参考连接: https://www.yuque.com/g/jesse-ztr2k/nkke46/ss4rlv/collaborator/join?token=XUVZNORisVWEWyst#
注意有些时候需要添加一个pb文件。 而不是tensorflow 提供的save 方法生成的一个目录里面包含了若干pb文件。
load时候直接填写这个目录即可。 但是有些时候需要合成一个pb文件。
1 目录结构
-assets
-variables
-variables.data-00000-of-00001
-variables.index
-saved_model.pb
2 作用
其中 variables 记录模型参数 , pb文件记录模型结构
tf2 都是保存的 权重和 结构分开的, 如果需要兼容tf V1的代码,即导入一个pb文件,就需要 1 )保存常量计算图 2)frozen graph pb格式。
环境准备:
tensorflow==1.15, tf-slim==1.1.0
https://github.com/tensorflow/models/tree/master/research/slim
注意 一定在tf v1 环境下生成pb
1 import cv2
2 import numpy as np
3 import tensorflow as tf
4 import os
5 from tensorflow.python.framework import graph_util
6
7 # 参考连接 https://blog.csdn.net/tensorflowforum/article/details/112352764 代码
8 # 参考连接 参数详解:https://blog.csdn.net/weixin_43529465/article/details/124721583
9 # https://blog.csdn.net/rain6789/article/details/78754516
10
11 class SingleCnn(tf.keras.Model):
12 def __init__(self):
13 super(SingleCnn, self).__init__()
14 # filters=1 卷积核数目,相当于卷积核的channel
15 self.conv = tf.keras.layers.Conv2D(filters=1,
16 kernel_size=[1, 1],
17 # valid表示不填充, same表示合理填充
18 padding='valid',
19 # data_format='channels_last',-> 表示HWC,输入可以定义批次
20 data_format='channels_last',
21 use_bias=False,
22 kernel_initializer=tf.keras.initializers.he_uniform(seed=None),
23 name="conv")
24
25 def call(self, inputs):
26 x = self.conv(inputs)
27 return x
28 if __name__ == "__main__":
29 # 构建场景输入数据
30
31 # images=tf.random.uniform((1, 300, 300, 3))
32
33 # 图像数据
34 imagefile = r"catanddog\cat\5.JPG"
35 img = cv2.imread(imagefile)
36 img = cv2.resize(img, (64, 64))
37 img = np.expand_dims(img, axis=0)
38 print(img.shape, type(img), img.dtype)
39
40 # 未量化的model不支持int32和int8
41 # img = img.astype(np.int32)
42 img = tf.convert_to_tensor(img, np.float32)
43 print(img.shape, type(img), img.dtype)
44 singlecnn = SingleCnn()
45
46 output = singlecnn(img)
47 print(output.shape, type(output))
48 print(output[0][2:10][2:6])
49 # =========== ckpt保存 with session的写法tf2 已不再使用 ===========
50 # with tf.Session(graph=tf.Graph()) as sess:
51 # constant_graph = graph_util.convert_variables_to_constants(sess, sess.graph_def, ['op_to_store'])
52
53 # 保存参考 https://zhuanlan.zhihu.com/p/146243327
54 # save_format='tf' 代表保存pb
55 # singlecnn.save('./pbmodel/singlecnn', save_format='tf')
56 # tf.saved_model.save(singlecnn, './pbmodel/singlecnn')
57 tf.keras.models.save_model(singlecnn, './pbmodel/singlecnn_0',
58 save_format="tf",
59 include_optimizer=False, save_traces=False)
60
61 # 加载模型 验证可以加载
62 new_model = tf.keras.models.load_model('./pbmodel/singlecnn_0', compile=False)
63 # new_model = tf.saved_model.load('./pbmodel/singlecnn_0')
64 # output_ = new_model(img)
65 # # print(output_.shape, output_[0][2:6][2:6])
66 # print(output_.shape)
67 #
68 # 查看结构
69 new_model.summary()
70
71 # print("----------------")
72 # # 加载模型
73 # saved_model = tf.saved_model.load('./pbmodel/singlecnn_0')
74 # # 将模型转换为pb格式 还是目录方法。
75 # converter = tf.saved_model.save(saved_model, "model.pb")
76
77 def change_pb(pretrained_model):
78 """tf v1 选用tf1 跑这个脚本生成pb"""
79 from tensorflow.python.framework.convert_to_constants import convert_variables_to_constants_v2
80 # 重点
81 # Convert Keras model to ConcreteFunction
82 # MobileNet is a function
83 full_model = tf.function(lambda x: pretrained_model(x))
84
85 # 指定shape和dtype对tf function进行重新追踪
86 full_model = full_model.get_concrete_function(
87 tf.TensorSpec(pretrained_model.inputs[0].shape, pretrained_model.inputs[0].dtype))
88
89 # Get frozen ConcreteFunction,将计算图中的变量及其取值通过常量的方式保存
90 frozen_func = convert_variables_to_constants_v2(full_model)
91 frozen_func.graph.as_graph_def()
92
93 layers = [op.name for op in frozen_func.graph.get_operations()]
94 print("-" * 50)
95 print("Frozen model layers: ")
96 for layer in layers:
97 print(layer)
98
99 print("-" * 50)
100 print("Frozen model inputs: ")
101 print(frozen_func.inputs)
102 print("Frozen model outputs: ")
103 print(frozen_func.outputs)
104
105 # Save frozen graph from frozen ConcreteFunction to hard drive
106 # as_text: If True, writes the graph as an ASCII proto; otherwise, The graph is written as a text proto
107 tf.io.write_graph(graph_or_graph_def=frozen_func.graph,
108 logdir="./frozen_models",
109 name="frozen_graph.pb",
110 as_text=True)
111
112
113 change_pb(new_model)
model_getpb
python download_and_convert_data.py --dataset_name=flowers --dataset_dir="tmp/dataset"
手机扫一扫
移动阅读更方便
你可能感兴趣的文章