-
Notifications
You must be signed in to change notification settings - Fork 31
/
prune.py
276 lines (236 loc) · 12.7 KB
/
prune.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
270
271
272
273
274
275
276
import torch
from torch.autograd import Variable
from torchvision import models
import cv2
import sys
import numpy as np
import math
def replace_layers(model, i, indexes, layers):
if i in indexes:
return layers[indexes.index(i)]
return model[i]
def prune_resnet18_conv_layer(model, layer_index, filter_index):
print("layer_index: ", layer_index)
print("filter_index: ", filter_index)
next_conv = None
next_new_conv = None
downin_conv = None
downout_conv = None
next_downin_conv = None
new_down_conv = None
if layer_index == 0:
_, conv = model.features._modules.items()[layer_index]
next_conv = model.features._modules.items()[4][1][0].conv1
if layer_index%2 == 0:
return model
if layer_index > 0 and layer_index < 5:
tt=1
kt=layer_index//3
pt=layer_index%2
if pt==1:
conv = model.features._modules.items()[3+tt][1][kt].conv1
next_conv = model.features._modules.items()[3+tt][1][kt].conv2
else:
if kt==0:
conv = model.features._modules.items()[3+tt][1][kt].conv2
next_conv = model.features._modules.items()[3+tt][1][kt+1].conv1
else:
conv = model.features._modules.items()[3+tt][1][kt].conv2
next_conv = model.features._modules.items()[3+tt+1][1][0].conv1
downin_conv = model.features._modules.items()[3+tt+1][1][0].downsample[0]
elif layer_index > 4 and layer_index < 9:
tt=2
kt=(layer_index-(tt-1)*4)//3
pt=(layer_index-(tt-1)*4)%2
if pt==1:
conv = model.features._modules.items()[3+tt][1][kt].conv1
next_conv = model.features._modules.items()[3+tt][1][kt].conv2
#downout_conv = model.features._modules.items()[3+tt][1][0].downsample[0]
else:
if kt==0:
conv = model.features._modules.items()[3+tt][1][kt].conv2
next_conv = model.features._modules.items()[3+tt][1][kt+1].conv1
downout_conv = model.features._modules.items()[3+tt][1][kt].downsample[0]
else:
conv = model.features._modules.items()[3+tt][1][kt].conv2
next_conv = model.features._modules.items()[3+tt+1][1][0].conv1
downin_conv = model.features._modules.items()[3+tt+1][1][0].downsample[0]
if layer_index > 8 and layer_index < 13:
tt=3
kt=(layer_index-(tt-1)*4)//3
pt=(layer_index-(tt-1)*4)%2
if pt==1:
conv = model.features._modules.items()[3+tt][1][kt].conv1
next_conv = model.features._modules.items()[3+tt][1][kt].conv2
#downout_conv = model.features._modules.items()[3+tt][1][0].downsample[0]
else:
if kt==0:
conv = model.features._modules.items()[3+tt][1][kt].conv2
next_conv = model.features._modules.items()[3+tt][1][kt+1].conv1
downout_conv = model.features._modules.items()[3+tt][1][kt].downsample[0]
else:
conv = model.features._modules.items()[3+tt][1][kt].conv2
next_conv = model.features._modules.items()[3+tt+1][1][0].conv1
downin_conv = model.features._modules.items()[3+tt+1][1][0].downsample[0]
if layer_index > 12 and layer_index < 17:
tt=4
kt=(layer_index-(tt-1)*4)//3
pt=(layer_index-(tt-1)*4)%2
if pt==1:
conv = model.features._modules.items()[3+tt][1][kt].conv1
next_conv = model.features._modules.items()[3+tt][1][kt].conv2
else:
if kt==0:
conv = model.features._modules.items()[3+tt][1][kt].conv2
next_conv = model.features._modules.items()[3+tt][1][kt+1].conv1
downout_conv = model.features._modules.items()[3+tt][1][kt].downsample[0]
else:
conv = model.features._modules.items()[3+tt][1][kt].conv2
#next_conv = model.features._modules.items()[5+1][1][0].conv1
if layer_index >= 17:
return model
new_conv = \
torch.nn.Conv2d(in_channels = conv.in_channels, \
out_channels = conv.out_channels - 1,
kernel_size = conv.kernel_size, \
stride = conv.stride,
padding = conv.padding,
dilation = conv.dilation,
groups = conv.groups,
bias = conv.bias)
old_weights = conv.weight.data.cpu().numpy()
new_weights = new_conv.weight.data.cpu().numpy()
new_weights[: filter_index, :, :, :] = old_weights[: filter_index, :, :, :]
new_weights[filter_index : , :, :, :] = old_weights[filter_index + 1 :, :, :, :]
new_conv.weight.data = torch.from_numpy(new_weights).cuda()
#bias_numpy = conv.bias.data.cpu().numpy()
#bias = np.zeros(shape = (bias_numpy.shape[0] - 1), dtype = np.float32)
#bias[:filter_index] = bias_numpy[:filter_index]
#bias[filter_index : ] = bias_numpy[filter_index + 1 :]
#new_conv.bias = torch.from_numpy(bias).cuda()
if not downout_conv is None:
new_down_conv = \
torch.nn.Conv2d(in_channels = downout_conv.in_channels, \
out_channels = downout_conv.out_channels - 1,
kernel_size = downout_conv.kernel_size, \
stride = downout_conv.stride,
padding = downout_conv.padding,
dilation = downout_conv.dilation,
groups = downout_conv.groups,
bias = downout_conv.bias)
old_weights = downout_conv.weight.data.cpu().numpy()
new_weights = new_down_conv.weight.data.cpu().numpy()
new_weights[: filter_index, :, :, :] = old_weights[: filter_index, :, :, :]
new_weights[filter_index : , :, :, :] = old_weights[filter_index + 1 :, :, :, :]
new_down_conv.weight.data = torch.from_numpy(new_weights).cuda()
if not next_conv is None:
next_new_conv = \
torch.nn.Conv2d(in_channels = next_conv.in_channels - 1,\
out_channels = next_conv.out_channels, \
kernel_size = next_conv.kernel_size, \
stride = next_conv.stride,
padding = next_conv.padding,
dilation = next_conv.dilation,
groups = next_conv.groups,
bias = next_conv.bias)
old_weights = next_conv.weight.data.cpu().numpy()
new_weights = next_new_conv.weight.data.cpu().numpy()
new_weights[:, : filter_index, :, :] = old_weights[:, : filter_index, :, :]
new_weights[:, filter_index : , :, :] = old_weights[:, filter_index + 1 :, :, :]
next_new_conv.weight.data = torch.from_numpy(new_weights).cuda()
#next_new_conv.bias = next_conv.bias
if not downin_conv is None:
next_downin_conv = \
torch.nn.Conv2d(in_channels = downin_conv.in_channels - 1,\
out_channels = downin_conv.out_channels, \
kernel_size = downin_conv.kernel_size, \
stride = downin_conv.stride,
padding = downin_conv.padding,
dilation = downin_conv.dilation,
groups = downin_conv.groups,
bias = downin_conv.bias)
old_weights = downin_conv.weight.data.cpu().numpy()
new_weights = next_downin_conv.weight.data.cpu().numpy()
new_weights[:, : filter_index, :, :] = old_weights[:, : filter_index, :, :]
new_weights[:, filter_index : , :, :] = old_weights[:, filter_index + 1 :, :, :]
next_downin_conv.weight.data = torch.from_numpy(new_weights).cuda()
if not next_conv is None:
if layer_index ==0:
features1 = torch.nn.Sequential(
*(replace_layers(model.features, i, [layer_index, layer_index], \
[new_conv, new_conv]) for i, _ in enumerate(model.features)))
del model.features
model.features = features1
model.features._modules.items()[4][1][0].conv1 = next_new_conv
else:
if pt==1:
model.features._modules.items()[3+tt][1][kt].conv1 = new_conv
model.features._modules.items()[3+tt][1][kt].conv2 = next_new_conv
#if tt > 1:
#ds = torch.nn.Sequential(
#*(replace_layers(model.features._modules.items()[3+tt][1][0].downsample, i, [0], \
#[new_down_conv]) for i, _ in enumerate(model.features._modules.items()[3+tt][1][0].downsample)))
#model.features._modules.items()[3+tt][1][kt].downsample = ds
#model.features._modules.items()[3+tt][1][kt].downsample[0] = next_downin_conv
else:
if kt==0:
model.features._modules.items()[3+tt][1][kt].conv2 = new_conv
model.features._modules.items()[3+tt][1][kt+1].conv1 = next_new_conv
if tt > 1:
ds = torch.nn.Sequential(
*(replace_layers(model.features._modules.items()[3+tt][1][kt].downsample, i, [0], \
[new_down_conv]) for i, _ in enumerate(model.features._modules.items()[3+tt][1][kt].downsample)))
model.features._modules.items()[3+tt][1][kt].downsample = ds
else:
model.features._modules.items()[3+tt][1][kt].conv2 = new_conv
model.features._modules.items()[3+tt+1][kt][0].conv1 = next_new_conv
#if tt == 1:
ds = torch.nn.Sequential(
*(replace_layers(model.features._modules.items()[3+tt+1][kt][0].downsample, i, [0], \
[next_downin_conv]) for i, _ in enumerate(model.features._modules.items()[3+tt+1][kt][0].downsample)))
model.features._modules.items()[3+tt+1][kt][0].downsample = ds
#model.features._modules.items()[3+tt+1][kt][0].downsample[0] = next_downin_conv
#model.features._modules.items()[3+tt+1][kt][0].downsample = ds
del conv
#model.features = features
else:
#Prunning the last conv layer. This affects the first linear layer of the classifier.
model.features._modules.items()[3+tt][1][kt].conv2 = new_conv
#model.features = torch.nn.Sequential(
#*(replace_layers(model.features, i, [layer_index], \
#[new_conv]) for i, _ in enumerate(model.features)))
layer_index = 0
old_linear_layer = None
for _, module in model.fc._modules.items():
if isinstance(module, torch.nn.Linear):
old_linear_layer = module
break
layer_index = layer_index + 1
if old_linear_layer is None:
raise BaseException("No linear laye found in classifier")
params_per_input_channel = old_linear_layer.in_features / conv.out_channels
new_linear_layer = \
torch.nn.Linear(old_linear_layer.in_features - params_per_input_channel,
old_linear_layer.out_features)
old_weights = old_linear_layer.weight.data.cpu().numpy()
new_weights = new_linear_layer.weight.data.cpu().numpy()
new_weights[:, : filter_index * params_per_input_channel] = \
old_weights[:, : filter_index * params_per_input_channel]
new_weights[:, filter_index * params_per_input_channel :] = \
old_weights[:, (filter_index + 1) * params_per_input_channel :]
new_linear_layer.bias.data = old_linear_layer.bias.data
new_linear_layer.weight.data = torch.from_numpy(new_weights).cuda()
fc = torch.nn.Sequential(
*(replace_layers(model.fc, i, [layer_index], \
[new_linear_layer]) for i, _ in enumerate(model.fc)))
del model.fc
del next_conv
del conv
model.fc = fc
return model
if __name__ == '__main__':
model = models.vgg16(pretrained=True)
model.train()
t0 = time.time()
model = prune_resnet18_conv_layer(model, 28, 10)
print "The prunning took", time.time() - t0