全局平均池化的示例

1.对一个3通道，5*5的矩阵，进行全局平均池化

每个矩阵的大小都是 5x5，假设这些矩阵代表一幅图像的三个不同通道。为简单起见，我们将这三个矩阵分别称为 A、B 和 C。合成图像将是一个三通道图像，每个通道由其中一个矩阵表示。

A = [[a11, a12, a13, a14, a15],
     [a21, a22, a23, a24, a25],
     [a31, a32, a33, a34, a35],
     [a41, a42, a43, a44, a45],
     [a51, a52, a53, a54, a55]]
 
B = [[b11, b12, b13, b14, b15],
     [b21, b22, b23, b24, b25],
     [b31, b32, b33, b34, b35],
     [b41, b42, b43, b44, b45],
     [b51, b52, b53, b54, b55]]
 
C = [[c11, c12, c13, c14, c15],
     [c21, c22, c23, c24, c25],
     [c31, c32, c33, c34, c35],
     [c41, c42, c43, c44, c45],
     [c51, c52, c53, c54, c55]]

现在，全局平均池操作将独立应用于每个通道。对于每个通道，它会计算该通道中所有元素的平均值。计算结果是一个向量，每个通道只有一个值。

avg_A = (a11 + a12 + ... + a55) / 25
avg_B = (b11 + b12 + ... + b55) / 25
avg_C = (c11 + c12 + ... + c55) / 25

这样，经过全局平均汇集后的合成图像将是一个 3 通道图像，每个通道由其原始矩阵的平均值表示： 

Composite Image = [[avg_A, avg_B, avg_C]]

2.torch示例

import torch
import torch.nn as nn
 
# Generate a random 3-channel matrix with integer values for a batch of size 2
torch.manual_seed(42)  # Setting seed for reproducibility
batch_size = 1
image_matrix = torch.randint(0, 10, (batch_size, 3, 5, 5), dtype=torch.float32)  # Batch size 2, 3 channels, 5x5 matrix
 
# Display the original matrix
print("Original Matrix:")
print(image_matrix)
 
# Apply global average pooling using nn.AdaptiveAvgPool2d
adaptive_avg_pool = nn.AdaptiveAvgPool2d(1)
global_avg_pooled = adaptive_avg_pool(image_matrix)
 
# Display the result after global average pooling
print("\nResult after Global Average Pooling:")
print(global_avg_pooled)

Original Matrix:
tensor([[[[2., 7., 6., 4., 6.],   
          [5., 0., 4., 0., 3.],   
          [8., 4., 0., 4., 1.],   
          [2., 5., 5., 7., 6.],   
          [9., 6., 3., 1., 9.]],  
 
         [[3., 1., 9., 7., 9.],   
          [2., 0., 5., 9., 3.],   
          [4., 9., 6., 2., 0.],   
          [6., 2., 7., 9., 7.],   
          [3., 3., 4., 3., 7.]],  
 
         [[0., 9., 0., 9., 6.],   
          [9., 5., 4., 8., 8.],   
          [6., 0., 0., 0., 0.],   
          [1., 3., 0., 1., 1.],   
          [7., 9., 4., 3., 8.]]]])
 
Result after Global Average Pooling:
tensor([[[[4.2800]],
 
         [[4.8000]],
 
         [[4.0400]]]])