av.read() exit gracefully on corrupt files?

[bjuncek]

I'm trying to load some videos using av.read and I am aware that dataset contains corrupt files. Most corrupt files can get filtered out by a simple trycatch as

# minimal reproducible example
for d in input_files[7161:7163]:
    try:
        vr = av.open(d)
    except:
        print("Skipping file {} - might be corrupted".format(d))

however, on a several, the file crashes with

Invalid NAL unit size (54396 > 13392).
Error splitting the input into NAL units.
Could not find codec parameters for stream 0 (Video: h264 (avc1 / 0x31637661), none, 640x360, 750 kb/s): unspecified pixel format

which is not caught in the exception, but rather completely breaks the code.
Is there a nicer way to gracefully exit or skip files upon this error?

I could do a binary search to find the videos in question but the dataset is quite large (2M examples) and it would take hours do do it by hand.

NOTE: trying to catch it with

except av.AVError:

fails as well.

[mikeboers]

Could you clarify what "crash" and "breaks the code" means? Does the Python interpreter exit? Or... does Python continue on as if there is no error?

[bjuncek]

There are two scenarios:

1. if I try to catch it in the error, without exception definition so just as
   except: some code, the code runs (although it prints the error above, the NAL file issue), but the output is nonsense, and if used within a larger framework such as pytorch the error is

 File "av/utils.pyx", line 27, in av.utils.AVError.__init__
TypeError: __init__() takes at least 3 positional arguments (2 given)

2. If I try and catch it with av.AVError, it fails immediately with
   TypeError: __init__() takes at least 3 positional arguments (2 given)

[mikeboers]

Interesting. It looks like the AVError construction is wrong. Can you post the full traceback?

[fmassa]

@mikeboers the constructor raises this error because av.AVError expects at least 2 arguments https://github.com/mikeboers/PyAV/blob/db30136b75e63d97cb8cb1a2e9ea4ef7ff777e2d/av/utils.pyx#L27
while in @bjuncek case, he was using PyTorch DataLoader, which wraps the exception and only passes a single argument to it, which is a string representing the full stack trace, see here and here for reference.

If one could make the constructor of AVError to accept a single argument (the message, as the first argument), then this would avoid @bjuncek errors that he mentioned. But this might be a BC-breaking change.

[jlaine]

I'm not sure I understand how PyTorch factors into this: is PyAV incorrectly constructing the AVError or is PyTorch?

Settings PyTorch aside, could we have a sample file which fails without properly raising an AVError?

[fmassa]

The problem is in PyTorch: it captures the exception message, and raises it afterwards, but only passing a single argument. I don't believe this is really a problem with PyAV, but more a limitation on how PyTorch assumes that exceptions should have their error message as the first argument

[fmassa]

Note that my last link above does not correspond to the right place anymore, here is the correct line now https://github.com/pytorch/pytorch/blob/863818e05a80b970e58a9dbce09c114ecbc6879e/torch/utils/data/dataloader.py#L608

[jlaine]

I'm sorry if I'm a bit slow here, but I don't understand the relation between PyTorch and PyAV : where is the code that uses DataLoader with PyAV?

[mikeboers]

I'm not following either, and it has been a long time. Re-open if you can show us what to change with our exceptions.