forked from miemie2013/Pytorch-PPYOLO
-
Notifications
You must be signed in to change notification settings - Fork 0
/
1_ppyolo_2x_2pytorch.py
329 lines (262 loc) · 10.4 KB
/
1_ppyolo_2x_2pytorch.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
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
#! /usr/bin/env python
# coding=utf-8
# ================================================================
#
# Author : miemie2013
# Created date: 2020-10-15 14:50:03
# Description : pytorch_ppyolo
#
# ================================================================
from config import *
from model.custom_layers import DCNv2
from model.ppyolo import PPYOLO
import paddle.fluid as fluid
use_gpu = True
cfg = PPYOLO_2x_Config()
# 该模型是COCO数据集上训练好的，所以强制改类别数为80
cfg.num_classes = 80
cfg.head['num_classes'] = cfg.num_classes
model_path = 'ppyolo.pdparams'
import torch
def load_weights(path):
state_dict = fluid.io.load_program_state(path)
return state_dict
state_dict = load_weights(model_path)
print('============================================================')
# 创建模型
Backbone = select_backbone(cfg.backbone_type)
backbone = Backbone(**cfg.backbone)
Head = select_head(cfg.head_type)
head = Head(yolo_loss=None, **cfg.head)
ppyolo = PPYOLO(backbone, head)
if use_gpu:
ppyolo = ppyolo.cuda()
ppyolo.eval() # 必须调用model.eval()来设置dropout和batch normalization layers在运行推理前，切换到评估模式. 不这样做的化会产生不一致的推理结果.
print('\nCopying...')
def copy_conv_bn(conv_unit, w, scale, offset, m, v):
conv_unit.conv.weight.data = torch.Tensor(w).cuda()
conv_unit.bn.weight.data = torch.Tensor(scale).cuda()
conv_unit.bn.bias.data = torch.Tensor(offset).cuda()
conv_unit.bn.running_mean.data = torch.Tensor(m).cuda()
conv_unit.bn.running_var.data = torch.Tensor(v).cuda()
def copy_conv(conv_layer, w, b):
conv_layer.weight.data = torch.Tensor(w).cuda()
conv_layer.bias.data = torch.Tensor(b).cuda()
# Resnet50Vd
w = state_dict['conv1_1_weights']
scale = state_dict['bnv1_1_scale']
offset = state_dict['bnv1_1_offset']
m = state_dict['bnv1_1_mean']
v = state_dict['bnv1_1_variance']
copy_conv_bn(backbone.stage1_conv1_1, w, scale, offset, m, v)
w = state_dict['conv1_2_weights']
scale = state_dict['bnv1_2_scale']
offset = state_dict['bnv1_2_offset']
m = state_dict['bnv1_2_mean']
v = state_dict['bnv1_2_variance']
copy_conv_bn(backbone.stage1_conv1_2, w, scale, offset, m, v)
w = state_dict['conv1_3_weights']
scale = state_dict['bnv1_3_scale']
offset = state_dict['bnv1_3_offset']
m = state_dict['bnv1_3_mean']
v = state_dict['bnv1_3_variance']
copy_conv_bn(backbone.stage1_conv1_3, w, scale, offset, m, v)
nums = [3, 4, 6, 3]
for nid, num in enumerate(nums):
stage_name = 'res' + str(nid + 2)
for kk in range(num):
block_name = stage_name + chr(ord("a") + kk)
conv_name1 = block_name + "_branch2a"
conv_name2 = block_name + "_branch2b"
conv_name3 = block_name + "_branch2c"
shortcut_name = block_name + "_branch1"
bn_name1 = 'bn' + conv_name1[3:]
bn_name2 = 'bn' + conv_name2[3:]
bn_name3 = 'bn' + conv_name3[3:]
shortcut_bn_name = 'bn' + shortcut_name[3:]
w = state_dict[conv_name1 + '_weights']
scale = state_dict[bn_name1 + '_scale']
offset = state_dict[bn_name1 + '_offset']
m = state_dict[bn_name1 + '_mean']
v = state_dict[bn_name1 + '_variance']
copy_conv_bn(backbone.get_block('stage%d_%d' % (2+nid, kk)).conv1, w, scale, offset, m, v)
if nid == 3: # DCNv2
conv_unit = backbone.get_block('stage%d_%d' % (2+nid, kk)).conv2
offset_w = state_dict[conv_name2 + '_conv_offset.w_0']
offset_b = state_dict[conv_name2 + '_conv_offset.b_0']
if isinstance(conv_unit.conv, DCNv2): # 如果是自实现的DCNv2
copy_conv(conv_unit.conv.conv_offset, offset_w, offset_b)
else:
copy_conv(conv_unit.conv.conv_offset_mask, offset_w, offset_b)
w = state_dict[conv_name2 + '_weights']
scale = state_dict[bn_name2 + '_scale']
offset = state_dict[bn_name2 + '_offset']
m = state_dict[bn_name2 + '_mean']
v = state_dict[bn_name2 + '_variance']
if isinstance(conv_unit.conv, DCNv2): # 如果是自实现的DCNv2
conv_unit.conv.dcn_weight.data = torch.Tensor(w).cuda()
conv_unit.bn.weight.data = torch.Tensor(scale).cuda()
conv_unit.bn.bias.data = torch.Tensor(offset).cuda()
conv_unit.bn.running_mean.data = torch.Tensor(m).cuda()
conv_unit.bn.running_var.data = torch.Tensor(v).cuda()
else:
copy_conv_bn(conv_unit, w, scale, offset, m, v)
else:
w = state_dict[conv_name2 + '_weights']
scale = state_dict[bn_name2 + '_scale']
offset = state_dict[bn_name2 + '_offset']
m = state_dict[bn_name2 + '_mean']
v = state_dict[bn_name2 + '_variance']
copy_conv_bn(backbone.get_block('stage%d_%d' % (2+nid, kk)).conv2, w, scale, offset, m, v)
w = state_dict[conv_name3 + '_weights']
scale = state_dict[bn_name3 + '_scale']
offset = state_dict[bn_name3 + '_offset']
m = state_dict[bn_name3 + '_mean']
v = state_dict[bn_name3 + '_variance']
copy_conv_bn(backbone.get_block('stage%d_%d' % (2+nid, kk)).conv3, w, scale, offset, m, v)
# 每个stage的第一个卷积块才有4个卷积层
if kk == 0:
w = state_dict[shortcut_name + '_weights']
scale = state_dict[shortcut_bn_name + '_scale']
offset = state_dict[shortcut_bn_name + '_offset']
m = state_dict[shortcut_bn_name + '_mean']
v = state_dict[shortcut_bn_name + '_variance']
copy_conv_bn(backbone.get_block('stage%d_%d' % (2+nid, kk)).conv4, w, scale, offset, m, v)
# head
conv_block_num = 2
num_classes = 80
anchors = [[10, 13], [16, 30], [33, 23],
[30, 61], [62, 45], [59, 119],
[116, 90], [156, 198], [373, 326]]
anchor_masks = [[6, 7, 8], [3, 4, 5], [0, 1, 2]]
batch_size = 1
norm_type = "bn"
coord_conv = True
iou_aware = True
iou_aware_factor = 0.4
block_size = 3
scale_x_y = 1.05
use_spp = True
drop_block = True
keep_prob = 0.9
clip_bbox = True
yolo_loss = None
downsample = [32, 16, 8]
in_channels = [2048, 1024, 512]
nms_cfg = None
is_train = False
bn = 0
gn = 0
af = 0
if norm_type == 'bn':
bn = 1
elif norm_type == 'gn':
gn = 1
elif norm_type == 'affine_channel':
af = 1
def copy_DetectionBlock(
_detection_block,
in_c,
channel,
coord_conv=True,
bn=0,
gn=0,
af=0,
conv_block_num=2,
is_first=False,
use_spp=True,
drop_block=True,
block_size=3,
keep_prob=0.9,
is_test=True,
name=''):
kkk = 0
for j in range(conv_block_num):
kkk += 1
conv_name = '{}.{}.0'.format(name, j)
w = state_dict[conv_name + '.conv.weights']
scale = state_dict[conv_name + '.bn.scale']
offset = state_dict[conv_name + '.bn.offset']
m = state_dict[conv_name + '.bn.mean']
v = state_dict[conv_name + '.bn.var']
copy_conv_bn(_detection_block.layers[kkk], w, scale, offset, m, v)
kkk += 1
if use_spp and is_first and j == 1:
kkk += 1
conv_name = '{}.{}.spp.conv'.format(name, j)
w = state_dict[conv_name + '.conv.weights']
scale = state_dict[conv_name + '.bn.scale']
offset = state_dict[conv_name + '.bn.offset']
m = state_dict[conv_name + '.bn.mean']
v = state_dict[conv_name + '.bn.var']
copy_conv_bn(_detection_block.layers[kkk], w, scale, offset, m, v)
kkk += 1
conv_name = '{}.{}.1'.format(name, j)
w = state_dict[conv_name + '.conv.weights']
scale = state_dict[conv_name + '.bn.scale']
offset = state_dict[conv_name + '.bn.offset']
m = state_dict[conv_name + '.bn.mean']
v = state_dict[conv_name + '.bn.var']
copy_conv_bn(_detection_block.layers[kkk], w, scale, offset, m, v)
kkk += 1
else:
conv_name = '{}.{}.1'.format(name, j)
w = state_dict[conv_name + '.conv.weights']
scale = state_dict[conv_name + '.bn.scale']
offset = state_dict[conv_name + '.bn.offset']
m = state_dict[conv_name + '.bn.mean']
v = state_dict[conv_name + '.bn.var']
copy_conv_bn(_detection_block.layers[kkk], w, scale, offset, m, v)
kkk += 1
if drop_block and j == 0 and not is_first:
kkk += 1
if drop_block and is_first:
kkk += 1
kkk += 1
conv_name = '{}.2'.format(name)
w = state_dict[conv_name + '.conv.weights']
scale = state_dict[conv_name + '.bn.scale']
offset = state_dict[conv_name + '.bn.offset']
m = state_dict[conv_name + '.bn.mean']
v = state_dict[conv_name + '.bn.var']
copy_conv_bn(_detection_block.layers[kkk], w, scale, offset, m, v)
kkk += 1
conv_name = '{}.tip'.format(name)
w = state_dict[conv_name + '.conv.weights']
scale = state_dict[conv_name + '.bn.scale']
offset = state_dict[conv_name + '.bn.offset']
m = state_dict[conv_name + '.bn.mean']
v = state_dict[conv_name + '.bn.var']
copy_conv_bn(_detection_block.tip_layers[1], w, scale, offset, m, v)
out_layer_num = len(downsample)
for i in range(out_layer_num):
copy_DetectionBlock(
head.detection_blocks[i],
in_c=in_channels[i],
channel=64 * (2**out_layer_num) // (2**i),
coord_conv=coord_conv,
bn=bn,
gn=gn,
af=af,
is_first=i == 0,
conv_block_num=conv_block_num,
use_spp=use_spp,
drop_block=drop_block,
block_size=block_size,
keep_prob=keep_prob,
is_test=(not is_train),
name="yolo_block.{}".format(i)
)
w = state_dict["yolo_output.{}.conv.weights".format(i)]
b = state_dict["yolo_output.{}.conv.bias".format(i)]
copy_conv(head.yolo_output_convs[i].conv, w, b)
if i < out_layer_num - 1:
conv_name = "yolo_transition.{}".format(i)
w = state_dict[conv_name + '.conv.weights']
scale = state_dict[conv_name + '.bn.scale']
offset = state_dict[conv_name + '.bn.offset']
m = state_dict[conv_name + '.bn.mean']
v = state_dict[conv_name + '.bn.var']
copy_conv_bn(head.upsample_layers[i*2], w, scale, offset, m, v)
torch.save(ppyolo.state_dict(), 'ppyolo_2x.pt')
print('\nDone.')