I think one can build a graph without importing any class from torch.autograd.
1 Like
I think the main point was that the use of Variable
was quite confusing for new users. Having everything as a Tensor
and simple context manager like torch.no_grad()
is actually enough to perform all operations.
So torch.autograd
is still used the same way as before, but explicit use of it has been reduce to simplify API and user code.
1 Like