Speed up random access time of PyAVReaderIndexed.

[keunhong]

Previously PyAVReaderIndexed decoded every packet to seek to a specific packet. This PR tries to speed this up by instead indexing the timestamps of each packet and seeking to it using container.seek().

This speeds up random access times by an order of magnitude. Sequential access time remains unchanged.

Test Code

test_reader = PyAVReaderIndexed(...)
inds = np.random.randint(low=0, high=len(test_reader), size=3)
print(test_reader)

def test():
    for i in inds:
        print(i)
        test_reader.get_frame(i)
    
%time test()

Before

<Frames>
Source: /local1/kpar/spatialaudiogen/data/preproc-hr/cA88rNC2uuM-video.mp4
Length: 9061 frames
Frame Shape: (1080, 1920, 3)

673
7158
5975
CPU times: user 41.1 s, sys: 7.8 ms, total: 41.1 s
Wall time: 41.1 s

After

<Frames>
Source: /local1/kpar/spatialaudiogen/data/preproc-hr/cA88rNC2uuM-video.mp4
Length: 9061 frames
Frame Shape: (1080, 1920, 3)

673
7158
5975
CPU times: user 1.78 s, sys: 8.01 ms, total: 1.79 s
Wall time: 1.78 s

[keunhong]

Hi, can anyone take a look at this? I've been using this for the past 2 weeks in my machine learning workflow and I haven't run into any issues yet.

You can even get more gains by re-encoding videos with more frequent keyframes since less frames have to be decoded after seeking.

[sofroniewn]

@keunhong i'm not a PIMS maintainer, but am very excited about this speed-up. Not sure if you've seen my post here dask/dask-image#134 asking about such things for usage with dask, but I'm curious if you'd expect to see speed-ups using dask-image imread which uses PIMs under the hood. I can give it a try and report back too if others are interested.

[caspervdw]

Thanks a lot for this improvement. I looked through the code and it looks good. Tests also pass, so this is good to merge. I'll leave it up to @rbnvrw or @nkeim to decide whether it should go in v0.5 or not.

[nkeim]

Thanks for submitting this great PR, @keunhong ! And thanks for the review, @caspervdw . I think this should go into v0.5 since it is such a major improvement. The logical alternative would be to wait until v0.6, and I'm scared of that 😉.

@sofroniewn Since this is for random access, I imagine this would help tremendously with any kind of distributed loading of images from PyAV. I hope I am right!