Heavy CQR object

[diegoglozano]

Hi! First of all, thank you for your work!
We are currently using your package, specifically, CQR class (using two xgboost inside a DualPredictor).

When I serialised the object using joblib, I realised that the object size was about 500MB.

Checking the code I discovered this line, where the data is saved as an attribute of the IdSplitter.

puncc/deel/puncc/api/splitting.py

Line 82 in d09f773

self._split = [(X_fit, y_fit, X_calib, y_calib)]

Is this an expected behavior? I fixed it by setting:

my_object.conformal_predictor.splitter._split = None

before saving the artifact.

[M-Mouhcine]

Hi @diegoglozano,

Thank you for your feedback.

We currently provide the method save for ConformalPredictor serialization. However, I understand that it may not address your concern, as it still serializes the data. To resolve this, I’ll introduce a flag argument that allows to specify whether the splitter should be saved or not.

In the meantime, your solution works well and can be used with no negative impact.