Graph Neural Networks for Graph Classification Benchmark

This project aims to compare the interactions between convolutional, pooling and readout layers on a graph classification task.

This is a simplified version of an original research project done for the M2DS course : ML research seminar (https://github.com/AntoineTSP/ML_research_seminar/tree/main).

To use training_script

python training_script.py -c configs/templates/MUTAG_test_template.yml

One has to properly fill the yml file. Don't forget to install the yaml module.

New layers of convolution or pooling have to be added to model\layer_selector.py.

New layers of local pooling should also contained a dictionary indicating the order of the variables returned by the forward of the pooling layer. For instance the MEWISPooling has the dictionary :

{'node_features':0,'edge_index':1,'batch':2, 'loss':3}

While the SAGPooling has the dictionary :

{'node_features':0,'edge_index':1,'batch':3}

Since the forward method of SAGPooling returns an edge_attributes values at the second place (which is useless to us), and similarly the MEWISPooling returns a loss at the third place that has to be taken into account.

Multiple models training

python pipeline.py

One need to place the basic config templates inside the folder for each different dataset

To do list

Remove the data from the git and put it on the cloud Create loggers (instead of print) Create a general config file which will monitor all the paths