将tensorflow模型打包成PB文件及PB文件读取方式

时间:2022-04-18 01:02:45

1. tensorflow模型文件打包PB文件

?
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
import tensorflow as tf
from tensorflow.python.tools import freeze_graph
 
with tf.Graph().as_default():
  with tf.device("/cpu:0"):
    config = tf.ConfigProto(allow_soft_placement=True)
    with tf.Session(config=config).as_default() as sess:
      model = Your_Model_Name()
      model.build_graph()
      sess.run(tf.initialize_all_variables())
      
      saver = tf.train.Saver()
      ckpt_path = "/your/model/path"
      saver.restore(sess, ckpt_path)
 
      graphdef = tf.get_default_graph().as_graph_def()
      tf.train.write_graph(sess.graph_def,"/your/save/path/","save_name.pb",as_text=False)
      frozen_graph = tf.graph_util.convert_variables_to_constants(sess,graphdef,['output/node/name'])
      frozen_graph_trim = tf.graph_util.remove_training_nodes(frozen_graph)
      freeze_graph.freeze_graph('/your/save/path/save_name.pb','',True, ckpt_path,'output/node/name','save/restore_all','save/Const:0','frozen_name.pb',True,"")

2. PB文件读取使用

?
1
2
3
4
5
6
7
8
9
10
output_graph_def = tf.GraphDef()
with open("your_name.pb","rb") as f:
  output_graph_def.ParseFromString(f.read())
  _ = tf.import_graph_def(output_graph_def, name="")
 
node_in = sess.graph.get_tensor_by_name("input_node_name")
model_out = sess.graph.get_tensor_by_name("out_node_name")
 
feed_dict = {node_in:in_data}
pred = sess.run(model_out, feed_dict)

以上这篇将tensorflow模型打包成PB文件及PB文件读取方式就是小编分享给大家的全部内容了,希望能给大家一个参考,也希望大家多多支持服务器之家。

原文链接:https://blog.csdn.net/hustchenze/article/details/83660960