search

yukara-ikemiya / friendly-stable-audio-tools

Refactored / updated version of `stable-audio-tools` which is an open-source code for audio/music generative models originally by Stability AI.
MIT License
136 stars 10 forks source link

readme

🐈 friendly-stable-audio-tools

Stable Audio 2.0

This repository is a refactored / updated version of stable-audio-tools which is an open-source code for audio/music generative models originally by Stability AI.

This repository contains the following additional features:

🔥 Stable Audio Open

Stability AI has now open-sourced the pre-trained model for Stable Audio.

For detailed instructions on how to use Stable Audio Open, please refer to this document if you are interested.

Requirements

Install

To run the training scripts or inference code, you'll need to clone this repository, navigate to the root directory, and then execute the pip command as follow:

$ git clone https://github.com/yukara-ikemiya/friendly-stable-audio-tools.git
$ cd friendly-stable-audio-tools
$ pip install .
$ # you may need to execute this to avoid Accelerate import error
$ pip uninstall -y transformer-engine

Building a training environment

To simplify setting up the training environment, I recommend to use container systems like Docker or Singularity instead of installing dependencies on each GPU machine. Below are the steps for creating Docker and Singularity containers.

All example scripts are stored at the container folder.

Please be sure that Docker and Singularity are installed in advance.

1. Create a Docker image

$ # create a Docker image
$ NAME=friendly-stable-audio-tools
$ docker build  -t ${NAME} -f ./container/${NAME}.Dockerfile .

2. Convert a Docker image to a Singularity container

$ # convert a Docker image to a Singularity container
$ singularity build friendly-stable-audio-tools.sif docker-daemon://friendly-stable-audio-tools

By running the above script, friendly-stable-audio-tools.sif should be created in the working directory.

Interface

A basic Gradio interface is provided to test out trained models.

For example, to create an interface for the stable-audio-open-1.0 model, once you've accepted the terms for the model on Hugging Face, you can run:

$ python3 ./run_gradio.py --pretrained-name stabilityai/stable-audio-open-1.0

If you need more detailed instruction on Stable Audio Open, I recommend referring to the Gradio interface part of the Stable Audio Open documentation.

The run_gradio.py script accepts the following command line arguments:

Logging

WandB setting

The training code also requires a Weights & Biases account to log the training outputs and demos. Create an account and log in with:

$ wandb login

Or you can also pass an API key as an environment variable WANDB_API_KEY. (You can obtain the API key from https://wandb.ai/authorize after logging in to your account.)

$ WANDB_API_KEY="12345x6789y..."

This method is convenient when you want to execute the code using containers such as Docker or Singularity.

Training

Configuration files

Before starting your training run, you have to prepare the following two configuration files.

For more information about those, refer to the Configuration section below.

Training from scratch

To start a training run, run the train.py script in the repo root with:

$ python3 train.py --dataset-config /path/to/dataset/config --model-config /path/to/model/config --name my_experiment

The --name parameter will set the project name for your Weights and Biases run.

Fine-tuning

Fine-tuning involves resuming a training run from a pre-trained checkpoint.

Unwrapping a model

stable-audio-tools uses PyTorch Lightning to facilitate multi-GPU and multi-node training.

When a model is being trained, it is wrapped in a "training wrapper", which is a pl.LightningModule that contains all of the relevant objects needed only for training. That includes things like discriminators for autoencoders, EMA copies of models, and all of the optimizer states.

The checkpoint files created during training include this training wrapper, which greatly increases the size of the checkpoint file.

unwrap_model.py takes in a wrapped model checkpoint and save a new checkpoint file including only the model itself.

That can be run with from the repo root with:

$ python3 unwrap_model.py --model-config /path/to/model/config --ckpt-path /path/to/wrapped/ckpt.ckpt --name /new/path/to/new_ckpt_name

Unwrapped model checkpoints are required for:

Configurations

Training and inference code for stable-audio-tools is based around JSON configuration files that define model hyperparameters, training settings, and information about your training dataset.

Model config

The model config file defines all of the information needed to load a model for training or inference. It also contains the training configuration needed to fine-tune a model or train from scratch.

The following properties are defined in the top level of the model configuration:

Dataset config

stable-audio-tools currently supports two kinds of data sources: local directories of audio files, and WebDataset datasets stored in Amazon S3. More information can be found in the dataset config documentation

Additional training flags

Additional optional flags for train.py include:

🔥 Let's train Stable Audio 2.0

Prerequisites

Prepare a checkpoint of CLAP encoder

To use CLAP encoder for conditioning music generation, you have to prepare a pretrained checkpoint file of CLAP.

  1. Download a pretrained CLAP checkpoint trained with music dataset (music_audioset_epoch_15_esc_90.14.pt) from the LAION CLAP repository.
  2. Store the checkpoint file to a directory of your choice.
  3. Edit a model config file of Stable Audio 2.0 as follows

= stable_audio_2_0.json =

...
"model": {
  ...
  "conditioning": {
            "configs": [
                {
                    ...
                    "config": {
                        ...
                        "clap_ckpt_path": "ckpt/clap/music_audioset_epoch_15_esc_90.14.pt",
                    ...

Prepare audio and metadata for training

Since Stable Audio uses text prompts as condition for music generation, you have to prepare them as metadata in addition to audio data.

When using a dataset in a local environment, I support the use of metadata in JSON format as follows.

  1. You can include any information as metadata in a JSON file, but you must always include the text data named prompt required for training of Stable Audio.

= music_2.json =

{
    "prompt": "This is an electronic song sending positive vibes."
}
  1. The metadata files must be placed in the same directory as corresponding audio files. And the file names must also be the same.
.
└── dataset/
    ├── music_1.wav
    ├── music_1.json
    ├── music_2.wav
    ├── music_2.json
    └── ...

Stage 1 : VAE-GAN (compression model)

Training

As the 1st stage of Stable Audio 2.0, you'll train a VAE-GAN which is a compression model for audio signal.

The model config file for a VAE-GAN is place in the configs directory. Regarding dataset configuration, please prepare a dataset config file corresponding to your own datasets.

Once you prepare configuration files, you can execute a training job like this:

CONTAINER_PATH="/path/to/sif/friendly-stable-audio-tools.sif"
ROOT_DIR="/path/to/friendly-stable-audio-tools/"
DATASET_DIR="/path/to/your/dataset/"
OUTPUT_DIR="/path/to/output/directory/"

MODEL_CONFIG="stable_audio_tools/configs/model_configs/autoencoders/stable_audio_2_0_vae.json"
DATASET_CONFIG="stable_audio_tools/configs/dataset_configs/local_training_example.json"

BATCH_SIZE=10 # WARNING : This is batch size per GPU
WANDB_API_KEY="12345x6789y..."
PORT=12345

# Singularity container case
# NOTE: Please change each configuration as you like

singularity exec --nv --pwd $ROOT_DIR -B $ROOT_DIR -B $DATASET_DIR \
  --env WANDB_API_KEY=$WANDB_API_KEY \
  ${CONTAINER_PATH} \
  torchrun --nproc_per_node gpu --master_port ${PORT} \
  ${ROOT_DIR}/train.py \
    --dataset-config ${DATASET_CONFIG} \
    --model-config ${MODEL_CONFIG} \
    --name "vae_training" \
    --num-gpus 8 \
    --batch-size ${BATCH_SIZE} \
    --num-workers 8 \
    --save-dir ${OUTPUT_DIR}

Model unwrapping

As described in the unwrapping-a-model section, after completing the training of VAE, you need to unwrap the model checkpoint for using the next stage training.

CKPT_PATH="/path/to/wrapped_ckpt/last.ckpt"
# NOTE: file extension ".ckpt" will be automatically added to the end of OUTPOUT_DIR name
OUTPUT_PATH="/path/to/output_name/unwrapped_last"

singularity exec --nv --pwd $ROOT_DIR -B $ROOT_DIR \
  --env WANDB_API_KEY=$WANDB_API_KEY \
  ${CONTAINER_PATH} \
  torchrun --nproc_per_node gpu --master_port ${PORT} \
    ${ROOT_DIR}/unwrap_model.py \
    --model-config ${MODEL_CONFIG} \
    --ckpt-path ${CKPT_PATH} \
    --name ${OUTPUT_PATH}

Reconstruction test

Once you finished the VAE training, you might want to test and evaluate reconstruction quality of the trained model.

I support reconstruction of audio files in a directory with reconstruct_audios.py, and you can use the reconstructed audios for your evaluation.

AUDIO_DIR="/path/to/original_audio/"
OUTPUT_DIR="/path/to/output_audio/"

FRAME_DURATION=1.0 # [sec]
OVERLAP_RATE=0.01
BATCH_SIZE=50

singularity exec --nv --pwd $ROOT_DIR -B $ROOT_DIR -B $DATASET_DIR \
  --env WANDB_API_KEY=$WANDB_API_KEY \
  ${CONTAINER_PATH} \
  torchrun --nproc_per_node gpu --master_port ${PORT} \
    ${ROOT_DIR}/reconstruct_audios.py \
    --model-config ${MODEL_CONFIG} \
    --ckpt-path ${UNWRAP_CKPT_PATH} \
    --audio-dir ${AUDIO_DIR} \
    --output-dir ${OUTPUT_DIR} \
    --frame-duration ${FRAME_DURATION} \
    --overlap-rate ${OVERLAP_RATE} \
    --batch-size ${BATCH_SIZE}

Stage 2 : Diffusion Transformer (DiT)

Training

As the 2nd stage of Stable Audio 2.0, you'll train a DiT which is a generative model in latent domain.

Before this part, please make sure that

Now, you can train a DiT model as follows:

CONTAINER_PATH="/path/to/sif/friendly-stable-audio-tools.sif"
ROOT_DIR="/path/to/friendly-stable-audio-tools/"
DATASET_DIR="/path/to/your/dataset/"
OUTPUT_DIR="/path/to/output/directory/"

MODEL_CONFIG="stable_audio_tools/configs/model_configs/txt2audio/stable_audio_2_0.json"
DATASET_CONFIG="stable_audio_tools/configs/dataset_configs/local_training_example.json"

# Pretrained checkpoint of VAE (Stage-1) model
PRETRANSFORM_CKPT="/path/to/vae_ckpt/unwrapped_last.ckpt"

BATCH_SIZE=10 # WARNING : This is batch size per GPU
WANDB_API_KEY="12345x6789y..."
PORT=12345

singularity exec --nv --pwd $ROOT_DIR -B $ROOT_DIR -B $DATASET_DIR \
  --env WANDB_API_KEY=$WANDB_API_KEY \
  ${CONTAINER_PATH} \
  torchrun --nproc_per_node gpu --master_port ${PORT} \
    ${ROOT_DIR}/train.py \
    --dataset-config ${DATASET_CONFIG} \
    --model-config ${MODEL_CONFIG} \
    --pretransform-ckpt-path ${PRETRANSFORM_CKPT} \
    --name "dit_training" \
    --num-gpus ${NUM_GPUS} \
    --batch-size ${BATCH_SIZE} \
    --save-dir ${OUTPUT_DIR}

Todo