Open Source Summit Japan & Automotive Linux Summit 2020

Deep learning technologies for images are adopted in various kinds of business applications such as autonomous driving and cancer prediction. Preprocessing massive amounts of images in a short time is necessary for building models that are applied to business applications. Popular deep learning frameworks such as PyTorch offer their preprocessing libraries such as torchvision, which are compatible with those deep learning frameworks.

We evaluated the performance of preprocessing libraries and the effect of the preprocessing tuning in deep learning frameworks.

We adopted MLPerf, which is the popular benchmark for machine learning applications, and confirmed that parallelization and queueing could reduce the preprocessing time dramatically in both torchvision and OpenCV case.

However, in PyTorch framework, preprocessing and training are conducted in parallel, the effect of the preprocessing tuning can be small when the training workload is larger than the preprocessing workload.

We concluded that whether we should tune preprocessing or training depends on the purpose of the machine learning system such as parameter search and transfer learning.