tensorflow随笔-训练检查点tf.train.Saver

#!/usr/bin/env python2
# -*- coding: utf-8 -*-


import tensorflow as tf
g1=tf.Graph()

with g1.as_default(): 
    with tf.name_scope("input_Variable"):        
        my_var=tf.Variable(1,dtype=tf.float32)
    with tf.name_scope("global_step"):
        my_step=tf.Variable(0,dtype=tf.int32)
    with tf.name_scope("update"):
        varop=tf.assign(my_var,tf.multiply(tf.log(tf.add(my_var,1)),1))
        stepop=tf.assign_add(my_step,1)
        addop=tf.group([varop,stepop])
    with tf.name_scope("summaries"):
        tf.summary.scalar('myvar',my_var)
    with tf.name_scope("global_ops"):
        init=tf.global_variables_initializer()
        merged_summaries=tf.summary.merge_all()

with tf.Session(graph=g1) as sess:  
    writer=tf.summary.FileWriter('sum_vars',sess.graph)
    sess.run(init)
    #---0
    step,var,summary=sess.run([my_step,my_var,merged_summaries])
    writer.add_summary(summary,global_step=step)
    print step,var
    saver=tf.train.Saver()
    #1-49
    for i in xrange(1,50):
        sess.run(addop)
        step,var,summary=sess.run([my_step,my_var,merged_summaries])
        writer.add_summary(summary,global_step=step)
        print step,var
        if i%5==0:
            saver.save(sess,'./myvar-model/myvar-model',global_step=i)
    saver.save(sess,'./myvar-model/myvar-model',global_step=49)

    writer.flush()
    writer.close()

38 0.0512373
39 0.04996785
40 0.048759546
41 0.04760808
42 0.04650955
43 0.045460388
44 0.04445735
45 0.04349747
46 0.042578023
47 0.041696515
48 0.040850647
49 0.04003831

保存数据流图的变量到二进制检查点文件。

#!/usr/bin/env python2
# -*- coding: utf-8 -*-


import tensorflow as tf
import os
g1=tf.Graph()

with g1.as_default(): 
    with tf.name_scope("input_Variable"):        
        my_var=tf.Variable(1,dtype=tf.float32)
    with tf.name_scope("global_step"):
        my_step=tf.Variable(0,dtype=tf.int32,trainable=False)
    with tf.name_scope("update"):
        varop=tf.assign(my_var,tf.multiply(tf.log(tf.add(my_var,1)),1))
        stepop=tf.assign_add(my_step,1)
        addop=tf.group([varop,stepop])
    with tf.name_scope("summaries"):
        tf.summary.scalar('myvar',my_var)
    with tf.name_scope("global_ops"):
        init=tf.global_variables_initializer()
        merged_summaries=tf.summary.merge_all()

with tf.Session(graph=g1) as sess:  
    writer=tf.summary.FileWriter('sum_vars',sess.graph)
    sess.run(init)

    saver=tf.train.Saver()

    #如果之前保存了检查点文件，则恢复模型后，继续
    init_step=0
    ckpt=tf.train.get_checkpoint_state(os.getcwd()+'/myvar-model')
    if ckpt and ckpt.model_checkpoint_path:
        saver.restore(sess,ckpt.model_checkpoint_path)
        init_step=int(ckpt.model_checkpoint_path.rsplit('-',1)[1])
        print "读取检查点文件..."
    for i in xrange(init_step,100):
        step,var,summary=sess.run([my_step,my_var,merged_summaries])
        writer.add_summary(summary,global_step=step)
        print step,var,init_step
        if i%5==0 and i<=50:
            print "保存检查点文件"
            saver.save(sess,'./myvar-model/myvar-model',global_step=i)
        sess.run(addop)

    writer.flush()
    writer.close()

上面代码跑第一次时，检查点文件被保存，跑第二次开始，检查点文件将被读取，循环次数从step=50开始。

跑第二次时

读取检查点文件…
50 0.03925755 50
保存检查点文件
51 0.038506564 50
52 0.037783686 50
53 0.03708737 50
54 0.036416177 50
55 0.035768777 50
56 0.03514393 50
…
…
…
93 0.021334965 50
94 0.02111056 50
95 0.02089082 50
96 0.0206756 50
97 0.020464761 50
98 0.020258171 50
99 0.020055704 50