added on_backward trainer callback
#5249
Conversation
ArjunSubramonian
requested review from
dirkgr,
epwalsh and
AkshitaB
and removed request for
dirkgr
There was a problem hiding this comment.
I think this design is too heavy. I don't like putting a core functionality like loss.backward() into a callback. It makes it too hard to see what's going on in the trainer.
Instead, can we use the normal TrainerCallback, and give it some extra methods, like pre_backward() and post_backward()? Can you solve your problem with that?
I can't do adversarial training without |
|
You could leave one |
@dirkgr I have made the revisions you suggested :) |
CHANGELOG.md
Outdated
| @@ -7,6 +7,10 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 | |||
|
|
|||
| ## Unreleased | |||
|
|
|||
| ### Added | |||
|
|
|||
| - Added `BackwardCallback`, a training callback which allows for control over backpropagation and gradient manipulation. | |||
There was a problem hiding this comment.
This comment isn't accurate anymore, is it?
| if not backward_called: | ||
| trainer._scaler.scale(loss).backward() # type: ignore | ||
| return True | ||
| return False |
There was a problem hiding this comment.
Is it an error if this gets called with backward_called == True? Should we throw an exception in that case?
tests/training/trainer_test.py
Outdated
| if not backward_called: | ||
| loss.backward() | ||
| for param in trainer.model.parameters(): | ||
| param.grad *= 0.0 |
There was a problem hiding this comment.
Is that really the best way to do that?
| param.grad *= 0.0 | |
| param.zero_() |
I don't know for sure, but I would guess that zero_() is faster.
ArjunSubramonian
changed the title
on_backward trainer callback
* added BackwardCallback * finished tests * fixed linting issue * revised design per Dirk's suggestion * added OnBackwardException, changed loss to batch_ouputs, etc. Co-authored-by: Arjun Subramonian <arjuns@Arjuns-MacBook-Pro.local>
Additions proposed in this pull request:
on_backwardtraining callback which allows for control over backpropagation and gradient manipulation.