AutoGL

AutoGL is an open source automated machine learning toolkit for graph data, developed by the Tsinghua University MN Lab (THUMNLab). It lets researchers and developers run AutoML on graph datasets and tasks: rather than hand picking a graph neural network and tuning it, you hand AutoGL a dataset and a task and its solvers search for a strong pipeline automatically. The library is aimed at researchers and machine learning engineers working on node classification, graph classification and related problems. Datasets are handled through AutoGL Dataset, which builds on PyTorch Geometric or the Deep Graph Library, either of which can serve as the backend. Each solver draws on a set of modules: an auto feature engineer that adds, removes and combines attributes and topological features; neural architecture search with several search spaces, algorithms and estimators; an auto model module holding common graph deep models with their hyperparameter spaces; hyperparameter optimization; and auto ensemble to combine the best models found. It installs with pip, from conda or from source, and requires Python 3.6 or later and PyTorch 1.6 or later. AutoGL is free under the Apache 2.0 license. There is no commercial edition or pricing. The current documentation is for a 0.3.0 release candidate, and the project describes itself as actively under development. It is a research library rather than a hosted service, so it assumes comfort with Python and graph learning frameworks.