零碎

突然想到的零碎point

在torch中对clone一个张量类似于获得该张量乘1的值，即：x_copy = x.clone() => x_copy = x

>>> import torch
>>> import torch.nn as nn
>>> a = nn.Parameter(torch.tensor([1.,]))
>>> b = nn.Parameter(torch.tensor([4.,]))
>>> x = nn.Parameter(torch.tensor([1.,]))
>>> x_c = x.clone()
>>> x_c
tensor([1.], grad_fn=<CloneBackward0>)
>>> y = a * x + b * x_c
>>> y
tensor([5.], grad_fn=<AddBackward0>)
>>> y.backward()
>>> x.grad
tensor([5.])

说明克隆的结果和原张量并不完全相互独立，克隆出来的结果更像是原张量的乘1结果
关于克隆出来的结果与原张量id一致，可参考：https://www.zhihu.com/question/275830564