Closed xgdgsc closed 3 years ago
Currently, the approach for saving/loading would be using BSON:
using BSON: @save, @load
@save "model_L1.bson" model
@load "model_L1.bson" model
I hadn't consider integrating at the moment. Would that be a useful feature to you?
Yes, though not with a high priority. Some applications may require c library integration. And I' d be curious how would julia compare with the code generated by treelite.
I just updated the benchmarks to include inference and EvoTrees appears to perform surprisingly well compared to XGBoost:
Dimensions / Algo | XGBoost Hist | EvoTrees | EvoTrees GPU |
---|---|---|---|
100K x 100 | 0.177s | 0.029s | 0.035s |
500K x 100 | 0.861s | 0.191s | 0.214s |
1M x 100 | 1.67s | 0.400s | 0.469s |
5M x 100 | 8.51s | 2.14s | 2.43s |
How do you deploy? Consider adding save/load functions? Or be able to export to some format of other packages like lightgbm/xgboost to leverage the package dmlc/treelite: model compiler for decision tree ensembles?