video clipping subsampler about video2dataset HOT 4 CLOSED

TODO's:

• add clip_column parameter, check if it's None
• create clipping subsampler (test on singular videos)
• if clip_column is not None, run the clipping subsampler
• iterate over each and write each sample

clipping subsampler TODO's:

• input: video stream, metadata
• output: list[fragmented_video_stream, adjusted_metadata]
• use ffmpeg to produce fragmented video_streams

from video2dataset.

current state: everything works but it's super slow since it needs to read the video every time it makes a CLIP

we can try to solve this by using the --segment-time parameter to make equally timed adjacent segments but this doesn't make much sense in the context of a dataset where this is selected somehow and it might not be adjacent or equally timed

from video2dataset.

in the case of HD-VILA100M it actually happens to be the case that the clips a re equally timed and adjacent so maybe this will be sufficient for now

from video2dataset.

solved here #52

from video2dataset.