Too slow to process the EPIC-KITCHENS-100 dataset

Hi, thanks for your work! When I use this code to process the process videos in the EPIC-KITCHENS-100 dataset to tfrecord file, I found that the processing speed is extremely slow (about 40 samples an hour, while the EPIC-KITCHENS-100 dataset contains more than 90,000 samples). The following below code is the main reason.
` cmd = (
ffmpeg
.input(video_path)
.trim(start=start, end=end)
.filter("fps", fps=fps)
.filter("scale", new_width, -1)
.output("pipe:", format="image2pipe")
)
jpeg_bytes, _ = cmd.run(capture_stdout=True, quiet=True)`

I guess when extract frames from the video, it will process all frames before the end time, instead of localize the start time at first, so how to solve the problem?