-
Notifications
You must be signed in to change notification settings - Fork 1
/
BANet.py
201 lines (176 loc) · 11.3 KB
/
BANet.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
# -*- coding: utf-8 -*-
# @File : BANet.py
import torch
import torch.nn as nn
import torch.nn.functional as F
from ..sync_batchnorm.batchnorm import SynchronizedBatchNorm2d
from .ASPP import *
from .base import *
from .criterion import *
class MultiTaskNet(nn.Module):
def __init__(self, n_class, in_channels, out_channels, edge_poolings, batchnorm=nn.BatchNorm2d):
super(MultiTaskNet, self).__init__()
self.reduce_conv = nn.Conv2d(in_channels, out_channels, 1, 1, bias=False)
self.edge_pooling1 = nn.AvgPool2d(edge_poolings[0], stride=1, padding=(edge_poolings[0] - 1) // 2)
self.edge_pooling2 = nn.AvgPool2d(edge_poolings[1], stride=1, padding=(edge_poolings[1] - 1) // 2)
self.fuse_conv = nn.Sequential(nn.Conv2d(out_channels * 3, out_channels, 1, 1, bias=False),
nn.ReLU(True),
nn.Conv2d(out_channels, out_channels, 3, 1, 1, bias=False, groups=out_channels),
batchnorm(out_channels),
nn.Conv2d(out_channels, out_channels, 1, 1, bias=False))
self.edge_conv = nn.Sequential(nn.Conv2d(out_channels, out_channels, 3, padding=1, bias=False,
groups=out_channels),
batchnorm(out_channels),
nn.Conv2d(out_channels, out_channels, 1, 1, bias=False))
self.att_edge = nn.Sequential(nn.Conv2d(out_channels, out_channels, 3, padding=1, bias=False,
groups=out_channels),
batchnorm(out_channels),
nn.Conv2d(out_channels, out_channels, 1, 1, bias=False))
self.seg_conv = nn.Sequential(nn.Conv2d(out_channels, out_channels, 3, padding=1, bias=False,
groups=out_channels),
batchnorm(out_channels),
nn.Conv2d(out_channels, out_channels, 1, 1, bias=False))
self.att_seg = nn.Sequential(nn.Conv2d(out_channels, out_channels, 3, padding=1, bias=False,
groups=out_channels),
batchnorm(out_channels),
nn.Conv2d(out_channels, out_channels, 1, 1, bias=False))
self.sigmoid = nn.Sigmoid()
self.edge_supervision = nn.Conv2d(out_channels, n_class, 3, padding=1)
self.seg_supervision = nn.Conv2d(out_channels, n_class, 3, padding=1)
def forward(self, x):
x = self.reduce_conv(x)
edge1 = self.edge_pooling1(x)
edge2 = self.edge_pooling2(x)
x = self.fuse_conv(torch.cat([x, edge1, edge2], dim=1))
edge_x = self.edge_conv(x)
seg_x = self.seg_conv(x)
edge_feats = edge_x + (1 - self.sigmoid(self.att_edge(edge_x))) * seg_x
seg_feats = seg_x + (1 - self.sigmoid(self.att_seg(seg_x))) * edge_x
edge_sup = self.edge_supervision(edge_feats)
seg_sup = self.seg_supervision(seg_feats)
return seg_feats, edge_feats, seg_sup, edge_sup
class FeatureCrossFusion(nn.Module):
def __init__(self, in_channels):
super(FeatureCrossFusion, self).__init__()
self.sigmoid = nn.Sigmoid()
self.att_conv1 = nn.Conv2d(in_channels, in_channels, 3, padding=1, bias=False)
self.att_conv2 = nn.Conv2d(in_channels, in_channels, 3, padding=1, bias=False)
self.att_conv3 = nn.Conv2d(in_channels, in_channels, 3, padding=1, bias=False)
self.att_conv4 = nn.Conv2d(in_channels, in_channels, 3, padding=1, bias=False)
def forward(self, x1, x2, x3, x4):
size = x1.size()[2:]
if x2.size()[2:] != size:
x2 = F.interpolate(x2, size, mode='bilinear', align_corners=True)
if x3.size()[2:] != size:
x3 = F.interpolate(x3, size, mode='bilinear', align_corners=True)
if x4.size()[2:] != size:
x4 = F.interpolate(x4, size, mode='bilinear', align_corners=True)
att1 = self.att_conv1(x1)
att2 = self.att_conv2(x2)
att3 = self.att_conv3(x3)
att4 = self.att_conv4(x4)
out = x1 + (1 - self.sigmoid(att1)) * (
self.sigmoid(att2) * x2 + self.sigmoid(att3) * x3 + self.sigmoid(att4) * x4)
return out
class StackMultiTaskFusion_head(nn.Module):
def __init__(self, n_class, in_channels, reduce_dim, batchnorm=nn.BatchNorm2d, has_aspp=False, aspp_dim=None):
super(StackMultiTaskFusion_head, self).__init__()
print('aspp: {}'.format(has_aspp))
self.has_aspp = has_aspp
if has_aspp:
self.aspp_combine = nn.Sequential(nn.ReLU(True),
nn.Conv2d(reduce_dim + aspp_dim, reduce_dim + aspp_dim, 3, 1, 1,
bias=False, groups=reduce_dim + aspp_dim),
batchnorm(reduce_dim + aspp_dim),
nn.Conv2d(reduce_dim + aspp_dim, reduce_dim, 1, bias=False))
self.mtl4 = MultiTaskNet(n_class, in_channels[3], reduce_dim, [3, 5], batchnorm)
self.cf4 = FeatureCrossFusion(reduce_dim)
self.up_conv4 = nn.Sequential(nn.ReLU(True),
nn.Conv2d(reduce_dim * 2, reduce_dim * 2, kernel_size=3, padding=1, bias=False,
groups=reduce_dim * 2),
batchnorm(reduce_dim * 2),
nn.Conv2d(reduce_dim * 2, reduce_dim, kernel_size=1, bias=False))
self.mtl3 = MultiTaskNet(n_class, in_channels[2], reduce_dim, [3, 5], batchnorm)
self.cf3 = FeatureCrossFusion(reduce_dim)
self.up_conv3 = nn.Sequential(nn.ReLU(True),
nn.Conv2d(reduce_dim * 2, reduce_dim * 2, kernel_size=3, padding=1, bias=False,
groups=reduce_dim * 2),
batchnorm(reduce_dim * 2),
nn.Conv2d(reduce_dim * 2, reduce_dim, kernel_size=1, bias=False))
self.mtl2 = MultiTaskNet(n_class, in_channels[1], reduce_dim, [5, 7], batchnorm)
self.cf2 = FeatureCrossFusion(reduce_dim)
self.up_conv2 = nn.Sequential(nn.ReLU(True),
nn.Conv2d(reduce_dim * 2, reduce_dim * 2, kernel_size=3, padding=1, bias=False,
groups=reduce_dim * 2),
batchnorm(reduce_dim * 2),
nn.Conv2d(reduce_dim * 2, reduce_dim, kernel_size=1, bias=False))
self.mtl1 = MultiTaskNet(n_class, in_channels[0], reduce_dim, [5, 7], batchnorm)
self.cf1 = FeatureCrossFusion(reduce_dim)
self.up_conv1 = nn.Sequential(nn.ReLU(True),
nn.Conv2d(reduce_dim * 2, reduce_dim * 2, kernel_size=3, padding=1, bias=False,
groups=reduce_dim * 2),
batchnorm(reduce_dim * 2),
nn.Conv2d(reduce_dim * 2, reduce_dim, kernel_size=1, bias=False))
self.conv1 = nn.Sequential(nn.Conv2d(reduce_dim, reduce_dim, 3, 1, 1),
batchnorm(reduce_dim),
nn.ReLU(True))
self.conv2 = nn.Conv2d(reduce_dim, n_class, 1)
def forward(self, x1, x2, x3, x4, aspp=None):
x1_size = x1.size()[2:]
mtl4_1, mtl4_2, mtl4_3, mtl4_4 = self.mtl4(x4)
mtl3_1, mtl3_2, mtl3_3, mtl3_4 = self.mtl3(x3)
mtl2_1, mtl2_2, mtl2_3, mtl2_4 = self.mtl2(x2)
mtl1_1, mtl1_2, mtl1_3, mtl1_4 = self.mtl1(x1)
d4 = self.cf4(mtl4_1, mtl1_1, mtl2_1, mtl3_1)
if aspp is not None:
d4 = self.aspp_combine(torch.cat((d4, aspp), dim=1))
d3 = self.cf3(mtl3_1, mtl1_1, mtl2_1, mtl4_1)
d3 = self.up_conv3(torch.cat((d3, d4), dim=1))
d2 = self.cf2(mtl2_1, mtl1_1, mtl3_1, mtl4_1)
d2 = self.up_conv2(torch.cat((d2, d3), dim=1))
d1 = self.cf1(mtl1_1, mtl2_1, mtl3_1, mtl4_1)
up_d2 = F.interpolate(d2, x1_size, mode='bilinear', align_corners=True)
d1 = self.up_conv1(torch.cat((d1, up_d2), dim=1))
out = self.conv1(d1)
out = self.conv2(out)
return out, mtl1_3, mtl1_4, mtl2_3, mtl2_4, mtl3_3, mtl3_4, mtl4_3, mtl4_4
class BANet(BaseNet):
def __init__(self, n_class, backbone, batchnorm, aspp_rates=None, edge_kernel=None, is_train=True,
test_size=[256, 256], aspp_out_dim=256, reduce_dim=128, pooling='max', **kwargs):
super(BANet, self).__init__(n_class, backbone, batchnorm=batchnorm, pooling=pooling,
**kwargs)
self.aspp_rate = aspp_rates
self.is_train = is_train
self.test_size = test_size
self.edge_kernel = edge_kernel
self.block_channels = [256, 512, 1024, 2048]
if aspp_rates is not None:
self.aspp = ASPP(self.block_channels[3], aspp_out_dim, aspp_rates, batchnorm)
self.head = StackMultiTaskFusion_head(n_class, self.block_channels, reduce_dim, has_aspp=True,
aspp_dim=aspp_out_dim)
else:
self.head = StackMultiTaskFusion_head(n_class, self.block_channels, reduce_dim, has_aspp=False)
def forward(self, x):
imsize = x.size()[2:] if self.is_train else self.test_size
# print(imsize)
c1, c2, c3, c4 = self.base_forward(x)
if self.aspp_rate is not None:
aspp = self.aspp(c4)
out, seg1, edge1, seg2, edge2, seg3, edge3, seg4, edge4 = self.head(c1, c2, c3, c4, aspp)
else:
out, seg1, edge1, seg2, edge2, seg3, edge3, seg4, edge4 = self.head(c1, c2, c3, c4)
outputs = F.interpolate(out, imsize, mode='bilinear', align_corners=True)
segout1 = F.interpolate(seg1, imsize, mode='bilinear', align_corners=True)
edgeout1 = F.interpolate(edge1, imsize, mode='bilinear', align_corners=True)
segout2 = F.interpolate(seg2, imsize, mode='bilinear', align_corners=True)
edgeout2 = F.interpolate(edge2, imsize, mode='bilinear', align_corners=True)
segout3 = F.interpolate(seg3, imsize, mode='bilinear', align_corners=True)
edgeout3 = F.interpolate(edge3, imsize, mode='bilinear', align_corners=True)
segout4 = F.interpolate(seg4, imsize, mode='bilinear', align_corners=True)
edgeout4 = F.interpolate(edge4, imsize, mode='bilinear', align_corners=True)
return outputs, segout1, edgeout1, segout2, edgeout2, segout3, edgeout3, segout4, edgeout4
def get_model(model_name='BANet', dataset='kvasir', backbone='resnet50', root='./pretrain_models', **kwargs):
model_dict = {'banet': BANet}
from encoding.dataset import datasets
model = model_dict[model_name.lower()](datasets[dataset.lower()].NUM_CLASS, backbone=backbone, root=root, **kwargs)
return model