深度学习（RepVGG重参数化）

 

 

 

 

 

 

 

 

 

RepVGG的核心思想是：在训练网络时使用复杂结构以获得更优的性能，在推理网络时则将其等价转换为一个简单结构以提高速度。

网络核心结构如下图：

如果是ResNet这样的网络，短连接中间有一个非线性层，则没有办法重参数化。

所以RepVGG中将3*3结构，1*1结构和直连结构并联在一起，从而在训练时能够学到更多参数。

在推理时则将三个分支参数合并为3*3结构，从而提高推理速度。

下面代码是按照自己的理解实现的重参数化Block，分为训练和部署两个分支，结果通过了allclose验证。

import torch
import torch.nn as nn

class RepVGGBlock(nn.Module):
    def __init__(self, channels, deploy):
        super(RepVGGBlock, self).__init__()

        self.deploy = deploy
        self.channels = channels

        self.conv3x3 = nn.Conv2d(channels, channels, kernel_size=3, stride=1, padding=1, bias=True)
        self.bn3x3 = nn.BatchNorm2d(channels)

        self.conv1x1 = nn.Conv2d(channels, channels, kernel_size=1, stride=1, padding=0, bias=True)
        self.bn1x1 = nn.BatchNorm2d(channels)
        
        self.bn = nn.BatchNorm2d(channels)

        if(deploy==False):
            self.conv3x3.weight.data = torch.randn(channels,channels,3,3)
            self.conv3x3.bias.data = torch.randn(channels)
            self.bn3x3.weight.data = torch.randn(channels)
            self.bn3x3.bias.data = torch.randn(channels)

            self.conv1x1.weight.data = torch.randn(channels,channels,1,1)
            self.conv1x1.bias.data = torch.randn(channels)
            self.bn1x1.weight.data = torch.randn(channels)
            self.bn1x1.bias.data = torch.randn(channels)
            
            self.bn.weight.data = torch.randn(channels)
            self.bn.bias.data = torch.randn(channels)

        # Fusion conv
        self.fusion_conv = nn.Conv2d(channels, channels, kernel_size=3, stride=1, padding=1, bias=True)
        self.relu = nn.ReLU(inplace=True)

    def forward(self, x):

        if self.deploy==False:
            x1 = self.conv3x3(x)
            x1 = self.bn3x3(x1)

            x2 = self.conv1x1(x)
            x2 = self.bn1x1(x2)

            x = self.bn(x)

            x = x1 + x2 + x
        else:
            x = self.fusion_conv(x)

        return self.relu(x)

    def reparam3x3(self):
        conv_w = self.conv3x3.weight
        conv_b = self.conv3x3.bias

        bn_w = self.bn3x3.weight
        bn_b = self.bn3x3.bias 

        bn_w = bn_w.div(torch.sqrt(self.bn3x3.eps + self.bn3x3.running_var))

        fusion_w = torch.mm(torch.diag(bn_w), conv_w.view(self.channels, -1)).view(self.channels,self.channels,3,3)
        fusion_b = bn_w * (conv_b - self.bn3x3.running_mean) + bn_b

        print(fusion_w.shape,fusion_b.shape)
        return fusion_w, fusion_b

    def reparam1x1(self):
        conv_w = self.conv1x1.weight
        conv_b = self.conv1x1.bias

        bn_w = self.bn1x1.weight
        bn_b = self.bn1x1.bias 

        bn_w = bn_w.div(torch.sqrt(self.bn1x1.eps + self.bn1x1.running_var))

        fusion_w = torch.mm(torch.diag(bn_w), conv_w.view(self.channels, -1)).view(self.channels,self.channels,1,1)

        w = torch.zeros(self.channels, self.channels, 3, 3)
        w[:,:,1,1] = fusion_w.squeeze(2).squeeze(2)

        fusion_b = bn_w * (conv_b - self.bn1x1.running_mean) + bn_b

        print(w.shape,fusion_b.shape)
        return w , fusion_b

    def reparamBn(self):

        bn_w = self.bn.weight
        bn_b = self.bn.bias

        bn_w = bn_w.div(torch.sqrt(self.bn.eps + self.bn.running_var))

        bn_w = torch.diag(bn_w).view(self.channels,self.channels,1,1)

        w = torch.zeros(self.channels, self.channels, 3, 3)
        w[:,:,1,1] = bn_w.squeeze(2).squeeze(2)
        print(w.shape,bn_b.shape)
        return w, bn_b
            
    def reparam(self):

        w_3x3, b_3x3 = self.reparam3x3()
        w_1x1, b_1x1 = self.reparam1x1()
        w, b = self.reparamBn()

        self.fusion_conv.weight.data = (w_3x3 + w_1x1+w).clone()
        self.fusion_conv.bias.data = (b_3x3 + b_1x1+b).clone()

x = torch.randn(1, 20, 224, 224)

net1 = RepVGGBlock(20,deploy=False)
torch.save(net1.state_dict(), "repvgg.pth")
net1.eval()
y1 = net1(x)

net2 = RepVGGBlock(20,deploy=True)
net2.load_state_dict(torch.load("repvgg.pth"))
net2.reparam()
net2.eval()
y2 = net2(x)

print(y1.shape,y2.shape)
print(torch.allclose(y1, y2, atol=1e-4))

torch.onnx.export(net1, x, "repvgg.onnx", input_names=['input'], output_names=['output'])
torch.onnx.export(net2, x, "repvgg_deploy.onnx", input_names=['input'], output_names=['output'])