Pipeline Parallelism in Distributed Deep Learning for Diabetic Retinopathy Classification

Deep learning (DL) has been used widely and successfully in a variety of industries. It is challenging for researchers to carry out their experiments to find the optimal learning model because training the deep learning model is a time-consuming operation. Distributed deep learning offers a technique for rapid training of DL models (DDL). It is difficult to create and implement DDL, nevertheless, so it is important to report on the performance analysis of different DDL algorithms for the benefit of the research community. The authors of this proposed study use two GPUs to design, construct, and analyse a pipeline parallelism in DDL and then present an analysis of its performance. As a case study for DDL implementation, the authors employ diabetic retinopathy classification. 0.834 is obtained as the validation F1 score. The performance of DDL is reported and examined by the authors, and it is demonstrated to improve training effectiveness in terms of both time and speed.