Installation¶
OpenSportsLib requires Python 3.12 or newer. Install the package first, then run its setup command to install a PyTorch profile for the current machine. The setup command replaces the installed Torch stack; use a dedicated environment.
Base installation¶
conda create -n osl python=3.12 pip
conda activate osl
python -m pip install --upgrade pip
pip install opensportslib
opensportslib setup
For the latest prerelease, use pip install --pre opensportslib. For source development, clone the repository and replace the install command with pip install -e ..
opensportslib setup detects nvidia-smi when available and installs a CPU wheel when it is not. It selects among CUDA 12.6, 12.8, and 13.0 wheel profiles from the reported driver and visible GPU compute capabilities; it is not a general CUDA toolkit installer.
Optional profiles¶
| Command | Use when | Important behavior |
|---|---|---|
opensportslib setup --dali |
Using DALI video loading on an NVIDIA GPU | Installs DALI/CuPy for the selected CUDA profile; DALI is not a CPU loader. |
opensportslib setup --pyg |
Using graph/tracking models | Replaces Torch with the pinned PyG-compatible Torch 2.12.1 profile and installs torch-geometric. |
opensportslib setup --pyg --pyg_extensions |
Optional compiled PyG extensions are required | Verifies binary wheels before installing extension packages. |
opensportslib setup --vqa_xvars |
Running X-VARS VQA | Replaces the Hugging Face dependency set with X-VARS pins. |
opensportslib setup --vqa_qwen |
Running Qwen VQA | Replaces the Hugging Face dependency set with Qwen pins. |
The X-VARS and Qwen profiles have incompatible dependency pins; use separate environments when both are needed. On CPU, compatible DALI-video configs are normalized to the OpenCV loader.
Verify and authenticate¶
python -c "from opensportslib.apis import ClassificationModel, LocalizationModel, VQAModel, Config; print('OpenSportsLib ready')"
opensportslib setup --help
Hugging Face access is required only for gated datasets or model repositories. Authenticate with hf auth login before such a workflow. The separately packaged inference server is not installed by pip install opensportslib; see the server guide.