full working demo

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 <[email protected]>. 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
+

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: <https://code.facebook.com/cla>
+
+## 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.

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.

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 = "<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()

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)

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;
+}

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

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),
+}

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:

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:

docs/docutils.conf

@@ -0,0 +1,2 @@
+[writers]
+option-limit=0

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 <https://arxiv.org/abs/1508.07909>`__, 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 <https://github.com/rsennrich/subword-nmt/blob/master/subword_nmt/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 <https://github.com/moses-smt/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 <https://github.com/pytorch/fairseq#pre-trained-models>`__ 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) <https://arxiv.org/abs/1806.00187>`__ 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
+<https://docs.nvidia.com/deeplearning/sdk/mixed-precision-training/index.html>`__.
+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
+<https://pytorch.org/docs/stable/distributed.html#launch-utility>`__ 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.

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.

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
+<http://pytorch.org/>`_ 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`

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:

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

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:

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:

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:

docs/overview.rst

@@ -0,0 +1,74 @@
+Overview
+========
+
+Fairseq can be extended through user-supplied `plug-ins
+<https://en.wikipedia.org/wiki/Plug-in_(computing)>`_. 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

docs/requirements.txt

@@ -0,0 +1,2 @@
+sphinx<2.0
+sphinx-argparse

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:

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 <https://pytorch.org/tutorials/intermediate/char_rnn_classification_tutorial.html>`_
+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 <https://dl.fbaipublicfiles.com/fairseq/data/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
+<https://pytorch.org/tutorials/intermediate/char_rnn_classification_tutorial.html#creating-the-network>`_.
+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

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)

examples/.gitignore

@@ -0,0 +1,2 @@
+!*/*.sh
+!*/*.md

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

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).

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)

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

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()

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

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)

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)

examples/adaptive_span/truncated_bptt_lm_task.py

@@ -0,0 +1 @@
+../truncated_bptt/truncated_bptt_lm_task.py

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 <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) | [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/wmt18.en-de.ensemble.tar.gz) <br> 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,
+}
+```

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()

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()

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

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 '<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

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

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

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>
+```
+`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))
+```

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 `<mask>` tokens in the input.
+```python
+bart = torch.hub.load('pytorch/fairseq', 'bart.base')
+bart.eval()
+bart.fill_mask(['The cat <mask> on the <mask>.'], 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 <mask> on the <mask>.'], 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 <mask> on the <mask>.', 'The dog <mask> on the <mask>.'], 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},
+}
+```

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
+```

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()

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 ([[email protected]](mailto:[email protected])),
+Kyunghyun Cho ([[email protected]](mailto:[email protected])),
+Jiatao Gu ([[email protected]](mailto:[email protected]))