Skip to content

Latest commit

 

History

History
107 lines (92 loc) · 4.75 KB

sklearn_and_cuml.md

File metadata and controls

107 lines (92 loc) · 4.75 KB
Raw

Scikit-Learn and cuML Support

NOTE: Due to a change in Scikit-Learn 1.2.0, forest models from version 1.2.0 and later are not currently supported. Support will be added in an upcoming release of Triton.

Model Serialization

While LightGBM and XGBoost have their own serialization formats that are directly supported by the Triton FIL backend, tree models trained with Scikit-Learn or cuML are generally serialized using Python's pickle module. In order to avoid a round-trip through Python in Triton, the FIL backend instead requires that these pickled models first be converted to Treelite's binary checkpoint format. Note that this also allows you to make use of any Treelite-supported model framework in Triton simply by exporting to the binary checkpoint format.

The FIL backend repo includes scripts for easy conversion from pickle-serialized cuML or Scikit-Learn models to Treelite checkpoints. You can download the relevant script for Scikit-Learn here and for cuML here.

Prerequisites

To use the Scikit-Learn conversion script, you must run it from within a Python environment containing both Scikit-Learn and Treelite. To use the cuML conversion script, you must run it from within a Python environment containing cuML.

For convenience, a conda environment config file is provided which will install all three of these prerequisites:

conda env create -f scripts/environment.yml
conda activate triton_scripts

Converting to Treelite checkpoints

NOTE: The following steps are not necessary for LightGBM or XGBoost models. The FIL backend supports the native serialization formats for these frameworks directly.

If you already have a Scikit-Learn or cuML RF model saved as a pickle file (model.pkl), place it in a directory structure as follows:

model_repository/
`-- fil
    |-- 1
    |   `-- model.pkl
    `-- config.pbtxt

Then perform the conversion by running either:

./convert_sklearn model_repository/fil/1/model.pkl

for Scikit-Learn models or

./convert_cuml.py model_repository/fil/1/model.pkl

for cuML models. This will generate a checkpoint.tl file in the model repository in the necessary location. You can then proceed as with any other model type, setting the model_type parameter in config.pbtxt to "treelite_checkpoint".

Note that Treelite did not guarantee compatibility between minor release versions for its binary checkpoint model until version 3.0.0 and does not guarantee compatibility between major releases, so it is recommended that you keep the original pickle file. If you later make use of a newer version of Treelite, you can simple re-run the conversion on this pickle file.