I hope you have great start of 2021! This is my first post of new year!
For there are many data which can be represented as graph. So graph based deep learning(GL) is very interesting area. In chemistry area, molecule can be represented as graph so GL is also attractive method for chemoinformatics.
I posted some topics about GL with DeepGraphLibrary and torch_geometric. Both packages are very useful for chemoinformatics area.
And today I would like to introduce a new package for GL named AutoGL which is developed for researchers and developers to quickly conduct autoML on the graph datasets & tasks.
The package can find following URL.
To install AutoGL, torch_geometric(PyG) is required. But to use current version of AutoGL version of PyG should be <1.6.1.
This package provides Auto node_classification and graph_classification methods. In the chemoinformatics, molecule seems as graph, so I have an interest to graph_classification.
So I tried to use the package for molecule property prediction. Following code is my first code of this year ;) I uploaded my code to gist. Solubility data is used for my test.
There are several way to define auto solver, from config file, from config and ad hoc. I used from_config to define solver. Regarding the readme.md in original repo, for grraph classification GIN and TopKPool are supported but GCN is not supported.
My trial seems that my model is overfitting.
In summary AutoGL is an interesting and useful package for Graph Learning because it wrap torch_geometric. It means that user can optimize model very easily.
However if user can define model with torch_geometic directly optuna seems another useful package for automated GL optimization. :)