tensorflow saver简介+Demo with linear-model

tf.train.Saver提供Save和Restore Tensorflow变量的功能，常用于保存、还原模型训练结果，这在自己的训练和迁移学习中都很有用。

训练、保存脚本：

import tensorflow as tf

checkpoint_dir = './ckpt/'

x_train = [1, 2, 3, 6, 8]
y_train = [4.8, 8.5, 10.4, 21.0, 25.3]

x = tf.placeholder(tf.float32, name='x')
y = tf.placeholder(tf.float32, name='y')

W = tf.Variable(1, dtype=tf.float32, name='W')
b = tf.Variable(0, dtype=tf.float32, name='b')

# 定义模型
linear_model = W * x + b

with tf.name_scope("loss-model"):
    loss = tf.reduce_sum(tf.square(linear_model - y))
    acc = tf.sqrt(loss)

train_step = tf.train.GradientDescentOptimizer(0.001).minimize(loss)

sess = tf.Session()
init = tf.global_variables_initializer()
sess.run(init)

variable_saver = tf.train.Saver(max_to_keep=3)

# 训练、保存variables
for i in range(1000):
    sess.run([train_step], {x: x_train, y: y_train})
    if i%10 == 0:
        variable_saver.save(sess, checkpoint_dir, i)

curr_W, curr_b, curr_loss, curr_acc = sess.run([W, b, loss, acc], {x: x_train, y: y_train})
print("After train W: %f, b: %f, loss: %f, acc: %f" % (curr_W, curr_b, curr_loss, curr_acc))

运行保存的文件如下

还原保存的变量：

import tensorflow as tf

checkpoint_dir = './ckpt/'

W = tf.Variable(1, dtype=tf.float32, name='W')
b = tf.Variable(0, dtype=tf.float32, name='b')

sess = tf.Session()
init = tf.global_variables_initializer()
sess.run(init)

variable_saver = tf.train.Saver(max_to_keep=3)
latest_checkpoint = tf.train.latest_checkpoint(checkpoint_dir)
if latest_checkpoint is not None:
    variable_saver.restore(sess, latest_checkpoint)

curr_W, curr_b = sess.run([W, b])
print("After train W: %f, b: %f" % (curr_W, curr_b))

参考了：https://blog.csdn.net/gzj_1101/article/details/80299610