pytorch之expand,gather,squeeze,sum,contiguous,softmax,max,argmax

目录

gather

squeeze 

expand

sum

contiguous

softmax

max

argmax


gather

torch.gather(input,dim,index,out=None)。对指定维进行索引。比如4*3的张量,对dim=1进行索引,那么index的取值范围就是0~2.

input是一个张量,index是索引张量。input和index的size要么全部维度都相同,要么指定的dim那一维度值不同。输出为和index大小相同的张量。

import torch
a=torch.tensor([[.1,.2,.3],
                [1.1,1.2,1.3],
                [2.1,2.2,2.3],
                [3.1,3.2,3.3]])
b=torch.LongTensor([[1,2,1],
                    [2,2,2],
                    [2,2,2],
                    [1,1,0]])
b=b.view(4,3)

print(a.gather(1,b))
print(a.gather(0,b))
c=torch.LongTensor([1,2,0,1])
c=c.view(4,1)
print(a.gather(1,c))

输出:

tensor([[ 0.2000,  0.3000,  0.2000],
        [ 1.3000,  1.3000,  1.3000],
        [ 2.3000,  2.3000,  2.3000],
        [ 3.2000,  3.2000,  3.1000]])
tensor([[ 1.1000,  2.2000,  1.3000],
        [ 2.1000,  2.2000,  2.3000],
        [ 2.1000,  2.2000,  2.3000],
        [ 1.1000,  1.2000,  0.3000]])
tensor([[ 0.2000],
        [ 1.3000],
        [ 2.1000],
        [ 3.2000]])

squeeze 

将维度为1的压缩掉。如size为(3,1,1,2),压缩之后为(3,2)

import torch
a=torch.randn(2,1,1,3)
print(a)
print(a.squeeze())

输出:

tensor([[[[-0.2320,  0.9513,  1.1613]]],


        [[[ 0.0901,  0.9613, -0.9344]]]])
tensor([[-0.2320,  0.9513,  1.1613],
        [ 0.0901,  0.9613, -0.9344]])

expand

扩展某个size为1的维度。如(2,2,1)扩展为(2,2,3)

import torch
x=torch.randn(2,2,1)
print(x)
y=x.expand(2,2,3)
print(y)

输出:

tensor([[[ 0.0608],
         [ 2.2106]],

        [[-1.9287],
         [ 0.8748]]])
tensor([[[ 0.0608,  0.0608,  0.0608],
         [ 2.2106,  2.2106,  2.2106]],

        [[-1.9287, -1.9287, -1.9287],
         [ 0.8748,  0.8748,  0.8748]]])

sum

size为(m,n,d)的张量,dim=1时,输出为size为(m,d)的张量

import torch
a=torch.tensor([[[1,2,3],[4,8,12]],[[1,2,3],[4,8,12]]])
print(a.sum())
print(a.sum(dim=1))

输出:

tensor(60)
tensor([[  5,  10,  15],
        [  5,  10,  15]])

contiguous

返回一个内存为连续的张量,如本身就是连续的,返回它自己。一般用在view()函数之前,因为view()要求调用张量是连续的。可以通过is_contiguous查看张量内存是否连续。

import torch
a=torch.tensor([[[1,2,3],[4,8,12]],[[1,2,3],[4,8,12]]])
print(a.is_contiguous)

print(a.contiguous().view(4,3))

输出:

<built-in method is_contiguous of Tensor object at 0x7f4b5e35afa0>
tensor([[  1,   2,   3],
        [  4,   8,  12],
        [  1,   2,   3],
        [  4,   8,  12]])

softmax

假设数组V有C个元素。对其进行softmax等价于将V的每个元素的指数除以所有元素的指数之和。这会使值落在区间(0,1)上,并且和为1。

S_{i}=\frac{e^{v_{i}}}{ \sum_{i=1}^{C} e^{v_{i}} }

import torch
import torch.nn.functional as F

a=torch.tensor([[1.,1],[2,1],[3,1],[1,2],[1,3]])
b=F.softmax(a,dim=1)
print(b)

输出:

tensor([[ 0.5000,  0.5000],
        [ 0.7311,  0.2689],
        [ 0.8808,  0.1192],
        [ 0.2689,  0.7311],
        [ 0.1192,  0.8808]])

max

返回最大值,或指定维度的最大值以及index

import torch
a=torch.tensor([[.1,.2,.3],
                [1.1,1.2,1.3],
                [2.1,2.2,2.3],
                [3.1,3.2,3.3]])
print(a.max(dim=1))
print(a.max())

输出:

(tensor([ 0.3000,  1.3000,  2.3000,  3.3000]), tensor([ 2,  2,  2,  2]))
tensor(3.3000)

argmax

返回最大值的index

import torch
a=torch.tensor([[.1,.2,.3],
                [1.1,1.2,1.3],
                [2.1,2.2,2.3],
                [3.1,3.2,3.3]])
print(a.argmax(dim=1))
print(a.argmax())

输出:

tensor([ 2,  2,  2,  2])
tensor(11)

猜你喜欢

转载自blog.csdn.net/hbu_pig/article/details/81454503
今日推荐