Pytorch SGD 和梯度

Question

I tried figure out how pytorch SGD works.

However I don't understand why I get this result.

Here are codes.

model = torchvision.models.resnet18(weights = None)

output = model(torch.rand([100,3,64,64]))
target = torch.ones([100,1000])

loss_fn = torch.nn.CrossEntropyLoss()
loss = loss_fn(output,target) / 1000

opt = torch.optim.Adam(model.parameters(), lr=2)

for i in model.parameters():
    print(i[0,0,0,0])
    break

# update
opt.zero_grad()
loss.backward()
opt.step()

for i in model.parameters():
    print(i.grad[0,0,0,0])
    break
for i in model.parameters():
    print(i[0,0,0,0])
    break

and output is

tensor(0.0358, grad_fn=<SelectBackward0>)
tensor(-0.0011)
tensor(2.0357, grad_fn=<SelectBackward0>)

It seems that the parameter is updated without being affected by gradient.
It means affected only by lr

Of course I think I'm wrong.

Thanks for letting me know where the error is!