30- padding 研究 (TensorFlow系列) (深度学习)-优快云博客

本文链接：https://blog.youkuaiyun.com/March_A/article/details/129346810

该文探讨了在TensorFlow的Keras库中，MaxPooling层使用valid和same两种填充方式时，如何改变输入图像尺寸。当padding=valid，图像尺寸会按池化窗口大小减小；而padding=same时，图像尺寸保持不变，边缘通过填充来保持尺寸。

摘要生成于 C知道，由 DeepSeek-R1 满血版支持，前往体验 >

知识要点

padding= valid和same, 图片大小的变化.
卷积图片尺寸的变换.

一 padding 研究

from tensorflow import keras
import tensorflow as tf
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt

1.1 padding = valid

$N = \frac{W - F + 1}{S}$ (向上取整)

model = keras.Sequential([keras.layers.Input(shape = (7, 7, 3), dtype = 'float32'),
                          keras.layers.MaxPool2D(pool_size = 3, strides = 2, padding = 'valid')])

model.summary()

model = keras.Sequential([keras.layers.Input(shape = (8, 8, 3), dtype = 'float32'),
                          keras.layers.MaxPool2D(pool_size = 4, strides = 2, padding = 'same')])

model.summary()

1.2 padding = same

$N = \frac{W}{S}$ (向上取整)

model = keras.Sequential([keras.layers.Input(shape = (8, 8, 3), dtype = 'float32'),
                          keras.layers.MaxPool2D(pool_size = 3, strides = 2, padding = 'valid')])

model.summary()

model = keras.Sequential([keras.layers.Input(shape = (8, 8, 3), dtype = 'float32'),
                          keras.layers.MaxPool2D(pool_size = 3, strides = 2, padding = 'same')])
model.summary()

model = keras.Sequential([keras.layers.Input(shape = (8, 8, 3), dtype = 'float32'),
                          keras.layers.MaxPool2D(pool_size = 3, strides = 2, padding = 'SAME')])

model.summary()