汇聚层

中文版 mxnet
1

https://zh.d2l.ai/chapter_convolutional-neural-networks/pooling.html

2 
import torch
from torch import nn

conv = nn.Conv2d(1, 1, kernel_size=(2, 2), bias=False)
pool = nn.AvgPool2d((2, 2), stride=1)

X = torch.rand((6, 6)).reshape((1, 1, 6, 6))
h_k, w_k = conv.kernel_size
conv.weight.data[:] = (1 / (h_k * w_k))

print(conv(X) == pool(X))

tensor([[[[True, True, True, True, True],
          [True, True, True, True, True],
          [True, True, True, True, True],
          [True, True, True, True, True],
          [True, True, True, True, True]]]])

平均pooling实际上就是conv kernel的权重为1/(h_p * w_p)

3

本节将介绍汇聚层，它具有双重目的：降低卷积层对位置的敏感性，同时降低对空间降采样表示的敏感性。

这里应该翻译错了，原文是：

This section introduces pooling layers, which serve the dual purposes of mitigating the sensitivity of convolutional layers to location and of spatially downsampling representations.

应该是 serve the dual purposes of A and of B
即

本节将介绍汇聚层，它具有双重目的：降低卷积层对位置的敏感性，以及在空间维度上对表示进行降采样。