tensorflow的基本用法(十)——保存神经网络参数和加载神经网络参数

#!/usr/bin/env python
# _*_ coding: utf-8 _*_

import tensorflow as tf
import numpy as np

# 保存神经网络参数
def save_para():
# 定义权重参数
W = tf.Variable([[1, 2, 3], [4, 5, 6]], dtype = tf.float32, name = 'weights')
# 定义偏置参数
b = tf.Variable([[1, 2, 3]], dtype = tf.float32, name = 'biases')
# 参数初始化
init = tf.global_variables_initializer()
# 定义保存参数的saver
saver = tf.train.Saver()

with tf.Session() as sess:
sess.run(init)
# 保存session中的数据
save_path = saver.save(sess, 'my_net/save_net.ckpt')
# 输出保存路径
print 'Save to path: ', save_path

# 恢复神经网络参数
def restore_para():
# 定义权重参数
W = tf.Variable(np.arange(6).reshape((2, 3)), dtype = tf.float32, name = 'weights')
# 定义偏置参数
b = tf.Variable(np.arange(3).reshape((1, 3)), dtype = tf.float32, name = 'biases')
# 定义提取参数的saver
saver = tf.train.Saver()

with tf.Session() as sess:
# 加载文件中的参数数据，会根据name加载数据并保存到变量W和b中
save_path = saver.restore(sess, 'my_net/save_net.ckpt')
# 输出保存路径
print 'Weights: ', sess.run(W)
print 'biases:  ', sess.run(b)

# save_para()
restore_para()

# save
Save to path:  my_net/save_net.ckpt

# restore
Weights:  [[ 1.  2.  3.]
[ 4.  5.  6.]]
biases:   [[ 1.  2.  3.]]

参考资料