I am learning pytorch and write a simple code as below.

import torch x = torch.randn(3,requires_grad=True).cuda() print(x) y = x * x print(y) y.backward(torch.tensor([1,1.0,1]).cuda()) print(x.grad) 

tensor([ 0.5934, -1.8813, -0.7817], device='cuda:0', grad_fn=<CopyBackwards>) tensor([0.3521, 3.5392, 0.6111], device='cuda:0', grad_fn=<MulBackward0>) None 

if I change the code as

from torch.autograd import Variable import torch # x = torch.randn(3,requires_grad=True).cuda() x = Variable(torch.randn(3).cuda(),requires_grad=True) print(x) y = x * x print(y) y.backward(torch.tensor([1,1.0,1]).cuda()) print(x.grad) 

tensor([0.9800, 0.3597, 1.6315], device='cuda:0', requires_grad=True) tensor([0.9605, 0.1294, 2.6617], device='cuda:0', grad_fn=<MulBackward0>) tensor([1.9601, 0.7194, 3.2630], device='cuda:0') 

The grad is ok. But why? I hate the Variable class.

env

python:3.8

pytorch:1.5

cuda :10.2