Design And Implementation Of Sampling Strategy For Graph Neural Network Algorithm Library

In recent years,graph neural network algorithms have been widely used in many fields.With the deepening of research on graph-related algorithms,developers and researchers seek open source graph neural network algorithm libraries to quickly develop and test new graph neural network algorithms,reducing RD costs and development time.Pengcheng Lab aims to use the deep learning framework TensorLayerX to develop an autonomous and controllable domestic open source graph neural network algorithm library GammaGL(Gamma Graph Library).The research focus of this thesis is to design and implement the sampling module and graph distributed training module in the GammaGL algorithm library.Firstly,the important role of graph sampling method and graph distributed training method in graph neural network algorithm is introduced,and then various graph sampling strategies are studied,focusing on the research of graph neighbor sampling algorithm.In addition,optimization methods for graph distributed training are investigated.Finally,a variety of graph sampling strategies and optimization methods for graph distributed training are integrated and encapsulated in the algorithm library.For the research on the graph neighbor sampling algorithm,this paper uses the sparse matrix storage graph structure in the CSR(Compressed Sparse Row)format,and uses the sampling method with/without replacement and the node renumbering method with lower time complexity to speed up the node renumbering method.The acquisition of neighbors improves the sampling speed of the graph neighbor sampling algorithm.For the optimization of graph distributed training in the algorithm library,this paper focuses on the optimization of the parallel training part in graph distributed training,using the method of GPU sampling and feature extraction based on UVA(Unified Virtual Addressing)technology,so that the training speed can increase with The increase of computing resources increases linearly.Finally,the requirements of graph sampling and graph distributed training in the graph neural network algorithm library are analyzed,and sampling algorithms corresponding to various graph sampling strategies and optimization methods for graph distributed training are designed and implemented,which are packaged and integrated into graph In the neural network algorithm library,and write the corresponding algorithm model to test it,and verify the usability.