DARTS代碼分析（Pytorch）

最近在看DARTS的代碼，有一個operations.py的文件，里面是對各類點與點之間操作的方法。

OPS = {
    'none': lambda C, stride, affine: Zero(stride),
    'avg_pool_3x3': lambda C, stride, affine: PoolBN('avg', C, 3, stride, 1, affine=affine),
    'max_pool_3x3': lambda C, stride, affine: PoolBN('max', C, 3, stride, 1, affine=affine),
    'skip_connect': lambda C, stride, affine: \
        Identity() if stride == 1 else FactorizedReduce(C, C, affine=affine),
    'sep_conv_3x3': lambda C, stride, affine: SepConv(C, C, 3, stride, 1, affine=affine),
    'sep_conv_5x5': lambda C, stride, affine: SepConv(C, C, 5, stride, 2, affine=affine),
    'sep_conv_7x7': lambda C, stride, affine: SepConv(C, C, 7, stride, 3, affine=affine),
    'dil_conv_3x3': lambda C, stride, affine: DilConv(C, C, 3, stride, 2, 2, affine=affine), # 5x5
    'dil_conv_5x5': lambda C, stride, affine: DilConv(C, C, 5, stride, 4, 2, affine=affine), # 9x9
    'conv_7x1_1x7': lambda C, stride, affine: FacConv(C, C, 7, stride, 3, affine=affine)
}

首先定義10個操作，依次解釋：

• class PoolBN(nn.Module):
      """
      AvgPool or MaxPool - BN
      """
      def __init__(self, pool_type, C, kernel_size, stride, padding, affine=True):
          """
          Args:
              pool_type: 'max' or 'avg'
          """
          super().__init__()
          if pool_type.lower() == 'max':
              self.pool = nn.MaxPool2d(kernel_size, stride, padding)
          elif pool_type.lower() == 'avg':
              self.pool = nn.AvgPool2d(kernel_size, stride, padding, count_include_pad=False)
          else:
              raise ValueError()
  
          self.bn = nn.BatchNorm2d(C, affine=affine)
  
      def forward(self, x):
          out = self.pool(x)
          out = self.bn(out)
          return out

  這是池化函數，有最大池化和平均池化方法，count_include_pad=False表示不把填充的0計算進去

• class Identity(nn.Module):
      def __init__(self):
          super().__init__()
  
      def forward(self, x):
          return x

  這個表示skip conncet

• class FactorizedReduce(nn.Module):
      """
      Reduce feature map size by factorized pointwise(stride=2).
      """
      def __init__(self, C_in, C_out, affine=True):
          super().__init__()
          self.relu = nn.ReLU()
          self.conv1 = nn.Conv2d(C_in, C_out // 2, 1, stride=2, padding=0, bias=False)
          self.conv2 = nn.Conv2d(C_in, C_out // 2, 1, stride=2, padding=0, bias=False)
          self.bn = nn.BatchNorm2d(C_out, affine=affine)
  
      def forward(self, x):
          x = self.relu(x)
          out = torch.cat([self.conv1(x), self.conv2(x[:, :, 1:, 1:])], dim=1)
          out = self.bn(out)
          return out

  這個表示將特征圖大小變為原來的一半

• class DilConv(nn.Module):
      """ (Dilated) depthwise separable conv
      ReLU - (Dilated) depthwise separable - Pointwise - BN
  
      If dilation == 2, 3x3 conv => 5x5 receptive field
                        5x5 conv => 9x9 receptive field
      """
      def __init__(self, C_in, C_out, kernel_size, stride, padding, dilation, affine=True):
          super().__init__()
          self.net = nn.Sequential(
              nn.ReLU(),
              nn.Conv2d(C_in, C_in, kernel_size, stride, padding, dilation=dilation, groups=C_in,
                        bias=False),
              nn.Conv2d(C_in, C_out, 1, stride=1, padding=0, bias=False),
              nn.BatchNorm2d(C_out, affine=affine)
          )
  
      def forward(self, x):
          return self.net(x)

  深度可分離卷積，groups=C_in，表示把輸入特種圖分成C_in（輸入通道數）那么多組，然后加C_out（輸出通道數）1*1的卷積，這樣可以對每個通道單獨提取特征，同時降低了參數量和計算量。

• class SepConv(nn.Module):
      """ Depthwise separable conv
      DilConv(dilation=1) * 2
      """
      def __init__(self, C_in, C_out, kernel_size, stride, padding, affine=True):
          super().__init__()
          self.net = nn.Sequential(
              DilConv(C_in, C_in, kernel_size, stride, padding, dilation=1, affine=affine),
              DilConv(C_in, C_out, kernel_size, 1, padding, dilation=1, affine=affine)
          )
  
      def forward(self, x):
          return self.net(x)

  深度可分離卷積，由兩個上面的深度分組卷積組成

• class FacConv(nn.Module):
      """ Factorized conv
      ReLU - Conv(Kx1) - Conv(1xK) - BN
      """
      def __init__(self, C_in, C_out, kernel_length, stride, padding, affine=True):
          super().__init__()
          self.net = nn.Sequential(
              nn.ReLU(),
              nn.Conv2d(C_in, C_in, (kernel_length, 1), stride, padding, bias=False),
              nn.Conv2d(C_in, C_out, (1, kernel_length), stride, padding, bias=False),
              nn.BatchNorm2d(C_out, affine=affine)
          )
  
      def forward(self, x):
          return self.net(x)

  這個表示長方形的卷積，增加了一點特征圖的長和寬

• class Zero(nn.Module):
      def __init__(self, stride):
          super().__init__()
          self.stride = stride
  
      def forward(self, x):
          if self.stride == 1:
              return x * 0.
  
          # re-sizing by stride
          return x[:, :, ::self.stride, ::self.stride] * 0.

  這個表示把特種圖的輸出變為全是0，但特征圖的大小會根據stride而改變