f073f90e639afa1ee2141dbed5d76e26ed31d5de,texar/modules/encoders/conv1d_discriminator.py,CNN,_build,#CNN#,54

Before Change

             / tf.cast(tf.shape(inputs)[1], tf.float32)
    if self._hparams.use_gate:
      proj = tf.tanh(self._gate_proj(inputs))
      scores = tf.nn.softmax(tf.reduce_sum(self._gate_u * proj, [2]) / gamma)
      inputs = tf.expand_dims(scores, 2) * inputs

    inputs = tf.nn.dropout(
      inputs, utils.switch_dropout(self._hparams.input_keep_prob))

    pooled_outputs = []
    for conv_layer in self._conv_layers:
      h = conv_layer(inputs)
      h = tf.nn.leaky_relu(h, alpha=self._hparams.leaky_relu_alpha)
      // pooling after conv
      h = tf.reduce_max(h, axis=1)
      pooled_outputs.append(h)

    outputs = tf.concat(pooled_outputs, 1)

After Change

        gamma = 1.
      scores = tf.reduce_sum(self._gate_u * proj, [2]) / gamma
      scores = scores * mask + ((1.0 - mask) * tf.float32.min)
      scores = tf.nn.softmax(scores)
      inputs = tf.expand_dims(scores, 2) * inputs
    else:
      inputs = tf.expands_dims(mask, 2) * inputs

    inputs = tf.nn.dropout(
      inputs, utils.switch_dropout(self._hparams.input_keep_prob))

    pooled_outputs = []
    for conv_layer in self._conv_layers:
      h = conv_layer(inputs)
      h = tf.nn.leaky_relu(h, alpha=self._hparams.leaky_relu_alpha)
      // pooling after conv
      h = tf.reduce_max(h, axis=1)
      pooled_outputs.append(h)

    outputs = tf.concat(pooled_outputs, 1)

Instances