MNIST数据集上的无条件生成对抗串行网络

Question

我正在使用tfgan库和tfgan估计器在MNIST数据集上训练无条件GAN。一切都很好，图像正在生成，see。生成器和鉴别器模型函数的助手函数是使用tf.layers编写的。但是，当我只更改助手函数并使用tf.keras编写它们时，相同的代码不起作用，也没有图像生成，see。有人能帮我解决这个问题吗？这两个脚本之间唯一的区别是帮助器函数从使用tf.layers改为使用tf.keras。使用tf.layers的助手函数：

def _dense(inputs, units, l2_weight):
  return tf.layers.dense(
      inputs, units, None,
      kernel_initializer=tf.keras.initializers.glorot_uniform,
      kernel_regularizer=tf.keras.regularizers.l2(l=l2_weight),
      bias_regularizer=tf.keras.regularizers.l2(l=l2_weight))

def _batch_norm(inputs, is_training):
  return tf.layers.batch_normalization(
      inputs, momentum=0.999, epsilon=0.001, training=is_training)

def _deconv2d(inputs, filters, kernel_size, stride, l2_weight):
  return tf.layers.conv2d_transpose(
      inputs, filters, [kernel_size, kernel_size], strides=[stride, stride], 
      activation=tf.nn.relu, padding='same',
      kernel_initializer=tf.keras.initializers.glorot_uniform,
      kernel_regularizer=tf.keras.regularizers.l2(l=l2_weight),
      bias_regularizer=tf.keras.regularizers.l2(l=l2_weight))

def _conv2d(inputs, filters, kernel_size, stride, l2_weight):
  return tf.layers.conv2d(
      inputs, filters, [kernel_size, kernel_size], strides=[stride, stride], 
      activation=None, padding='same',
      kernel_initializer=tf.keras.initializers.glorot_uniform,
      kernel_regularizer=tf.keras.regularizers.l2(l=l2_weight),
      bias_regularizer=tf.keras.regularizers.l2(l=l2_weight)) 

使用tf.keras的助手函数：

def _dense(inputs, units, l2_weight):
  return Dense(units,
      kernel_initializer=tf.keras.initializers.glorot_uniform,
      kernel_regularizer=tf.keras.regularizers.l2(l=l2_weight),
      bias_regularizer=tf.keras.regularizers.l2(l=l2_weight))(inputs)

def _batch_norm(inputs, is_training):
  return BatchNormalization(momentum=0.999, epsilon=0.001)(inputs, training = is_training)


def _deconv2d(inputs, filters, kernel_size, stride, l2_weight):
  return Conv2DTranspose(filters=filters, kernel_size=[kernel_size, kernel_size], strides=[stride, stride],
                                      activation=keras.activations.relu, padding='same',
                                      kernel_initializer=keras.initializers.glorot_uniform,
                                      kernel_regularizer=keras.regularizers.l2(l=l2_weight),
                                      bias_regularizer=keras.regularizers.l2(l=l2_weight))(inputs)

def _conv2d(inputs, filters, kernel_size, stride, l2_weight):
  return Conv2D(filters=filters, kernel_size=[kernel_size, kernel_size], strides=[stride, stride], padding='same',
      kernel_initializer=tf.keras.initializers.glorot_uniform,
      kernel_regularizer=tf.keras.regularizers.l2(l=l2_weight),
      bias_regularizer=tf.keras.regularizers.l2(l=l2_weight))(inputs)

Answer 1

不幸的是，tfgan目前依赖于variable_scopes才能正常工作，而Keras层并不尊重variable_scopes。我们有一个重新设计的总体计划，将支持Keras，但不幸的是，目前我们没有任何东西可以展示它或ETA。欢迎贡献代码！