I’m working on a video classif problem where dataset is a sequence of jpg images. Is it possible to pack that in recordio? the im2rec tool seems designed for single images (.lst files map to one picture per row)
Yes im2rec is designed for single images but if you make sure that it contains the correct order of images, then it should work for videos, too.
I looked into the recommended best practices for loading of video datasets: Nvidia provides the library NVVL that accelerates ML on video datasets. According to the Github repo NVVL is part of Nvidia’s Data Loading Library (DALI). DALI supports MXNet, so it might be worthwhile using NVVL instead of im2rec.