add UperNet model

[jawi289o]

kindly add UperNet model

[GrantorShadow]

commenting to be a part of the discussion

[qubvel]

Hi! Thanks for the interest! Would anyone like to add this model? I will appreciate any contribution, just let me know if you are interested and need any guidance 🙂

[GrantorShadow]

I would love to take this on, any guidance is much appreciated
@qubvel

[qubvel]

Thanks for your interest @GrantorShadow!

You can fork the repo, clone it, and copy-paste the Unet model folder in decoders/ and modify it to Upernet. I suppose this will be the easiest way to adopt a new model. You have to rewrite decoder.py file, while model.py most probably will be similar or with minor changes.
Then you can open a PR for initial review and feedback.

In terms of testing would be nice to fine-tune it with an example notebook in the repo to check if the model is trainable and add some basic tests under /tests folder.

If you will face any particular difficulties we can discuss it here or under the PR!
Looking forward to collaborating with you, thanks!

[jawi289o]

@GrantorShadow Hi are you working on it?

[GrantorShadow]

Yes sorry I got a little delayed but will be getting it done in the next 2 weeks

[yurithefury]

Given Hugging Face integration, I believe you can simply use UperNetForSemanticSegmentation from transformers:

from transformers import UperNetForSemanticSegmentation
self.model = UperNetForSemanticSegmentation.from_pretrained(
    pretrained_model_name_or_path='openmmlab/upernet-convnext-small'
    )