GCINT: Dynamic Quantization Algorithm for Training Graph Convolution Neural Networks Using Only Integers

Qizhe Wu, Letian Zhao, Huawen Liang, Xiaotian Wang, LinFeng Tao, Teng Tian, Tingxin Wang, Zerong He, Wei Wu, Xi Jin

Keywords: Quantization, Graph Neural Networks, Acceleration Training, Integers Networks

TL;DR: Robust and Adaptive Quantization Algorithm for Training Graph Convolution Neural Networks Using Only Integers

Abstract: Quantization approaches can minimize storage costs while decreasing the computational complexity of a model, although there is minimal study in the GNN field on quantization networks. We studied the four primary reasons why existing quantization approaches cannot be employed extensively with GNNs: (1)Quantifying the distinctions between data sources; (2)Quantifying the distinctions between data streams; (3)Quantifying the distinctions between concentrations; (4)QAT’s Limitations. Based on this, we propose GCINT, which is an efficient quantization framework prepared for GNN training. The entire forward, backward, optimizer, and loss functions are calculated using integer data. We achieved a training acceleration ratio of nearly 10× compared to FP32 Cuda Core in RTX 2080TI INT8 Tensor Core. Our quantization is independent of the dataset and weight distribution, and more than 2,000 randomized trials have been undertaken on the 8 popular GNN benchmark datasets, with all achieving errors within 1% of the FP32.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Deep Learning and representational learning