Visualization Techniques And Systems For Computational Graphs In Deep Learning

In recent years,deep learning has achieved rapid development and gradually developed into the mainstream research direction in the field of artificial intelligence.The problem with model complexity is the lack of interpretability of models,that is,they are black box models,which can be extremely confusing to developers,users,and regulators.Visual analysis is an analytical and reasoning science based on visual interactive interfaces.In recent years,great progress has been made in improving the interpretability of deep learning models.Visualization of deep learning computational graphs is an important technology combining visual analysis and deep learning,which enables users to intuitively explore the overall structure of the model and debug the model system.With the expansion of model size,the complexity of computational graphs increases exponentially,which brings two major challenges to the visualization of computational graphs.1.Poor layout effect and rendering performance: due to a large number of nodes and edge structures in the large model,existing computational graph visualization tools have many visual confusion problems such as edge crossing and scattered node arrangement when displaying multi-level computational graphs.Most of the current computational graph visualization tools are third-party libraries based on SVG,and the performance of visualizing large model is not high.2.Difficult to fit parallel training computational graph: multi-device parallel training is a crucial method to solve the bottleneck of large model training performance.Paral-lel training includes data parallelism,model parallelism,pipeline parallelism,etc.,which involves the segmentation of model or training data and the communication between par-allel devices,which further increases the complexity of computational graph structure.At present,there is no reasonable and effective visualization solution for parallel training computational graph.To solve the problem 1,a visualization system is developed based on Huawei's open source deep learning framework Mind Spore.Optimization methods such as edge binding and isomorphic subgraph stacking are used to simplify the presentation of computational graphs,and ELK's hierarchical orthogonal layout algorithm is used for layout,making the data flow of computational graphs more clear and readable.On the other hand,in order to solve the performance problem of the computational graph display,Web GL related technology is adopted to accelerate the rendering.Finally,the test results of classical models such as RESNET-50,Bert and VGG16 show that our method is superior to the existing methods in four aspects: the simplicity of graph structure visualization,the logical rationality of model structure,the fluency of front-end loading and the ease of user interaction.To solve the problem 2,a novel visualization solution of computational graph is proposed.The computational graph of the parallel training model is processed for a computationcommunication bipartite graph by using the minimum cut algorithm and hash algorithm.Stacking algorithm is used to merge the same subgraph structure.We use res NET-50 model under parallel training to test the system.Experimental results show that our method simplifies the presentation of model structure and is beneficial to locate communication operators and observe the fusion strategy of communication operators quickly.