Simpletransformers: sliding window for Multi-Label classification

[jonas-nothnagel]

Are there plans to implement the sliding_window function for multi-label classification as well?
It would be great to be able to work with larger documents as well here!

[ThilinaRajapakse]

I'm afraid I don't have immediate plans to extend it to multi-label classification at the moment. But, I will add this to feature requests and see if I can do something about it.

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[jonas-nothnagel]

I would also still love to have this!

[empty-id]

@jonas-nothnagel It already has.

[jonas-nothnagel]

@null-id and @ThilinaRajapakse this is great to hear! Thank you for keep improving this library!

I just tried to enable sliding window in the MultiLabelClassificationArgs() but always ran into following error during evaluation:

TypeError: zip argument #2 must support iteration

Anything to keep in mind?

[ThilinaRajapakse]

Unfortunately, it's not done yet. The bot marked it as done, not me. 😅

[empty-id]

Ah... But Why I just used it normally for MultiLabelClassification?😂 @ThilinaRajapakse

model = MultiLabelClassificationModel('bert', 'bert-base-uncased', num_labels=5, args={'sliding_window':True, 'stride':0.8})

[ThilinaRajapakse]

You can set it, but I don't think it will be used. Unless I added the feature but have absolutely no memory of doing so. 😆

[empty-id]

Emm... However, It seems the result makes sense... The trained model can predict long text and return a list of predictions (the length is the number of segmentation for this piece of text).

[jonas-nothnagel]

@null-id How do you use the trained model for evaluation?

You can set and also run model.train_model() with sliding-window, however it always fails for me in the evaluation step.

@ThilinaRajapakse perhaps save us from speculating and implement :) ?

[empty-id]

@jonas-nothnagel Yeah, I skip the evaluation step and just predict.

[ThilinaRajapakse]

Well, damn lol. Either of you guys interested in taking a look and seeing if you can get it to work with evaluation so that we can add it officially? 😉

[empty-id]

I can try to mean the list of segments to get the result to evaluate. But is this an acceptable version to merge?

[ThilinaRajapakse]

Yes. That is how it is implemented in ClassificationModel as well. You should be able to just follow that spec I think.

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[jonas-nothnagel]

Wow is it done now?

[ThilinaRajapakse]

I'm afraid not. Apparently, it does work with training and prediction though (no guarantees).