Open steven-johnson opened 5 years ago
@BachiLi Do you think we need an explicit drop_gradients intrinsic? Or should we just silence this? What do you think it most usable for people using autodiff?
(the intrinsic would drop gradients silently)
It might not be a good idea to silence it just yet. In this case, I believe the dropping warnings come from the fast_log() call. It is not very obvious that some gradients would be dropped by just calling an approximate log function. Let the warnings be there by default and have an option to turn them off might be better.
It is not very obvious that some gradients would be dropped by just calling an approximate log function.
Indeed, quite so! Which is the entire point of my complaint :-) The warning needs to be made actionable somehow.
fast_log should probably be lowered later than it is, like lerp
When building with the ml-autoscheduler, you'll often get a lot of buildspam of the form:
(etc)
Although emitted as a "user" warning, it's likely that the typical user of the autoscheduler has no idea what this is actually warning about, whether they should be concerned, or how they can fix the warning. If we can't upgrade the text to something user actionable, we should remove it (or downgrade it to debug(1) or lower).