parser.parse_args([]) + args.criterion = "" + model = LSTMModel.build_model(args, task) + model.eval() + scripted_model = torch.jit.script(model) + scripted_model.eval() + idx = len(task.source_dictionary) + iter = 100 + # Inject random input and check output + seq_len_tensor = torch.randint(1, 10, (iter,)) + num_samples_tensor = torch.randint(1, 10, (iter,)) + for i in range(iter): + seq_len = seq_len_tensor[i] + num_samples = num_samples_tensor[i] + src_token = (torch.randint(0, idx, (num_samples, seq_len)),) + src_lengths = torch.randint(1, seq_len + 1, (num_samples,)) + src_lengths, _ = torch.sort(src_lengths, descending=True) + # Force the first sample to have seq_len + src_lengths[0] = seq_len + prev_output_token = (torch.randint(0, idx, (num_samples, 1)),) + result = model(src_token[0], src_lengths, prev_output_token[0], None) + scripted_result = scripted_model( + src_token[0], src_lengths, prev_output_token[0], None + ) + self.assertTensorEqual(result[0], scripted_result[0]) + self.assertTensorEqual(result[1], scripted_result[1]) + + +if __name__ == "__main__": + unittest.main() diff --git a/tests/test_memory_efficient_fp16.py b/tests/test_memory_efficient_fp16.py new file mode 100644 index 0000000000000000000000000000000000000000..2bf2f29888d6027896128930626b1aafe7f18475 --- /dev/null +++ b/tests/test_memory_efficient_fp16.py @@ -0,0 +1,78 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import logging +import unittest + +import torch +from fairseq.optim.adam import FairseqAdam +from fairseq.optim.fp16_optimizer import MemoryEfficientFP16Optimizer +from omegaconf import OmegaConf + + +@unittest.skipIf(not torch.cuda.is_available(), "test requires a GPU") +class TestMemoryEfficientFP16(unittest.TestCase): + def setUp(self): + logging.disable(logging.CRITICAL) + + def tearDown(self): + logging.disable(logging.NOTSET) + + def test_load_state_dict(self): + # define simple FP16 model + model = torch.nn.Linear(5, 5).cuda().half() + params = list(model.parameters()) + + # initialize memory efficient FP16 optimizer + # with pseudo DictConfigs + optimizer = FairseqAdam( + cfg=OmegaConf.create( + vars( + argparse.Namespace( + adam_betas="(0.9, 0.999)", + adam_eps=1e-8, + weight_decay=0.0, + lr=[0.00001], + ) + ) + ), + params=params, + ) + me_optimizer = MemoryEfficientFP16Optimizer( + cfg=OmegaConf.create( + { + "common": vars( + argparse.Namespace( + fp16_init_scale=1, + fp16_scale_window=1, + fp16_scale_tolerance=1, + threshold_loss_scale=1, + min_loss_scale=1e-4, + ) + ) + } + ), + params=params, + optimizer=optimizer, + ) + + # optimizer state is created in the first step + loss = model(torch.rand(5).cuda().half()).sum() + me_optimizer.backward(loss) + me_optimizer.step() + + # reload state + state = me_optimizer.state_dict() + me_optimizer.load_state_dict(state) + for k, v in me_optimizer.optimizer.state.items(): + self.assertTrue(k.dtype == torch.float16) + for v_i in v.values(): + if torch.is_tensor(v_i): + self.assertTrue(v_i.dtype == torch.float32) + + +if __name__ == "__main__": + unittest.main() diff --git a/tests/test_metrics.py b/tests/test_metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..2de6969cf4445bc6cda44dacf6de765ea30d5f5b --- /dev/null +++ b/tests/test_metrics.py @@ -0,0 +1,77 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import unittest +import uuid + +from fairseq import metrics + + +class TestMetrics(unittest.TestCase): + def test_nesting(self): + with metrics.aggregate() as a: + metrics.log_scalar("loss", 1) + with metrics.aggregate() as b: + metrics.log_scalar("loss", 2) + + self.assertEqual(a.get_smoothed_values()["loss"], 1.5) + self.assertEqual(b.get_smoothed_values()["loss"], 2) + + def test_new_root(self): + with metrics.aggregate() as a: + metrics.log_scalar("loss", 1) + with metrics.aggregate(new_root=True) as b: + metrics.log_scalar("loss", 2) + + self.assertEqual(a.get_smoothed_values()["loss"], 1) + self.assertEqual(b.get_smoothed_values()["loss"], 2) + + def test_nested_new_root(self): + with metrics.aggregate() as layer1: + metrics.log_scalar("loss", 1) + with metrics.aggregate(new_root=True) as layer2: + metrics.log_scalar("loss", 2) + with metrics.aggregate() as layer3: + metrics.log_scalar("loss", 3) + with metrics.aggregate(new_root=True) as layer4: + metrics.log_scalar("loss", 4) + metrics.log_scalar("loss", 1.5) + + self.assertEqual(layer4.get_smoothed_values()["loss"], 4) + self.assertEqual(layer3.get_smoothed_values()["loss"], 3) + self.assertEqual(layer2.get_smoothed_values()["loss"], 2.5) + self.assertEqual(layer1.get_smoothed_values()["loss"], 1.25) + + def test_named(self): + name = str(uuid.uuid4()) + metrics.reset_meters(name) + + with metrics.aggregate(name): + metrics.log_scalar("loss", 1) + + metrics.log_scalar("loss", 3) + + with metrics.aggregate(name): + metrics.log_scalar("loss", 2) + + self.assertEqual(metrics.get_smoothed_values(name)["loss"], 1.5) + + def test_nested_duplicate_names(self): + name = str(uuid.uuid4()) + metrics.reset_meters(name) + + with metrics.aggregate(name): + metrics.log_scalar("loss", 1) + with metrics.aggregate() as other: + with metrics.aggregate(name): + metrics.log_scalar("loss", 2) + metrics.log_scalar("loss", 6) + + self.assertEqual(metrics.get_smoothed_values(name)["loss"], 3) + self.assertEqual(other.get_smoothed_values()["loss"], 2) + + +if __name__ == "__main__": + unittest.main() diff --git a/tests/test_multi_corpus_dataset.py b/tests/test_multi_corpus_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..5a79f4b680e5bc2c7374ec6dd8ea525c47b40985 --- /dev/null +++ b/tests/test_multi_corpus_dataset.py @@ -0,0 +1,79 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import unittest +from collections import OrderedDict + +import torch +from fairseq.data import LanguagePairDataset, TokenBlockDataset +from fairseq.data.multi_corpus_dataset import MultiCorpusDataset +from tests.test_train import mock_dict + + +class TestMultiCorpusDataset(unittest.TestCase): + def setUp(self): + d = mock_dict() + tokens_1 = torch.LongTensor([i for i in range(1, 5000, 2)]).view(1, -1) + tokens_ds1 = TokenBlockDataset( + tokens_1, + sizes=[tokens_1.size(-1)], + block_size=1, + pad=0, + eos=1, + include_targets=False, + ) + self.dataset_1 = LanguagePairDataset( + tokens_ds1, tokens_ds1.sizes, d, shuffle=False + ) + tokens_2 = torch.LongTensor([i for i in range(0, 5000, 2)]).view(1, -1) + tokens_ds2 = TokenBlockDataset( + tokens_2, + sizes=[tokens_2.size(-1)], + block_size=1, + pad=0, + eos=1, + include_targets=False, + ) + self.dataset_2 = LanguagePairDataset( + tokens_ds2, tokens_ds2.sizes, d, shuffle=False + ) + + def _test_sample_helper( + self, + distribution, + ): + m = MultiCorpusDataset( + OrderedDict({0: self.dataset_1, 1: self.dataset_2}), + distribution=distribution, + seed=0, + sort_indices=True, + ) + m.set_epoch(1) + indices = m.ordered_indices() + count_sample_from_first_dataset = 0 + items = set() + for i in indices: + item = m[i]["source"].item() + if item % 2 == 1: + count_sample_from_first_dataset += 1 + + items.add(item) + sample_from_first_ds_percentage = ( + 1.0 * count_sample_from_first_dataset / len(indices) + ) + self.assertLess( + abs(sample_from_first_ds_percentage - distribution[0]), + 0.01, + ) + self.assertEqual( + len(items), + int(min(len(self.dataset_1), len(indices) * distribution[0]) + + min(len(self.dataset_1), len(indices) * distribution[1])) + ) + print(distribution) + + def test_multi_corpus_dataset(self): + for distribution in [[0.5, 0.5], [0.1, 0.9], [0.9, 0.1]]: + self._test_sample_helper(distribution=distribution) diff --git a/tests/test_multi_corpus_sampled_dataset.py b/tests/test_multi_corpus_sampled_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..05b20328c5605178767d138cc75e070824679842 --- /dev/null +++ b/tests/test_multi_corpus_sampled_dataset.py @@ -0,0 +1,95 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import unittest +from collections import OrderedDict + +import numpy as np +import torch +from fairseq.data import LanguagePairDataset, TokenBlockDataset +from fairseq.data.multi_corpus_sampled_dataset import MultiCorpusSampledDataset +from tests.test_train import mock_dict + + +class TestMultiCorpusSampledDataset(unittest.TestCase): + def setUp(self): + d = mock_dict() + tokens_1 = torch.LongTensor([1]).view(1, -1) + tokens_ds1 = TokenBlockDataset( + tokens_1, + sizes=[tokens_1.size(-1)], + block_size=1, + pad=0, + eos=1, + include_targets=False, + ) + self.dataset_1 = LanguagePairDataset( + tokens_ds1, tokens_ds1.sizes, d, shuffle=False + ) + tokens_2 = torch.LongTensor([2]).view(1, -1) + tokens_ds2 = TokenBlockDataset( + tokens_2, + sizes=[tokens_2.size(-1)], + block_size=1, + pad=0, + eos=1, + include_targets=False, + ) + self.dataset_2 = LanguagePairDataset( + tokens_ds2, tokens_ds2.sizes, d, shuffle=False + ) + + def _test_sample_helper( + self, + expected_sample_from_first_ds_percentage, + num_samples=1000, + sampling_func=None, + ): + # To make sure test is not flaky + np.random.seed(0) + if sampling_func is None: + m = MultiCorpusSampledDataset( + OrderedDict({0: self.dataset_1, 1: self.dataset_2}), + ) + else: + m = MultiCorpusSampledDataset( + OrderedDict({0: self.dataset_1, 1: self.dataset_2}), + sampling_func=sampling_func, + ) + m.ordered_indices() + count_sample_from_first_dataset = 0 + for _ in range(num_samples): + if m.collater([m[0], m[1]])["net_input"]["src_tokens"][0] == 1: + count_sample_from_first_dataset += 1 + sample_from_first_ds_percentage = ( + 1.0 * count_sample_from_first_dataset / num_samples + ) + self.assertLess( + abs( + sample_from_first_ds_percentage + - expected_sample_from_first_ds_percentage + ), + 0.01, + ) + + def test_multi_corpus_sampled_dataset_uniform_sample(self): + self._test_sample_helper(expected_sample_from_first_ds_percentage=0.5) + + def test_multi_corpus_sampled_dataset_weighted_sample(self): + def naive_weighted_sample(weights): + def f(l): + v = np.random.random() + agg = 0 + for i, weight in enumerate(weights): + agg += weight + if agg > v: + return i + + return f + + self._test_sample_helper( + expected_sample_from_first_ds_percentage=0.9, + sampling_func=naive_weighted_sample(weights=[0.9, 0.1]), + ) diff --git a/tests/test_multihead_attention.py b/tests/test_multihead_attention.py new file mode 100644 index 0000000000000000000000000000000000000000..620a2d679147bbbb8d15f3323374a39939686ec2 --- /dev/null +++ b/tests/test_multihead_attention.py @@ -0,0 +1,73 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import unittest + +import torch +from fairseq.modules.multihead_attention import MultiheadAttention + + +class TestMultiheadAttention(unittest.TestCase): + def test_append_prev_key_padding_mask(self): + bsz = 1 + src_len = 4 + + cases = [ + # no padding mask + (None, None, None), + # current padding mask only + ( + torch.tensor([[1]]).bool(), + None, + torch.tensor([[0, 0, 0, 1]]).bool(), + ), + # previous padding mask only + ( + None, + torch.tensor([[0, 1, 0]]).bool(), + torch.tensor([[0, 1, 0, 0]]).bool(), + ), + # both padding masks + ( + torch.tensor([[1]]).bool(), + torch.tensor([[0, 1, 0]]).bool(), + torch.tensor([[0, 1, 0, 1]]).bool(), + ), + # prev_key_padding_mask already full + ( + torch.tensor([[0, 1, 0, 1]]).bool(), + None, + torch.tensor([[0, 1, 0, 1]]).bool(), + ), + # key_padding_mask already full + ( + None, + torch.tensor([[0, 1, 0, 1]]).bool(), + torch.tensor([[0, 1, 0, 1]]).bool(), + ), + ] + for c in cases: + key_padding_mask = MultiheadAttention._append_prev_key_padding_mask( + c[0], + c[1], + batch_size=bsz, + src_len=src_len, + static_kv=False, + ) + + if key_padding_mask is not None: + self.assertTrue( + torch.all(torch.eq(key_padding_mask, c[2])), + f"Unexpected resultant key padding mask: {key_padding_mask}" + f" given current: {c[0]} and previous: {c[1]}", + ) + self.assertEqual(key_padding_mask.size(0), bsz) + self.assertEqual(key_padding_mask.size(1), src_len) + else: + self.assertIsNone(c[2]) + + +if __name__ == "__main__": + unittest.main() diff --git a/tests/test_noising.py b/tests/test_noising.py new file mode 100644 index 0000000000000000000000000000000000000000..b3d0d123c42eaca6f79371aa268049e668fcfcce --- /dev/null +++ b/tests/test_noising.py @@ -0,0 +1,530 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import unittest +from typing import Dict, List + +import tests.utils as test_utils +import torch +from fairseq import utils +from fairseq.data import ( + Dictionary, + LanguagePairDataset, + TransformEosDataset, + data_utils, + noising, +) + + +class TestDataNoising(unittest.TestCase): + def _get_test_data_with_bpe_cont_marker(self, append_eos=True): + """ + Args: + append_eos: if True, each input sentence in the source tokens tensor + will have an EOS appended to the end. + + Returns: + vocabs: BPE vocab with continuation markers as suffixes to denote + non-end of word tokens. This is the standard BPE format used in + fairseq's preprocessing. + x: input tensor containing numberized source tokens, with EOS at the + end if append_eos is true + src_lengths: and source lengths. + """ + vocab = Dictionary() + vocab.add_symbol("he@@") + vocab.add_symbol("llo") + vocab.add_symbol("how") + vocab.add_symbol("are") + vocab.add_symbol("y@@") + vocab.add_symbol("ou") + vocab.add_symbol("n@@") + vocab.add_symbol("ew") + vocab.add_symbol("or@@") + vocab.add_symbol("k") + + src_tokens = [ + ["he@@", "llo", "n@@", "ew", "y@@", "or@@", "k"], + ["how", "are", "y@@", "ou"], + ] + x, src_lengths = x, src_lengths = self._convert_src_tokens_to_tensor( + vocab=vocab, src_tokens=src_tokens, append_eos=append_eos + ) + return vocab, x, src_lengths + + def _get_test_data_with_bpe_end_marker(self, append_eos=True): + """ + Args: + append_eos: if True, each input sentence in the source tokens tensor + will have an EOS appended to the end. + + Returns: + vocabs: BPE vocab with end-of-word markers as suffixes to denote + tokens at the end of a word. This is an alternative to fairseq's + standard preprocessing framework and is not generally supported + within fairseq. + x: input tensor containing numberized source tokens, with EOS at the + end if append_eos is true + src_lengths: and source lengths. + """ + vocab = Dictionary() + vocab.add_symbol("he") + vocab.add_symbol("llo_EOW") + vocab.add_symbol("how_EOW") + vocab.add_symbol("are_EOW") + vocab.add_symbol("y") + vocab.add_symbol("ou_EOW") + vocab.add_symbol("n") + vocab.add_symbol("ew_EOW") + vocab.add_symbol("or") + vocab.add_symbol("k_EOW") + + src_tokens = [ + ["he", "llo_EOW", "n", "ew_EOW", "y", "or", "k_EOW"], + ["how_EOW", "are_EOW", "y", "ou_EOW"], + ] + x, src_lengths = x, src_lengths = self._convert_src_tokens_to_tensor( + vocab=vocab, src_tokens=src_tokens, append_eos=append_eos + ) + return vocab, x, src_lengths + + def _get_test_data_with_word_vocab(self, append_eos=True): + """ + Args: + append_eos: if True, each input sentence in the source tokens tensor + will have an EOS appended to the end. + + Returns: + vocabs: word vocab + x: input tensor containing numberized source tokens, with EOS at the + end if append_eos is true + src_lengths: and source lengths. + """ + vocab = Dictionary() + + vocab.add_symbol("hello") + vocab.add_symbol("how") + vocab.add_symbol("are") + vocab.add_symbol("you") + vocab.add_symbol("new") + vocab.add_symbol("york") + src_tokens = [ + ["hello", "new", "york", "you"], + ["how", "are", "you", "new", "york"], + ] + x, src_lengths = self._convert_src_tokens_to_tensor( + vocab=vocab, src_tokens=src_tokens, append_eos=append_eos + ) + return vocab, x, src_lengths + + def _convert_src_tokens_to_tensor( + self, vocab: Dictionary, src_tokens: List[List[str]], append_eos: bool + ): + src_len = [len(x) for x in src_tokens] + # If we have to append EOS, we include EOS in counting src length + if append_eos: + src_len = [length + 1 for length in src_len] + + x = torch.LongTensor(len(src_tokens), max(src_len)).fill_(vocab.pad()) + for i in range(len(src_tokens)): + for j in range(len(src_tokens[i])): + x[i][j] = vocab.index(src_tokens[i][j]) + if append_eos: + x[i][j + 1] = vocab.eos() + + x = x.transpose(1, 0) + return x, torch.LongTensor(src_len) + + def assert_eos_at_end(self, x, x_len, eos): + """Asserts last token of every sentence in x is EOS """ + for i in range(len(x_len)): + self.assertEqual( + x[x_len[i] - 1][i], + eos, + ( + "Expected eos (token id {eos}) at the end of sentence {i} " + "but got {other} instead" + ).format(i=i, eos=eos, other=x[i][-1]), + ) + + def assert_word_dropout_correct(self, x, x_noised, x_len, l_noised): + # Expect only the first word (2 bpe tokens) of the first example + # was dropped out + self.assertEqual(x_len[0] - 2, l_noised[0]) + for i in range(l_noised[0]): + self.assertEqual(x_noised[i][0], x[i + 2][0]) + + def test_word_dropout_with_eos(self): + vocab, x, x_len = self._get_test_data_with_bpe_cont_marker(append_eos=True) + + with data_utils.numpy_seed(1234): + noising_gen = noising.WordDropout(vocab) + x_noised, l_noised = noising_gen.noising(x, x_len, 0.2) + self.assert_word_dropout_correct( + x=x, x_noised=x_noised, x_len=x_len, l_noised=l_noised + ) + self.assert_eos_at_end(x=x_noised, x_len=l_noised, eos=vocab.eos()) + + def assert_word_blanking_correct(self, x, x_noised, x_len, l_noised, unk): + # Expect only the first word (2 bpe tokens) of the first example + # was blanked out + self.assertEqual(x_len[0], l_noised[0]) + for i in range(l_noised[0]): + if i < 2: + self.assertEqual(x_noised[i][0], unk) + else: + self.assertEqual(x_noised[i][0], x[i][0]) + + def test_word_blank_with_eos(self): + vocab, x, x_len = self._get_test_data_with_bpe_cont_marker(append_eos=True) + + with data_utils.numpy_seed(1234): + noising_gen = noising.WordDropout(vocab) + x_noised, l_noised = noising_gen.noising(x, x_len, 0.2, vocab.unk()) + self.assert_word_blanking_correct( + x=x, x_noised=x_noised, x_len=x_len, l_noised=l_noised, unk=vocab.unk() + ) + self.assert_eos_at_end(x=x_noised, x_len=l_noised, eos=vocab.eos()) + + def generate_unchanged_shuffle_map(self, length): + return {i: i for i in range(length)} + + def assert_word_shuffle_matches_expected( + self, + x, + x_len, + max_shuffle_distance: int, + vocab: Dictionary, + expected_shufle_maps: List[Dict[int, int]], + expect_eos_at_end: bool, + bpe_end_marker=None, + ): + """ + This verifies that with a given x, x_len, max_shuffle_distance, and + vocab, we get the expected shuffle result. + + Args: + x: Tensor of shape (T x B) = (sequence_length, batch_size) + x_len: Tensor of length B = batch_size + max_shuffle_distance: arg to pass to noising + expected_shuffle_maps: List[mapping] where mapping is a + Dict[old_index, new_index], mapping x's elements from their + old positions in x to their new positions in x. + expect_eos_at_end: if True, check the output to make sure there is + an EOS at the end. + bpe_end_marker: str denoting the BPE end token. If this is not None, we + set the BPE cont token to None in the noising classes. + """ + bpe_cont_marker = None + if bpe_end_marker is None: + bpe_cont_marker = "@@" + + with data_utils.numpy_seed(1234): + word_shuffle = noising.WordShuffle( + vocab, bpe_cont_marker=bpe_cont_marker, bpe_end_marker=bpe_end_marker + ) + x_noised, l_noised = word_shuffle.noising( + x, x_len, max_shuffle_distance=max_shuffle_distance + ) + + # For every example, we have a different expected shuffle map. We check + # that each example is shuffled as expected according to each + # corresponding shuffle map. + for i in range(len(expected_shufle_maps)): + shuffle_map = expected_shufle_maps[i] + for k, v in shuffle_map.items(): + self.assertEqual(x[k][i], x_noised[v][i]) + + # Shuffling should not affect the length of each example + for pre_shuffle_length, post_shuffle_length in zip(x_len, l_noised): + self.assertEqual(pre_shuffle_length, post_shuffle_length) + if expect_eos_at_end: + self.assert_eos_at_end(x=x_noised, x_len=l_noised, eos=vocab.eos()) + + def test_word_shuffle_with_eos(self): + vocab, x, x_len = self._get_test_data_with_bpe_cont_marker(append_eos=True) + + # Assert word shuffle with max shuffle distance 0 causes input to be + # unchanged + self.assert_word_shuffle_matches_expected( + x=x, + x_len=x_len, + max_shuffle_distance=0, + vocab=vocab, + expected_shufle_maps=[ + self.generate_unchanged_shuffle_map(example_len) + for example_len in x_len + ], + expect_eos_at_end=True, + ) + + # Assert word shuffle with max shuffle distance 3 matches our expected + # shuffle order + self.assert_word_shuffle_matches_expected( + x=x, + x_len=x_len, + vocab=vocab, + max_shuffle_distance=3, + expected_shufle_maps=[ + self.generate_unchanged_shuffle_map(x_len[0]), + {0: 0, 1: 3, 2: 1, 3: 2}, + ], + expect_eos_at_end=True, + ) + + def test_word_shuffle_with_eos_nonbpe(self): + """The purpose of this is to test shuffling logic with word vocabs""" + vocab, x, x_len = self._get_test_data_with_word_vocab(append_eos=True) + + # Assert word shuffle with max shuffle distance 0 causes input to be + # unchanged + self.assert_word_shuffle_matches_expected( + x=x, + x_len=x_len, + max_shuffle_distance=0, + vocab=vocab, + expected_shufle_maps=[ + self.generate_unchanged_shuffle_map(example_len) + for example_len in x_len + ], + expect_eos_at_end=True, + ) + + # Assert word shuffle with max shuffle distance 3 matches our expected + # shuffle order + self.assert_word_shuffle_matches_expected( + x=x, + x_len=x_len, + vocab=vocab, + max_shuffle_distance=3, + expected_shufle_maps=[ + {0: 0, 1: 1, 2: 3, 3: 2}, + {0: 0, 1: 2, 2: 1, 3: 3, 4: 4}, + ], + expect_eos_at_end=True, + ) + + def test_word_shuffle_without_eos(self): + """Same result as word shuffle with eos except no EOS at end""" + vocab, x, x_len = self._get_test_data_with_bpe_cont_marker(append_eos=False) + + # Assert word shuffle with max shuffle distance 0 causes input to be + # unchanged + self.assert_word_shuffle_matches_expected( + x=x, + x_len=x_len, + max_shuffle_distance=0, + vocab=vocab, + expected_shufle_maps=[ + self.generate_unchanged_shuffle_map(example_len) + for example_len in x_len + ], + expect_eos_at_end=False, + ) + + # Assert word shuffle with max shuffle distance 3 matches our expected + # shuffle order + self.assert_word_shuffle_matches_expected( + x=x, + x_len=x_len, + vocab=vocab, + max_shuffle_distance=3, + expected_shufle_maps=[ + self.generate_unchanged_shuffle_map(x_len[0]), + {0: 0, 1: 3, 2: 1, 3: 2}, + ], + expect_eos_at_end=False, + ) + + def test_word_shuffle_without_eos_with_bpe_end_marker(self): + """Same result as word shuffle without eos except using BPE end token""" + vocab, x, x_len = self._get_test_data_with_bpe_end_marker(append_eos=False) + + # Assert word shuffle with max shuffle distance 0 causes input to be + # unchanged + self.assert_word_shuffle_matches_expected( + x=x, + x_len=x_len, + max_shuffle_distance=0, + vocab=vocab, + expected_shufle_maps=[ + self.generate_unchanged_shuffle_map(example_len) + for example_len in x_len + ], + expect_eos_at_end=False, + bpe_end_marker="_EOW", + ) + + # Assert word shuffle with max shuffle distance 3 matches our expected + # shuffle order + self.assert_word_shuffle_matches_expected( + x=x, + x_len=x_len, + vocab=vocab, + max_shuffle_distance=3, + expected_shufle_maps=[ + self.generate_unchanged_shuffle_map(x_len[0]), + {0: 0, 1: 3, 2: 1, 3: 2}, + ], + expect_eos_at_end=False, + bpe_end_marker="_EOW", + ) + + def assert_no_eos_at_end(self, x, x_len, eos): + """Asserts that the last token of each sentence in x is not EOS """ + for i in range(len(x_len)): + self.assertNotEqual( + x[x_len[i] - 1][i], + eos, + "Expected no eos (token id {eos}) at the end of sentence {i}.".format( + eos=eos, i=i + ), + ) + + def test_word_dropout_without_eos(self): + """Same result as word dropout with eos except no EOS at end""" + vocab, x, x_len = self._get_test_data_with_bpe_cont_marker(append_eos=False) + + with data_utils.numpy_seed(1234): + noising_gen = noising.WordDropout(vocab) + x_noised, l_noised = noising_gen.noising(x, x_len, 0.2) + self.assert_word_dropout_correct( + x=x, x_noised=x_noised, x_len=x_len, l_noised=l_noised + ) + self.assert_no_eos_at_end(x=x_noised, x_len=l_noised, eos=vocab.eos()) + + def test_word_blank_without_eos(self): + """Same result as word blank with eos except no EOS at end""" + vocab, x, x_len = self._get_test_data_with_bpe_cont_marker(append_eos=False) + + with data_utils.numpy_seed(1234): + noising_gen = noising.WordDropout(vocab) + x_noised, l_noised = noising_gen.noising(x, x_len, 0.2, vocab.unk()) + self.assert_word_blanking_correct( + x=x, x_noised=x_noised, x_len=x_len, l_noised=l_noised, unk=vocab.unk() + ) + self.assert_no_eos_at_end(x=x_noised, x_len=l_noised, eos=vocab.eos()) + + def _get_noising_dataset_batch( + self, + src_tokens_no_pad, + src_dict, + append_eos_to_tgt=False, + ): + """ + Constructs a NoisingDataset and the corresponding + ``LanguagePairDataset(NoisingDataset(src), src)``. If + *append_eos_to_tgt* is True, wrap the source dataset in + :class:`TransformEosDataset` to append EOS to the clean source when + using it as the target. + """ + src_dataset = test_utils.TestDataset(data=src_tokens_no_pad) + + noising_dataset = noising.NoisingDataset( + src_dataset=src_dataset, + src_dict=src_dict, + seed=1234, + max_word_shuffle_distance=3, + word_dropout_prob=0.2, + word_blanking_prob=0.2, + noising_class=noising.UnsupervisedMTNoising, + ) + tgt = src_dataset + language_pair_dataset = LanguagePairDataset( + src=noising_dataset, tgt=tgt, src_sizes=None, src_dict=src_dict + ) + language_pair_dataset = TransformEosDataset( + language_pair_dataset, + src_dict.eos(), + append_eos_to_tgt=append_eos_to_tgt, + ) + + dataloader = torch.utils.data.DataLoader( + dataset=language_pair_dataset, + batch_size=2, + collate_fn=language_pair_dataset.collater, + ) + denoising_batch_result = next(iter(dataloader)) + return denoising_batch_result + + def test_noising_dataset_with_eos(self): + src_dict, src_tokens, _ = self._get_test_data_with_bpe_cont_marker( + append_eos=True + ) + + # Format data for src_dataset + src_tokens = torch.t(src_tokens) + src_tokens_no_pad = [] + for src_sentence in src_tokens: + src_tokens_no_pad.append( + utils.strip_pad(tensor=src_sentence, pad=src_dict.pad()) + ) + denoising_batch_result = self._get_noising_dataset_batch( + src_tokens_no_pad=src_tokens_no_pad, src_dict=src_dict + ) + + eos, pad = src_dict.eos(), src_dict.pad() + + # Generated noisy source as source + expected_src = torch.LongTensor( + [[4, 5, 10, 11, 8, 12, 13, eos], [pad, pad, pad, 6, 8, 9, 7, eos]] + ) + # Original clean source as target (right-padded) + expected_tgt = torch.LongTensor( + [[4, 5, 10, 11, 8, 12, 13, eos], [6, 7, 8, 9, eos, pad, pad, pad]] + ) + generated_src = denoising_batch_result["net_input"]["src_tokens"] + tgt_tokens = denoising_batch_result["target"] + + self.assertTensorEqual(expected_src, generated_src) + self.assertTensorEqual(expected_tgt, tgt_tokens) + + def test_noising_dataset_without_eos(self): + """ + Similar to test noising dataset with eos except that we have to set + *append_eos_to_tgt* to ``True``. + """ + + src_dict, src_tokens, _ = self._get_test_data_with_bpe_cont_marker( + append_eos=False + ) + + # Format data for src_dataset + src_tokens = torch.t(src_tokens) + src_tokens_no_pad = [] + for src_sentence in src_tokens: + src_tokens_no_pad.append( + utils.strip_pad(tensor=src_sentence, pad=src_dict.pad()) + ) + denoising_batch_result = self._get_noising_dataset_batch( + src_tokens_no_pad=src_tokens_no_pad, + src_dict=src_dict, + append_eos_to_tgt=True, + ) + + eos, pad = src_dict.eos(), src_dict.pad() + + # Generated noisy source as source + expected_src = torch.LongTensor( + [[4, 5, 10, 11, 8, 12, 13], [pad, pad, pad, 6, 8, 9, 7]] + ) + # Original clean source as target (right-padded) + expected_tgt = torch.LongTensor( + [[4, 5, 10, 11, 8, 12, 13, eos], [6, 7, 8, 9, eos, pad, pad, pad]] + ) + + generated_src = denoising_batch_result["net_input"]["src_tokens"] + tgt_tokens = denoising_batch_result["target"] + + self.assertTensorEqual(expected_src, generated_src) + self.assertTensorEqual(expected_tgt, tgt_tokens) + + def assertTensorEqual(self, t1, t2): + self.assertEqual(t1.size(), t2.size(), "size mismatch") + self.assertEqual(t1.ne(t2).long().sum(), 0) + + +if __name__ == "__main__": + unittest.main() diff --git a/tests/test_online_backtranslation.py b/tests/test_online_backtranslation.py new file mode 100644 index 0000000000000000000000000000000000000000..0ae7e773da0ff838b3c8151bc14b84a6a9238a72 --- /dev/null +++ b/tests/test_online_backtranslation.py @@ -0,0 +1,206 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import tempfile +import unittest +from pathlib import Path +from typing import Any, Dict, Sequence + +import fairseq.data.indexed_dataset as indexed_dataset +import fairseq.options +import fairseq.tasks.online_backtranslation as obt +import torch +from tests import utils + + +def mk_sample(tokens: Sequence[int], batch_size: int = 2) -> Dict[str, Any]: + batch = torch.stack([torch.tensor(tokens, dtype=torch.long)] * batch_size) + sample = { + "net_input": { + "src_tokens": batch, + "prev_output_tokens": batch, + "src_lengths": torch.tensor([len(tokens)] * batch_size, dtype=torch.long), + }, + "target": batch[:, 1:], + } + return sample + + +def mk_dataset(num_samples: int, max_len: int, output: Path): + output.parent.mkdir(exist_ok=True) + idx = indexed_dataset.IndexedDatasetBuilder(str(output)) + data = torch.randint(5, 100, (num_samples, max_len)) + lengths = torch.randint(3, max_len, (num_samples,)) + for d, l in zip(data, lengths): + d[0] = 0 + idx.add_item(d[:l]) + idx.finalize(output.with_suffix(".idx")) + assert output.exists() + assert output.with_suffix(".idx").exists() + + +class OnlineBacktranslationTest(unittest.TestCase): + + tmp_dir = Path(tempfile.mkdtemp(suffix="OnlineBacktranslationTest")) + + @classmethod + def obt_task( + cls, languages: Sequence[str], data: Path = None, language_mapping: str = None + ): + dict_path = cls.tmp_dir / "dict.txt" + if not dict_path.exists(): + dictionary = utils.dummy_dictionary(100) + dictionary.save(str(dict_path)) + + if data is not None: + (data / "dict.txt").write_text(dict_path.read_text()) + else: + data = cls.tmp_dir + assert len(languages) >= 2 + + kwargs = { + "arch": "transformer", + # --max-sentences=1 for better predictability of batches + "max_sentences": 1, + # Use characteristics dimensions + "encoder_layers": 3, + "encoder_embed_dim": 12, + "encoder_ffn_embed_dim": 14, + "encoder_attention_heads": 4, + "decoder_layers": 3, + "decoder_embed_dim": 12, + "decoder_output_dim": 12, + "decoder_ffn_embed_dim": 14, + "decoder_attention_heads": 4, + # Disable dropout so we have comparable tests. + "dropout": 0, + "attention_dropout": 0, + "activation_dropout": 0, + "encoder_layerdrop": 0, + } + + args = fairseq.options.get_args( + data, + task="online_backtranslation", + mono_langs=",".join(languages), + valid_lang_pairs=f"{languages[0]}-{languages[1]}", + tokens_per_sample=256, + language_mapping=language_mapping, + **kwargs, + ) + task = obt.OnlineBackTranslationTask.setup_task(args) + # we need to build the model to have the correct dictionary + model = task.build_model(task.args) + return task, model + + def tmp_path(self, test_case: str) -> Path: + return Path(tempfile.mkdtemp(test_case, dir=self.tmp_dir)) + + def test_lang_tokens(self): + task, model = self.obt_task(["en", "ro", "zh"]) + assert obt._lang_token("en") in task.dictionary + assert obt._lang_token("ro") in task.dictionary + assert obt._lang_token("zh") in task.dictionary + + en_bos = obt._lang_token_index(task.common_dict, "en") + assert "en" == task.common_dict[en_bos].strip("_") + zh_bos = obt._lang_token_index(task.common_dict, "zh") + assert "zh" == task.common_dict[zh_bos].strip("_") + zh_sample = mk_sample([zh_bos, 16, 14, 12, 10]) + + # we expect to receive the bos token for translation + assert task.get_bos_token_from_sample(zh_sample) == en_bos + + def test_backtranslate_sample(self): + task, model = self.obt_task(["en", "ro", "zh"]) + + en_bos = obt._lang_token_index(task.common_dict, "en") + zh_bos = obt._lang_token_index(task.common_dict, "zh") + sample = mk_sample([zh_bos, 16, 14, 12, 10]) + + task.backtranslate_sample(sample, "zh", "en") + target_zh = list(sample["target"][0]) + assert target_zh == [16, 14, 12, 10] # original zh sentence + generated_en = sample["net_input"]["src_tokens"][0] + assert generated_en[0] == en_bos + + def test_train_dataset(self): + data = self.tmp_path("test_train_dataset") + mk_dataset(20, 10, data / "en" / "train.bin") + mk_dataset(10, 10, data / "zh" / "train.bin") + task, model = self.obt_task(["en", "zh"], data) + task.load_dataset("train") + + en_bos = obt._lang_token_index(task.common_dict, "en") + zh_bos = obt._lang_token_index(task.common_dict, "zh") + + train = task.datasets["train"] + train.ordered_indices() + train.prefetch([0, 19]) + sample_0 = train[0] + sample_19 = train[19] + self.assertEqual( + set(sample_0.keys()), {"en-BT", "en-DENOISE", "zh-BT", "zh-DENOISE"} + ) + for sample in (sample_0, sample_19): + self.assertEqual(sample["en-BT"]["source"][0], en_bos) + # bt target isn't ready to look at. + self.assertEqual(sample["en-DENOISE"]["source"][0], en_bos) + # TODO What could we check on the target side ? + + for i in range(10): + # Zh dataset is shorter, and is wrapped around En dataset. + train.prefetch([i, i + 10]) + self.assertEqual( + list(train[i]["zh-DENOISE"]["source"]), + list(train[i + 10]["zh-DENOISE"]["source"]), + ) + self.assertEqual(train[i]["zh-DENOISE"]["source"][0].item(), zh_bos) + + # Sorted by increasing len + self.assertLess( + len(sample_0["en-BT"]["source"]), len(sample_19["en-BT"]["source"]) + ) + + def test_valid_dataset(self): + data = self.tmp_path("test_valid_dataset") + mk_dataset(10, 21, data / "valid.en-zh.en.bin") + mk_dataset(10, 21, data / "valid.en-zh.zh.bin") + + task, model = self.obt_task(["en", "zh"], data) + valid = task.load_dataset("valid") + en_bos = obt._lang_token_index(task.common_dict, "en") + + assert valid is not None + valid.prefetch(range(10)) + sample_0 = valid[0] + sample_9 = valid[9] + self.assertEqual(sample_0["id"], 0) + self.assertEqual(sample_9["id"], 9) + self.assertEqual(sample_0["source"][0], en_bos) + self.assertEqual(sample_9["source"][0], en_bos) + # TODO: could we test the target side ? + + def assertFnMatch(self, fn, values): + for x, y in values.items(): + fn_x = fn(x) + self.assertEqual(fn_x, y, f"Fn has wrong value: fn({x}) = {fn_x} != {y}") + + def test_piecewise_linear_fn(self): + self.assertFnMatch( + obt.PiecewiseLinearFn.from_string("1.0"), {0: 1, 100: 1, 500: 1, 1000: 1} + ) + self.assertFnMatch( + obt.PiecewiseLinearFn.from_string("0:1,1000:0"), + {0: 1, 500: 0.5, 1000: 0, 2000: 0}, + ) + self.assertFnMatch( + obt.PiecewiseLinearFn.from_string("0:0,1000:1"), + {0: 0, 500: 0.5, 1000: 1, 2000: 1}, + ) + self.assertFnMatch( + obt.PiecewiseLinearFn.from_string("0:0,1000:1,2000:0"), + {0: 0, 500: 0.5, 1000: 1, 1500: 0.5, 2000: 0, 3000: 0}, + ) diff --git a/tests/test_plasma_utils.py b/tests/test_plasma_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..e6344c2a5a73fcb2fb81376e7bd43470963b3674 --- /dev/null +++ b/tests/test_plasma_utils.py @@ -0,0 +1,126 @@ +import contextlib +import unittest +import tempfile +from io import StringIO + +import numpy as np + +from tests.utils import create_dummy_data, preprocess_lm_data, train_language_model + +try: + from pyarrow import plasma + from fairseq.data.plasma_utils import PlasmaView, PlasmaStore + + PYARROW_AVAILABLE = True +except ImportError: + PYARROW_AVAILABLE = False + +dummy_path = "dummy" + + +@unittest.skipUnless(PYARROW_AVAILABLE, "") +class TestPlasmaView(unittest.TestCase): + def setUp(self) -> None: + self.tmp_file = tempfile.NamedTemporaryFile() # noqa: P201 + self.path = self.tmp_file.name + self.server = PlasmaStore.start(path=self.path, nbytes=10000) + self.client = plasma.connect(self.path, num_retries=10) + + def tearDown(self) -> None: + self.client.disconnect() + self.tmp_file.close() + self.server.kill() + + def test_two_servers_do_not_share_object_id_space(self): + data_server_1 = np.array([0, 1]) + data_server_2 = np.array([2, 3]) + server_2_path = self.path + with tempfile.NamedTemporaryFile() as server_1_path: + server = PlasmaStore.start(path=server_1_path.name, nbytes=10000) + arr1 = PlasmaView( + data_server_1, dummy_path, 1, plasma_path=server_1_path.name + ) + assert len(arr1.client.list()) == 1 + assert (arr1.array == data_server_1).all() + arr2 = PlasmaView(data_server_2, dummy_path, 1, plasma_path=server_2_path) + assert (arr2.array == data_server_2).all() + assert (arr1.array == data_server_1).all() + server.kill() + + def test_hash_collision(self): + data_server_1 = np.array([0, 1]) + data_server_2 = np.array([2, 3]) + arr1 = PlasmaView(data_server_1, dummy_path, 1, plasma_path=self.path) + assert len(arr1.client.list()) == 1 + arr2 = PlasmaView(data_server_2, dummy_path, 1, plasma_path=self.path) + assert len(arr1.client.list()) == 1 + assert len(arr2.client.list()) == 1 + assert (arr2.array == data_server_1).all() + # New hash key based on tuples + arr3 = PlasmaView( + data_server_2, dummy_path, (1, 12312312312, None), plasma_path=self.path + ) + assert ( + len(arr2.client.list()) == 2 + ), "No new object was created by using a novel hash key" + assert ( + arr3.object_id in arr2.client.list() + ), "No new object was created by using a novel hash key" + assert ( + arr3.object_id in arr3.client.list() + ), "No new object was created by using a novel hash key" + del arr3, arr2, arr1 + + @staticmethod + def _assert_view_equal(pv1, pv2): + np.testing.assert_array_equal(pv1.array, pv2.array) + + def test_putting_same_array_twice(self): + data = np.array([4, 4, 4]) + arr1 = PlasmaView(data, dummy_path, 1, plasma_path=self.path) + assert len(self.client.list()) == 1 + arr1b = PlasmaView( + data, dummy_path, 1, plasma_path=self.path + ) # should not change contents of store + arr1c = PlasmaView( + None, dummy_path, 1, plasma_path=self.path + ) # should not change contents of store + + assert len(self.client.list()) == 1 + self._assert_view_equal(arr1, arr1b) + self._assert_view_equal(arr1, arr1c) + PlasmaView( + data, dummy_path, 2, plasma_path=self.path + ) # new object id, adds new entry + assert len(self.client.list()) == 2 + + new_client = plasma.connect(self.path) + assert len(new_client.list()) == 2 # new client can access same objects + assert isinstance(arr1.object_id, plasma.ObjectID) + del arr1b + del arr1c + + def test_plasma_store_full_raises(self): + with tempfile.NamedTemporaryFile() as new_path: + server = PlasmaStore.start(path=new_path.name, nbytes=10000) + with self.assertRaises(plasma.PlasmaStoreFull): + # 2000 floats is more than 2000 bytes + PlasmaView( + np.random.rand(10000, 1), dummy_path, 1, plasma_path=new_path.name + ) + server.kill() + + def test_object_id_overflow(self): + PlasmaView.get_object_id("", 2 ** 21) + + def test_training_lm_plasma(self): + with contextlib.redirect_stdout(StringIO()): + with tempfile.TemporaryDirectory("test_transformer_lm") as data_dir: + create_dummy_data(data_dir) + preprocess_lm_data(data_dir) + train_language_model( + data_dir, + "transformer_lm", + ["--use-plasma-view", "--plasma-path", self.path], + run_validation=True, + ) diff --git a/tests/test_reproducibility.py b/tests/test_reproducibility.py new file mode 100644 index 0000000000000000000000000000000000000000..94931b2a0721c4adfee8899c89cac24f45973d17 --- /dev/null +++ b/tests/test_reproducibility.py @@ -0,0 +1,150 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import contextlib +import json +import os +import tempfile +import unittest +from io import StringIO + +import torch + +from . import test_binaries + + +class TestReproducibility(unittest.TestCase): + def _test_reproducibility( + self, + name, + extra_flags=None, + delta=0.0001, + resume_checkpoint="checkpoint1.pt", + max_epoch=3, + ): + def get_last_log_stats_containing_string(log_records, search_string): + for log_record in logs.records[::-1]: + if isinstance(log_record.msg, str) and search_string in log_record.msg: + return json.loads(log_record.msg) + + if extra_flags is None: + extra_flags = [] + + with tempfile.TemporaryDirectory(name) as data_dir: + with self.assertLogs() as logs: + test_binaries.create_dummy_data(data_dir) + test_binaries.preprocess_translation_data(data_dir) + + # train epochs 1 and 2 together + with self.assertLogs() as logs: + test_binaries.train_translation_model( + data_dir, + "fconv_iwslt_de_en", + [ + "--dropout", + "0.0", + "--log-format", + "json", + "--log-interval", + "1", + "--max-epoch", + str(max_epoch), + ] + + extra_flags, + ) + train_log = get_last_log_stats_containing_string(logs.records, "train_loss") + valid_log = get_last_log_stats_containing_string(logs.records, "valid_loss") + + # train epoch 2, resuming from previous checkpoint 1 + os.rename( + os.path.join(data_dir, resume_checkpoint), + os.path.join(data_dir, "checkpoint_last.pt"), + ) + with self.assertLogs() as logs: + test_binaries.train_translation_model( + data_dir, + "fconv_iwslt_de_en", + [ + "--dropout", + "0.0", + "--log-format", + "json", + "--log-interval", + "1", + "--max-epoch", + str(max_epoch), + ] + + extra_flags, + ) + train_res_log = get_last_log_stats_containing_string( + logs.records, "train_loss" + ) + valid_res_log = get_last_log_stats_containing_string( + logs.records, "valid_loss" + ) + + for k in ["train_loss", "train_ppl", "train_num_updates", "train_gnorm"]: + self.assertAlmostEqual( + float(train_log[k]), float(train_res_log[k]), delta=delta + ) + for k in [ + "valid_loss", + "valid_ppl", + "valid_num_updates", + "valid_best_loss", + ]: + self.assertAlmostEqual( + float(valid_log[k]), float(valid_res_log[k]), delta=delta + ) + + def test_reproducibility(self): + self._test_reproducibility("test_reproducibility") + + @unittest.skipIf(not torch.cuda.is_available(), "test requires a GPU") + def test_reproducibility_fp16(self): + self._test_reproducibility( + "test_reproducibility_fp16", + [ + "--fp16", + "--fp16-init-scale", + "4096", + ], + delta=0.011, + ) + + @unittest.skipIf(not torch.cuda.is_available(), "test requires a GPU") + def test_reproducibility_memory_efficient_fp16(self): + self._test_reproducibility( + "test_reproducibility_memory_efficient_fp16", + [ + "--memory-efficient-fp16", + "--fp16-init-scale", + "4096", + ], + ) + + @unittest.skipIf(not torch.cuda.is_available(), "test requires a GPU") + def test_reproducibility_amp(self): + self._test_reproducibility( + "test_reproducibility_amp", + [ + "--amp", + "--fp16-init-scale", + "4096", + ], + delta=0.011, + ) + + def test_mid_epoch_reproducibility(self): + self._test_reproducibility( + "test_mid_epoch_reproducibility", + ["--save-interval-updates", "3"], + resume_checkpoint="checkpoint_1_3.pt", + max_epoch=1, + ) + + +if __name__ == "__main__": + unittest.main() diff --git a/tests/test_resampling_dataset.py b/tests/test_resampling_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..ccb53a253ce6ca0d8e972adfa708144b4299b3cb --- /dev/null +++ b/tests/test_resampling_dataset.py @@ -0,0 +1,103 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import collections +import unittest + +import numpy as np +from fairseq.data import ListDataset, ResamplingDataset + + +class TestResamplingDataset(unittest.TestCase): + def setUp(self): + self.strings = ["ab", "c", "def", "ghij"] + self.weights = [4.0, 2.0, 7.0, 1.5] + self.size_ratio = 2 + self.dataset = ListDataset( + self.strings, np.array([len(s) for s in self.strings]) + ) + + def _test_common(self, resampling_dataset, iters): + assert len(self.dataset) == len(self.strings) == len(self.weights) + assert len(resampling_dataset) == self.size_ratio * len(self.strings) + + results = {"ordered_by_size": True, "max_distribution_diff": 0.0} + + totalfreqs = 0 + freqs = collections.defaultdict(int) + + for epoch_num in range(iters): + resampling_dataset.set_epoch(epoch_num) + + indices = resampling_dataset.ordered_indices() + assert len(indices) == len(resampling_dataset) + + prev_size = -1 + + for i in indices: + cur_size = resampling_dataset.size(i) + # Make sure indices map to same sequences within an epoch + assert resampling_dataset[i] == resampling_dataset[i] + + # Make sure length of sequence is correct + assert cur_size == len(resampling_dataset[i]) + + freqs[resampling_dataset[i]] += 1 + totalfreqs += 1 + + if prev_size > cur_size: + results["ordered_by_size"] = False + + prev_size = cur_size + + assert set(freqs.keys()) == set(self.strings) + for s, weight in zip(self.strings, self.weights): + freq = freqs[s] / totalfreqs + expected_freq = weight / sum(self.weights) + results["max_distribution_diff"] = max( + results["max_distribution_diff"], abs(expected_freq - freq) + ) + + return results + + def test_resampling_dataset_batch_by_size_false(self): + resampling_dataset = ResamplingDataset( + self.dataset, + self.weights, + size_ratio=self.size_ratio, + batch_by_size=False, + seed=0, + ) + + results = self._test_common(resampling_dataset, iters=1000) + + # For batch_by_size = False, the batches should be returned in + # arbitrary order of size. + assert not results["ordered_by_size"] + + # Allow tolerance in distribution error of 2%. + assert results["max_distribution_diff"] < 0.02 + + def test_resampling_dataset_batch_by_size_true(self): + resampling_dataset = ResamplingDataset( + self.dataset, + self.weights, + size_ratio=self.size_ratio, + batch_by_size=True, + seed=0, + ) + + results = self._test_common(resampling_dataset, iters=1000) + + # For batch_by_size = True, the batches should be returned in + # increasing order of size. + assert results["ordered_by_size"] + + # Allow tolerance in distribution error of 2%. + assert results["max_distribution_diff"] < 0.02 + + +if __name__ == "__main__": + unittest.main() diff --git a/tests/test_roberta.py b/tests/test_roberta.py new file mode 100644 index 0000000000000000000000000000000000000000..b0b9cfd31e8cb1e03ae74403886d2fb5266e0443 --- /dev/null +++ b/tests/test_roberta.py @@ -0,0 +1,314 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import functools +import unittest +from typing import Any, Dict, Sequence + +import fairseq +import fairseq.options +import fairseq.tasks +import torch +from tests.utils import dummy_dictionary + +VOCAB_SIZE = 100 + + +@fairseq.tasks.register_task("fake_task") +class FakeTask(fairseq.tasks.LegacyFairseqTask): + def __init__(self, args): + super().__init__(args) + self.dictionary = dummy_dictionary(VOCAB_SIZE - 4) + assert len(self.dictionary) == VOCAB_SIZE + + @property + def source_dictionary(self): + return self.dictionary + + @property + def target_dictionary(self): + return self.dictionary + + +@functools.lru_cache() +def get_toy_model( + device: str, + architecture: str = "roberta_enc_dec", + **extra_args: Any, +): + assert device in ("gpu", "cpu") + kwargs = { + "arch": architecture, + # Use characteristics dimensions + "encoder_layers": 3, + "encoder_embed_dim": 12, + "encoder_ffn_embed_dim": 14, + "encoder_attention_heads": 4, + "decoder_layers": 3, + "decoder_embed_dim": 12, + "decoder_ffn_embed_dim": 14, + "decoder_attention_heads": 4, + # Disable dropout so we have comparable tests. + "dropout": 0, + "attention_dropout": 0, + "activation_dropout": 0, + "encoder_layerdrop": 0, + # required args + "tokens_per_sample": 256, + "data": "/tmp/test_roberta", + } + kwargs.update(extra_args) + fake_task = FakeTask(kwargs) + args = fairseq.options.get_args( + task="online_backtranslation", + mono_langs="en,ro", + valid_lang_pairs="en-ro", + **kwargs, + ) + torch.manual_seed(0) + model = fake_task.build_model(args) + if device == "gpu": + model.cuda() + return fake_task, model + + +def mk_sample( + lang: str, device: str, tok: Sequence[int] = None, batch_size: int = 2 +) -> Dict[str, Any]: + assert device in ("gpu", "cpu") + if not tok: + if lang == "en": + tok = [10, 11, 12, 13, 14, 15, 2] + else: + tok = [20, 21, 22, 23, 24, 25, 26, 27, 2] + + batch = torch.stack([torch.tensor(tok, dtype=torch.long)] * batch_size) + if device == "gpu": + batch = batch.cuda() + sample = { + "net_input": { + "src_tokens": batch, + "prev_output_tokens": batch, + "src_lengths": torch.tensor( + [len(tok)] * batch_size, dtype=torch.long, device=batch.device + ), + }, + "target": batch[:, 1:], + } + return sample + + +def cpu_gpu(fn): + def helper(self): + fn(self, "cpu") + if torch.cuda.is_available(): + fn(self, "gpu") + + return helper + + +def architectures(fn): + def helper(self): + for arch in ["roberta_enc_dec", "transformer"]: + fn(self, arch) + + return helper + + +class RobertaTest(unittest.TestCase): + def assertTensorEqual(self, t1, t2, delta: float = 1e-6): + self.assertEqual(t1.size(), t2.size(), "size mismatch") + if delta == 0.0: + self.assertEqual(t1.ne(t2).long().sum(), 0) + else: + self.assertEqual(((t2 - t1).abs() > delta).long().sum(), 0) + + def assertSharing(self, model, link_groups: Sequence[Sequence[str]]): + ids = {} + for group in link_groups: + group_ids = {name: id(params(model, name)) for name in group} + shared_id = group_ids[group[0]] + self.assertEqual(group_ids, {name: shared_id for name in group}) + self.assertNotIn(shared_id, ids) + ids[shared_id] = group + + def test_roberta_shared_params(self): + _, roberta = get_toy_model("cpu", architecture="roberta") + self.assertSharing( + roberta, + [ + [ + "encoder.sentence_encoder.embed_tokens.weight", + "encoder.lm_head.weight", + ] + ], + ) + + _, roberta = get_toy_model( + "cpu", architecture="roberta", untie_weights_roberta=True + ) + self.assertSharing( + roberta, + [ + ["encoder.sentence_encoder.embed_tokens.weight"], + ["encoder.lm_head.weight"], + ], + ) + + def test_roberta_enc_dec_shared_params(self): + # 3 distinct embeddings + _, enc_dec = get_toy_model("cpu", architecture="roberta_enc_dec") + self.assertSharing( + enc_dec, + [ + ["encoder.embed_tokens.weight"], + ["decoder.embed_tokens.weight"], + ["decoder.output_projection.weight"], + ], + ) + + # 2 distinct embeddings, one for encoder, one for decoder + _, enc_dec = get_toy_model( + "cpu", architecture="roberta_enc_dec", share_decoder_input_output_embed=True + ) + self.assertSharing( + enc_dec, + [ + ["encoder.embed_tokens.weight"], + [ + "decoder.embed_tokens.weight", + "decoder.output_projection.weight", + ], + ], + ) + + # shared embeddings + _, enc_dec = get_toy_model( + "cpu", architecture="roberta_enc_dec", share_all_embeddings=True + ) + self.assertSharing( + enc_dec, + [ + [ + "encoder.embed_tokens.weight", + "decoder.embed_tokens.weight", + "decoder.output_projection.weight", + ] + ], + ) + + def test_roberta_max_positions_is_correctly_set(self): + device = "cpu" + task, model = get_toy_model(device) + max_pos = model.max_decoder_positions() + self.assertEqual(max_pos, 256) + self.assertEqual(max_pos, model.decoder.max_positions()) + self.assertEqual(max_pos, model.encoder.max_positions()) + self.assertEqual(max_pos, model.encoder.embed_positions.max_positions) + + sentence = [31 for _ in range(max_pos)] + sample = mk_sample("en", device, sentence, batch_size=1) + self.assertEqual(list(sample["net_input"]["src_lengths"]), [max_pos]) + self.assertEqual(len(sample["net_input"]["src_tokens"][0]), max_pos) + x, _ = model.forward(**sample["net_input"]) + self.assertEqual(x.shape, (1, max_pos, VOCAB_SIZE)) + + @cpu_gpu + def test_roberta_forward_backward(self, device: str): + _, model = get_toy_model(device) + sample = mk_sample("en", device) + en_tokens = sample["net_input"]["src_tokens"] + (bs, l) = en_tokens.shape + # Forward + logits, _ = model(**sample["net_input"]) + self.assertEqual(logits.shape, (bs, l, VOCAB_SIZE)) + + # Backward + loss = logits.sum() + loss.backward() + + @cpu_gpu + def test_roberta_forward_backward_bs1(self, device: str): + _, model = get_toy_model(device) + sample = mk_sample("en", device, batch_size=1) + o, _ = model.forward(**sample["net_input"]) + loss = o.sum() + sample2 = mk_sample("ro", device, batch_size=1) + o, _ = model.forward(**sample2["net_input"]) + loss += o.sum() + loss.backward() + + @cpu_gpu + def test_roberta_batching(self, device: str): + """ + Checks that the batch of size 2 give twice the same results than the batch of size 1. + """ + _, model = get_toy_model(device) + sample = mk_sample("en", device, batch_size=1) + slen = sample["net_input"]["src_lengths"][0] + sample2 = mk_sample("en", device, batch_size=2) + with torch.no_grad(): + z = model.encoder.forward( + sample["net_input"]["src_tokens"], sample["net_input"]["src_lengths"] + ) + z = z["encoder_out"][-1] + logits, _ = model.forward(**sample["net_input"]) + + z2 = model.encoder.forward( + sample2["net_input"]["src_tokens"], sample["net_input"]["src_lengths"] + ) + z2 = z2["encoder_out"][-1] + logits2, _ = model.forward(**sample2["net_input"]) + + self.assertEqual(z.shape, (slen, 1, 12)) + self.assertEqual(z2.shape, (slen, 2, 12)) + self.assertTensorEqual(logits2[0], logits2[1]) + self.assertTensorEqual(logits[0], logits2[0]) + + @cpu_gpu + def test_roberta_incremental_decoder(self, device: str): + """ + Checks that incremental decoding yields the same result than non incremental one. + """ + task, model = get_toy_model(device) + + en_sample = mk_sample("en", device) + en_tokens = en_sample["net_input"]["src_tokens"] + ro_sample = mk_sample("ro", device) + ro_tokens = ro_sample["net_input"]["src_tokens"] + + en_enc = model.encoder.forward( + en_tokens, src_lengths=en_sample["net_input"]["src_lengths"] + ) + (bs, tgt_len) = ro_tokens.shape + + # Decode without incremental state + ro_dec, _ = model.decoder.forward(ro_tokens, encoder_out=en_enc) + self.assertEqual(ro_dec.shape, (bs, tgt_len, VOCAB_SIZE)) + self.assertTensorEqual(ro_dec[0], ro_dec[1]) + + # Decode with incremental state + inc_state = {} + ro_dec_inc = [] + for l in range(tgt_len): + ro, _ = model.decoder.forward( + ro_tokens[:, : l + 1], encoder_out=en_enc, incremental_state=inc_state + ) + self.assertEqual(ro.shape, (bs, 1, VOCAB_SIZE)) + ro_dec_inc.append(ro) + + for l in range(tgt_len): + # Intra-batch + self.assertTensorEqual(ro_dec_inc[l][0], ro_dec_inc[l][1]) + # Incremental vs non-incremental + self.assertTensorEqual(ro_dec_inc[l][:, 0], ro_dec[:, l]) + + +def params(model, name): + if "." not in name: + return getattr(model, name) + + prefix, name = name.split(".", 1) + return params(getattr(model, prefix), name) diff --git a/tests/test_sequence_generator.py b/tests/test_sequence_generator.py new file mode 100644 index 0000000000000000000000000000000000000000..afbdfb6c2cde139dfc7e8c48fdbf889375c8d4e1 --- /dev/null +++ b/tests/test_sequence_generator.py @@ -0,0 +1,745 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import tempfile +import unittest +import math +import numpy as np + + +import tests.utils as test_utils +import torch +from fairseq import search +from fairseq.data.dictionary import Dictionary +from fairseq.models.transformer import TransformerModel +from fairseq.sequence_generator import EnsembleModel, SequenceGenerator +from fairseq.ngram_repeat_block import NGramRepeatBlock +from fairseq.tasks.fairseq_task import LegacyFairseqTask + + +DEFAULT_TEST_VOCAB_SIZE = 100 + + +class DummyTask(LegacyFairseqTask): + def __init__(self, args): + super().__init__(args) + self.dictionary = get_dummy_dictionary() + if getattr(self.args, "ctc", False): + self.dictionary.add_symbol("<ctc_blank>") + self.src_dict = self.dictionary + self.tgt_dict = self.dictionary + + @property + def source_dictionary(self): + return self.src_dict + + @property + def target_dictionary(self): + return self.dictionary + + +def get_dummy_dictionary(vocab_size=DEFAULT_TEST_VOCAB_SIZE): + dummy_dict = Dictionary() + # add dummy symbol to satisfy vocab size + for id, _ in enumerate(range(vocab_size)): + dummy_dict.add_symbol("{}".format(id), n=1000) + return dummy_dict + + +def get_dummy_task_and_parser(): + """ + to build a fariseq model, we need some dummy parse and task. This function + is used to create dummy task and parser to faciliate model/criterion test + + Note: we use FbSpeechRecognitionTask as the dummy task. You may want + to use other task by providing another function + """ + parser = argparse.ArgumentParser( + description="test_dummy_s2s_task", argument_default=argparse.SUPPRESS + ) + DummyTask.add_args(parser) + args = parser.parse_args([]) + task = DummyTask.setup_task(args) + return task, parser + + +class TestJitSequenceGeneratorBase(unittest.TestCase): + def setUp(self): + self.task, self.parser = get_dummy_task_and_parser() + eos = self.task.tgt_dict.eos() + src_tokens = torch.randint(3, 50, (2, 10)).long() + src_tokens = torch.cat((src_tokens, torch.LongTensor([[eos], [eos]])), -1) + src_lengths = torch.LongTensor([2, 10]) + self.sample = { + "net_input": {"src_tokens": src_tokens, "src_lengths": src_lengths} + } + TransformerModel.add_args(self.parser) + args = self.parser.parse_args([]) + args.encoder_layers = 2 + args.decoder_layers = 1 + self.transformer_model = TransformerModel.build_model(args, self.task) + + def assertOutputEqual(self, hypo, pos_probs): + pos_scores = torch.FloatTensor(pos_probs).log() + self.assertTensorSizeEqual(hypo["positional_scores"], pos_scores) + self.assertTensorSizeEqual(pos_scores.numel(), hypo["tokens"].numel()) + + def assertTensorSizeEqual(self, t1, t2): + self.assertEqual(t1.size(), t2.size(), "size mismatch") + + def assertAlmostEqual(self, t1, t2): + self.assertEqual(t1.size(), t2.size(), "size mismatch") + self.assertLess((t1 - t2).abs().max(), 1e-4) + + def assertTensorEqual(self, t1, t2): + self.assertEqual(t1.size(), t2.size(), "size mismatch") + self.assertEqual(t1.ne(t2).long().sum(), 0) + + def assertHypoEqual(self, h1, h2): + "Check two hypos are equal" + self.assertTensorEqual(h1["tokens"], h2["tokens"]) + self.assertAlmostEqual(h1["positional_scores"], h2["positional_scores"]) + self.assertLess(abs(h1["score"] - h2["score"]), 1e-6) + self.assertAlmostEqual(h1["attention"], h2["attention"]) + + def _test_save_and_load(self, scripted_module): + with tempfile.NamedTemporaryFile() as f: + scripted_module.save(f.name) + torch.jit.load(f.name) + + +JIT_MSG = "Targeting OSS scriptability for the 1.6 release" + + +@unittest.skipIf(torch.__version__ < "1.6.0", JIT_MSG) +class TestJitSequenceGenerator(TestJitSequenceGeneratorBase): + def test_export_transformer(self): + model = self.transformer_model + torch.jit.script(model) + + def test_ensemble_sequence_generator(self): + model = self.transformer_model + generator = SequenceGenerator( + [model], + self.task.tgt_dict, + beam_size=2, + no_repeat_ngram_size=2, + max_len_b=10, + ) + scripted_model = torch.jit.script(generator) + self._test_save_and_load(scripted_model) + + def test_export_ensemble_model(self): + model = self.transformer_model + ensemble_models = EnsembleModel([model]) + torch.jit.script(ensemble_models) + + +class TestExportSearch(unittest.TestCase): + def setUp(self): + task, _ = get_dummy_task_and_parser() + self.tgt_dict = task.tgt_dict + self.min_top1_prob = 0.4 + + def test_export_diverse_bs(self): + search_strategy = search.DiverseBeamSearch( + self.tgt_dict, num_groups=2, diversity_strength=0.0 + ) + torch.jit.script(search_strategy) + + def test_export_sampling(self): + low_sampling_topp = self.min_top1_prob / 2.0 + search_strategy = search.Sampling( + self.tgt_dict, sampling_topp=low_sampling_topp + ) + torch.jit.script(search_strategy) + + def test_export_diverse_siblings_search(self): + search_strategy = search.DiverseSiblingsSearch( + self.tgt_dict, diversity_rate=0.5 + ) + torch.jit.script(search_strategy) + + +class TestSequenceGeneratorBase(unittest.TestCase): + def assertHypoTokens(self, hypo, tokens): + self.assertTensorEqual(hypo["tokens"], torch.LongTensor(tokens)) + + def assertHypoScore(self, hypo, pos_probs, normalized=True, lenpen=1.0): + pos_scores = torch.FloatTensor(pos_probs).log() + self.assertAlmostEqual(hypo["positional_scores"], pos_scores) + self.assertEqual(pos_scores.numel(), hypo["tokens"].numel()) + score = pos_scores.sum() + if normalized: + score /= pos_scores.numel() ** lenpen + self.assertLess(abs(score - hypo["score"]), 1e-6) + + def assertAlmostEqual(self, t1, t2): + self.assertEqual(t1.size(), t2.size(), "size mismatch") + self.assertLess((t1 - t2).abs().max(), 1e-4) + + def assertTensorEqual(self, t1, t2): + self.assertEqual(t1.size(), t2.size(), "size mismatch") + self.assertEqual(t1.ne(t2).long().sum(), 0) + + +class TestSequenceGenerator(TestSequenceGeneratorBase): + def setUp(self): + ( + self.tgt_dict, + self.w1, + self.w2, + src_tokens, + src_lengths, + self.model, + ) = test_utils.sequence_generator_setup() + self.sample = { + "net_input": {"src_tokens": src_tokens, "src_lengths": src_lengths} + } + + def test_with_normalization(self): + generator = SequenceGenerator([self.model], self.tgt_dict, beam_size=2) + hypos = generator.forward(self.sample) + eos, w1, w2 = self.tgt_dict.eos(), self.w1, self.w2 + # sentence 1, beam 1 + self.assertHypoTokens(hypos[0][0], [w1, eos]) + self.assertHypoScore(hypos[0][0], [0.9, 1.0]) + # sentence 1, beam 2 + self.assertHypoTokens(hypos[0][1], [w2, w1, w2, eos]) + self.assertHypoScore(hypos[0][1], [0.1, 0.9, 0.9, 1.0]) + # sentence 2, beam 1 + self.assertHypoTokens(hypos[1][0], [w1, w2, w1, eos]) + self.assertHypoScore(hypos[1][0], [0.7, 0.4, 0.4, 1.0]) + # sentence 2, beam 2 + self.assertHypoTokens(hypos[1][1], [w1, w2, eos]) + self.assertHypoScore(hypos[1][1], [0.7, 0.4, 0.6]) + + def test_without_normalization(self): + # Sentence 1: unchanged from the normalized case + # Sentence 2: beams swap order + generator = SequenceGenerator( + [self.model], self.tgt_dict, beam_size=2, normalize_scores=False + ) + hypos = generator.forward(self.sample) + eos, w1, w2 = self.tgt_dict.eos(), self.w1, self.w2 + # sentence 1, beam 1 + self.assertHypoTokens(hypos[0][0], [w1, eos]) + self.assertHypoScore(hypos[0][0], [0.9, 1.0], normalized=False) + # sentence 1, beam 2 + self.assertHypoTokens(hypos[0][1], [w2, w1, w2, eos]) + self.assertHypoScore(hypos[0][1], [0.1, 0.9, 0.9, 1.0], normalized=False) + # sentence 2, beam 1 + self.assertHypoTokens(hypos[1][0], [w1, w2, eos]) + self.assertHypoScore(hypos[1][0], [0.7, 0.4, 0.6], normalized=False) + # sentence 2, beam 2 + self.assertHypoTokens(hypos[1][1], [w1, w2, w1, eos]) + self.assertHypoScore(hypos[1][1], [0.7, 0.4, 0.4, 1.0], normalized=False) + + def test_with_lenpen_favoring_short_hypos(self): + lenpen = 0.6 + generator = SequenceGenerator( + [self.model], self.tgt_dict, beam_size=2, len_penalty=lenpen + ) + hypos = generator.forward(self.sample) + eos, w1, w2 = self.tgt_dict.eos(), self.w1, self.w2 + # sentence 1, beam 1 + self.assertHypoTokens(hypos[0][0], [w1, eos]) + self.assertHypoScore(hypos[0][0], [0.9, 1.0], lenpen=lenpen) + # sentence 1, beam 2 + self.assertHypoTokens(hypos[0][1], [w2, w1, w2, eos]) + self.assertHypoScore(hypos[0][1], [0.1, 0.9, 0.9, 1.0], lenpen=lenpen) + # sentence 2, beam 1 + self.assertHypoTokens(hypos[1][0], [w1, w2, eos]) + self.assertHypoScore(hypos[1][0], [0.7, 0.4, 0.6], lenpen=lenpen) + # sentence 2, beam 2 + self.assertHypoTokens(hypos[1][1], [w1, w2, w1, eos]) + self.assertHypoScore(hypos[1][1], [0.7, 0.4, 0.4, 1.0], lenpen=lenpen) + + def test_with_lenpen_favoring_long_hypos(self): + lenpen = 5.0 + generator = SequenceGenerator( + [self.model], self.tgt_dict, beam_size=2, len_penalty=lenpen + ) + hypos = generator.forward(self.sample) + eos, w1, w2 = self.tgt_dict.eos(), self.w1, self.w2 + # sentence 1, beam 1 + self.assertHypoTokens(hypos[0][0], [w2, w1, w2, eos]) + self.assertHypoScore(hypos[0][0], [0.1, 0.9, 0.9, 1.0], lenpen=lenpen) + # sentence 1, beam 2 + self.assertHypoTokens(hypos[0][1], [w1, eos]) + self.assertHypoScore(hypos[0][1], [0.9, 1.0], lenpen=lenpen) + # sentence 2, beam 1 + self.assertHypoTokens(hypos[1][0], [w1, w2, w1, eos]) + self.assertHypoScore(hypos[1][0], [0.7, 0.4, 0.4, 1.0], lenpen=lenpen) + # sentence 2, beam 2 + self.assertHypoTokens(hypos[1][1], [w1, w2, eos]) + self.assertHypoScore(hypos[1][1], [0.7, 0.4, 0.6], lenpen=lenpen) + + def test_maxlen(self): + generator = SequenceGenerator( + [self.model], self.tgt_dict, beam_size=2, max_len_b=2 + ) + hypos = generator.forward(self.sample) + eos, w1, w2 = self.tgt_dict.eos(), self.w1, self.w2 + # sentence 1, beam 1 + self.assertHypoTokens(hypos[0][0], [w1, eos]) + self.assertHypoScore(hypos[0][0], [0.9, 1.0]) + # sentence 1, beam 2 + self.assertHypoTokens(hypos[0][1], [w2, w2, eos]) + self.assertHypoScore(hypos[0][1], [0.1, 0.1, 0.6]) + # sentence 2, beam 1 + self.assertHypoTokens(hypos[1][0], [w1, w2, eos]) + self.assertHypoScore(hypos[1][0], [0.7, 0.4, 0.6]) + # sentence 2, beam 2 + self.assertHypoTokens(hypos[1][1], [w2, w2, eos]) + self.assertHypoScore(hypos[1][1], [0.3, 0.9, 0.01]) + + def test_encoder_with_different_output_len(self): + args = self.model.encoder.args + task = test_utils.TestTranslationTask.setup_task( + args, self.tgt_dict, self.tgt_dict + ) + reshaping_model = test_utils.TestReshapingModel.build_model(args, task) + generator = SequenceGenerator( + [reshaping_model], self.tgt_dict, beam_size=2, max_len_b=2 + ) + hypos = generator.forward(self.sample) + for sent in [0, 1]: + for beam in [0, 1]: + assert hypos[sent][beam]["attention"] is not None + + def test_generation_with_additional_input(self): + args = self.model.encoder.args + task = test_utils.TestTranslationTask.setup_task( + args, self.tgt_dict, self.tgt_dict + ) + add_input_model = test_utils.TestAdditionalInputModel.build_model(args, task) + generator = SequenceGenerator([add_input_model], self.tgt_dict, beam_size=2) + sample = self.sample.copy() + sample["net_input"]["fancy_other_input"] = sample["net_input"]["src_tokens"] + hypos = generator.forward(self.sample) + eos, w1, w2 = self.tgt_dict.eos(), self.w1, self.w2 + # sentence 1, beam 1 + self.assertHypoTokens(hypos[0][0], [w1, eos]) + self.assertHypoScore(hypos[0][0], [0.9, 1.0]) + + +@unittest.skipUnless(torch.cuda.is_available(), "") +class TestRepeatNgramBlocking(TestSequenceGeneratorBase): + @classmethod + def setUpClass(cls): + ( + cls.tgt_dict, + cls.w1, + cls.w2, + src_tokens, + src_lengths, + cls.model, + ) = test_utils.sequence_generator_setup() + return cls + + def test_finds_repetitive_tokens(self): + bsz, vocab_size, beam_size, step = 2, 4, 1, 3 + generated_tok = torch.tensor( + [[2, 2, 2, 2], [3, 3, 3, 3]], dtype=torch.long, device="cuda" + ) + lprobs = torch.zeros((beam_size * bsz, vocab_size), device="cuda") + desired_result = lprobs.new_tensor( + [[0.0, 0.0, -math.inf, 0.0], [0.0, 0.0, 0.0, -math.inf]] + ) + + cuda_ext_result, baseline_result = self._compare_cuda_ext_to_default_implem( + bsz, beam_size, generated_tok, lprobs, step, 2 + ) + self.assertTensorEqual(cuda_ext_result, desired_result) + self.assertTensorEqual(baseline_result, desired_result) + + @unittest.skipIf(torch.__version__ < "1.6.0", JIT_MSG) + def test_jit_no_extension(self): + bsz, vocab_size, beam_size, step = 2, 4, 1, 3 + generated_tok = torch.tensor( + [[2, 2, 2, 2], [3, 3, 3, 3]], dtype=torch.long, device="cuda" + ) + lprobs = torch.zeros((beam_size * bsz, vocab_size), device="cuda") + blocker = NGramRepeatBlock(2, use_extension=False) + base_result = blocker(generated_tok, lprobs.clone(), bsz, beam_size, step) + scripted_blocker = torch.jit.script(blocker) + jit_result = scripted_blocker( + generated_tok, lprobs.clone(), bsz, beam_size, step + ) + self.assertTensorEqual(base_result, jit_result) + + def test_ngram_blocking_same_as_default_implem(self): + """Test that cuda extension returns same things as default impl in many settings.""" + vocab_size = 4 + step = 6 + for _ in range(2): + block_param = np.random.choice([1, 2, 3, 4]) + batch_size = np.random.randint(1, 8) + beam_size = np.random.choice([1, 2, 4, 8]) + lprobs = torch.zeros((beam_size * batch_size, vocab_size), device="cuda") + + generated_tok = torch.tensor( + np.random.randint( + 0, vocab_size, size=(batch_size * beam_size, step + 1) + ), + device="cuda", + dtype=torch.long, + ) + self._compare_cuda_ext_to_default_implem( + batch_size, + beam_size, + generated_tok, + lprobs, + step, + block_param, + ) + + def _compare_cuda_ext_to_default_implem( + self, bsz, beam_size, generated_tok, lprobs, step, block_param + ): + """Assert that cuda extension and default implem return the same thing.""" + blocker = NGramRepeatBlock(block_param) + assert blocker.use_extension, "Extension not compiled" + cuda_ext_result = blocker( + generated_tok, + lprobs.clone(), + bsz, + beam_size, + step, + ) + blocker.use_extension = False + baseline_result = blocker( + generated_tok, + lprobs.clone(), + bsz, + beam_size, + step, + ) + self.assertTensorEqual(cuda_ext_result, baseline_result) + blocker.use_extension = True + return cuda_ext_result, baseline_result + + +class TestDiverseBeamSearch(TestSequenceGeneratorBase): + def setUp(self): + # construct dummy dictionary + d = test_utils.dummy_dictionary(vocab_size=2) + self.assertEqual(d.pad(), 1) + self.assertEqual(d.eos(), 2) + self.assertEqual(d.unk(), 3) + self.eos = d.eos() + self.w1 = 4 + self.w2 = 5 + + # construct source data + self.src_tokens = torch.LongTensor( + [ + [self.w1, self.w2, self.eos], + [self.w1, self.w2, self.eos], + ] + ) + self.src_lengths = torch.LongTensor([2, 2]) + + args = argparse.Namespace() + unk = 0.0 + args.beam_probs = [ + # step 0: + torch.FloatTensor( + [ + # eos w1 w2 + # sentence 1: + [0.0, unk, 0.9, 0.1], # beam 1 + [0.0, unk, 0.9, 0.1], # beam 2 + # sentence 2: + [0.0, unk, 0.7, 0.3], + [0.0, unk, 0.7, 0.3], + ] + ), + # step 1: + torch.FloatTensor( + [ + # eos w1 w2 + # sentence 1: + [0.0, unk, 0.6, 0.4], + [0.0, unk, 0.6, 0.4], + # sentence 2: + [0.25, unk, 0.35, 0.4], + [0.25, unk, 0.35, 0.4], + ] + ), + # step 2: + torch.FloatTensor( + [ + # eos w1 w2 + # sentence 1: + [1.0, unk, 0.0, 0.0], + [1.0, unk, 0.0, 0.0], + # sentence 2: + [0.9, unk, 0.1, 0.0], + [0.9, unk, 0.1, 0.0], + ] + ), + ] + + task = test_utils.TestTranslationTask.setup_task(args, d, d) + self.model = task.build_model(args) + self.tgt_dict = task.target_dictionary + + def test_diverse_beam_search(self): + search_strategy = search.DiverseBeamSearch( + self.tgt_dict, num_groups=2, diversity_strength=0.0 + ) + generator = SequenceGenerator( + [self.model], + self.tgt_dict, + beam_size=2, + search_strategy=search_strategy, + ) + sample = { + "net_input": { + "src_tokens": self.src_tokens, + "src_lengths": self.src_lengths, + } + } + hypos = generator.forward(sample) + eos, w1, w2 = self.eos, self.w1, self.w2 + # sentence 1, beam 1 + self.assertHypoTokens(hypos[0][0], [w1, w1, eos]) + self.assertHypoScore(hypos[0][0], [0.9, 0.6, 1.0]) + # sentence 1, beam 2 + self.assertHypoTokens(hypos[0][1], [w1, w1, eos]) + self.assertHypoScore(hypos[0][1], [0.9, 0.6, 1.0]) + # sentence 2, beam 1 + self.assertHypoTokens(hypos[1][0], [w1, w2, eos]) + self.assertHypoScore(hypos[1][0], [0.7, 0.4, 0.9]) + # sentence 2, beam 2 + self.assertHypoTokens(hypos[1][1], [w1, w2, eos]) + self.assertHypoScore(hypos[1][1], [0.7, 0.4, 0.9]) + + +class TestDiverseSiblingsSearch(TestDiverseBeamSearch): + def assertHypoScore( + self, hypo, pos_probs, sibling_rank, diversity_rate, normalized=True, lenpen=1.0 + ): + pos_scores = torch.FloatTensor(pos_probs).log() + pos_scores.sub_(torch.Tensor(sibling_rank) * diversity_rate) + self.assertAlmostEqual(hypo["positional_scores"], pos_scores) + self.assertEqual(pos_scores.numel(), hypo["tokens"].numel()) + score = pos_scores.sum() + if normalized: + score /= pos_scores.numel() ** lenpen + self.assertLess(abs(score - hypo["score"]), 1e-6) + + def test_diverse_beam_search(self): + search_strategy = search.DiverseSiblingsSearch( + self.tgt_dict, diversity_rate=0.5 + ) + generator = SequenceGenerator( + [self.model], self.tgt_dict, beam_size=2, search_strategy=search_strategy + ) + sample = { + "net_input": { + "src_tokens": self.src_tokens, + "src_lengths": self.src_lengths, + } + } + hypos = generator.forward(sample) + eos, w1, w2 = self.eos, self.w1, self.w2 + # sentence 1, beam 1 + self.assertHypoTokens(hypos[0][0], [w1, w1, eos]) + self.assertHypoScore(hypos[0][0], [0.9, 0.6, 1.0], [0, 1, 1], 0.5) + # sentence 1, beam 2 + self.assertHypoTokens(hypos[0][1], [w1, w2, eos]) + self.assertHypoScore(hypos[0][1], [0.9, 0.4, 1.0], [0, 2, 1], 0.5) + # sentence 2, beam 1 + self.assertHypoTokens(hypos[1][0], [w1, w2, eos]) + self.assertHypoScore(hypos[1][0], [0.7, 0.4, 0.9], [0, 1, 1], 0.5) + # sentence 2, beam 2 + self.assertHypoTokens(hypos[1][1], [w1, w1, eos]) + self.assertHypoScore(hypos[1][1], [0.7, 0.35, 0.9], [0, 2, 1], 0.5) + + +class TestTopPSamplingSearch(TestSequenceGeneratorBase): + def setUp(self): + # construct dummy dictionary + d = test_utils.dummy_dictionary(vocab_size=2) + self.assertEqual(d.pad(), 1) + self.assertEqual(d.eos(), 2) + self.assertEqual(d.unk(), 3) + self.eos = d.eos() + self.w1 = 4 + self.w2 = 5 + + # construct source data + self.src_tokens = torch.LongTensor( + [ + [self.w1, self.w2, self.eos], + [self.w1, self.w2, self.eos], + ] + ) + self.src_lengths = torch.LongTensor([2, 2]) + + args = argparse.Namespace() + unk = 0.0 + # The minimal probability of top 2 tokens. + self.min_top2_prob = 0.75 + # The minimal probability of the top 1 token. + self.min_top1_prob = 0.4 + + w1_prob = self.min_top1_prob + w2_prob = self.min_top2_prob - self.min_top1_prob + eos_prob = 1 - self.min_top2_prob + + args.beam_probs = [ + # step 0: + torch.FloatTensor( + [ + # eos w1 w2 + [0.0, unk, 1.0, 0.0], + [0.0, unk, 1.0, 0.0], + [0.0, unk, 1.0, 0.0], + [0.0, unk, 1.0, 0.0], + ] + ), + # step 1: + torch.FloatTensor( + [ + # eos w1 w2 + [eos_prob, unk, w1_prob, w2_prob], + [eos_prob, unk, w1_prob, w2_prob], + [eos_prob, unk, w1_prob, w2_prob], + [eos_prob, unk, w1_prob, w2_prob], + ] + ), + # step 2: + torch.FloatTensor( + [ + # eos w1 w2 + [1.0, unk, 0.0, 0.0], + [1.0, unk, 0.0, 0.0], + [1.0, unk, 0.0, 0.0], + [1.0, unk, 0.0, 0.0], + ] + ), + ] + + task = test_utils.TestTranslationTask.setup_task(args, d, d) + self.model = task.build_model(args) + self.tgt_dict = task.target_dictionary + + def test_topp_sampling_search_low_prob(self): + # Given a prob low enough to top-P sampling, we expect only the top + # 1 token to be sampled, which always results in the same output. + low_sampling_topp = self.min_top1_prob / 2.0 + search_strategy = search.Sampling( + self.tgt_dict, sampling_topp=low_sampling_topp + ) + generator = SequenceGenerator( + [self.model], self.tgt_dict, beam_size=2, search_strategy=search_strategy + ) + sample = { + "net_input": { + "src_tokens": self.src_tokens, + "src_lengths": self.src_lengths, + } + } + hypos = generator.forward(sample) + eos, w1 = self.eos, self.w1 + # sentence 1, beam 1 + self.assertHypoTokens(hypos[0][0], [w1, w1, eos]) + self.assertHypoScore(hypos[0][0], [1.0, 0.4, 1.0]) + # sentence 1, beam 2 + self.assertHypoTokens(hypos[0][1], [w1, w1, eos]) + self.assertHypoScore(hypos[0][1], [1.0, 0.4, 1.0]) + # sentence 2, beam 1 + self.assertHypoTokens(hypos[1][0], [w1, w1, eos]) + self.assertHypoScore(hypos[1][0], [1.0, 0.4, 1.0]) + # sentence 2, beam 2 + self.assertHypoTokens(hypos[1][1], [w1, w1, eos]) + self.assertHypoScore(hypos[1][1], [1.0, 0.4, 1.0]) + + def test_topp_sampling_search_high_prob(self): + # Given a prob high enough to top-P sampling, any of the top 2 + # tokens could be sampled. This can cause different outputs. + high_sampling_topp = (self.min_top1_prob + self.min_top2_prob) / 2.0 + search_strategy = search.Sampling( + self.tgt_dict, sampling_topp=high_sampling_topp + ) + generator = SequenceGenerator( + [self.model], self.tgt_dict, beam_size=2, search_strategy=search_strategy + ) + sample = { + "net_input": { + "src_tokens": self.src_tokens, + "src_lengths": self.src_lengths, + } + } + hypos = generator.forward(sample) + eos, w1, w2 = self.eos, self.w1, self.w2 + # sentence 1, beam 1 + self.assertTrue( + self.hypoTokens(hypos[0][0], [w1, w1, eos]) + or self.hypoTokens(hypos[0][0], [w1, w2, eos]) + ) + self.assertTrue( + self.hypoScore(hypos[0][0], [1.0, 0.4, 1.0]) + or self.hypoScore(hypos[0][0], [1.0, 0.35, 1.0]) + ) + + # sentence 1, beam 2 + self.assertTrue( + self.hypoTokens(hypos[0][1], [w1, w1, eos]) + or self.hypoTokens(hypos[0][1], [w1, w2, eos]) + ) + self.assertTrue( + self.hypoScore(hypos[0][1], [1.0, 0.4, 1.0]) + or self.hypoScore(hypos[0][1], [1.0, 0.35, 1.0]) + ) + + # sentence 2, beam 1 + self.assertTrue( + self.hypoTokens(hypos[1][0], [w1, w1, eos]) + or self.hypoTokens(hypos[1][0], [w1, w2, eos]) + ) + self.assertTrue( + self.hypoScore(hypos[1][0], [1.0, 0.4, 1.0]) + or self.hypoScore(hypos[1][0], [1.0, 0.35, 1.0]) + ) + + # sentence 2, beam 2 + self.assertTrue( + self.hypoTokens(hypos[1][1], [w1, w1, eos]) + or self.hypoTokens(hypos[1][1], [w1, w2, eos]) + ) + self.assertTrue( + self.hypoScore(hypos[1][1], [1.0, 0.4, 1.0]) + or self.hypoScore(hypos[1][1], [1.0, 0.35, 1.0]) + ) + + def hypoTokens(self, hypo, tokens): + return self.tensorEqual(hypo["tokens"], torch.LongTensor(tokens)) + + def hypoScore(self, hypo, pos_probs, normalized=True, lenpen=1.0): + pos_scores = torch.FloatTensor(pos_probs).log() + if not self.almostEqual(hypo["positional_scores"], pos_scores): + return False + if pos_scores.numel() != hypo["tokens"].numel(): + return False + score = pos_scores.sum() + if normalized: + score /= pos_scores.numel() ** lenpen + return abs(score - hypo["score"]) < 1e-6 + + def almostEqual(self, t1, t2): + return t1.size() == t2.size() and (t1 - t2).abs().max() < 1e-4 + + def tensorEqual(self, t1, t2): + return t1.size() == t2.size() and t1.ne(t2).long().sum() == 0 + + +if __name__ == "__main__": + unittest.main() diff --git a/tests/test_sequence_scorer.py b/tests/test_sequence_scorer.py new file mode 100644 index 0000000000000000000000000000000000000000..42f9447b599bcd7a9913aec37d94ea5078ff43a3 --- /dev/null +++ b/tests/test_sequence_scorer.py @@ -0,0 +1,120 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import unittest + +import tests.utils as test_utils +import torch +from fairseq.sequence_scorer import SequenceScorer + + +class TestSequenceScorer(unittest.TestCase): + def test_sequence_scorer(self): + # construct dummy dictionary + d = test_utils.dummy_dictionary(vocab_size=2) + self.assertEqual(d.pad(), 1) + self.assertEqual(d.eos(), 2) + self.assertEqual(d.unk(), 3) + eos = d.eos() + w1 = 4 + w2 = 5 + + # construct dataloader + data = [ + { + "source": torch.LongTensor([w1, w2, eos]), + "target": torch.LongTensor([w1, w2, w1, eos]), + }, + { + "source": torch.LongTensor([w2, eos]), + "target": torch.LongTensor([w2, w1, eos]), + }, + { + "source": torch.LongTensor([w2, eos]), + "target": torch.LongTensor([w2, eos]), + }, + ] + data_itr = test_utils.dummy_dataloader(data) + + # specify expected output probabilities + args = argparse.Namespace() + unk = 0.0 + args.beam_probs = [ + # step 0: + torch.FloatTensor( + [ + # eos w1 w2 + [0.0, unk, 0.6, 0.4], # sentence 1 + [0.0, unk, 0.4, 0.6], # sentence 2 + [0.0, unk, 0.7, 0.3], # sentence 3 + ] + ), + # step 1: + torch.FloatTensor( + [ + # eos w1 w2 + [0.0, unk, 0.2, 0.7], # sentence 1 + [0.0, unk, 0.8, 0.2], # sentence 2 + [0.7, unk, 0.1, 0.2], # sentence 3 + ] + ), + # step 2: + torch.FloatTensor( + [ + # eos w1 w2 + [0.10, unk, 0.50, 0.4], # sentence 1 + [0.15, unk, 0.15, 0.7], # sentence 2 + [0.00, unk, 0.00, 0.0], # sentence 3 + ] + ), + # step 3: + torch.FloatTensor( + [ + # eos w1 w2 + [0.9, unk, 0.05, 0.05], # sentence 1 + [0.0, unk, 0.00, 0.0], # sentence 2 + [0.0, unk, 0.00, 0.0], # sentence 3 + ] + ), + ] + expected_scores = [ + [0.6, 0.7, 0.5, 0.9], # sentence 1 + [0.6, 0.8, 0.15], # sentence 2 + [0.3, 0.7], # sentence 3 + ] + + task = test_utils.TestTranslationTask.setup_task(args, d, d) + model = task.build_model(args) + scorer = SequenceScorer(task.target_dictionary) + for sample in data_itr: + hypos = task.inference_step(scorer, [model], sample) + for id, hypos_id in zip(sample["id"].tolist(), hypos): + self.assertHypoTokens(hypos_id[0], data[id]["target"]) + self.assertHypoScore(hypos_id[0], expected_scores[id]) + + def assertHypoTokens(self, hypo, tokens): + self.assertTensorEqual(hypo["tokens"], torch.LongTensor(tokens)) + + def assertHypoScore(self, hypo, pos_probs, normalized=True, lenpen=1.0): + pos_scores = torch.FloatTensor(pos_probs).log() + self.assertAlmostEqual(hypo["positional_scores"], pos_scores) + self.assertEqual(pos_scores.numel(), hypo["tokens"].numel()) + score = pos_scores.sum() + if normalized: + score /= pos_scores.numel() ** lenpen + self.assertLess(abs(score - hypo["score"]), 1e-6) + + def assertAlmostEqual(self, t1, t2): + self.assertEqual(t1.size(), t2.size(), "size mismatch") + self.assertLess((t1 - t2).abs().max(), 1e-4) + + def assertTensorEqual(self, t1, t2): + self.assertEqual(t1.size(), t2.size(), "size mismatch") + self.assertEqual(t1.ne(t2).long().sum(), 0) + + +if __name__ == "__main__": + unittest.main() diff --git a/tests/test_sparse_multihead_attention.py b/tests/test_sparse_multihead_attention.py new file mode 100644 index 0000000000000000000000000000000000000000..3e32b25a7fb1e12295b84d0c65064f8e42b7bdd3 --- /dev/null +++ b/tests/test_sparse_multihead_attention.py @@ -0,0 +1,114 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import unittest + +import torch +from fairseq.modules.sparse_multihead_attention import SparseMultiheadAttention + + +class TestSparseMultiheadAttention(unittest.TestCase): + def test_sparse_multihead_attention(self): + attn_weights = torch.randn(1, 8, 8) + bidirectional_sparse_mask = torch.tensor( + [ + [0, 0, 0, 0, 0, float("-inf"), float("-inf"), 0], + [0, 0, 0, 0, 0, float("-inf"), float("-inf"), 0], + [0, 0, 0, 0, 0, float("-inf"), float("-inf"), 0], + [0, 0, 0, 0, 0, float("-inf"), float("-inf"), 0], + [float("-inf"), float("-inf"), float("-inf"), 0, 0, 0, 0, 0], + [float("-inf"), float("-inf"), float("-inf"), 0, 0, 0, 0, 0], + [float("-inf"), float("-inf"), float("-inf"), 0, 0, 0, 0, 0], + [float("-inf"), float("-inf"), float("-inf"), 0, 0, 0, 0, 0], + ] + ) + + bidirectional_attention = SparseMultiheadAttention( + 16, 1, stride=4, expressivity=1, is_bidirectional=True + ) + bidirectional_attention_sparse_mask = ( + bidirectional_attention.buffered_sparse_mask(attn_weights, 8, 8) + ) + torch.all( + torch.eq(bidirectional_attention_sparse_mask, bidirectional_sparse_mask) + ) + + sparse_mask = torch.tensor( + [ + [ + 0, + float("-inf"), + float("-inf"), + float("-inf"), + float("-inf"), + float("-inf"), + float("-inf"), + float("-inf"), + ], + [ + 0, + 0, + float("-inf"), + float("-inf"), + float("-inf"), + float("-inf"), + float("-inf"), + float("-inf"), + ], + [ + 0, + 0, + 0, + float("-inf"), + float("-inf"), + float("-inf"), + float("-inf"), + float("-inf"), + ], + [ + 0, + 0, + 0, + 0, + float("-inf"), + float("-inf"), + float("-inf"), + float("-inf"), + ], + [0, 0, 0, 0, 0, float("-inf"), float("-inf"), float("-inf")], + [ + float("-inf"), + float("-inf"), + float("-inf"), + 0, + 0, + 0, + float("-inf"), + float("-inf"), + ], + [ + float("-inf"), + float("-inf"), + float("-inf"), + 0, + 0, + 0, + 0, + float("-inf"), + ], + [float("-inf"), float("-inf"), float("-inf"), 0, 0, 0, 0, 0], + ] + ) + + attention = SparseMultiheadAttention( + 16, 1, stride=4, expressivity=1, is_bidirectional=False + ) + attention_sparse_mask = attention.buffered_sparse_mask(attn_weights, 8, 8) + + torch.all(torch.eq(attention_sparse_mask, sparse_mask)) + + +if __name__ == "__main__": + unittest.main() diff --git a/tests/test_token_block_dataset.py b/tests/test_token_block_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..c4d7b76dcd55fe7869dbb1fa188f7b36fb639bda --- /dev/null +++ b/tests/test_token_block_dataset.py @@ -0,0 +1,92 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import unittest + +import tests.utils as test_utils +import torch +from fairseq.data import TokenBlockDataset + + +class TestTokenBlockDataset(unittest.TestCase): + def _build_dataset(self, data, **kwargs): + sizes = [len(x) for x in data] + underlying_ds = test_utils.TestDataset(data) + return TokenBlockDataset(underlying_ds, sizes, **kwargs) + + def test_eos_break_mode(self): + data = [ + torch.tensor([5, 4, 3, 2, 1], dtype=torch.long), + torch.tensor([1], dtype=torch.long), + torch.tensor([8, 7, 6, 1], dtype=torch.long), + ] + ds = self._build_dataset(data, block_size=None, pad=0, eos=1, break_mode="eos") + self.assertEqual(ds[0].tolist(), [5, 4, 3, 2, 1]) + self.assertEqual(ds[1].tolist(), [1]) + self.assertEqual(ds[2].tolist(), [8, 7, 6, 1]) + + data = [ + torch.tensor([5, 4, 3, 2, 1], dtype=torch.long), + torch.tensor([8, 7, 6, 1], dtype=torch.long), + torch.tensor([1], dtype=torch.long), + ] + ds = self._build_dataset(data, block_size=None, pad=0, eos=1, break_mode="eos") + self.assertEqual(ds[0].tolist(), [5, 4, 3, 2, 1]) + self.assertEqual(ds[1].tolist(), [8, 7, 6, 1]) + self.assertEqual(ds[2].tolist(), [1]) + + def test_block_break_mode(self): + data = [ + torch.tensor([5, 4, 3, 2, 1], dtype=torch.long), + torch.tensor([8, 7, 6, 1], dtype=torch.long), + torch.tensor([9, 1], dtype=torch.long), + ] + ds = self._build_dataset(data, block_size=3, pad=0, eos=1, break_mode="none") + self.assertEqual(ds[0].tolist(), [5, 4, 3]) + self.assertEqual(ds[1].tolist(), [2, 1, 8]) + self.assertEqual(ds[2].tolist(), [7, 6, 1]) + self.assertEqual(ds[3].tolist(), [9, 1]) + + def test_complete_break_mode(self): + data = [ + torch.tensor([5, 4, 3, 2, 1], dtype=torch.long), + torch.tensor([8, 7, 6, 1], dtype=torch.long), + torch.tensor([9, 1], dtype=torch.long), + ] + ds = self._build_dataset( + data, block_size=6, pad=0, eos=1, break_mode="complete" + ) + self.assertEqual(ds[0].tolist(), [5, 4, 3, 2, 1]) + self.assertEqual(ds[1].tolist(), [8, 7, 6, 1, 9, 1]) + + data = [ + torch.tensor([4, 3, 2, 1], dtype=torch.long), + torch.tensor([5, 1], dtype=torch.long), + torch.tensor([1], dtype=torch.long), + torch.tensor([6, 1], dtype=torch.long), + ] + ds = self._build_dataset( + data, block_size=3, pad=0, eos=1, break_mode="complete" + ) + self.assertEqual(ds[0].tolist(), [4, 3, 2, 1]) + self.assertEqual(ds[1].tolist(), [5, 1, 1]) + self.assertEqual(ds[2].tolist(), [6, 1]) + + def test_4billion_tokens(self): + """Regression test for numpy type promotion issue https://github.com/numpy/numpy/issues/5745""" + data = [torch.tensor(list(range(10000)), dtype=torch.long)] * 430000 + ds = self._build_dataset( + data, block_size=6, pad=0, eos=1, break_mode="complete" + ) + ds[-1] # __getitem__ works + start, end = ds.slice_indices[-1] + assert end > 4294967295 # data must be sufficiently large to overflow uint32 + assert not isinstance( + end + 1, float + ) # this would also raise, since np.uint64(1) + 1 => 2.0 + + +if __name__ == "__main__": + unittest.main() diff --git a/tests/test_train.py b/tests/test_train.py new file mode 100644 index 0000000000000000000000000000000000000000..65f4683bc67ca80c81bf1d2c27be621b57f7df94 --- /dev/null +++ b/tests/test_train.py @@ -0,0 +1,246 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import contextlib +import logging +import unittest +from io import StringIO +from unittest.mock import MagicMock, patch + +import torch +from fairseq import checkpoint_utils, data +from omegaconf import OmegaConf + + +def mock_trainer(epoch, num_updates, iterations_in_epoch): + trainer = MagicMock() + trainer.load_checkpoint.return_value = { + "train_iterator": { + "epoch": epoch, + "iterations_in_epoch": iterations_in_epoch, + "shuffle": False, + }, + } + trainer.get_num_updates.return_value = num_updates + return trainer + + +def mock_dict(): + d = MagicMock() + d.pad.return_value = 1 + d.eos.return_value = 2 + d.unk.return_value = 3 + return d + + +def get_trainer_and_epoch_itr(epoch, epoch_size, num_updates, iterations_in_epoch): + tokens = torch.LongTensor(list(range(epoch_size))).view(1, -1) + tokens_ds = data.TokenBlockDataset( + tokens, + sizes=[tokens.size(-1)], + block_size=1, + pad=0, + eos=1, + include_targets=False, + ) + trainer = mock_trainer(epoch, num_updates, iterations_in_epoch) + dataset = data.LanguagePairDataset( + tokens_ds, tokens_ds.sizes, mock_dict(), shuffle=False + ) + epoch_itr = data.EpochBatchIterator( + dataset=dataset, + collate_fn=dataset.collater, + batch_sampler=[[i] for i in range(epoch_size)], + ) + return trainer, epoch_itr + + +def get_mock_cfg(finetune_from_model): + cfg_mock = OmegaConf.create( + { + "checkpoint": { + "optimizer_overrides": "{}", + "reset_dataloader": False, + "reset_meters": False, + "reset_optimizer": False, + "reset_lr_scheduler": False, + "finetune_from_model": finetune_from_model, + "model_parallel_size": 1, + "restore_file": "checkpoint_last.pt", + }, + "common": { + "model_parallel_size": 1, + }, + } + ) + return cfg_mock + + +class TestLoadCheckpoint(unittest.TestCase): + def setUp(self): + self.cfg_mock = get_mock_cfg(None) + self.patches = { + "os.makedirs": MagicMock(), + "os.path.join": MagicMock(), + "os.path.isfile": MagicMock(return_value=True), + "os.path.isabs": MagicMock(return_value=False), + "fairseq.file_io.PathManager.exists": MagicMock(return_value=False), + } + self.applied_patches = [patch(p, d) for p, d in self.patches.items()] + [p.start() for p in self.applied_patches] + logging.disable(logging.CRITICAL) + + def tearDown(self): + patch.stopall() + logging.disable(logging.NOTSET) + + def test_load_partial_checkpoint(self): + with contextlib.redirect_stdout(StringIO()): + trainer, epoch_itr = get_trainer_and_epoch_itr(2, 150, 200, 50) + trainer.get_train_iterator = MagicMock(return_value=epoch_itr) + + _, epoch_itr = checkpoint_utils.load_checkpoint( + self.cfg_mock.checkpoint, trainer + ) + + self.assertEqual(epoch_itr.epoch, 2) + self.assertEqual(epoch_itr.iterations_in_epoch, 50) + + itr = epoch_itr.next_epoch_itr(shuffle=False) + self.assertEqual(epoch_itr.epoch, 2) + self.assertEqual(epoch_itr.iterations_in_epoch, 50) + + self.assertEqual(next(itr)["net_input"]["src_tokens"][0].item(), 50) + self.assertEqual(epoch_itr.iterations_in_epoch, 51) + + for _ in range(150 - 52): + next(itr) + self.assertEqual(epoch_itr.iterations_in_epoch, 149) + self.assertTrue(itr.has_next()) + next(itr) + self.assertFalse(itr.has_next()) + + itr = epoch_itr.next_epoch_itr(shuffle=False) + self.assertTrue(itr.has_next()) + self.assertEqual(epoch_itr.epoch, 3) + self.assertEqual(epoch_itr.iterations_in_epoch, 0) + + def test_load_full_checkpoint(self): + with contextlib.redirect_stdout(StringIO()): + trainer, epoch_itr = get_trainer_and_epoch_itr(2, 150, 300, 150) + trainer.get_train_iterator = MagicMock(return_value=epoch_itr) + + _, epoch_itr = checkpoint_utils.load_checkpoint( + self.cfg_mock.checkpoint, trainer + ) + itr = epoch_itr.next_epoch_itr(shuffle=False) + + self.assertEqual(epoch_itr.epoch, 3) + self.assertEqual(epoch_itr.iterations_in_epoch, 0) + self.assertEqual(next(itr)["net_input"]["src_tokens"][0].item(), 0) + + def test_load_no_checkpoint(self): + with contextlib.redirect_stdout(StringIO()): + trainer, epoch_itr = get_trainer_and_epoch_itr(1, 150, 0, 0) + trainer.get_train_iterator = MagicMock(return_value=epoch_itr) + self.patches["os.path.isfile"].return_value = False + + _, epoch_itr = checkpoint_utils.load_checkpoint( + self.cfg_mock.checkpoint, trainer + ) + itr = epoch_itr.next_epoch_itr(shuffle=False) + + self.assertEqual(epoch_itr.epoch, 1) + self.assertEqual(epoch_itr.iterations_in_epoch, 0) + self.assertEqual(next(itr)["net_input"]["src_tokens"][0].item(), 0) + + def test_finetune_from_model_args_conflict(self): + with contextlib.redirect_stdout(StringIO()): + trainer, epoch_itr = get_trainer_and_epoch_itr(1, 150, 0, 0) + trainer.get_train_iterator = MagicMock(return_value=epoch_itr) + + for arg in [ + "reset_optimizer", + "reset_lr_scheduler", + "reset_meters", + "reset_dataloader", + ]: + with self.subTest(arg=arg): + cfg_mock = get_mock_cfg("/temp/checkpoint_pretrained.pt") + cfg_mock["checkpoint"][arg] = True + with self.assertRaises(Exception) as context: + _, _ = checkpoint_utils.load_checkpoint( + cfg_mock.checkpoint, trainer + ) + + self.assertTrue( + "--finetune-from-model can not be set together with either --reset-optimizer" + " or reset_lr_scheduler or reset_meters or reset_dataloader" + in str(context.exception) + ) + + def test_finetune_from_model(self): + with contextlib.redirect_stdout(StringIO()): + trainer, epoch_itr = get_trainer_and_epoch_itr(1, 150, 0, 0) + trainer.get_train_iterator = MagicMock(return_value=epoch_itr) + from_model_path = "/temp/checkpoint_pretrained.pt" + + def mock_finetune_exist(path): + if path == from_model_path: + return True + else: + return False + + self.patches[ + "fairseq.file_io.PathManager.exists" + ].side_effect = mock_finetune_exist + cfg_mock = get_mock_cfg(from_model_path) + cfg_mock.checkpoint.restore_file = "checkpoint_last.pt" + _, _ = checkpoint_utils.load_checkpoint(cfg_mock.checkpoint, trainer) + ( + checkpoint_path, + reset_optimizer, + reset_lr_scheduler, + optimizer_overrides, + ) = trainer.load_checkpoint.call_args[0] + reset_meters = trainer.load_checkpoint.call_args[1]["reset_meters"] + self.assertTrue(reset_optimizer) + self.assertTrue(reset_lr_scheduler) + self.assertTrue(reset_meters) + + def test_finetune_from_model_resume(self): + with contextlib.redirect_stdout(StringIO()): + trainer, epoch_itr = get_trainer_and_epoch_itr(1, 150, 0, 0) + trainer.get_train_iterator = MagicMock(return_value=epoch_itr) + from_model_path = "/temp/checkpoint_pretrained.pt" + + # launch second time + # both restore_file=checkpoint_last.pt and finetune_from_model are set + def mock_finetune_exist(path): + if path == from_model_path or path.endsWith("checkpoint_last.pt"): + return True + else: + return False + + self.patches[ + "fairseq.file_io.PathManager.exists" + ].side_effect = mock_finetune_exist + cfg_mock = get_mock_cfg(from_model_path) + cfg_mock.checkpoint.restore_file = "checkpoint_last.pt" + _, _ = checkpoint_utils.load_checkpoint(cfg_mock.checkpoint, trainer) + ( + checkpoint_path, + reset_optimizer, + reset_lr_scheduler, + optimizer_overrides, + ) = trainer.load_checkpoint.call_args[0] + reset_meters = trainer.load_checkpoint.call_args[1]["reset_meters"] + self.assertFalse(reset_optimizer) + self.assertFalse(reset_lr_scheduler) + self.assertFalse(reset_meters) + + +if __name__ == "__main__": + unittest.main() diff --git a/tests/test_transformer.py b/tests/test_transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..de5c5bdbd49692e63fb1cb50108a791304425dc1 --- /dev/null +++ b/tests/test_transformer.py @@ -0,0 +1,65 @@ +import argparse +import unittest +from typing import Any, Dict, Sequence + +import torch +from fairseq.models import transformer + +from tests.test_roberta import FakeTask + + +def mk_sample(tok: Sequence[int] = None, batch_size: int = 2) -> Dict[str, Any]: + if not tok: + tok = [10, 11, 12, 13, 14, 15, 2] + + batch = torch.stack([torch.tensor(tok, dtype=torch.long)] * batch_size) + sample = { + "net_input": { + "src_tokens": batch, + "prev_output_tokens": batch, + "src_lengths": torch.tensor( + [len(tok)] * batch_size, dtype=torch.long, device=batch.device + ), + }, + "target": batch[:, 1:], + } + return sample + + +def mk_transformer(**extra_args: Any): + overrides = { + # Use characteristics dimensions + "encoder_embed_dim": 12, + "encoder_ffn_embed_dim": 14, + "decoder_embed_dim": 12, + "decoder_ffn_embed_dim": 14, + # Disable dropout so we have comparable tests. + "dropout": 0, + "attention_dropout": 0, + "activation_dropout": 0, + "encoder_layerdrop": 0, + } + overrides.update(extra_args) + # Overrides the defaults from the parser + args = argparse.Namespace(**overrides) + transformer.tiny_architecture(args) + + torch.manual_seed(0) + task = FakeTask(args) + return transformer.TransformerModel.build_model(args, task) + + +class TransformerTestCase(unittest.TestCase): + def test_forward_backward(self): + model = mk_transformer(encoder_embed_dim=12, decoder_embed_dim=12) + sample = mk_sample() + o, _ = model.forward(**sample["net_input"]) + loss = o.sum() + loss.backward() + + def test_different_encoder_decoder_embed_dim(self): + model = mk_transformer(encoder_embed_dim=12, decoder_embed_dim=16) + sample = mk_sample() + o, _ = model.forward(**sample["net_input"]) + loss = o.sum() + loss.backward() diff --git a/tests/test_utils.py b/tests/test_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..79195903e0f34372a24fa50312a6e00170c14471 --- /dev/null +++ b/tests/test_utils.py @@ -0,0 +1,114 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import unittest + +import torch +from fairseq import utils + + +class TestUtils(unittest.TestCase): + def test_convert_padding_direction(self): + pad = 1 + left_pad = torch.LongTensor( + [ + [2, 3, 4, 5, 6], + [1, 7, 8, 9, 10], + [1, 1, 1, 11, 12], + ] + ) + right_pad = torch.LongTensor( + [ + [2, 3, 4, 5, 6], + [7, 8, 9, 10, 1], + [11, 12, 1, 1, 1], + ] + ) + + self.assertAlmostEqual( + right_pad, + utils.convert_padding_direction( + left_pad, + pad, + left_to_right=True, + ), + ) + self.assertAlmostEqual( + left_pad, + utils.convert_padding_direction( + right_pad, + pad, + right_to_left=True, + ), + ) + + def test_make_positions(self): + pad = 1 + left_pad_input = torch.LongTensor( + [ + [9, 9, 9, 9, 9], + [1, 9, 9, 9, 9], + [1, 1, 1, 9, 9], + ] + ) + left_pad_output = torch.LongTensor( + [ + [2, 3, 4, 5, 6], + [1, 2, 3, 4, 5], + [1, 1, 1, 2, 3], + ] + ) + right_pad_input = torch.LongTensor( + [ + [9, 9, 9, 9, 9], + [9, 9, 9, 9, 1], + [9, 9, 1, 1, 1], + ] + ) + right_pad_output = torch.LongTensor( + [ + [2, 3, 4, 5, 6], + [2, 3, 4, 5, 1], + [2, 3, 1, 1, 1], + ] + ) + + self.assertAlmostEqual( + left_pad_output, + utils.make_positions(left_pad_input, pad), + ) + self.assertAlmostEqual( + right_pad_output, + utils.make_positions(right_pad_input, pad), + ) + + def test_clip_grad_norm_(self): + params = torch.nn.Parameter(torch.zeros(5)).requires_grad_(False) + grad_norm = utils.clip_grad_norm_(params, 1.0) + self.assertTrue(torch.is_tensor(grad_norm)) + self.assertEqual(grad_norm, 0.0) + + params = [torch.nn.Parameter(torch.zeros(5)) for i in range(3)] + for p in params: + p.grad = torch.full((5,), fill_value=2.0) + grad_norm = utils.clip_grad_norm_(params, 1.0) + exp_grad_norm = torch.full((15,), fill_value=2.0).norm() + self.assertTrue(torch.is_tensor(grad_norm)) + self.assertEqual(grad_norm, exp_grad_norm) + + grad_norm = utils.clip_grad_norm_(params, 1.0) + self.assertAlmostEqual(grad_norm, torch.tensor(1.0)) + + def test_resolve_max_positions_with_tuple(self): + resolved = utils.resolve_max_positions(None, (2000, 100, 2000), 12000) + self.assertEqual(resolved, (2000, 100, 2000)) + + def assertAlmostEqual(self, t1, t2): + self.assertEqual(t1.size(), t2.size(), "size mismatch") + self.assertLess(utils.item((t1 - t2).abs().max()), 1e-4) + + +if __name__ == "__main__": + unittest.main() diff --git a/tests/test_valid_subset_checks.py b/tests/test_valid_subset_checks.py new file mode 100644 index 0000000000000000000000000000000000000000..3e9191bda66fccfebba34920f88bf7b1efea5f7e --- /dev/null +++ b/tests/test_valid_subset_checks.py @@ -0,0 +1,138 @@ +import os +import shutil +import tempfile +import unittest + +from fairseq import options +from fairseq.dataclass.utils import convert_namespace_to_omegaconf +from fairseq.data.data_utils import raise_if_valid_subsets_unintentionally_ignored +from .utils import create_dummy_data, preprocess_lm_data, train_language_model + + +def make_lm_config( + data_dir=None, + extra_flags=None, + task="language_modeling", + arch="transformer_lm_gpt2_tiny", +): + task_args = [task] + if data_dir is not None: + task_args += [data_dir] + train_parser = options.get_training_parser() + train_args = options.parse_args_and_arch( + train_parser, + [ + "--task", + *task_args, + "--arch", + arch, + "--optimizer", + "adam", + "--lr", + "0.0001", + "--max-tokens", + "500", + "--tokens-per-sample", + "500", + "--save-dir", + data_dir, + "--max-epoch", + "1", + ] + + (extra_flags or []), + ) + cfg = convert_namespace_to_omegaconf(train_args) + return cfg + + +def write_empty_file(path): + with open(path, "w"): + pass + assert os.path.exists(path) + + +class TestValidSubsetsErrors(unittest.TestCase): + """Test various filesystem, clarg combinations and ensure that error raising happens as expected""" + + def _test_case(self, paths, extra_flags): + with tempfile.TemporaryDirectory() as data_dir: + [ + write_empty_file(os.path.join(data_dir, f"{p}.bin")) + for p in paths + ["train"] + ] + cfg = make_lm_config(data_dir, extra_flags=extra_flags) + raise_if_valid_subsets_unintentionally_ignored(cfg) + + def test_default_raises(self): + with self.assertRaises(ValueError): + self._test_case(["valid", "valid1"], []) + with self.assertRaises(ValueError): + self._test_case( + ["valid", "valid1", "valid2"], ["--valid-subset", "valid,valid1"] + ) + + def partially_specified_valid_subsets(self): + with self.assertRaises(ValueError): + self._test_case( + ["valid", "valid1", "valid2"], ["--valid-subset", "valid,valid1"] + ) + # Fix with ignore unused + self._test_case( + ["valid", "valid1", "valid2"], + ["--valid-subset", "valid,valid1", "--ignore-unused-valid-subsets"], + ) + + def test_legal_configs(self): + self._test_case(["valid"], []) + self._test_case(["valid", "valid1"], ["--ignore-unused-valid-subsets"]) + self._test_case(["valid", "valid1"], ["--combine-val"]) + self._test_case(["valid", "valid1"], ["--valid-subset", "valid,valid1"]) + self._test_case(["valid", "valid1"], ["--valid-subset", "valid1"]) + self._test_case( + ["valid", "valid1"], ["--combine-val", "--ignore-unused-valid-subsets"] + ) + self._test_case( + ["valid1"], ["--valid-subset", "valid1"] + ) # valid.bin doesn't need to be ignored. + + def test_disable_validation(self): + self._test_case([], ["--disable-validation"]) + self._test_case(["valid", "valid1"], ["--disable-validation"]) + + def test_dummy_task(self): + cfg = make_lm_config(task="dummy_lm") + raise_if_valid_subsets_unintentionally_ignored(cfg) + + def test_masked_dummy_task(self): + cfg = make_lm_config(task="dummy_masked_lm") + raise_if_valid_subsets_unintentionally_ignored(cfg) + + +class TestCombineValidSubsets(unittest.TestCase): + def _train(self, extra_flags): + with self.assertLogs() as logs: + with tempfile.TemporaryDirectory("test_transformer_lm") as data_dir: + create_dummy_data(data_dir, num_examples=20) + preprocess_lm_data(data_dir) + + shutil.copyfile(f"{data_dir}/valid.bin", f"{data_dir}/valid1.bin") + shutil.copyfile(f"{data_dir}/valid.idx", f"{data_dir}/valid1.idx") + train_language_model( + data_dir, + "transformer_lm", + ["--max-update", "0", "--log-format", "json"] + extra_flags, + run_validation=False, + ) + return [x.message for x in logs.records] + + def test_combined(self): + flags = ["--combine-valid-subsets"] + logs = self._train(flags) + assert any(["valid1" in x for x in logs]) # loaded 100 examples from valid1 + assert not any(["valid1_ppl" in x for x in logs]) # metrics are combined + + def test_subsets(self): + flags = ["--valid-subset", "valid,valid1"] + logs = self._train(flags) + assert any(["valid_ppl" in x for x in logs]) # loaded 100 examples from valid1 + assert any(["valid1_ppl" in x for x in logs]) # metrics are combined diff --git a/tests/utils.py b/tests/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..6e0c709517aea570acb36901dd47bc12a3025b07 --- /dev/null +++ b/tests/utils.py @@ -0,0 +1,717 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import json +import os +import random +import sys +from io import StringIO + +import torch +import torch.nn.functional as F +from fairseq import options, utils +from fairseq.data import Dictionary +from fairseq.data.language_pair_dataset import collate +from fairseq.models import ( + FairseqEncoder, + FairseqEncoderDecoderModel, + FairseqIncrementalDecoder, +) +from fairseq.models.fairseq_encoder import EncoderOut +from fairseq.tasks import LegacyFairseqTask +from fairseq_cli import generate, interactive, preprocess, train, validate +import fairseq.distributed.utils as distributed_utils +from fairseq.dataclass.utils import convert_namespace_to_omegaconf + + +def dummy_dictionary(vocab_size, prefix="token_"): + d = Dictionary() + for i in range(vocab_size): + token = prefix + str(i) + d.add_symbol(token) + d.finalize(padding_factor=1) # don't add extra padding symbols + return d + + +def dummy_dataloader( + samples, padding_idx=1, eos_idx=2, batch_size=None, +): + if batch_size is None: + batch_size = len(samples) + + # add any missing data to samples + for i, sample in enumerate(samples): + if "id" not in sample: + sample["id"] = i + + # create dataloader + dataset = TestDataset(samples) + dataloader = torch.utils.data.DataLoader( + dataset, + batch_size=batch_size, + collate_fn=(lambda samples: collate(samples, padding_idx, eos_idx)), + ) + return iter(dataloader) + + +def sequence_generator_setup(): + # construct dummy dictionary + d = dummy_dictionary(vocab_size=2) + + eos = d.eos() + w1 = 4 + w2 = 5 + + # construct source data + src_tokens = torch.LongTensor([[w1, w2, eos], [w1, w2, eos]]) + src_lengths = torch.LongTensor([2, 2]) + + args = argparse.Namespace() + unk = 0.0 + args.beam_probs = [ + # step 0: + torch.FloatTensor( + [ + # eos w1 w2 + # sentence 1: + [0.0, unk, 0.9, 0.1], # beam 1 + [0.0, unk, 0.9, 0.1], # beam 2 + # sentence 2: + [0.0, unk, 0.7, 0.3], + [0.0, unk, 0.7, 0.3], + ] + ), + # step 1: + torch.FloatTensor( + [ + # eos w1 w2 prefix + # sentence 1: + [1.0, unk, 0.0, 0.0], # w1: 0.9 (emit: w1 <eos>: 0.9*1.0) + [0.0, unk, 0.9, 0.1], # w2: 0.1 + # sentence 2: + [0.25, unk, 0.35, 0.4], # w1: 0.7 (don't emit: w1 <eos>: 0.7*0.25) + [0.00, unk, 0.10, 0.9], # w2: 0.3 + ] + ), + # step 2: + torch.FloatTensor( + [ + # eos w1 w2 prefix + # sentence 1: + [0.0, unk, 0.1, 0.9], # w2 w1: 0.1*0.9 + [ + 0.6, + unk, + 0.2, + 0.2, + ], # w2 w2: 0.1*0.1 (emit: w2 w2 <eos>: 0.1*0.1*0.6) + # sentence 2: + [ + 0.60, + unk, + 0.4, + 0.00, + ], # w1 w2: 0.7*0.4 (emit: w1 w2 <eos>: 0.7*0.4*0.6) + [0.01, unk, 0.0, 0.99], # w2 w2: 0.3*0.9 + ] + ), + # step 3: + torch.FloatTensor( + [ + # eos w1 w2 prefix + # sentence 1: + [ + 1.0, + unk, + 0.0, + 0.0, + ], # w2 w1 w2: 0.1*0.9*0.9 (emit: w2 w1 w2 <eos>: 0.1*0.9*0.9*1.0) + [ + 1.0, + unk, + 0.0, + 0.0, + ], # w2 w1 w1: 0.1*0.9*0.1 (emit: w2 w1 w1 <eos>: 0.1*0.9*0.1*1.0) + # sentence 2: + [ + 0.1, + unk, + 0.5, + 0.4, + ], # w2 w2 w2: 0.3*0.9*0.99 (emit: w2 w2 w2 <eos>: 0.3*0.9*0.99*0.1) + [ + 1.0, + unk, + 0.0, + 0.0, + ], # w1 w2 w1: 0.7*0.4*0.4 (emit: w1 w2 w1 <eos>: 0.7*0.4*0.4*1.0) + ] + ), + ] + + task = TestTranslationTask.setup_task(args, d, d) + model = task.build_model(args) + tgt_dict = task.target_dictionary + + return tgt_dict, w1, w2, src_tokens, src_lengths, model + + +def create_dummy_data(data_dir, num_examples=100, maxlen=20, alignment=False): + def _create_dummy_data(filename): + data = torch.rand(num_examples * maxlen) + data = 97 + torch.floor(26 * data).int() + with open(os.path.join(data_dir, filename), "w") as h: + offset = 0 + for _ in range(num_examples): + ex_len = random.randint(1, maxlen) + ex_str = " ".join(map(chr, data[offset : offset + ex_len])) + print(ex_str, file=h) + offset += ex_len + + def _create_dummy_alignment_data(filename_src, filename_tgt, filename): + with open(os.path.join(data_dir, filename_src), "r") as src_f, open( + os.path.join(data_dir, filename_tgt), "r" + ) as tgt_f, open(os.path.join(data_dir, filename), "w") as h: + for src, tgt in zip(src_f, tgt_f): + src_len = len(src.split()) + tgt_len = len(tgt.split()) + avg_len = (src_len + tgt_len) // 2 + num_alignments = random.randint(avg_len // 2, 2 * avg_len) + src_indices = torch.floor(torch.rand(num_alignments) * src_len).int() + tgt_indices = torch.floor(torch.rand(num_alignments) * tgt_len).int() + ex_str = " ".join( + [ + "{}-{}".format(src, tgt) + for src, tgt in zip(src_indices, tgt_indices) + ] + ) + print(ex_str, file=h) + + _create_dummy_data("train.in") + _create_dummy_data("train.out") + _create_dummy_data("valid.in") + _create_dummy_data("valid.out") + _create_dummy_data("test.in") + _create_dummy_data("test.out") + + if alignment: + _create_dummy_alignment_data("train.in", "train.out", "train.align") + _create_dummy_alignment_data("valid.in", "valid.out", "valid.align") + _create_dummy_alignment_data("test.in", "test.out", "test.align") + + +def preprocess_lm_data(data_dir): + preprocess_parser = options.get_preprocessing_parser() + preprocess_args = preprocess_parser.parse_args( + [ + "--only-source", + "--trainpref", + os.path.join(data_dir, "train.out"), + "--validpref", + os.path.join(data_dir, "valid.out"), + "--testpref", + os.path.join(data_dir, "test.out"), + "--destdir", + data_dir, + ] + ) + preprocess.main(preprocess_args) + + +def preprocess_translation_data(data_dir, extra_flags=None): + preprocess_parser = options.get_preprocessing_parser() + preprocess_args = preprocess_parser.parse_args( + [ + "--source-lang", + "in", + "--target-lang", + "out", + "--trainpref", + os.path.join(data_dir, "train"), + "--validpref", + os.path.join(data_dir, "valid"), + "--testpref", + os.path.join(data_dir, "test"), + "--thresholdtgt", + "0", + "--thresholdsrc", + "0", + "--destdir", + data_dir, + ] + + (extra_flags or []), + ) + preprocess.main(preprocess_args) + + +def preprocess_summarization_data(data_dir, extra_flags=None): + preprocess_parser = options.get_preprocessing_parser() + preprocess_args = preprocess_parser.parse_args( + [ + "--source-lang", + "in", + "--target-lang", + "out", + "--trainpref", + os.path.join(data_dir, "train"), + "--validpref", + os.path.join(data_dir, "valid"), + "--testpref", + os.path.join(data_dir, "test"), + "--thresholdtgt", + "0", + "--thresholdsrc", + "0", + "--joined-dictionary", + "--destdir", + data_dir, + ] + + (extra_flags or []), + ) + preprocess.main(preprocess_args) + + +def create_laser_data_and_config_json(data_dir): + src_langs = ["de", "fr", "ru", "tr", "zh"] + tgt_langs = ["en", "es"] + config_json = {} + config_train_json = [] + src_vocab = None + tgt_vocab = None + + for src_lang in src_langs: + for tgt_lang in tgt_langs: + langpair_folder = f"{src_lang}-{tgt_lang}" + + langpair_path = os.path.join(data_dir, langpair_folder) + os.mkdir(langpair_path) + create_dummy_data(langpair_path) + preprocess_translation_data(langpair_path, ["--dataset-impl", "cached"]) + + src_vocab = os.path.join(langpair_path, "dict.in.txt") + tgt_vocab = os.path.join(langpair_path, "dict.out.txt") + config_train_json.append( + { + "id": 0 if tgt_lang == "en" else 1, + "src": os.path.join(langpair_path, "train.in-out.in"), + "tgt": os.path.join(langpair_path, "train.in-out.out"), + } + ) + + config_json["src_vocab"] = src_vocab + config_json["tgt_vocab"] = tgt_vocab + config_json["train"] = config_train_json + + with open(os.path.join(data_dir, "laserconfig.json"), "w") as config_file: + json.dump(config_json, config_file) + + return config_file + + +def train_translation_model( + data_dir, + arch, + extra_flags=None, + task="translation", + run_validation=False, + lang_flags=None, + extra_valid_flags=None, + world_size=1, +): + if lang_flags is None: + lang_flags = [ + "--source-lang", + "in", + "--target-lang", + "out", + ] + train_parser = options.get_training_parser() + train_args = options.parse_args_and_arch( + train_parser, + [ + "--task", + task, + data_dir, + "--save-dir", + data_dir, + "--arch", + arch, + "--optimizer", + "nag", + "--lr", + "0.05", + "--max-tokens", + "500", + "--max-epoch", + "1", + "--no-progress-bar", + "--distributed-world-size", + str(world_size), + "--num-workers", + "0", + ] + + lang_flags + + (extra_flags or []), + ) + + cfg = convert_namespace_to_omegaconf(train_args) + distributed_utils.call_main(cfg, train.main) + + if run_validation: + # test validation + validate_parser = options.get_validation_parser() + validate_args = options.parse_args_and_arch( + validate_parser, + [ + "--task", + task, + data_dir, + "--path", + os.path.join(data_dir, "checkpoint_last.pt"), + "--valid-subset", + "valid", + "--max-tokens", + "500", + "--no-progress-bar", + "--num-workers", + "0", + ] + + lang_flags + + (extra_valid_flags or []), + ) + validate.main(validate_args) + + +def generate_main(data_dir, extra_flags=None, path=None): + if extra_flags is None: + extra_flags = [ + "--print-alignment", + ] + if path is None: + path = os.path.join(data_dir, "checkpoint_last.pt") + generate_parser = options.get_generation_parser() + generate_args = options.parse_args_and_arch( + generate_parser, + [ + data_dir, + "--path", + path, + "--beam", + "3", + "--batch-size", + "64", + "--max-len-b", + "5", + "--gen-subset", + "valid", + "--no-progress-bar", + "--num-workers", + "0", + ] + + (extra_flags or []), + ) + + # evaluate model in batch mode + generate.main(generate_args) + + # evaluate model interactively + generate_args.buffer_size = 0 + generate_args.input = "-" + generate_args.batch_size = None + orig_stdin = sys.stdin + sys.stdin = StringIO("h e l l o\n") + interactive.main(generate_args) + sys.stdin = orig_stdin + + +class TestDataset(torch.utils.data.Dataset): + def __init__(self, data): + super().__init__() + self.data = data + self.sizes = None + + def __getitem__(self, index): + return self.data[index] + + def __len__(self): + return len(self.data) + + +class TestTranslationTask(LegacyFairseqTask): + def __init__(self, args, src_dict, tgt_dict, model): + super().__init__(args) + self.src_dict = src_dict + self.tgt_dict = tgt_dict + self.model = model + + @classmethod + def setup_task(cls, args, src_dict=None, tgt_dict=None, model=None): + return cls(args, src_dict, tgt_dict, model) + + def build_model(self, args): + return TestModel.build_model(args, self) + + @property + def source_dictionary(self): + return self.src_dict + + @property + def target_dictionary(self): + return self.tgt_dict + + +class TestModel(FairseqEncoderDecoderModel): + def __init__(self, encoder, decoder): + super().__init__(encoder, decoder) + + @classmethod + def build_model(cls, args, task): + encoder = TestEncoder(args, task.source_dictionary) + decoder = TestIncrementalDecoder(args, task.target_dictionary) + return cls(encoder, decoder) + + +class TestEncoder(FairseqEncoder): + def __init__(self, args, dictionary): + super().__init__(dictionary) + self.args = args + + def forward(self, src_tokens, src_lengths=None, **kwargs): + return EncoderOut( + encoder_out=src_tokens, + encoder_padding_mask=None, + encoder_embedding=None, + encoder_states=None, + src_tokens=None, + src_lengths=None, + ) + + def reorder_encoder_out(self, encoder_out, new_order): + return EncoderOut( + encoder_out=encoder_out.encoder_out.index_select(0, new_order), + encoder_padding_mask=None, + encoder_embedding=None, + encoder_states=None, + src_tokens=None, + src_lengths=None, + ) + + +class TestIncrementalDecoder(FairseqIncrementalDecoder): + def __init__(self, args, dictionary): + super().__init__(dictionary) + assert hasattr(args, "beam_probs") or hasattr(args, "probs") + args.max_decoder_positions = getattr(args, "max_decoder_positions", 100) + self.args = args + + def forward(self, prev_output_tokens, encoder_out=None, incremental_state=None): + if incremental_state is not None: + prev_output_tokens = prev_output_tokens[:, -1:] + bbsz = prev_output_tokens.size(0) + vocab = len(self.dictionary) + src_len = encoder_out.encoder_out.size(1) + tgt_len = prev_output_tokens.size(1) + + # determine number of steps + if incremental_state is not None: + # cache step number + step = utils.get_incremental_state(self, incremental_state, "step") + if step is None: + step = 0 + utils.set_incremental_state(self, incremental_state, "step", step + 1) + steps = [step] + else: + steps = list(range(tgt_len)) + + # define output in terms of raw probs + if hasattr(self.args, "probs"): + assert ( + self.args.probs.dim() == 3 + ), "expected probs to have size bsz*steps*vocab" + probs = self.args.probs.index_select(1, torch.LongTensor(steps)) + else: + probs = torch.FloatTensor(bbsz, len(steps), vocab).zero_() + for i, step in enumerate(steps): + # args.beam_probs gives the probability for every vocab element, + # starting with eos, then unknown, and then the rest of the vocab + if step < len(self.args.beam_probs): + probs[:, i, self.dictionary.eos() :] = self.args.beam_probs[step] + else: + probs[:, i, self.dictionary.eos()] = 1.0 + + # random attention + attn = torch.rand(bbsz, tgt_len, src_len) + + dev = prev_output_tokens.device + return probs.to(dev), {"attn": [attn.to(dev)]} + + def get_normalized_probs(self, net_output, log_probs, _): + # the decoder returns probabilities directly + probs = net_output[0] + if log_probs: + return probs.log() + else: + return probs + + def max_positions(self): + return self.args.max_decoder_positions + + +class TestReshapingEncoder(FairseqEncoder): + def __init__(self, args, dictionary): + super().__init__(dictionary) + self.args = args + + def forward(self, src_tokens, src_lengths=None, **kwargs): + b_sz, t_sz = src_tokens.shape + padding_needed = t_sz % 2 + x = src_tokens + if padding_needed > 0: + padding_needed = 2 - padding_needed + x = F.pad(x, (0, padding_needed)) + + return EncoderOut( + encoder_out=x.view(b_sz, -1, 2), + encoder_padding_mask=None, + encoder_embedding=None, + encoder_states=None, + src_tokens=None, + src_lengths=None, + ) + + def reorder_encoder_out(self, encoder_out, new_order): + return EncoderOut( + encoder_out=encoder_out.encoder_out.index_select(0, new_order), + encoder_padding_mask=None, + encoder_embedding=None, + encoder_states=None, + src_tokens=None, + src_lengths=None, + ) + + +class TestReshapingModel(FairseqEncoderDecoderModel): + def __init__(self, encoder, decoder): + super().__init__(encoder, decoder) + + @classmethod + def build_model(cls, args, task): + encoder = TestReshapingEncoder(args, task.source_dictionary) + decoder = TestIncrementalDecoder(args, task.target_dictionary) + return cls(encoder, decoder) + + +class TestAdditionalInputEncoder(FairseqEncoder): + def __init__(self, args, dictionary): + super().__init__(dictionary) + self.args = args + + def forward(self, src_tokens, src_lengths=None, **kwargs): + assert "fancy_other_input" in kwargs + assert kwargs["fancy_other_input"] is not None + return EncoderOut( + encoder_out=src_tokens, + encoder_padding_mask=None, + encoder_embedding=None, + encoder_states=None, + src_tokens=None, + src_lengths=None, + ) + + def reorder_encoder_out(self, encoder_out, new_order): + return EncoderOut( + encoder_out=encoder_out.encoder_out.index_select(0, new_order), + encoder_padding_mask=None, + encoder_embedding=None, + encoder_states=None, + src_tokens=None, + src_lengths=None, + ) + + +class TestAdditionalInputModel(FairseqEncoderDecoderModel): + def __init__(self, encoder, decoder): + super().__init__(encoder, decoder) + + @classmethod + def build_model(cls, args, task): + encoder = TestAdditionalInputEncoder(args, task.source_dictionary) + decoder = TestIncrementalDecoder(args, task.target_dictionary) + return cls(encoder, decoder) + + def forward(self, src_tokens, src_lengths, prev_output_tokens, **kwargs): + encoder_out = self.encoder(src_tokens, src_lengths=src_lengths, **kwargs) + decoder_out = self.decoder( + prev_output_tokens, encoder_out=encoder_out, **kwargs + ) + return decoder_out + + +def train_language_model( + data_dir, + arch, + extra_flags=None, + run_validation=False, + extra_valid_flags=None, + task="language_modeling", + world_size=1, +): + train_parser = options.get_training_parser() + train_args = options.parse_args_and_arch( + train_parser, + [ + "--task", + task, + data_dir, + "--arch", + arch, + "--optimizer", + "adam", + "--lr", + "0.0001", + "--max-tokens", + "500", + "--tokens-per-sample", + "500", + "--save-dir", + data_dir, + "--max-epoch", + "1", + "--no-progress-bar", + "--distributed-world-size", + str(world_size), + "--ddp-backend", + "no_c10d", + "--num-workers", + "0", + ] + + (extra_flags or []), + ) + cfg = convert_namespace_to_omegaconf(train_args) + distributed_utils.call_main(cfg, train.main) + + if run_validation: + # test validation + validate_parser = options.get_validation_parser() + validate_args = options.parse_args_and_arch( + validate_parser, + [ + "--task", + task, + data_dir, + "--path", + os.path.join(data_dir, "checkpoint_last.pt"), + "--valid-subset", + "valid", + "--max-tokens", + "500", + "--no-progress-bar", + "--num-workers", + "0", + ] + + (extra_valid_flags or []), + ) + validate.main(validate_args) diff --git a/train.py b/train.py new file mode 100644 index 0000000000000000000000000000000000000000..321de3d9b53f8194b58c26f5cb2c03281afc2bb1 --- /dev/null +++ b/train.py @@ -0,0 +1,14 @@ +#!/usr/bin/env python3 -u +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +Legacy entry point. Use fairseq_cli/train.py or fairseq-train instead. +""" + +from fairseq_cli.train import cli_main + + +if __name__ == "__main__": + cli_main()