Auto-Generating Code Algorithm Based On Deep Learning

Coding according to design drawings is a typical task for developers to build websites and mobile applications for customers,but coding is a very tedious and repetitive task,which will consume a lot of time and energy of programmers.Pix2 code is an end-to-end deep learning model framework,generating code automatically by inputting a graphical user interface with a correct rate of 77%.But the pix2 code model is composed of relatively simple algorithms,and the data set is too small to be put into practical application.In this thesis,we firstly discuss the traditional automatic code generation technology and the basic theory of deep learning,then try to use more complex algorithms and some regularization methods to improve the model.Pix2 code is mainly divided into three modules,visual model utilizing Convolutional Neural Networks,language model and decoder layer utilizing Long Short-Term Memory,combining the advantages of the two algorithms to generate code.The correct rate of test set has increased from 77% of the original model to 85% in new algorithms.The results also show that:(1)Residual Neural Network has a powerful ability to extract features,which improves the stability of the overall system,enhances the generalization ability,but the increasing of the accuracy of the test set less obviously.Without increasing the number of training epochs,the accuracy of the train set has been greatly improved,which increases the possibility of improvement of the system in the future.(2)After Bi-directional Long Short-Term Memory optimize the visual model and decoder layer,the robustness of the system is improved obviously,the model becomes more stable,and the accuracy of the test set is greatly improved.Regularization can also alleviate over-fitting.Adding dropout regularization to the decoder layer of the original system has no obvious effect.So Bi-directional Long Short-Term Memory has obvious advantages in dealing with such contextual content tasks with obvious structure and content relevance.Although the generation of code based on deep learning algorithm is still in the initial stage of research,this study has shown great potential.With the continuous research,ultimately,it will become a reality to write large-scale programs that can run online by computer system entirely.