TensorFlow reuse variable with tf.layers.conv2d

问题:

I am trying to make 2 conv layers share the same weights, however, it seems the API does not work.

import tensorflow as tf  x = tf.random_normal(shape=[10, 32, 32, 3])  with tf.variable_scope('foo') as scope:     conv1 = tf.contrib.layers.conv2d(x, 3, [2, 2], padding='SAME', reuse=True, scope=scope)     print(conv1.name)      conv2 = tf.contrib.layers.conv2d(x, 3, [2, 2], padding='SAME', reuse=True, scope=scope)     print(conv2.name) 

It prints out

foo/foo/Relu:0 foo/foo_1/Relu:0 

Changing from tf.contrib.layers.conv2d to tf.layers.conv2d does not solve the problem.

It has the same problem with tf.layers.conv2d:

import tensorflow as tf  x = tf.random_normal(shape=[10, 32, 32, 3])  conv1 = tf.layers.conv2d(x, 3, [2, 2], padding='SAME', reuse=None, name='conv') print(conv1.name) conv2 = tf.layers.conv2d(x, 3, [2, 2], padding='SAME', reuse=True, name='conv') print(conv2.name) 

gives

conv/BiasAdd:0 conv_2/BiasAdd:0 

回答1:

In the code you wrote, variables do get reused between the two convolution layers. Try this :

import tensorflow as tf  x = tf.random_normal(shape=[10, 32, 32, 3])  conv1 = tf.layers.conv2d(x, 3, [2, 2], padding='SAME', reuse=None, name='conv')  conv2 = tf.layers.conv2d(x, 3, [2, 2], padding='SAME', reuse=True, name='conv')  print([x.name for x in tf.global_variables()])  # prints # [u'conv/kernel:0', u'conv/bias:0'] 

Note that only one weight and one bias tensor has been created. Even though they share the weights, the layers do not share the actual computation. Hence you see the two different names for the operations.

文章来源: TensorFlow reuse variable with tf.layers.conv2d