Spaces:
Running
A newer version of the Gradio SDK is available:
5.4.0
Jets Recipe
In this recipe, we will show how to train Jets using Amphion's infrastructure. Jets is an end-to-end text-to-speech (E2E-TTS) model which jointly trains FastSpeech2 and HiFi-GAN.
There are four stages in total:
- Data preparation
- Features extraction
- Training
- Inference
NOTE: You need to run every command of this recipe in the
Amphion
root path:
cd Amphion
1. Data Preparation
Dataset Download
You can use LJSpeech to train TTS model. How to download dataset is detailed here.
Configuration
After downloading the dataset, you can set the dataset paths in exp_config.json
. Note that you can change the dataset
list to use your preferred datasets.
"dataset": [
"LJSpeech",
],
"dataset_path": {
// TODO: Fill in your dataset path
"LJSpeech": "[LJSpeech dataset path]",
},
2. Features Extraction
Configuration
Specify the processed_dir
and the log_dir
and for saving the processed data and the checkpoints in exp_config.json
:
// TODO: Fill in the output log path
"log_dir": "ckpts/tts",
"preprocess": {
// TODO: Fill in the output data path
"processed_dir": "data",
...
},
Run
Run the run.sh
as the preproces stage (set --stage 1
):
sh egs/tts/Jets/run.sh --stage 1
3. Training
Configuration
We provide the default hyparameters in the exp_config.json
. They can work on single NVIDIA-24g GPU. You can adjust them based on your GPU machines.
"train": {
"batch_size": 16,
}
Run
Run the run.sh
as the training stage (set --stage 2
). Specify a experimental name to run the following command. The tensorboard logs and checkpoints will be saved in ckpts/tts/[YourExptName]
.
sh egs/tts/Jets/run.sh --stage 2 --name [YourExptName]
NOTE: The
CUDA_VISIBLE_DEVICES
is set as"0"
in default. We recommend you to only use one GPU for training.
4. Inference
Configuration
For inference, you need to specify the following configurations when running run.sh
:
Parameters | Description | Example |
---|---|---|
--infer_expt_dir |
The experimental directory which contains checkpoint |
ckpts/tts/[YourExptName] |
--infer_output_dir |
The output directory to save inferred audios. | ckpts/tts/[YourExptName]/result |
--infer_mode |
The inference mode, e.g., "batch ". |
batch " to generate a batch of speech at a time. |
--infer_dataset |
The dataset used for inference. | For LJSpeech dataset, the inference dataset would be LJSpeech . |
--infer_testing_set |
The subset of the inference dataset used for inference, e.g., test | For LJSpeech dataset, the testing set would be "test " split from LJSpeech at the feature extraction |
Run
For example, if you want to generate speech of all testing set split from LJSpeech, just run:
sh egs/tts/Jets/run.sh --stage 3 \
--infer_expt_dir ckpts/tts/[YourExptName] \
--infer_output_dir ckpts/tts/[YourExptName]/result \
--infer_mode "batch" \
--infer_dataset "LJSpeech" \
--infer_testing_set "test"
ISSUES and Solutions
NotImplementedError: Using RTX 3090 or 4000 series doesn't support faster communication broadband via P2P or IB. Please set `NCCL_P2P_DISABLE="1"` and `NCCL_IB_DISABLE="1" or use `accelerate launch` which will do this automatically.
2024-02-24 10:57:49 | INFO | torch.distributed.distributed_c10d | Rank 0: Completed store-based barrier for key:store_based_barrier_key:1 with 2 nodes.
The error message is related to an incompatibility issue with the NVIDIA RTX 3090 or 4000 series GPUs when trying to use peer-to-peer (P2P) communication or InfiniBand (IB) for faster communication. This incompatibility arises within the PyTorch accelerate library, which facilitates distributed training and inference.
To fix this issue, before running your script, you can set the environment variables in your terminal:
export NCCL_P2P_DISABLE=1
export NCCL_IB_DISABLE=1
Noted
Extensive logging messages related to torch._subclasses.fake_tensor
and torch._dynamo.output_graph
may be observed during inference. Despite attempts to ignore these logs, no effective solution has been found. However, it does not impact the inference process.
@article{lim2022jets,
title={JETS: Jointly training FastSpeech2 and HiFi-GAN for end to end text to speech},
author={Lim, Dan and Jung, Sunghee and Kim, Eesung},
journal={arXiv preprint arXiv:2203.16852},
year={2022}
}