PyTorch Geometric is a deep learning library for irregularly structured data such as graphs, point clouds, and manifolds, built on PyTorch. It provides a unified framework for various graph-related methods, including convolutional and pooling layers, and supports both CPU and GPU computations. The library uses efficient mini-batch handling, sparse GPU acceleration, and dedicated CUDA kernels to achieve high data throughput. It also includes a MessagePassing interface for rapid prototyping of new research ideas. The library supports a wide range of graph neural network (GNN) methods, including GCN, GAT, GIN, APPNP, DNA, and relational GCN. It also provides methods for point cloud and manifold learning, such as PointNet++, PointCNN, MPNN, MoNet, SplineCNN, and EdgeCNN. Additionally, it includes high-level implementations for tasks like maximizing mutual information, autoencoding graphs, aggregating jumping knowledge, and predicting temporal events in knowledge graphs. PyTorch Geometric supports graph-level outputs through various readout functions, including global pooling, set-to-set, sort pooling, and global soft attention. It also provides hierarchical pooling methods, such as Graclus, voxel grid pooling, iterative farthest point sampling, and differentiable pooling mechanisms like DiffPool and top-k pooling. The library supports a variety of benchmark datasets, including citation networks, social networks, bioinformatics datasets, and 3D object datasets. It includes over 60 graph kernel benchmark datasets, such as PROTEINS, IMDB-BINARY, Cora, CiteSeer, PubMed, and Cora-Full, as well as molecule datasets, protein-protein interaction graphs, and temporal datasets. Empirical evaluations show that PyTorch Geometric achieves high performance, with models trained up to 40 times faster than the Deep Graph Library (DGL) v0.2. It also provides efficient runtime for various GNN methods, including GAT, with up to 7 times faster performance using optimized sparse softmax kernels. The authors conclude that PyTorch Geometric is a powerful framework for fast representation learning on graphs, point clouds, and manifolds, and they invite researchers and software engineers to collaborate in extending its scope.PyTorch Geometric is a deep learning library for irregularly structured data such as graphs, point clouds, and manifolds, built on PyTorch. It provides a unified framework for various graph-related methods, including convolutional and pooling layers, and supports both CPU and GPU computations. The library uses efficient mini-batch handling, sparse GPU acceleration, and dedicated CUDA kernels to achieve high data throughput. It also includes a MessagePassing interface for rapid prototyping of new research ideas. The library supports a wide range of graph neural network (GNN) methods, including GCN, GAT, GIN, APPNP, DNA, and relational GCN. It also provides methods for point cloud and manifold learning, such as PointNet++, PointCNN, MPNN, MoNet, SplineCNN, and EdgeCNN. Additionally, it includes high-level implementations for tasks like maximizing mutual information, autoencoding graphs, aggregating jumping knowledge, and predicting temporal events in knowledge graphs. PyTorch Geometric supports graph-level outputs through various readout functions, including global pooling, set-to-set, sort pooling, and global soft attention. It also provides hierarchical pooling methods, such as Graclus, voxel grid pooling, iterative farthest point sampling, and differentiable pooling mechanisms like DiffPool and top-k pooling. The library supports a variety of benchmark datasets, including citation networks, social networks, bioinformatics datasets, and 3D object datasets. It includes over 60 graph kernel benchmark datasets, such as PROTEINS, IMDB-BINARY, Cora, CiteSeer, PubMed, and Cora-Full, as well as molecule datasets, protein-protein interaction graphs, and temporal datasets. Empirical evaluations show that PyTorch Geometric achieves high performance, with models trained up to 40 times faster than the Deep Graph Library (DGL) v0.2. It also provides efficient runtime for various GNN methods, including GAT, with up to 7 times faster performance using optimized sparse softmax kernels. The authors conclude that PyTorch Geometric is a powerful framework for fast representation learning on graphs, point clouds, and manifolds, and they invite researchers and software engineers to collaborate in extending its scope.