Installation¶

Use Python 3.10 or newer in a fresh environment. COATI is an independent package: its training engine is included, and the original TraInf repositories are not required.

Install directly from GitHub¶

python -m pip install "coati-trajectory[visualization] @ git+https://github.com/shankswang953/COATI.git"

This installs COATI plus AnnData and UMAP support. For training only, omit [visualization]; add notebooks to the extras if you use Jupyter. Import the package as coati. The distribution name is coati-trajectory; it is not yet published on PyPI.

Editable installation and notebooks¶

Clone the repository to obtain the example scripts and notebook files:

git clone https://github.com/shankswang953/COATI.git
cd COATI
python -m venv .venv
source .venv/bin/activate
python -m pip install -e '.[notebooks,visualization]'

On Windows activate with .venv\Scripts\activate. For training only, install -e .. AnnData and notebook support are optional extras; the core numerical dependencies install with the package. Import name: coati. Distribution name: coati-trajectory. No PyPI release is currently assumed.

CPU and GPU¶

The public default is device="cpu". Use Config(device="cuda:0") for a compatible CUDA environment, or device="mps" explicitly on supported Macs. device="auto" preserves the original CUDA/MPS/CPU selection. Only CPU has been checked in this release.

Website source¶

Website source, assets and publishing tools are maintained separately in COATI_website. See that repository for local preview instructions. They are not part of the COATI package installation.

Existing conda environments¶

Prefer the fresh environment above. Mixed conda/pip OpenMP runtimes can conflict on macOS. COATI does not set KMP_DUPLICATE_LIB_OK, change global tensor defaults or silently select a different solver to mask environment issues.

For the AnnData/UMAP velocity tutorial:

python -m pip install -e '.[visualization,notebooks]'

UMAP fitting is optional: plotting an existing embedding only needs the base plotting dependencies and AnnData if that input type is used.