Graph Ladling: Shockingly Simple Parallel GNN Training without Intermediate Communication

Authors: Ajay Kumar Jaiswal, Shiwei Liu, Tianlong Chen, Ying Ding, Zhangyang Wang

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our extensive experiments across many real-world small and large graphs, illustrate the effectiveness of our approach and point towards a promising orthogonal direction for GNN scaling.
Researcher Affiliation Academia Ajay Jaiswal 1 Shiwei Liu 1 Tianlong Chen 1 Ying Ding 1 Zhangyang Wang 1 *Equal contribution 1University of Texas at Austin. Correspondence to: Ajay Jaiswal <ajayjaiswal@utexas.edu>.
Pseudocode Yes Algorithm 1 Greedy Interpolation Soup Procedure Algorithm 2 Model soup with Graph Sampling Algorithm 3 Model soup with Graph Partitioning
Open Source Code Yes Codes are available at: https://github. com/VITA-Group/graph_ladling.
Open Datasets Yes For our small-scale experiments, we use three standard citation network datasets: Cora, Citeseer, and Pubmed, while our large-scale experiments are conducted with four popular large-scale open benchmarks: Flickr, Reddit, OGBN-Ar Xiv, and OGBN-products (Appendix A).
Dataset Splits Yes For our evaluation on our chosen datasets, we have closely followed the data split settings and metrics reported by the recent benchmark (Duan et al., 2022; Chen et al., 2021). ... Table 9. The searched optimal hyperparameters for all tested methods for data-centric soup. Split: 0.50/0.25/0.25 Split: 0.66 / 0.10 / 0.24 Split: 0.10 / 0.02 / 0.88
Hardware Specification Yes Our experiments are conducted on two GPU servers equipped with RTX A6000 and RTX 3090 GPUs.
Software Dependencies No The paper does not list specific version numbers for software dependencies used in the experiments. It mentions PyTorch Geometric in Appendix A for Flickr and Reddit datasets, but without a version number.
Experiment Setup Yes The hyperparameters for soup ingredients corresponding to different datasets training are selected via our built-in efficient parameter sweeping functionalities from pre-defined ranges nearby our optimal setting in Appendix B. ... we have used 50 candidate ingredients for small-scale graphs and 30 candidate ingredients for large-scale graphs experiments and our model soup is prepared using interpolation hyperparameter α [0 1] with a step size of 0.01 in Algorithm 1. ... Table 9. The searched optimal hyperparameters for all tested methods for data-centric soup. LR: 0.0001, WD: 0.0001, DP: 0.5, EP: 50, HD: 512, L: 4, BS: 1000