philgzl / wavenet

A PyTorch implementation of DeepMind's WaveNet

wavenet

A PyTorch implementation of DeepMind's WaveNet

Installation

1. Clone and change directory

git clone https://github.com/philgzl/wavenet.git
cd wavenet

2. Create a virtual environment and activate

python -m venv venv
source venv/bin/activate

3. Install requirements

pip install -r requirements.txt

4. Install project

python setup.py develop

How to use

Initializing a model

Models are initialized with the scripts/init_model.py script. This creates a new directory under models/ containing the configuration file for the model. The model directory is named after a unique ID generated from the configuration file.

The script takes as arguments the hyperparameters for the model. You can use python scripts/init_model.py --help for a list of available hyperparameters.

Example:

python scripts/init_model.py --layers 10

Training a model

To train an initialized model, simply call the scripts/train.py script with the model path as argument:

python scripts/train.py models/{model_id}/

A checkpoint file checkpoint.pt will be created in the model directory, next to the configuration file.

Submitting jobs to GPU cluster

If you are on a server equipped with GPUs and able to send jobs through the bsub command, you can submit a training job using:

bash jobs/train.sh models/{model_id}/

Logs will be stored under jobs/logs/.

To do

• Evaluation script
• Global/local conditioning
• Speaker selection