How Attentive are Graph Attention Networks?

The project is structured to facilitate the replication of experiments across different deep learning frameworks. Key subdirectories include `arxiv_mag_products_collab_citation2_noise` for node-prediction, link-prediction, and robustness to noise experiments; `proteins` for ogbn-proteins node-prediction results; `dictionary_lookup` for the DictionaryLookup benchmark; and `tf-gnn-samples` for VarMisuse and QM9 dataset experiments. This modular design allows users to easily navigate and utilize the code relevant to their specific research interests.

The GATv2 model, a core component of the research, is made accessible through integrations with popular graph neural network libraries. It is available within PyTorch Geometric, DGL, and Google's TensorFlow GNN library, with direct links provided for easy access to their respective documentation and implementations. This widespread availability ensures flexibility for users working with different preferred tools.

Requirements for running the code are generally consistent, with dependencies on PyTorch (version 1.7.1) and PyTorch Geometric (version 1.7.0). Specific subdirectories may have additional dependencies, such as DGL (version 0.6.0) for the `proteins` subdirectory or TensorFlow (version 1.13) for `tf-gnn-samples`. The project notes that experiments can be executed on either GPU or CPU hardware, offering accessibility for a wide range of computational resources.

This repository is invaluable for anyone interested in understanding the nuances of attention in graph neural networks, replicating cutting-edge research in the field, or applying advanced GATv2 models to their own graph-based machine learning tasks. The comprehensive code structure and framework integrations make it a robust resource for the graph learning community.