GTC 2020 S21997
Presenters: Anas Abidin ,NVIDIA ; Christopher Bridge,Center for Clinical Data Science
Abstract
Deep dive into the various considerations that could be critical to substantially improve existing medical imaging workflows. We’ll motivate our discussions through a detailed analysis of two common medical imaging workflows: 2D classification and 3D segmentation. Our results suggest that controlling for aspects such as i/o format, selection of hyper-parameters, and mixed-precision training could all be key to maximizing hardware performance and reducing turnaround time for experiments. Furthermore, we aim to discuss other strategies, such as learning-rate warmup and scaling, effect of optimizers, scaling to multiple GPUs, and their potential effects on training throughput. We were able to show a 5x improvement for the 2D classification task through our ablation experiments. We’ll use these results suggest best practices learned, and also build a sizing calculator for providing quantitative insights for hardware investments.
Watch this session
Join in the conversation below.