get_batch() returns mmco: unref short failure, even for a valid frame

[ffahmed]

get_batch returns " [h264 @ 0x561c38325ac0] mmco: unref short failure ", for some frames in the middle of the video. Video plays fine and there is no other issue.

To recreate the problem for one of the many videos do the following:

1. download this https://www.rocq.inria.fr/cluster-willow/amiech/Youcook2_val.zip
2. extract and download Youcook2_val/validation/201/mZwK0TBI1iY.mkv
3. Load the video using vr=VideoReader(..)
4. Do a vr.get_batch([257]).shape
5. Although it has 7939 frames, it throws error while getting frame 257.

Could you please help me debug this?

Thanks

[dxli94]

Any update on this, pleaes?

[frelwx]

I meet this, too. It seems to be neither a bug or warning.😭

[DrJimFan]

I have the same issue, but it didn't crash the program. Also I visualized the extracted tensor, it actually looks fine despite the very serious-looking warning message? Do you guys also see a normal execution and output? I can't figure out how to suppress that warning though. It's spamming my training loop constantly. @ffahmed @dxli94 @frelwx

[dxli94]

I have the same issue, but it didn't crash the program. Also I visualized the extracted tensor, it actually looks fine despite the very serious-looking warning message? Do you guys also see a normal execution and output? I can't figure out how to suppress that warning though. It's spamming my training loop constantly. @ffahmed @dxli94 @frelwx

Same observation. Training was not interrupted. It might be due to corrupt video files but I can't be sure.

[cecilia930426]

Any update?

[scottfleming]

Same issue here. Different video, different dataset, but basically the exact same problem as @ffahmed. Doesn't affect runtime - despite the ominous error the reading/loading appears to work just fine. But it does raise concerns about potential problems underneath the hood with unexpected behavior/results downstream. Happy to provide more details if it would be helpful, but if you can reproduce the OP's issue, I'm fairly certain it would be the same for me.

[SCZwangxiao]

Same dataset, same issue.

I've tried to re-encode the original *.mkv files into *.mp4 files, but the message still keeps going out.

[kahnchana]

I'm facing the same issue - any updates?

[zhanwenchen]

I get 4x [h264 @ 0x5651802bc400] mmco: unref short failure when I process v_UvPUywSVy1k.mp4 from ActivityNet using cpu, but it hangs with gpu. This is the code in question:

def load_video(vis_path, device, num_frm=100):
    with open(vis_path, 'rb') as file_in:
        # vr = VideoReader(file_in, ctx=gpu(0), num_threads=0)
        vr = VideoReader(file_in, ctx=cpu(0), num_threads=0)
    total_frame_num = len(vr)
    total_num_frm = min(total_frame_num, num_frm)
    frame_idx = get_seq_frames(total_frame_num, total_num_frm)
    # img_array = vr.get_batch(frame_idx).asnumpy()  # (n_clips*num_frm, H, W, 3)
    img_array = vr.get_batch(frame_idx) # (n_clips*num_frm, H, W, 3)
    del vr

[patrick-tssn]

same question

[zealota]

I met the same issue, during processing ActivityNet dataset.