Bugfix: now works correctly even if the input Tensor(s) do not require grad.
This works by adding a fake input to the forward()
of torch.autograd.Function
.
Also adds an automatic detection of prologue (first layers with no trainable parameters)
Prochaines maintenances programmées: lundi 06/05, lundi 03/06, lundi 01/07
Pour plus d'informations: https://doc-si.inria.fr/display/SU/Gitlab
This works by adding a fake input to the forward()
of torch.autograd.Function
.
Also adds an automatic detection of prologue (first layers with no trainable parameters)