1 import tensorflow as tf
2 import onnx
3 import onnxsim
4 import numpy as np
5 import torch
6 from model.facedetector_model import mobilenetv2_yolov3
7
8 #提取pb模型中的参数
9 def extract_params_from_pb():
10 constant_values = {}
11 with tf.compat.v1.Session() as sess:
12 with tf.io.gfile.GFile('model/FaceDetector.pb', 'rb') as f:
13 graph_def = tf.compat.v1.GraphDef()
14 graph_def.ParseFromString(f.read())
15 sess.graph.as_default()
16 tf.import_graph_def(graph_def, name='')
17 # # input
18 # input_x = sess.graph.get_tensor_by_name('input/input_data:0')
19 # # output
20 # output = sess.graph.get_tensor_by_name('pred_bbox/Reshape:0')
21 # sess.run(output, feed_dict={'input/input_data:0': inputimage})
22
23 constant_ops = [op for op in sess.graph.get_operations()]#[op for op in sess.graph.get_operations() if op.type == "Const"]
24 for constant_op in constant_ops:
25 if constant_op.op_def.name == "Const":
26 if "Shape" in constant_op.name or "pred" in constant_op.name:
27 continue
28 constant_values[constant_op.name] = sess.run(constant_op.outputs[0])
29 return constant_values
30
31 #过滤提取出来的params
32 def filter_params(constant_values):
33 total = 0
34 prompt = []
35 res = {}
36 forbidden = ['shape','stack']
37
38 for k,v in constant_values.items():
39 # filtering some by checking ndim and name
40 if v.ndim<1: continue
41 if v.ndim==1:
42 token = k.split(r'/')[-1]
43 flag = False
44 for word in forbidden:
45 if token.find(word)!=-1:
46 flag = True
47 break
48 if flag:
49 continue
50
51 shape = v.shape
52 cnt = 1
53 for dim in shape:
54 cnt *= dim
55 prompt.append('{} with shape {} has {}'.format(k, shape, cnt))
56 res[k] = v
57 print(prompt[-1])
58 total += cnt
59 prompt.append('totaling {}'.format(total))
60 # print(prompt[-1])
61 return res
62
63 #将Tensorflow的张量转换成PyTorch的张量
64 def trans_tensor_pb2pth(k,a):
65
66 v = tf.convert_to_tensor(a).numpy()
67 # tensorflow weights to pytorch weights
68 if len(v.shape) == 4:
69 if "depthwise_weights" in k:#防止深度可分离卷积
70 return np.ascontiguousarray(v.transpose(2,3,0,1))
71 return np.ascontiguousarray(v.transpose(3,2,0,1))
72 elif len(v.shape) == 2:
73 return np.ascontiguousarray(v.transpose())
74 return v
75
76 #将pb的对应params名字转换为pth对应参数名
77 def trans_name_pb2pth(trans_weights):
78 model_dict = {}
79 for name,para in trans_weights.items():
80 name = name.replace('/',".")
81
82 if "MobilenetV2.Conv" in name:#处理MobilenetV2.Conv
83 name = name.replace('weights',"0.weight")
84 name = name.replace('BatchNorm',"1")
85 name = name.replace('gamma',"weight")
86 name = name.replace('beta',"bias")
87 name = name.replace('moving_mean',"running_mean")
88 name = name.replace('moving_variance',"running_var")
89 elif "MobilenetV2.expanded_conv." in name:#处理MobilenetV2.expanded_conv.
90 name = name.replace('depthwise.',"0.")
91 name = name.replace('project',"1")
92 name = name.replace('depthwise_weights',"0.weight")
93 name = name.replace('weights',"0.weight")
94 name = name.replace('BatchNorm',"1")
95 name = name.replace('gamma',"weight")
96 name = name.replace('beta',"bias")
97 name = name.replace('moving_mean',"running_mean")
98 name = name.replace('moving_variance',"running_var")
99 elif "MobilenetV2.expanded_conv_" in name:#处理MobilenetV2.expanded_conv_*
100 name = name.replace('expand.',"0.")
101 name = name.replace('depthwise.',"1.")
102 name = name.replace('project',"2")
103 name = name.replace('depthwise_weights',"0.weight")
104 name = name.replace('weights',"0.weight")
105 name = name.replace('BatchNorm',"1")
106 name = name.replace('gamma',"weight")
107 name = name.replace('beta',"bias")
108 name = name.replace('moving_mean',"running_mean")
109 name = name.replace('moving_variance',"running_var")
110 elif "yolo-v3" in name:
111 if "bbox" in name:
112 continue
113 name = name.replace('yolo-v3',"yolo_v3")
114 name = name.replace('weight',"0.weight")
115 name = name.replace('kernel',"weight")
116 name = name.replace('batch_normalization',"1")
117 name = name.replace('gamma',"weight")
118 name = name.replace('beta',"bias")
119 name = name.replace('moving_mean',"running_mean")
120 name = name.replace('moving_variance',"running_var")
121 print(name)
122 model_dict[name] = torch.Tensor(para)
123 return model_dict
124
125 #将pb参数copy给pth模型
126 def copy_pbParams2pthParams():
127 constant_values = extract_params_from_pb()
128 TF_weights = filter_params(constant_values)
129 trans_weights = {k:trans_tensor_pb2pth(k,v) for (k, v) in TF_weights.items() }
130
131 #创建pytorch模型
132 PyTorchModel = mobilenetv2_yolov3()
133 model_dict = trans_name_pb2pth(trans_weights)
134 # model_dict = PyTorchModel.state_dict()
135 # for name in model_dict.keys():
136 # print(name)
137 PyTorchModel.load_state_dict(model_dict)
138 PyTorchModel.cuda().eval()
139 dummy_input = torch.rand(1,1,224,224,device="cuda").float()
140 # out = PyTorchModel(dummy_input)
141 torch.onnx.export(PyTorchModel,dummy_input,"P3mNet.onnx",verbose = True,opset_version = 11)
142 print("====> Simplifying...")
143 model_opt,_ = onnxsim.simplify("P3mNet.onnx")
144 onnx.save(model_opt, 'P3mNet_sim.onnx')
145 print("onnx model simplify Ok!")
146 copy_pbParams2pthParams()
标签：sess,constant,name,graph,模型,torch,pb,import,op From： https://www.cnblogs.com/zukang/p/16778294.html