|
a |
|
b/model/MyModel.py |
|
|
1 |
import torch |
|
|
2 |
import torch.nn as nn |
|
|
3 |
import torch.nn.functional as F |
|
|
4 |
from torch.nn import init |
|
|
5 |
from features.feature import * |
|
|
6 |
from gcn.layers import GConv |
|
|
7 |
|
|
|
8 |
class Self_Attn(nn.Module): |
|
|
9 |
""" Self attention Layer""" |
|
|
10 |
def __init__(self,in_dim,activation): |
|
|
11 |
super(Self_Attn,self).__init__() |
|
|
12 |
self.chanel_in = in_dim |
|
|
13 |
self.activation = activation |
|
|
14 |
|
|
|
15 |
self.query_conv = nn.Conv2d(in_channels = in_dim , out_channels = in_dim//8 , kernel_size= 1) |
|
|
16 |
self.key_conv = nn.Conv2d(in_channels = in_dim , out_channels = in_dim//8 , kernel_size= 1) |
|
|
17 |
self.value_conv = nn.Conv2d(in_channels = in_dim , out_channels = in_dim , kernel_size= 1) |
|
|
18 |
self.gamma = nn.Parameter(torch.zeros(1)) |
|
|
19 |
|
|
|
20 |
self.softmax = nn.Softmax(dim=-1) # |
|
|
21 |
def forward(self,x): |
|
|
22 |
m_batchsize,C,width ,height = x.size() |
|
|
23 |
proj_query = self.query_conv(x).view(m_batchsize,-1,width*height).permute(0,2,1) # B X CX(N) |
|
|
24 |
proj_key = self.key_conv(x).view(m_batchsize,-1,width*height) # B X C x (*W*H) |
|
|
25 |
energy = torch.bmm(proj_query,proj_key) # transpose check |
|
|
26 |
attention = self.softmax(energy) # BX (N) X (N) |
|
|
27 |
proj_value = self.value_conv(x).view(m_batchsize,-1,width*height) # B X C X N |
|
|
28 |
|
|
|
29 |
out = torch.bmm(proj_value,attention.permute(0,2,1) ) |
|
|
30 |
out = out.view(m_batchsize,C,width,height) |
|
|
31 |
|
|
|
32 |
|
|
|
33 |
|
|
|
34 |
out = self.gamma*out + x |
|
|
35 |
return out,attention |
|
|
36 |
|
|
|
37 |
def init_weights(net, init_type='normal', gain=0.02): |
|
|
38 |
def init_func(m): |
|
|
39 |
classname = m.__class__.__name__ |
|
|
40 |
if hasattr(m, 'weight') and (classname.find('Conv') != -1 or classname.find('Linear') != -1): |
|
|
41 |
if init_type == 'normal': |
|
|
42 |
init.normal_(m.weight.data, 0.0, gain) |
|
|
43 |
elif init_type == 'xavier': |
|
|
44 |
init.xavier_normal_(m.weight.data, gain=gain) |
|
|
45 |
elif init_type == 'kaiming': |
|
|
46 |
init.kaiming_normal_(m.weight.data, a=0, mode='fan_in') |
|
|
47 |
elif init_type == 'orthogonal': |
|
|
48 |
init.orthogonal_(m.weight.data, gain=gain) |
|
|
49 |
else: |
|
|
50 |
raise NotImplementedError('initialization method [%s] is not implemented' % init_type) |
|
|
51 |
if hasattr(m, 'bias') and m.bias is not None: |
|
|
52 |
init.constant_(m.bias.data, 0.0) |
|
|
53 |
elif classname.find('BatchNorm2d') != -1: |
|
|
54 |
init.normal_(m.weight.data, 1.0, gain) |
|
|
55 |
init.constant_(m.bias.data, 0.0) |
|
|
56 |
|
|
|
57 |
print('initialize network with %s' % init_type) |
|
|
58 |
net.apply(init_func) |
|
|
59 |
|
|
|
60 |
class conv_block(nn.Module): |
|
|
61 |
def __init__(self,ch_in,ch_out): |
|
|
62 |
super(conv_block,self).__init__() |
|
|
63 |
self.conv = nn.Sequential( |
|
|
64 |
nn.Conv2d(ch_in, ch_out, kernel_size=3,stride=1,padding=1,bias=True), |
|
|
65 |
nn.BatchNorm2d(ch_out), |
|
|
66 |
nn.ReLU(inplace=True), |
|
|
67 |
nn.Conv2d(ch_out, ch_out, kernel_size=3,stride=1,padding=1,bias=True), |
|
|
68 |
nn.BatchNorm2d(ch_out), |
|
|
69 |
nn.ReLU(inplace=True) |
|
|
70 |
) |
|
|
71 |
|
|
|
72 |
|
|
|
73 |
def forward(self,x): |
|
|
74 |
x = self.conv(x) |
|
|
75 |
return x |
|
|
76 |
|
|
|
77 |
class up_conv(nn.Module): |
|
|
78 |
def __init__(self,ch_in,ch_out): |
|
|
79 |
super(up_conv,self).__init__() |
|
|
80 |
self.up = nn.Sequential( |
|
|
81 |
nn.Upsample(scale_factor=2), |
|
|
82 |
nn.Conv2d(ch_in,ch_out,kernel_size=3,stride=1,padding=1,bias=True), |
|
|
83 |
nn.BatchNorm2d(ch_out), |
|
|
84 |
nn.ReLU(inplace=True) |
|
|
85 |
) |
|
|
86 |
|
|
|
87 |
def forward(self,x): |
|
|
88 |
x = self.up(x) |
|
|
89 |
return x |
|
|
90 |
|
|
|
91 |
class Recurrent_block(nn.Module): |
|
|
92 |
def __init__(self,ch_out,t=2): |
|
|
93 |
super(Recurrent_block,self).__init__() |
|
|
94 |
self.t = t |
|
|
95 |
self.ch_out = ch_out |
|
|
96 |
self.conv = nn.Sequential( |
|
|
97 |
nn.Conv2d(ch_out,ch_out,kernel_size=3,stride=1,padding=1,bias=True), |
|
|
98 |
nn.BatchNorm2d(ch_out), |
|
|
99 |
nn.ReLU(inplace=True) |
|
|
100 |
) |
|
|
101 |
|
|
|
102 |
def forward(self,x): |
|
|
103 |
for i in range(self.t): |
|
|
104 |
|
|
|
105 |
if i==0: |
|
|
106 |
x1 = self.conv(x) |
|
|
107 |
|
|
|
108 |
x1 = self.conv(x+x1) |
|
|
109 |
return x1 |
|
|
110 |
|
|
|
111 |
class RRCNN_block(nn.Module): |
|
|
112 |
def __init__(self,ch_in,ch_out,t=2): |
|
|
113 |
super(RRCNN_block,self).__init__() |
|
|
114 |
self.RCNN = nn.Sequential( |
|
|
115 |
Recurrent_block(ch_out,t=t), |
|
|
116 |
Recurrent_block(ch_out,t=t) |
|
|
117 |
) |
|
|
118 |
self.Conv_1x1 = nn.Conv2d(ch_in,ch_out,kernel_size=1,stride=1,padding=0) |
|
|
119 |
|
|
|
120 |
def forward(self,x): |
|
|
121 |
x = self.Conv_1x1(x) |
|
|
122 |
x1 = self.RCNN(x) |
|
|
123 |
return x+x1 |
|
|
124 |
|
|
|
125 |
class single_conv(nn.Module): |
|
|
126 |
def __init__(self,ch_in,ch_out): |
|
|
127 |
super(single_conv,self).__init__() |
|
|
128 |
self.conv = nn.Sequential( |
|
|
129 |
nn.Conv2d(ch_in, ch_out, kernel_size=3,stride=1,padding=1,bias=True), |
|
|
130 |
nn.BatchNorm2d(ch_out), |
|
|
131 |
nn.ReLU(inplace=True) |
|
|
132 |
) |
|
|
133 |
|
|
|
134 |
def forward(self,x): |
|
|
135 |
x = self.conv(x) |
|
|
136 |
return x |
|
|
137 |
|
|
|
138 |
class Attention_block(nn.Module): |
|
|
139 |
def __init__(self,F_g,F_l,F_int): |
|
|
140 |
super(Attention_block,self).__init__() |
|
|
141 |
self.W_g = nn.Sequential( |
|
|
142 |
nn.Conv2d(F_g, F_int, kernel_size=1,stride=1,padding=0,bias=True), |
|
|
143 |
nn.BatchNorm2d(F_int) |
|
|
144 |
) |
|
|
145 |
|
|
|
146 |
self.W_x = nn.Sequential( |
|
|
147 |
nn.Conv2d(F_l, F_int, kernel_size=1,stride=1,padding=0,bias=True), |
|
|
148 |
nn.BatchNorm2d(F_int) |
|
|
149 |
) |
|
|
150 |
|
|
|
151 |
self.psi = nn.Sequential( |
|
|
152 |
nn.Conv2d(F_int, 1, kernel_size=1,stride=1,padding=0,bias=True), |
|
|
153 |
nn.BatchNorm2d(1), |
|
|
154 |
nn.Sigmoid() |
|
|
155 |
) |
|
|
156 |
|
|
|
157 |
self.relu = nn.ReLU(inplace=True) |
|
|
158 |
|
|
|
159 |
def forward(self,g,x): |
|
|
160 |
g1 = self.W_g(g) |
|
|
161 |
x1 = self.W_x(x) |
|
|
162 |
psi = self.relu(g1+x1) |
|
|
163 |
psi = self.psi(psi) |
|
|
164 |
|
|
|
165 |
return x*psi |
|
|
166 |
|
|
|
167 |
def dootsu(img): |
|
|
168 |
se = img |
|
|
169 |
_, c, _, _ = img.size() |
|
|
170 |
# squeezelayer = nn.Conv2d(c, c // 16, kernel_size=1) |
|
|
171 |
# squeezelayer.cuda() |
|
|
172 |
# img = squeezelayer(img) |
|
|
173 |
img = img.cpu().detach() |
|
|
174 |
channel = list(img.size())[1] |
|
|
175 |
batch = list(img.size())[0] |
|
|
176 |
imgfolder = img.chunk(batch, dim=0) |
|
|
177 |
chw_output = [] |
|
|
178 |
for index in range(batch): |
|
|
179 |
bchw = imgfolder[index] |
|
|
180 |
chw = bchw.squeeze() |
|
|
181 |
chwfolder = chw.chunk(channel, dim=0) |
|
|
182 |
hw_output = [] |
|
|
183 |
for i in range(channel): |
|
|
184 |
hw = chwfolder[i].squeeze() |
|
|
185 |
hw = np.transpose(hw.detach().numpy(), (0, 1)) |
|
|
186 |
hw_otsu = otsu(hw) |
|
|
187 |
hw_otsu = torch.from_numpy(hw_otsu) |
|
|
188 |
hw_output.append(hw_otsu) |
|
|
189 |
chw_otsu = torch.stack(hw_output, dim=0) |
|
|
190 |
chw_output.append(chw_otsu) |
|
|
191 |
bchw_otsu = torch.stack(chw_output, dim=0).cuda() |
|
|
192 |
# result = torch.cat([se.float().cuda(), bchw_otsu.float().cuda()],dim=1) |
|
|
193 |
return bchw_otsu |
|
|
194 |
|
|
|
195 |
class Self_Attn_OTSU(nn.Module): |
|
|
196 |
""" Self attention otsu Layer""" |
|
|
197 |
def __init__(self,in_dim,activation): |
|
|
198 |
super(Self_Attn_OTSU,self).__init__() |
|
|
199 |
self.chanel_in = in_dim |
|
|
200 |
self.activation = activation |
|
|
201 |
|
|
|
202 |
self.query_conv = nn.Conv2d(in_channels = in_dim , out_channels = in_dim//8 , kernel_size= 1) |
|
|
203 |
self.key_conv = nn.Conv2d(in_channels = in_dim , out_channels = in_dim//8 , kernel_size= 1) |
|
|
204 |
self.value_conv = nn.Conv2d(in_channels = in_dim , out_channels = in_dim , kernel_size= 1) |
|
|
205 |
self.gamma = nn.Parameter(torch.zeros(1)) |
|
|
206 |
|
|
|
207 |
self.softmax = nn.Softmax(dim=-1) # |
|
|
208 |
def forward(self,x): |
|
|
209 |
m_batchsize,C,width ,height = x.size() |
|
|
210 |
proj_query = self.query_conv(x).view(m_batchsize,-1,width*height).permute(0,2,1) # B X CX(N) |
|
|
211 |
proj_key = self.key_conv(x).view(m_batchsize,-1,width*height) # B X C x (*W*H) |
|
|
212 |
energy = torch.bmm(proj_query,proj_key) # transpose check |
|
|
213 |
attention = self.softmax(energy) # BX (N) X (N) |
|
|
214 |
proj_value = self.value_conv(x).view(m_batchsize,-1,width*height) # B X C X N |
|
|
215 |
|
|
|
216 |
out = torch.bmm(proj_value,attention.permute(0,2,1) ) |
|
|
217 |
out = out.view(m_batchsize,C,width,height) |
|
|
218 |
|
|
|
219 |
out = dootsu(out) |
|
|
220 |
|
|
|
221 |
out = self.gamma*out + x |
|
|
222 |
return out,attention |
|
|
223 |
|
|
|
224 |
def edge_conv2d(im): |
|
|
225 |
in_channel = list(im.size())[1] |
|
|
226 |
out_channel = in_channel |
|
|
227 |
# 用nn.Conv2d定义卷积操作 |
|
|
228 |
conv_op = nn.Conv2d(in_channel, out_channel, kernel_size=3, padding=1, bias=False) |
|
|
229 |
'''定义sobel算子参数,所有值除以3----有人觉得出来的图更好些;但我感觉应该是概率问题,没啥用''' |
|
|
230 |
# sobel_kernel = np.array([[-1, -1, -1], [-1, 8, -1], [-1, -1, -1]], dtype='float32') / 3 |
|
|
231 |
sobel_kernel = np.array([[-1, -1, -1], [-1, 8, -1], [-1, -1, -1]], dtype='float32') |
|
|
232 |
# 将sobel算子转换为适配卷积操作的卷积核 |
|
|
233 |
sobel_kernel = sobel_kernel.reshape((1, 1, 3, 3)) |
|
|
234 |
# 卷积输出out_channle |
|
|
235 |
sobel_kernel = np.repeat(sobel_kernel, in_channel, axis=1) |
|
|
236 |
# 输入图的通道in_channle |
|
|
237 |
sobel_kernel = np.repeat(sobel_kernel, out_channel, axis=0) |
|
|
238 |
|
|
|
239 |
# 给卷积操作的卷积核赋值 |
|
|
240 |
conv_op.weight.data = torch.from_numpy(sobel_kernel).cuda() |
|
|
241 |
# print(conv_op.weight.size()) |
|
|
242 |
# print(im.size()) |
|
|
243 |
# print(conv_op, '\n') |
|
|
244 |
|
|
|
245 |
edge_detect = conv_op(im) |
|
|
246 |
# print(torch.max(edge_detect)) |
|
|
247 |
# 将输出转换为图片格式 |
|
|
248 |
edge_detect = edge_detect.squeeze().detach() |
|
|
249 |
return edge_detect |
|
|
250 |
|
|
|
251 |
class Self_Attn_EDGE(nn.Module): |
|
|
252 |
""" Self attention edge Layer""" |
|
|
253 |
def __init__(self,in_dim,activation): |
|
|
254 |
super(Self_Attn_EDGE,self).__init__() |
|
|
255 |
self.chanel_in = in_dim |
|
|
256 |
self.activation = activation |
|
|
257 |
|
|
|
258 |
self.query_conv = nn.Conv2d(in_channels = in_dim , out_channels = in_dim//8 , kernel_size= 1) |
|
|
259 |
self.key_conv = nn.Conv2d(in_channels = in_dim , out_channels = in_dim//8 , kernel_size= 1) |
|
|
260 |
self.value_conv = nn.Conv2d(in_channels = in_dim , out_channels = in_dim , kernel_size= 1) |
|
|
261 |
self.gamma = nn.Parameter(torch.zeros(1)) |
|
|
262 |
|
|
|
263 |
self.softmax = nn.Softmax(dim=-1) # |
|
|
264 |
def forward(self,x): |
|
|
265 |
m_batchsize,C,width ,height = x.size() |
|
|
266 |
proj_query = self.query_conv(x).view(m_batchsize,-1,width*height).permute(0,2,1) # B X CX(N) |
|
|
267 |
proj_key = self.key_conv(x).view(m_batchsize,-1,width*height) # B X C x (*W*H) |
|
|
268 |
energy = torch.bmm(proj_query,proj_key) # transpose check |
|
|
269 |
attention = self.softmax(energy) # BX (N) X (N) |
|
|
270 |
proj_value = self.value_conv(x).view(m_batchsize,-1,width*height) # B X C X N |
|
|
271 |
|
|
|
272 |
out = torch.bmm(proj_value,attention.permute(0,2,1) ) |
|
|
273 |
out = out.view(m_batchsize,C,width,height) |
|
|
274 |
|
|
|
275 |
out = edge_conv2d(out) |
|
|
276 |
|
|
|
277 |
out = self.gamma*out + x |
|
|
278 |
return out,attention |
|
|
279 |
|
|
|
280 |
class Self_Attn_GABOR(nn.Module): |
|
|
281 |
""" Self attention gabor Layer""" |
|
|
282 |
def __init__(self,in_dim,activation): |
|
|
283 |
super(Self_Attn_GABOR,self).__init__() |
|
|
284 |
self.chanel_in = in_dim |
|
|
285 |
self.activation = activation |
|
|
286 |
|
|
|
287 |
self.query_conv = nn.Conv2d(in_channels = in_dim , out_channels = in_dim//8 , kernel_size= 1) |
|
|
288 |
self.key_conv = nn.Conv2d(in_channels = in_dim , out_channels = in_dim//8 , kernel_size= 1) |
|
|
289 |
self.value_conv = nn.Conv2d(in_channels = in_dim , out_channels = in_dim , kernel_size= 1) |
|
|
290 |
self.gaborlayer = GConv(in_dim, out_channels = in_dim//4, kernel_size= 1, stride=1, M=4, nScale=1, bias=False, expand=True) |
|
|
291 |
self.gamma = nn.Parameter(torch.zeros(1)) |
|
|
292 |
|
|
|
293 |
self.softmax = nn.Softmax(dim=-1) # |
|
|
294 |
def forward(self,x): |
|
|
295 |
m_batchsize,C,width ,height = x.size() |
|
|
296 |
proj_query = self.query_conv(x).view(m_batchsize,-1,width*height).permute(0,2,1) # B X CX(N) |
|
|
297 |
proj_key = self.key_conv(x).view(m_batchsize,-1,width*height) # B X C x (*W*H) |
|
|
298 |
energy = torch.bmm(proj_query,proj_key) # transpose check |
|
|
299 |
attention = self.softmax(energy) # BX (N) X (N) |
|
|
300 |
proj_value = self.value_conv(x).view(m_batchsize,-1,width*height) # B X C X N |
|
|
301 |
|
|
|
302 |
out = torch.bmm(proj_value,attention.permute(0,2,1) ) |
|
|
303 |
out = out.view(m_batchsize,C,width,height) |
|
|
304 |
# print(out.size()) |
|
|
305 |
|
|
|
306 |
out = self.gaborlayer(out) |
|
|
307 |
# print(out.size()) |
|
|
308 |
|
|
|
309 |
out = self.gamma*out + x |
|
|
310 |
return out,attention |
|
|
311 |
|
|
|
312 |
class Mymodel(nn.Module): |
|
|
313 |
def __init__(self,img_ch=3,output_ch=1): |
|
|
314 |
super(Mymodel,self).__init__() |
|
|
315 |
|
|
|
316 |
self.Maxpool = nn.MaxPool2d(kernel_size=2,stride=2) |
|
|
317 |
|
|
|
318 |
self.Conv1 = conv_block(ch_in=img_ch,ch_out=64) |
|
|
319 |
self.Conv2 = conv_block(ch_in=64,ch_out=128) |
|
|
320 |
self.Conv3 = conv_block(ch_in=128,ch_out=256) |
|
|
321 |
self.Conv4 = conv_block(ch_in=256,ch_out=512) |
|
|
322 |
self.Conv5 = conv_block(ch_in=512,ch_out=1024) |
|
|
323 |
|
|
|
324 |
self.Up5 = up_conv(ch_in=1024,ch_out=512) |
|
|
325 |
self.Att5 = Attention_block(F_g=512,F_l=512,F_int=256) |
|
|
326 |
self.Up_conv5 = conv_block(ch_in=1024, ch_out=512) |
|
|
327 |
|
|
|
328 |
self.Up4 = up_conv(ch_in=512,ch_out=256) |
|
|
329 |
self.Att4 = Attention_block(F_g=256,F_l=256,F_int=128) |
|
|
330 |
self.Up_conv4 = conv_block(ch_in=512, ch_out=256) |
|
|
331 |
|
|
|
332 |
self.Up3 = up_conv(ch_in=256,ch_out=128) |
|
|
333 |
self.Att3 = Attention_block(F_g=128,F_l=128,F_int=64) |
|
|
334 |
self.Up_conv3 = conv_block(ch_in=256, ch_out=128) |
|
|
335 |
|
|
|
336 |
self.Up2 = up_conv(ch_in=128,ch_out=64) |
|
|
337 |
self.Att2 = Attention_block(F_g=64,F_l=64,F_int=32) |
|
|
338 |
self.Up_conv2 = conv_block(ch_in=128, ch_out=64) |
|
|
339 |
|
|
|
340 |
# # gabor |
|
|
341 |
# self.gaborlayer = GConv(in_channels=64, out_channels = 16, kernel_size= 1, stride=1, M=4, nScale=1, bias=False, expand=True) |
|
|
342 |
|
|
|
343 |
self.pred1 = nn.Conv2d(32,output_ch,kernel_size=1,stride=1,padding=0) |
|
|
344 |
self.pred2 = nn.Conv2d(32,output_ch,kernel_size=1,stride=1,padding=0) |
|
|
345 |
self.pred3 = nn.Conv2d(32,output_ch,kernel_size=1,stride=1,padding=0) |
|
|
346 |
|
|
|
347 |
# self-attention |
|
|
348 |
self.pred_1 = single_conv(ch_in=64, ch_out=32) |
|
|
349 |
self.pred_2 = single_conv(ch_in=64, ch_out=32) |
|
|
350 |
self.pred_3 = single_conv(ch_in=64, ch_out=32) |
|
|
351 |
self.conv_Atten1 = Self_Attn_OTSU(32,'relu') |
|
|
352 |
self.conv_Atten2 = Self_Attn_GABOR(32,'relu') |
|
|
353 |
# self.conv_Atten3 = Self_Attn(32,'relu') |
|
|
354 |
self.conv_Atten3 = Self_Attn_GABOR(32,'relu') |
|
|
355 |
|
|
|
356 |
# fusion module |
|
|
357 |
self.conv_fusion1 = conv_block(ch_in=160, ch_out=64) |
|
|
358 |
self.conv_fusion2 = nn.Conv2d(64, output_ch, kernel_size=1, stride=1, padding=0) # v1 |
|
|
359 |
|
|
|
360 |
|
|
|
361 |
def forward(self,x): |
|
|
362 |
# encoding path |
|
|
363 |
x1 = self.Conv1(x) |
|
|
364 |
|
|
|
365 |
x2 = self.Maxpool(x1) |
|
|
366 |
x2 = self.Conv2(x2) |
|
|
367 |
|
|
|
368 |
x3 = self.Maxpool(x2) |
|
|
369 |
x3 = self.Conv3(x3) |
|
|
370 |
|
|
|
371 |
x4 = self.Maxpool(x3) |
|
|
372 |
x4 = self.Conv4(x4) |
|
|
373 |
|
|
|
374 |
x5 = self.Maxpool(x4) |
|
|
375 |
x5 = self.Conv5(x5) |
|
|
376 |
|
|
|
377 |
# decoding + concat path |
|
|
378 |
d5 = self.Up5(x5) |
|
|
379 |
x4 = self.Att5(g=d5,x=x4) |
|
|
380 |
d5 = torch.cat((x4,d5),dim=1) |
|
|
381 |
d5 = self.Up_conv5(d5) |
|
|
382 |
|
|
|
383 |
d4 = self.Up4(d5) |
|
|
384 |
x3 = self.Att4(g=d4,x=x3) |
|
|
385 |
d4 = torch.cat((x3,d4),dim=1) |
|
|
386 |
d4 = self.Up_conv4(d4) |
|
|
387 |
|
|
|
388 |
d3 = self.Up3(d4) |
|
|
389 |
x2 = self.Att3(g=d3,x=x2) |
|
|
390 |
d3 = torch.cat((x2,d3),dim=1) |
|
|
391 |
d3 = self.Up_conv3(d3) |
|
|
392 |
|
|
|
393 |
d2 = self.Up2(d3) |
|
|
394 |
x1 = self.Att2(g=d2,x=x1) |
|
|
395 |
d2 = torch.cat((x1,d2),dim=1) |
|
|
396 |
d2 = self.Up_conv2(d2) |
|
|
397 |
|
|
|
398 |
# d2_gabor = self.gaborlayer(d2) |
|
|
399 |
|
|
|
400 |
pred_1 = self.pred_1(d2) |
|
|
401 |
pred_2 = self.pred_2(d2) |
|
|
402 |
pred_3 = self.pred_3(d2) |
|
|
403 |
pred1 = self.pred1(pred_1) |
|
|
404 |
pred2 = self.pred2(pred_2) |
|
|
405 |
pred3 = self.pred3(pred_3) |
|
|
406 |
|
|
|
407 |
# self-attention |
|
|
408 |
attention_higher, _ = self.conv_Atten1(pred_1) |
|
|
409 |
attention_lower, _ = self.conv_Atten2(pred_2) |
|
|
410 |
attention_all, _ = self.conv_Atten3(pred_3) |
|
|
411 |
|
|
|
412 |
# fusion module |
|
|
413 |
y = torch.cat((attention_higher, attention_all, attention_lower, d2), dim=1) |
|
|
414 |
y = self.conv_fusion1(y) |
|
|
415 |
pred = self.conv_fusion2(y) |
|
|
416 |
|
|
|
417 |
return [pred1, pred2, pred3, pred] |