tensorflow在代码不同位置调用神经网络想要共享权重

69 阅读 0 评论 46 点赞

我是靠谱客的博主迷人水壶，最近开发中收集的这篇文章主要介绍tensorflow在代码不同位置调用神经网络想要共享权重，觉得挺不错的，现在分享给大家，希望可以做个参考。

概述

一个简单的代码：
cnn网络定义了四个卷积层，权重为随机初始化的权重，输入为随机的（50,64,64,3）的向量
x1，x2的输出都调用了cnn这个函数

> import tensorflow as tf
>
> def cnn(hidden):
>
kwargs = dict(strides=2, activation=tf.nn.relu)
>
hidden = tf.layers.conv2d(hidden, 32, 6, **kwargs)
>
hidden = tf.layers.conv2d(hidden, 64, 4, **kwargs)
>
hidden = tf.layers.conv2d(hidden, 128, 4, **kwargs)
>
hidden = tf.layers.conv2d(hidden, 256, 6, **kwargs)
>
return hidden
>
>
> x = tf.random_uniform(shape=(1,64,64,3)) x1 = cnn(x) x2 = cnn(x)
>
>
> with tf.Session() as sess:
>
sess.run(tf.global_variables_initializer())
>
x1,x2 = sess.run((x1,x2))

此时程序的输出为
x1：


[[[[0.02014228 0.
0.
0.
0.
0.01977238
0.03838697 0.
0.02983167 0.02799959 0.
0.06397045
0.00736133 0.
0.03821873 0.0057195
0.
0.
0.04449557 0.
0.00071993 0.01928715 0.
0.00838473
0.01739994 0.0649211
0.02711494 0.04341522 0.
0.
0.
0.
0.00908231 0.
0.07131048 0.10422069
0.04077441 0.
0.
0.
0.
0.
0.
0.03626646 0.0288934
0.05571283 0.
0.04144262
0.
0.
0.
0.0188169
0.07488646 0.
0.02443938 0.03237902 0.
0.
0.05923894 0.
0.03736772 0.03978631 0.
0.
0.02545817 0.05508033
0.
0.02230431 0.
0.07633069 0.
0.
0.08178918 0.
0.
0.
0.00414837 0.
0.05754855 0.03113919 0.
0.
0.05155324 0.
0.09053664 0.
0.0110682
0.0217151
0.06611994 0.02389411
0.02607417...

x2：

[[[[0.00000000e+00 4.19012196e-02 0.00000000e+00 0.00000000e+00
0.00000000e+00 0.00000000e+00 0.00000000e+00 2.27831006e-02
0.00000000e+00 0.00000000e+00 1.51350833e-02 0.00000000e+00
0.00000000e+00 0.00000000e+00 3.25169414e-02 0.00000000e+00
0.00000000e+00 0.00000000e+00 0.00000000e+00 0.00000000e+00
0.00000000e+00 1.05397804e-02 4.10364196e-03 0.00000000e+00
0.00000000e+00 2.62358692e-02 0.00000000e+00 0.00000000e+00
0.00000000e+00 1.30762607e-02 0.00000000e+00 2.36116201e-02
5.30595221e-02 0.00000000e+00 2.96120755e-02 4.43295240e-02
6.72412962e-02 1.44396229e-02 6.79600835e-02 0.00000000e+00
4.11668187e-03 2.66320305e-03 0.00000000e+00 6.97394609e-02
0.00000000e+00 8.66800826e-03 4.97324169e-02 1.58384461e-02
6.17823116e-02 2.54427711e-03 4.39084396e-02 7.62259439e-02
3.68163846e-02 0.00000000e+00 4.27600034e-02 1.58234145e-02
0.00000000e+00 1.29800113e-02 0.00000000e+00 4.92318310e-02
0.00000000e+00 0.00000000e+00 0.0000...

若想让x1和x2输出一样的值：

import tensorflow as tf
def cnn(hidden):
kwargs = dict(strides=2, activation=tf.nn.relu)
hidden = tf.layers.conv2d(hidden, 32, 6, **kwargs)
hidden = tf.layers.conv2d(hidden, 64, 4, **kwargs)
hidden = tf.layers.conv2d(hidden, 128, 4, **kwargs)
hidden = tf.layers.conv2d(hidden, 256, 6, **kwargs)
return hidden
x = tf.random_uniform(shape=(1,64,64,3))
with tf.variable_scope("cnn"):
x1 = cnn(x)
with tf.variable_scope("cnn",reuse=True):
x2 = cnn(x)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
x1,x2 = sess.run((x1,x2))

此时x1和x2输出都是

[[[[0.
0.
0.
0.00473333 0.
0.
0.
0.02189207 0.
0.02265558 0.
0.0203282
0.
0.07767714 0.01926881 0.
0.
0.01530384
0.05857469 0.
0.07027114 0.05844937 0.
0.00194241
0.03735007 0.
0.05242127 0.0738355
0.
0.
0.00370752 0.00985879 0.01877109 0.
0.
0.
0.09586586 0.
0.
0.06655864 0.03449909 0.
0.
0.04952947 0.05264677 0.0590784
0.
0.10032271
0.
0.
0.
0.
0.
0.
0.05604002 0.04438177 0.02950861 0.04137567 0.
0.05043553
0.
0.
0.00433128 0.01197986 0.03473867 0.
0.03830672 0.00620859 0.00809103 0.
0.
0.01776852
0.05212681 0.
0.03483381 0.
0.
0.01069366
0.02201606 0.01776179 0.
0.
0.
0.0233967
0.02833988 0.
0.
0.
0.
0.02155995
...