Import a SetFit model (MPNetForSequenceClassification) into Spark-NLP

[dfustes]

Hi folks! I'm trying to integrate a given classification model with Spark NLP, but I'm finding some difficulties.

The model has been trained using SetFit. In this case, the model is built with MP-Net, plus a classification head, using Pytorch. From what I've understood from the documentation, to load this kind of model in Spark-NLP, I would need to convert it to a Tensorflow SavedModel using TFMPNetForSequenceClassification, and then use the corresponding annotator in Spark-NLP (I guess it would be MPNetForSequenceClassification, which does not exist as of today).

Could you tell if this is feasible and which are the steps needed in order to achieve this integration?

[maziyarpanahi]

Hi @dfustes,

Your steps are correct and you are also correct that the MPNet for sequence and token classifications do not exist.

We can add them as a feature request for the next release.

[dfustes]

Hey @maziyarpanahi,

thanks a lot for your quick feedback. I would love to see that feature in Spark-NLP, so please add that feature request. I'm a Software Engineer myself, so I could contribute as well if you give me some guidance.

Cheers!

[maziyarpanahi]

#14060