Boost Your Medical Deep-Learning Training By Lazy Loading

Chenglong Wang, Chengxiu Zhang, Yun Liu, Guang Yang

Keywords: Deep Learning Efficiency, Memory-Mapping, Data I/O Optimization

Abstract: In recent years, the growing volume size of medical datasets has posed a significant challenge for medical deep learning training pipelines, often leading to inefficiencies stemming from data I/O bottlenecks. Addressing this issue, we present a simply yet effective trick, lazy loading strategy, leveraging memory-mapping mechanisms to boost training processes. By dynamically loading only the target slices of large medical datasets into active memory, our method minimizes the reading time and conserves memory. This paper mainly aims to remind community to realize the advantages of the lazy loading strategy, which could substantially boost the efficiency of deep learning training process in the medical domain.

Submission Number: 26