is likely to be unexpected: the tuple returned by torch.autograd.grad is repeated twice, so gradients[0] is not multipled by 2.
This is not a major issue, but may bother someone who likes accurate gradients (e.g. who is using default MTT parameters ).
Hi, very inspiring work! The behavior of equation (6) implementation:
https://github.com/justincui03/tesla/blob/b9180b334bb726b5fb978abc8f1e0eedd82260b9/distill.py#L418-L420
is likely to be unexpected: the tuple returned by
torch.autograd.grad
is repeated twice, sogradients[0]
is not multipled by 2. This is not a major issue, but may bother someone who likes accurate gradients (e.g. who is using default MTT parameters ).