@article{Parés Pont_Megias_Garcia-Gasulla_Garcia-Gasulla_Ayguadé_Labarta_2021, title={Size & Shape Matters: The Need of HPC Benchmarks of High Resolution Image Training for Deep Learning}, volume={8}, url={https://superfri.org/index.php/superfri/article/view/365}, DOI={10.14529/jsfi210103}, abstractNote={<p>One of the purposes of HPC benchmarks is to identify limitations and bottlenecks in hardware. <span style="font-size: 10px;">This functionality is particularly influential when assessing performance on emerging tasks, </span><span style="font-size: 10px;">the nature and requirements of which may not yet be fully understood. In this setting, a proper </span><span style="font-size: 10px;">benchmark can steer the design of next generation hardware by properly identifying said requirements, </span><span style="font-size: 10px;">and quicken the deployment of novel solutions. With the increasing popularity of deep </span><span style="font-size: 10px;">learning workloads, benchmarks for this family of tasks have been gaining popularity. Particularly </span><span style="font-size: 10px;">for image based tasks, which rely on the most well established family of deep learning models: </span><span style="font-size: 10px;">Convolutional Neural Networks. Significantly, most benchmarks for CNN use low-resolution and </span><span style="font-size: 10px;">fixed-shape (LR&FS) images. While this sort of inputs have been very successful for certain purposes, </span><span style="font-size: 10px;">they are insufficient for some domains of special interest (e.g., medical image diagnosis or </span><span style="font-size: 10px;">autonomous driving) where one requires higher resolutions and variable-shape (HR&VS) images to </span><span style="font-size: 10px;">avoid loss of information and deformation. As of today, it is still unclear how does image resolution </span><span style="font-size: 10px;">and shape variability affect the nature of the problem from a computational perspective. In this </span><span style="font-size: 10px;">paper we assess the differences between training with LR&FS and HR&VS, as means to justify </span><span style="font-size: 10px;">the importance of building benchmarks specific for the latter. Our results on three different HPC </span><span style="font-size: 10px;">clusters show significant variations in time, resources and memory management, highlighting the </span><span style="font-size: 10px;">differences between LR&FS and HR&VS image deep learning.</span></p>}, number={1}, journal={Supercomputing Frontiers and Innovations}, author={Parés Pont, Ferran and Megias, Pedro and Garcia-Gasulla, Dario and Garcia-Gasulla, Marta and Ayguadé, Eduard and Labarta, Jesús}, year={2021}, month={Apr.}, pages={28–44} }