Performance and Accuracy Evaluation of an Image Classifier using Multiple Machine Learning Models and Multiple GPUs

A performance and accuracy comparison is presented using six different machine learning models with five different NVIDIA GPUs to perform image classification of flower images. The image classifier is an adaptation based on the 2019 machine learning work by Bonner for flower image prediction. The image classifier was run employing six machine learning models using five types of GPUs accessed via Google Colaboratory and the University of Tennessee ISAAC Open OnDemand environment. In the Google Colaboratory environment, 16 GB NVIDIA T4 and P100 GPUs were used. A 32 GB NVIDIA V100S, 48 GB NVIDIA A40, and 80 GB NVIDIA H100 were used in the ISAAC Open OnDemand environment. Both environments were used to perform model load and training, perform image classifier runs, and generate results. Each environment that provided GPU access had its own challenges for access and running the models which will be discussed. The machine learning models each have their own advantages and disadvantages described in detail in their related publications with the focus of this work on results and analysis on accuracy of training, accuracy of flower image prediction, resource utilization, and performance. Our paper describes the image classifier, the machine learning approach using multiple models, GPUs used, analysis of performance, accuracy of results, conclusions, and suggestions for future work.