-
Notifications
You must be signed in to change notification settings - Fork 12
/
prune_by_gamma.py
215 lines (185 loc) · 9.76 KB
/
prune_by_gamma.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
# coding:utf-8
import caffe
import numpy as np
import shutil
from copy import deepcopy
# model = "dw_mobilenet_v2/dw_mobilenet_v2_sl_0.001l1_part_right_iter_42500.caffemodel"
# prototxt = "dw_mobilenet_v2_deploy.prototxt"
model = "weights/2_prune0.3_mobilenet_v2.caffemodel"
prototxt = "models/2_prune0.3_mobilenet_v2_deploy.prototxt"
# model = "ResNet50/sl_ResNet50_sl_0.00005l1_right_iter_5000.caffemodel"
# prototxt = "ResNet_50_deploy.prototxt"
caffe.set_mode_gpu()
net = caffe.Net(prototxt, model, caffe.TEST)
percent = 0.3
new_prototxt = prototxt.replace('_mobilenet_v2', '{:.1f}_mobilenet_v2'.format(percent))
shutil.copyfile(prototxt, new_prototxt)
new_net = caffe.Net(new_prototxt, caffe.TEST)
total = 0
for name, layer in zip(net._layer_names, net.layers):
# print(layer.type)
if 'dwise' not in name and 'linear' not in name and layer.type == 'Scale':
total += net.params[name][0].data.shape[0]
# print(total)
bn = np.zeros(total)
index = 0
for name, layer in zip(net._layer_names, net.layers):
if 'dwise' not in name and 'linear' not in name and layer.type == 'Scale':
size = net.params[name][0].data.shape[0]
bn[index:(index+size)] = deepcopy(np.abs(net.params[name][0].data))
index += size
# print(total)
# print(len(bn))
sorted_bn = np.sort(bn)
threshold_index = int(total * percent)
threshold = sorted_bn[threshold_index]
pruned = 0
cfg = []
cfg_mask = []
for k, (name, layer) in enumerate(zip(net._layer_names, net.layers)):
if 'dwise' not in name and 'linear' not in name and layer.type == 'Scale':
weight_copy = deepcopy(np.abs(net.params[name][0].data))
one_index = np.squeeze(np.argwhere(weight_copy > threshold))
mask = np.zeros_like(weight_copy)
mask[one_index] = 1
pruned = pruned + mask.shape[0] - np.sum(mask)
net.params[name][0].data[...] *= mask
net.params[name][1].data[...] *= mask
cfg.append(int(np.sum(mask)))
cfg_mask.append(deepcopy(mask))
print('total channel: {:d} \t remaining channel: {:d} \t prune ratio: {:f}'.
format(mask.shape[0], int(np.sum(mask)), np.sum(mask)/mask.shape[0]))
elif layer.type == 'Scale':
cfg.append(net.params[name][0].data.shape[0])
elif layer.type == 'Pooling':
cfg.append('M')
pruned_ratio = pruned/total
print('Pre-processing Successful!')
print(cfg)
print(threshold_index, threshold)
new_net = caffe.Net(prototxt, model, caffe.TEST)
layer_id_in_cfg = 0
start_mask = np.ones(3) # 一开始的卷积核是3个通道
end_mask = cfg_mask[layer_id_in_cfg]
for name, layer in zip(net._layer_names, net.layers):
if 'dwise' not in name and 'linear' not in name and layer.type == 'Scale':
# print(name)
idx1 = np.squeeze(np.argwhere(np.asarray(end_mask))) # argwhere：找到非0索引；squeeze：去除维数为1的维度
if idx1.size == 1:
idx1 = np.resize(idx1,(1,))
#### prune bn ####
bn_name = name.replace('scale', 'bn')
new_weight = deepcopy(net.params[name][0].data[idx1.tolist()])
new_bias = deepcopy(net.params[name][1].data[idx1.tolist()])
new_running_mean = deepcopy(net.params[bn_name][0].data[idx1.tolist()])
new_running_var = deepcopy(net.params[bn_name][1].data[idx1.tolist()])
new_net.params[name][0].reshape(*new_weight.shape)
new_net.params[name][1].reshape(*new_bias.shape)
new_net.params[bn_name][0].reshape(*new_running_mean.shape)
new_net.params[bn_name][1].reshape(*new_running_var.shape)
new_net.params[name][0].data[...] = deepcopy(new_weight)
new_net.params[name][1].data[...] = deepcopy(new_bias)
new_net.params[bn_name][0].data[...] = deepcopy(new_running_mean)
new_net.params[bn_name][1].data[...] = deepcopy(new_running_var)
#### prune depthwise convolution ####
dwise_conv_name = name.replace('expand/scale', 'dwise')
if 'dwise' in dwise_conv_name and dwise_conv_name in net._layer_names:
new_weight = deepcopy(net.params[dwise_conv_name][0].data[idx1.tolist(), :, :, :])
new_net.params[dwise_conv_name][0].reshape(*new_weight.shape)
new_net.params[dwise_conv_name][0].data[...] = new_weight
output_size = new_weight.shape[0]
prune_ratio = 1 - output_size*1.0/net.params[dwise_conv_name][0].data.shape[0]
print('{:s}\tOut shape {:d}\t prune ratio:{:.2f}'.format(dwise_conv_name, output_size, prune_ratio))
dwise_bn_name = name.replace('expand/scale', 'dwise/bn')
dwise_scale_name = name.replace('expand/scale', 'dwise/scale')
new_weight = deepcopy(net.params[dwise_scale_name][0].data[idx1.tolist()])
new_bias = deepcopy(net.params[dwise_scale_name][1].data[idx1.tolist()])
new_running_mean = deepcopy(net.params[dwise_bn_name][0].data[idx1.tolist()])
new_running_var = deepcopy(net.params[dwise_bn_name][1].data[idx1.tolist()])
new_net.params[dwise_scale_name][0].reshape(*new_weight.shape)
new_net.params[dwise_scale_name][1].reshape(*new_bias.shape)
new_net.params[dwise_bn_name][0].reshape(*new_running_mean.shape)
new_net.params[dwise_bn_name][1].reshape(*new_running_var.shape)
new_net.params[dwise_scale_name][0].data[...] = deepcopy(new_weight)
new_net.params[dwise_scale_name][1].data[...] = deepcopy(new_bias)
new_net.params[dwise_bn_name][0].data[...] = deepcopy(new_running_mean)
new_net.params[dwise_bn_name][1].data[...] = deepcopy(new_running_var)
#### prune output pointwise convolution ####
linear_conv_name = name.replace('expand/scale', 'linear')
if 'linear' in linear_conv_name and linear_conv_name in net._layer_names:
# print(linear_name)
# print(net.params[linear_name][0].data.shape)
new_weight = deepcopy(net.params[linear_conv_name][0].data[:, idx1.tolist(), :, :])
new_net.params[linear_conv_name][0].reshape(*new_weight.shape)
new_net.params[linear_conv_name][0].data[...] = new_weight
output_size = new_weight.shape[0]
prune_ratio = 1 - output_size*1.0/net.params[linear_conv_name][0].data.shape[0]
print('{:s}\tOut shape {:d}\t prune ratio:{:.2f}'.format(linear_conv_name, output_size, prune_ratio))
layer_id_in_cfg += 1
start_mask = deepcopy(end_mask)
if layer_id_in_cfg < len(cfg_mask): # do not change in Final FC
end_mask = cfg_mask[layer_id_in_cfg]
#### prune input pointwise convolution or conv1 ####
elif 'linear' not in name and layer.type == 'Convolution': # dwise为DepthwiseConvolution
idx1 = np.squeeze(np.argwhere(np.asarray(end_mask)))
# print('In shape: {:d}, Out shape {:d}.'.format(idx0.size, idx1.size))
if idx1.size == 1:
idx1 = np.resize(idx1, (1,))
#### mobilenet v2的卷积层没有bias ####
if new_net.params[name][0].shape[0] == 143: # 如果是全连接层，不改变其输出个数
idx0 = np.squeeze(np.argwhere(np.asarray(start_mask)))
if idx0.size == 1:
idx0 = np.resize(idx0, (1,))
new_weight = deepcopy(net.params[name][0].data[:, idx0.tolist(), :, :])
else: # 如果是卷积层
if 'conv2_1/expand' in name: # 上一层是conv1，输出通道数可能会变
idx0 = np.squeeze(np.argwhere(np.asarray(start_mask)))
if idx0.size == 1:
idx0 = np.resize(idx0, (1,))
new_weight = deepcopy(net.params[name][0].data[:, idx0.tolist(), :, :])
new_weight = deepcopy(new_weight[idx1.tolist(), :, :, :])
else:
new_weight = deepcopy(net.params[name][0].data[idx1.tolist(), :, :, :])
new_net.params[name][0].reshape(*new_weight.shape)
new_net.params[name][0].data[...] = new_weight
output_size = new_weight.shape[0]
prune_ratio = 1 - output_size*1.0/net.params[name][0].data.shape[0]
print('{:s}\tOut shape {:d}\t prune ratio:{:.2f}'.format(name, output_size, prune_ratio))
layer_outputs = {}
for name, layer in zip(new_net._layer_names, new_net.layers):
if 'Convolution' in layer.type:
layer_outputs[name] = new_net.params[name][0].data.shape[0]
# print(name, new_net.params[name][0].data.shape)
# print(name, new_net.params[name][0].data.shape[0], net.params[name][0].data.shape[0])
# print(layer_outputs)
with open(new_prototxt, 'r') as f:
lines = f.readlines()
new_prototxt = new_prototxt.replace('_mobilenet_v2', '{:.1f}_mobilenet_v2'.format(percent))
with open(new_prototxt, 'w') as f:
for line in lines:
if 'name:' in line:
layer_name = line.split('"')[1]
if 'num_output' in line:
line = ' num_output: {:d}\n'.format(layer_outputs[layer_name])
if 'group' in line:
line = ' group: {:d}\n'.format(layer_outputs[layer_name])
f.write(line)
train_prototxt = prototxt.replace('deploy', 'train')
with open(train_prototxt, 'r') as f:
lines = f.readlines()
new_train_prototxt = new_prototxt.replace('deploy', 'train')
with open(new_train_prototxt, 'w') as f:
for line in lines:
if 'name:' in line:
layer_name = line.split('"')[1]
if 'num_output' in line:
line = ' num_output: {:d}\n'.format(layer_outputs[layer_name])
if 'group' in line:
line = ' group: {:d}\n'.format(layer_outputs[layer_name])
# if 'l1_lambda' in line:
# continue
# if 'l1_lambda' in line and ('dwise' in layer_name or 'linear' in layer_name):
# line = ' l1_lambda: 0.001\n'
f.write(line)
new_net.save("prune{:.1f}_{:s}".format(percent, model.split('/')[-1]))
print(threshold_index, threshold)