TensorFlow将ckpt模型转为pb模型

# -*- coding: utf-8 -*-
import tensorflow as tf
from tensorflow.python.framework import graph_util
import os
os.environ['CUDA_DEVICE_ORDER'] = 'PCI_BUS_ID'
os.environ['CUDA_VISIBLE_DEVICES'] = '1'
import cv2
import numpy as np


def freeze_graph_test(pb_path, image_path):
    '''
    :param pb_path:pb文件的路径
    :param image_path:测试图片的路径
    :return:
    '''
    with tf.Graph().as_default():
        output_graph_def = tf.GraphDef()
        with open(pb_path, "rb") as f:
            output_graph_def.ParseFromString(f.read())
            tf.import_graph_def(output_graph_def, name="")
        with tf.Session() as sess:
            sess.run(tf.global_variables_initializer())

            # 定义输入的张量名称,对应网络结构的输入张量
            # input:0作为输入图像,keep_prob:0作为dropout的参数,测试时值为1,is_training:0训练参数
            input_image_tensor = sess.graph.get_tensor_by_name("input:0")

            # 定义输出的张量名称
            output_tensor_name = sess.graph.get_tensor_by_name("Sigmoid:0")  # xuNet/Logits/logits

            # 读取测试图片
            im = np.reshape(cv2.imread(image_path, 0), (1, 512, 512, -1))
            # 测试读出来的模型是否正确，注意这里传入的是输出和输入节点的tensor的名字，不是操作节点的名字
            # out=sess.run("InceptionV3/Logits/SpatialSqueeze:0", feed_dict={'input:0': im,'keep_prob:0':1.0,'is_training:0':False})
            out = sess.run(output_tensor_name, feed_dict={input_image_tensor: get_input(im)})
            print("out:{}".format(out))
            score = tf.nn.softmax(out, name='pre')
            class_id = tf.argmax(score, 1)
            pre_class = sess.run(class_id)
            print("pre class_id:{}".format(pre_class))


def get_input(data):
    input = []
    for i in range(64):
        input.append(data)
    return np.reshape(input, (64, 1, 512, 512))


def freeze_graph(input_checkpoint, output_graph):
    '''
    :param input_checkpoint:
    :param output_graph: PB模型保存路径
    :return:
    '''
    # 指定输出的节点名称,该节点名称必须是原模型中存在的节点
    output_node_names = "Sigmoid"
    saver = tf.train.import_meta_graph(input_checkpoint + '.meta', clear_devices=True)
    graph = tf.get_default_graph()  # 获得默认的图
    input_graph_def = graph.as_graph_def()  # 返回一个序列化的图代表当前的图

    with tf.Session() as sess:
        saver.restore(sess, input_checkpoint)  # 恢复图并得到数据
        node_names = [n.name for n in input_graph_def.node]
        print(node_names)
        with open('./nodes.txt', 'a+') as w:
            for node in node_names:
                w.write(node + '\n')

        output_graph_def = graph_util.convert_variables_to_constants(  # 模型持久化，将变量值固定
            sess=sess,
            input_graph_def=input_graph_def,  # 等于:sess.graph_def
            output_node_names=output_node_names.split(","))  # 如果有多个输出节点，以逗号隔开

        with tf.gfile.GFile(output_graph, "wb") as f:  # 保存模型
            f.write(output_graph_def.SerializeToString())  # 序列化输出
        print("%d ops in the final graph." % len(output_graph_def.node))  # 得到当前图有几个操作节点


if __name__ == '__main__':
    # 输入ckpt模型路径
    input_checkpoint = './model.ckpt-10000'
    # 输出pb模型的路径
    out_pb_path = "../model/frozen_model_10000.pb"
    # 调用freeze_graph将ckpt转为pb
    freeze_graph(input_checkpoint, out_pb_path)
    freeze_graph_test(out_pb_path, './1.pgm')

reference1
reference2
reference3