[FEA] Support categorical features when serving XGBoost models

[gfalcone]

Hello !

XGBoost recently enabled developers to use categorical features in its models (Nvidia did an article on that : https://developer.nvidia.com/blog/categorical-features-in-xgboost-without-manual-encoding/).

From what I understand, we can load a XGBoost model trained on categorical features within the FIL_BACKEND.

However, the FIL_BACKEND only supports float in inputs, which means that we have to do some kind of ensemble (PYTHON_BACKEND + FIL_BACKEND) to accepts strings (steps described here : https://github.com/triton-inference-server/fil_backend/blob/main/notebooks/categorical-fraud-detection/Fraud_Detection_Example.ipynb)

It would make things easier to accept strings in the FIL_BACKEND. Would it be possible to do that ?

[hcho3]

It is currently not possible to accept string inputs in the FIL backend. The feature will require substantial amount of developer resource on our end. (The current model storage does not store label encoding, so it needs to be revamped). I will mark this as a feature request.

[gfalcone]

Thank you for your reply @hcho3 !