StackHourGlass的pytorch实现

import torch

# conv_BN_ReLu
class ConvolutionalLayer(nn.Module):

    def __init__(self, in_channels, out_channels, kernel_size, stride, padding, bias=False):
        super(ConvolutionalLayer, self).__init__()

        self.sub_module = nn.Sequential(
            nn.Conv2d(in_channels, out_channels, kernel_size, stride, padding, bias=bias),
            nn.BatchNorm2d(out_channels),
            nn.LeakyReLU()
        )

    def forward(self, x):
        return self.sub_module(x)

# 残差1x1 -> 3x3 -> 1x1
class Residual(nn.Module):

    def __init__(self, in_channels, out_channels):
        super(Residual, self).__init__()

        self.sub_module = nn.Sequential(
            ConvolutionalLayer(in_channels, out_channels, 1, 1, 0),
            ConvolutionalLayer(out_channels, out_channels, 3, 1, 1),
            ConvolutionalLayer(out_channels, in_channels, 1, 1, 0),
        )

    def forward(self, x):
        return x + self.sub_module(x)

# 预处理，减小1/4
class Preprocessing(nn.Module):

    def __init__(self, in_channels, out_channels):
        super(Preprocessing, self).__init__()

        self.sub_module = nn.Sequential(
            ConvolutionalLayer(in_channels, out_channels, 3, 2, 1),
            Residual(out_channels, out_channels),
            nn.MaxPool2d(2, 2)
        )

    def forward(self, x):
        return self.sub_module(x)

# 下采样1/2
class DownSampling(nn.Module):

    def __init__(self, in_channels, out_channels):
        super(DownSampling, self).__init__()

        self.layer = nn.Sequential(
            nn.Conv2d(in_channels, out_channels, 3, 2, 1),
            nn.LeakyReLU()
        )

    def forward(self, x):
        return self.layer(x)

# 上采样x2
class UpSampling(nn.Module):

    def __init__(self, in_channels, out_channels):
        super(UpSampling, self).__init__()

        self.layer = nn.Sequential(
            nn.ConvTranspose2d(in_channels, out_channels, 3, 2, 1, 1)
        )

    def forward(self, x):
        return self.layer(x)


class HourGlass(nn.Module):

    def __init__(self, nChannels, numReductions=4, numModules=2):
        super(HourGlass, self).__init__()
        self.nChannels = nChannels
        self.numReductions = numReductions
        self.numModules = numModules

        skip = []
        skip.append(nn.Conv2d(self.nChannels[self.numReductions], self.nChannels[self.numReductions-1], 1, 1))
        for _ in range(self.numModules):
            skip.append(Residual(self.nChannels[self.numReductions-1], self.nChannels[self.numReductions-1]))
        self.skip = nn.Sequential(*skip)

        self.down = DownSampling(self.nChannels[self.numReductions], self.nChannels[self.numReductions-1])

        afterpool = []
        for _ in range(self.numModules):
            afterpool.append(Residual(self.nChannels[self.numReductions-1], self.nChannels[self.numReductions-1]))
        self.afterpool = nn.Sequential(*afterpool)

        if numReductions > 1:
            self.hg = HourGlass(self.nChannels, self.numReductions - 1, self.numModules)
        else:
            num1res = []
            for _ in range(self.numModules):
                num1res.append(Residual(self.nChannels[self.numReductions-1], self.nChannels[self.numReductions-1]))
            self.num1res = nn.Sequential(*num1res)

        lowers = []
        lowers.append(nn.Conv2d(self.nChannels[self.numReductions-1], self.nChannels[self.numReductions], 1, 1))
        for _ in range(self.numModules):
            lowers.append(Residual(self.nChannels[self.numReductions], self.nChannels[self.numReductions]))
        self.lowers = nn.Sequential(*lowers)

        self.up = UpSampling(self.nChannels[self.numReductions], self.nChannels[self.numReductions-1])
        self.merge = nn.Conv2d(self.nChannels[self.numReductions-1], self.nChannels[self.numReductions], 1, 1)

    def forward(self, x):
        out1 = self.skip(x) # 2个残差
        out2 = self.down(x) # 下采样1/2
        out2 = self.afterpool(out2) # 2个残差
        if self.numReductions > 1:
            out2 = self.hg(out2) # 递归调用
        else:
            out2 = self.num1res(out2) # 2个残差
        out2 = self.lowers(out2) # 2个残差
        out2 = self.up(out2) # 上采样x2
        out2 = out1 + out2
        out3 = self.merge(out2)

        return out3

class StackHourGlass(nn.Module):

    def __init__(self, stack_num, nChannels):
        super(StackHourGlass, self).__init__()
        self.stack_num = stack_num
        self.nChannels = nChannels
        self.hg = HourGlass([nChannels*8, nChannels*4, nChannels*2, nChannels, nChannels])
        self.res = ConvolutionalLayer(self.nChannels, 1, 1, 1, 0)
        self.joints = ConvolutionalLayer(1, self.nChannels, 1, 1, 0)

    def forward(self, x):
        out = []
        for i in range(self.stack_num):
            x1 = self.hg(x)
            x1 = self.res(x1)
            out.append(x1)
            if i != self.stack_num - 1:
                x = x + self.joints(x1)
        return out

class Main(nn.Module):

    def __init__(self, stack_num=2, nChannels=8):
        super(Main, self).__init__()

        self.pre_process = Preprocessing(in_channels=3, out_channels=nChannels)
        self.hg = StackHourGlass(stack_num, nChannels)

    def forward(self, x):
        pre_process = self.pre_process(x)
        hg = self.hg(pre_process)

        return hg


if __name__ == "__main__":
    hg = Main().cuda()
    x = torch.randn((2, 3, 512, 512)).cuda()
    out = hg(x)
    print(out)