Research And Application Of Automatic Code Generation Technology Based On Neural Network

In recent years,with the advent of the Internet revolution,more and more enterprises have joined the Internet industry after the transition,therefore,the demand for developers in the society is increasing year by year.Among such a large number of developers,many people are engaged in the task of completing interface code according to the designer's planned graphical user interface.Different platforms have their own development languages,so cross-platform development is hard to realize,and a great amount of developers see repetition in the work being done.With the rapid development of artificial intelligence,deep learning has been successful in natural language processing and computer vision.Accordingly,using the excellent performance of deep learning in image and natural language,and sharing part of the work for developers becomes research hotspot.The system represented by the pix2code provides a successful case of automatically generating code according to graphical user interface.It makes a groundbreaking attempt in this field.However,there are still some shortcomings in the application.For example,in the process of the training,the visual model is placed in the network,and this way is easy to ignore the feature of the image itself.Moreover,the encoding of the language model cannot provide related information between words,so its application is restricted.Combined with the various issues raised above and relevant research,this thesis designs a code automatic generation model that can be used to generate code based on graphical user interface.The work of the thesis are as follows:firstly,the visual model is improved by using autoencoder based on deconvolution network.It places the encoding process of image part outside the network.The output coding vector not only meets the requirements,but also accurately covers the characteristic information of the network.Secondly,the language model adopts the word embedding.It transforms the words into low-dimensional continuous values,and further,the words with similar meanings are located at close positions in vector space,so that the potential connection between words can be mined.Finally,the process of inferring code uses the beam search.In the thesis,we study the value of search width,and analyze the effect of k on the generated code results.Experiments show that the model can generate effective code and show a relatively perfect graphical interface.Based on the mentioned model,the thesis constructs a front-end tool which can automatically generate code based on graphical user interface.Furthermore,the thesis analyzes and designs the overall architecture and sub-module functions of the system.