【pytorch官方文档】：https://pytorch.org/docs/stable/generated/torch.nn.AvgPool2d.html?highlight=avgpool2d#torch.nn.AvgPool2d
torch.nn.AvgPool2d()
作用
在由多通道组成的输入特征中进行2D平均池化计算
函数
|
1
|
torch.nn.AvgPool2d(kernel_size, stride=None, padding=0, ceil_mode=False, count_include_pad=True, divisor_override=None) |
参数
Args:
kernel_size: 滑窗(池化核)大小
stride: 滑窗的移动步长, 默认值为kernel_size
padding: 在输入信号两侧的隐式零填充数量
ceil_mode: 决定计算输出的形状时是向上取整还是向下取整, 默认为False(向下取整)
count_include_pad: 在平均池化计算中是否包含零填充, 默认为True(包含零填充)
divisor_override: 如果指定了, 它将被作为平均池化计算中的除数, 否则将使用池化区域的大小作为平均池化计算的除数
公式
代码实例
假设输入特征为S，输出特征为D
情况一
ceil_mode=False, count_include_pad=True(计算时包含零填充)
|
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
|
import torchimport torch.nn as nnimport numpy as np # 生成一个形状为1*1*3*3的张量x1 = np.array([ [1,2,3], [4,5,6], [7,8,9] ])x1 = torch.from_numpy(x1).float()x1 = x1.unsqueeze(0).unsqueeze(0) # 实例化二维平均池化avgpool1 = nn.AvgPool2d(kernel_size=3, stride=2, padding=1, ceil_mode=False, count_include_pad=True)y1 = avgpool1(x1)print(y1) # 打印结果'''tensor([[[[1.3333, 1.7778], [2.6667, 3.1111]]]])''' |
计算过程：
输出形状= floor[(3 - 3 + 2) / 2] + 1 = 2,
D[1,1] = (0+0+0+0+1+2+0+4+5) / 9 = 1.3333,
D[1,2] = (0+0+0+2+3+0+5+6+0) / 9 = 1.7778,
D[2,1] = (0+4+5+0+7+8+0+0+0) / 9 = 2.6667,
D[2,2] = (5+6+0+8+9+0+0+0+0) / 9 = 3.1111.
情况二
ceil_mode=False, count_include_pad=False(计算时不包含零填充)
|
1
2
3
4
5
6
7
8
9
10
|
avgpool2 = nn.AvgPool2d(kernel_size=3, stride=2, padding=1, ceil_mode=False, count_include_pad=False) y2 = avgpool2(x1)print(y2) # 打印结果'''tensor([[[[3., 4.], [6., 7.]]]])''' |
计算过程：
输出形状= floor[(3 - 3 + 2) / 2] + 1 = 2,
D[1,1] = (1+2+4+5) / 4 = 3,
D[1,2] = (2+3+5+6) / 4 = 4,
D[2,1] = (4+5+7+8) / 4 = 6,
D[2,2] = (5+6+8+9) / 4 = 7.
情况三
ceil_mode=False, count_include_pad=False, divisor_override=2(将计算平均池化时的除数指定为2)
|
1
2
3
4
5
6
7
8
9
10
|
avgpool3 = nn.AvgPool2d(kernel_size=3, stride=2, padding=1, ceil_mode=False, count_include_pad=False, divisor_override=2) y3 = avgpool3(x1)print(y3) # 打印结果'''tensor([[[[ 6., 8.], [12., 14.]]]])''' |
计算过程：
输出形状= floor[(3 - 3 + 2) / 2] + 1 = 2,
D[1,1] = (1+2+4+5) / 2 = 6,
D[1,2] = (2+3+5+6) / 2 = 8,
D[2,1] = (4+5+7+8) / 2 = 12,
D[2,2] = (5+6+8+9) / 2 = 14.
情况四
ceil_mode=True, count_include_pad=True, divisor_override=None(在计算输出的形状时向上取整)
|
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
|
x2 = np.array([ [1,2,3,4], [5,6,7,8], [9,10,11,12], [13,14,15,16] ])x2 = torch.from_numpy(x2).reshape(1,1,4,4).float()avgpool4 = nn.AvgPool2d(kernel_size=3, stride=2, padding=1, ceil_mode=True)y4 = avgpool4(x2)print(y4) # 打印结果'''tensor([[[[ 1.5556, 3.3333, 2.0000], [ 6.3333, 11.0000, 6.0000], [ 4.5000, 7.5000, 4.0000]]]])''' |
计算过程：
输出形状 = ceil[(4 - 3 + 2) / 2] + 1 = 3,
D[1,1] = (0+0+0+0+1+2+0+5+6) / 9 = 1.5556,
D[1,2] = (0+0+0+2+3+4+6+7+8) / 9 = 3.3333,
D[1,3] = (0+0+4+0+8+0) / 6 = 2,
D[2,1] = (0+5+6+0+9+10+0+13+14) / 9 = 6.3333,
D[2,2] = (6+7+8+10+11+12+14+15+16) / 9 = 11,
D[2,3] = (8+0+12+0+16+0) / 6 = 6,
D[3,1] = (0+13+14+0+0+0) / 6 = 4.5,
D[3,2] = (14+15+16+0+0+0) / 6 = 7.5,
D[3,3] = (16+0+0+0) / 4 = 4.
总结
到此这篇关于Pytorch平均池化nn.AvgPool2d()使用的文章就介绍到这了,更多相关Pytorch平均池化nn.AvgPool2d()使用内容请搜索服务器之家以前的文章或继续浏览下面的相关文章希望大家以后多多支持服务器之家！
原文链接：https://blog.csdn.net/qq_38964360/article/details/129148451
