Search

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 graph topology,we utilize the graph neural network to predict the subgraph matching feasibility and positions.We implement the framework using Python,which is compatible with the mainstream deep learning systems.The experimental results show that:1)compared to the total graph substitution rules,the proposed rule can reduce the search space by up to 92%;2)compared to the existing heuristic algorithms,GSO can complete the subgraph replacement process of the computational graph 2 times faster.The optimized computation graph is up to 34% faster the original graph.