[pytorch]极易犯错的问题

一 x.clone()

错误：

import torch
x = torch.randn(5)
print(x)
y = x.clone()
x =x*12

print(y)

tensor([-0.0455,  0.9827, -0.7915,  0.8110,  1.1524])
tensor([-0.0455,  0.9827, -0.7915,  0.8110,  1.1524])

正确：

import torch
x = torch.randn(5)
print(x)
y = x.clone()
z =x*12

print(z)

tensor([ 1.7743, -2.0879,  0.4297, -0.3849, -0.6147])
tensor([ 21.2920, -25.0553,   5.1562,  -4.6190,  -7.3768])

class A(nn.Module):
    def __init__(self):
        super().__init__()
        self.m = nn.Identity()  #nn.Linear(12, 12)

    def forward(self, x):
        print(f'old x:{x}')
        y = x  #==> .clone()
        x = self.m(x)*12
        print(f'new x:{x}')
        print(f'res y:{y}')

        return

A()(torch.randn(12))

old x:tensor([-0.0204,  0.5557,  2.1049, -1.2744, -1.0640,  1.2480, -2.2674, -0.6850,
         0.0507, -0.9022, -0.2756, -0.1572])
new x:tensor([ -0.2450,   6.6683,  25.2586, -15.2927, -12.7677,  14.9765, -27.2085,
         -8.2195,   0.6080, -10.8261,  -3.3074,  -1.8865])
res y:tensor([-0.0204,  0.5557,  2.1049, -1.2744, -1.0640,  1.2480, -2.2674, -0.6850,
         0.0507, -0.9022, -0.2756, -0.1572])