PyTorch Geometric (PyG) is a library designed for deep learning on irregularly structured data such as graphs, point clouds, and manifolds, built on top of PyTorch. It offers a comprehensive set of tools and methods for graph representation learning, including various convolutional and pooling layers, neighborhood aggregation schemes, and global pooling techniques. PyG leverages sparse GPU acceleration and dedicated CUDA kernels to achieve high data throughput, making it suitable for large-scale and complex datasets. The library supports both CPU and GPU computations and provides a flexible and efficient framework for researchers and practitioners. The paper presents a detailed overview of PyG, including its architecture, implementation details, and a comprehensive comparative study of the implemented methods in homogeneous evaluation scenarios. The evaluation covers semi-supervised node classification, graph classification, and point cloud classification, demonstrating the library's performance and versatility.PyTorch Geometric (PyG) is a library designed for deep learning on irregularly structured data such as graphs, point clouds, and manifolds, built on top of PyTorch. It offers a comprehensive set of tools and methods for graph representation learning, including various convolutional and pooling layers, neighborhood aggregation schemes, and global pooling techniques. PyG leverages sparse GPU acceleration and dedicated CUDA kernels to achieve high data throughput, making it suitable for large-scale and complex datasets. The library supports both CPU and GPU computations and provides a flexible and efficient framework for researchers and practitioners. The paper presents a detailed overview of PyG, including its architecture, implementation details, and a comprehensive comparative study of the implemented methods in homogeneous evaluation scenarios. The evaluation covers semi-supervised node classification, graph classification, and point cloud classification, demonstrating the library's performance and versatility.