Highlights

• Chronos: an overhaul of the previous Zouwu time-series analysis library, with:

  • Built-in support of ~100 algorithms for time series forecast (e.g., TCN, seq2seq, ARIMA, Prophet, etc.), anomaly detection (e.g., DBScan, AutoEncoder etc.), and feature transformations (using TSDataset).
  • Automatic tuning of built-in models (e.g., AutoProphet, AutoARIMA, AutoXGBoost, etc.) using AutoML
  • Simple APIs for tuning user-defined models (including PyTorch and Keras) with AutoML
  • Improved APIs, documentation, quick start examples, etc.

• Reference implementation of large-scale feature transformation pipelines for recommendation systems (e.g., DLRM, DIEN, W&D, etc.)

• Enhancements to Orca (scaling TF/PyTorch models to distributed Big Data) for end-to-end computer vision pipelines (distributed image preprocessing, training and inference); for more information, please see our CPVR 2021 tutorial.

• Initial Python and PySpark application support for PPML (privacy preserving big data and machine learning)