Design And Implementation Of Research-process-oriented Task Universal Deep Learning Framework

With the rapid development of deep learning technologies,many deep learning frameworks have been proposed,such as PyTorch,TensorFlow,Keras,etc.These frameworks are designed to provide researchers and engineers with a fast,convenient,and efficient deep learning APIs,which can better promote the research.However,due to the complexity of the framework,the different operating mechanisms,and the lack of related auxiliary research tools,the deep learning researchers need to undertake the long and boring auxiliary work.Therefore,this paper proposes a research-oriented task universal deep learning framework based on PyTorch,named Jdit.Generally,deep learning works are relatively single and fixed,we implement several common task type templates in this framework.Selecting a corresponding template can reduce the work of training process,and the template can be customized and reused on others similar tasks with no pain.In addition,the framework also defines indispensable auxiliary tools for research,such as the data storage and visualization of loss curves,models,and training processes with the help of TensorBoard.Jdit decouples the core work of researchers from peripheral auxiliary work,enabling researchers to focus on the core content and improve research efficiency.At the end of this article,we use an example of electron microscope image research to illustrate the use of Jdit and its performance in actual projects.Jdit will also be a faster,convenient and efficient deep learning research tool proposed by future researchers.