2020-08-13

git init testdir
cd testdir/
git config core.sparsecheckout true
#echo 'model_zoo/official/cv/mobilenetv2' >> .git/info/sparse-checkout
echo 'model_zoo/official/cv/mobilenetv2' >> .git/info/sparse-checkout
git remote add -f origin git@gitee.com:mindspore/mindspore.git
#git remote add -f origin https://github.com/mindspore-ai/mindspore.git
git pull origin master
mv model_zoo/official/cv/mobilenetv2 ../mobilenetv2/
rm ../testdir -rf

https://s3.cn-north-1.amazonaws.com.cn/files.datafountain.cn/uploads/images/competitions/557/%E8%B5%9B%E9%A2%98%E6%95%B0%E6%8D%AE%EF%BC%88data%29-430493.ziphttps://s3.cn-north-1.amazonaws.com.cn/files.datafountain.cn/uploads/images/competitions/557/%E5%BB%BA%E6%A8%A1%E6%A0%B7%E4%BE%8B%EF%BC%88pytorch_reference_model%EF%BC%89-577358.ziphttps://s3.cn-north-1.amazonaws.com.cn/files.datafountain.cn/uploads/images/competitions/557/%E8%B5%9B%E9%A2%98%E4%BB%A3%E7%A0%81%EF%BC%88Pytorch_evaluation_inference%EF%BC%89-532530.zip

class DownBlock(nn.Module):
    def __init__(self, in_features, out_features):
        super(DownBlock, self).__init__()
        self.down = nn.Sequential(
            nn.Conv2d(in_features, out_features, 3, 2, padding=1),
            nn.LeakyReLU(),
            nn.Dropout(p=0.25)
        )

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


class Discriminator(nn.Module):
    def __init__(self, num_blocks=4):
        super(Discriminator, self).__init__()
        zero_padding = nn.Sequential(
            nn.ZeroPad2d((32, 32, 0, 0)),
        )
        down_blocks = []
        for i in range(num_blocks):
            down_blocks.append(
                DownBlock(1 if i == 0 else i * 8, (i + 1) * 8))

        linear_block = nn.Sequential(
            nn.Linear(64, 1),
        )
        self.zero_padding = zero_padding
        self.conv_block = nn.ModuleList(down_blocks)
        self.linear_block = linear_block

    def forward(self, input):
        output = self.zero_padding(input)
        output = self.conv_block(output)
        output = output.view(-1, 64)
        output = self.linear_block(output)
        return output

class UpSamplingBlock(nn.Module):
    def __init__(self, in_features, out_features, v):
        super(UpSamplingBlock, self).__init__()
        self.cov = nn.Conv2d(in_features, out_features, 3, 1, padding=1)
        self.bn = nn.BatchNorm1d(out_features)
        self.v = v
        self.act = nn.LeakyReLU(negative_slope=0.2, inplace=True) if self.v == 0 else nn.Tanh()

    def forward(self, x):
        x = F.interpolate(x, scale_factor=2)
        x = self.cov(x)
        x = self.bn(x)
        x = self.act(x)
        return x
class Generator(nn.Module):
    def __init__(self, num_blocks=4):
        super(Generator, self).__init__()
        linear_block = nn.Sequential(
            nn.Linear(128, 8192),
            nn.BatchNorm1d(8192),
            nn.LeakyReLU(),
        )
        up_blocks = []
        m = [1024, 1024, 512, 256, 128, 2]
        v = [0, 0, 0, 0, 1]
        for i in range(num_blocks):
            up_blocks.append(
                UpSamplingBlock(m[i], m[i+1], v[i]))
        self.linear_block = linear_block
        self.conv_block = nn.ModuleList(up_blocks)

    def forward(self, input):
        output = self.linear_block(input)
        output = output.view(-1, 64, 4, 2)
        for up in self.conv_block:
            output = up(output)
        output = output[:, :, self.crop_size:-self.crop_size]
        return output