Research On RNN·Based Automatic Code Generation And Visual Analysis

In recent years,with the rapid development of the economy,the advent of the era of big data,and the enhancement of computing power,artificial intelligence has developed rapidly.As a representative of the new generation of artificial intelligence technology,deep learning has made great breakthroughs in many fields or tasks.Automatic code generation has been a task with research value and challenges in the computer field for many years.The current research focus has also shifted to methods based on deep learning,and has made great improvements in terms of code generation quality and portability.However,the inherent black box characteristic of deep learning makes it difficult to explain the working mechanism of automatic code generation model,and it is difficult to further optimize and improve the model in a targeted manner.Combining the above issues,this paper uses Char-RNN-based code to automaticallygenerate models as the basis,and uses visual analysis to study the changes in the state of neurons in the recurrent neural network to explore the working mechanism of the model and improve the interpretability of deep learning models.Finally optimize the model through the results of visual analysis.The main work of this paper:First,establish a Char-RNN language model,train with C and Python code sets as data sets,and save the state of neurons during training;then use the visual analysis tool LSTMVis to train the process The state changes of the neurons are visually displayed and analyzed from multiple angles.Finally,the Dropout optimization algorithm is selected based on the analysis results.The Dropout algorithm is improved,and the model is optimized to achieve better results.Through the visual analysis tool LSTMVis,the programming language's if syntax andprint()function are used as observation points.In the experiments using the converged model as the experimental object,it was found that there is a strong relationship between the specific neurons in the model and the prediction of if syntax points or print()function points;In experiments using the same type of model at different training time points as experimental objects,it was found that there are neurons inside the model that continue to affect the prediction of specific observation points throughout the training.Finally,based on these findings,specific improvements were made to the Dropout optimization algorithm,and compared with the original Dropout algorithm and parameter optimization,the code generation results were improved.