Fix ViTForMaskedImageModeling doc example

[ydshieh]

What does this PR do?

Same #22185

[amyeroberts]

@ydshieh @alaradirik @fxmarty The issue coming from #22152 was an oversight on my part about breaking changes. Perhaps we should revert that PR first and then agree how to introduce this change as it is intended to be added to other vision model?

[ydshieh]

Oh, I thought it was a new model head! Indeed a breaking change there. Good for me to revert that PR, but would be nice to talk to Sylvain or Lysandre first (if you feel necessary). I will leave you judge.

Regarding a solution if we really want to have this new attribute and the new name MaskedImageCompletionOutput, adding a new property (named logits) to MaskedImageCompletionOutput might be a way, but I didn't think about this deeply.

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[ydshieh]

Converted to draft for now

[amyeroberts]

@ydshieh Yep - let's get @LysandreJik and @sgugger 's opinions.

I think having the logits param is probably the best solution. As far as I know, it's very rare to check against the model output type itself. I believe reconstruction was chosen because of the ImageSuperResolutionOutput data class. As mentioned in the original PR - we probably do want a different model type to be returned as the documented shapes are incorrect.

@alaradirik - could you open a PR to revert the changes?

[amyeroberts]

Actually, it's late for @alaradirik. I'll open the PR now.

[sgugger]

Yes we can't rename the parameter in the outputs like that for a model that has been around for a bit. What is even more annoying is that the commit was in the release, so we will need to make a patch with the fix.

[ydshieh]

Close this PR as it's clear we will and have to definitely use the original logits.

[alaradirik]

Sorry for being late to comment, I added the MaskedImageCompletionOutput to replace the inaccurate MaskedLMOutput class used by the masked image modeling heads (ViT and DeiT). Neither of these models have any checkpoints on the hub as mentioned in #22152 . Swin's MIM head has its own output class but no fine-tuned checkpoints for the MIM task either.

With that said, ViT and Swin's MIM heads are implementations of SimMIM and SimMIM have recently released fine-tuned checkpoints for these two models (as opposed to the base model weights on the hub for Swin MIM head). I'm planning to convert these checkpoints and add a masked-image-completion pipeline after @sheonhan merges the ICT PR (a contemporary, better performing MIM model). It'd be great to add an output class and fix inaccurate class output (listed as a language model in the docs) before that. While logits is not an accurate output name in this case as the model returns full reconstructed images, I could replace reconstruction with logits and open a new PR.

What do you think @amyeroberts @sgugger?
CC @LysandreJik @ydshieh