CNN lightweight model mobilenet v1

mobilenet v1

Reading papers

Papers Address: https://arxiv.org/abs/1704.04861

The core idea is to replace ordinary conv by depthwise conv. About depthwise conv can refer https://www.cnblogs.com/sdu20112013/p/11759928.html

Model structure: similar to the structure of such a stack vgg.

Each layer of computation can be seen, the amount of computation is not absolutely proportional to the number of parameters , of course, the overall trend, the parameters of the model will be smaller amount of computation faster.

Code

https://github.com/marvis/pytorch-mobilenet

Network architecture:

class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()

        def conv_bn(inp, oup, stride):
            return nn.Sequential(
                nn.Conv2d(inp, oup, 3, stride, 1, bias=False),
                nn.BatchNorm2d(oup),
                nn.ReLU(inplace=True)
            )

        def conv_dw(inp, oup, stride):
            return nn.Sequential(
                nn.Conv2d(inp, inp, 3, stride, 1, groups=inp, bias=False),
                nn.BatchNorm2d(inp),
                nn.ReLU(inplace=True),
    
                nn.Conv2d(inp, oup, 1, 1, 0, bias=False),
                nn.BatchNorm2d(oup),
                nn.ReLU(inplace=True),
            )

        self.model = nn.Sequential(
            conv_bn(  3,  32, 2), 
            conv_dw( 32,  64, 1),
            conv_dw( 64, 128, 2),
            conv_dw(128, 128, 1),
            conv_dw(128, 256, 2),
            conv_dw(256, 256, 1),
            conv_dw(256, 512, 2),
            conv_dw(512, 512, 1),
            conv_dw(512, 512, 1),
            conv_dw(512, 512, 1),
            conv_dw(512, 512, 1),
            conv_dw(512, 512, 1),
            conv_dw(512, 1024, 2),
            conv_dw(1024, 1024, 1),
            nn.AvgPool2d(7),
        )
        self.fc = nn.Linear(1024, 1000)

    def forward(self, x):
        x = self.model(x)
        x = x.view(-1, 1024)
        x = self.fc(x)
        return x

Dissertation reference structure, the first layer is a common layer of convolution, convolution back contact are separable.

Here noted that the usage parameter groups when the number of groups = input channel, i.e. for each channel to do the convolution. Default groups =. 1, this case is the ordinary convolution.

Training pseudo-code

# create model
model = Net()

# define loss function (criterion) and optimizer
criterion = nn.CrossEntropyLoss().cuda()

optimizer = torch.optim.SGD(model.parameters(), args.lr,
                            momentum=args.momentum,
                            weight_decay=args.weight_decay)


# load data
train_loader = torch.utils.data.DataLoader()

# train
for every epoch:
    input,target=get_from_data
    
    #前向传播得到预测值
    output = model(input_var)
    
    #计算loss
    loss = criterion(output, target_var)
        
    #反向传播更新网络参数
    optimizer.zero_grad()
    loss.backward()
    optimizer.step()