Returning outputs only when asked for for MaskFormer. #15936
Conversation
Narsil
requested review from
FrancescoSaverioZuppichini and
sgugger
and removed request for
FrancescoSaverioZuppichini
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. |
| @@ -2421,6 +2428,7 @@ def forward( | |||
| mask_labels: Optional[Tensor] = None, | |||
| class_labels: Optional[Tensor] = None, | |||
| pixel_mask: Optional[Tensor] = None, | |||
| output_auxiliary_logits: Optional[bool] = None, | |||
There was a problem hiding this comment.
This should probably go with a config argument (like output_hidden_states etc) that would give a default. Otherwise defaulting to None makes no sense.
There was a problem hiding this comment.
Ok, I made the change but will wait for @FrancescoSaverioZuppichini advice to make sure this is sound.
There was a problem hiding this comment.
Thanks! Just a little bit of context, the auxiliary_logits are consumed by the loss if use_auxiliary_loss was set to true. So we could remove them from the output or add a flag in the config to return them
There was a problem hiding this comment.
So if someone uses both output_auxiliary_logits=True and use_auxiliary_loss=True then it wouldn't work currently is that it ? (If so we need to change the consuming behavior into a copy I think)
There was a problem hiding this comment.
Friendly ping @FrancescoSaverioZuppichini
There was a problem hiding this comment.
It will work :)
What does this PR do?
Change the return output from
()toNonewhich seems more aligned withthe rest of the library.
Also
auxiliary_logitsseem optional and don't seem to be used by the featureextractor, so this PR makes them optional too.
Note: I couldn't test the modeling tests, there seem to be no fast tests, and the slow tests are failing for reasons seemingly unrelated to this PR.
Fixes # (issue)
Before submitting
Pull Request section?
to it if that's the case.
documentation guidelines, and
here are tips on formatting docstrings.
Who can review?
Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.