The traffic prediction library trying to end benchmark chaos

What it does

LibCity is a PyTorch-based framework that collects traffic prediction tasks—traffic flow, speed, ETA, map matching, trajectory prediction, and more—into a single modular pipeline. It provides standardized data storage via its own atomic files format, unified model interfaces, and evaluation metrics, aiming to replace the usual scatter of one-off research scripts. The library currently hosts 74 reproduced models across 9 tasks and 52 datasets, treating traffic research less like a custom art project and more like an assembly line.

The interesting bit

The real value is the bureaucracy: by forcing every model through the same configuration, data, and evaluation modules, LibCity attempts to solve the reproducibility problem that plagues spatial-temporal research. It even spawned a MindSpore twin, M-LibCity, suggesting the authors believe the pipeline matters more than the backend.

Key highlights

• 74 reproduced models covering traffic state, trajectory, ETA, map matching, and road network representation learning
• 52 built-in datasets with a unified atomic files storage format
• Modular five-part architecture: configuration, data loading, model execution, evaluation, and training
• TensorBoard integration for loss visualization during training
• A separate web-based experiment management tool for visual workflow control

Caveats

• Installation is source-only; there is no PyPI package, so you will be cloning and configuring manually
• Datasets must be downloaded separately from BaiduDisk or Google Drive and converted to the atomic files format before use
• The README is heavy on news and paper citations but light on concrete benchmark comparisons between the included models

Verdict

Researchers in urban computing and traffic forecasting who are tired of rewriting data loaders should look here; practitioners seeking a polished, pip-installable production tool will find it too academic.