GSO:A GNN-based Deep Learning Computation Graph Substitutions Optimization Framework

• Published in: Ji suan ji ke xue Vol. 49; no. 3; pp. 86 - 91
• Main Authors: Miao, Xu-peng, Zhou, Yue, Shao, Ying-xia, Cui, Bin
• Format: Journal Article
• Language: Chinese
• Published: Chongqing Guojia Kexue Jishu Bu 01.03.2022; Editorial office of Computer Science
• Subjects: Algorithms; Computer networks; Deep learning; Graph matching; Graph neural networks; Graph theory; Graphs; Heuristic; Heuristic methods; Machine learning; Neural networks; Optimization; optimizing dnn computation|graph substitutions|deep learning|graph neural network; Searching; Substitutes; Topology
• ISSN: 1002-137X
• DOI: 10.11896/jsjkx.210700199

Deep learning has achieved great success in various practical applications.How to effectively improve the model execution efficiency is one of the important research issues in this field.The existing deep learning frameworks usually model deep learning in the form of computational graphs, try to optimize computational graphs through subgraph substitution rules designed by experts and mainly use heuristic algorithms to search substitution sequences.Their shortcomings mainly include: 1)the exis-ting subgraph substitution rules result in a large search space and the heuristic algorithms are not efficient; 2)these algorithms are not scalable for large computation graphs; 3)cannot utilize the history optimization results.In order to solve the above problem, we propose GSO,a graph neural network-based deep learning computation graph optimization framework.We transfer the graph substitution optimization problem as the subgraph matching problem.Based on the feature information from the operators and the computation g