Tensorflow 中的交互式会话 - 卷积运算符的不同输出

Question

我是 tensorflow 的新手，我遇到了 InteractiveSession 的问题。

在下面的代码中：

import tensorflow as tf

def weight_variable(shape):
  initial = tf.random_uniform(shape, 0, 10, seed=1, dtype="int32")
  print("weights=\n",initial.eval())
  return tf.Variable(tf.to_float(initial))

def conv2d(x, W):
  return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')


# first dimension: Number of examples to train on, 2nd and 3rd: example width and height, 
# last one is: the number of channels 

x = tf.to_float(tf.Variable([[[[1],     [4],    [5],    [6],    [7]],  
                             [[10],   [11],   [22],    [9],    [8]],  
                             [[24],   [25],   [20],   [21],   [19]],  
                             [[14],   [12],   [13],    [3],   [18]],  
                             [[15],   [16],   [19],   [18],   [17]]]])) # 1 example of 5x5 one channel image


sess = tf.InteractiveSession()

# The first two dimensions are the patch size, the next is the number of input channels, 
# and the last is the number of output channels. 
W_conv1 = weight_variable([2, 2, 1, 1]) #[3,3,3,64]

conv = conv2d(x, W_conv1)


sess.run(tf.initialize_all_variables())

print(sess.run(conv))

sess.close()

当我评论这行时：

print("weights=\n",initial.eval())

打印卷积时得到不同的结果print(sess.run(conv))。我知道关键字 eval 与会话交互，但我的理解是无论我是否使用它都不会改变输出。

这是我使用 initial.eval():

时得到的输出

[[[[7]]

[[9]]]

[[[3]]

[[2]]]] [[[[ 156.] [ 209.] [ 278.] [ 167.] [ 79.]]

[[ 389.] [ 472.] [ 337.] [ 319.] [ 179.]]

[[ 386.] [ 332.] [ 314.] [ 254.] [ 181.]]

[[ 293.] [ 317.] [ 262.] [ 360.] [ 171.]]

[[ 143.] [ 168.] [ 163.] [ 154.] [ 17.]]]]

当我评论该行时，我得到：

[[[[ 95.] [ 150.] [ 173.] [ 148.] [ 73.]]

[[ 291.] [ 390.] [ 337.] [ 236.] [ 113.]]

[[ 459.] [ 417.] [ 374.] [ 363.] [ 187.]]

[[ 283.] [ 287.] [ 211.] [ 271.] [ 177.]]

[[ 249.] [ 283.] [ 295.] [ 279.] [ 119.]]]]

注意把156改成95和剩下的卷积输出。

Answer 1

这是因为 RNG 的种子运作方式。在 tf.random_uniform 中设置 op-level seed 为伪 RNG 提供了一个固定的起点，但并不意味着对 op 的重复评估将产生相同的随机数。如果您查看 tf.set_random_seed 并通过调用 eval() 两次并打印输出的玩具示例，可以在文档中看到这一点：

In [2]: initial = tf.random_uniform((5,), 0, 10, seed=1, dtype="int32")
   ...: print("weights=\n",initial.eval())
   ...: print("weights=\n",initial.eval())
   ...: 
('weights=\n', array([7, 9, 3, 2, 7], dtype=int32))
('weights=\n', array([3, 5, 5, 4, 9], dtype=int32))

In [3]: initial = tf.random_uniform((5,), 0, 10, seed=1, dtype="int32")
   ...: print("weights=\n",initial.eval())
   ...: print("weights=\n",initial.eval())
   ...: 
('weights=\n', array([7, 9, 3, 2, 7], dtype=int32))
('weights=\n', array([3, 5, 5, 4, 9], dtype=int32))

Tensorflow 中的交互式会话 - 卷积运算符的不同输出

Interactive Session in Tensorflow - different output for convolution operator

python-3.5

tensorflow