mxnet的网络结构，
https://github.com/zma-c-137/VarGFaceNet
pytorch:
作者自己改的：
https://www.cnblogs.com/wanghui-garcia/p/12582953.html
pytorch，也是自己写的：
https://blog.csdn.net/shanglianlm/article/details/104586878
加上l2 norm好像能收敛。
收敛一段时间又loss为nan了， 又加了bn1d,又能继续训练了。
https://github.com/shanglianlm0525/PyTorch-Networks/blob/24ac3e6b5fdca7fefcd8d0313adc88d4244f0576/FaceDetectorAndRecognition/VarGFaceNet.py
import torch
import torch.nn as nn
import torchvisionclass SqueezeAndExcite(nn.Module):def __init__(self, in_channels, out_channels, divide=4):super(SqueezeAndExcite, self).__init__()mid_channels = in_channels // divideself.pool = nn.AdaptiveAvgPool2d(1)self.SEblock = nn.Sequential(nn.Linear(in_features=in_channels, out_features=mid_channels),nn.ReLU6(inplace=True),nn.Linear(in_features=mid_channels, out_features=out_channels),nn.ReLU6(inplace=True),)def forward(self, x):b, c, h, w = x.size()out = self.pool(x)out = out.view(b, -1)out = self.SEblock(out)out = out.view(b, c, 1, 1)return out * xdef Conv1x1BNReLU(in_channels,out_channels):return nn.Sequential(nn.Conv2d(in_channels=in_channels, out_channels=out_channels, kernel_size=1, stride=1, padding=0, bias=False),nn.BatchNorm2d(out_channels),nn.ReLU6(inplace=True))def Conv3x3BNReLU(in_channels,out_channels,stride):return nn.Sequential(nn.Conv2d(in_channels=in_channels, out_channels=out_channels, kernel_size=3, stride=stride, padding=1, bias=False),nn.BatchNorm2d(out_channels),nn.ReLU6(inplace=True))def VarGConv(in_channels,out_channels,kernel_size,stride,S):return nn.Sequential(nn.Conv2d(in_channels, out_channels, kernel_size, stride, padding=kernel_size // 2, groups=in_channels // S,bias=False),nn.BatchNorm2d(out_channels),nn.PReLU(),)def VarGPointConv(in_channels, out_channels,stride,S,isRelu):return nn.Sequential(nn.Conv2d(in_channels, out_channels, 1, stride, padding=0, groups=in_channels // S,bias=False),nn.BatchNorm2d(out_channels),nn.PReLU() if isRelu else nn.Sequential(),)class VarGBlock_S1(nn.Module):def __init__(self, in_plances,kernel_size, stride=1, S=8):super(VarGBlock_S1, self).__init__()plances = 2 * in_plancesself.varGConv1 = VarGConv(in_plances, plances, kernel_size, stride, S)self.varGPointConv1 = VarGPointConv(plances, in_plances, stride, S, isRelu=True)self.varGConv2 = VarGConv(in_plances, plances, kernel_size, stride, S)self.varGPointConv2 = VarGPointConv(plances, in_plances, stride, S, isRelu=False)self.se = SqueezeAndExcite(in_plances,in_plances)self.prelu = nn.PReLU()def forward(self, x):out = xx = self.varGPointConv1(self.varGConv1(x))x = self.varGPointConv2(self.varGConv2(x))x = self.se(x)out += xreturn self.prelu(out)class VarGBlock_S2(nn.Module):def __init__(self, in_plances,kernel_size, stride=2, S=8):super(VarGBlock_S2, self).__init__()plances = 2 * in_plancesself.varGConvBlock_branch1 = nn.Sequential(VarGConv(in_plances, plances, kernel_size, stride, S),VarGPointConv(plances, plances, 1, S, isRelu=True),)self.varGConvBlock_branch2 = nn.Sequential(VarGConv(in_plances, plances, kernel_size, stride, S),VarGPointConv(plances, plances, 1, S, isRelu=True),)self.varGConvBlock_3 = nn.Sequential(VarGConv(plances, plances*2, kernel_size, 1, S),VarGPointConv(plances*2, plances, 1, S, isRelu=False),)self.shortcut = nn.Sequential(VarGConv(in_plances, plances, kernel_size, stride, S),VarGPointConv(plances, plances, 1, S, isRelu=False),)self.prelu = nn.PReLU()def forward(self, x):out = self.shortcut(x)x1 = x2 = xx1= self.varGConvBlock_branch1(x1)x2 = self.varGConvBlock_branch2(x2)x_new = x1 + x2x_new = self.varGConvBlock_3(x_new)out += x_newreturn self.prelu(out)class HeadBlock(nn.Module):def __init__(self, in_plances, kernel_size, S=8):super(HeadBlock, self).__init__()self.varGConvBlock = nn.Sequential(VarGConv(in_plances, in_plances, kernel_size, 2, S),VarGPointConv(in_plances, in_plances, 1, S, isRelu=True),VarGConv(in_plances, in_plances, kernel_size, 1, S),VarGPointConv(in_plances, in_plances, 1, S, isRelu=False),)self.shortcut = nn.Sequential(VarGConv(in_plances, in_plances, kernel_size, 2, S),VarGPointConv(in_plances, in_plances, 1, S, isRelu=False),)def forward(self, x):out = self.shortcut(x)x = self.varGConvBlock(x)out += xreturn outclass TailEmbedding(nn.Module):def __init__(self, in_plances, plances=512, S=8):super(TailEmbedding, self).__init__()self.embedding = nn.Sequential(Conv1x1BNReLU(in_plances, 1024),nn.Conv2d(1024, 1024, 7, 1, padding=0, groups=1024 // S,bias=False),nn.Conv2d(1024, 512, 1, 1, padding=0, groups=512, bias=False),)self.fc = nn.Linear(in_features=512,out_features=plances)def forward(self, x):x = self.embedding(x)x = x.view(x.size(0),-1)out = self.fc(x)return outclass VarGFaceNet(nn.Module):def __init__(self, num_classes=512):super(VarGFaceNet, self).__init__()S = 8self.conv1 = Conv3x3BNReLU(3, 40, 1)self.head = HeadBlock(40,3)self.stage2 = nn.Sequential(VarGBlock_S2(40,3,2),VarGBlock_S1(80, 3, 1),VarGBlock_S1(80, 3, 1),)self.stage3 = nn.Sequential(VarGBlock_S2(80, 3, 2),VarGBlock_S1(160, 3, 1),VarGBlock_S1(160, 3, 1),VarGBlock_S1(160, 3, 1),VarGBlock_S1(160, 3, 1),VarGBlock_S1(160, 3, 1),VarGBlock_S1(160, 3, 1),)self.stage4 = nn.Sequential(VarGBlock_S2(160, 3, 2),VarGBlock_S1(320, 3, 1),VarGBlock_S1(320, 3, 1),VarGBlock_S1(320, 3, 1),)self.tail = TailEmbedding(320,num_classes)def init_params(self):for m in self.modules():if isinstance(m, nn.Conv2d):nn.init.kaiming_normal_(m.weight)nn.init.constant_(m.bias, 0)elif isinstance(m, nn.Linear) or isinstance(m, nn.BatchNorm2d):nn.init.constant_(m.weight, 1)nn.init.constant_(m.bias, 0)def forward(self, x):x = self.conv1(x)x = self.head(x)x = self.stage2(x)x = self.stage3(x)x = self.stage4(x)out= self.tail(x)return outif __name__ == '__main__':model = VarGFaceNet()print(model)input = torch.randn(1, 3, 112, 112)out = model(input)print(out.shape)