Weird optimizer behaviour

autograd
1

why is this code not crashing

import torch
x=2.0
y =2.0
param  =  torch.nn.Parameter(torch.tensor(x))
opt = torch.optim.SGD([param],lr =.1)
loss = torch.nn.Parameter(torch.tensor(y))
loss.backward()
opt.step()
print('done')

Should not the optimizer complain saying param was not used in the computation ?
Thanks in advance!

2

Hi,

Optimizers don’t do that in general no. They will ignore Tensors whose .grad field is None and use the gradient that is there if there is one.

1 Like