Non-pickle export option?

[nickeubank]

Related to #308 , expansion of #201

It'd be great to have a non-pickle export option. As in #308, some people want to move models to other languages, but even for those who are staying in Python, an export option that isn't pickle-dependent would be really nice given how brittle pickles can be to version differences (having trouble not imaging a brittle pickle now...) and the fact some environments don't support their use for security reasons.

[dswah]

@nickeubank Interesting!
What do you thinkwould be a good export format?

I see that scikit-learn supports pickle, joblib (similar to pickle), and onnx.

The Onnx approach would be cool, but requires some custom work. Maybe we can fork the sklearn-onnx repo to get a head start?

[fernandochafim]

Hi all! Thanks for raising this – I totally agree that having a non-pickle export option would be a great enhancement, especially for portability and security.

ONNX is definitely an interesting direction and could open the door to cross-language support, which would be very valuable. Another approach that might be more lightweight and flexible could be similar to what Keras does: separating model architecture and weights. For example, being able to export the model’s parameters (e.g., as .h5, .json) and reload them later with a load_weights-like method would improve compatibility across environments and mitigate some of the version fragility of pickle-based serialization.

Happy to help brainstorm or prototype if this seems feasible!