Reference Guide

23 Dell EMC Ready Solutions for AI Deep Learning with NVIDIA | v1.0
stress the I/O to be able to demonstrate notable difference in performance between the storage options. But, it
turns out that the performance with the TFRecrods database is much better than using raw JPEG images. The
performance delta between these two different image formats is much larger in the less complicated neural
networks like AlexNet. For instance, the performance advantage when using TFRecords over raw JPEG images
is ~23%-40% for VGG16, but around 5x (500%!) for AlexNet. This is because the TFRecords format packs
many raw JPEG images together making the data access more efficient, and additionally, TensorFlow includes
input pipeline optimization for TFRecords format but no corresponding optimization API for raw images.
Because of this reason, the performance variation with JPEG images in different storages is quite large,
especially in VGG16. Profiling and further exploration is ongoing to understand the variation with JPEG images.
The recommended format to use would be TFRecords.
(a) AlexNet
(b) Resnet50