-
Notifications
You must be signed in to change notification settings - Fork 136
/
layer_prune.py
211 lines (172 loc) · 8.77 KB
/
layer_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
from models import *
from utils.utils import *
import torch
import numpy as np
from copy import deepcopy
from test import test
from terminaltables import AsciiTable
import time
from utils.utils import *
from utils.prune_utils import *
import argparse
def prune_and_eval(model, prune_shortcuts=[]):
model_copy = deepcopy(model)
for idx in prune_shortcuts:
for i in [idx, idx - 1]:
bn_module = model_copy.module_list[i][1]
mask = torch.zeros(bn_module.weight.data.shape[0]).cuda()
bn_module.weight.data.mul_(mask)
with torch.no_grad():
mAP = eval_model(model_copy)[0][2]
print(f'simply mask the BN Gama of to_be_pruned CBL as zero, now the mAP is {mAP:.4f}')
# %%
def obtain_filters_mask(model, CBL_idx, prune_shortcuts):
filters_mask = []
for idx in CBL_idx:
bn_module = model.module_list[idx][1]
mask = np.ones(bn_module.weight.data.shape[0], dtype='float32')
filters_mask.append(mask.copy())
CBLidx2mask = {idx: mask for idx, mask in zip(CBL_idx, filters_mask)}
for idx in prune_shortcuts:
for i in [idx, idx - 1]:
bn_module = model.module_list[i][1]
mask = np.zeros(bn_module.weight.data.shape[0], dtype='float32')
CBLidx2mask[i] = mask.copy()
return CBLidx2mask
def obtain_avg_forward_time(input, model, repeat=200):
model.eval()
start = time.time()
with torch.no_grad():
for i in range(repeat):
output = model(input)
avg_infer_time = (time.time() - start) / repeat
return avg_infer_time, output
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--cfg', type=str, default='cfg/yolov3-hand.cfg', help='cfg file path')
parser.add_argument('--data', type=str, default='data/oxfordhand.data', help='*.data file path')
parser.add_argument('--weights', type=str, default='weights/last.pt', help='sparse model weights')
parser.add_argument('--shortcuts', type=int, default=8, help='how many shortcut layers will be pruned,\
pruning one shortcut will also prune two CBL,yolov3 has 23 shortcuts')
parser.add_argument('--img-size', type=int, default=416, help='inference size (pixels)')
parser.add_argument('--batch-size', type=int, default=16, help='batch-size')
opt = parser.parse_args()
print(opt)
assert opt.cfg.find("mobilenet") == -1, "Mobilenet doesn't support layer pruning!"
img_size = opt.img_size
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = Darknet(opt.cfg, (img_size, img_size)).to(device)
if opt.weights.endswith(".pt"):
model.load_state_dict(torch.load(opt.weights, map_location=device)['model'])
else:
load_darknet_weights(model, opt.weights)
print('\nloaded weights from ', opt.weights)
eval_model = lambda model: test(model=model, cfg=opt.cfg, data=opt.data, batch_size=opt.batch_size, imgsz=img_size,
rank=-1)
obtain_num_parameters = lambda model: sum([param.nelement() for param in model.parameters()])
with torch.no_grad():
print("\nlet's test the original model first:")
origin_model_metric = eval_model(model)
origin_nparameters = obtain_num_parameters(model)
CBL_idx, Conv_idx, shortcut_idx = parse_module_defs4(model.module_defs)
print('all shortcut_idx:', [i + 1 for i in shortcut_idx])
# highest_thre = torch.zeros(len(shortcut_idx))
# for i, idx in enumerate(shortcut_idx):
# highest_thre[i] = model.module_list[idx][1].weight.data.abs().max().clone()
# _, sorted_index_thre = torch.sort(highest_thre)
# 这里更改了选层策略,由最大值排序改为均值排序,均值一般表现要稍好,但不是绝对,可以自己切换尝试;前面注释的四行为原策略。
bn_mean = torch.zeros(len(shortcut_idx))
for i, idx in enumerate(shortcut_idx):
bn_mean[i] = model.module_list[idx][1].weight.data.abs().mean().clone()
_, sorted_index_thre = torch.sort(bn_mean)
prune_shortcuts = torch.tensor(shortcut_idx)[[sorted_index_thre[:opt.shortcuts]]]
prune_shortcuts = [int(x) for x in prune_shortcuts]
index_all = list(range(len(model.module_defs)))
index_prune = []
for idx in prune_shortcuts:
index_prune.extend([idx - 1, idx, idx + 1])
index_remain = [idx for idx in index_all if idx not in index_prune]
print('These shortcut layers and corresponding CBL will be pruned :', index_prune)
prune_and_eval(model, prune_shortcuts)
CBLidx2mask = obtain_filters_mask(model, CBL_idx, prune_shortcuts)
pruned_model = prune_model_keep_size(model, CBL_idx, CBL_idx, CBLidx2mask)
with torch.no_grad():
mAP = eval_model(pruned_model)[0][2]
print("after transfering the offset of pruned CBL's activation, map is {}".format(mAP))
compact_module_defs = deepcopy(model.module_defs)
for j, module_def in enumerate(compact_module_defs):
if module_def['type'] == 'route':
from_layers = [int(s) for s in module_def['layers']]
if len(from_layers) == 1 and from_layers[0] > 0:
count = 0
for i in index_prune:
if i <= from_layers[0]:
count += 1
from_layers[0] = from_layers[0] - count
# from_layers = str(from_layers[0])
module_def['layers'] = from_layers
elif len(from_layers) == 2:
count = 0
if from_layers[1] > 0:
for i in index_prune:
if i <= from_layers[1]:
count += 1
from_layers[1] = from_layers[1] - count
else:
for i in index_prune:
if i > j + from_layers[1] and i < j:
count += 1
from_layers[1] = from_layers[1] + count
# from_layers = ', '.join([str(s) for s in from_layers])
module_def['layers'] = from_layers
compact_module_defs = [compact_module_defs[i] for i in index_remain]
compact_model = Darknet([model.hyperparams.copy()] + compact_module_defs, (img_size, img_size)).to(device)
compact_nparameters = obtain_num_parameters(compact_model)
# init_weights_from_loose_model(compact_model, pruned_model, CBL_idx, Conv_idx, CBLidx2mask)
random_input = torch.rand((1, 3, img_size, img_size)).to(device)
pruned_forward_time, pruned_output = obtain_avg_forward_time(random_input, pruned_model)
compact_forward_time, compact_output = obtain_avg_forward_time(random_input, compact_model)
# 在测试集上测试剪枝后的模型, 并统计模型的参数数量
with torch.no_grad():
compact_model_metric = eval_model(compact_model)
# 比较剪枝前后参数数量的变化、指标性能的变化
metric_table = [
["Metric", "Before", "After"],
["mAP", f'{origin_model_metric[0][2]:.6f}', f'{compact_model_metric[0][2]:.6f}'],
["Parameters", f"{origin_nparameters}", f"{compact_nparameters}"],
["Inference", f'{pruned_forward_time:.4f}', f'{compact_forward_time:.4f}']
]
print(AsciiTable(metric_table).table)
# 生成剪枝后的cfg文件并保存模型
pruned_cfg_name = opt.cfg.replace('/', f'/layer_prune_{opt.shortcuts}_shortcut_')
# 创建存储目录
dir_name = pruned_cfg_name.split('/')[0] + '/' + pruned_cfg_name.split('/')[1]
if not os.path.isdir(dir_name):
os.makedirs(dir_name)
# 由于原始的compact_module_defs将anchor从字符串变为了数组,因此这里将anchors重新变为字符串
file = open(opt.cfg, 'r')
lines = file.read().split('\n')
for line in lines:
if line.split(' = ')[0] == 'anchors':
anchor = line.split(' = ')[1]
break
if line.split('=')[0] == 'anchors':
anchor = line.split('=')[1]
break
file.close()
for item in compact_module_defs:
if item['type'] == 'shortcut':
item['from'] = str(item['from'][0])
elif item['type'] == 'route':
item['layers'] = ",".join('%s' % i for i in item['layers'])
elif item['type'] == 'yolo':
item['mask'] = ",".join('%s' % i for i in item['mask'])
item['anchors'] = anchor
pruned_cfg_file = write_cfg(pruned_cfg_name, [model.hyperparams.copy()] + compact_module_defs)
print(f'Config file has been saved: {pruned_cfg_file}')
weights_dir_name = dir_name.replace('cfg', 'weights')
if not os.path.isdir(weights_dir_name):
os.makedirs(weights_dir_name)
compact_model_name = weights_dir_name + f'/layer_prune_{str(opt.shortcuts)}_shortcuts.weights'
save_weights(compact_model, path=compact_model_name)
print(f'Compact model has been saved: {compact_model_name}')