Pytorch softmax:使用什么维度?

Pytorch softmax: What dimension to use?

函数torch.nn.functional.softmax有两个参数:inputdim。根据其文档,softmax 操作应用于 input 沿指定 dim 的所有切片,并将重新缩放它们,以便元素位于 (0, 1) 范围内并且总和为 1。

设输入为:

input = torch.randn((3, 4, 5, 6))

假设我想要以下内容,以便该数组中的每个条目都是 1:

sum = torch.sum(input, dim = 3) # sum's size is (3, 4, 5, 1)

我应该如何应用softmax?

softmax(input, dim = 0) # Way Number 0
softmax(input, dim = 1) # Way Number 1
softmax(input, dim = 2) # Way Number 2
softmax(input, dim = 3) # Way Number 3

我的直觉告诉我那是最后一个,但我不确定。英语不是我的第一语言,因此 along 这个词的使用让我感到困惑。

我不是很清楚 "along" 是什么意思,所以我将使用一个可以澄清事情的例子。假设我们有一个大小为 (s1, s2, s3, s4) 的张量,我希望它发生

让我们考虑二维的例子

x = [[1,2],
    [3,4]]

你希望你的最终结果是

y = [[0.27,0.73],
    [0.27,0.73]]

y = [[0.12,0.12],
    [0.88,0.88]]

如果是第一个选项,那么您需要 dim = 1。如果是第二个选项,您需要 dim = 0。

请注意,在第二个示例中对列或第零维进行了规范化,因此它沿第零维进行了规范化。

2018-07-10 更新:反映第零维指的是 pytorch 中的列。

我能想到的让您理解的最简单方法是:假设给定了一个形状为 (s1, s2, s3, s4) 的张量,并且正如您提到的,您希望获得沿最后一个轴的所有条目的总和是 1.

sum = torch.sum(input, dim = 3) # input is of shape (s1, s2, s3, s4)

那么你应该将 softmax 称为:

softmax(input, dim = 3)

为了便于理解,您可以将形状为 (s1, s2, s3, s4) 的 4d 张量视为形状为 (s1*s2*s3, s4) 的 2d 张量或矩阵。现在,如果您希望矩阵在每行 (axis=0) 或列 (axis=1) 中包含总和为 1 的值,那么,您可以简单地在 2d 张量上调用 softmax 函数,如下所示:

softmax(input, dim = 0) # normalizes values along axis 0
softmax(input, dim = 1) # normalizes values along axis 1

你可以看看史蒂文在他的中提到的例子。

不正确。请参阅下面的快照。其实是相反的方式。

图片转录为代码:

>>> x = torch.tensor([[1,2],[3,4]],dtype=torch.float)
>>> F.softmax(x,dim=0)
tensor([[0.1192, 0.1192],
        [0.8808, 0.8808]])
>>> F.softmax(x,dim=1)
tensor([[0.2689, 0.7311],
        [0.2689, 0.7311]])

我不是 100% 确定你的问题是什么意思,但我认为你的困惑只是因为你不明白 dim 参数的意思。因此,我将对其进行解释并提供示例。

如果我们有:

m0 = nn.Softmax(dim=0)

这意味着 m0 将沿它接收的张量的第零坐标对元素进行归一化。正式地,如果给定一个大小为 (d0,d1) 的张量 b,则以下内容为真:

sum^{d0}_{i0=1} b[i0,i1] = 1, forall i1 \in {0,...,d1}

您可以使用 Pytorch 示例轻松检查:

>>> b = torch.arange(0,4,1.0).view(-1,2)
>>> b 
tensor([[0., 1.],
        [2., 3.]])
>>> m0 = nn.Softmax(dim=0) 
>>> b0 = m0(b)
>>> b0 
tensor([[0.1192, 0.1192],
        [0.8808, 0.8808]])

既然 dim=0 意味着遍历 i0 \in {0,1}(即遍历行),如果我们选择任何列 i1 并对它的元素(即行)求和,那么我们应该得到1.检查一下:

>>> b0[:,0].sum()
tensor(1.0000)
>>> b0[:,1].sum()
tensor(1.0000)

符合预期。

请注意,我们确实通过 "summing out the rows" 和 torch.sum(b0,dim=0) 使所有行总和为 1,检查一下:

>>> torch.sum(b0,0)
tensor([1.0000, 1.0000])

我们可以创建一个更复杂的例子来确保它真的很清楚。

a = torch.arange(0,24,1.0).view(-1,3,4)
>>> a
tensor([[[ 0.,  1.,  2.,  3.],
         [ 4.,  5.,  6.,  7.],
         [ 8.,  9., 10., 11.]],

        [[12., 13., 14., 15.],
         [16., 17., 18., 19.],
         [20., 21., 22., 23.]]])
>>> a0 = m0(a)
>>> a0[:,0,0].sum()
tensor(1.0000)
>>> a0[:,1,0].sum()
tensor(1.0000)
>>> a0[:,2,0].sum()
tensor(1.0000)
>>> a0[:,1,0].sum()
tensor(1.0000)
>>> a0[:,1,1].sum()
tensor(1.0000)
>>> a0[:,2,3].sum()
tensor(1.0000)

所以正如我们预期的那样,如果我们将沿第一个坐标从第一个值到最后一个值的所有元素相加,我们得到 1。因此所有内容都沿第一个维度(或第一个坐标 i0)归一化。

>>> torch.sum(a0,0)
tensor([[1.0000, 1.0000, 1.0000, 1.0000],
        [1.0000, 1.0000, 1.0000, 1.0000],
        [1.0000, 1.0000, 1.0000, 1.0000]])

此外,沿维度 0 表示您沿该维度改变坐标并考虑每个元素。有点像让 for 循环遍历第一个坐标可以采用的值,即

for i0 in range(0,d0):
    a[i0,b,c,d]
import torch
import torch.nn.functional as F

x = torch.tensor([[1, 2], [3, 4]], dtype=torch.float)

s1 = F.softmax(x, dim=0)
tensor([[0.1192, 0.1192],
    [0.8808, 0.8808]])

s2 = F.softmax(x, dim=1)
tensor([[0.2689, 0.7311],
    [0.2689, 0.7311]])

torch.sum(s1, dim=0)
tensor([1., 1.])

torch.sum(s2, dim=1)
tensor([1., 1.])

想一想 softmax 试图实现的目标。它输出一种结果相对于另一种结果的概率。假设您正在尝试预测两个结果:是 A 还是 B。如果 p(A) 大于 p(B),那么下一步是将结果转换为布尔值(即如果 p,则结果将是 A (A) > 50% or B if p(B) > 50% 因为我们处理的是概率,它们应该加起来为 1。 因此,您想要的是每行的总和概率为 1。因此,您指定 dim=1 或 row sum

另一方面,如果您的模型旨在预测两个以上的变量,则输出张量将类似于 [p(a), p(b), p(c)...p(i)] 这里重要的是 p(a) + p(b) + p(c) +...p(i) = 1 那么你会使用 dim = 0

这完全取决于您如何定义输出层。