import torch
x=torch.randn(3,3,4)
print(x)
print(x.argmax(dim=0))
输出:tensor([[[ 0.5128, 0.3717, 0.3606, -0.0286],
[ 0.0933, -1.4781, -0.3561, -0.2652],
[-0.8861, 0.6988, 1.1243, -1.1301]],
[[-0.0246, 0.0917, -0.0623, -1.4874],
[-2.0169, -0.8390, 0.1292, -0.1190],
[-0.4090, -0.9277, -0.9193, 2.1597]],
[[ 0.7030, -1.3322, -2.0856, 0.3122],
[ 1.4656, -1.1145, 0.8208, -1.8250],
[-0.8721, 0.4831, 1.5668, 0.1656]]])
tensor([[2, 0, 0, 2],
[2, 1, 2, 1],
[1, 0, 2, 1]])
(1)dim=0,比较的是torch.randn(3,3,4)中第一个第一个3所在维度的比较,也就是:
[ 0.5128, 0.3717, 0.3606, -0.0286],
[ 0.0933, -1.4781, -0.3561, -0.2652], 下标为0
[-0.8861, 0.6988, 1.1243, -1.1301]
[-0.0246, 0.0917, -0.0623, -1.4874],
[-2.0169, -0.8390, 0.1292, -0.1190], 下标为1
[-0.4090, -0.9277, -0.9193, 2.1597]]
[ 0.7030, -1.3322, -2.0856, 0.3122],
[ 1.4656, -1.1145, 0.8208, -1.8250], 下标为2
[-0.8721, 0.4831, 1.5668, 0.1656]]
三个矩阵(下标从0开始,)对应元素的比较,
所以print(x.argmax(dim=0)),就是0.5128和-0.0246和0.7030比较,输出最大值所在的下标(0.7030所在的张量下标为2,所以输出为2);同理0.3717和0.0917和-1.3322进行比较,输出最大值的下标0(最大值为0.3717,下标为0);以此类推。
所以dim=0 的输出为:
tensor([[2, 0, 0, 2],
[2, 1, 2, 1],
[1, 0, 2, 1]]) 。所以看到输出你就很容易理解那个矩阵对应元素的值最大。
(2)dim=1,及(3,3,4)对第二个维度的元素进行比较(第二个3)。
import torch
x=[[[ 0.5128, 0.3717, 0.3606, -0.0286],
[ 0.0933, -1.4781, -0.3561, -0.2652],
[-0.8861, 0.6988, 1.1243, -1.1301]],
[[-0.0246, 0.0917, -0.0623, -1.4874],
[-2.0169, -0.8390, 0.1292, -0.1190],
[-0.4090, -0.9277, -0.9193, 2.1597]],
[[ 0.7030, -1.3322, -2.0856, 0.3122],
[ 1.4656, -1.1145, 0.8208, -1.8250],
[-0.8721, 0.4831, 1.5668, 0.1656]]]
x=torch.tensor(x)
print(x.argmax(dim=1))
输出:
tensor([[0, 2, 2, 0],
[0, 0, 1, 2],
[1, 2, 2, 0]])
[ 0.5128, 0.3717, 0.3606, -0.0286], 行标为0
[ 0.0933, -1.4781, -0.3561, -0.2652], 行标为1
[-0.8861, 0.6988, 1.1243, -1.1301] 行标为2
对行元素进行比较 0.5128,0.0933,-0.8861进行比较,输出最大值所在的下标,最大值为0.5128,行标为0,所以输出为0;0.3717,-1.4781,0.6988进行比较,输出最大值得下标2(0.6988最大,行下标为2),以此类推第一个矩阵输出[0, 2, 2, 0],第二个矩阵输出
[0, 0, 1, 2],第三个矩阵输出[1, 2, 2, 0]。即dim=1输出tensor([[0, 2, 2, 0],
[0, 0, 1, 2],
[1, 2, 2, 0]])
(3)dim=2,(3,3,4)对第三个维度的元素进行比较(4)。
import torch
x=[[[ 0.5128, 0.3717, 0.3606, -0.0286],
[ 0.0933, -1.4781, -0.3561, -0.2652],
[-0.8861, 0.6988, 1.1243, -1.1301]],
[[-0.0246, 0.0917, -0.0623, -1.4874],
[-2.0169, -0.8390, 0.1292, -0.1190],
[-0.4090, -0.9277, -0.9193, 2.1597]],
[[ 0.7030, -1.3322, -2.0856, 0.3122],
[ 1.4656, -1.1145, 0.8208, -1.8250],
[-0.8721, 0.4831, 1.5668, 0.1656]]]
x=torch.tensor(x)
print(x.argmax(dim=2))
输出:
tensor([[0, 0, 2],
[1, 2, 3],
[0, 0, 2]])
第0列 第1列 第2列 第3列
[ 0.5128, 0.3717, 0.3606, -0.0286],
[ 0.0933, -1.4781, -0.3561, -0.2652],
[-0.8861, 0.6988, 1.1243, -1.1301]
第一行 [ 0.5128, 0.3717, 0.3606, -0.0286]中最大的数0.5128所在的列标号为0,
第二行 [ 0.0933, -1.4781, -0.3561, -0.2652]最大元素0.0933所在的列标号为0,
第三行 [-0.8861, 0.6988, 1.1243, -1.1301] 最大元素1.1243所在的列标号为2,
所以输出[0, 0, 2]
[-0.0246, 0.0917, -0.0623, -1.4874],
[-2.0169, -0.8390, 0.1292, -0.1190],
[-0.4090, -0.9277, -0.9193, 2.1597]
同理,输出[1, 2, 3],
[ 0.7030, -1.3322, -2.0856, 0.3122],
[ 1.4656, -1.1145, 0.8208, -1.8250],
[-0.8721, 0.4831, 1.5668, 0.1656]
输出[0, 0, 2]
所以dim=2,总的输出为tensor([[0, 0, 2],
[1, 2, 3],
[0, 0, 2]])