Pass arguments to callback

I would like to be able to give heavy objects computed in optimize(trial) as parameters to the callback(study, trial).

Motivation

I want to save predictions when I have an improvement, that is when study.best_trial.number == trial.number.

Description

My optimize(trial) function performs cross-validation, and I would like to be able to reuse the fitted estimators to make my predictions.

The models are too heavy to be stored in trial.user_attrs.

Propositions

I think there are 2 options:

• provide a field callback_data in FrozenTrial that can be set from objective(trial) with trial.set_callback_data(value) or trial.set_callback_data(key, value)

• allow objective(trial) to return more complex values, e.g. through an object Result(value, *args, **kwargs) that is then used on callback(study, trial, *args, **kwargs)

[nlgranger]

Yes. Contrary to a lambda function, the output of partial is picklable if its arguments are (the function is importable and the arguments are pickable) so I believe it should work. If I understand you problem correctly, I don’t think the arguments will be sent to the process anyway in this case, they will just be there in memory after forking the main process.

[nlgranger]

FYI functools.partial might be a simpler and more idiomatic solution as well.