Data set discussion

Recently, we released audtorch, an audio for PyTorch package that we started some time ago. It contains a few audio data sets that might be worth integrating here: Mozilla Common Voice, AudioSet, VoxCeleb1, LibriSpeech.

But before doing a pull request, there are a few things that I would like to discuss as I'm not completely happy with our current implementation:

Inherit from an Audio Base class or not?

• To not repeat lots of code we use an audio data set base class, from which others can inherit. It handles data loading, sampling rate handling, check for existence, and applies transforms.
• This approach is different from the two data sets VCTK and YesNo that are currently part of torchaudio
• torchvision has a vision base class, but this is handling more or less only transforms.

What is the best way to handle the sampling rate?

• Functions/classes using the data set have to be aware of the underlying sampling rate. At the moment we solved that by using a sampling_rate() property
• Transforms can change the sampling rate of a data set, we added a semi-optimal check for that
• Data sets can contain different sampling rates. At the moment we force the sampling rate to be the same
• The two data sets currently part of torchaudio just return data and the user has to know the sampling rate.

Should we handle failures during data loading?

• In order to not crash the training process we directly implemented error handling in our load function.

Note, our data sets currently return the data as numpy arrays as we use a lot of numpy transforms. But this can easily be changed.

Read the linked audtorch datasets/base.py and utils.py alongside torchaudio's vctk.py and yesno.py. Resolve the base-class, sampling-rate, failure-handling, and NumPy-versus-Tensor questions before proposing dataset integrations; done is an agreed design or a follow-up implementation issue.