服务器之家

服务器之家 > 正文

将TensorFlow的模型网络导出为单个文件的方法

时间:2021-02-04 00:55     来源/作者:EncodeTS

有时候,我们需要将TensorFlow的模型导出为单个文件(同时包含模型架构定义与权重),方便在其他地方使用(如在c++中部署网络)。利用tf.train.write_graph()默认情况下只导出了网络的定义(没有权重),而利用tf.train.Saver().save()导出的文件graph_def与权重是分离的,因此需要采用别的方法。

我们知道,graph_def文件中没有包含网络中的Variable值(通常情况存储了权重),但是却包含了constant值,所以如果我们能把Variable转换为constant,即可达到使用一个文件同时存储网络架构与权重的目标。

我们可以采用以下方式冻结权重并保存网络:

?
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
import tensorflow as tf
from tensorflow.python.framework.graph_util import convert_variables_to_constants
 
# 构造网络
a = tf.Variable([[3],[4]], dtype=tf.float32, name='a')
b = tf.Variable(4, dtype=tf.float32, name='b')
# 一定要给输出tensor取一个名字!!
output = tf.add(a, b, name='out')
 
# 转换Variable为constant,并将网络写入到文件
with tf.Session() as sess:
  sess.run(tf.global_variables_initializer())
  # 这里需要填入输出tensor的名字
  graph = convert_variables_to_constants(sess, sess.graph_def, ["out"])
  tf.train.write_graph(graph, '.', 'graph.pb', as_text=False)

当恢复网络时,可以使用如下方式:

?
1
2
3
4
5
6
7
import tensorflow as tf
with tf.Session() as sess:
  with open('./graph.pb', 'rb') as f:
    graph_def = tf.GraphDef()
    graph_def.ParseFromString(f.read())
    output = tf.import_graph_def(graph_def, return_elements=['out:0'])
    print(sess.run(output))

输出结果为:

[array([[ 7.],
       [ 8.]], dtype=float32)]

可以看到之前的权重确实保存了下来!!

问题来了,我们的网络需要能有一个输入自定义数据的接口啊!不然这玩意有什么用。。别急,当然有办法。

?
1
2
3
4
5
6
7
8
9
10
11
import tensorflow as tf
from tensorflow.python.framework.graph_util import convert_variables_to_constants
a = tf.Variable([[3],[4]], dtype=tf.float32, name='a')
b = tf.Variable(4, dtype=tf.float32, name='b')
input_tensor = tf.placeholder(tf.float32, name='input')
output = tf.add((a+b), input_tensor, name='out')
 
with tf.Session() as sess:
  sess.run(tf.global_variables_initializer())
  graph = convert_variables_to_constants(sess, sess.graph_def, ["out"])
  tf.train.write_graph(graph, '.', 'graph.pb', as_text=False)

用上述代码重新保存网络至graph.pb,这次我们有了一个输入placeholder,下面来看看怎么恢复网络并输入自定义数据。

?
1
2
3
4
5
6
7
8
import tensorflow as tf
 
with tf.Session() as sess:
  with open('./graph.pb', 'rb') as f:
    graph_def = tf.GraphDef()
    graph_def.ParseFromString(f.read())
    output = tf.import_graph_def(graph_def, input_map={'input:0':4.}, return_elements=['out:0'], name='a')
    print(sess.run(output))

输出结果为:

[array([[ 11.],
       [ 12.]], dtype=float32)]

可以看到结果没有问题,当然在input_map那里可以替换为新的自定义的placeholder,如下所示:

?
1
2
3
4
5
6
7
8
9
10
import tensorflow as tf
 
new_input = tf.placeholder(tf.float32, shape=())
 
with tf.Session() as sess:
  with open('./graph.pb', 'rb') as f:
    graph_def = tf.GraphDef()
    graph_def.ParseFromString(f.read())
    output = tf.import_graph_def(graph_def, input_map={'input:0':new_input}, return_elements=['out:0'], name='a')
    print(sess.run(output, feed_dict={new_input:4}))

看看输出,同样没有问题。

[array([[ 11.],
       [ 12.]], dtype=float32)]

另外需要说明的一点是,在利用tf.train.write_graph写网络架构的时候,如果令as_text=True了,则在导入网络的时候,需要做一点小修改。

?
1
2
3
4
5
6
7
8
9
10
11
import tensorflow as tf
from google.protobuf import text_format
 
with tf.Session() as sess:
  # 不使用'rb'模式
  with open('./graph.pb', 'r') as f:
    graph_def = tf.GraphDef()
    # 不使用graph_def.ParseFromString(f.read())
    text_format.Merge(f.read(), graph_def)
    output = tf.import_graph_def(graph_def, return_elements=['out:0'])
    print(sess.run(output))

参考资料

Is there an example on how to generate protobuf files holding trained Tensorflow graphs

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服务器之家。

原文链接:https://blog.csdn.net/encodets/article/details/54428456

标签:

相关文章

热门资讯

2020微信伤感网名听哭了 让对方看到心疼的伤感网名大全
2020微信伤感网名听哭了 让对方看到心疼的伤感网名大全 2019-12-26
Intellij idea2020永久破解,亲测可用!!!
Intellij idea2020永久破解,亲测可用!!! 2020-07-29
背刺什么意思 网络词语背刺是什么梗
背刺什么意思 网络词语背刺是什么梗 2020-05-22
yue是什么意思 网络流行语yue了是什么梗
yue是什么意思 网络流行语yue了是什么梗 2020-10-11
苹果12mini价格表官网报价 iPhone12mini全版本价格汇总
苹果12mini价格表官网报价 iPhone12mini全版本价格汇总 2020-11-13
返回顶部