output.topk()函数解读

代码如下

>>> a=torch.randn((4,6))
>>> print(a)
tensor([[ 0.7042,  0.2533,  1.1596, -0.7436,  0.5264,  0.2085],
        [ 0.2641,  0.9683,  0.4469, -1.9215, -0.7564,  1.1776],
        [ 1.0520, -1.6003, -0.8634,  1.7596, -0.8464,  0.7166],
        [-0.0492, -0.7746,  1.2592, -0.8273,  0.1266,  1.0450]])
>>> maxk=max((1,3))
>>> _, pred=a.topk(maxk,1,True,True)
>>> print(_)
tensor([[1.1596, 0.7042, 0.5264],
        [1.1776, 0.9683, 0.4469],
        [1.7596, 1.0520, 0.7166],
        [1.2592, 1.0450, 0.1266]])
>>> print(pred)
tensor([[2, 0, 4],
        [5, 1, 2],
        [3, 0, 5],
        [2, 5, 4]])
>>> _, pred=a.topk(1,1,True,True)
>>> print(_)
tensor([[1.1596],
        [1.1776],
        [1.7596],
        [1.2592]])
>>> print(pred)
tensor([[2],
        [5],
        [3],
        [2]])

如上我们可以看到，topk()函数取指定维度上的最大值(或最大几个)，第二个参数dim=1，为按行取，dim=0，为按列取，如下：

>>> _, pred=a.topk(1,0,True,True)
>>> print(_)
tensor([[1.0520, 0.9683, 1.2592, 1.7596, 0.5264, 1.1776]])
>>> print(pred)
tensor([[2, 1, 3, 2, 0, 1]])