TL;DR - Is there a way to access the numerical
and analytical
results from a torch.autograd.gradcheck
call?
Hi All,
I’ve recently being trying to validate a custom torch.autograd.Function
object via torch.autograd.gradcheck
, however, my manually defined backward is incorrect. The torch.autograd.gradcheck
function works well at telling me what parts of my jacobian are wrong, but annoyingly I can’t seem to access the numerical
and analtyical
results from the torch.autograd.gradcheck
itself as they’re simply printed within the error function and not returned.
Is there a way to access the numerical
and analytical
tensors, so I can do an element-wise difference to check what values are correct/wrong?