Pytorch 实现下采样的方法（卷积和池化）

最新推荐文章于 2024-08-20 20:55:16 发布

whetherfailbuttry

最新推荐文章于 2024-08-20 20:55:16 发布

阅读量1.2w

点赞数 1

文章标签： python pytorch

本文链接：https://blog.csdn.net/qq_40602790/article/details/123692342

版权

本文详细探讨了卷积下采样（通过Max Pooling和卷积层）在深度学习中的应用，以及如何通过不同卷积核大小配合padding保持图像尺寸。通过实例代码展示了3x3、5x5和7x7卷积核的使用，并解释了卷积参数计算方法。

摘要由CSDN通过智能技术生成

# 卷积核大小和下采样实现

标签（空格分隔）：深度学习
https://www.zhihu.com/question/307839910(卷积下采样和池化下采样的区别)
---

### 1.下采样的实现：maxpooling 或者 conv,k=2,s=2 实现下采样两倍(池化下采样比较粗暴，可能将有用的信息滤除掉，而卷积下采样过程控制了步进大小，信息融合较好)
```
self.conv_downsampling = nn.Conv2d(3,3,kernel_size=2,stride=2)
self.max_pooling = nn.MaxPool2d(kernel_size=2)
输出：torch.Size([1, 3, 128, 128]) torch.Size([1, 3, 128, 128])

```

### 2.卷积核大小为3，5，7时，padding为1，2，3 时，图像卷积后大小不变(Pytorch默认conv的s=1，p=0)
```
self.conv33 = nn.Conv2d(3,3,kernel_size=3,padding=1)
self.conv55 = nn.Conv2d(3,3,kernel_size=5,padding=2)
self.conv77 = nn.Conv2d(3,3,kernel_size=7,padding=3)
输出：torch.Size([1, 3, 256, 256]) torch.Size([1, 3, 256, 256]) torch.Size([1, 3, 256, 256])

```
### 3.卷积核参数计算方法：卷积核大小(n*d*w) * inchanels * outchannels

## Code
```
import torch
from torch import nn

class test(nn.Module):
def __init__(self):
super(test, self).__init__()
self.conv_downsampling = nn.Conv2d(3,3,kernel_size=2,stride=2)
self.max_pooling = nn.MaxPool2d(kernel_size=2)
self.conv33 = nn.Conv2d(3,3,kernel_size=3,padding=1)
self.conv55 = nn.Conv2d(3,3,kernel_size=5,padding=2)
self.conv77 = nn.Conv2d(3,3,kernel_size=7,padding=3)
def forward(self,x):
out1 = self.conv_downsampling(x)
out2 = self.max_pooling(x)
out3 = self.conv33(x)
out4 = self.conv55(x)
out5 = self.conv77(x)
print(out1.shape,out2.shape,\
out3.shape,out4.shape,\
out5.shape)

if __name__ == '__main__':
data = torch.randn(1,3,256,256)
test_func = test()
test_func(data)

```