Skip to content

Installation

YAIB supports Python 3.10 to 3.13 on Linux, macOS and Windows. It is installed from source with uv, which reads pyproject.toml and the locked versions in uv.lock.

Install from source

Clone the repository. The newest features and fixes are on the development branch:

git clone https://github.com/rvandewater/YAIB.git
cd YAIB
git checkout development

Install uv if you don't have it:

curl -LsSf https://astral.sh/uv/install.sh | sh

Install Python and the project with its dependencies into .venv:

uv python install 3.12
uv sync

Check that the command-line interface works:

uv run icu-benchmarks train -h

Every command in this documentation runs through uv run, which uses the project's environment. You can also activate it with source .venv/bin/activate and drop the uv run prefix.

Optional dependencies

Command Adds
uv sync --extra imputation The hyperimpute and PyPOTS imputation baselines (MissForest, GAIN, MIWAE, MIRACLE, Sinkhorn, SoftImpute, BRITS, SAITS, Attention) and the SSSD models. About 100 extra packages, including TensorFlow.
uv sync --extra explain Captum, for the feature attributions of deep learning models with --explain-features.
uv sync --dev Tests, coverage, linting and pre-commit, for development.
uv sync --only-group docs The tools to build this documentation, see Documentation.

Platform notes

  • GPUs: on Linux and Windows, uv installs the CUDA build of PyTorch from the PyTorch index. YAIB lists the GPUs it finds at the start of each run. Pass --cpu to train on the CPU anyway.
  • macOS: uv installs the default PyPI build of PyTorch. On Macs with Metal Performance Shaders, set PYTORCH_ENABLE_MPS_FALLBACK=1 if a model uses an operation that MPS does not support.
  • Windows: replace the line continuation \ in the examples with ^ (Command Prompt) or ` (PowerShell).
  • Mamba: the Mamba model runs in plain PyTorch. On NVIDIA GPUs it uses the fused kernels of mamba-ssm if you install it separately.

PyPI

YAIB is also published on PyPI as yaib. The CLI reads its configs from the configs/ folder of the current directory, so for running experiments a clone of the repository is the simplest setup.