Python深度学习之使用Pytorch搭建ShuffleNetv2

一、model.py

1.1 Channel Shuffle

def channel_shuffle(x: Tensor, groups: int) -> Tensor:

batch_size, num_channels, height, width = x.size()
   channels_per_group = num_channels // groups

# reshape
   # [batch_size, num_channels, height, width] -> [batch_size, groups, channels_per_group, height, width]
   x = x.view(batch_size, groups, channels_per_group, height, width)

x = torch.transpose(x, 1, 2).contiguous()

# flatten
   x = x.view(batch_size, -1, height, width)

return x

1.2 block

class InvertedResidual(nn.Module):
   def __init__(self, input_c: int, output_c: int, stride: int):
       super(InvertedResidual, self).__init__()

if stride not in [1, 2]:
           raise ValueError("illegal stride value.")
       self.stride = stride

assert output_c ％ 2 == 0
       branch_features = output_c // 2
       # 当stride为1时，input_channel应该是branch_features的两倍
       # python中 '<<' 是位运算，可理解为计算×2的快速方法
       assert (self.stride != 1) or (input_c == branch_features << 1)

if self.stride == 2:
           self.branch1 = nn.Sequential(
               self.depthwise_conv(input_c, input_c, kernel_s=3, stride=self.stride, padding=1),
               nn.BatchNorm2d(input_c),
               nn.Conv2d(input_c, branch_features, kernel_size=1, stride=1, padding=0, bias=False),
               nn.BatchNorm2d(branch_features),
               nn.ReLU(inplace=True)
           )
       else:
           self.branch1 = nn.Sequential()

self.branch2 = nn.Sequential(
           nn.Conv2d(input_c if self.stride > 1 else branch_features, branch_features, kernel_size=1,
                     stride=1, padding=0, bias=False),
           nn.BatchNorm2d(branch_features),
           nn.ReLU(inplace=True),
           self.depthwise_conv(branch_features, branch_features, kernel_s=3, stride=self.stride, padding=1),
           nn.BatchNorm2d(branch_features),
           nn.Conv2d(branch_features, branch_features, kernel_size=1, stride=1, padding=0, bias=False),
           nn.BatchNorm2d(branch_features),
           nn.ReLU(inplace=True)
       )

@staticmethod
   def depthwise_conv(input_c: int,
                      output_c: int,
                      kernel_s: int,
                      stride: int = 1,
                      padding: int = 0,
                      bias: bool = False) -> nn.Conv2d:
       return nn.Conv2d(in_channels=input_c, out_channels=output_c, kernel_size=kernel_s,
                        stride=stride, padding=padding, bias=bias, groups=input_c)

def forward(self, x: Tensor) -> Tensor:
       if self.stride == 1:
           x1, x2 = x.chunk(2, dim=1)
           out = torch.cat((x1, self.branch2(x2)), dim=1)
       else:
           out = torch.cat((self.branch1(x), self.branch2(x)), dim=1)

out = channel_shuffle(out, 2)

return out

1.3 shufflenet v2

class ShuffleNetV2(nn.Module):
   def __init__(self,
                stages_repeats: List[int],
                stages_out_channels: List[int],
                num_classes: int = 1000,
                inverted_residual: Callable[..., nn.Module] = InvertedResidual):
       super(ShuffleNetV2, self).__init__()

if len(stages_repeats) != 3:
           raise ValueError("expected stages_repeats as list of 3 positive ints")
       if len(stages_out_channels) != 5:
           raise ValueError("expected stages_out_channels as list of 5 positive ints")
       self._stage_out_channels = stages_out_channels

# input RGB image
       input_channels = 3
       output_channels = self._stage_out_channels[0]

self.conv1 = nn.Sequential(
           nn.Conv2d(input_channels, output_channels, kernel_size=3, stride=2, padding=1, bias=False),
           nn.BatchNorm2d(output_channels),
           nn.ReLU(inplace=True)
       )
       input_channels = output_channels

self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)

# Static annotations for mypy
       self.stage2: nn.Sequential
       self.stage3: nn.Sequential
       self.stage4: nn.Sequential

stage_names = ["stage{}".format(i) for i in [2, 3, 4]]
       for name, repeats, output_channels in zip(stage_names, stages_repeats,
                                                 self._stage_out_channels[1:]):
           seq = [inverted_residual(input_channels, output_channels, 2)]
           for i in range(repeats - 1):
               seq.append(inverted_residual(output_channels, output_channels, 1))
           setattr(self, name, nn.Sequential(*seq))
           input_channels = output_channels

output_channels = self._stage_out_channels[-1]
       self.conv5 = nn.Sequential(
           nn.Conv2d(input_channels, output_channels, kernel_size=1, stride=1, padding=0, bias=False),
           nn.BatchNorm2d(output_channels),
           nn.ReLU(inplace=True)
       )

self.fc = nn.Linear(output_channels, num_classes)

def _forward_impl(self, x: Tensor) -> Tensor:
       # See note [TorchScript super()]
       x = self.conv1(x)
       x = self.maxpool(x)
       x = self.stage2(x)
       x = self.stage3(x)
       x = self.stage4(x)
       x = self.conv5(x)
       x = x.mean([2, 3])  # global pool
       x = self.fc(x)
       return x

def forward(self, x: Tensor) -> Tensor:
       return self._forward_impl(x)

二、train.py

来源：https://blog.csdn.net/weixin_43154149/article/details/116267653