EXACT: Scalable Graph Neural Networks Training via Extreme Activation Compression

Authors: Zirui Liu, Kaixiong Zhou, Fan Yang, Li Li, Rui Chen, Xia Hu

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 5 EXPERIMENTS The experiments are designed to answer the following research questions. RQ1: How effective is EXACT in terms of model performance at different compression rates (Section 5.1)? RQ2: Is the training process of deeper GNNs also robust to the noise introduced by EXACT (Section 5.1)? RQ3: How sensitive is EXACT to its key hyperparameters (Appendix I.4)? RQ4: What is the running time overhead of EXACT (Section 5.2)? RQ5: Is the convergence speed of GNNs impacted by EXACT (Appendix I.3)? RQ6: To what extent can EXACT reduce the hardware requirement for training GNNs on large graphs (Section 5.3)?
Researcher Affiliation Collaboration Zirui Liu1, Kaixiong Zhou1, Fan Yang1, Li Li2, Rui Chen2 , and Xia Hu1 1Rice University, 2Samsung Research America
Pseudocode Yes Algorithm 1: Forward Pass of the lth GCN layer
Open Source Code Yes The code is available at https://github.com/warai-0toko/Exact.
Open Datasets Yes To evaluate the scalability of EXACT, we adopt five common large-scale graph benchmark datasets from different domains. Namely, Reddit, Flickr, Yelp, ogbn-arxiv, and ogbn-products. ... Table 10: Dataset Statistics. (contains URLs for each dataset)
Dataset Splits Yes We report the test accuracy associated with the highest validation score.
Hardware Specification Yes To answer RQ4, we compare the training throughput of EXACT with the baseline using a single RTX 3090 (24GB) GPU. ... train a full-batch Graph SAGE on ogbn-products on a single GTX 1080 Ti with 11GB memory
Software Dependencies Yes Table 9: Package configurations of our experiments. Package Version CUDA 11.1 pytorch sparse 0.6.12 pytorch scatter 2.0.8 pytorch geometric 1.7.2 pytorch 1.9.0 OGB 1.3.1
Experiment Setup Yes G. EXPERIMENT SETTINGS ... G.2 MODEL HYPERPARAMETER CONFIGURATIONS OF TABLE 1 AND TABLE 2 ... Table 11: Training configuration of Full-Batch GCN, Graph SAGE, and GAT in Table 1 and Table 2. ... Table 13: Training configuration of Cluster-GCN in Table 3.