Can't retrieve lightgbm model's best_iteration after Optuna optimization about optuna

[nzw0301]

I think we can close this issue because the workaround is discussed on the lightgbm issue helpfully. Alternatively we might save the best mode as a file and update the best model's file name similarly to the shared script.

Personally I appreciate you for your minimal reproducible code and investigating the copy logics in lightgbm repo!

[YingJie-Zhao]

BTW，this problem has been issued at microsoft/LightGBM#6384 (comment) and currently didn't found any code related to lightgbm could cause this situation. So I think may be turning into Optuna developer's help could be better.

[HideakiImamura]

Perhaps this is due to the migration of the integration module to the optuna_integration package. We have supported backward compatibility to function without re-writing the code, but you will need to install the optuna_integration package. Could you please run pip install optuna_integration?

[YingJie-Zhao]

The reproducible example above was tested with both optuna 3.6.0 and optuna_integration 3.6.0 installed. So I think may be this was not about the version.

[nzw0301]

I suspect that this is lightgbm issue; when I copy model with copy.copy or copy.deepcopy, both copy models have best_iteration=-1.

[YingJie-Zhao]

Thanks to your reply.

I tried printing best_iteration both inside and outside the objective function, seems like the inner print works well but the outer print returns -1. See microsoft/LightGBM#6384 (comment)

[nzw0301]

Thank you for sharing it. I've left the comment on this as microsoft/LightGBM#6384 (comment).