Video Loader Benchmark

Key Insight

Most video-training pipelines are bottlenecked on decoding the video — turning a compressed file back into frames — not on the GPU doing math, so the library you pick to read frames off disk quietly sets your training speed. This project times four common readers — decord, torchvision.io, PyAV, and ffmpeg-python — on the same folder of .mp4s and reports decode time per clip. They differ by large factors because each makes different trade-offs around the underlying video codec, CPU threading, and how directly it hands you a tensor instead of a generic image. Finding the winner for your data turns a slow pipeline into a fast one without touching the model at all.