forked from JuttaZhang/ASTER
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathSFPres50.py
175 lines (166 loc) · 7.6 KB
/
SFPres50.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
import os
import argparse
import torch
import torch.nn as nn
from torch.autograd import Variable
from torchvision import datasets, transforms
#用于resnet50模型的软剪枝方法
import numpy as np
import sys
sys.path.append('../')
#from models.channel_selection import *
from resnet import *
def softpruning(args, model, thres, end_flag, output_size):
sub_flops = []
total = 0 # out_channel数量
for m in model.modules():
if isinstance(m, nn.BatchNorm2d):
total += m.weight.data.shape[0]
bn = torch.zeros(total) # 把所以bn层的｜weight｜合到同一个tensor中
index = 0
for m in model.modules():
if isinstance(m, nn.BatchNorm2d):
size = m.weight.data.shape[0]
bn[index:(index + size)] = m.weight.data.abs().clone()
index += size
y, i = torch.sort(bn) # y表示排序后的原tensor，i表示排在第几位
y = y.cuda()
thre_index = 0
for i in range(len(y)):
if abs(y[i]) < thres:
thre_index += 1
ratio = thre_index / total
thre = y[thre_index] # 低于thre的全部裁剪掉
pruned = 0
#要有一个打印prune rate的地方8223
cfg_mask = []
for k, m in enumerate(model.modules()):
if isinstance(m, nn.BatchNorm2d):
weight_copy = m.weight.data.clone()
mask = weight_copy.abs().gt(thre).float().cuda()
# torch.sum(mask)统计大于thre的数，mask.shape[0] - torch.sum(mask)为小于thre需要被裁剪掉的数目
pruned = pruned + mask.shape[0] - torch.sum(mask)
# 对于需要被裁剪掉的channel,w,b直接置零
m.weight.data.mul_(mask)
m.bias.data.mul_(mask)
m.running_mean.mul_(mask)
m.running_var.mul_(mask)
cfg_mask.append(mask.clone())
if (int(torch.sum(mask))/mask.shape[0]<=0.05):
end_flag = True
print('layer index: {:d} \t total channel: {:d} \t remaining channel: {:d}'.
format(k, mask.shape[0], int(torch.sum(mask))))
conv_mask = []
linear_mask = []
old_modules = list(model.modules())
layer_id_in_cfg = 0
start_mask = torch.ones(3)
end_mask = cfg_mask[layer_id_in_cfg]
conv_count = 0
conv_id = 0
layers = [8,10,12, 19,21,23, 27,29,31,36,38,40,
47,49,51, 55,57,59, 63,65,67, 72,74,76,
83,85,87, 91,93,95, 99,101,103, 107,109,111,
115,117,119, 124,126,128, 135,137,139, 143,145,147]
for layer_id in range(len(old_modules)):
m0 = old_modules[layer_id]
if isinstance(m0, nn.BatchNorm2d):
# idx1表示mask=1的index
idx1 = np.squeeze(np.argwhere(np.asarray(end_mask.cpu().numpy())))
layer_id_in_cfg += 1
start_mask = end_mask.clone()
if layer_id_in_cfg < len(cfg_mask): # do not change in Final FC
end_mask = cfg_mask[layer_id_in_cfg]
elif isinstance(m0, nn.Conv2d):
output_height, output_width = output_size[conv_id]
if conv_count == 0:
conv_count += 1
#continue
if layer_id in layers:
#if layer_id in []:
#if isinstance(old_modules[layer_id - 1], channel_selection) or isinstance(old_modules[layer_id - 1],
# nn.BatchNorm2d):
# This convers the convolutions in the residual block.
# The convolutions are either after the channel selection layer or after the batch normalization layer.
#print(layer_id,m0,conv_count)
conv_count += 1
idx0 = np.squeeze(np.argwhere(np.asarray(start_mask.cpu().numpy())))
idx1 = np.squeeze(np.argwhere(np.asarray(end_mask.cpu().numpy())))
#print('shape: {} shape:{}'.format(idx0.shape, idx1.shape))
if idx0.shape == ():
inshape = 1
else:
inshape = idx0.shape[0]
if idx1.shape == ():
outshape = 1
else:
outshape = idx1.shape[0]
print('In shape: {:d} Out shape:{:d}'.format(inshape, outshape))
for i in range(m0.weight.data.shape[1]):
if i not in idx0:
m0.weight.data[:, i, :, :] = 0
if conv_count % 3 != 1:
for j in range(m0.weight.data.shape[0]):
if j not in idx1 and i in idx0:
m0.weight.data[j, i, :, :] = 0
mask = m0.weight.data.ne(0).float().cuda()
conv_mask.append(mask.clone())
#print(m0.kernel_size) # [0], [1]
#print(m0.in_channels)
#print(m0.out_channels)
kernel_ops = m0.kernel_size[0] * m0.kernel_size[1] * (inshape / m0.groups)
flops = kernel_ops * outshape * output_height * output_width
else:
kernel_ops = m0.kernel_size[0] * m0.kernel_size[1] * (m0.in_channels / m0.groups)
flops = kernel_ops * m0.out_channels * output_height * output_width
conv_id += 1
sub_flops.append(flops)
elif isinstance(m0, nn.Linear):
idx0 = np.squeeze(np.argwhere(np.asarray(start_mask.cpu().numpy())))
for i in range(m0.weight.data.shape[1]):
if i not in idx0:
m0.weight.data[:, i] = 0
mask = m0.weight.data.ne(0).float().cuda()
linear_mask.append(mask.clone())
flops = idx0.shape[0] * m0.out_features
sub_flops.append(flops)
total_flops = sum(sub_flops)
print("total_flops========>{}".format(total_flops))
return model, total_flops, ratio, cfg_mask, conv_mask, linear_mask, end_flag
def do_Mask(model,cfg_mask,conv_mask,linear_mask):
i = 0
j = 0
l = 0
conv_count = 0
layers = [8, 10, 12, 19, 21, 23, 27, 29, 31, 36, 38, 40,
47, 49, 51, 55, 57, 59, 63, 65, 67, 72, 74, 76,
83, 85, 87, 91, 93, 95, 99, 101, 103, 107, 109, 111,
115, 117, 119, 124, 126, 128, 135, 137, 139, 143, 145, 147]
old_modules = list(model.modules())
for layer_id in range(len(old_modules)):
m = old_modules[layer_id]
if isinstance(m, nn.BatchNorm2d) and i < len(cfg_mask):
#print('m.weight.data shape:{}, bn_mask shape:{} '.format(m.weight.data.shape, cfg_mask[i].shape))
m.weight.data.mul_(cfg_mask[i])
m.bias.data.mul_(cfg_mask[i])
m.running_mean.mul_(cfg_mask[i])
m.running_var.mul_(cfg_mask[i])
i+=1
elif isinstance(m, nn.Conv2d) and j < len(conv_mask):
if conv_count == 0:
conv_count += 1
continue
elif layer_id in layers:
#elif isinstance(old_modules[layer_id-1], channel_selection) or isinstance(old_modules[layer_id-1], nn.BatchNorm2d):
#print('m.weight.data shape:{}, conv_mask shape:{} '.format(m.weight.data.shape, conv_mask[j].shape))
m.weight.data.mul_(conv_mask[j])
if m.bias is not None:
m.bias.data.mul_(conv_mask[j])
j+=1
elif isinstance(m, nn.Linear) and l < len(linear_mask):
#print('m.weight.data shape:{}, linear_mask shape:{} '.format(m.weight.data.shape, linear_mask[l].shape))
m.weight.data.mul_(linear_mask[l])
if l != len(linear_mask)-1:
m.bias.data.mul_(linear_mask[l])
l+=1
return model