I would avoid using the .data attribute anymore and instead wrap the code in a
with torch.no_grad():
...
block if necessary.
Autograd cannot warn you, if you manipulate the underlying tensor.data, which might lead to wrong results.
3 Likes