-
Notifications
You must be signed in to change notification settings - Fork 43
/
Copy pathunet.py
237 lines (218 loc) · 10.5 KB
/
unet.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
# Copyright (c) OpenMMLab. All rights reserved.
import warnings
from typing import Dict
import torch
from torch.nn import Upsample as Upsample
import torch.nn as nn
import torch.utils.checkpoint as cp
from torch.nn.modules.batchnorm import _BatchNorm
from torch.nn.modules.instancenorm import _InstanceNorm
SyncBatchNorm_ = nn.SyncBatchNorm
from terratorch.models.backbones.utils import UpConvBlock, BasicConvBlock
from terratorch.models.decoders.utils import ConvModule
from terratorch.registry import TERRATORCH_BACKBONE_REGISTRY
@TERRATORCH_BACKBONE_REGISTRY.register
class UNet(nn.Module):
"""UNet backbone.
This backbone is the implementation of `U-Net: Convolutional Networks
for Biomedical Image Segmentation <https://arxiv.org/abs/1505.04597>`_.
Args:
in_channels (int): Number of input image channels. Default" 3.
out_channels (int): Number of base channels of each stage.
The output channels of the first stage. Default: 64.
num_stages (int): Number of stages in encoder, normally 5. Default: 5.
strides (Sequence[int 1 | 2]): Strides of each stage in encoder.
len(strides) is equal to num_stages. Normally the stride of the
first stage in encoder is 1. If strides[i]=2, it uses stride
convolution to downsample in the correspondence encoder stage.
Default: (1, 1, 1, 1, 1).
enc_num_convs (Sequence[int]): Number of convolutional layers in the
convolution block of the correspondence encoder stage.
Default: (2, 2, 2, 2, 2).
dec_num_convs (Sequence[int]): Number of convolutional layers in the
convolution block of the correspondence decoder stage.
Default: (2, 2, 2, 2).
downsamples (Sequence[int]): Whether use MaxPool to downsample the
feature map after the first stage of encoder
(stages: [1, num_stages)). If the correspondence encoder stage use
stride convolution (strides[i]=2), it will never use MaxPool to
downsample, even downsamples[i-1]=True.
Default: (True, True, True, True).
enc_dilations (Sequence[int]): Dilation rate of each stage in encoder.
Default: (1, 1, 1, 1, 1).
dec_dilations (Sequence[int]): Dilation rate of each stage in decoder.
Default: (1, 1, 1, 1).
with_cp (bool): Use checkpoint or not. Using checkpoint will save some
memory while slowing down the training speed. Default: False.
conv_cfg (dict | None): Config dict for convolution layer.
Default: None.
norm_cfg (dict | None): Config dict for normalization layer.
Default: dict(type='BN').
act_cfg (dict | None): Config dict for activation layer in ConvModule.
Default: dict(type='ReLU').
upsample_cfg (dict): The upsample config of the upsample module in
decoder. Default: dict(type='InterpConv').
norm_eval (bool): Whether to set norm layers to eval mode, namely,
freeze running stats (mean and var). Note: Effect on Batch Norm
and its variants only. Default: False.
dcn (bool): Use deformable convolution in convolutional layer or not.
Default: None.
plugins (dict): plugins for convolutional layers. Default: None.
pretrained (str, optional): model pretrained path. Default: None
init_cfg (dict or list[dict], optional): Initialization config dict.
Default: None
Notice:
The input image size should be divisible by the whole downsample rate
of the encoder. More detail of the whole downsample rate can be found
in UNet._check_input_divisible.
"""
def __init__(self,
in_channels=3,
out_channels=64,
num_stages=5,
strides=(1, 1, 1, 1, 1),
enc_num_convs=(2, 2, 2, 2, 2),
dec_num_convs=(2, 2, 2, 2),
downsamples=(True, True, True, True),
enc_dilations=(1, 1, 1, 1, 1),
dec_dilations=(1, 1, 1, 1),
with_cp=False,
conv_cfg=None,
norm_cfg=dict(type='BN'),
act_cfg=dict(type='ReLU'),
upsample_cfg=None,
norm_eval=False,
dcn=None,
plugins=None,
pretrained=None,
init_cfg=None):
super(UNet, self).__init__()
self.pretrained = pretrained
assert not (init_cfg and pretrained), \
'init_cfg and pretrained cannot be setting at the same time'
if isinstance(pretrained, str):
warnings.warn('DeprecationWarning: pretrained is a deprecated, '
'please use "init_cfg" instead')
self.init_cfg = dict(type='Pretrained', checkpoint=pretrained)
elif pretrained is None:
if init_cfg is None:
self.init_cfg = [
dict(type='Kaiming', layer='Conv2d'),
dict(
type='Constant',
val=1,
layer=['_BatchNorm', 'GroupNorm'])
]
else:
raise TypeError('pretrained must be a str or None')
assert dcn is None, 'Not implemented yet.'
assert plugins is None, 'Not implemented yet.'
assert len(strides) == num_stages, \
'The length of strides should be equal to num_stages, '\
f'while the strides is {strides}, the length of '\
f'strides is {len(strides)}, and the num_stages is '\
f'{num_stages}.'
assert len(enc_num_convs) == num_stages, \
'The length of enc_num_convs should be equal to num_stages, '\
f'while the enc_num_convs is {enc_num_convs}, the length of '\
f'enc_num_convs is {len(enc_num_convs)}, and the num_stages is '\
f'{num_stages}.'
assert len(dec_num_convs) == (num_stages-1), \
'The length of dec_num_convs should be equal to (num_stages-1), '\
f'while the dec_num_convs is {dec_num_convs}, the length of '\
f'dec_num_convs is {len(dec_num_convs)}, and the num_stages is '\
f'{num_stages}.'
assert len(downsamples) == (num_stages-1), \
'The length of downsamples should be equal to (num_stages-1), '\
f'while the downsamples is {downsamples}, the length of '\
f'downsamples is {len(downsamples)}, and the num_stages is '\
f'{num_stages}.'
assert len(enc_dilations) == num_stages, \
'The length of enc_dilations should be equal to num_stages, '\
f'while the enc_dilations is {enc_dilations}, the length of '\
f'enc_dilations is {len(enc_dilations)}, and the num_stages is '\
f'{num_stages}.'
assert len(dec_dilations) == (num_stages-1), \
'The length of dec_dilations should be equal to (num_stages-1), '\
f'while the dec_dilations is {dec_dilations}, the length of '\
f'dec_dilations is {len(dec_dilations)}, and the num_stages is '\
f'{num_stages}.'
self.num_stages = num_stages
self.strides = strides
self.downsamples = downsamples
self.norm_eval = norm_eval
self.out_channels = num_stages*[out_channels]
self.encoder = nn.ModuleList()
self.decoder = nn.ModuleList()
for i in range(num_stages):
enc_conv_block = []
if i != 0:
if strides[i] == 1 and downsamples[i - 1]:
enc_conv_block.append(nn.MaxPool2d(kernel_size=2))
upsample = (strides[i] != 1 or downsamples[i - 1])
self.decoder.append(
UpConvBlock(
conv_block=BasicConvBlock,
in_channels=out_channels * 2**i,
skip_channels=out_channels * 2**(i - 1),
out_channels=out_channels * 2**(i - 1),
num_convs=dec_num_convs[i - 1],
stride=1,
dilation=dec_dilations[i - 1],
with_cp=with_cp,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=act_cfg,
upsample_cfg=upsample_cfg if upsample else None,
dcn=None,
plugins=None))
enc_conv_block.append(
BasicConvBlock(
in_channels=in_channels,
out_channels=out_channels * 2**i,
num_convs=enc_num_convs[i],
stride=strides[i],
dilation=enc_dilations[i],
with_cp=with_cp,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=act_cfg,
dcn=None,
plugins=None))
self.encoder.append((nn.Sequential(*enc_conv_block)))
in_channels = out_channels * 2**i
def forward(self, x):
# We can check just the first image, since the batch
# already was approved by the stackability test, which means
# all images has the same dimensions.
self._check_input_divisible(x[0])
enc_outs = []
for enc in self.encoder:
x = enc(x)
enc_outs.append(x)
dec_outs = [x]
for i in reversed(range(len(self.decoder))):
x = self.decoder[i](enc_outs[i], x)
dec_outs.append(x)
return dec_outs
def train(self, mode=True):
"""Convert the model into training mode while keep normalization layer
freezed."""
super(UNet, self).train(mode)
if mode and self.norm_eval:
for m in self.modules():
# trick: eval have effect on BatchNorm only
if isinstance(m, _BatchNorm):
m.eval()
def _check_input_divisible(self, x):
h, w = x.shape[-2:]
whole_downsample_rate = 1
for i in range(1, self.num_stages):
if self.strides[i] == 2 or self.downsamples[i - 1]:
whole_downsample_rate *= 2
assert (h % whole_downsample_rate == 0) \
and (w % whole_downsample_rate == 0),\
f'The input image size {(h, w)} should be divisible by the whole '\
f'downsample rate {whole_downsample_rate}, when num_stages is '\
f'{self.num_stages}, strides is {self.strides}, and downsamples '\
f'is {self.downsamples}.'