个人中心
gpt4 book ai didi

tensorflow1.0学习之模型的保存与恢复(Saver)

转载 作者：qq735679552 更新时间：2022-09-28 22:32:09 28 4
gpt4 key购买 nike

CFSDN坚持开源创造价值，我们致力于搭建一个资源共享平台，让每一个IT人在这里找到属于你的精彩世界.

这篇CFSDN的博客文章tensorflow1.0学习之模型的保存与恢复(Saver)由作者收集整理，如果你对这篇文章有兴趣，记得点赞哟.

将训练好的模型参数保存起来，以便以后进行验证或测试，这是我们经常要做的事情。tf里面提供模型保存的是tf.train.Saver()模块.

模型保存，先要创建一个Saver对象：如 。

?
1
saver = tf.train.Saver()

在创建这个Saver对象的时候，有一个参数我们经常会用到，就是 max_to_keep 参数，这个是用来设置保存模型的个数，默认为5，即 max_to_keep=5，保存最近的5个模型。如果你想每训练一代（epoch)就想保存一次模型，则可以将 max_to_keep设置为None或者0，如:

?
1
saver = tf.train.Saver(max_to_keep = 0 )

但是这样做除了多占用硬盘，并没有实际多大的用处，因此不推荐.

当然，如果你只想保存最后一代的模型，则只需要将max_to_keep设置为1即可，即 。

?
1
saver = tf.train.Saver(max_to_keep = 1 )

创建完saver对象后，就可以保存训练好的模型了，如:

?
1
saver.save(sess, 'ckpt/mnist.ckpt' ,global_step = step)

第一个参数sess,这个就不用说了。第二个参数设定保存的路径和名字，第三个参数将训练的次数作为后缀加入到模型名字中.

saver.save(sess, 'my-model', global_step=0) ==>      filename: 'my-model-0'  ...  saver.save(sess, 'my-model', global_step=1000) ==> filename: 'my-model-1000' 。

看一个mnist实例:

?
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
# -*- coding: utf-8 -*-
"""
Created on Sun Jun 4 10:29:48 2017
 
@author: Administrator
"""
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets( "MNIST_data/" , one_hot = False )
 
x = tf.placeholder(tf.float32, [ None , 784 ])
y_ = tf.placeholder(tf.int32,[ None ,])
 
dense1 = tf.layers.dense(inputs = x,
            units = 1024 ,
            activation = tf.nn.relu,
            kernel_initializer = tf.truncated_normal_initializer(stddev = 0.01 ),
            kernel_regularizer = tf.nn.l2_loss)
dense2 = tf.layers.dense(inputs = dense1,
            units = 512 ,
            activation = tf.nn.relu,
            kernel_initializer = tf.truncated_normal_initializer(stddev = 0.01 ),
            kernel_regularizer = tf.nn.l2_loss)
logits = tf.layers.dense(inputs = dense2,
             units = 10 ,
             activation = None ,
             kernel_initializer = tf.truncated_normal_initializer(stddev = 0.01 ),
             kernel_regularizer = tf.nn.l2_loss)
 
loss = tf.losses.sparse_softmax_cross_entropy(labels = y_,logits = logits)
train_op = tf.train.AdamOptimizer(learning_rate = 0.001 ).minimize(loss)
correct_prediction = tf.equal(tf.cast(tf.argmax(logits, 1 ),tf.int32), y_) 
acc = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
 
sess = tf.InteractiveSession()
sess.run(tf.global_variables_initializer())
 
saver = tf.train.Saver(max_to_keep = 1 )
for i in range ( 100 ):
  batch_xs, batch_ys = mnist.train.next_batch( 100 )
  sess.run(train_op, feed_dict = {x: batch_xs, y_: batch_ys})
  val_loss,val_acc = sess.run([loss,acc], feed_dict = {x: mnist.test.images, y_: mnist.test.labels})
  print ( 'epoch:%d, val_loss:%f, val_acc:%f' % (i,val_loss,val_acc))
  saver.save(sess, 'ckpt/mnist.ckpt' ,global_step = i + 1 )
sess.close()

代码中红色部分就是保存模型的代码，虽然我在每训练完一代的时候，都进行了保存，但后一次保存的模型会覆盖前一次的，最终只会保存最后一次。因此我们可以节省时间，将保存代码放到循环之外（仅适用max_to_keep=1,否则还是需要放在循环内). 。

在实验中，最后一代可能并不是验证精度最高的一代，因此我们并不想默认保存最后一代，而是想保存验证精度最高的一代，则加个中间变量和判断语句就可以了.

?
1
2
3
4
5
6
7
8
9
10
11
saver = tf.train.Saver(max_to_keep = 1 )
max_acc = 0
for i in range ( 100 ):
  batch_xs, batch_ys = mnist.train.next_batch( 100 )
  sess.run(train_op, feed_dict = {x: batch_xs, y_: batch_ys})
  val_loss,val_acc = sess.run([loss,acc], feed_dict = {x: mnist.test.images, y_: mnist.test.labels})
  print ( 'epoch:%d, val_loss:%f, val_acc:%f' % (i,val_loss,val_acc))
  if val_acc>max_acc:
    max_acc = val_acc
    saver.save(sess, 'ckpt/mnist.ckpt' ,global_step = i + 1 )
sess.close()

如果我们想保存验证精度最高的三代，且把每次的验证精度也随之保存下来，则我们可以生成一个txt文件用于保存.

?
1
2
3
4
5
6
7
8
9
10
11
12
13
14
saver = tf.train.Saver(max_to_keep = 3 )
max_acc = 0
f = open ( 'ckpt/acc.txt' , 'w' )
for i in range ( 100 ):
  batch_xs, batch_ys = mnist.train.next_batch( 100 )
  sess.run(train_op, feed_dict = {x: batch_xs, y_: batch_ys})
  val_loss,val_acc = sess.run([loss,acc], feed_dict = {x: mnist.test.images, y_: mnist.test.labels})
  print ( 'epoch:%d, val_loss:%f, val_acc:%f' % (i,val_loss,val_acc))
  f.write( str (i + 1 ) + ', val_acc: ' + str (val_acc) + ' ' )
  if val_acc>max_acc:
    max_acc = val_acc
    saver.save(sess, 'ckpt/mnist.ckpt' ,global_step = i + 1 )
f.close()
sess.close()

模型的恢复用的是restore()函数，它需要两个参数restore(sess, save_path)，save_path指的是保存的模型路径。我们可以使用tf.train.latest_checkpoint（）来自动获取最后一次保存的模型。如:

?
1
2
model_file = tf.train.latest_checkpoint( 'ckpt/' )
saver.restore(sess,model_file)

则程序后半段代码我们可以改为:

?
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
sess = tf.InteractiveSession()
sess.run(tf.global_variables_initializer())
is_train = False
saver = tf.train.Saver(max_to_keep = 3 )
 
#训练阶段
if is_train:
   max_acc = 0
   f = open ( 'ckpt/acc.txt' , 'w' )
   for i in range ( 100 ):
    batch_xs, batch_ys = mnist.train.next_batch( 100 )
    sess.run(train_op, feed_dict = {x: batch_xs, y_: batch_ys})
    val_loss,val_acc = sess.run([loss,acc], feed_dict = {x: mnist.test.images, y_: mnist.test.labels})
    print ( 'epoch:%d, val_loss:%f, val_acc:%f' % (i,val_loss,val_acc))
    f.write( str (i + 1 ) + ', val_acc: ' + str (val_acc) + ' ' )
    if val_acc>max_acc:
      max_acc = val_acc
      saver.save(sess, 'ckpt/mnist.ckpt' ,global_step = i + 1 )
   f.close()
 
#验证阶段
else :
   model_file = tf.train.latest_checkpoint( 'ckpt/' )
   saver.restore(sess,model_file)
   val_loss,val_acc = sess.run([loss,acc], feed_dict = {x: mnist.test.images, y_: mnist.test.labels})
   print ( 'val_loss:%f, val_acc:%f' % (val_loss,val_acc))
sess.close()

标红的地方，就是与保存、恢复模型相关的代码。用一个bool型变量is_train来控制训练和验证两个阶段.

整个源程序:

?
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
# -*- coding: utf-8 -*-
"""
Created on Sun Jun 4 10:29:48 2017
 
@author: Administrator
"""
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets( "MNIST_data/" , one_hot = False )
 
x = tf.placeholder(tf.float32, [ None , 784 ])
y_ = tf.placeholder(tf.int32,[ None ,])
 
dense1 = tf.layers.dense(inputs = x,
            units = 1024 ,
            activation = tf.nn.relu,
            kernel_initializer = tf.truncated_normal_initializer(stddev = 0.01 ),
            kernel_regularizer = tf.nn.l2_loss)
dense2 = tf.layers.dense(inputs = dense1,
            units = 512 ,
            activation = tf.nn.relu,
            kernel_initializer = tf.truncated_normal_initializer(stddev = 0.01 ),
            kernel_regularizer = tf.nn.l2_loss)
logits = tf.layers.dense(inputs = dense2,
             units = 10 ,
             activation = None ,
             kernel_initializer = tf.truncated_normal_initializer(stddev = 0.01 ),
             kernel_regularizer = tf.nn.l2_loss)
 
loss = tf.losses.sparse_softmax_cross_entropy(labels = y_,logits = logits)
train_op = tf.train.AdamOptimizer(learning_rate = 0.001 ).minimize(loss)
correct_prediction = tf.equal(tf.cast(tf.argmax(logits, 1 ),tf.int32), y_) 
acc = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
 
sess = tf.InteractiveSession()
sess.run(tf.global_variables_initializer())
 
is_train = True
saver = tf.train.Saver(max_to_keep = 3 )
 
#训练阶段
if is_train:
   max_acc = 0
   f = open ( 'ckpt/acc.txt' , 'w' )
   for i in range ( 100 ):
    batch_xs, batch_ys = mnist.train.next_batch( 100 )
    sess.run(train_op, feed_dict = {x: batch_xs, y_: batch_ys})
    val_loss,val_acc = sess.run([loss,acc], feed_dict = {x: mnist.test.images, y_: mnist.test.labels})
    print ( 'epoch:%d, val_loss:%f, val_acc:%f' % (i,val_loss,val_acc))
    f.write( str (i + 1 ) + ', val_acc: ' + str (val_acc) + ' ' )
    if val_acc>max_acc:
      max_acc = val_acc
      saver.save(sess, 'ckpt/mnist.ckpt' ,global_step = i + 1 )
   f.close()
 
#验证阶段
else :
   model_file = tf.train.latest_checkpoint( 'ckpt/' )
   saver.restore(sess,model_file)
   val_loss,val_acc = sess.run([loss,acc], feed_dict = {x: mnist.test.images, y_: mnist.test.labels})
   print ( 'val_loss:%f, val_acc:%f' % (val_loss,val_acc))
sess.close()

参考文章：http://www.zzvips.com/article/138370.html 。

以上就是本文的全部内容，希望对大家的学习有所帮助，也希望大家多多支持我.

原文链接：https://www.cnblogs.com/denny402/p/6940134.html 。

最后此篇关于tensorflow1.0学习之模型的保存与恢复(Saver)的文章就讲到这里了,如果你想了解更多关于tensorflow1.0学习之模型的保存与恢复(Saver)的内容请搜索CFSDN的文章或继续浏览相关文章，希望大家以后支持我的博客！ 。

28 4 0
文章推荐： python基于pygame实现响应游戏中事件的方法(附源码)
文章推荐： 使用tensorflow实现线性svm
文章推荐： Python基于pygame实现的弹力球效果(附源码)
文章推荐： Python中pygame安装方法图文详解
qq735679552
个人简介

我是一名优秀的程序员,十分优秀！

滴滴打车优惠券免费领取
滴滴打车优惠券
Copyright 2021 - 2024 cfsdn All Rights Reserved 蜀ICP备2022000587号
广告合作:1813099741@qq.com 6ren.com