久久久99999,久久久久久夜,亚洲伊人久久影院

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文件讀取使用

									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)