PyTorch：卷积/padding/pooling api

﹏ヽ暗。殇╰゛Y 2022-10-27 06:24 177阅读 0赞

## 填充padding ##

torch.nn.[ConstantPad2d][](padding: Union\[T, Tuple\[T, T, T, T\]\], value: float)

参数：padding (int, tuple) – the size of the padding. If is int, uses the same padding in all boundaries. If a 4-tuple, uses padding\_left , padding\_right , padding\_top , padding\_bottom )

![20210203214631315.png][]

示例：

>>> m = nn.ConstantPad2d(2, 3.5)  
>>> input = torch.randn(1, 2, 2)  
>>> input  
tensor(\[\[\[ 1.6585,  0.4320\],  
         \[-0.8701, -0.4649\]\]\])  
>>> m(input)  
tensor(\[\[\[ 3.5000,  3.5000,  3.5000,  3.5000,  3.5000,  3.5000\],  
         \[ 3.5000,  3.5000,  3.5000,  3.5000,  3.5000,  3.5000\],  
         \[ 3.5000,  3.5000,  1.6585,  0.4320,  3.5000,  3.5000\],  
         \[ 3.5000,  3.5000, -0.8701, -0.4649,  3.5000,  3.5000\],  
         \[ 3.5000,  3.5000,  3.5000,  3.5000,  3.5000,  3.5000\],  
         \[ 3.5000,  3.5000,  3.5000,  3.5000,  3.5000,  3.5000\]\]\])  
>>> \# using different paddings for different sides  
>>> m = nn.ConstantPad2d((3, 0, 2, 1), 3.5)  
>>> m(input)  
tensor(\[\[\[ 3.5000,  3.5000,  3.5000,  3.5000,  3.5000\],  
         \[ 3.5000,  3.5000,  3.5000,  3.5000,  3.5000\],  
         \[ 3.5000,  3.5000,  3.5000,  1.6585,  0.4320\],  
         \[ 3.5000,  3.5000,  3.5000, -0.8701, -0.4649\],  
         \[ 3.5000,  3.5000,  3.5000,  3.5000,  3.5000\]\]\])

## 二维巻积CONV2D ##

torch.nn.[Conv2d][](in\_channels: int, out\_channels: int, kernel\_size: Union\[T, Tuple\[T, T\]\], stride: Union\[T, Tuple\[T, T\]\] = 1, padding: Union\[T, Tuple\[T, T\]\] = 0, dilation: Union\[T, Tuple\[T, T\]\] = 1, groups: int = 1, bias: bool = True, padding\_mode: str = 'zeros')

参数：in\_channels：输入通道数。一般第1层是1个通道，如果是多层cnn堆叠，从第二层起通道数为上层的输出feature maps数=即上一层的out\_channels。

Note:每层参数量计算。第l层参数量：C\_in\*C\_out\*kernel\_size\[l\]\[0\]\*kernel\_size\[l\]\[1\]，其中C\_in即in\_channels\[l\]。

![20210203214908314.png][]

示例：

>>> \# With square kernels and equal stride  
>>> m = nn.Conv2d(16, 33, 3, stride=2)

Conv2d(1, 8, kernel\_size=\[3, 3\], stride=(1, 1))

Conv2d(8, 16, kernel\_size=\[3, 3\], stride=(1, 1))

## 自适应池化Adaptive Pooling ##

torch.nn.[AdaptiveAvgPool2d][](output\_size: Union\[T, Tuple\[T, ...\]\])

普通Max/AvgPooling计算公式为：output\_size = ceil ( (input\_size+2∗padding−kernel\_size)/stride）+1

当我们使用Adaptive Pooling时，这个问题就变成了由已知量input\_size，output\_size求解kernel\_size与stride。你只需要告诉torch你需要什么样的输出结果。

为了简化问题，我们将padding设为0（后面我们可以发现源码里也是这样操作的[c++源码部分][c]）

stride = floor ( (input\_size / (output\_size) )

kernel\_size = input\_size − (output\_size−1) \* stride

padding = 0

示例：

>>> \# target output size of 5x7  
>>> m = nn.AdaptiveAvgPool2d((5,7))  
>>> input = torch.randn(1, 64, 8, 9)  
>>> output = m(input)

>>> \# target output size of 7x7 (square)  
>>> m = nn.AdaptiveAvgPool2d(7)  
>>> input = torch.randn(1, 64, 10, 9)  
>>> output = m(input)

\[[\[开发技巧\]·AdaptivePooling与Max/AvgPooling相互转换][AdaptivePooling_Max_AvgPooling]\]

## 一维巻积CONV1D ##

torch.nn.[Conv1d][](in\_channels, out\_channels, kernel\_size, stride=1, padding=0, dilation=1, groups=1, bias=True, padding\_mode='zeros')

in\_channels(int) – 输入信号的通道。在文本分类中，即为词向量的维度  
out\_channels(int) – 卷积产生的通道。有多少个out\_channels，就需要多少个1维卷积  
kernel\_size(int or tuple) - 卷积核的尺寸，卷积核的大小为(k,)，第二个维度是由in\_channels来决定的，所以实际上卷积大小为kernel\_size\*in\_channels  
stride(int or tuple, optional) - 卷积步长  
padding (int or tuple, optional)- 输入的每一条边补充0的层数  
dilation(int or tuple, \`optional\`\`) – 卷积核元素之间的间距  
groups(int, optional) – 从输入通道到输出通道的阻塞连接数  
bias(bool, optional) - 如果bias=True，添加偏置

### 示例 ###

常用于textcnn\[[深度学习：文本CNN-textcnn][CNN-textcnn]\]

conv1 = nn.Conv1d(in\_channels=256，out\_channels=100,kernel\_size=2)  
input = torch.randn(32,35,256)  
\# batch\_size x text\_len x embedding\_size -> batch\_size x embedding\_size x text\_len  
input = input.permute(0,2,1)  
out = conv1(input)  
print(out.size())

这里32为batch\_size，35为句子最大长度，256为词向量。在输入一维卷积的时候，需要将32\*35\*256变换为32\*256\*35，因为一维卷积是在最后维度上扫的，最后out的大小即为：32\*100\*（35-2+1）=32\*100\*34。

\[[pytorch之nn.Conv1d详解][pytorch_nn.Conv1d]\]\[\]

from: [\-柚子皮-][-_-]

ref:

[ConstantPad2d]: https://pytorch.org/docs/stable/generated/torch.nn.ConstantPad2d.html
[20210203214631315.png]: /images/20221024/a0b23ecb77c34156b67ce9fa51396f7f.png
[Conv2d]: https://pytorch.org/docs/stable/generated/torch.nn.Conv2d.html
[20210203214908314.png]: /images/20221024/ce67e659ee7a416e83595ab21f2b8cfa.png
[AdaptiveAvgPool2d]: https://pytorch.org/docs/stable/generated/torch.nn.AdaptiveAvgPool2d.html
[c]: https://github.com/pytorch/pytorch/blob/65b00aa5972e23b2a70aa60dec5125671a3d7153/aten/src/ATen/native/AdaptiveAveragePooling.cpp
[AdaptivePooling_Max_AvgPooling]: https://blog.csdn.net/xiaosongshine/article/details/89453037
[Conv1d]: https://pytorch.org/docs/stable/generated/torch.nn.Conv1d.html
[CNN-textcnn]: https://blog.csdn.net/pipisorry/article/details/85076712?ops_request_misc=%257B%2522request%255Fid%2522%253A%2522161581419416780262535475%2522%252C%2522scm%2522%253A%252220140713.130102334.pc%255Fblog.%2522%257D&request_id=161581419416780262535475&biz_id=0&utm_medium=distribute.pc_search_result.none-task-blog-2~blog~first_rank_v2~rank_v29-1-85076712.pc_v2_rank_blog_default&utm_term=textcnn
[pytorch_nn.Conv1d]: https://blog.csdn.net/sunny_xsc1994/article/details/82969867
[-_-]: https://me.csdn.net/pipisorry