tensorflow 获取变量&打印权值的实例讲解

379人浏览 / 0人评论

在使用tensorflow中，我们常常需要获取某个变量的值，比如：打印某一层的权重，通常我们可以直接利用变量的name属性来获取，但是当我们利用一些第三方的库来构造神经网络的layer时，存在一种情况：就是我们自己无法定义该层的变量，因为是自动进行定义的。

比如用tensorflow的slim库时：

def resnet_stack(images, output_shape, hparams, scope=None):
 """Create a resnet style transfer block.

 Args:
 images: [batch-size, height, width, channels] image tensor to feed as input
 output_shape: output image shape in form [height, width, channels]
 hparams: hparams objects
 scope: Variable scope

 Returns:
 Images after processing with resnet blocks.
 """
 end_points = {}
 if hparams.noise_channel:
 # separate the noise for visualization
 end_points['noise'] = images[:, :, :, -1]
 assert images.shape.as_list()[1:3] == output_shape[0:2]

 with tf.variable_scope(scope, 'resnet_style_transfer', [images]):
 with slim.arg_scope(
  [slim.conv2d],
  normalizer_fn=slim.batch_norm,
  kernel_size=[hparams.generator_kernel_size] * 2,
  stride=1):
  net = slim.conv2d(
   images,
   hparams.resnet_filters,
   normalizer_fn=None,
   activation_fn=tf.nn.relu)
  for block in range(hparams.resnet_blocks):
  net = resnet_block(net, hparams)
  end_points['resnet_block_{}'.format(block)] = net

  net = slim.conv2d(
   net,
   output_shape[-1],
   kernel_size=[1, 1],
   normalizer_fn=None,
   activation_fn=tf.nn.tanh,
   scope='conv_out')
  end_points['transferred_images'] = net
 return net, end_points

我们希望获取第一个卷积层的权重weight，该怎么办呢？？

在训练时，这些可训练的变量会被tensorflow保存在 tf.trainable_variables() 中，于是我们就可以通过打印 tf.trainable_variables() 来获取该卷积层的名称（或者你也可以自己根据scope来看出来该变量的name ），然后利用tf.get_default_grap().get_tensor_by_name 来获取该变量。

举个简单的例子：

import tensorflow as tf
with tf.variable_scope("generate"):
 with tf.variable_scope("resnet_stack"):
  #简单起见，这里没有用第三方库来说明，
  bias = tf.Variable(0.0,name="bias")
  weight = tf.Variable(0.0,name="weight")

for tv in tf.trainable_variables():
 print (tv.name)

b = tf.get_default_graph().get_tensor_by_name("generate/resnet_stack/bias:0")
w = tf.get_default_graph().get_tensor_by_name("generate/resnet_stack/weight:0")

with tf.Session() as sess:
 tf.global_variables_initializer().run()
 print(sess.run(b))
 print(sess.run(w))