TorchSpatiotemporal/tsl

What it solves

It provides a unified framework for neural spatiotemporal data processing and forecasting, simplifying the transition from data preprocessing to model prototyping. It specifically addresses the complexity of handling spatiotemporal data, such as sensor networks and environmental data, by providing tools to manage irregularities, missing data, and varying network structures.

How it works

Built on PyTorch, PyTorch Geometric (PyG), and PyTorch Lightning, the library offers high-level APIs to create custom models and datasets. It includes specialized modular neural layers and streamlined preprocessing methods for scaling, resampling, and clustering time series. It also integrates with Hydra for reproducible experiments and leverages PyTorch Lightning for scalability across CPUs and GPUs.

Who it’s for

Researchers and developers working with spatiotemporal data mining and neural forecasting, particularly those using Graph Neural Networks (GNNs).

Highlights

• Comprehensive Ecosystem: Access to a wide range of existing spatiotemporal datasets and state-of-the-art baseline models.
• Robot-ready Data Handling: Tools to handle missing data and irregularities in data streams.
• Integrated Preprocessing: Built-in methods for time series scaling, resampling, and clustering.
• Scalable Infrastructure: Seamless integration with PyTorch Lightning for efficient training and inference across various hardware configurations.