tensorflow模型的save与restore,及checkpoint中读取变量方式

Tanya ·

更新时间:2024-09-21

· 804 次阅读

创建一个NN

import tensorflow as tf
import numpy as np


#fake data
x = np.linspace(-1, 1, 100)[:, np.newaxis] #shape(100,1)
noise = np.random.normal(0, 0.1, size=x.shape)
y = np.power(x, 2) + noise  #shape(100,1) + noise
tf_x = tf.placeholder(tf.float32, x.shape) #input x
tf_y = tf.placeholder(tf.float32, y.shape) #output y
l = tf.layers.dense(tf_x, 10, tf.nn.relu) #hidden layer
o = tf.layers.dense(l, 1)     #output layer
loss = tf.losses.mean_squared_error(tf_y, o ) #compute loss
train_op = tf.train.GradientDescentOptimizer(learning_rate=0.5).minimize(loss)

1.使用save对模型进行保存


sess= tf.Session()
sess.run(tf.global_variables_initializer())  #initialize var in graph
saver = tf.train.Saver() # define a saver for saving and restoring
for step in range(100):   #train
 sess.run(train_op,{tf_x:x, tf_y:y})
saver.save(sess, 'params/params.ckpt', write_meta_graph=False) # mate_graph is not recommend

生成三个文件，分别是checkpoint，.ckpt.data-00000-of-00001,.ckpt.index

2.使用restore对提取模型

在提取模型时，需要将模型结构再定义一遍，再将各参数加载出来


#bulid entire net again and restore
tf_x = tf.placeholder(tf.float32, x.shape)
tf_y = tf.placeholder(tf.float32, y.shape)
l_ = tf.layers.dense(tf_x, 10, tf.nn.relu)
o_ = tf.layers.dense(l_, 1)
loss_ = tf.losses.mean_squared_error(tf_y, o_)
sess = tf.Session()
# don't need to initialize variables, just restoring trained variables
saver = tf.train.Saver() # define a saver for saving and restoring
saver.restore(sess, './params/params.ckpt')

3.有时会报错Not found：b1 not found in checkpoint

这时我们想知道我在文件中到底保存了什么内容，即需要读取出checkpoint中的tensor


import os
from tensorflow.python import pywrap_tensorflow
checkpoint_path = os.path.join('params','params.ckpt')
# Read data from checkpoint file
reader = pywrap_tensorflow.NewCheckpointReader(checkpoint_path)
var_to_shape_map = reader.get_variable_to_shape_map()
# Print tensor name and value
f = open('params.txt','w')
for key in var_to_shape_map: # write tensors' names and values in file
 print(key,file=f)
 print(reader.get_tensor(key),file=f)
f.close()

运行后生成一个params.txt文件，在其中可以看到模型的参数。

补充知识：TensorFlow按时间保存检查点

一实例

介绍一种更简便地保存检查点功能的方法——tf.train.MonitoredTrainingSession函数，该函数可以直接实现保存及载入检查点模型的文件。

演示使用MonitoredTrainingSession函数来自动管理检查点文件。

二代码


import tensorflow as tf
tf.reset_default_graph()
global_step = tf.train.get_or_create_global_step()
step = tf.assign_add(global_step, 1)
with tf.train.MonitoredTrainingSession(checkpoint_dir='log/checkpoints',save_checkpoint_secs = 2) as sess:
 print(sess.run([global_step]))
 while not sess.should_stop():
  i = sess.run( step)
  print( i)

三运行结果

1 第一次运行后，会发现log文件夹下产生如下文件

2 第二次运行后，结果如下：

INFO:tensorflow:Create CheckpointSaverHook.
INFO:tensorflow:Restoring parameters from log/checkpoints\model.ckpt-15147
INFO:tensorflow:Saving checkpoints for 15147 into log/checkpoints\model.ckpt.
[15147]
15148
15149
15150
15151
15152
15153
15154
15155
15156
15157
15158
15159