RepVGG

# !/usr/bin/env python
# -- coding: utf-8 --

import numpy as np
import torch
import torch.nn as nn


def Conv1x1BN(in_channels,out_channels, stride=1, groups=1, bias=False):
    return nn.Sequential(
            nn.Conv2d(in_channels=in_channels, out_channels=out_channels, kernel_size=1, stride=stride, padding=0, groups=groups, bias=bias),
            nn.BatchNorm2d(out_channels)
        )

def Conv3x3BN(in_channels,out_channels, stride=1, groups=1, bias=False):
    return nn.Sequential(
        nn.Conv2d(in_channels=in_channels,out_channels=out_channels,kernel_size=3,stride=stride,padding=1, groups=groups, bias=bias),
        nn.BatchNorm2d(out_channels)
    )


class RepVGGBlock(nn.Module):
    def __init__(self, in_channels, out_channels, stride=1, groups=1, deploy=False):
        super(RepVGGBlock, self).__init__()
        self.deploy = deploy

        if self.deploy:
            self.conv = nn.Conv2d(in_channels=in_channels, out_channels=out_channels,
                                         kernel_size=3, stride=stride, padding=1, dilation=1, groups=groups, bias=True)
        else:
            self.conv1 = Conv3x3BN(in_channels, out_channels, stride=stride, groups=groups, bias=False)
            self.conv2 = Conv1x1BN(in_channels, out_channels, stride=stride, groups=groups, bias=False)

            self.identity = nn.BatchNorm2d(in_channels) if out_channels == in_channels and stride == 1 else None

        self.act = nn.ReLU(inplace=True)

    def forward(self, x):
        if self.deploy:
            return self.act(self.conv(x))
        if self.identity is None:
            return self.act(self.conv1(x) + self.conv2(x))
        else:
            return self.act(self.conv1(x) + self.conv2(x) + self.identity(x))


class RepVGG(nn.Module):
    def __init__(self, block_nums, width_multiplier=None, group=1, num_classes=1000, deploy=False):
        super(RepVGG, self).__init__()
        self.deploy = deploy
        self.group = group
        assert len(width_multiplier) == 4

        self.stage0 = RepVGGBlock(in_channels=3,out_channels=min(64, int(64 * width_multiplier[0])), stride=2, deploy=self.deploy)
        self.cur_layer_idx = 1
        self.stage1 = self._make_layers(in_channels=min(64, int(64 * width_multiplier[0])), out_channels= int(64 * width_multiplier[0]), stride=2, block_num=block_nums[0])
        self.stage2 = self._make_layers(in_channels=int(64 * width_multiplier[0]), out_channels=int(128 * width_multiplier[1]), stride=2, block_num=block_nums[1])
        self.stage3 = self._make_layers(in_channels=int(128 * width_multiplier[1]), out_channels=int(256 * width_multiplier[2]), stride=2, block_num=block_nums[2])
        self.stage4 = self._make_layers(in_channels=int(256 * width_multiplier[2]), out_channels=int(512 * width_multiplier[3]), stride=2, block_num=block_nums[3])
        self.avg_pool = nn.AdaptiveAvgPool2d(output_size=1)
        self.linear = nn.Linear(int(512 * width_multiplier[3]), num_classes)

        self._init_params()

    def _make_layers(self, in_channels, out_channels, stride, block_num):
        layers = []
        layers.append(RepVGGBlock(in_channels,out_channels, stride=stride, groups=self.group if self.cur_layer_idx%2==0 else 1, deploy=self.deploy))
        self.cur_layer_idx += 1
        for i in range(block_num):
            layers.append(RepVGGBlock(out_channels,out_channels, stride=1, groups=self.group if self.cur_layer_idx%2==0 else 1, deploy=self.deploy))
            self.cur_layer_idx += 1
        return nn.Sequential(*layers)

    def _init_params(self):
        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
            elif isinstance(m, nn.BatchNorm2d):
                nn.init.constant_(m.weight, 1)
                nn.init.constant_(m.bias, 0)

    def forward(self, x):
        x = self.stage0(x)
        x = self.stage1(x)
        x = self.stage2(x)
        x = self.stage3(x)
        x = self.stage4(x)
        x = self.avg_pool(x)
        x = x.view(x.size(0), -1)
        out = self.linear(x)
        return out

def RepVGG_A0(deploy=False):
    return RepVGG(block_nums=[2, 4, 14, 1], num_classes=1000,
                  width_multiplier=[0.75, 0.75, 0.75, 2.5], group=1, deploy=deploy)

def RepVGG_A1(deploy=False):
    return RepVGG(block_nums=[2, 4, 14, 1], num_classes=1000,
                  width_multiplier=[1, 1, 1, 2.5], group=1, deploy=deploy)

def RepVGG_A2(deploy=False):
    return RepVGG(block_nums=[2, 4, 14, 1], num_classes=1000,
                  width_multiplier=[1.5, 1.5, 1.5, 2.75], group=1, deploy=deploy)

def RepVGG_B0(deploy=False):
    return RepVGG(block_nums=[4, 6, 16, 1], num_classes=1000,
                  width_multiplier=[1, 1, 1, 2.5], group=1, deploy=deploy)

def RepVGG_B1(deploy=False):
    return RepVGG(block_nums=[4, 6, 16, 1], num_classes=1000,
                  width_multiplier=[2, 2, 2, 4], group=1, deploy=deploy)

def RepVGG_B1g2(deploy=False):
    return RepVGG(block_nums=[4, 6, 16, 1], num_classes=1000,
                  width_multiplier=[2, 2, 2, 4], group=2, deploy=deploy)

def RepVGG_B1g4(deploy=False):
    return RepVGG(block_nums=[4, 6, 16, 1], num_classes=1000,
                  width_multiplier=[2, 2, 2, 4], group=4, deploy=deploy)


def RepVGG_B2(deploy=False):
    return RepVGG(block_nums=[4, 6, 16, 1], num_classes=1000,
                  width_multiplier=[2.5, 2.5, 2.5, 5], group=1, deploy=deploy)

def RepVGG_B2g2(deploy=False):
    return RepVGG(block_nums=[4, 6, 16, 1], num_classes=1000,
                  width_multiplier=[2.5, 2.5, 2.5, 5], group=2, deploy=deploy)

def RepVGG_B2g4(deploy=False):
    return RepVGG(block_nums=[4, 6, 16, 1], num_classes=1000,
                  width_multiplier=[2.5, 2.5, 2.5, 5], group=4, deploy=deploy)


def RepVGG_B3(deploy=False):
    return RepVGG(block_nums=[1, 4, 6, 16, 1], num_classes=1000,
                  width_multiplier=[3, 3, 3, 5], group=1, deploy=deploy)

def RepVGG_B3g2(deploy=False):
    return RepVGG(block_nums=[1, 4, 6, 16, 1], num_classes=1000,
                  width_multiplier=[3, 3, 3, 5], group=2, deploy=deploy)

def RepVGG_B3g4(deploy=False):
    return RepVGG(block_nums=[1, 4, 6, 16, 1], num_classes=1000,
                  width_multiplier=[3, 3, 3, 5], group=4, deploy=deploy)


if __name__ == '__main__':
    model = RepVGG_A1()
    print(model)

    input = torch.randn(1,3,224,224)
    out = model(input)
    print(out.shape)

  

 

  
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
 

文章来源: wanghao.blog.csdn.net，作者：AI浩，版权归原作者所有，如需转载，请联系作者。

原文链接：wanghao.blog.csdn.net/article/details/121573980