-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathcoordconv.py
187 lines (148 loc) · 7.47 KB
/
coordconv.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
"""
This code is the implementation of coordconv in pytorch.
https://github.com/walsvid/CoordConv/blob/master/coordconv.py
"""
import torch
import torch.nn as nn
import torch.nn.modules.conv as conv
class AddCoords(nn.Module):
def __init__(self, rank, with_r=False, use_cuda=True):
super(AddCoords, self).__init__()
self.rank = rank
self.with_r = with_r
self.use_cuda = use_cuda
def forward(self, input_tensor):
"""
:param input_tensor: shape (N, C_in, H, W)
:return:
"""
if self.rank == 1:
batch_size_shape, channel_in_shape, dim_x = input_tensor.shape
xx_range = torch.arange(dim_x, dtype=torch.int32)
xx_channel = xx_range[None, None, :]
xx_channel = xx_channel.float() / (dim_x - 1)
xx_channel = xx_channel * 2 - 1
xx_channel = xx_channel.repeat(batch_size_shape, 1, 1)
if torch.cuda.is_available and self.use_cuda:
input_tensor = input_tensor.cuda()
xx_channel = xx_channel.cuda()
out = torch.cat([input_tensor, xx_channel], dim=1)
if self.with_r:
rr = torch.sqrt(torch.pow(xx_channel - 0.5, 2))
out = torch.cat([out, rr], dim=1)
elif self.rank == 2:
batch_size_shape, channel_in_shape, dim_y, dim_x = input_tensor.shape
xx_ones = torch.ones([1, 1, 1, dim_x], dtype=torch.int32)
yy_ones = torch.ones([1, 1, 1, dim_y], dtype=torch.int32)
xx_range = torch.arange(dim_y, dtype=torch.int32)
yy_range = torch.arange(dim_x, dtype=torch.int32)
xx_range = xx_range[None, None, :, None]
yy_range = yy_range[None, None, :, None]
xx_channel = torch.matmul(xx_range, xx_ones)
yy_channel = torch.matmul(yy_range, yy_ones)
# transpose y
yy_channel = yy_channel.permute(0, 1, 3, 2)
xx_channel = xx_channel.float() / (dim_y - 1)
yy_channel = yy_channel.float() / (dim_x - 1)
xx_channel = xx_channel * 2 - 1
yy_channel = yy_channel * 2 - 1
xx_channel = xx_channel.repeat(batch_size_shape, 1, 1, 1)
yy_channel = yy_channel.repeat(batch_size_shape, 1, 1, 1)
if torch.cuda.is_available and self.use_cuda:
input_tensor = input_tensor.cuda()
xx_channel = xx_channel.cuda()
yy_channel = yy_channel.cuda()
out = torch.cat([input_tensor, xx_channel, yy_channel], dim=1)
if self.with_r:
rr = torch.sqrt(torch.pow(xx_channel - 0.5, 2) + torch.pow(yy_channel - 0.5, 2))
out = torch.cat([out, rr], dim=1)
elif self.rank == 3:
batch_size_shape, channel_in_shape, dim_z, dim_y, dim_x = input_tensor.shape
xx_ones = torch.ones([1, 1, 1, 1, dim_x], dtype=torch.int32)
yy_ones = torch.ones([1, 1, 1, 1, dim_y], dtype=torch.int32)
zz_ones = torch.ones([1, 1, 1, 1, dim_z], dtype=torch.int32)
xy_range = torch.arange(dim_y, dtype=torch.int32)
xy_range = xy_range[None, None, None, :, None]
yz_range = torch.arange(dim_z, dtype=torch.int32)
yz_range = yz_range[None, None, None, :, None]
zx_range = torch.arange(dim_x, dtype=torch.int32)
zx_range = zx_range[None, None, None, :, None]
xy_channel = torch.matmul(xy_range, xx_ones)
xx_channel = torch.cat([xy_channel + i for i in range(dim_z)], dim=2)
xx_channel = xx_channel.repeat(batch_size_shape, 1, 1, 1, 1)
yz_channel = torch.matmul(yz_range, yy_ones)
yz_channel = yz_channel.permute(0, 1, 3, 4, 2)
yy_channel = torch.cat([yz_channel + i for i in range(dim_x)], dim=4)
yy_channel = yy_channel.repeat(batch_size_shape, 1, 1, 1, 1)
zx_channel = torch.matmul(zx_range, zz_ones)
zx_channel = zx_channel.permute(0, 1, 4, 2, 3)
zz_channel = torch.cat([zx_channel + i for i in range(dim_y)], dim=3)
zz_channel = zz_channel.repeat(batch_size_shape, 1, 1, 1, 1)
if torch.cuda.is_available and self.use_cuda:
input_tensor = input_tensor.cuda()
xx_channel = xx_channel.cuda()
yy_channel = yy_channel.cuda()
zz_channel = zz_channel.cuda()
out = torch.cat([input_tensor, xx_channel, yy_channel, zz_channel], dim=1)
if self.with_r:
rr = torch.sqrt(torch.pow(xx_channel - 0.5, 2) +
torch.pow(yy_channel - 0.5, 2) +
torch.pow(zz_channel - 0.5, 2))
out = torch.cat([out, rr], dim=1)
else:
raise NotImplementedError
return out
class CoordConv1d(conv.Conv1d):
def __init__(self, in_channels, out_channels, kernel_size, stride=1,
padding=0, dilation=1, groups=1, bias=True, with_r=False, use_cuda=True):
super(CoordConv1d, self).__init__(in_channels, out_channels, kernel_size,
stride, padding, dilation, groups, bias)
self.rank = 1
self.addcoords = AddCoords(self.rank, with_r, use_cuda=use_cuda)
self.conv = nn.Conv1d(in_channels + self.rank + int(with_r), out_channels,
kernel_size, stride, padding, dilation, groups, bias)
def forward(self, input_tensor):
"""
input_tensor_shape: (N, C_in,H,W)
output_tensor_shape: N,C_out,H_out,W_out）
:return: CoordConv2d Result
"""
out = self.addcoords(input_tensor)
out = self.conv(out)
return out
class CoordConv2d(conv.Conv2d):
def __init__(self, in_channels, out_channels, kernel_size, stride=1,
padding=0, dilation=1, groups=1, bias=True, with_r=False, use_cuda=True):
super(CoordConv2d, self).__init__(in_channels, out_channels, kernel_size,
stride, padding, dilation, groups, bias)
self.rank = 2
self.addcoords = AddCoords(self.rank, with_r, use_cuda=use_cuda)
self.conv = nn.Conv2d(in_channels + self.rank + int(with_r), out_channels,
kernel_size, stride, padding, dilation, groups, bias)
def forward(self, input_tensor):
"""
input_tensor_shape: (N, C_in,H,W)
output_tensor_shape: N,C_out,H_out,W_out）
:return: CoordConv2d Result
"""
out = self.addcoords(input_tensor)
out = self.conv(out)
return out
class CoordConv3d(conv.Conv3d):
def __init__(self, in_channels, out_channels, kernel_size, stride=1,
padding=0, dilation=1, groups=1, bias=True, with_r=False, use_cuda=True):
super(CoordConv3d, self).__init__(in_channels, out_channels, kernel_size,
stride, padding, dilation, groups, bias)
self.rank = 3
self.addcoords = AddCoords(self.rank, with_r, use_cuda=use_cuda)
self.conv = nn.Conv3d(in_channels + self.rank + int(with_r), out_channels,
kernel_size, stride, padding, dilation, groups, bias)
def forward(self, input_tensor):
"""
input_tensor_shape: (N, C_in,H,W)
output_tensor_shape: N,C_out,H_out,W_out）
:return: CoordConv2d Result
"""
out = self.addcoords(input_tensor)
out = self.conv(out)
return out