wavenet

WaveNet minimal reimplementation

About this repo

This reimplementation uses tensorflow and keras. I've added a synthetic dataset which serves as a way to quickly demonstrate the training convergence. The data consists of either sine, square or sawtooth waveforms. The shape of the waveform can be included as global context by providing a global_cond when invoking the training script.

Code structure

• train.py: contains the code to setup the training
• wavenet.py: contains the model
• datasets.py: contains a SimpleWaveForm dataset that generates sines, sawtooths and square waveforms.
• util.py: some non-relevant utilities

Usage

The following invocation will at least produce some nice TensorBoard visualizations within a few epochs:

python train.py --dilation_stacks 2 --dilation_pow2 5 --train_size 1000 --test_size 100 --sequence_len 10000 --global_cond