将tensorflow模型打包成PB文件及PB文件读取方式

1. tensorflow模型文件打包成PB文件

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

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)

来源：https://blog.csdn.net/hustchenze/article/details/83660960