在Tengine上面验证PB模型是否能正常运行

Tengine的官方代码：https://github.com/OAID/Tengine 

验证TensorFlow的模型，只需要修改掉test_tf_mobilenet测试程序的模型文件格式即可 ，流程如下：

1，修改test_tf_mobilenet文件的如下部分 

const char* model_file = "./models/Test_Model.pb";

2，将Test_Model.pb拷贝到Tengine的主目录的models目录下。

3，修改makefile.config文件，使能CONFIG_TF_SERIALIZER，然后编译 ，编译参照：https://github.com/OAID/Tengine/blob/master/doc/install.md 

4，编译完成之后运行：

./build/tests/bin/test_tf_mobilenet 即可。

5，用TensorFlow框架进行验证的Python代码

说明：              如下脚本可直接测试PB模型是否正常。

该脚本输入：

1. 输入tensor和输出tensor
2. 测试图片
3. PB格式的模型文件

获取输入输出tensor的方式有两种：

1. 通过工具：TensorFlow的自带的工具summarize_graph 
2. 通过脚本列出所有的tensor ，如附件Show_All_Tensor.py
3. 通过可视化工具Netron来进行查看。

import tensorflow as tf import  numpy as np import PIL.Image as Image from skimage import io, transform   #model_name = 'D:/Project/Tengine-D/TF_pb/classify_image_graph_def.pb' image_name = 'D:/Project/Tengine-D/TF_pb/ssd_dog.jpg'   #model_name = 'classify_image_graph_def.pb' #input_tensor = 'DecodeJpeg/contents:0' #output_tensor = 'softmax:0'   model_name = 'frozen_dcscn.pb' input_tensor = 'Placeholder:0' output_tensor = 'srnetwork/Reshape:0'   def recognize(jpg_path, pb_file_path):     with tf.Graph().as_default():         output_graph_def = tf.GraphDef()           with open(pb_file_path, "rb") as f:             output_graph_def.ParseFromString(f.read())             _ = tf.import_graph_def(output_graph_def, name="")           with tf.Session() as sess:             init = tf.global_variables_initializer()             sess.run(init)               input_x = sess.graph.get_tensor_by_name(input_tensor)             print(input_x)             out_softmax = sess.graph.get_tensor_by_name(output_tensor)             print(out_softmax)           #  out_label = sess.graph.get_tensor_by_name("output:0")           #  print(out_label)               img = io.imread(jpg_path)             img = transform.resize(img, (240, 426, 3))             img_out_softmax = sess.run(out_softmax, feed_dict={input_x:np.reshape(img, [-1, 240, 426, 3])})               print("img_out_softmax:",img_out_softmax)             prediction_labels = np.argmax(img_out_softmax, axis=1)             print("label:",prediction_labels)   recognize(image_name, model_name)

Show_All_Tensor.py

# -*- coding: utf-8 -*-   import tensorflow as tf import os   #model_dir = 'C:/Users/Dexter/Documents/ML_files/171009_Retrain_with_Inception_v3' #model_name = 'output_graph.pb'   model_dir = 'D:/Project/Tengine-D/TF_pb/' #model_name = 'frozen_dcscn.pb' model_name = 'frozen_mobilenet_v1_224.pb'   # 读取并创建一个图graph来存放Google训练好的Inception_v3模型（函数） def create_graph():     with tf.gfile.FastGFile(os.path.join(             model_dir, model_name), 'rb') as f:         # 使用tf.GraphDef()定义一个空的Graph         graph_def = tf.GraphDef()         graph_def.ParseFromString(f.read())         # Imports the graph from graph_def into the current default Graph.         tf.import_graph_def(graph_def, name='')   # 创建graph create_graph()   tensor_name_list = [tensor.name for tensor in tf.get_default_graph().as_graph_def().node] for tensor_name in tensor_name_list:     print(tensor_name,'\n')