PyTorch torch.nn.AdaptiveMaxPool2d 函数
PyTorch torch.nn 参考手册torch.nn.AdaptiveMaxPool2d 是 PyTorch 中的自适应最大池化模块。
它将输入池化到指定尺寸，保留最大值而非平均值。
函数定义
torch.nn.AdaptiveMaxPool2d(output_size, return_indices=False)
参数
output_size: 输出尺寸return_indices: 是否返回索引
使用示例
示例 1: 基本用法
实例
import torch
import torch.nn as nn
# 全局最大池化
gap = nn.AdaptiveMaxPool2d(1)
# 输入不同尺寸
x1 = torch.randn(1, 64, 32, 32)
x2 = torch.randn(1, 64, 16, 16)
print("32x32 ->", gap(x1).shape)
print("16x16 ->", gap(x2).shape)
import torch.nn as nn
# 全局最大池化
gap = nn.AdaptiveMaxPool2d(1)
# 输入不同尺寸
x1 = torch.randn(1, 64, 32, 32)
x2 = torch.randn(1, 64, 16, 16)
print("32x32 ->", gap(x1).shape)
print("16x16 ->", gap(x2).shape)
示例 2: 返回索引
实例
import torch
import torch.nn as nn
gap = nn.AdaptiveMaxPool2d(1, return_indices=True)
x = torch.randn(1, 64, 8, 8)
output, indices = gap(x)
print("输出形状:", output.shape)
print("索引形状:", indices.shape)
print("索引值:", indices.item())
import torch.nn as nn
gap = nn.AdaptiveMaxPool2d(1, return_indices=True)
x = torch.randn(1, 64, 8, 8)
output, indices = gap(x)
print("输出形状:", output.shape)
print("索引形状:", indices.shape)
print("索引值:", indices.item())
示例 3: 与 AdaptiveAvgPool2d 对比
实例
import torch
import torch.nn as nn
x = torch.tensor([[[
[1, 2, 3, 4],
[5, 6, 7, 8],
[9, 10, 11, 12],
[13, 14, 15, 16]
]]], dtype=torch.float32)
avgpool = nn.AdaptiveAvgPool2d(1)
maxpool = nn.AdaptiveMaxPool2d(1)
print("输入:n", x[0, 0])
print("平均池化:", avgpool(x).item())
print("最大池化:", maxpool(x).item())
import torch.nn as nn
x = torch.tensor([[[
[1, 2, 3, 4],
[5, 6, 7, 8],
[9, 10, 11, 12],
[13, 14, 15, 16]
]]], dtype=torch.float32)
avgpool = nn.AdaptiveAvgPool2d(1)
maxpool = nn.AdaptiveMaxPool2d(1)
print("输入:n", x[0, 0])
print("平均池化:", avgpool(x).item())
print("最大池化:", maxpool(x).item())
使用场景
- 全局最大池化: 提取最显著特征
- 特征汇聚: 保留关键信息
- 分类网络: 替代 FC 层
提示：最大池化保留最显著特征，平均池化更平滑。