Closed ahmadsharif1 closed 1 month ago
This is great! Are we also okay with comparing to Decord?
Should we also compare a call to .get_frames_at_indices
(or pts)? I suspect torchcodec will shine compared to (at least) torchvision?
I'm happy to commit this version and work on follow-ups later. I think we'll probably spend a bunch of time tweaking the presentation of the graph, and iterating on the exact experiments we run.
I merged this as is.
I'll create an issue for follow-up work on benchmarking -- throughput benchmarks, sampler benchmarks, better video selection, etc.
This PR: