-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathmodel.py
199 lines (166 loc) · 7.15 KB
/
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
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
import torch
import torch.nn as nn
import torchvision
import numpy as np
import torch.nn.functional as F
from lib.nn import SynchronizedBatchNorm2d
import resnet
from graphModule import GCU
class SegmentationModuleBase(nn.Module):
def __init__(self):
super(SegmentationModuleBase, self).__init__()
def pixel_acc(self, pred, label):
_, preds = torch.max(pred, dim=1)
valid = (label >= 0).long()
acc_sum = torch.sum(valid * (preds == label).long())
pixel_sum = torch.sum(valid)
acc = acc_sum.float() / (pixel_sum.float() + 1e-10)
return acc
class SegmentationModule(SegmentationModuleBase):
def __init__(self, net_enc, gcu,crit, tr): #graphconv is a list
super(SegmentationModule, self).__init__()
self.encoder = net_enc
self.crit = crit
self.conv1 = torch.nn.Conv2d(2304,150 , kernel_size=3, stride=1, padding=1).cuda(1)
self.tr= tr
self.gcu = gcu
self.gcu.cuda(1)
self.encoder.cuda(0)
def forward(self, feed_dict):
# training
#feed_dict = feed_dict[0]
enc_out1 = self.encoder(feed_dict['img_data'].cuda(0))
# print(enc_out1.shape)
enc_out2 = enc_out1.cuda(1)
enc_out1 = self.gcu(enc_out2)
up = nn.Upsample(scale_factor=8)
pred = up(enc_out1)
pred = self.conv1(pred);pred = nn.functional.log_softmax(pred, dim=1)
if self.tr:
loss = self.crit(pred, feed_dict['seg_label'].cuda(1)) #NLLL Loss
acc = self.pixel_acc(pred, feed_dict['seg_label'].cuda(1))
return loss,acc
# # inference
else:
return pred
def conv3x3(in_planes, out_planes, stride=1, has_bias=False):
"3x3 convolution with padding"
return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride,
padding=1, bias=has_bias)
def conv3x3_bn_relu(in_planes, out_planes, stride=1):
return nn.Sequential(
conv3x3(in_planes, out_planes, stride),
SynchronizedBatchNorm2d(out_planes),
nn.ReLU(inplace=True),
)
class ModelBuilder():
# custom weights initialization
def weights_init(self, m):
classname = m.__class__.__name__
if classname.find('Conv') != -1:
nn.init.kaiming_normal_(m.weight.data)
elif classname.find('BatchNorm') != -1:
m.weight.data.fill_(1.)
m.bias.data.fill_(1e-4)
#elif classname.find('Linear') != -1:
# m.weight.data.normal_(0.0, 0.0001)
def build_encoder(self, arch='resnet50dilated', fc_dim=512, weights=''):
pretrained = True if len(weights) == 0 else False
arch = arch.lower()
if arch == 'mobilenetv2dilated':
orig_mobilenet = mobilenet.__dict__['mobilenetv2'](pretrained=pretrained)
net_encoder = MobileNetV2Dilated(orig_mobilenet, dilate_scale=8)
elif arch == 'resnet18':
orig_resnet = resnet.__dict__['resnet18'](pretrained=pretrained)
net_encoder = Resnet(orig_resnet)
elif arch == 'resnet18dilated':
orig_resnet = resnet.__dict__['resnet18'](pretrained=pretrained)
net_encoder = ResnetDilated(orig_resnet, dilate_scale=8)
elif arch == 'resnet34':
raise NotImplementedError
orig_resnet = resnet.__dict__['resnet34'](pretrained=pretrained)
net_encoder = Resnet(orig_resnet)
elif arch == 'resnet34dilated':
raise NotImplementedError
orig_resnet = resnet.__dict__['resnet34'](pretrained=pretrained)
net_encoder = ResnetDilated(orig_resnet, dilate_scale=8)
elif arch == 'resnet50':
orig_resnet = resnet.__dict__['resnet50'](pretrained=pretrained)
net_encoder = Resnet(orig_resnet)
elif arch == 'resnet50dilated':
orig_resnet = resnet.__dict__['resnet50'](pretrained=pretrained)
net_encoder = ResnetDilated(orig_resnet, dilate_scale=8)
elif arch == 'resnet101':
orig_resnet = resnet.__dict__['resnet101'](pretrained=pretrained)
net_encoder = Resnet(orig_resnet)
elif arch == 'resnet101dilated':
orig_resnet = resnet.__dict__['resnet101'](pretrained=pretrained)
net_encoder = ResnetDilated(orig_resnet, dilate_scale=8)
elif arch == 'resnext101':
orig_resnext = resnext.__dict__['resnext101'](pretrained=pretrained)
net_encoder = Resnet(orig_resnext) # we can still use class Resnet
else:
raise Exception('Architecture undefined!')
# net_encoder.apply(self.weights_init)
if len(weights) > 0:
print('Loading weights for net_encoder')
net_encoder.load_state_dict(
torch.load(weights, map_location=lambda storage, loc: storage), strict=False)
return net_encoder
class ResnetDilated(nn.Module):
def __init__(self, orig_resnet, dilate_scale=8):
super(ResnetDilated, self).__init__()
from functools import partial
if dilate_scale == 8:
orig_resnet.layer3.apply(
partial(self._nostride_dilate, dilate=2))
orig_resnet.layer4.apply(
partial(self._nostride_dilate, dilate=4))
elif dilate_scale == 16:
orig_resnet.layer4.apply(
partial(self._nostride_dilate, dilate=2))
# take pretrained resnet, except AvgPool and FC
self.conv1 = orig_resnet.conv1
self.bn1 = orig_resnet.bn1
self.relu1 = orig_resnet.relu1
self.conv2 = orig_resnet.conv2
self.bn2 = orig_resnet.bn2
self.relu2 = orig_resnet.relu2
self.conv3 = orig_resnet.conv3
self.bn3 = orig_resnet.bn3
self.relu3 = orig_resnet.relu3
self.maxpool = orig_resnet.maxpool
self.layer1 = orig_resnet.layer1
self.layer2 = orig_resnet.layer2
self.layer3 = orig_resnet.layer3
self.layer4 = orig_resnet.layer4
def _nostride_dilate(self, m, dilate):
classname = m.__class__.__name__
if classname.find('Conv') != -1:
# the convolution with stride
if m.stride == (2, 2):
m.stride = (1, 1)
if m.kernel_size == (3, 3):
m.dilation = (dilate//2, dilate//2)
m.padding = (dilate//2, dilate//2)
# other convoluions
else:
if m.kernel_size == (3, 3):
m.dilation = (dilate, dilate)
m.padding = (dilate, dilate)
def forward(self, x, return_feature_maps=False):
conv_out = []
x = self.relu1(self.bn1(self.conv1(x)))
x = self.relu2(self.bn2(self.conv2(x)))
x = self.relu3(self.bn3(self.conv3(x)))
x = self.maxpool(x)
x = self.layer1(x); conv_out.append(x);
x = self.layer2(x); conv_out.append(x);
x = self.layer3(x); conv_out.append(x);
x = self.layer4(x); conv_out.append(x);
if return_feature_maps:
return conv_out
return x
#binliena interpolation to resize image
#conv to give an output of 3C image
#loss crosetropy + regularization