-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathBasicModules.py
252 lines (204 loc) · 10.5 KB
/
BasicModules.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
import torch
from torch import nn
import torch.nn.functional as F
import time
def conv3x3(in_planes, out_planes, stride=1):
"3x3 convolution with padding"
return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride,
padding=1, bias=False)
class BasicConv2d(nn.Module):
def __init__(self, in_planes, out_planes, kernel_size, stride=1, padding=0, dilation=1):
super(BasicConv2d, self).__init__()
self.conv = nn.Conv2d(in_planes, out_planes,
kernel_size=kernel_size, stride=stride,
padding=padding, dilation=dilation, bias=False)
self.bn = nn.BatchNorm2d(out_planes)
self.relu = nn.ReLU(inplace=True)
def forward(self, x):
x = self.conv(x)
x = self.bn(x)
return x
class SPLayer(nn.Module):
def __init__(self):
super(SPLayer, self).__init__()
def forward(self, list_x):
rgb0 = list_x[0][0].unsqueeze(dim=0)
tma0 = list_x[0][1].unsqueeze(dim=0)
rgb1 = list_x[1][0].unsqueeze(dim=0)
tma1 = list_x[1][1].unsqueeze(dim=0)
rgb2 = list_x[2][0].unsqueeze(dim=0)
tma2 = list_x[2][1].unsqueeze(dim=0)
rgb3 = list_x[3][0].unsqueeze(dim=0)
tma3 = list_x[3][1].unsqueeze(dim=0)
return rgb0, rgb1, rgb2, rgb3, tma0, tma1, tma2, tma3
class CA(nn.Module):
def __init__(self, in_planes, ratio=4):
super(CA, self).__init__()
self.avg_pool = nn.AdaptiveAvgPool2d(1)
self.max_pool = nn.AdaptiveMaxPool2d(1)
self.sharedMLP = nn.Sequential(
nn.Conv2d(in_planes, in_planes // ratio, 1, bias=False), nn.ReLU(),
nn.Conv2d(in_planes // ratio, in_planes, 1, bias=False))
self.sigmoid = nn.Sigmoid()
def forward(self, x):
avgout = self.sharedMLP(self.avg_pool(x))
maxout = self.sharedMLP(self.max_pool(x))
return self.sigmoid(avgout + maxout)
class CCE(nn.Module):
def __init__(self, in_dim=2048, sr_ratio=1):
super(CCE, self).__init__()
input_dim = in_dim
self.chanel_in = input_dim
self.query_convrd = nn.Conv2d(in_channels=input_dim, out_channels=input_dim, kernel_size=1)
self.key_convrd = nn.Conv2d(in_channels=input_dim, out_channels=input_dim, kernel_size=1)
self.value_convrd = nn.Conv2d(in_channels=input_dim, out_channels=input_dim, kernel_size=1)
self.query_convdr = nn.Conv2d(in_channels=input_dim, out_channels=input_dim, kernel_size=1)
self.key_convdr = nn.Conv2d(in_channels=input_dim, out_channels=input_dim, kernel_size=1)
self.value_convdr = nn.Conv2d(in_channels=input_dim, out_channels=input_dim, kernel_size=1)
self.sr_ratio = sr_ratio
dim = in_dim
if sr_ratio > 1:
self.sr_k = nn.Conv2d(dim, dim, kernel_size=sr_ratio, stride=sr_ratio)
self.sr_v = nn.Conv2d(dim, dim, kernel_size=sr_ratio, stride=sr_ratio)
self.norm_k = nn.BatchNorm2d(dim, eps=1e-05, momentum=0.1, affine=True)
self.norm_v = nn.BatchNorm2d(dim, eps=1e-05, momentum=0.1, affine=True)
self.sr_kk = nn.Conv2d(dim, dim, kernel_size=sr_ratio, stride=sr_ratio)
self.sr_vv = nn.Conv2d(dim, dim, kernel_size=sr_ratio, stride=sr_ratio)
self.norm_kk = nn.BatchNorm2d(dim, eps=1e-05, momentum=0.1, affine=True)
self.norm_vv = nn.BatchNorm2d(dim, eps=1e-05, momentum=0.1, affine=True)
self.gamma_rd = nn.Parameter(torch.zeros(1))
self.gamma_dr = nn.Parameter(torch.zeros(1))
self.gamma_x = nn.Parameter(torch.zeros(1))
self.softmax = nn.Softmax(dim=-1)
self.relu1 = nn.ReLU()
self.relu2 = nn.ReLU()
self.softmax = nn.Softmax(dim=1)
self.relu = nn.ReLU(inplace=True)
self.sigmoid = nn.Sigmoid()
self.fc1 = nn.Conv2d(dim*2, dim//2, kernel_size = 1)
self.fc2 = nn.Conv2d(dim//2, dim*2, kernel_size = 1)
self.merge_conv1x1 = nn.Sequential(
nn.Conv2d(dim*2, dim, 1, 1), self.relu)
def forward(self, x):
xr, xd = x[0].unsqueeze(dim=0), x[1].unsqueeze(dim=0)
m_batchsize, C, width, height = xr.size()
query_r = self.query_convrd(xr).view(m_batchsize, -1, width * height).permute(0, 2, 1)
if self.sr_ratio > 1:
key_d = self.norm_k(self.sr_k(xd))
value_d = self.norm_v(self.sr_v(xd))
key_d = self.key_convrd(key_d).view(m_batchsize, -1, width//self.sr_ratio * height//self.sr_ratio)
value_d = self.value_convrd(value_d).view(m_batchsize, -1, width//self.sr_ratio * height//self.sr_ratio)
else:
key_d = self.key_convrd(xd).view(m_batchsize, -1, width * height)
value_d = self.value_convrd(xd).view(m_batchsize, -1, width * height)
attention_rd = self.softmax(torch.bmm(query_r, key_d))
out_rd = torch.bmm(value_d, attention_rd.permute(0, 2, 1))
out_rd = out_rd.view(m_batchsize, C, width, height)
query_d = self.query_convdr(xd).view(m_batchsize, -1, width * height).permute(0, 2, 1)
if self.sr_ratio > 1:
key_r = self.norm_kk(self.sr_kk(xr))
value_r = self.norm_vv(self.sr_vv(xr))
key_r = self.key_convdr(key_r).view(m_batchsize, -1, width//self.sr_ratio * height//self.sr_ratio)
value_r = self.value_convdr(value_r).view(m_batchsize, -1, width//self.sr_ratio * height//self.sr_ratio)
else:
key_r = self.key_convdr(xr).view(m_batchsize, -1, width * height)
value_r = self.value_convdr(xr).view(m_batchsize, -1, width * height)
attention_dr = self.softmax(torch.bmm(query_d, key_r))
out_dr = torch.bmm(value_r, attention_dr.permute(0, 2, 1))
out_dr = out_dr.view(m_batchsize, C, width, height)
out_rd = self.gamma_rd * out_rd + xr
out_dr = self.gamma_dr * out_dr + xd
out_rd = self.relu(out_rd)
out_dr = self.relu(out_dr)
rgb_gap = nn.AvgPool2d(out_rd.shape[2:])(out_rd).view(len(out_rd), C, 1, 1)
hha_gap = nn.AvgPool2d(out_dr.shape[2:])(out_dr).view(len(out_dr), C, 1, 1)
stack_gap = torch.cat([rgb_gap, hha_gap], dim=1)
stack_gap = self.fc1(stack_gap)
stack_gap = self.relu(stack_gap)
stack_gap = self.fc2(stack_gap)
rgb_ = stack_gap[:, 0:C, :, :] * out_rd
hha_ = stack_gap[:, C:2*C, :, :] * out_dr
merge_feature = torch.cat([rgb_, hha_], dim=1)
merge_feature = self.merge_conv1x1(merge_feature)
rgb_out = (xr + merge_feature) / 2
hha_out = (xd + merge_feature) / 2
rgb_out = self.relu1(rgb_out)
hha_out = self.relu2(hha_out)
out_x = torch.cat([rgb_out, hha_out], dim=0)
return out_x
class aggregation_scale(nn.Module):
def __init__(self, in_dim, out_dim, dilation=[1,2,3], residual=False):
super(aggregation_scale, self).__init__()
if in_dim == out_dim:
residual=True
self.use_res_connect = residual
mid_dim = out_dim*2
self.conv1 = BasicConv2d(in_dim, mid_dim, kernel_size=1)
self.hidden_conv1 = nn.Conv2d(mid_dim, mid_dim, kernel_size=3, padding=1, groups=mid_dim, dilation=1)
self.hidden_conv2 = nn.Conv2d(mid_dim, mid_dim, kernel_size=3, padding=2, groups=mid_dim, dilation=2)
self.hidden_conv3 = nn.Conv2d(mid_dim, mid_dim, kernel_size=3, padding=3, groups=mid_dim, dilation=3)
self.hidden_bnact = nn.Sequential(nn.BatchNorm2d(mid_dim), nn.ReLU(inplace=True))
self.out_conv = nn.Sequential(nn.Conv2d(mid_dim, out_dim, 1, 1, 0, bias=False))
def forward(self, input_):
x = self.conv1(input_)
x1 = self.hidden_conv1(x)
x2 = self.hidden_conv2(x)
x3 = self.hidden_conv3(x)
intra = self.hidden_bnact(x1+x2+x3)
output = self.out_conv(intra)
if self.use_res_connect:
output = input_ + output
return output
class aggregation_cross(nn.Module):
def __init__(self, channel):
super(aggregation_cross, self).__init__()
self.relu = nn.ReLU(True)
self.upsample = nn.Upsample(scale_factor=2, mode='bilinear', align_corners=True)
self.conv_upsample1 = BasicConv2d(channel, channel, 3, padding=1)
self.conv_upsample2 = BasicConv2d(channel, channel, 3, padding=1)
self.conv_upsample3 = BasicConv2d(channel, channel, 3, padding=1)
self.conv_upsample4 = BasicConv2d(channel, channel, 3, padding=1)
self.conv_upsample5 = BasicConv2d(2*channel, 2*channel, 3, padding=1)
self.conv_concat2 = BasicConv2d(2*channel, 2*channel, 3, padding=1)
self.conv_concat3 = BasicConv2d(3*channel, 3*channel, 3, padding=1)
self.conv4 = BasicConv2d(3*channel, channel, 3, padding=1)
self.conv5 = nn.Conv2d(channel, 1, 1)
def forward(self, x1, x2, x3):
x1_1 = x1
x2_1 = self.conv_upsample1(self.upsample(x1)) * x2
x3_1 = self.conv_upsample2(self.upsample(self.upsample(x1))) \
* self.conv_upsample3(self.upsample(x2)) * x3
x2_2 = torch.cat((x2_1, self.conv_upsample4(self.upsample(x1_1))), 1)
x3_2 = torch.cat((x3_1, self.conv_upsample5(self.upsample(x2_2))), 1)
x3_2 = self.conv_concat3(x3_2)
x_k = self.conv4(x3_2)
x = self.conv5(x_k)
return x_k, x
class TransBasicBlock(nn.Module):
expansion = 1
def __init__(self, inplanes, planes, stride=1, upsample=None, **kwargs):
super(TransBasicBlock, self).__init__()
self.conv1 = conv3x3(inplanes, inplanes)
self.bn1 = nn.BatchNorm2d(inplanes)
self.relu = nn.ReLU(inplace=True)
if upsample is not None and stride != 1:
self.conv2 = nn.ConvTranspose2d(inplanes, planes,
kernel_size=3, stride=stride, padding=1,
output_padding=1, bias=False)
else:
self.conv2 = conv3x3(inplanes, planes, stride)
self.bn2 = nn.BatchNorm2d(planes)
self.upsample = upsample
self.stride = stride
def forward(self, x):
residual = x
out = self.conv1(x)
out = self.bn1(out)
out = self.relu(out)
out = self.conv2(out)
out = self.bn2(out)
if self.upsample is not None:
residual = self.upsample(x)
out += residual
out = self.relu(out)
return out