Classic network learning RESNET code implementation

Preface

Based on the previous theoretical analysis , Today, let's talk about learning ResNet Code implementation of , If you haven't seen << Classic online learning -ResNet>> It's suggested to take a look at . Before I write this , I also investigated other online implementations , Are all behind pytorch The official source code is well implemented , So the official version explains how to achieve resNet

• pytorch resnet Source link

ResNet framework

Here is still the architecture diagram in the paper ：

Each layer in the figure is actually BasicBlock perhaps BotteNeck structure . Here is given ResNet-34 The structure diagram is shown in the figure , The dashed connecting line in the figure indicates that the number of channels is different , Channel adjustment required Use zero padding or 1x1 To achieve this goal .

## Residual structure The residual structure diagram is as follows ：

The code is interpreted as ：
conv->bn->relu->conv->bn->shortcut->relu

BasicBlock structure

#  be used for resnet18 and resnet34 Basic residual structure block 
#downsample The residual structure corresponding to the dotted line 
# # Downsampling is performed by conv3_1, conv4_1, and conv5_1 with a stride of 2
class BasicBlock(nn.Module):
    # Channel expansion factor , The base number is 64
    expansion: int = 1

    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        stride: int = 1,
        downsample: Optional[nn.Module] = None,
        groups: int = 1,
        base_width: int = 64,
        dilation: int = 1,
        norm_layer: Optional[Callable[..., nn.Module]] = None,
    ) -> None:
        super().__init__()
        if norm_layer is None:
            norm_layer = nn.BatchNorm2d
        if groups != 1 or base_width != 64:
            raise ValueError("BasicBlock only supports groups=1 and base_width=64")
        if dilation > 1:
            raise NotImplementedError("Dilation > 1 not supported in BasicBlock")
        # Both self.conv1 and self.downsample layers downsample the input when stride != 1
        self.conv1 = conv3x3(in_channels, out_channels, stride)
        self.bn1 = norm_layer(out_channels)
        self.relu = nn.ReLU(inplace=True)
        self.conv2 = conv3x3(out_channels, out_channels)
        self.bn2 = norm_layer(out_channels)
        self.downsample = downsample
        self.stride = stride

    def forward(self, x: Tensor) -> Tensor:
        identity = x

        out = self.conv1(x)
        out = self.bn1(out)
        out = self.relu(out)

        out = self.conv2(out)
        out = self.bn2(out)

        # The dotted line of the model architecture in the paper , Down sampling is required 
        if self.downsample is not None:
            identity = self.downsample(x)
        
        #shortcut Connect 
        out += identity
        out = self.relu(out)

        return out

In code conv3x3 Definition

# Definition 3x3 belt padding Convolution of 
def conv3x3(in_planes: int, out_planes: int, stride: int = 1, groups: int = 1, dilation: int = 1) -> nn.Conv2d:
    """3x3 convolution with padding"""
    return nn.Conv2d(
        in_planes,
        out_planes,
        kernel_size=3,
        stride=stride,
        padding=dilation,
        groups=groups,
        bias=False,
        dilation=dilation,
    )

• What needs to be noted is ：bias = False

therefore , After convolution , If you want to pick up BN operation , It is better not to set the offset , Because it doesn't work , And occupy the memory of the graphics card .

BottleNeck structure

class Bottleneck(nn.Module):
    # pytorch  Realization  Bottleneck  Is in 3x3 Convolution (self.conv2) Set up stride = 2
    #  The original paper (https://arxiv.org/abs/1512.03385) To realize  Bottleneck  Is in 1x1 Convolution (self.conv1) Set up stride = 2
    #  This improves the accuracy .  This variant is also known as ResNet V1.5  Reference resources https://ngc.nvidia.com/catalog/model-scripts/nvidia:resnet_50_v1_5_for_pytorch.

    # Channel expansion factor 
    expansion: int = 4

    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        stride: int = 1,
        downsample: Optional[nn.Module] = None,
        groups: int = 1,
        base_width: int = 64,
        dilation: int = 1,
        norm_layer: Optional[Callable[..., nn.Module]] = None,
    ) -> None:
        super().__init__()
        if norm_layer is None:
            norm_layer = nn.BatchNorm2d
        width = int(out_channels * (base_width / 64.0)) * groups
        # Both self.conv2 and self.downsample layers downsample the input when stride != 1
        self.conv1 = conv1x1(in_channels, width)
        self.bn1 = norm_layer(width)
        self.conv2 = conv3x3(width, width, stride, groups, dilation)
        self.bn2 = norm_layer(width)
        self.conv3 = conv1x1(width, out_channels * self.expansion)
        self.bn3 = norm_layer(out_channels * self.expansion)
        self.relu = nn.ReLU(inplace=True)
        self.downsample = downsample
        self.stride = stride

    def forward(self, x: Tensor) -> Tensor:
        identity = x

        out = self.conv1(x)
        out = self.bn1(out)
        out = self.relu(out)

        out = self.conv2(out)
        out = self.bn2(out)
        out = self.relu(out)

        out = self.conv3(out)
        out = self.bn3(out)

        if self.downsample is not None:
            identity = self.downsample(x)

        out += identity
        out = self.relu(out)

        return out

ResNet Code implementation

class ResNet(nn.Module):
    def __init__(
        self,
        block: Type[Union[BasicBlock, Bottleneck]],
        layers: List[int],
        num_classes: int = 1000,
        zero_init_residual: bool = False,
        groups: int = 1,
        width_per_group: int = 64,
        replace_stride_with_dilation: Optional[List[bool]] = None,
        norm_layer: Optional[Callable[..., nn.Module]] = None,
    ) -> None:
        super().__init__()
        if norm_layer is None:
            norm_layer = nn.BatchNorm2d
        self._norm_layer = norm_layer

        self.in_channels = 64
        self.dilation = 1
        if replace_stride_with_dilation is None:
            # each element in the tuple indicates if we should replace
            # the 2x2 stride with a dilated convolution instead
            replace_stride_with_dilation = [False, False, False]
        if len(replace_stride_with_dilation) != 3:
            raise ValueError(
                "replace_stride_with_dilation should be None "
                f"or a 3-element tuple, got {
      replace_stride_with_dilation}"
            )
        self.groups = groups
        self.base_width = width_per_group
        self.conv1 = nn.Conv2d(3, self.in_channels, kernel_size=7, stride=2, padding=3, bias=False)
        self.bn1 = norm_layer(self.in_channels)
        self.relu = nn.ReLU(inplace=True)
        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)

        self.layer1 = self._make_layer(block, 64, layers[0])
        self.layer2 = self._make_layer(block, 128, layers[1], stride=2, dilate=replace_stride_with_dilation[0])
        self.layer3 = self._make_layer(block, 256, layers[2], stride=2, dilate=replace_stride_with_dilation[1])
        self.layer4 = self._make_layer(block, 512, layers[3], stride=2, dilate=replace_stride_with_dilation[2])
        self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
        self.fc = nn.Linear(512 * block.expansion, num_classes)

        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                nn.init.kaiming_normal_(m.weight, mode="fan_out", nonlinearity="relu")
            elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
                nn.init.constant_(m.weight, 1)
                nn.init.constant_(m.bias, 0)

        # Zero-initialize the last BN in each residual branch( The last of each residual block BN Initialize with zero ),
        # so that the residual branch starts with zeros, and each residual block behaves like an identity.( So each residual block starts from zero , It's like identity)
        # This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677 ( Improved accuracy 0.2~0.3)
        if zero_init_residual:
            for m in self.modules():
                if isinstance(m, Bottleneck):
                    nn.init.constant_(m.bn3.weight, 0)  # type: ignore[arg-type]
                elif isinstance(m, BasicBlock):
                    nn.init.constant_(m.bn2.weight, 0)  # type: ignore[arg-type]

    #  establish conv2_x,conv3_x,conv4_x,conv5_x layer 
    # channel:conv2/3/4/5 The number of the first convolution kernel on the main branch of the residual structure corresponding to various depths / The channel number 
    #  The number of residual structures of a convolution 
    def _make_layer(
        self,
        # Residual block type ： It can be BasicBlock perhaps Bottleneck
        block: Type[Union[BasicBlock, Bottleneck]],
        # The residual is faster than the input channel of the first convolution 
        channels: int,
        # Number of residual blocks 
        blocks: int,
        stride: int = 1,
        dilate: bool = False,
    ) -> nn.Sequential:
        norm_layer = self._norm_layer
        downsample = None
        previous_dilation = self.dilation
        if dilate:
            self.dilation *= stride
            stride = 1
        #  about resnet50/101/152 Layer structure , The first layer is the dotted line residual , Take the next sample 
        #  about resnet18/34 Layer network will skip this judgment , Because input and output shape Agreement , No down sampling 
        # conv2_x The first layer of lower sampling only needs to be increased channel, Don't change the height and width (stride = 1) Because input and output shape All for 64×64
        if stride != 1 or self.in_channels != channels * block.expansion:
            downsample = nn.Sequential(
                conv1x1(self.in_channels, channels * block.expansion, stride),
                norm_layer(channels * block.expansion),
            )

        layers = []
        layers.append(
            block(
                self.in_channels, channels, stride, downsample, self.groups, self.base_width, previous_dilation, norm_layer
            )
        )
        self.in_channels = channels * block.expansion
        for _ in range(1, blocks):
            layers.append(
                block(
                    self.in_channels,
                    channels,
                    groups=self.groups,
                    base_width=self.base_width,
                    dilation=self.dilation,
                    norm_layer=norm_layer,
                )
            )

        #Sequential Class to implement a simple sequential connection model 
        return nn.Sequential(*layers)

    def _forward_impl(self, x: Tensor) -> Tensor:
        ''' Forward propagation implementation function '''

        x = self.conv1(x)
        x = self.bn1(x)
        x = self.relu(x)
        x = self.maxpool(x)

        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)
        x = self.layer4(x)

        # Global average pooling 
        x = self.avgpool(x)
        x = torch.flatten(x, 1)
        #  The last full connection layer 
        x = self.fc(x)

        return x

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

The above codes are from pytorch Source code , There are deletions for easy understanding , All the above have been put into github Warehouse