saltedfishlz / torchstream Goto Github PK

Hi, all

Although the original Kinetics400 dataset provides MP4 videos, which can work correctly, I found that decoding mp4 files consumes a lot of CPU resources and makes data preprocessing the main performance bottleneck.
AVI files usually require more disk space than MP4 files. But they are easier to decode and won't make data preprocessing the performance bottleneck. There is a trade-off here. And I hope that we can support AVI version of Kinetics.

This feature requires:

• supporting multiple file types in datasets (it requires collecting multiple types of metadata)

• an AVI version of the dataset (I have a simple map-reduce library in torchstream which can be used here)

Best Regards
Zheng Liang