-
Notifications
You must be signed in to change notification settings - Fork 21
/
fusion.py
336 lines (273 loc) · 15.1 KB
/
fusion.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
323
324
325
326
327
328
329
330
331
332
333
334
335
import torch
import torch.nn as nn
from torch.autograd import Variable
import torch.nn.functional as F
'''
fusion: consits of DRB, DMSW, RAM.
'''
class ConvLSTMCell(nn.Module):
def __init__(self, input_channels, hidden_channels, kernel_size, bias=True):
super(ConvLSTMCell, self).__init__()
assert hidden_channels % 2 == 0
self.input_channels = input_channels
self.hidden_channels = hidden_channels
self.bias = bias
self.kernel_size = kernel_size
self.num_features = 4
self.padding = (kernel_size - 1) //2
self.conv = nn.Conv2d(self.input_channels + self.hidden_channels, 4 * self.hidden_channels, self.kernel_size, 1,
self.padding)
self._initialize_weights()
def _initialize_weights(self):
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.normal(m.weight.data, std=0.01)
if m.bias is not None:
m.bias.data.zero_()
def forward(self, input, h, c):
combined = torch.cat((input, h), dim=1)
A = self.conv(combined)
(ai, af, ao, ag) = torch.split(A, A.size()[1] // self.num_features, dim=1)
i = torch.sigmoid(ai) #input gate
f = torch.sigmoid(af) #forget gate
o = torch.sigmoid(ao) #output
g = torch.tanh(ag) #update_Cell
new_c = f * c + i * g
new_h = o * torch.tanh(new_c)
return new_h, new_c, o
@staticmethod
def init_hidden(batch_size, hidden_c, shape):
return (Variable(torch.zeros(batch_size, hidden_c, shape[0], shape[1])).cuda(),
Variable(torch.zeros(batch_size, hidden_c, shape[0], shape[1])).cuda())
class ConvLSTM(nn.Module):
def __init__(self, input_channels, hidden_channels, kernel_size, step=1, effective_step=[1], bias=True):
super(ConvLSTM, self).__init__()
self.input_channels = [input_channels] + hidden_channels
self.hidden_channels = hidden_channels
self.kernel_size = kernel_size
self.num_layers = len(hidden_channels)
self.step = step
self.bias = bias
self.effective_step = effective_step
self._all_layers = []
for i in range(self.num_layers):
name = 'cell{}'.format(i)
cell = ConvLSTMCell(self.input_channels[i], self.hidden_channels[i], self.kernel_size, self.bias)
setattr(self, name, cell)
self._all_layers.append(cell)
# --------------------------- Depth Refinement Block -------------------------- #
# DRB 1
self.conv_refine1_1 = nn.Conv2d(64, 64, 3, padding=1)
self.bn_refine1_1 = nn.BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True)
self.relu_refine1_1 = nn.PReLU()
self.conv_refine1_2 = nn.Conv2d(64, 64, 3, padding=1)
self.bn_refine1_2 = nn.BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True)
self.relu_refine1_2 = nn.PReLU()
self.conv_refine1_3 = nn.Conv2d(64, 64, 3, padding=1)
self.bn_refine1_3 = nn.BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True)
self.relu_refine1_3 = nn.PReLU()
self.down_2_1 = nn.MaxPool2d(2, stride=2, ceil_mode=True)
self.down_2_2 = nn.MaxPool2d(2, stride=2, ceil_mode=True)
# DRB 2
self.conv_refine2_1 = nn.Conv2d(128, 128, 3, padding=1)
self.bn_refine2_1 = nn.BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True)
self.relu_refine2_1 = nn.PReLU()
self.conv_refine2_2 = nn.Conv2d(128, 128, 3, padding=1)
self.bn_refine2_2 = nn.BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True)
self.relu_refine2_2 = nn.PReLU()
self.conv_refine2_3 = nn.Conv2d(128, 128, 3, padding=1)
self.bn_refine2_3 = nn.BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True)
self.relu_refine2_3 = nn.PReLU()
self.conv_r2_1 = nn.Conv2d(128, 64, 3, padding=1)
self.bn_r2_1 = nn.BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True)
self.relu_r2_1 = nn.PReLU()
# DRB 3
self.conv_refine3_1 = nn.Conv2d(256, 256, 3, padding=1)
self.bn_refine3_1 = nn.BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True)
self.relu_refine3_1 = nn.PReLU()
self.conv_refine3_2 = nn.Conv2d(256, 256, 3, padding=1)
self.bn_refine3_2 = nn.BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True)
self.relu_refine3_2 = nn.PReLU()
self.conv_refine3_3 = nn.Conv2d(256, 256, 3, padding=1)
self.bn_refine3_3 = nn.BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True)
self.relu_refine3_3 = nn.PReLU()
self.conv_r3_1 = nn.Conv2d(256, 64, 3, padding=1)
self.bn_r3_1 = nn.BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True)
self.relu_r3_1 = nn.PReLU()
# DRB 4
self.conv_refine4_1 = nn.Conv2d(512, 512, 3, padding=1)
self.bn_refine4_1 = nn.BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True)
self.relu_refine4_1 = nn.PReLU()
self.conv_refine4_2 = nn.Conv2d(512, 512, 3, padding=1)
self.bn_refine4_2 = nn.BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True)
self.relu_refine4_2 = nn.PReLU()
self.conv_refine4_3 = nn.Conv2d(512, 512, 3, padding=1)
self.bn_refine4_3 = nn.BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True)
self.relu_refine4_3 = nn.PReLU()
self.conv_r4_1 = nn.Conv2d(512, 64, 3, padding=1)
self.bn_r4_1 = nn.BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True)
self.relu_r4_1 = nn.PReLU()
# DRB 5
self.conv_refine5_1 = nn.Conv2d(512, 512, 3, padding=1)
self.bn_refine5_1 = nn.BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True)
self.relu_refine5_1 = nn.PReLU()
self.conv_refine5_2 = nn.Conv2d(512, 512, 3, padding=1)
self.bn_refine5_2 = nn.BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True)
self.relu_refine5_2 = nn.PReLU()
self.conv_refine5_3 = nn.Conv2d(512, 512, 3, padding=1)
self.bn_refine5_3 = nn.BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True)
self.relu_refine5_3 = nn.PReLU()
self.conv_r5_1 = nn.Conv2d(512, 64, 3, padding=1)
self.bn_r5_1 = nn.BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True)
self.relu_r5_1 = nn.PReLU()
# ----------------------------- Multi-scale ----------------------------- #
# Add new structure: ASPP Atrous spatial Pyramid Pooling based on DeepLab v3
# part0: 1*1*64 Conv
self.conv5_conv_1 = nn.Conv2d(64, 64, 1, padding=0) # size: 64*64*64
self.bn5_conv_1 = nn.BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True)
self.relu5_conv_1 = nn.ReLU(inplace=True)
# part1: 3*3*64 Conv
self.conv5_conv = nn.Conv2d(64, 64, 3, padding=1) # size: 64*64*64
self.bn5_conv = nn.BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True)
self.relu5_conv = nn.ReLU(inplace=True)
# part2: 3*3*64 (dilated=7) Atrous Conv
self.Atrous_conv_1 = nn.Conv2d(64, 64, 3, padding=7, dilation=7) # size: 64*64*64
self.Atrous_bn5_1 = nn.BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True)
self.Atrous_relu_1 = nn.ReLU(inplace=True)
# part3: 3*3*64 (dilated=5) Atrous Conv
self.Atrous_conv_2 = nn.Conv2d(64, 64, 3, padding=5, dilation=5) # size: 64*64*64
self.Atrous_bn5_2 = nn.BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True)
self.Atrous_relu_2 = nn.ReLU(inplace=True)
# part4: 3*3*64 (dilated=3) Atrous Conv
self.Atrous_conv_5 = nn.Conv2d(64, 64, 3, padding=3, dilation=3) # size: 64*64*64
self.Atrous_bn5_5 = nn.BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True)
self.Atrous_relu_5 = nn.ReLU(inplace=True)
# part5: Max_pooling # size: 16*16*64
self.Atrous_pooling = nn.MaxPool2d(2, stride=2, ceil_mode=True)
self.Atrous_conv_pool = nn.Conv2d(64, 64, 1, padding=0)
self.Atrous_bn_pool = nn.BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True)
self.Atrous_relu_pool = nn.ReLU(inplace=True)
# ----------------------------- Channel-wise Attention ----------------------------- #
self.conv_c = nn.Conv2d(64, 64, 3, padding=1)
self.conv_h = nn.Conv2d(64, 64, 3, padding=1)
self.pool_avg = nn.AvgPool2d(64, stride=2, ceil_mode=True) # 1/8
# ----------------------------- Sptatial-wise Attention ----------------------------- #
self.conv_s1 = nn.Conv2d(64 * self.num_layers, 64, 1, padding=0)
self.conv_s2 = nn.Conv2d(64 * self.num_layers, 1, 1, padding=0)
# ----------------------------- Prediction ----------------------------- #
self.conv_pred = nn.Conv2d(64, 2, 1, padding=0)
self._initialize_weights()
def _initialize_weights(self):
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.normal(m.weight.data, std=0.01)
if m.bias is not None:
m.bias.data.zero_()
def forward(self,depth_vector,h1,h2,h3,h4,h5,d1,d2,d3,d4,d5):
internal_state = []
# -------- apply DRB --------- #
# drb 1
d1_1 = self.relu_refine1_1(self.bn_refine1_1(self.conv_refine1_1(d1)))
d1_2 = self.relu_refine1_2(self.bn_refine1_2(self.conv_refine1_2(d1_1)))
d1_2 = d1_2 + h1 # (256x256)*64
d1_2 = self.down_2_2(self.down_2_1(d1_2))
d1_2_0 = d1_2
d1_3 = self.relu_refine1_3(self.bn_refine1_3(self.conv_refine1_3(d1_2)))
drb1 = d1_2_0 + d1_3 # (64 x 64)*64
# drb 2
d2_1 = self.relu_refine2_1(self.bn_refine2_1(self.conv_refine2_1(d2)))
d2_2 = self.relu_refine2_2(self.bn_refine2_2(self.conv_refine2_2(d2_1)))
d2_2 = d2_2 + h2 # (128x128)*128
d2_2 = self.down_2_1(d2_2)
d2_2_0 = d2_2
d2_3 = self.relu_refine2_3(self.bn_refine2_3(self.conv_refine2_3(d2_2)))
drb2 = d2_2_0 + d2_3
drb2 = self.relu_r2_1(self.bn_r2_1(self.conv_r2_1(drb2))) # (64 x 64)*64
# drb 3
d3_1 = self.relu_refine3_1(self.bn_refine3_1(self.conv_refine3_1(d3)))
d3_2 = self.relu_refine3_2(self.bn_refine3_2(self.conv_refine3_2(d3_1)))
d3_2 = d3_2 + h3 # (64 x 64)*256
d3_2_0 = d3_2
d3_3 = self.relu_refine3_3(self.bn_refine3_3(self.conv_refine3_3(d3_2)))
drb3 = d3_2_0 + d3_3
drb3 = self.relu_r3_1(self.bn_r3_1(self.conv_r3_1(drb3))) # (64 x 64)*64
# drb 4
d4_1 = self.relu_refine4_1(self.bn_refine4_1(self.conv_refine4_1(d4)))
d4_2 = self.relu_refine4_2(self.bn_refine4_2(self.conv_refine4_2(d4_1)))
d4_2 = d4_2 + h4 # (32 x 32)*512
d4_2 = F.upsample(d4_2, scale_factor=2, mode='bilinear')
d4_2_0 = d4_2
d4_3 = self.relu_refine4_3(self.bn_refine4_3(self.conv_refine4_3(d4_2)))
drb4 = d4_2_0 + d4_3
drb4 = self.relu_r4_1(self.bn_r4_1(self.conv_r4_1(drb4))) # (64 x 64)*64
# drb 5
d5_1 = self.relu_refine5_1(self.bn_refine5_1(self.conv_refine5_1(d5)))
d5_2 = self.relu_refine5_2(self.bn_refine5_2(self.conv_refine5_2(d5_1)))
d5_2 = d5_2 + h5 # (16 x 16)*64
d5_2 = F.upsample(d5_2, scale_factor=4, mode='bilinear')
d5_2_0 = d5_2
d5_3 = self.relu_refine5_3(self.bn_refine5_3(self.conv_refine5_3(d5_2)))
drb5 = d5_2_0 + d5_3
drb5 = self.relu_r5_1(self.bn_r5_1(self.conv_r5_1(drb5))) # (64 x 64)*64
drb_fusion = drb1 +drb2 + drb3 +drb4 +drb5 # (64 x 64)*64
# --------------------- obtain multi-scale ----------------------- #
f1 = self.relu5_conv_1(self.bn5_conv_1(self.conv5_conv_1(drb_fusion)))
f2 = self.relu5_conv(self.bn5_conv(self.conv5_conv(drb_fusion)))
f3 = self.Atrous_relu_1(self.Atrous_bn5_1(self.Atrous_conv_1(drb_fusion)))
f4 = self.Atrous_relu_2(self.Atrous_bn5_2(self.Atrous_conv_2(drb_fusion)))
f5 = self.Atrous_relu_5(self.Atrous_bn5_5(self.Atrous_conv_5(drb_fusion)))
f6 = F.upsample(
self.Atrous_relu_pool(self.Atrous_bn_pool(self.Atrous_conv_pool(self.Atrous_pooling(self.Atrous_pooling(drb_fusion))))),
scale_factor=4, mode='bilinear')
fusion = torch.cat([f1,f2,f3,f4,f5,f6],dim=0) # 6x64x64x64
fusion_o = fusion
input = torch.cat(torch.chunk(fusion, 6, dim=0), dim=1)
for step in range(self.step):
depth = depth_vector # 1x 6 x 1 x1
if step == 0:
basize, _, height, width = input.size()
(h_step, c) = ConvLSTMCell.init_hidden(basize, self.hidden_channels[self.num_layers-1],(height, width))
# Feature-wise Attention
depth = torch.mul(F.softmax(depth,dim=1), 6)
basize, dime, h, w = depth.size()
depth = depth.view(1, basize, dime, h, w).transpose(0,1).transpose(1,2)
depth = torch.cat(torch.chunk(depth, basize, dim=0), dim=1).view(basize*dime, 1, 1, 1)
depth = torch.mul(fusion_o, depth).view(1, basize*dime, 64, 64, 64)
depth = torch.cat(torch.chunk(depth, basize, dim=1), dim=0)
F_sum = torch.sum(depth, 1, keepdim=False)#.squeeze()
# Channel-wise Attention
depth_fw_ori = F_sum
depth = self.conv_c(F_sum)
h_c = self.conv_h(h_step)
depth = depth + h_c
depth = self.pool_avg(depth)
depth = torch.mul(F.softmax(depth, dim=1), 64)
F_sum_wt = torch.mul(depth_fw_ori, depth)
x = F_sum_wt
if step < self.step-1:
for i in range(self.num_layers):
# all cells are initialized in the first step
if step == 0:
bsize, _, height, width = x.size()
(h, c) = ConvLSTMCell.init_hidden(bsize, self.hidden_channels[i], (height, width))
internal_state.append((h, c))
# do forward
name = 'cell{}'.format(i)
(h, c) = internal_state[i]
h_step = h
x, new_c, new_o = getattr(self, name)(x, h, c) # ConvLSTMCell forward
internal_state[i] = (x, new_c)
# only record effective steps
#if step in self.effective_step:
if step == 0:
outputs_o = new_o
else:
outputs_o = torch.cat((outputs_o, new_o), dim=1)
# ---------------> Spatial-wise Attention Module <----------------- #
outputs = self.conv_s1(outputs_o)
spatial_weight = F.sigmoid(self.conv_s2(outputs_o))
outputs = torch.mul(outputs,spatial_weight)
# -------------------------> Prediction <-------------------------- #
outputs = self.conv_pred(outputs)
output = F.upsample(outputs, scale_factor=4, mode='bilinear')
return output