wavenet

This is a Python module for WaveNet in TensorFlow. It implements both a traditional convolutional API, and an RNNCell for fast stepping.

Usage

You can install the package with pip like so:

$ pip install wavenet

The example is a complete program which trains a WaveNet on synthetic data from the macOS speech synthesizer. In this README, we only show the basics of applying WaveNets to sequences.

Once you have the wavenet package, it's trivial to construct and use a WaveNet. Here's an example:

import tensorflow as tf
from wavenet import Conv, Network

# Produce a [batch x timesteps x depth] input sequence.
# Only `depth` needs to be known ahead of time.
inputs = ...

# Create a new network and all its variables.
# In this case, the receptive field is 16.
num_channels = inputs.get_shape()[-1].value
network = Network([Conv(channels=num_channels, dilation=2**i) for i in range(4)])

# Apply the model to the inputs, yielding outputs for
# every timestep of the input.
outputs = network.apply(inputs)

If you need to step the model efficiently, or if you want to use a WaveNet in place of another recurrent neural network, you can get an RNNCell directly from a Network by calling network.cell(). In the above example, you could replace network.apply(inputs) with this code that uses an RNNCell:

rnn_cell = network.cell()
outputs, _ = tf.nn.dynamic_rnn(rnn_cell, inputs, dtype=tf.float32)