chengtan9907/OpenSTL

Implements a modular three-layer architecture (core, algorithm, user interface) supporting 20+ spatio-temporal prediction methods across diverse tasks including video prediction, weather forecasting, traffic flow, and human motion. Offers unified training APIs with flexible method decomposition into trainable strategies, network architectures, and reusable modules, plus PyTorch Lightning support for streamlined experiment management. Includes standardized benchmarks across datasets like Moving MNIST, KTH, and real-world scenarios with pre-trained model zoo on Hugging Face.

1,075 stars.