-
Notifications
You must be signed in to change notification settings - Fork 93
/
Copy pathnew_prune.py
199 lines (192 loc) · 9.59 KB
/
new_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
import torch
import argparse
from yolomodel import *
import torch.nn.functional as F
def arg_parse():
parser = argparse.ArgumentParser(description="YOLO v3 Prune")
parser.add_argument("--cfg",dest="cfgfile",help="网络模型",
default=r"D:/yolotest/cfg/yolov3.cfg",type=str)
parser.add_argument("--weights",dest="weightsfile",help="权重文件",
default=r"D:/yolotest/yolov3.weights",type=str)
parser.add_argument('--percent', type=float, default=0.5,help='剪枝的比例')
return parser.parse_args()
#alpha = 0.1
args = arg_parse()
start = 0
CUDA = torch.cuda.is_available()
print("load network")
model = Darknet(args.cfgfile)
print("done!")
print("load weightsfile")
model.load_weights(args.weightsfile)
if CUDA:
model.cuda()
# 根据shortcut找到不应该被裁的，并记录其序号，放在donntprune中
nnlist = model.module_list
donntprune = dontprune(model)
total = 0
for i in range(len(nnlist)):
if 'conv_with_bn' in list(nnlist[i].named_children())[0][0]:
if i not in donntprune:
total += list(nnlist[i].named_children())[1][1].weight.data.shape[0]
bn = torch.zeros(total)
index = 0
for i in range(len(nnlist)):
if 'conv_with_bn' in list(nnlist[i].named_children())[0][0]:
if i not in donntprune:
size = list(nnlist[i].named_children())[1][1].weight.data.shape[0]
bn[index:(index+size)] = list(nnlist[i].named_children())[1][1].weight.data.abs().clone()
index += size
y, i = torch.sort(bn)
thre_index = int(total * args.percent)
thre = y[thre_index].cuda()
print(y)
pruned = 0
cfg = []
cfg_mask = []
print(thre)
print('--'*30)
print("Pre-processing...")
# 处理bias值
remain_bias_list = dict()
for i in range(len(nnlist)):
if i not in donntprune:
for name in nnlist[i].named_children():
if "_".join(name[0].split("_")[0:-1]) == 'batch_norm':
weight_copy = name[1].weight.data.abs().clone()
mask = weight_copy.gt(thre).float().cuda() # 掩模
if int(torch.sum(mask)) == 0: # 如果该层所有都被剪掉的时候
mask[int(torch.argmax(weight_copy))]=1.
pruned = pruned + mask.shape[0] - torch.sum(mask)
name[1].weight.data.mul_(mask) # 直接修改γ，
cfg.append(int(torch.sum(mask)))
cfg_mask.append(mask.clone())
print('layer index: {:d} \t total channel: {:d} \t remaining channel: {:d}'.
format(i, mask.shape[0], int(torch.sum(mask))))
bias_mask = torch.ones_like(mask) - mask
remain_bias = bias_mask*name[1].bias.data
remain_bias_list[i] = remain_bias
for next_name in nnlist[i+1].named_children():
if "_".join(next_name[0].split("_")[0:-1]) == 'conv_with_bn':
activations = torch.mm(F.relu(remain_bias).view(1,-1),next_name[1].weight.data.sum(dim = [2,3]).transpose(1,0).contiguous())
mean = nnlist[i+1][1].running_mean-activations
mean = mean.view(-1)
nnlist[i + 1][1].running_mean = mean
elif "_".join(next_name[0].split("_")[0:-1]) == 'conv_without_bn':
activations = torch.mm(F.relu(remain_bias).view(1,-1),next_name[1].weight.data.sum(dim = [2,3]).transpose(1,0).contiguous())
bias = next_name[1].bias.data + activations
bias = bias.view(-1)
next_name[1].bias.data = bias
elif next_name[0].split("_")[0] == 'maxpool':
activations = torch.mm(F.relu(remain_bias).view(1, -1) , nnlist[i + 2][0].weight.sum(dim=[2, 3]).transpose(1,0).contiguous())
mean = nnlist[i + 2][1].running_mean - activations
mean = mean.view(-1)
nnlist[i + 2][1].running_mean = mean
elif next_name[0].split("_")[0] == 'reorg':
stride = next_name[1].stride
remain_bias_list[i+1] = torch.squeeze(remain_bias.expand(int(stride*stride),int(remain_bias.size(0))).transpose(1,0).contiguous().view(1,-1))
elif name[0].split("_")[0] == 'route':
prev_1 = name[1].layers[0]+i
have_prev_2 = False
if name[1].layers[1] !=0:
prev_2 = name[1].layers[1] + i
have_prev_2 = True
if isinstance(nnlist[prev_1][0],nn.Conv2d):
if not have_prev_2:
remain_bias = remain_bias_list[prev_1]
else:
remain_bias = torch.cat((remain_bias_list[prev_1],remain_bias_list[prev_2]),0)
activations = torch.mm(F.relu(remain_bias).view(1, -1),nnlist[i + 1][0].weight.sum(dim=[2, 3]).transpose(1, 0).contiguous())
mean = nnlist[i + 1][1].running_mean - activations
mean = mean.view(-1)
nnlist[i + 1][1].running_mean = mean
else:
for name in nnlist[i].named_children():
if "_".join(name[0].split("_")[0:-1]) == 'batch_norm':
dontp = name[1].weight.data.numel()
mask = torch.ones(name[1].weight.data.shape)
print('layer index: {:d} \t total channel: {:d} \t remaining channel: {:d}'.
format(i, dontp, int(dontp)))
cfg.append(int(dontp))
cfg_mask.append(mask.clone())
pruned_ratio = pruned/total
print('Pre-processing Successful!')
print('--'*30)
#print(cfg)
# 写出被减枝的cfg文件
prunecfg = write_cfg(args.cfgfile,cfg)
newmodel = Darknet(prunecfg)
newmodel.header_info = model.header_info
if CUDA:
newmodel.cuda()
old_modules = list(model.modules())
new_modules = list(newmodel.modules())
layer_id_in_cfg = 0
start_mask = torch.ones(3)
end_mask = cfg_mask[layer_id_in_cfg]
print("pruning...")
v=0
for layer_id in range(len(old_modules)):
m0 = old_modules[layer_id]
m1 = new_modules[layer_id]
if isinstance(m0, nn.BatchNorm2d):# 向新模型中写入
idx1 = np.squeeze(np.argwhere(np.asarray(end_mask.cpu().numpy())))
if idx1.size == 1:
idx1 = np.resize(idx1,(1,))
m1.weight.data = m0.weight.data[idx1.tolist()].clone()
m1.bias.data = m0.bias.data[idx1.tolist()].clone() #去掉的bias导致精度大量下降
m1.running_mean = m0.running_mean[idx1.tolist()].clone()
m1.running_var = m0.running_var[idx1.tolist()].clone()
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.Sequential):
for name in m0.named_children():
if name[0].split("_")[0] == 'route':
ind = v+old_modules[layer_id + 1].layers[0]
cfg_mask1 = cfg_mask[route_problem(model, ind)]
if old_modules[layer_id + 1].layers[1]!=0:
ind =v + old_modules[layer_id + 1].layers[1]
cfg_mask1 = cfg_mask1.unsqueeze(0)
cfg_mask2 = cfg_mask[route_problem(model, ind)].unsqueeze(0).cuda()
cfg_mask3 = torch.cat((cfg_mask1,cfg_mask2),1)
cfg_mask1 = cfg_mask3.squeeze(0)
start_mask = cfg_mask1.clone()
elif name[0].split("_")[0] == 'reorg':
stride = name[1].stride
cfg_mask[layer_id_in_cfg - 1] = torch.squeeze(
start_mask.expand(int(stride * stride), int(start_mask.size(0))).transpose(1, 0).contiguous().view(
1, -1))
elif "_".join(name[0].split("_")[0:-1]) == 'conv_with_bn':
idx0 = np.squeeze(np.argwhere(np.asarray(start_mask.cpu().numpy())))
idx1 = np.squeeze(np.argwhere(np.asarray(end_mask.cpu().numpy())))
print('Conv In shape: {:d}, Out shape {:d}.'.format(idx0.size, idx1.size))
if idx0.size == 1:
idx0 = np.resize(idx0, (1,))
if idx1.size == 1:
idx1 = np.resize(idx1, (1,))
w1 = old_modules[layer_id + 1].weight.data[:, idx0.tolist(), :, :].clone()
w1 = w1[idx1.tolist(), :, :, :].clone()
new_modules[layer_id + 1].weight.data = w1.clone()
elif "_".join(name[0].split("_")[0:-1]) == 'conv_without_bn':
idx0 = np.squeeze(np.argwhere(np.asarray(start_mask.cpu().numpy())))
w1 = old_modules[layer_id + 1].weight.data[:, idx0.tolist(), :, :].clone()
new_modules[layer_id + 1].weight.data = w1.clone()
new_modules[layer_id + 1].bias.data = old_modules[layer_id + 1].bias.data.clone()
print('Detect: In shape: {:d}, Out shape {:d}.'.format(new_modules[layer_id + 1].weight.data.size(1),
new_modules[layer_id + 1].weight.data.size(0)))
v=v+1
elif isinstance(m0, nn.Linear):
idx0 = np.squeeze(np.argwhere(np.asarray(start_mask.cpu().numpy())))
if idx0.size == 1:
idx0 = np.resize(idx0, (1,))
m1.weight.data = m0.weight.data[:, idx0].clone()
m1.bias.data = m0.bias.data.clone()
print('--'*30)
print('prune done!')
print('pruned ratio %.3f'%pruned_ratio)
prunedweights = os.path.join('\\'.join(args.weightsfile.split("/")[0:-1]),"prune_"+args.weightsfile.split("/")[-1])
print('save weights file in %s'%prunedweights)
newmodel.save_weights(prunedweights)
print('done!')