CatBoost Roadmap: January-March 2023

[kizill]

Our main goals for January-March 2023 are solving long lasting problems with CatBoost binaries build and to make CatBoost with popular tools, like Optuna and others.

Building CatBoost

• Builds for Linux & MacOS should be more open and understandable for all the developers, so we are switching to CMake + Conan +ninja for building and releasing artifacts
• Python wheels now build with mk_wheel.py, we'd like to make setup.py and use it for building, packing and releasing wheels for Linux & MacOS

External tools and CatBoost

• Better support for parameters tuning: integration with hyperopt & optuna, make more examples
• Support for experiment tracking: better callbacks for mlflow, W&B support

Dask support

We'd like to make some steps to make distributed CatBoost training easier and make minimally working version in Dask environment.

[celestinoxp]

Hi,

Builds for Linux & MacOS should be more open...

you talk about Linux and MacOS... I ask and Windows too, right? :)

I think the next few months are going to be very interesting! I'm with you :) Thank you

[kizill]

Yes, we are still fixing Windows build problems. We are trying to fully switch to CMake or to MSVS solution files & msbuild. We can't fully migrate to clang toolkit on windows as we have CUDA code and it's compilation tightly bound with MSVC toolkits.

[makart19]

Hi,

As part of your migration to CMake can you please add support of:

• building without CUDA (smth like ‘-DHAVE_CUDA=no’ as it was for ‘ya make’).
• building for riscv64 CPU.

Thanks in advance!

[andrey-khropov]

Hi,

As part of your migration to CMake can you please add support of:

* building without CUDA (smth like  ‘-DHAVE_CUDA=no’ as it was for ‘ya make’).

Yes, there's an issue about that: #2298

* building for riscv64 CPU.

Please create an issue about that. But I'm not sure that it will be supported soon.

Thanks in advance!

[makart19]

Thank you for quick response, here it is #2314.

[kesav02]

Hello, are you planing to extend the C languge api to include training model functionality?