Resbench: an Analysis of Deep Learning Frameworks for Image Classification

School Name

South Carolina Governor's School for Science and Mathematics

Grade Level

12th Grade

Presentation Topic

Computer Science

Presentation Type

Mentored

Abstract

Deep learning frameworks are of particular note, as they have been shown to drastically alter the performance of Deep Learning while training, as well as the final result of the training, despite the theoretically identical math being done while training. This study expands upon the works of “DLBench: a comprehensive experimental evaluation of deep learning frameworks” by Radwa Elshawi, Abdul Wahab, Ahmed Barnawi & Sherif Sakr through testing only ResNet50v1.5 with the frameworks TensorFlow, PyTorch, and MxNet, without changing the architecture. This gives us less variation between framework tests, as the DLBench study used different architectures as well as different databases when testing, leaving the reason for changes in performance up to multiple indistinguishable factors. Our study performed tests using 2 Tesla V100 GPUs from Nvidia, and was evaluated with each framework on 2 datasets: CIFAR100 and ImageNet.

Location

HSS 209

Start Date

4-2-2022 9:30 AM

Presentation Format

Oral and Written

Group Project

Yes

COinS
 
Apr 2nd, 9:30 AM

Resbench: an Analysis of Deep Learning Frameworks for Image Classification

HSS 209

Deep learning frameworks are of particular note, as they have been shown to drastically alter the performance of Deep Learning while training, as well as the final result of the training, despite the theoretically identical math being done while training. This study expands upon the works of “DLBench: a comprehensive experimental evaluation of deep learning frameworks” by Radwa Elshawi, Abdul Wahab, Ahmed Barnawi & Sherif Sakr through testing only ResNet50v1.5 with the frameworks TensorFlow, PyTorch, and MxNet, without changing the architecture. This gives us less variation between framework tests, as the DLBench study used different architectures as well as different databases when testing, leaving the reason for changes in performance up to multiple indistinguishable factors. Our study performed tests using 2 Tesla V100 GPUs from Nvidia, and was evaluated with each framework on 2 datasets: CIFAR100 and ImageNet.