-
Notifications
You must be signed in to change notification settings - Fork 0
/
convert_efficient.py
170 lines (166 loc) · 7.05 KB
/
convert_efficient.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
import torch
import torch.nn as nn
import DenseNet_Efficient
def densenet_264_k32_efficient_model():
model_dict = torch.load('densenet_264_k32.pth')
v1 =[]
k1 =[]
for k,v in model_dict.items():
v1.append(v)
k1.append(k)
# generate efficient model
growth_rate = 32
block_config=(6,12,64,48)
model = DenseNet_Efficient.DenseNetEfficient(num_init_features=64,
growth_rate = growth_rate,
block_config = block_config,
num_classes = 1000,
cifar = False)
k2 =[]
v2 =[]
for key, value in model.state_dict().items():
k2.append(key)
v2.append(value)
pretrained_dict = {}
idx = 0
while(idx < len(k2)):
if k2[idx].endswith('bn.conv_weight'):
pretrained_dict[k2[idx]] = v1[idx+2]
pretrained_dict[k2[idx+1]] = v1[idx]
pretrained_dict[k2[idx+2]] = v1[idx+1]
print('{}: {}->{} {}-->{}'.format(idx,k1[idx+2],k2[idx],v1[idx+2].shape,v2[idx].shape))
print('{}: {}->{} {}-->{}'.format(idx+1,k1[idx],k2[idx+1],v1[idx].shape,v2[idx+1].shape))
print('{}: {}->{} {}-->{}'.format(idx+2, k1[idx+1],k2[idx+2],v1[idx+1].shape,v2[idx+2].shape))
idx += 3
else:
pretrained_dict[k2[idx]] = v1[idx]
print('{}: {}->{} {}-->{}'.format(idx, k1[idx],k2[idx],v1[idx].shape,v2[idx].shape))
idx += 1
model.load_state_dict(pretrained_dict)
torch.save(model.state_dict(),'densenet_264_k32_eff.pth')
return
def densenet_232_k48_efficient_model():
model_dict = torch.load('densenet_232_k48.pth')
v1 =[]
k1 =[]
for k,v in model_dict.items():
v1.append(v)
k1.append(k)
# generate efficient model
growth_rate = 48
block_config=(6,12,48,48)
model = DenseNet_Efficient.DenseNetEfficient(num_init_features=96,
growth_rate = growth_rate,
block_config = block_config,
num_classes = 1000,
cifar = False)
k2 =[]
v2 =[]
for key, value in model.state_dict().items():
k2.append(key)
v2.append(value)
pretrained_dict = {}
idx = 0
while(idx < len(k2)):
if k2[idx].endswith('bn.conv_weight'):
pretrained_dict[k2[idx]] = v1[idx+2]
pretrained_dict[k2[idx+1]] = v1[idx]
pretrained_dict[k2[idx+2]] = v1[idx+1]
print('{}: {}->{} {}-->{}'.format(idx,k1[idx+2],k2[idx],v1[idx+2].shape,v2[idx].shape))
print('{}: {}->{} {}-->{}'.format(idx+1,k1[idx],k2[idx+1],v1[idx].shape,v2[idx+1].shape))
print('{}: {}->{} {}-->{}'.format(idx+2, k1[idx+1],k2[idx+2],v1[idx+1].shape,v2[idx+2].shape))
idx += 3
else:
pretrained_dict[k2[idx]] = v1[idx]
print('{}: {}->{} {}-->{}'.format(idx, k1[idx],k2[idx],v1[idx].shape,v2[idx].shape))
idx += 1
model.load_state_dict(pretrained_dict)
torch.save(model.state_dict(),'densenet_232_k48_eff.pth')
return
def densenet_cosine_264_k32_efficient_model():
model_dict = torch.load('densenet_cosine_264_k32.pth')
v1 =[]
k1 =[]
for k,v in model_dict.items():
v1.append(v)
k1.append(k)
# generate efficient model
growth_rate = 32
block_config=(6,12,64,48)
model = DenseNet_Efficient.DenseNetEfficient(num_init_features=64,
growth_rate = growth_rate,
block_config = block_config,
num_classes = 1000,
cifar = False)
k2 =[]
v2 =[]
for key, value in model.state_dict().items():
k2.append(key)
v2.append(value)
pretrained_dict = {}
idx = 0
while(idx < len(k2)):
if k2[idx].endswith('bn.conv_weight'):
pretrained_dict[k2[idx]] = v1[idx+2]
pretrained_dict[k2[idx+1]] = v1[idx]
pretrained_dict[k2[idx+2]] = v1[idx+1]
print('{}: {}->{} {}-->{}'.format(idx,k1[idx+2],k2[idx],v1[idx+2].shape,v2[idx].shape))
print('{}: {}->{} {}-->{}'.format(idx+1,k1[idx],k2[idx+1],v1[idx].shape,v2[idx+1].shape))
print('{}: {}->{} {}-->{}'.format(idx+2, k1[idx+1],k2[idx+2],v1[idx+1].shape,v2[idx+2].shape))
idx += 3
else:
pretrained_dict[k2[idx]] = v1[idx]
print('{}: {}->{} {}-->{}'.format(idx, k1[idx],k2[idx],v1[idx].shape,v2[idx].shape))
idx += 1
model.load_state_dict(pretrained_dict)
torch.save(model.state_dict(),'densenet_cosine_264_k32_eff.pth')
return
def densenet_cosine_264_k48_efficient_model():
model_dict = torch.load('densenet_cosine_264_k48.pth')
v1 =[]
k1 =[]
for k,v in model_dict.items():
v1.append(v)
k1.append(k)
# generate efficient model
growth_rate = 48
block_config=(6,12,64,48)
model = DenseNet_Efficient.DenseNetEfficient(num_init_features=96,
growth_rate = growth_rate,
block_config = block_config,
num_classes = 1000,
cifar = False)
k2 =[]
v2 =[]
for key, value in model.state_dict().items():
k2.append(key)
v2.append(value)
pretrained_dict = {}
idx = 0
while(idx < len(k2)):
if k2[idx].endswith('bn.conv_weight'):
pretrained_dict[k2[idx]] = v1[idx+2]
pretrained_dict[k2[idx+1]] = v1[idx]
pretrained_dict[k2[idx+2]] = v1[idx+1]
print('{}: {}->{} {}-->{}'.format(idx,k1[idx+2],k2[idx],v1[idx+2].shape,v2[idx].shape))
print('{}: {}->{} {}-->{}'.format(idx+1,k1[idx],k2[idx+1],v1[idx].shape,v2[idx+1].shape))
print('{}: {}->{} {}-->{}'.format(idx+2, k1[idx+1],k2[idx+2],v1[idx+1].shape,v2[idx+2].shape))
idx += 3
else:
pretrained_dict[k2[idx]] = v1[idx]
print('{}: {}->{} {}-->{}'.format(idx, k1[idx],k2[idx],v1[idx].shape,v2[idx].shape))
idx += 1
model.load_state_dict(pretrained_dict)
torch.save(model.state_dict(),'densenet_cosine_264_k48_eff.pth')
return
def main():
####### convert densenet_cosine_264_k48 to efficient model
densenet_cosine_264_k48_efficient_model()
####### convert densenet_cosine_264_k32 to efficient model
# densenet_cosine_264_k32_efficient_model()
####### convert densenet_232_k48 to efficient model
# densenet_232_k48_efficient_model()
####### convert densenet_264_k32 to efficient model
# densenet_264_k32_efficient_model()
if __name__ == '__main__':
main()