Add `multiclass` fine-tuning example

[qubvel]

It would be great to have more examples for fine-tuning in the library!

Similar to current examples for binary-segmentaion and multi-label segmentation would be great to have multi-class segmentation example (the whole image is annotated). The same camvid dataset can be used.

The example should showcase how to

• fine-tune a model with pytorch-ligtning (or any other fine-tuning framework, even a plain pytorch)
• use loss function with ignore index for unlabeled class
• compute metrics per-class
• visualize results

In case anyone wants to work on this you are welcome! Just notify in this issue, and then ping me to review a PR when it's ready.

[abdksyed]

I can take this up.

Do we want to fine-tune on all the classes or should I limit to say 5 most common classes from the dataset?

[qubvel]

Hi @abdksyed, thanks a lot for your interest to this issue!
Let's have all the classes from the dataset.

[TimbusCalin]

@qubvel I have managed to create a multiclass classsification example.

Check the PR here : #922

[abdksyed]

looks like I am little late. I was about to create a PR, with the notebook. But looks like @TimbusCalin has already done it.

[qubvel]

Hi @abdksyed, If you already have the work done you can open a PR, I will prioritize it because you were first on this issue (sorry @TimbusCalin). Or, to save your contribution, you can implement additional features, like per-clas metrics computing for multiclass segmentation, and join with @TimbusCalin PR.

Next time I would recommend opening a PR in draft mode and linking it to the issue as early as possible to let others know that work is in progress 🙂

Thank you both for your contributions, I hope we will resolve this issue in a good way for both of you!

[abdksyed]

I am happy with @TimbusCalin's PR, we can merge that. As for the class wise metric, I will try to implement it, but if @TimbusCalin does it before me, he can push the update.

[TimbusCalin]

Yes @qubvel we also had a brief discussion in the PR thread where we decided that it's optimum to integrate the PR now, and then @abdksyed will take on to implement the class-wise metric on its own feature branch.

@qubvel thanks for the suggestion, I will do that accordingly in the next situation (I've been working one entire day on it without having the draft open, and just opened the PR).

Cheers!