-
Notifications
You must be signed in to change notification settings - Fork 0
/
lateral_net.py
313 lines (265 loc) · 12.3 KB
/
lateral_net.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
import torch
import torch.nn as nn
import ResNeXt as ResNeXt
import resnext_weights_helper as resnext_utils
import mobilenetv2_weight_helper as mobilenet_utils
import MobileNetV2 as MobileNetV2
from torch.nn import functional as F
import math
def lateral_resnext50_32x4d_body_stride16(cfg):
return lateral(cfg,ResNeXt.ResNeXt50_32x4d_body_stride16)
def lateral_resnext101_32x4d_body_stride16(cfg):
return lateral(cfg,ResNeXt.ResNeXt101_32x4d_body_stride16)
def lateral_mobilenetv2_body_stride8(cfg):
return lateral(cfg,MobileNetV2.MobileNetV2_body_stride8)
class lateral(nn.Module):
def __init__(self, cfg,conv_body_func):
super().__init__()
self.cfg = cfg
self.dim_in = self.cfg['RESNET_BOTTLENECK_DIM']
self.dim_in = self.dim_in[-1:0:-1]
self.dim_out = self.cfg['LATERAL_OUT']
self.num_lateral_stages = len(self.dim_in)
self.topdown_lateral_modules = nn.ModuleList()
for i in range(self.num_lateral_stages):
self.topdown_lateral_modules.append(
lateral_block(self.dim_in[i], self.dim_out[i]))
self.bottomup = conv_body_func(self.cfg)
dilation_rate = [4, 8, 12] if 'stride_8' in self.cfg['ENCODER'] else [2, 4, 6]
encoder_stride = 8 if 'stride8' in self.cfg['ENCODER'] else 16
if 'mobilenetv2' in self.cfg['ENCODER']:
self.bottomup_top = Global_pool_block(self.dim_in[0], self.dim_out[0], encoder_stride)
else:
self.bottomup_top = ASPP_block(self.dim_in[0], self.dim_out[0], dilation_rate, encoder_stride)
self._init_modules(self.cfg['INIT_TYPE'])
def _init_modules(self, init_type):
if self.cfg['LOAD_IMAGENET_PRETRAINED_WEIGHTS']:
if 'resnext' in self.cfg['ENCODER'].lower():
resnext_utils.load_pretrained_imagenet_resnext_weights(self.bottomup)
elif 'mobilenetv2' in self.cfg['ENCODER'].lower():
mobilenet_utils.load_pretrained_imagenet_resnext_weights(self.bottomup)
self._init_weights(init_type)
def _init_weights(self, init_type='xavier'):
def init_func(m):
if isinstance(m, nn.Conv2d):
if init_type == 'xavier':
nn.init.xavier_normal_(m.weight)
if init_type == 'kaiming':
nn.init.kaiming_normal_(m.weight)
if init_type == 'gaussian':
nn.init.normal_(m.weight, std=0.01)
if m.bias is not None:
nn.init.constant_(m.bias, 0.0)
elif isinstance(m, nn.BatchNorm2d):
nn.init.constant_(m.weight.data, 1.0)
nn.init.constant_(m.bias.data, 0.0)
def init_model_weight(m):
for child_m in m.children():
if not isinstance(child_m, nn.ModuleList):
child_m.apply(init_func)
if self.cfg['LOAD_IMAGENET_PRETRAINED_WEIGHTS']:
init_model_weight(self.topdown_lateral_modules)
init_model_weight(self.bottomup_top)
else:
init_model_weight(self)
def forward(self, x):
_, _, h, w = x.shape
backbone_stage_size = [(math.ceil(h/(2.0**i)), math.ceil(w/(2.0**i))) for i in range(5, 0, -1)]
backbone_stage_size.append((h, w))
bottemup_blocks_out = [self.bottomup.res1(x)]
for i in range(1, self.bottomup.convX):
bottemup_blocks_out.append(
getattr(self.bottomup, 'res%d' % (i + 1))(bottemup_blocks_out[-1])
)
bottemup_top_out = self.bottomup_top(bottemup_blocks_out[-1])
lateral_blocks_out = [bottemup_top_out]
for i in range(self.num_lateral_stages):
lateral_blocks_out.append(self.topdown_lateral_modules[i](
bottemup_blocks_out[-(i + 1)]
))
return lateral_blocks_out, backbone_stage_size
class Global_pool_block(nn.Module):
def __init__(self, dim_in, dim_out, output_stride):
super().__init__()
self.dim_in = dim_in
self.dim_out = dim_out
self.globalpool_conv1x1 = nn.Conv2d(self.dim_in, self.dim_out, 1, stride=1, padding=0, bias=False)
self.globalpool = nn.AdaptiveAvgPool2d((1, 1))
self.globalpool_bn = nn.BatchNorm2d(self.dim_out, momentum=0.9)
self.unpool = nn.AdaptiveAvgPool2d((int(self.cfg['CROP_SIZE'][0] / output_stride), int(self.cfg['CROP_SIZE'][1] / output_stride)))
def forward(self, x):
out = self.globalpool_conv1x1(x)
out = self.globalpool_bn(out)
out = self.globalpool(out)
out = self.unpool(out)
return out
class ASPP_block(nn.Module):
def __init__(self, dim_in, dim_out, dilate_rates, output_stride):
super().__init__()
self.dim_in = dim_in
self.dim_out = dim_out
self.dilate_rates = dilate_rates
self.aspp_conv1x1 = nn.Conv2d(self.dim_in, self.dim_out, 1, stride=1, padding=0, bias=False)
self.aspp_conv3_1 = nn.Conv2d(self.dim_in, self.dim_out, 3, stride=1, padding=self.dilate_rates[0],
dilation=self.dilate_rates[0], bias=False)
self.aspp_conv3_2 = nn.Conv2d(self.dim_in, self.dim_out, 3, stride=1, padding=self.dilate_rates[1],
dilation=self.dilate_rates[1], bias=False)
self.aspp_conv3_3 = nn.Conv2d(self.dim_in, self.dim_out, 3, stride=1, padding=self.dilate_rates[2],
dilation=self.dilate_rates[2], bias=False)
self.aspp_bn1x1 = nn.BatchNorm2d(self.dim_out, momentum=0.5)
self.aspp_bn3_1 = nn.BatchNorm2d(self.dim_out, momentum=0.5)
self.aspp_bn3_2 = nn.BatchNorm2d(self.dim_out, momentum=0.5)
self.aspp_bn3_3 = nn.BatchNorm2d(self.dim_out, momentum=0.5)
self.globalpool = nn.AdaptiveAvgPool2d((1, 1))
self.globalpool_conv1x1 = nn.Conv2d(self.dim_in, self.dim_out, 1, stride=1, padding=0, bias=False)
self.globalpool_bn = nn.BatchNorm2d(self.dim_out, momentum=0.5)
def forward(self, x):
x1 = self.aspp_conv1x1(x)
x1 = self.aspp_bn1x1(x1)
x2 = self.aspp_conv3_1(x)
x2 = self.aspp_bn3_1(x2)
x3 = self.aspp_conv3_2(x)
x3 = self.aspp_bn3_2(x3)
x4 = self.aspp_conv3_3(x)
x4 = self.aspp_bn3_3(x4)
x5 = self.globalpool(x)
x5 = self.globalpool_conv1x1(x5)
x5 = self.globalpool_bn(x5)
w, h = x1.size(2), x1.size(3)
x5 = F.upsample(input=x5, size=(w, h), mode='bilinear', align_corners=True)
out = torch.cat([x1, x2, x3, x4, x5], 1)
return out
class lateral_block(nn.Module):
def __init__(self, dim_in, dim_out):
super().__init__()
self.dim_in = dim_in
self.dim_out = dim_out
self.lateral = FTB_block(dim_in, dim_out)
def forward(self, x):
out = self.lateral(x)
return out
class fcn_topdown(nn.Module):
def __init__(self, cfg,conv_body_func):
super().__init__()
self.cfg = cfg
self.dim_in = self.cfg['FCN_DIM_IN']
self.dim_out = self.cfg['FCN_DIM_OUT'] + [self.cfg['DECODER_OUTPUT_C']]
self.num_fcn_topdown = len(self.dim_in)
aspp_blocks_num = 1 if 'mobilenetv2' in self.cfg['ENCODER'] else 5
self.top = nn.Sequential(
nn.Conv2d(self.dim_in[0] * aspp_blocks_num, self.dim_in[0], 1, stride=1, padding=0, bias=False),
nn.BatchNorm2d(self.dim_in[0], 0.5)
)
self.topdown_fcn1 = fcn_topdown_block(self.dim_in[0], self.dim_out[0])
self.topdown_fcn2 = fcn_topdown_block(self.dim_in[1], self.dim_out[1])
self.topdown_fcn3 = fcn_topdown_block(self.dim_in[2], self.dim_out[2])
self.topdown_fcn4 = fcn_topdown_block(self.dim_in[3], self.dim_out[3])
self.topdown_fcn5 = fcn_last_block(self.dim_in[4], self.dim_out[4])
self.topdown_predict = fcn_topdown_predict(self.dim_in[5], self.dim_out[5])
self.init_type = self.cfg['INIT_TYPE']
self._init_modules(self.init_type)
def _init_modules(self, init_type):
self._init_weights(init_type)
def _init_weights(self, init_type='xavier'):
def init_func(m):
if isinstance(m, nn.Conv2d):
if init_type == 'xavier':
nn.init.xavier_normal_(m.weight)
if init_type == 'kaiming':
nn.init.kaiming_normal(m.weight)
if init_type == 'gaussian':
nn.init.normal_(m.weight, std=0.01)
if m.bias is not None:
nn.init.constant_(m.bias, 0.0)
elif isinstance(m, nn.BatchNorm2d):
nn.init.normal_(m.weight.data, 1.0, 0.0)
nn.init.constant_(m.bias.data, 0.0)
for child_m in self.children():
child_m.apply(init_func)
def forward(self, laterals, backbone_stage_size):
x = self.top(laterals[0])
x1 = self.topdown_fcn1(laterals[1], x)
x2 = self.topdown_fcn2(laterals[2], x1)
x3 = self.topdown_fcn3(laterals[3], x2)
x4 = self.topdown_fcn4(laterals[4], x3)
x5 = self.topdown_fcn5(x4, backbone_stage_size)
x6 = self.topdown_predict(x5)
return x6
class fcn_topdown_block(nn.Module):
def __init__(self, dim_in, dim_out):
super().__init__()
self.dim_in = dim_in
self.dim_out = dim_out
self.afa_block = AFA_block(dim_in)
self.ftb_block = FTB_block(self.dim_in, self.dim_out)
def forward(self, lateral, top, size=None):
if lateral.shape != top.shape:
h, w = lateral.size(2), lateral.size(3)
top = F.interpolate(input=top, size=(h, w), mode='bilinear',align_corners=True)
out = self.afa_block(lateral, top)
out = self.ftb_block(out)
return out
class fcn_topdown_predict(nn.Module):
def __init__(self, dim_in, dim_out):
super().__init__()
self.dim_in = dim_in
self.dim_out = dim_out
self.dropout = nn.Dropout2d(0.0)
self.conv1 = nn.Conv2d(self.dim_in, self.dim_out, 3, stride=1, padding=2, dilation=2, bias=True)
self.softmax = nn.Softmax(dim=1)
def forward(self, x):
x = self.dropout(x)
x = self.conv1(x)
x_softmax = self.softmax(x)
return x, x_softmax
class FTB_block(nn.Module):
def __init__(self, dim_in, dim_out):
super().__init__()
self.dim_in = dim_in
self.dim_out = dim_out
self.conv1 = nn.Conv2d(self.dim_in, self.dim_out, 1, stride=1, padding=0, bias=False)
self.conv2 = nn.Conv2d(self.dim_out, self.dim_out, 3, stride=1, padding=2, dilation=2, bias=True)
self.bn1 = nn.BatchNorm2d(self.dim_out, momentum=0.5)
self.relu = nn.ReLU(inplace=True)
self.conv3 = nn.Conv2d(self.dim_out, self.dim_out, 3, stride=1, padding=2, dilation=2, bias=False)
def forward(self, x):
x = self.conv1(x)
residual = x
out = self.conv2(x)
out = self.bn1(out)
out = self.relu(out)
out = self.conv3(out)
out += residual
out = self.relu(out)
return out
class AFA_block(nn.Module):
def __init__(self, dim):
super().__init__()
self.dim_in = dim * 2
self.dim_out = dim
self.dim_mid = int(dim / 8)
self.globalpool = nn.AdaptiveAvgPool2d(1)
self.conv1 = nn.Conv2d(self.dim_in, self.dim_mid, 1, stride=1, padding=0, bias=False)
self.relu = nn.ReLU(inplace=True)
self.conv2 = nn.Conv2d(self.dim_mid, self.dim_out, 1, stride=1, padding=0, bias=False)
self.sigmd = nn.Sigmoid()
def forward(self, lateral, top):
w = torch.cat([lateral, top], 1)
w = self.globalpool(w)
w = self.conv1(w)
w = self.relu(w)
w = self.conv2(w)
w = self.sigmd(w)
out = w * lateral + top
return out
class fcn_last_block(nn.Module):
def __init__(self, dim_in, dim_out):
super().__init__()
self.dim_in = dim_in
self.dim_out = dim_out
self.ftb = FTB_block(dim_in, dim_out)
def forward(self, input, backbone_stage_size):
out = F.upsample(input=input, size=(backbone_stage_size[4][0], backbone_stage_size[4][1]), mode='bilinear', align_corners=True)
out = self.ftb(out)
out = F.upsample(input=out, size=(backbone_stage_size[5][0], backbone_stage_size[5][1]), mode='bilinear', align_corners=True)
return out