-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathextract_torch_t7.py
269 lines (244 loc) · 13.8 KB
/
extract_torch_t7.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
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
import os
import tensorflow as tf
import numpy as np
import cPickle as pickle
import torchfile # pip install torchfile
import resnet
# FLAGS(?)
T7_PATH = './resnet-18.t7'
INIT_CHECKPOINT_DIR = './init'
# Open ResNet-18 torch checkpoint
print('Open ResNet-18 torch checkpoint: %s' % T7_PATH)
o = torchfile.load(T7_PATH)
# Load weights in a brute-force way
print('Load weights in a brute-force way')
conv1_weights = o.modules[0].weight
conv1_bn_gamma = o.modules[1].weight
conv1_bn_beta = o.modules[1].bias
conv1_bn_mean = o.modules[1].running_mean
conv1_bn_var = o.modules[1].running_var
conv2_1_weights_1 = o.modules[4].modules[0].modules[0].modules[0].modules[0].weight
conv2_1_bn_1_gamma = o.modules[4].modules[0].modules[0].modules[0].modules[1].weight
conv2_1_bn_1_beta = o.modules[4].modules[0].modules[0].modules[0].modules[1].bias
conv2_1_bn_1_mean = o.modules[4].modules[0].modules[0].modules[0].modules[1].running_mean
conv2_1_bn_1_var = o.modules[4].modules[0].modules[0].modules[0].modules[1].running_var
conv2_1_weights_2 = o.modules[4].modules[0].modules[0].modules[0].modules[3].weight
conv2_1_bn_2_gamma = o.modules[4].modules[0].modules[0].modules[0].modules[4].weight
conv2_1_bn_2_beta = o.modules[4].modules[0].modules[0].modules[0].modules[4].bias
conv2_1_bn_2_mean = o.modules[4].modules[0].modules[0].modules[0].modules[4].running_mean
conv2_1_bn_2_var = o.modules[4].modules[0].modules[0].modules[0].modules[4].running_var
conv2_2_weights_1 = o.modules[4].modules[1].modules[0].modules[0].modules[0].weight
conv2_2_bn_1_gamma = o.modules[4].modules[1].modules[0].modules[0].modules[1].weight
conv2_2_bn_1_beta = o.modules[4].modules[1].modules[0].modules[0].modules[1].bias
conv2_2_bn_1_mean = o.modules[4].modules[1].modules[0].modules[0].modules[1].running_mean
conv2_2_bn_1_var = o.modules[4].modules[1].modules[0].modules[0].modules[1].running_var
conv2_2_weights_2 = o.modules[4].modules[1].modules[0].modules[0].modules[3].weight
conv2_2_bn_2_gamma = o.modules[4].modules[1].modules[0].modules[0].modules[4].weight
conv2_2_bn_2_beta = o.modules[4].modules[1].modules[0].modules[0].modules[4].bias
conv2_2_bn_2_mean = o.modules[4].modules[1].modules[0].modules[0].modules[4].running_mean
conv2_2_bn_2_var = o.modules[4].modules[1].modules[0].modules[0].modules[4].running_var
conv3_1_weights_skip = o.modules[5].modules[0].modules[0].modules[1].weight
conv3_1_weights_1 = o.modules[5].modules[0].modules[0].modules[0].modules[0].weight
conv3_1_bn_1_gamma = o.modules[5].modules[0].modules[0].modules[0].modules[1].weight
conv3_1_bn_1_beta = o.modules[5].modules[0].modules[0].modules[0].modules[1].bias
conv3_1_bn_1_mean = o.modules[5].modules[0].modules[0].modules[0].modules[1].running_mean
conv3_1_bn_1_var = o.modules[5].modules[0].modules[0].modules[0].modules[1].running_var
conv3_1_weights_2 = o.modules[5].modules[0].modules[0].modules[0].modules[3].weight
conv3_1_bn_2_gamma = o.modules[5].modules[0].modules[0].modules[0].modules[4].weight
conv3_1_bn_2_beta = o.modules[5].modules[0].modules[0].modules[0].modules[4].bias
conv3_1_bn_2_mean = o.modules[5].modules[0].modules[0].modules[0].modules[4].running_mean
conv3_1_bn_2_var = o.modules[5].modules[0].modules[0].modules[0].modules[4].running_var
conv3_2_weights_1 = o.modules[5].modules[1].modules[0].modules[0].modules[0].weight
conv3_2_bn_1_gamma = o.modules[5].modules[1].modules[0].modules[0].modules[1].weight
conv3_2_bn_1_beta = o.modules[5].modules[1].modules[0].modules[0].modules[1].bias
conv3_2_bn_1_mean = o.modules[5].modules[1].modules[0].modules[0].modules[1].running_mean
conv3_2_bn_1_var = o.modules[5].modules[1].modules[0].modules[0].modules[1].running_var
conv3_2_weights_2 = o.modules[5].modules[1].modules[0].modules[0].modules[3].weight
conv3_2_bn_2_gamma = o.modules[5].modules[1].modules[0].modules[0].modules[4].weight
conv3_2_bn_2_beta = o.modules[5].modules[1].modules[0].modules[0].modules[4].bias
conv3_2_bn_2_mean = o.modules[5].modules[1].modules[0].modules[0].modules[4].running_mean
conv3_2_bn_2_var = o.modules[5].modules[1].modules[0].modules[0].modules[4].running_var
conv4_1_weights_skip = o.modules[6].modules[0].modules[0].modules[1].weight
conv4_1_weights_1 = o.modules[6].modules[0].modules[0].modules[0].modules[0].weight
conv4_1_bn_1_gamma = o.modules[6].modules[0].modules[0].modules[0].modules[1].weight
conv4_1_bn_1_beta = o.modules[6].modules[0].modules[0].modules[0].modules[1].bias
conv4_1_bn_1_mean = o.modules[6].modules[0].modules[0].modules[0].modules[1].running_mean
conv4_1_bn_1_var = o.modules[6].modules[0].modules[0].modules[0].modules[1].running_var
conv4_1_weights_2 = o.modules[6].modules[0].modules[0].modules[0].modules[3].weight
conv4_1_bn_2_gamma = o.modules[6].modules[0].modules[0].modules[0].modules[4].weight
conv4_1_bn_2_beta = o.modules[6].modules[0].modules[0].modules[0].modules[4].bias
conv4_1_bn_2_mean = o.modules[6].modules[0].modules[0].modules[0].modules[4].running_mean
conv4_1_bn_2_var = o.modules[6].modules[0].modules[0].modules[0].modules[4].running_var
conv4_2_weights_1 = o.modules[6].modules[1].modules[0].modules[0].modules[0].weight
conv4_2_bn_1_gamma = o.modules[6].modules[1].modules[0].modules[0].modules[1].weight
conv4_2_bn_1_beta = o.modules[6].modules[1].modules[0].modules[0].modules[1].bias
conv4_2_bn_1_mean = o.modules[6].modules[1].modules[0].modules[0].modules[1].running_mean
conv4_2_bn_1_var = o.modules[6].modules[1].modules[0].modules[0].modules[1].running_var
conv4_2_weights_2 = o.modules[6].modules[1].modules[0].modules[0].modules[3].weight
conv4_2_bn_2_gamma = o.modules[6].modules[1].modules[0].modules[0].modules[4].weight
conv4_2_bn_2_beta = o.modules[6].modules[1].modules[0].modules[0].modules[4].bias
conv4_2_bn_2_mean = o.modules[6].modules[1].modules[0].modules[0].modules[4].running_mean
conv4_2_bn_2_var = o.modules[6].modules[1].modules[0].modules[0].modules[4].running_var
conv5_1_weights_skip = o.modules[7].modules[0].modules[0].modules[1].weight
conv5_1_weights_1 = o.modules[7].modules[0].modules[0].modules[0].modules[0].weight
conv5_1_bn_1_gamma = o.modules[7].modules[0].modules[0].modules[0].modules[1].weight
conv5_1_bn_1_beta = o.modules[7].modules[0].modules[0].modules[0].modules[1].bias
conv5_1_bn_1_mean = o.modules[7].modules[0].modules[0].modules[0].modules[1].running_mean
conv5_1_bn_1_var = o.modules[7].modules[0].modules[0].modules[0].modules[1].running_var
conv5_1_weights_2 = o.modules[7].modules[0].modules[0].modules[0].modules[3].weight
conv5_1_bn_2_gamma = o.modules[7].modules[0].modules[0].modules[0].modules[4].weight
conv5_1_bn_2_beta = o.modules[7].modules[0].modules[0].modules[0].modules[4].bias
conv5_1_bn_2_mean = o.modules[7].modules[0].modules[0].modules[0].modules[4].running_mean
conv5_1_bn_2_var = o.modules[7].modules[0].modules[0].modules[0].modules[4].running_var
conv5_2_weights_1 = o.modules[7].modules[1].modules[0].modules[0].modules[0].weight
conv5_2_bn_1_gamma = o.modules[7].modules[1].modules[0].modules[0].modules[1].weight
conv5_2_bn_1_beta = o.modules[7].modules[1].modules[0].modules[0].modules[1].bias
conv5_2_bn_1_mean = o.modules[7].modules[1].modules[0].modules[0].modules[1].running_mean
conv5_2_bn_1_var = o.modules[7].modules[1].modules[0].modules[0].modules[1].running_var
conv5_2_weights_2 = o.modules[7].modules[1].modules[0].modules[0].modules[3].weight
conv5_2_bn_2_gamma = o.modules[7].modules[1].modules[0].modules[0].modules[4].weight
conv5_2_bn_2_beta = o.modules[7].modules[1].modules[0].modules[0].modules[4].bias
conv5_2_bn_2_mean = o.modules[7].modules[1].modules[0].modules[0].modules[4].running_mean
conv5_2_bn_2_var = o.modules[7].modules[1].modules[0].modules[0].modules[4].running_var
fc_weights = o.modules[10].weight
fc_biases = o.modules[10].bias
model_weights_temp = {
'conv1/conv/kernel': conv1_weights,
'conv1/bn/mu': conv1_bn_mean,
'conv1/bn/sigma': conv1_bn_var,
'conv1/bn/beta': conv1_bn_beta,
'conv1/bn/gamma': conv1_bn_gamma,
'conv2_1/conv_1/kernel': conv2_1_weights_1,
'conv2_1/bn_1/mu': conv2_1_bn_1_mean,
'conv2_1/bn_1/sigma': conv2_1_bn_1_var,
'conv2_1/bn_1/beta': conv2_1_bn_1_beta,
'conv2_1/bn_1/gamma': conv2_1_bn_1_gamma,
'conv2_1/conv_2/kernel': conv2_1_weights_2,
'conv2_1/bn_2/mu': conv2_1_bn_2_mean,
'conv2_1/bn_2/sigma': conv2_1_bn_2_var,
'conv2_1/bn_2/beta': conv2_1_bn_2_beta,
'conv2_1/bn_2/gamma': conv2_1_bn_2_gamma,
'conv2_2/conv_1/kernel': conv2_2_weights_1,
'conv2_2/bn_1/mu': conv2_2_bn_1_mean,
'conv2_2/bn_1/sigma': conv2_2_bn_1_var,
'conv2_2/bn_1/beta': conv2_2_bn_1_beta,
'conv2_2/bn_1/gamma': conv2_2_bn_1_gamma,
'conv2_2/conv_2/kernel': conv2_2_weights_2,
'conv2_2/bn_2/mu': conv2_2_bn_2_mean,
'conv2_2/bn_2/sigma': conv2_2_bn_2_var,
'conv2_2/bn_2/beta': conv2_2_bn_2_beta,
'conv2_2/bn_2/gamma': conv2_2_bn_2_gamma,
'conv3_1/shortcut/kernel': conv3_1_weights_skip,
'conv3_1/conv_1/kernel': conv3_1_weights_1,
'conv3_1/bn_1/mu': conv3_1_bn_1_mean,
'conv3_1/bn_1/sigma': conv3_1_bn_1_var,
'conv3_1/bn_1/beta': conv3_1_bn_1_beta,
'conv3_1/bn_1/gamma': conv3_1_bn_1_gamma,
'conv3_1/conv_2/kernel': conv3_1_weights_2,
'conv3_1/bn_2/mu': conv3_1_bn_2_mean,
'conv3_1/bn_2/sigma': conv3_1_bn_2_var,
'conv3_1/bn_2/beta': conv3_1_bn_2_beta,
'conv3_1/bn_2/gamma': conv3_1_bn_2_gamma,
'conv3_2/conv_1/kernel': conv3_2_weights_1,
'conv3_2/bn_1/mu': conv3_2_bn_1_mean,
'conv3_2/bn_1/sigma': conv3_2_bn_1_var,
'conv3_2/bn_1/beta': conv3_2_bn_1_beta,
'conv3_2/bn_1/gamma': conv3_2_bn_1_gamma,
'conv3_2/conv_2/kernel': conv3_2_weights_2,
'conv3_2/bn_2/mu': conv3_2_bn_2_mean,
'conv3_2/bn_2/sigma': conv3_2_bn_2_var,
'conv3_2/bn_2/beta': conv3_2_bn_2_beta,
'conv3_2/bn_2/gamma': conv3_2_bn_2_gamma,
'conv4_1/shortcut/kernel': conv4_1_weights_skip,
'conv4_1/conv_1/kernel': conv4_1_weights_1,
'conv4_1/bn_1/mu': conv4_1_bn_1_mean,
'conv4_1/bn_1/sigma': conv4_1_bn_1_var,
'conv4_1/bn_1/beta': conv4_1_bn_1_beta,
'conv4_1/bn_1/gamma': conv4_1_bn_1_gamma,
'conv4_1/conv_2/kernel': conv4_1_weights_2,
'conv4_1/bn_2/mu': conv4_1_bn_2_mean,
'conv4_1/bn_2/sigma': conv4_1_bn_2_var,
'conv4_1/bn_2/beta': conv4_1_bn_2_beta,
'conv4_1/bn_2/gamma': conv4_1_bn_2_gamma,
'conv4_2/conv_1/kernel': conv4_2_weights_1,
'conv4_2/bn_1/mu': conv4_2_bn_1_mean,
'conv4_2/bn_1/sigma': conv4_2_bn_1_var,
'conv4_2/bn_1/beta': conv4_2_bn_1_beta,
'conv4_2/bn_1/gamma': conv4_2_bn_1_gamma,
'conv4_2/conv_2/kernel': conv4_2_weights_2,
'conv4_2/bn_2/mu': conv4_2_bn_2_mean,
'conv4_2/bn_2/sigma': conv4_2_bn_2_var,
'conv4_2/bn_2/beta': conv4_2_bn_2_beta,
'conv4_2/bn_2/gamma': conv4_2_bn_2_gamma,
'conv5_1/shortcut/kernel': conv5_1_weights_skip,
'conv5_1/conv_1/kernel': conv5_1_weights_1,
'conv5_1/bn_1/mu': conv5_1_bn_1_mean,
'conv5_1/bn_1/sigma': conv5_1_bn_1_var,
'conv5_1/bn_1/beta': conv5_1_bn_1_beta,
'conv5_1/bn_1/gamma': conv5_1_bn_1_gamma,
'conv5_1/conv_2/kernel': conv5_1_weights_2,
'conv5_1/bn_2/mu': conv5_1_bn_2_mean,
'conv5_1/bn_2/sigma': conv5_1_bn_2_var,
'conv5_1/bn_2/beta': conv5_1_bn_2_beta,
'conv5_1/bn_2/gamma': conv5_1_bn_2_gamma,
'conv5_2/conv_1/kernel': conv5_2_weights_1,
'conv5_2/bn_1/mu': conv5_2_bn_1_mean,
'conv5_2/bn_1/sigma': conv5_2_bn_1_var,
'conv5_2/bn_1/beta': conv5_2_bn_1_beta,
'conv5_2/bn_1/gamma': conv5_2_bn_1_gamma,
'conv5_2/conv_2/kernel': conv5_2_weights_2,
'conv5_2/bn_2/mu': conv5_2_bn_2_mean,
'conv5_2/bn_2/sigma': conv5_2_bn_2_var,
'conv5_2/bn_2/beta': conv5_2_bn_2_beta,
'conv5_2/bn_2/gamma': conv5_2_bn_2_gamma,
'logits/fc/weights': fc_weights,
'logits/fc/biases': fc_biases,
}
# Transpose conv and fc weights
model_weights = {}
for k, v in model_weights_temp.items():
if len(v.shape) == 4:
model_weights[k] = np.transpose(v, (2, 3, 1, 0))
elif len(v.shape) == 2:
model_weights[k] = np.transpose(v)
else:
model_weights[k] = v
# Build ResNet-18 model and save parameters
with tf.Graph().as_default():
global_step = tf.Variable(0, trainable=False, name='global_step')
images = [tf.placeholder(tf.float32, [2, 224, 224, 3])]
labels = [tf.placeholder(tf.int32, [2])]
# Build model
print("Build ResNet-18 model")
hp = resnet.HParams(batch_size=2,
num_gpus=1,
num_classes=1000,
weight_decay=0.001,
momentum=0.9,
finetune=False)
network_train = resnet.ResNet(hp, images, labels, global_step, name="train")
network_train.build_model()
print('Number of Weights: %d' % network_train._weights)
print('FLOPs: %d' % network_train._flops)
# Build an initialization operation to run below.
init = tf.global_variables_initializer()
# Start running operations on the Graph.
sess = tf.Session(config=tf.ConfigProto(
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.96),
allow_soft_placement=True,
log_device_placement=False))
sess.run(init)
# Set variables values
print('Set variables to loaded weights')
all_vars = tf.trainable_variables()
for v in all_vars:
print('\t' + v.op.name)
assign_op = v.assign(model_weights[v.op.name])
sess.run(assign_op)
# Save as checkpoint
print('Save as checkpoint: %s' % INIT_CHECKPOINT_DIR)
if not os.path.exists(INIT_CHECKPOINT_DIR):
os.mkdir(INIT_CHECKPOINT_DIR)
saver = tf.train.Saver(tf.global_variables())
saver.save(sess, os.path.join(INIT_CHECKPOINT_DIR, 'model.ckpt'))
print('Done!')