| Repo |     |
| Package |    |
| Testing |    |
| Contact |  |
HyperTorch is a library for hypergraph learning and benchmarking. It provides a standardized workflow for loading hypergraph datasets, training models, evaluating them under comparable settings, and reporting results. The current release focuses on Hyperlink Prediction, with ready-to-run pipelines for established hypergraph baselines.
The library is built around extensibility: datasets are represented in HIF format and converted into typed tensor objects, models can be implemented as standard Lightning modules, and benchmarking is handled through reusable trainers, samplers, metrics, loggers, and result exporters (Markdown/LaTeX). HyperTorch includes preloaded datasets, mini-batch and full-hypergraph data loading, negative sampling utilities, structural feature enrichers, neural components, and built-in models such as HGNN, HNHN, HyperGCN, GCN, MLP/SLP, NHP, Node2Vec, VilLain, and more.
Use HyperTorch to:
- Benchmark existing models across a shared collection of hypergraph datasets.
- Develop custom PyTorch or PyTorch Lightning models and train and compare them against the built-in baselines.
- Integrate new datasets through the HIF format and run the same training, evaluation, and reporting pipeline on them.
| Feature | What you can do | Highlights | Location |
|---|---|---|---|
| Dataset management | Load, process, and validate hypergraph datasets | HIF loader/processor, built-in datasets such as Algebra, Cora, Pubmed, DBLP, Amazon, and IMDB | hypertorch.data |
| Splitting, sampling, and batching | Prepare train/validation/test data and mini-batches | Dataset splitters, node and hyperedge samplers, negative samplers, data loaders | hypertorch.data |
| Feature enrichment | Enrich node and hyperedge features before training | Laplacian positional encodings, Node2Vec features, hyperedge weights and attributes | hypertorch.data |
| Neural components (NN) | Build models and pipelines | Layers, aggregators, losses, and activation/normalization helpers | hypertorch.nn |
| Models | Access hypergraph models | HGNN, HGNNP, HNHN, HyperGCN, GCN, MLP/SLP, NHP, Node2Vec, VilLain, CommonNeighbors | hypertorch.models |
| Hyperlink prediction (HLP) pipelines | Use ready-to-train hyperlink prediction modules | HLP modules with encoders, configs, losses, and stage metrics for multiple models | hypertorch.hyperlink_prediction |
| Node classification (NC) pipelines | Use ready-to-train node classification modules | NC modules with encoders, configs, losses, and stage metrics for multiple models | hypertorch.node_classification |
| Training and benchmarking | Train, compare, checkpoint, and report model runs | Multi-model trainer, schedulers, TensorBoard support, CSV/Markdown/LaTeX result tables | hypertorch.train |
HyperTorch can be installed from PyPI when you want to use it as a dependency, or from source when you want to contribute or run the latest repository version.
CI pipelines validate CPU installs on Python 3.10 through 3.14 for Linux x86_64, Linux ARM/aarch64, macOS arm64, and Windows x64. Install the matching PyTorch and PyG wheels for your platform (e.g., CUDA) before installing HyperTorch.
For more detailed instructions, see the installation guide.
git clone https://github.com/hypernetwork-research-group/hypertorch.git
cd hypertorch
make setupSee the installation guide for platform notes and dependency ranges.
To include TensorBoard support, also run HyperTorch install command with the TensorBoard extra:
uv pip install "hypertorch[tensorboard]"When installing from source, run the command:
make setup-tensorboardYou can download the examples directory and run the example scripts to get started.
With Python:
python3 examples/hyperlink_prediction/nhp.pyOr with uv:
uv run examples/hyperlink_prediction/nhp.pySee CONTRIBUTING.md for details on contributing to the project.
You can find the extensive documentation here.
Alternatively, you can build the documentation locally with the following commands:
make docs
# With explicit commands
uv run zensical build --clean -f zensical.toml
uv run zensical serve -f zensical.toml -a 127.0.0.1:8000and open the browser at http://localhost:8000 to access the documentation.
See LICENSE.
Most development discussions take place on GitHub in this repo, via the GitHub issue tracker.