-
Notifications
You must be signed in to change notification settings - Fork 15
/
resnet.py
207 lines (161 loc) · 7.26 KB
/
resnet.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
import torch
import torch.nn as nn
import torch.nn.functional as F
import math
class FrozenBN(nn.Module):
def __init__(self, num_channels, momentum=0.1, eps=1e-5):
super(FrozenBN, self).__init__()
self.num_channels = num_channels
self.momentum = momentum
self.eps = eps
self.params_set = False
def set_params(self, scale, bias, running_mean, running_var):
self.register_buffer('scale', scale)
self.register_buffer('bias', bias)
self.register_buffer('running_mean', running_mean)
self.register_buffer('running_var', running_var)
self.params_set = True
def forward(self, x):
assert self.params_set, 'model.set_params(...) must be called before the forward pass'
return torch.batch_norm(x, self.scale, self.bias, self.running_mean, self.running_var, False, self.momentum, self.eps, torch.backends.cudnn.enabled)
def __repr__(self):
return 'FrozenBN(%d)'%self.num_channels
def freeze_bn(m, name):
for attr_str in dir(m):
target_attr = getattr(m, attr_str)
if type(target_attr) == torch.nn.BatchNorm3d:
frozen_bn = FrozenBN(target_attr.num_features, target_attr.momentum, target_attr.eps)
frozen_bn.set_params(target_attr.weight.data, target_attr.bias.data, target_attr.running_mean, target_attr.running_var)
setattr(m, attr_str, frozen_bn)
for n, ch in m.named_children():
freeze_bn(ch, n)
#-----------------------------------------------------------------------------------------------#
class Bottleneck(nn.Module):
expansion = 4
def __init__(self, inplanes, planes, stride, downsample, temp_conv, temp_stride, use_nl=False):
super(Bottleneck, self).__init__()
self.conv1 = nn.Conv3d(inplanes, planes, kernel_size=(1 + temp_conv * 2, 1, 1), stride=(temp_stride, 1, 1), padding=(temp_conv, 0, 0), bias=False)
self.bn1 = nn.BatchNorm3d(planes)
self.conv2 = nn.Conv3d(planes, planes, kernel_size=(1, 3, 3), stride=(1, stride, stride), padding=(0, 1, 1), bias=False)
self.bn2 = nn.BatchNorm3d(planes)
self.conv3 = nn.Conv3d(planes, planes * 4, kernel_size=1, stride=1, padding=0, bias=False)
self.bn3 = nn.BatchNorm3d(planes * 4)
self.relu = nn.ReLU(inplace=True)
self.downsample = downsample
self.stride = stride
outplanes = planes * 4
self.nl = NonLocalBlock(outplanes, outplanes, outplanes//2) if use_nl else None
def forward(self, x):
residual = x
out = self.conv1(x)
out = self.bn1(out)
out = self.relu(out)
out = self.conv2(out)
out = self.bn2(out)
out = self.relu(out)
out = self.conv3(out)
out = self.bn3(out)
if self.downsample is not None:
residual = self.downsample(x)
out += residual
out = self.relu(out)
if self.nl is not None:
out = self.nl(out)
return out
class NonLocalBlock(nn.Module):
def __init__(self, dim_in, dim_out, dim_inner):
super(NonLocalBlock, self).__init__()
self.dim_in = dim_in
self.dim_inner = dim_inner
self.dim_out = dim_out
self.theta = nn.Conv3d(dim_in, dim_inner, kernel_size=(1,1,1), stride=(1,1,1), padding=(0,0,0))
self.maxpool = nn.MaxPool3d(kernel_size=(1,2,2), stride=(1,2,2), padding=(0,0,0))
self.phi = nn.Conv3d(dim_in, dim_inner, kernel_size=(1,1,1), stride=(1,1,1), padding=(0,0,0))
self.g = nn.Conv3d(dim_in, dim_inner, kernel_size=(1,1,1), stride=(1,1,1), padding=(0,0,0))
self.out = nn.Conv3d(dim_inner, dim_out, kernel_size=(1,1,1), stride=(1,1,1), padding=(0,0,0))
self.bn = nn.BatchNorm3d(dim_out)
def forward(self, x):
residual = x
batch_size = x.shape[0]
mp = self.maxpool(x)
theta = self.theta(x)
phi = self.phi(mp)
g = self.g(mp)
theta_shape_5d = theta.shape
theta, phi, g = theta.view(batch_size, self.dim_inner, -1), phi.view(batch_size, self.dim_inner, -1), g.view(batch_size, self.dim_inner, -1)
theta_phi = torch.bmm(theta.transpose(1, 2), phi) # (8, 1024, 784) * (8, 1024, 784) => (8, 784, 784)
theta_phi_sc = theta_phi * (self.dim_inner**-.5)
p = F.softmax(theta_phi_sc, dim=-1)
t = torch.bmm(g, p.transpose(1, 2))
t = t.view(theta_shape_5d)
out = self.out(t)
out = self.bn(out)
out = out + residual
return out
#-----------------------------------------------------------------------------------------------#
class I3Res50(nn.Module):
def __init__(self, block=Bottleneck, layers=[3, 4, 6, 3], num_classes=400, use_nl=False):
self.inplanes = 64
super(I3Res50, self).__init__()
self.conv1 = nn.Conv3d(3, 64, kernel_size=(5, 7, 7), stride=(2, 2, 2), padding=(2, 3, 3), bias=False)
self.bn1 = nn.BatchNorm3d(64)
self.relu = nn.ReLU(inplace=True)
self.maxpool1 = nn.MaxPool3d(kernel_size=(2, 3, 3), stride=(2, 2, 2), padding=(0, 0, 0))
self.maxpool2 = nn.MaxPool3d(kernel_size=(2, 1, 1), stride=(2, 1, 1), padding=(0, 0, 0))
nonlocal_mod = 2 if use_nl else 1000
self.layer1 = self._make_layer(block, 64, layers[0], stride=1, temp_conv=[1, 1, 1], temp_stride=[1, 1, 1])
self.layer2 = self._make_layer(block, 128, layers[1], stride=2, temp_conv=[1, 0, 1, 0], temp_stride=[1, 1, 1, 1], nonlocal_mod=nonlocal_mod)
self.layer3 = self._make_layer(block, 256, layers[2], stride=2, temp_conv=[1, 0, 1, 0, 1, 0], temp_stride=[1, 1, 1, 1, 1, 1], nonlocal_mod=nonlocal_mod)
self.layer4 = self._make_layer(block, 512, layers[3], stride=2, temp_conv=[0, 1, 0], temp_stride=[1, 1, 1])
self.avgpool = nn.AdaptiveAvgPool3d((1, 1, 1))
self.fc = nn.Linear(512 * block.expansion, num_classes)
self.drop = nn.Dropout(0.5)
for m in self.modules():
if isinstance(m, nn.Conv3d):
m.weight = nn.init.kaiming_normal_(m.weight, mode='fan_out')
elif isinstance(m, nn.BatchNorm3d):
m.weight.data.fill_(1)
m.bias.data.zero_()
def _make_layer(self, block, planes, blocks, stride, temp_conv, temp_stride, nonlocal_mod=1000):
downsample = None
if stride != 1 or self.inplanes != planes * block.expansion or temp_stride[0]!=1:
downsample = nn.Sequential(
nn.Conv3d(self.inplanes, planes * block.expansion, kernel_size=(1, 1, 1), stride=(temp_stride[0], stride, stride), padding=(0, 0, 0), bias=False),
nn.BatchNorm3d(planes * block.expansion)
)
layers = []
layers.append(block(self.inplanes, planes, stride, downsample, temp_conv[0], temp_stride[0], False))
self.inplanes = planes * block.expansion
for i in range(1, blocks):
layers.append(block(self.inplanes, planes, 1, None, temp_conv[i], temp_stride[i], i%nonlocal_mod==nonlocal_mod-1))
return nn.Sequential(*layers)
def forward_single(self, x):
x = self.conv1(x)
x = self.bn1(x)
x = self.relu(x)
x = self.maxpool1(x)
x = self.layer1(x)
x = self.maxpool2(x)
x = self.layer2(x)
x = self.layer3(x)
x = self.layer4(x)
x = self.avgpool(x)
return x
def forward(self, batch):
if batch['frames'].dim() == 5:
feat = self.forward_single(batch['frames'])
return feat
#-----------------------------------------------------------------------------------------------#
def i3_res50(num_classes, pretrainedpath):
net = I3Res50(num_classes=num_classes, use_nl=False)
state_dict = torch.load(pretrainedpath)
net.load_state_dict(state_dict)
print("Received Pretrained model..")
# freeze_bn(net, "net") # Only needed for finetuning. For validation, .eval() works.
return net
def i3_res50_nl(num_classes, pretrainedpath):
net = I3Res50(num_classes=num_classes, use_nl=True)
state_dict = torch.load(pretrainedpath)
net.load_state_dict(state_dict)
# freeze_bn(net, "net") # Only needed for finetuning. For validation, .eval() works.
return net