Yet another Tsetlin Machine

A Python implementation of the Tsetlin Machine, running on CPU and CUDA.

• Optimized for large datasets, convolution, and thermometer-encoded discretized inputs.
• Binary, multi-class, multi-output, and regression TM classes.
• Binarizer module.
• Pickle support for model saving/loading.
• Local (WAC) and global (WIC) interpretation helpers.
• Timer, Profiler, and other small utils.

Implemented TM variants:

• Vanilla TM
• Weighted TM
• Regression TM
• Coalesced TM
• Convolutional TM
• ****** TM

Install

Requires Python >= 3.11.

• Using pip:

  pip install "ytm @ git+https://gitlab.com/cair/YTM.git"          # CPU
  pip install "ytm[cuda] @ git+https://gitlab.com/cair/YTM.git"    # CPU + CUDA

• Using uv:

  uv add "ytm @ git+https://gitlab.com/cair/YTM.git"          # CPU
  uv add "ytm[cuda] @ git+https://gitlab.com/cair/YTM.git"    # CPU + CUDA

• Using pixi (Recommended):

  pixi add clang llvm-openmp                                                  # C compiler + OpenMP for CPU
  pixi add --pypi ytm --git "https://gitlab.com/cair/YTM.git"            # CPU only
  pixi add --pypi "ytm[cuda]" --git "https://gitlab.com/cair/YTM.git"    # CPU+CUDA

Note:

• CPU:

  • C kernels are compiled at runtime, so clang or gcc must be on PATH. OpenMP (libomp/libgomp) is needed for multithreading, without it training runs single-threaded with a warning.

  • clang or gcc can be installed using:

    # Arch
    sudo pacman -S clang # or gcc
    
    # Ubuntu
    sudo apt install clang # or gcc
    
    # macOS should have clang by default, if not:
    xcode-select --install
    
    # Pixi, if using pixi to manage python environment
    pixi add clang

  • OpenMP can be installed using:

    # Arch (clang)
    sudo pacman -S openmp
    
    # Ubuntu (clang)
    sudo apt install libomp-dev
    
    # macOS (Homebrew, https://brew.sh)
    brew install libomp
    
    # Pixi
    pixi add llvm-openmp

• CUDA: Optional, not installed by default.

  • Use the cuda extra, it installs cupy-cuda12x[ctk], which bundles the CUDA runtime and NVRTC (kernels are compiled at runtime), so no system CUDA toolkit is needed.
  • If facing problems with CUDA/cupy, don't use the cuda extra, and follow the CuPy installation instructions to get it working in the environment first.
  • Requires an NVIDIA GPU with a driver supporting CUDA 12 (check CUDA Version in nvidia-smi).

Local Development

The environment needed is defined in pyproject.toml.

git clone https://gitlab.com/cair/YTM.git && cd YTM
pixi install              # CPU, includes clang and llvm-openmp
pixi install -e cuda      # CUDA

# or with uv, needs a system C compiler with OpenMP (see Install)
uv sync --group dev                 # CPU
uv sync --group dev --extra cuda    # CUDA

Tests

pixi run pytest tests                     # CPU only
pixi run -e cuda pytest tests             # CPU + CUDA (CUDA tests skip if no usable GPU)
pixi run pytest tests -m statistical      # RNG and distribution/frequency tests
pixi run pytest tests -m sanitize         # ASan/UBSan, plus compute-sanitizer when CUDA is usable

Inspired by

This implementation is inspired by:

• TMU
• PySparseCoalescedTsetlinMachineCUDA