A mature machine learning pipeline includes components, such as feature engineering, model training, hyperparameter tuning, and model serving. With huge recommendation models with sparse input data available in Angel 2.x, this time, our new Angel 3.0, aiming at a full-stack machine learning platform, further completes the other components. First, the auto feature engineering (AFE) is supported. Second, we provide a type of auto hyperparameter tuning based on Bayesian optimization. Third, we also provide a cross-platform model serving system. It can serve the models from Angel, Spark, XGBoost, and PyTorch. Apart from completing the pipeline, a new PyTorch engine for Angel is introduced. PyTorch is used for forward and backward propagation to obtain gradients, while Angel parameter server stores, synchronizes and updates parameters. Consequently, we provide a variety of graph embedding and GNN algorithms. Moreover, we make Spark ON Angel adapt to Spark 2.4 and support Kubernetes. Hence, the DataFrame API and Spark Pipeline are supported.
