-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathseg_model.py
159 lines (130 loc) · 4.96 KB
/
seg_model.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
import loss
import torch
import torch.nn as nn
import torch.nn.functional as F
class CNNEncoder(nn.Module):
def __init__(
self,
input_shape,
out_channels_per_layer,
kernel_sizes_per_layer,
maxpool_per_layer,
):
super().__init__()
self.out_channels_per_layer = out_channels_per_layer
self.kernel_sizes_per_layer = kernel_sizes_per_layer
self.maxpool_per_layer = maxpool_per_layer
in_features, h, w = input_shape
layers = nn.ModuleList()
for i in range(len(out_channels_per_layer)):
block = [
nn.Conv2d(
in_channels=in_features,
out_channels=out_channels_per_layer[i],
kernel_size=kernel_sizes_per_layer[i],
padding='same',
),
nn.ReLU(),
]
if maxpool_per_layer[i]:
block.append(nn.MaxPool2d(kernel_size=maxpool_per_layer[i]))
h //= maxpool_per_layer[i]
w //= maxpool_per_layer[i]
layers.append(nn.Sequential(*block))
in_features = out_channels_per_layer[i]
self.input_shape = input_shape
self.output_shape = (out_channels_per_layer[-1], h, w)
self.layers = nn.Sequential(*layers)
def forward(self, x):
x = self.layers(x)
return x
class CNNDecoder(nn.Module):
def __init__(
self,
input_shape,
out_channels_per_layer,
kernel_sizes_per_layer,
upsample_per_layer,
):
super().__init__()
self.out_channels_per_layer = out_channels_per_layer
self.kernel_sizes_per_layer = kernel_sizes_per_layer
self.upsample_per_layer = upsample_per_layer
in_features, h, w = input_shape
layers = nn.ModuleList()
for i in range(len(out_channels_per_layer)):
block = [
nn.Conv2d(
in_channels=in_features,
out_channels=out_channels_per_layer[i],
kernel_size=kernel_sizes_per_layer[i],
padding='same',
),
nn.ReLU(),
]
if upsample_per_layer[i]:
block.append(nn.Upsample(scale_factor=upsample_per_layer[i], mode='bilinear'))
h *= upsample_per_layer[i]
w *= upsample_per_layer[i]
layers.append(nn.Sequential(*block))
in_features = out_channels_per_layer[i]
self.input_shape = input_shape
self.output_shape = (in_features, h, w)
self.layers = nn.Sequential(*layers)
def forward(self, x):
x = self.layers(x)
return x
class GenWeakSegNet(nn.Module):
def __init__(self, classifier, num_classes=2, hparams=None):
super().__init__()
self.input_shape = (3, 224, 224)
self.num_classes = num_classes
self.hparams = hparams
out_channels = [64, 64, 128, 128, 256, 256, 512, 512]
kernel_sizes = [3, 3, 3, 3, 3, 3, 3, 3]
maxpool = [0, 2, 0, 2, 0, 2, 0, 2]
self.encoder = CNNEncoder(
input_shape=self.input_shape,
out_channels_per_layer=out_channels,
kernel_sizes_per_layer=kernel_sizes,
maxpool_per_layer=maxpool,
)
out_channels = out_channels[::-1]
kernel_sizes = kernel_sizes[::-1]
upsample = maxpool
self.decoder = CNNDecoder(
input_shape=self.encoder.output_shape,
out_channels_per_layer=out_channels,
kernel_sizes_per_layer=kernel_sizes,
upsample_per_layer=upsample,
)
self.op_cls_img = nn.Conv2d(out_channels[-1], 3*num_classes, kernel_size=3, padding='same')
self.op_cls_mask = nn.Conv2d(out_channels[-1], num_classes, kernel_size=3, padding='same')
self.recon_loss_fn = loss.ReconLoss(L=1)
self.mask_reg_loss_fn = loss.MaskRegLoss(num_classes)
for params in classifier.parameters():
params.requires_grad = False
self.cls_guide_loss_fn = loss.ClsGuideLoss(classifier)
def forward(self, x):
x = self.encoder(x)
x = self.decoder(x)
y_img = self.op_cls_img(x).view(x.shape[0], self.num_classes, *self.input_shape)
y_img = F.sigmoid(y_img)
y_mask = self.op_cls_mask(x)
return (y_img, y_mask)
def loss_fn(self, x, label, y_img, y_mask):
recon = self.recon_loss_fn(x, y_img, y_mask)
mask_reg = self.mask_reg_loss_fn(label, y_mask)
cls_guide = self.cls_guide_loss_fn(label, y_img, y_mask)
loss = (
self.hparams['recon']*recon +
self.hparams['mask_reg']*mask_reg +
self.hparams['cls_guide']*cls_guide
)
loss_dict = {
'loss': loss,
'recon': recon,
'mask_reg': mask_reg,
'cls_guide': cls_guide,
}
return (loss, loss_dict)