torch.mean()、torch.max() 和 nn.AdaptiveAvgPool2d()、nn.AdaptiveMaxPool2d() 用法小结
Author ：Horizon Max
✨ 编程技巧篇：各种操作小结
机器视觉篇：会变魔术 OpenCV
深度学习篇：简单入门 PyTorch
神经网络篇：经典网络模型
算法篇：再忙也别忘了 LeetCode
文章目录
- torch.randn()
- torch.mean()
- torch.max()
- nn.AdaptiveAvgPool2d()
- nn.AdaptiveMaxPool2d()
torch.randn()
首先创建1个 2个通道（3 x 5）大小的张量 ：
x = torch.randn(1, 2, 3, 5)
print(x)
输出结果：
tensor([[[[ 0.9529, -0.3196, 0.3037, 0.0507, 0.3179],
[-2.0785, -0.8697, 0.2880, 0.4416, 0.7148],
[-1.7157, 0.6767, -0.5997, 1.1221, 0.2088]],
[[-1.1475, 0.2046, -0.7609, -1.2099, -0.5255],
[-0.1989, -0.6862, -0.6264, 0.1123, 0.3257],
[ 1.5529, -0.6140, -0.0471, -0.6173, -0.9239]]]])
torch.mean()
dim = 0 ：按行求平均值，返回结果为（1，列），通道数不会发生变化
dim = 0 ：按列求平均值，返回结果为（行，1），通道数会发生变化
默认 ：输出所有值的平均值
avg_out = torch.mean(x)
print(avg_out)
avg_out0 = torch.mean(x, dim=0, keepdim=True)
print(avg_out0)
avg_out1 = torch.mean(x, dim=1, keepdim=True)
print(avg_out1)
输出结果：
tensor(-0.1889)
tensor([[[[ 0.9529, -0.3196, 0.3037, 0.0507, 0.3179],
[-2.0785, -0.8697, 0.2880, 0.4416, 0.7148],
[-1.7157, 0.6767, -0.5997, 1.1221, 0.2088]],
[[-1.1475, 0.2046, -0.7609, -1.2099, -0.5255],
[-0.1989, -0.6862, -0.6264, 0.1123, 0.3257],
[ 1.5529, -0.6140, -0.0471, -0.6173, -0.9239]]]])
tensor([[[[-0.0973, -0.0575, -0.2286, -0.5796, -0.1038],
[-1.1387, -0.7779, -0.1692, 0.2770, 0.5203],
[-0.0814, 0.0313, -0.3234, 0.2524, -0.3575]]]])
torch.max()
dim = 0 ：按行求最大值，返回结果为（1，列），通道数不会发生变化
dim = 0 ：按列求最大值，返回结果为（行，1），通道数会发生变化
默认 ：输出所有值的最大值
max_out = torch.max(x)
print(max_out)
max_out0, _ = torch.max(x, dim=0, keepdim=True)
print(max_out0)
max_out1, _ = torch.max(x, dim=1, keepdim=True)
print(max_out1)
输出结果：
tensor(1.5529)
tensor([[[[ 0.9529, -0.3196, 0.3037, 0.0507, 0.3179],
[-2.0785, -0.8697, 0.2880, 0.4416, 0.7148],
[-1.7157, 0.6767, -0.5997, 1.1221, 0.2088]],
[[-1.1475, 0.2046, -0.7609, -1.2099, -0.5255],
[-0.1989, -0.6862, -0.6264, 0.1123, 0.3257],
[ 1.5529, -0.6140, -0.0471, -0.6173, -0.9239]]]])
tensor([[[[ 0.9529, 0.2046, 0.3037, 0.0507, 0.3179],
[-0.1989, -0.6862, 0.2880, 0.4416, 0.7148],
[ 1.5529, 0.6767, -0.0471, 1.1221, 0.2088]]]])
另外：
out = torch.max(x, dim=0, keepdim=True)
print(out)
输出结果：
torch.return_types.max(
values=tensor([[[[ 0.9529, -0.3196, 0.3037, 0.0507, 0.3179],
[-2.0785, -0.8697, 0.2880, 0.4416, 0.7148],
[-1.7157, 0.6767, -0.5997, 1.1221, 0.2088]],
[[-1.1475, 0.2046, -0.7609, -1.2099, -0.5255],
[-0.1989, -0.6862, -0.6264, 0.1123, 0.3257],
[ 1.5529, -0.6140, -0.0471, -0.6173, -0.9239]]]]),
indices=tensor([[[[0, 0, 0, 0, 0],
[0, 0, 0, 0, 0],
[0, 0, 0, 0, 0]],
[[0, 0, 0, 0, 0],
[0, 0, 0, 0, 0],
[0, 0, 0, 0, 0]]]]))
nn.AdaptiveAvgPool2d()
给定输出 feature map 的大小，输出前后的 通道数不会发生变化 ：
avg_pool0 = nn.AdaptiveAvgPool2d(1)
avg_pool0 = avg_pool0(x)
print(avg_pool0)
avg_pool1 = nn.AdaptiveAvgPool2d((2, 2))
avg_pool1 = avg_pool1(x)
print(avg_pool1)
输出结果：
tensor([[[[-0.0337]],
[[-0.3442]]]])
tensor([[[[-0.2872, 0.3528],
[-0.7165, 0.3626]],
[[-0.5359, -0.4475],
[-0.1033, -0.2961]]]])
nn.AdaptiveMaxPool2d()
给定输出 feature map 的大小，输出前后的 通道数不会发生变化 ：
max_pool0 = nn.AdaptiveMaxPool2d(1)
max_pool0 = max_pool0(x)
print(max_pool0)
max_pool1 = nn.AdaptiveMaxPool2d((2, 2))
max_pool1 = max_pool1(x)
print(max_pool1)
输出结果：
tensor([[[[1.1221]],
[[1.5529]]]])
tensor([[[[0.9529, 0.7148],
[0.6767, 1.1221]],
[[0.2046, 0.3257],
[1.5529, 0.3257]]]])