-
Notifications
You must be signed in to change notification settings - Fork 0
/
channel_layer_prune.py
174 lines (145 loc) · 8.05 KB
/
channel_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
from models.pose_model import PoseModel
from utils.prune_utils import *
from models.utils.utils import write_cfg
from tester import Tester
import torch
from utils.utils import get_option_path, get_corresponding_cfg
posenet = PoseModel(device="cpu")
class ChannelLayerPruner:
def __init__(self, model_path, model_cfg, compact_model_path="", compact_model_cfg=""):
self.model_path = model_path
self.model_cfg = model_cfg
posenet.build(model_cfg)
self.model = posenet.model
posenet.load(model_path)
self.backbone = posenet.backbone
self.kps = posenet.kps
self.se_ratio = posenet.se_ratio
self.block_num = posenet.block_nums
self.first_conv = posenet.first_conv
self.residual = posenet.residual
self.head_channel = posenet.head_channel
self.head_type = posenet.head
if not compact_model_path or not compact_model_cfg:
self.compact_model_path = "buffer/all_{}.pth".format(self.backbone)
self.compact_model_cfg = "buffer/cfg_all_{}.json".format(self.backbone)
else:
self.compact_model_path = compact_model_path
self.compact_model_cfg = compact_model_cfg
if self.backbone == "seresnet50" or self.backbone == "seresnet101":
from utils.prune_utils import obtain_prune_idx_50 as obtain_prune
self.init_weight_channel = init_weights_from_loose_model_shortcut50
self.init_weight_layer = init_weights_from_loose_model_layer
else:
raise ValueError("{} is not supported for layer pruning! ".format(self.backbone))
self.obtain_prune_idx = obtain_prune
def obtain_block_idx(self, shortcut_idx, prune_shortcuts):
shortcuts_location = []
candidate_block_num = [i-1 for i in self.block_num]
for shortcut in prune_shortcuts:
shortcuts_location.append(self.get_layer_block(shortcut_idx.index(shortcut), candidate_block_num))
return shortcuts_location
@staticmethod
def get_layer_block(idx, candidate_block_num):
for i in range(len(candidate_block_num)):
if idx < sum(candidate_block_num[:i+1]):
return i
raise ValueError("Wrong index! ")
def run(self, threshold, layer_num):
print("------------------------- Prune channels first -------------------------------")
all_bn_id, normal_idx, shortcut_idx, downsample_idx, head_idx = self.obtain_prune_idx(self.model)
prune_idx = all_bn_id
sorted_bn = sort_bn(self.model, prune_idx)
threshold = obtain_bn_threshold(self.model, sorted_bn, threshold / 100)
pruned_filters, pruned_maskers = obtain_filters_mask(self.model, prune_idx, threshold)
CBLidx2mask_channel = {idx - 1: mask.astype('float32') for idx, mask in zip(all_bn_id, pruned_maskers)}
CBLidx2filter = {idx - 1: filter_num for idx, filter_num in zip(all_bn_id, pruned_filters)}
final_layer_groups = [downsample_idx[-1] - 1] + [shortcut_idx[-1] - 1, shortcut_idx[-2] - 1]
mask_groups = [
[shortcut_idx[sum(self.block_num[:0]) + i] for i in range(self.block_num[0])] + [downsample_idx[0]],
[shortcut_idx[sum(self.block_num[:1]) + i] for i in range(self.block_num[1])] + [downsample_idx[1]],
[shortcut_idx[sum(self.block_num[:2]) + i] for i in range(self.block_num[2])] + [downsample_idx[2]],
[shortcut_idx[sum(self.block_num[:3]) + i] for i in range(self.block_num[3])] + [downsample_idx[3]]]
if self.backbone == "seresnet50" or self.backbone == "seresnet101":
final_layer_groups.append(shortcut_idx[-3]-1)
merge_mask(CBLidx2mask_channel, CBLidx2filter, mask_groups)
adjust_final_mask(CBLidx2mask_channel, CBLidx2filter, self.model, final_layer_groups)
for head in head_idx:
adjust_mask(CBLidx2mask_channel, CBLidx2filter, self.model, head)
valid_filter = {k: v for k, v in CBLidx2filter.items() if k + 1 in prune_idx}
channel_str = ",".join(map(lambda x: str(x), valid_filter.values()))
print(channel_str, file=open("buffer/cfg_all_{}.txt".format(self.backbone), "w"))
m_channel_cfg = {
'backbone': self.backbone,
'kps': self.kps,
'se_ratio': self.se_ratio,
"first_conv": valid_filter[all_bn_id[0] - 1],
'residual': get_residual_channel([filt for _, filt in valid_filter.items()], self.backbone),
'channels': get_channel_dict([filt for _, filt in valid_filter.items()], self.backbone),
"head_type": "pixel_shuffle",
"head_channel": [CBLidx2filter[i - 1] for i in head_idx]
}
write_cfg(m_channel_cfg, self.compact_model_cfg)
posenet.build(self.compact_model_cfg)
compact_channel_model = posenet.model
self.init_weight_channel(compact_channel_model, self.model, CBLidx2mask_channel, valid_filter, downsample_idx,
head_idx)
print("\n------------------------- Prune layers -------------------------------")
all_bn_id, other_idx, shortcut_idx, downsample_idx = obtain_prune_idx_layer(self.model)
bn_mean = torch.zeros(len(shortcut_idx))
for i, idx in enumerate(shortcut_idx):
bn_mean[i] = list(self.model.named_modules())[idx][1].weight.data.abs().mean().clone()
_, sorted_index_thre = torch.sort(bn_mean)
prune_shortcuts = torch.tensor(shortcut_idx)[[sorted_index_thre[:layer_num]]]
prune_shortcuts = [int(x) for x in prune_shortcuts]
print_mean(bn_mean, shortcut_idx, prune_shortcuts)
prune_layers = []
for prune_shortcut in prune_shortcuts:
target_idx = all_bn_id.index(prune_shortcut)
for i in range(3):
prune_layers.append(all_bn_id[target_idx - i])
CBLidx2mask_layer = obtain_layer_filters_mask(compact_channel_model, all_bn_id, prune_layers)
pruned_locations = self.obtain_block_idx(shortcut_idx, prune_shortcuts)
blocks = self.block_num
for pruned_location in pruned_locations:
blocks[pruned_location] -= 1
m_layer_cfg = {
'backbone': self.backbone,
'kps': self.kps,
'se_ratio': self.se_ratio,
"first_conv": m_channel_cfg["first_conv"],
'residual': m_channel_cfg["residual"],
'channels': obtain_all_prune_channels(sorted_index_thre[:layer_num].tolist(), m_channel_cfg["channels"]),
"head_type": self.head_type,
"head_channel": m_channel_cfg["head_channel"],
}
write_cfg(m_layer_cfg, self.compact_model_cfg)
posenet.build(self.compact_model_cfg)
compact_layer_model = posenet.model
# compact_all_bn = [idx for idx, mod in enumerate(list(compact_model.named_modules()))
# if isinstance(mod[1], torch.nn.BatchNorm2d)]
compact_all_bn_idx, compact_other_idx, compact_shortcut_idx, compact_downsample_idx = \
obtain_prune_idx_layer(compact_layer_model)
init_weights_from_loose_model_layer(compact_layer_model, compact_channel_model, CBLidx2mask_layer, compact_all_bn_idx)
torch.save(compact_layer_model.state_dict(), self.compact_model_path)
def test(self, data_cfg, dataset_name):
print("Testing with dataset {}".format(dataset_name[0].keys()))
tester = Tester(self.compact_model_cfg, self.compact_model_path, dataset_name, data_cfg)
tester.test()
tester.get_benchmark()
return tester.summarize()
if __name__ == '__main__':
model_path = "weights/pose_prune_asset/resnet50/latest.pth"
model_cfg = ""
data_cfg = ""
if not model_cfg or not data_cfg:
model_cfg, data_cfg, _ = get_corresponding_cfg(model_path, check_exist=["data", "model"])
thresh = 80
layer_num = 2
CLP = ChannelLayerPruner(model_path, model_cfg)
CLP.run(thresh, layer_num)
dataset_name = "coco"
if data_cfg:
from config.config import datasets_info
data_info = [{dataset_name: datasets_info[dataset_name]}]
CLP.test(data_cfg, data_info)