allardvm / lightgbm.jl Goto Github PK

This could also be viewed as a documentation problem, but I will admit I was surprised by this behavior which suggests it might be considered a bug instead.

The fit! and train functions honor early stopping. But although the best iteration is reported to the console and the scores are truncated to the best iteration (shrinkresults!), it is apparently up to the user to read out the truncated scores and save a truncated model—by default, saving a model after training will save the last iteration, not the best.

But if the scores are truncated, shouldn't the model be truncated back to the best as well? There's a strange mismatch there. At a minimum it'd be great if the README had an example of reading the truncated scores to find the best iteration and explicitly truncate the model on save. Thanks.

Hi

Thank you for this package. What is the current plan regarding support for ranking?

Any plans to update this to Julia v1.0 in the near future?

Thanks for great work!
Do you have plan to make this as a official package ?

Hello, first thanks for this amazing package :)

Recently we've done some changes in the LightGBM Api (Microsoft's upstream) to allow for early stopping at prediction time, which can speed up predictions significantly (https://github.com/Microsoft/LightGBM/blob/master/include/LightGBM/c_api.h#L644, the parameter string is what is new here)

I will be gone for two weeks so I won't be able to work on this, but I wanted to leave the issue open here to warn you about the possible incompatibility with the latest upstream, and also that it would be very useful to have early stopping working with LightGBM.jl.

Hello,

I am confused! There is another library with the same name as yours.
The other one also work as an interface.

So what is the difference?
Are you both working together?

Which one should I use?

Thank you!

Title says it all - it would be nice if scikit-learn interface would be supported through ScikitLearnBase.jl.