As discussed in #438, we are spinning off the C codegen ("compiler") part of Treelite into a separate project (dubbed tl2cgen).

Goals

• Keep Treelite small. It should do one thing well, namely serve as a universal exchange and serialization format for decision tree forests.
• Hand off the control of tl2cgen to other contributor(s), as I do not have bandwidth to maintain it in the future.
• Make it easy for other projects to adopt Treelite as serialization format.
• Future-proof Treelite 4.0 by making breaking changes that will enable additional types of decision tree forests.

Functions to be migrated to tl2cgen + removed from Treelite 4.0

• treelite.Annotator
• treelite.Model.compile
• treelite.Model.export_lib
• treelite.Model.export_srcpkg
• treelite.generate_makefile
• The entire treelite_runtime package will be renamed to tl2cgen_runtime and will be moved to the tl2cgen project.
  • treelite_runtime.DMatrix
  • treelite_runtime.Predictor
• The entire Java runtime will be moved to tl2cgen project.

Create new project tl2cgen

• Create new repository dmlc/tl2cgen
• Set up directory skeleton for C++ source files
• Set up CMake build
• Set up directory skeleton for Python package
• Set up pyproject.toml and PEP 517 backend
• Set up pytests. In particular, there should be integration tests to ensure that C codegen functions of Treelite can be replaced 1-to-1 with those in tl2cgen.
• Make dmlc/tl2cgen public
• Migrate relevant issues from dmlc/treelite to dmlc/tl2cgen
• Set up a working Sphinx doc
• Create a 0.1 release for tl2cgen
• Publish tl2cgen to PyPI
• Submit a pull request to Conda-forge's staged recipes repo

Release Treelite 3.9

• Write a migration guide.
• Add deprecation warnings to Model.export_lib and other related functions. Warnings should direct users to the migration guide.
• Publish Treelite 3.9 to PyPI and Conda-forge

Release Treelite 4.0

• Remove the functions in the list above.
• Address the blockers (see below).
• Create tl2cgen 0.2 to handle Treelite 4.0+ serialization format.
• Publish Treelite 4.0 and tl2cgen 0.2 to PyPI and Conda-forge

Tasks for Treelite 4.0

• Replace setup.py with pyproject.toml #489
• Document Treelite serialization format. #503
• Implement v4 serialization format #515
• Documentation for library writers #469
• Breaking change: Update task_type enum to distinguish between regressors and binary classifiers #463
• Breaking change: Allow vector global_bias #445
• Breaking change: Add first-class support for multi-output models in Treelite 4.0 #467
• Support boosting from the average in sckit-learn #446
• Set up pre-commit hook to automatically check formatting #429
• Reformat Python code with black and isort #426
• Breaking change: API refactors #468
• Treelite crashes with XGBoost 2.0 dev #492
• Revamp JSON importer to make it easy to use #517