Reproduce “Deep Graph Library: A Graph-Centric, Highly-Performant Package for Graph Neural Networks” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Minjie Wang, Da Zheng, Zihao Ye, Quan Gan, Mufei Li, Xiang Song, Jinjing Zhou, Chao Ma and 7 more

The Deep Graph Library (DGL) is an open-source framework designed to bridge the gap between graph-centric GNN computations and tensor-centric deep learning frameworks, providing a high-performance, memory-efficient, and framework-neutral solution. It achieves substantial speedups, such as 1.68x faster GAT training on OGBN-ARXIV and up to 64x faster CPU performance over PyTorch-Geometric, while significantly reducing memory consumption.

Key insights

  • All GNN message-passing operations, including their gradients, can be systematically unified and optimized using generalized sparse matrix multiplication (g-SpMM) and generalized sampled dense-dense matrix multiplication (g-SDDMM).
  • Treating graphs as first-class citizens in a programming abstraction allows the system to apply transparent optimizations and manage complex data structures, simplifying GNN development.
  • A pragmatic framework-neutral design, where critical sparse tensor operations are owned by the library and non-critical functionalities are delegated, achieves high performance and broad interoperability across deep learning ecosystems.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.