[v0.4.2] Release Tracker

[fmassa]

We will be cutting a 0.4.2 release on top of the 0.4.1 branch.

It will contain the new C++ video reader backend, which improves the speed for decoding videos significantly. Due to packaging issues of ffmpeg, for users to be able to benefit from this faster version they will need to compile torchvision from source (and they need to have ffmpeg available in their system).

By comparing all the PRs that were merged since 0.4.1 (link available at v0.4.0...master), here is the proposed list of PRs that will be integrated in the 0.4.2 release.

• [video reader] inception commit ([video reader] inception commit #1303) 31fad34
• Expose frame-rate and cache to video datasets (Expose frame-rate and cache to video datasets #1356) 85ffd93
• Expose num_workers in VideoClips (Expose num_workers in VideoClips #1359) 02a8c0a
• Fix randomresized params flaky (Fix randomresized params flaky #1282) 7c9bbf5
• Video transforms (Video transforms #1353) 64917bc
• add _backend argument to init() of class VideoClips (add _backend argument to __init__() of class VideoClips #1363) 7874374
• Video clips workers (Video clips workers #1369) 0982395
• modified code of io.read_video and io.read_video_timestamps to intepret pts values in seconds (modified code of io.read_video and io.read_video_timestamps to intepret pts values in seconds #1331) 17e355f
• add metadata to video dataset classes. bug fix. more robustness (add metadata to video dataset classes. bug fix. more robustness #1376) 49b01e3
• move sampler into TV core. Update UniformClipSampler (move sampler into TV core. Update UniformClipSampler #1408) f0d3daa
• remove hardcoded video extension in kinetics400 dataset (remove hardcoded video extension in kinetics400 dataset #1418) 929c81d
• Fix hmdb51 and ucf101 typo (Fix hmdb51 and ucf101 typo #1420) b13931a
• fix a bug related to audio_end_pts (fix a bug related to audio_end_pts #1431) 1258bb7
• expose more io api (expose more io api #1423) e48b958
• Make video transforms private (Make video transforms private #1429) 79daca1
• extend video reader to support fast video probing (extend video reader to support fast video probing #1437) ed5b2dc
• Better handle corrupted videos (Better handle corrupted videos #1463) da89dad
• Temporary fix to remove ffmpeg from build time (Temporary fix to remove ffmpeg from build time #1475) ed04dee
• fix a bug when video decoding fails and empty frames are returned (fix a bug when video decoding fails and empty frames are returned #1506) 2804c12
• extend DistributedSampler to support group_size (extend DistributedSampler to support group_size #1512) 355e9d2
• Unify video backend (Unify video backend #1514) 97b53f9
• Unify video metadata in VideoClips (Unify video metadata in VideoClips #1527) 7d509c5
• Fixed compute_clips docstring (Fixed compute_clips docstring #1543) b438d32

cc @stephenyan1231 for review

[fmassa]

Here is the cherry-picked branch for the aforementioned commits: https://github.com/pytorch/vision/tree/v0.4.2

I had to fix a few minor conflicts on the way. Here is the list of commits that had conflicts:

• [video reader] inception commit ([video reader] inception commit #1303) 31fad34
• add _backend argument to init() of class VideoClips (add _backend argument to __init__() of class VideoClips #1363) 7874374
• Temporary fix to remove ffmpeg from build time (Temporary fix to remove ffmpeg from build time #1475) ed04dee
• Unify video backend (Unify video backend #1514) 97b53f9
• Unify video metadata in VideoClips (Unify video metadata in VideoClips #1527) 7d509c5