Method _create_network

wavenet/model.py:395–442

Construct the WaveNet network.

(self, input_batch, global_condition_batch)

        return skip_contribution, input_batch + transformed

    def _create_network(self, input_batch, global_condition_batch):
        '''Construct the WaveNet network.'''
        outputs = []
        current_layer = input_batch

        # Pre-process the input with a regular convolution
        current_layer = self._create_causal_layer(current_layer)

        output_width = tf.shape(input_batch)[1] - self.receptive_field + 1

        # Add all defined dilation layers.
        with tf.name_scope('dilated_stack'):
            for layer_index, dilation in enumerate(self.dilations):
                with tf.name_scope('layer{}'.format(layer_index)):
                    output, current_layer = self._create_dilation_layer(
                        current_layer, layer_index, dilation,
                        global_condition_batch, output_width)
                    outputs.append(output)

        with tf.name_scope('postprocessing'):
            # Perform (+) -> ReLU -> 1x1 conv -> ReLU -> 1x1 conv to
            # postprocess the output.
            w1 = self.variables['postprocessing']['postprocess1']
            w2 = self.variables['postprocessing']['postprocess2']
            if self.use_biases:
                b1 = self.variables['postprocessing']['postprocess1_bias']
                b2 = self.variables['postprocessing']['postprocess2_bias']

            if self.histograms:
                tf.histogram_summary('postprocess1_weights', w1)
                tf.histogram_summary('postprocess2_weights', w2)
                if self.use_biases:
                    tf.histogram_summary('postprocess1_biases', b1)
                    tf.histogram_summary('postprocess2_biases', b2)

            # We skip connections from the outputs of each layer, adding them
            # all up here.
            total = sum(outputs)
            transformed1 = tf.nn.relu(total)
            conv1 = tf.nn.conv1d(transformed1, w1, stride=1, padding="SAME")
            if self.use_biases:
                conv1 = tf.add(conv1, b1)
            transformed2 = tf.nn.relu(conv1)
            conv2 = tf.nn.conv1d(transformed2, w2, stride=1, padding="SAME")
            if self.use_biases:
                conv2 = tf.add(conv2, b2)

        return conv2

    def _create_generator(self, input_batch, global_condition_batch):
        '''Construct an efficient incremental generator.'''