pytorch中tensor.expand()和tensor.expand_as()函数

tensor.expand()

>>> import torch
>>> a = torch.tensor([[1],[2],[3]])
>>> print(a.size())
torch.Size([3, 1])
>>> a.expand(3,4)
tensor([[1, 1, 1, 1],
        [2, 2, 2, 2],
        [3, 3, 3, 3]])
>>> a
tensor([[1],
        [2],
        [3]])
>>> a.expand(-1, 4)  # -1 means not changing the size of that dimension
tensor([[ 1,  1,  1,  1],
        [ 2,  2,  2,  2],
        [ 3,  3,  3,  3]])

tensor.expand_as()

>>> b=torch.tensor([[4,7],[5,8],[6,9]])
>>> b
tensor([[ 4,  7],
        [ 5,  8],
        [ 6,  9]])
>>> print(b.size())
torch.Size([3, 2])
>>> a.expand_as(b)
tensor([[ 1,  1],
        [ 2,  2],
        [ 3,  3]])
>>> a
tensor([[ 1],
        [ 2],
        [ 3]])

总结：

tensor.expand()函数是把一个tensor变形为尺寸是括号内所给大小的tensor；

tensor.expand_as()函数是把一个tensor变成和函数括号内所给tensor同样形状的tensor；

expand括号里为size，expand_as括号里为其他tensor。

两个函数均是原来的tensor和变形后tensor不共享内存 的，如需使用变形后的，需重新赋值。