-
Notifications
You must be signed in to change notification settings - Fork 4
/
Copy pathmodules.py
346 lines (263 loc) · 13.1 KB
/
modules.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
336
337
338
339
340
341
342
343
344
345
346
import torch
import torch.nn.functional as F
from torch import nn
class DGNL(nn.Module):
def __init__(self, in_channels):
super(DGNL, self).__init__()
self.eps = 1e-6
self.sigma_pow2 = 100
self.theta = nn.Conv2d(in_channels, int(in_channels / 2), kernel_size=1)
self.phi = nn.Conv2d(in_channels, int(in_channels / 2), kernel_size=1)
self.g = nn.Conv2d(in_channels, int(in_channels / 2), kernel_size=1)
self.down = nn.Conv2d(in_channels, in_channels, kernel_size=4, stride=4, groups=in_channels, bias=False)
self.down.weight.data.fill_(1. / 16)
self.z = nn.Conv2d(int(in_channels / 2), in_channels, kernel_size=1)
def forward(self, x, depth_map):
n, c, h, w = x.size()
x_down = self.down(x)
# [n, (h / 8) * (w / 8), c / 2]
g = F.max_pool2d(self.g(x_down), kernel_size=2, stride=2).view(n, int(c / 2), -1).transpose(1, 2)
### appearance relation map
# [n, (h / 4) * (w / 4), c / 2]
theta = self.theta(x_down).view(n, int(c / 2), -1).transpose(1, 2)
# [n, c / 2, (h / 8) * (w / 8)]
phi = F.max_pool2d(self.phi(x_down), kernel_size=2, stride=2).view(n, int(c / 2), -1)
# [n, (h / 4) * (w / 4), (h / 8) * (w / 8)]
Ra = F.softmax(torch.bmm(theta, phi), 2)
### depth relation map
depth1 = F.interpolate(depth_map, size=[int(h / 4), int(w / 4)], mode='bilinear', align_corners = True).view(n, 1, int(h / 4)*int(w / 4)).transpose(1,2)
depth2 = F.interpolate(depth_map, size=[int(h / 8), int(w / 8)], mode='bilinear', align_corners = True).view(n, 1, int(h / 8)*int(w / 8))
# n, (h / 4) * (w / 4), (h / 8) * (w / 8)
depth1_expand = depth1.expand(n, int(h / 4) * int(w / 4), int(h / 8) * int(w / 8))
depth2_expand = depth2.expand(n, int(h / 4) * int(w / 4), int(h / 8) * int(w / 8))
Rd = torch.min(depth1_expand / (depth2_expand + self.eps), depth2_expand / (depth1_expand + self.eps))
# normalization: depth relation map [n, (h / 4) * (w / 4), (h / 8) * (w / 8)]
# Rd = Rd / (torch.sum(Rd, 2).view(n, int(h / 4) * int(w / 4), 1) + self.eps)
Rd = F.softmax(Rd, 2)
# ### position relation map
# position_h = torch.Tensor(range(h)).cuda().view(h, 1).expand(h, w)
# position_w = torch.Tensor(range(w)).cuda().view(1, w).expand(h, w)
#
# position_h1 = F.interpolate(position_h.unsqueeze(0).unsqueeze(0), size=[int(h / 4), int(w / 4)], mode='bilinear', align_corners=True).view(1, 1, int(h / 4) * int(w / 4)).transpose(1,2)
# position_h2 = F.interpolate(position_h.unsqueeze(0).unsqueeze(0), size=[int(h / 8), int(w / 8)], mode='bilinear', align_corners=True).view(1, 1, int(h / 8) * int(w / 8))
# position_h1_expand = position_h1.expand(n, int(h / 4) * int(w / 4), int(h / 8) * int(w / 8))
# position_h2_expand = position_h2.expand(n, int(h / 4) * int(w / 4), int(h / 8) * int(w / 8))
# h_distance = (position_h1_expand - position_h2_expand).pow(2)
#
# position_w1 = F.interpolate(position_w.unsqueeze(0).unsqueeze(0), size=[int(h / 4), int(w / 4)], mode='bilinear', align_corners=True).view(1, 1, int(h / 4) * int(w / 4)).transpose(1, 2)
# position_w2 = F.interpolate(position_w.unsqueeze(0).unsqueeze(0), size=[int(h / 8), int(w / 8)], mode='bilinear', align_corners=True).view(1, 1, int(h / 8) * int(w / 8))
# position_w1_expand = position_w1.expand(n, int(h / 4) * int(w / 4), int(h / 8) * int(w / 8))
# position_w2_expand = position_w2.expand(n, int(h / 4) * int(w / 4), int(h / 8) * int(w / 8))
# w_distance = (position_w1_expand - position_w2_expand).pow(2)
#
# Rp = 1 / (2 * 3.14159265 * self.sigma_pow2) * torch.exp(-0.5 * (h_distance / self.sigma_pow2 + w_distance / self.sigma_pow2))
#
# Rp = Rp / (torch.sum(Rp, 2).view(n, int(h / 4) * int(w / 4), 1) + self.eps)
### overal relation map
#S = F.softmax(Ra * Rd * Rp, 2)
S = F.softmax(Ra * Rd, 2)
# [n, c / 2, h / 4, w / 4]
y = torch.bmm(S, g).transpose(1, 2).contiguous().view(n, int(c / 2), int(h / 4), int(w / 4))
return x + F.upsample(self.z(y), size=x.size()[2:], mode='bilinear', align_corners = True)
class NLB(nn.Module):
def __init__(self, in_channels):
super(NLB, self).__init__()
self.theta = nn.Conv2d(in_channels, int(in_channels / 2), kernel_size=1)
self.phi = nn.Conv2d(in_channels, int(in_channels / 2), kernel_size=1)
self.g = nn.Conv2d(in_channels, int(in_channels / 2), kernel_size=1)
self.down = nn.Conv2d(in_channels, in_channels, kernel_size=4, stride=4, groups=in_channels, bias=False)
self.down.weight.data.fill_(1. / 16)
self.z = nn.Conv2d(int(in_channels / 2), in_channels, kernel_size=1)
def forward(self, x):
n, c, h, w = x.size()
x_down = self.down(x)
# [n, (h / 4) * (w / 4), c / 2]
theta = self.theta(x_down).view(n, int(c / 2), -1).transpose(1, 2)
# [n, c / 2, (h / 8) * (w / 8)]
phi = F.max_pool2d(self.phi(x_down), kernel_size=2, stride=2).view(n, int(c / 2), -1)
# [n, (h / 8) * (w / 8), c / 2]
g = F.max_pool2d(self.g(x_down), kernel_size=2, stride=2).view(n, int(c / 2), -1).transpose(1, 2)
# [n, (h / 4) * (w / 4), (h / 8) * (w / 8)]
f = F.softmax(torch.bmm(theta, phi), 2)
# [n, c / 2, h / 4, w / 4]
y = torch.bmm(f, g).transpose(1, 2).contiguous().view(n, int(c / 2), int(h / 4), int(w / 4))
return x + F.upsample(self.z(y), size=x.size()[2:], mode='bilinear', align_corners=True)
class DepthWiseDilatedResidualBlock(nn.Module):
def __init__(self, reduced_channels, channels, dilation):
super(DepthWiseDilatedResidualBlock, self).__init__()
self.conv0 = nn.Sequential(
# pw
nn.Conv2d(channels, channels * 2, 1, 1, 0, 1, bias=False),
nn.ReLU6(inplace=True),
# dw
nn.Conv2d(channels*2, channels*2, kernel_size=3, padding=dilation, dilation=dilation, groups=channels, bias=False),
nn.ReLU6(inplace=True),
# pw-linear
nn.Conv2d(channels*2, channels, 1, 1, 0, 1, 1, bias=False)
)
self.conv1 = nn.Sequential(
# pw
# nn.Conv2d(channels, channels * 2, 1, 1, 0, 1, bias=False),
# nn.ReLU6(inplace=True),
# dw
nn.Conv2d(channels, channels, kernel_size=3, padding=dilation, dilation=dilation, groups=channels,
bias=False),
nn.ReLU6(inplace=True),
# pw-linear
nn.Conv2d(channels, channels, 1, 1, 0, 1, 1, bias=False)
)
def forward(self, x):
res = self.conv1(self.conv0(x))
return res + x
class DilatedResidualBlock(nn.Module):
def __init__(self, channels, dilation):
super(DilatedResidualBlock, self).__init__()
self.conv0 = nn.Sequential(
nn.Conv2d(channels, channels, kernel_size=3, padding=dilation, dilation=dilation), nn.ReLU()
)
self.conv1 = nn.Conv2d(channels, channels, kernel_size=3, padding=dilation, dilation=dilation)
def forward(self, x):
conv0 = self.conv0(x)
conv1 = self.conv1(conv0)
return x + conv1
class SpatialRNN(nn.Module):
"""
SpatialRNN model for one direction only
"""
def __init__(self, alpha = 1.0, channel_num = 1, direction = "right"):
super(SpatialRNN, self).__init__()
self.alpha = nn.Parameter(torch.Tensor([alpha] * channel_num))
self.direction = direction
def __getitem__(self, item):
return self.alpha[item]
def __len__(self):
return len(self.alpha)
def forward(self, x):
"""
:param x: (N,C,H,W)
:return:
"""
height = x.size(2)
weight = x.size(3)
x_out = []
# from left to right
if self.direction == "right":
x_out = [x[:, :, :, 0].clamp(min=0)]
for i in range(1, weight):
temp = (self.alpha.unsqueeze(1) * x_out[i - 1] + x[:, :, :, i]).clamp(min=0)
x_out.append(temp) # a list of tensor
return torch.stack(x_out, 3) # merge into one tensor
# from right to left
elif self.direction == "left":
x_out = [x[:, :, :, -1].clamp(min=0)]
for i in range(1, weight):
temp = (self.alpha.unsqueeze(1) * x_out[i - 1] + x[:, :, :, -i - 1]).clamp(min=0)
x_out.append(temp)
x_out.reverse()
return torch.stack(x_out, 3)
# from up to down
elif self.direction == "down":
x_out = [x[:, :, 0, :].clamp(min=0)]
for i in range(1, height):
temp = (self.alpha.unsqueeze(1) * x_out[i - 1] + x[:, :, i, :]).clamp(min=0)
x_out.append(temp)
return torch.stack(x_out, 2)
# from down to up
elif self.direction == "up":
x_out = [x[:, :, -1, :].clamp(min=0)]
for i in range(1, height):
temp = (self.alpha.unsqueeze(1) * x_out[i - 1] + x[:, :, -i - 1, :]).clamp(min=0)
x_out.append(temp)
x_out.reverse()
return torch.stack(x_out, 2)
else:
print("Invalid direction in SpatialRNN!")
return KeyError
class TVLoss(nn.Module):
def __init__(self, tv_loss_weight=1):
super(TVLoss, self).__init__()
self.tv_loss_weight = tv_loss_weight
def forward(self, x):
batch_size = x.size()[0]
h_x = x.size()[2]
w_x = x.size()[3]
count_h = self.tensor_size(x[:, :, 1:, :])
count_w = self.tensor_size(x[:, :, :, 1:])
h_tv = torch.pow((x[:, :, 1:, :] - x[:, :, :h_x - 1, :]), 2).sum()
w_tv = torch.pow((x[:, :, :, 1:] - x[:, :, :, :w_x - 1]), 2).sum()
return self.tv_loss_weight * 2 * (h_tv / count_h + w_tv / count_w) / batch_size
@staticmethod
def tensor_size(t):
return t.size()[1] * t.size()[2] * t.size()[3]
class NLB(nn.Module):
def __init__(self, in_channels):
super(NLB, self).__init__()
self.theta = nn.Conv2d(in_channels, int(in_channels / 2), kernel_size=1)
self.phi = nn.Conv2d(in_channels, int(in_channels / 2), kernel_size=1)
self.g = nn.Conv2d(in_channels, int(in_channels / 2), kernel_size=1)
self.down = nn.Conv2d(in_channels, in_channels, kernel_size=4, stride=4, groups=in_channels, bias=False)
self.down.weight.data.fill_(1. / 16)
self.z = nn.Conv2d(int(in_channels / 2), in_channels, kernel_size=1)
def forward(self, x):
n, c, h, w = x.size()
x_down = self.down(x)
# [n, (h / 4) * (w / 4), c / 2]
theta = self.theta(x_down).view(n, int(c / 2), -1).transpose(1, 2)
# [n, c / 2, (h / 8) * (w / 8)]
phi = F.max_pool2d(self.phi(x_down), kernel_size=2, stride=2).view(n, int(c / 2), -1)
# [n, (h / 8) * (w / 8), c / 2]
g = F.max_pool2d(self.g(x_down), kernel_size=2, stride=2).view(n, int(c / 2), -1).transpose(1, 2)
# [n, (h / 4) * (w / 4), (h / 8) * (w / 8)]
f = F.softmax(torch.bmm(theta, phi), 2)
# [n, c / 2, h / 4, w / 4]
y = torch.bmm(f, g).transpose(1, 2).contiguous().view(n, int(c / 2), int(h / 4), int(w / 4))
return x + F.upsample(self.z(y), size=x.size()[2:], mode='bilinear', align_corners=True)
# class DGNLB(nn.Module):
# def __init__(self, in_channels):
# super(DGNLB, self).__init__()
#
# self.roll = nn.Conv2d(1, int(in_channels / 2), kernel_size=1)
# self.ita = nn.Conv2d(1, int(in_channels / 2), kernel_size=1)
#
# self.theta = nn.Conv2d(in_channels, int(in_channels / 2), kernel_size=1)
# self.phi = nn.Conv2d(in_channels, int(in_channels / 2), kernel_size=1)
# self.g = nn.Conv2d(in_channels, int(in_channels / 2), kernel_size=1)
#
# self.down = nn.Conv2d(in_channels, in_channels, kernel_size=4, stride=4, groups=in_channels, bias=False)
# self.down.weight.data.fill_(1. / 16)
#
# #self.down_depth = nn.Conv2d(1, 1, kernel_size=4, stride=4, groups=in_channels, bias=False)
# #self.down_depth.weight.data.fill_(1. / 16)
#
# self.z = nn.Conv2d(int(in_channels / 2), in_channels, kernel_size=1)
#
# def forward(self, x, depth):
# n, c, h, w = x.size()
# x_down = self.down(x)
#
# depth_down = F.avg_pool2d(depth, kernel_size=(4,4))
#
# # [n, (h / 4) * (w / 4), c / 2]
# #roll = self.roll(depth_down).view(n, int(c / 2), -1).transpose(1, 2)
# # [n, c / 2, (h / 4) * (w / 4)]
# #ita = self.ita(depth_down).view(n, int(c / 2), -1)
# # [n, (h / 4) * (w / 4), (h / 4) * (w / 4)]
#
# depth_correlation = F.softmax(torch.bmm(depth_down.view(n, 1, -1).transpose(1, 2), depth_down.view(n, 1, -1)), 2)
#
#
# # [n, (h / 4) * (w / 4), c / 2]
# theta = self.theta(x_down).view(n, int(c / 2), -1).transpose(1, 2)
# # [n, c / 2, (h / 8) * (w / 8)]
# phi = F.max_pool2d(self.phi(x_down), kernel_size=2, stride=2).view(n, int(c / 2), -1)
# # [n, (h / 8) * (w / 8), c / 2]
# g = F.max_pool2d(self.g(x_down), kernel_size=2, stride=2).view(n, int(c / 2), -1).transpose(1, 2)
# # [n, (h / 4) * (w / 4), (h / 8) * (w / 8)]
# f_correlation = F.softmax(torch.bmm(theta, phi), 2)
# # [n, (h / 4) * (w / 4), (h / 8) * (w / 8)]
# final_correlation = F.softmax(torch.bmm(depth_correlation, f_correlation), 2)
#
# # [n, c / 2, h / 4, w / 4]
# y = torch.bmm(final_correlation, g).transpose(1, 2).contiguous().view(n, int(c / 2), int(h / 4), int(w / 4))
#
# return x + F.upsample(self.z(y), size=x.size()[2:], mode='bilinear', align_corners=True)