卷积核尺度大小，步长，填充，膨胀率，对输出尺度的影响

卷积

代码

import torch
import torch.nn as nn


class BasicConv(nn.Module):
    def __init__(self, in_planes, out_planes, kernel_size, stride=1, padding=0, dilation=1, groups=1, relu=True, bn=True, bias=False):
        super(BasicConv, self).__init__()
        if in_planes <= 0 or out_planes <= 0:
            raise ValueError("Input and output channels must be greater than zero")
        self.out_channels = out_planes
        self.conv = nn.Conv2d(in_planes, out_planes, kernel_size=kernel_size, stride=stride, padding=padding, dilation=dilation, groups=groups, bias=bias)
        self.bn = nn.BatchNorm2d(out_planes, eps=1e-5, momentum=0.01, affine=True) if bn else None
        self.relu = nn.ReLU(inplace=True) if relu else None

    def forward(self, x):
        x = self.conv(x)
        if self.bn is not None:
            x = self.bn(x)
        if self.relu is not None:
            x = self.relu(x)
        return x


class Test(nn.Module):
    def __init__(self):
        super(Test, self).__init__()
        # Branch 1: Dilated Convolution with dilation rate 3 and stride 2 for downsampling
        self.branch1 = nn.Sequential(
            BasicConv(3, 3, kernel_size=3, stride=1, padding=1, dilation=1),
        )
    def forward(self, x):
        x1 = self.branch1(x)
        return x1


# Example usage
if __name__ == "__main__":
    inputs = torch.randn(1, 3, 608, 608)  # Example input tensor (batch_size, channels, height, width)
    # net = BasicRFB(in_planes=3, out_planes=3,stride=4)  # Use stride=4 for aggressive downsampling
    # out = net(inputs)
    net =Test()
    out = net(inputs)

    print("Input shape:", inputs.shape)
    print("Output shape:", out.shape)

padding（当 stride=1时，padding=kernel_size//2,保持原尺寸）

BasicConv(3, 3, kernel_size=3, stride=1, padding=1, dilation=1),

Input shape: torch.Size([1, 3, 608, 608])

Output shape: torch.Size([1, 3, 608, 608])

BasicConv(3, 3, kernel_size=3, stride=1, padding=2, dilation=1),

Input shape: torch.Size([1, 3, 608, 608])

Output shape: torch.Size([1, 3, 610, 610])

BasicConv(3, 3, kernel_size=3, stride=1, padding=3, dilation=1),

Input shape: torch.Size([1, 3, 608, 608])

Output shape: torch.Size([1, 3, 612, 612])

padding（当 stride=2时，padding=kernel_size//2,保持原尺寸）

BasicConv(3, 3, kernel_size=3, stride=2, padding=1, dilation=1),

Input shape: torch.Size([1, 3, 608, 608])

Output shape: torch.Size([1, 3, 304, 304])

BasicConv(3, 3, kernel_size=3, stride=2, padding=2, dilation=1),

Input shape: torch.Size([1, 3, 608, 608])

Output shape: torch.Size([1, 3, 305, 305])

BasicConv(3, 3, kernel_size=3, stride=2, padding=3, dilation=1),

Input shape: torch.Size([1, 3, 608, 608])

Output shape: torch.Size([1, 3, 306, 306])

padding（当 stride=3时，padding=kernel_size//2,保持原尺寸）

BasicConv(3, 3, kernel_size=3, stride=3, padding=1, dilation=1),

Input shape: torch.Size([1, 3, 608, 608])

Output shape: torch.Size([1, 3, 203, 203])

BasicConv(3, 3, kernel_size=3, stride=3, padding=2, dilation=1),

Input shape: torch.Size([1, 3, 608, 608])

Output shape: torch.Size([1, 3, 204, 204])

BasicConv(3, 3, kernel_size=3, stride=3, padding=3, dilation=1),

Input shape: torch.Size([1, 3, 608, 608])

Output shape: torch.Size([1, 3, 204, 204])

dilation：膨胀率（）卷积核每隔几个。

BasicConv(3, 3, kernel_size=3, stride=4, padding=1, dilation=1),

Input shape: torch.Size([1, 3, 608, 608])

Output shape: torch.Size([1, 3, 152, 152])

BasicConv(3, 3, kernel_size=3, stride=4, padding=1, dilation=2),

Input shape: torch.Size([1, 3, 608, 608])

Output shape: torch.Size([1, 3, 152, 152])

BasicConv(3, 3, kernel_size=4, stride=4, padding=2, dilation=2),

Input shape: torch.Size([1, 3, 608, 608])

Output shape: torch.Size([1, 3, 152, 152])

BasicConv(3, 3, kernel_size=4, stride=4, padding=3, dilation=3),

Input shape: torch.Size([1, 3, 608, 608])

Output shape: torch.Size([1, 3, 152, 152])

BasicConv(3, 3, kernel_size=4, stride=4, padding=6, dilation=5),

Input shape: torch.Size([1, 3, 608, 608])

Output shape: torch.Size([1, 3, 152, 152])

总结

stride控制尺度变化的整体倍数

kernel_size控制卷积核的大小。主体

padding则能微弱调节经过多少次卷积。微调最终输出的尺寸

dilation怎能控制卷积核的感受野，配合padding操作，可以实现相同尺度下的不同感受野。