Fast Graph Representation Learning with PyTorch Geometric

The paper introduces PyTorch Geometric, a library built on top of PyTorch for deep learning on irregularly structured input data such as graphs, point clouds, and manifolds. Beyond general-purpose graph data structures and processing methods, it packages a range of recently published techniques from relational learning and 3D data processing, giving practitioners a single framework for geometric deep learning.

To keep such workloads efficient, the library leverages sparse GPU acceleration, ships dedicated CUDA kernels, and introduces efficient mini-batch handling for input examples of differing sizes, yielding high data throughput. The authors describe the library in detail and carry out a comprehensive comparative study of its implemented methods under homogeneous evaluation scenarios, providing both a usable toolkit and a consistent benchmark across models.