Closed ogencoglu closed 2 months ago
@hcho3, any thoughts on how hard this would be to support for Treelite? I don't see any obvious dealbreakers offhand.
For context, @ogencoglu, the FIL backend uses Treelite under the hood for all of its model ingest and conversion logic. Assuming we can get this supported in Treelite, it should be straightforward to make it available in the FIL backend as well. Thanks for raising the issue!
@wphicks Treelite added a limited support for HistGradientBoosting (https://github.com/dmlc/treelite/pull/444). The support is not yet complete: HistGradientBoostingClassifier
with n_classes > 2
is not supported. To complete this support, we need to release Treelite 4.0 version.
Ah right! The global bias issue. I'd forgotten that we needed that for HistGradientBoosting
. Thanks, @hcho3!
@ogencoglu, given this, you can expect to see support for this in the FIL backend shortly after the release of Treelite 4.0. We'll keep this issue up-to-date and close it once Treelite 4.0 is integrated with the FIL backend.
HistGradientBoostingClassifier
is fully supported in Treelite 4.1. Based on the release schedule, the new Treelite will be incorporate into the FIL backend in release 24.05.
HistGradientBoostingClassifier
is now fully supported in the latest FIL backend (24.05).
Would be great to have support for scikit-learn's Histogram-based Gradient Boosting which is much simpler and faster than regular gradient boosting. Classifier, Regression