Neural networks - use of maximum pooling

Official documentation of the pooling layer Many pooling methods are introduced in , But the most common one is MaxPool2d, Here we also use MaxPool2d To explain , Other similar , The key is to learn to watch Official documents
summary ：
The purpose of maximum pooling is to Keep the input characteristics , But at the same time, reduce the amount of data , After maximum pooling, the amount of data is reduced , For the whole network , There are fewer parameters to calculate , Will train faster .
It is equivalent to watching videos on the Internet , Video again 1080P Of ,720P Of ,360P Of , You know what ,1080P It is equivalent to inputting video ,720P Is equivalent to the video after maximum pooling ,720P It can also meet the needs , When the net doesn't work, you can watch it 720P Yeah .

What is maximum pooling ？
The maximum pool operation is equivalent to the core moving on the image , Filter out the maximum value of the area covered by the core , Be careful The moving step of the kernel is kernel_size.

Speaking of MaxPool2d The interface is used here , The key is to learn how to communicate .

torch.nn.MaxPool2d(kernel_size, stride=None, padding=0, dilation=1, return_indices=False, ceil_mode=False)

Parameter description

• kernel_size – the size of the window to take a max over
• stride – the stride of the window. Default value is kernel_size
• padding – implicit zero padding to be added on both sides
• dilation – a parameter that controls the stride of elements in the window
• return_indices – if True, will return the max indices along with the outputs. Useful for torch.nn.MaxUnpool2d later
• ceil_mode – when True, will use ceil instead of floor to compute the output shape

Pick out some common ones and tell you ：

• kernel_size When you pass in a constant , A size of kernel_size X kernel_size The size of the nuclear ,kernel_size When a tuple is passed in , Then a core with specified height and width will be generated .
• Speaking of Ceil_model Parameters , In some cases , The core cannot be completely covered on the image , Only part of the image is covered , The other part has gone outside the image ,Ceil_model It was decided at this time , Do you want to perform a maximum pool operation on the block that only covers the part , Take the maximum , if Ceil_model by True, Then the maximum pool operation is performed , Take the maximum , Otherwise, the maximum pool operation will not be performed , Do not take the maximum value . You can see the figure below Ceil_model Parameters ：

The process ：

result ：

MaxPool2d The sample code used is as follows ：

import torch
import torchvision
from torch import nn
from torch.nn import MaxPool2d
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter
dataset = torchvision.datasets.CIFAR10("CIFAR10",train=False,transform=torchvision.transforms.ToTensor(),download=True)
#  Be careful dataset in transform The parameter receives an object , So we need to add parentheses , In addition, when using neural network for operation, the data type required is tensor type , therefore transforms Parameters to add .
dataloader = DataLoader(dataset,batch_size=64)

#  Building neural networks 
class Booze(nn.Module):

    #  Inherit nn.Module The initialization 
    def __init__(self):
        super(Booze, self).__init__()
        self.maxpool1 = MaxPool2d(kernel_size=3,ceil_mode=True)

    #  rewrite forward function 
    def forward(self,x):
        output = self.maxpool1(x)
        return output


obj = Booze()

#  Use tensorboard Visualization process 
writer = SummaryWriter("logs")
step = 0
for data in dataloader:
    imgs,targets = data
    writer.add_images("input",imgs,step)
    output = obj(imgs)
    writer.add_images("output",output,step)
    step+=1

writer.close()

The above code is finished , stay pycharm Below terminal Window type tensorboard --logdir=logs Enter to check ：

The effect is as follows ：

It can be seen from the results above , After the maximum pool processing, the part is obviously more blurred than the original image , But when compared with the original image, at least the original features can be seen .