Enabling passing unpickled model directly rather than just through the mode.txt file

[faithfulalabi]

• Hey, thanks for your great work on this. I'm currently looking to use your tool for optimizing a model in production. When the model is trained, it doesn't save to a model.txt file; instead, it's dumped as a pickle object. Would it be possible to load the unpickled model directly into the lleaves.Model function?

[siboehm]

Hi! That's currently not possible and I probably won't support it in the future either. It makes the implementation much simpler to load the model.txt instead of loading a pickled model :) Pickle is very Python specific, and I'm considering rewriting lleaves from scratch in Cpp.

You could try storing the model.txt in an in-memory tmp-file, using code roughly like this:

import lightgbm as lgb
import tempfile

with tempfile.NamedTemporaryFile(mode='w+b', delete=False, suffix='.txt') as tmp_file:
    model.save_model(tmp_file.name)
    print(f"Model saved to: {tmp_file.name}")

loaded_model = lleaves.Model(tmp_file.name)

[faithfulalabi]

Thanks for the fast response and sample code. Was going to do that as a work around but thought I'd ask first. Thanks again 😄