mobile-vsr/model.py at master · ncomly/mobile-vsr · GitHub

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
import math
import numpy as np

import torch
from torch import nn
from depthwise import *

def LipRes(alpha=2, reduction=1, num_classes=256):
    block = lambda in_planes, planes, stride: \
        LipResBlock(in_planes, planes, stride, reduction=reduction)
    return ResNet(block, [alpha, alpha, alpha, alpha], num_classes=num_classes)


class ResNet(nn.Module):
    def __init__(self, block, num_blocks, reduction=1, num_classes=256):
        super(ResNet, self).__init__()
        self.reduction = float(reduction) ** 0.5
        self.num_classes = num_classes
        self.in_planes = 64 #int(16 / self.reduction)

        self.layer1 = self._make_layer(block, self.in_planes, num_blocks[0], stride=1)
        self.layer2 = self._make_layer(block, 128, num_blocks[1], stride=2)
        self.layer3 = self._make_layer(block, 256, num_blocks[2], stride=2)
        self.layer4 = self._make_layer(block, 512, num_blocks[3], stride=2)
        self.fc = nn.Linear(512, num_classes)
        self.bnfc = nn.BatchNorm1d(num_classes)
        self.avgpool = nn.AvgPool2d(2)

        for m in self.modules():
            if isinstance(m, torch.nn.Conv2d):
                torch.nn.init.kaiming_uniform_(m.weight)
            elif isinstance(m, (nn.BatchNorm2d, nn.BatchNorm1d)):
                m.weight.data.fill_(1)
                m.bias.data.zero_()

    def _make_layer(self, block, planes, num_blocks, stride):
        strides = [stride] + [1]*(num_blocks-1)
        layers = []
        planes = int(planes)
        for stride in strides:
            layers.append(block(self.in_planes, planes, stride))
            self.in_planes = planes
        return nn.Sequential(*layers)

    def forward(self, x):

        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)
        x = self.layer4(x)

        x = self.avgpool(x)
        # 464 512 1 1
        x = x.view(x.size(0), -1)
        # 464 512
        x = self.fc(x)
        x = self.bnfc(x)

        return x


class LipNext(nn.Module):
    def __init__(self, inputDim=256, hiddenDim=512, nClasses=500, frameLen=29, alpha=2):
        super(LipNext, self).__init__()
        self.inputDim = inputDim
        self.hiddenDim = hiddenDim
        self.nClasses = nClasses
        self.frameLen = frameLen
        self.nLayers = 2
        self.alpha = alpha
        # frontend3D
        self.frontend3D = nn.Sequential( # potential optimizable
                nn.Conv3d(1, 64, kernel_size=(3, 3, 3), stride=(1, 2, 2), padding=(1, 1, 1), groups=1, bias=False),
                nn.BatchNorm3d(64),
                nn.ReLU(True),
                # group convolution
                nn.Conv3d(64, 64, kernel_size=(3, 3, 3), stride=(1, 2, 2), padding=(1, 1, 1), groups=64, bias=False),
                nn.Conv3d(64, 64, kernel_size=(3, 1, 1), stride=(1, 1, 1), padding=(1, 0, 0), groups=1, bias=False),
            )
        # resnet
        self.resnet34 = LipRes(self.alpha)
        # backend_conv
        self.backend_conv1 = nn.Sequential(
                nn.Conv1d(self.inputDim, 2*self.inputDim, 5, 2, 0, bias=False),
                nn.BatchNorm1d(2*self.inputDim),
                nn.ReLU(True),
                nn.MaxPool1d(2, 2),
                nn.Conv1d(2*self.inputDim, 4*self.inputDim, 5, 2, 0, bias=False),
                nn.BatchNorm1d(4*self.inputDim),
                nn.ReLU(True),
            )

        self.backend_conv2 = nn.Sequential(
                nn.Linear(4*self.inputDim, self.inputDim),
                nn.BatchNorm1d(self.inputDim),
                nn.ReLU(True),
                nn.Linear(self.inputDim, self.nClasses)
            )
        # initialize
        self._initialize_weights()

    def forward(self, x):

        x = self.frontend3D(x)
        # 16, 64, 29, 22,22
        x = x.transpose(1, 2)
        # 16, 29, 64 , 22, 22
        x = x.contiguous()

        x = x.view(-1, 64, x.size(3), x.size(4))
        # 464, 64, 22, 22
        x = self.resnet34(x)
        # 464 256
        x = x.view(-1, self.frameLen, self.inputDim)
        # 16 29 256
        x = x.transpose(1, 2)
        # 16 256 29
        x = self.backend_conv1(x)
        x = torch.mean(x,2)
        # x = x.view(-1, 4, 16, 16)
        x = self.backend_conv2(x)
        # x = x.view(-1, self.nClasses)
        return x

    def _initialize_weights(self):
        for m in self.modules():
            if isinstance(m, (nn.Conv3d, nn.Conv2d, nn.Conv1d)):
                torch.nn.init.kaiming_uniform_(m.weight)
                if m.bias is not None:
                    m.bias.data.zero_()

            elif isinstance(m, (nn.BatchNorm3d, nn.BatchNorm2d, nn.BatchNorm1d)):
                m.weight.data.fill_(1)
                m.bias.data.zero_()


def lipnext(inputDim=256, hiddenDim=512, nClasses=500, frameLen=29, alpha=2):
    model = LipNext(inputDim=inputDim, hiddenDim=hiddenDim, nClasses=nClasses, frameLen=frameLen, alpha=alpha)
    return model