feat: Using torchaudio for IO

[mogwai]

Closes #484

[mogwai]

Looks like the inference needs refactoring as well

[mogwai]

Running the tests I'm getting when the inference notebook is being run.

ValueError                                Traceback (most recent call last)

    ~/miniconda3/envs/v2/lib/python3.8/site-packages/IPython/core/formatters.py in __call__(self, obj)
    343             method = get_real_method(obj, self.print_method)
    344             if method is not None:
    --> 345                 return method()
    346             return None
    347         else:

    ~/miniconda3/envs/v2/lib/python3.8/site-packages/pyannote/core/feature.py in _repr_png_(self)
    238         from .notebook import repr_feature
    239 
    --> 240         return repr_feature(self)
    241 
    242     _HANDLED_TYPES = (np.ndarray, numbers.Number)

    ~/miniconda3/envs/v2/lib/python3.8/site-packages/pyannote/core/notebook.py in repr_feature(feature)
    321     plt.rcParams['figure.figsize'] = (notebook.width, 2)
    322     fig, ax = plt.subplots()
    --> 323     notebook.plot_feature(feature, ax=ax)
    324     data = print_figure(fig, 'png')
    325     plt.close(fig)

    ~/miniconda3/envs/v2/lib/python3.8/site-packages/pyannote/core/notebook.py in plot_feature(self, feature, ax, time, ylim)
    264 
    265         if ylim is None:
    --> 266             m = np.nanmin(data)
    267             M = np.nanmax(data)
    268             ylim = (m - 0.1 * (M - m), M + 0.1 * (M - m))

    <__array_function__ internals> in nanmin(*args, **kwargs)

    ~/miniconda3/envs/v2/lib/python3.8/site-packages/numpy/lib/nanfunctions.py in nanmin(a, axis, out, keepdims)
    317         # Fast, but not safe for subclasses of ndarray, or object arrays,
    318         # which do not implement isnan (gh-9009), or fmin correctly (gh-8975)
    --> 319         res = np.fmin.reduce(a, axis=axis, out=out, **kwargs)
    320         if np.isnan(res).any():
    321             warnings.warn("All-NaN slice encountered", RuntimeWarning,

    ValueError: zero-size array to reduction operation fmin which has no identity

[hbredin]

Looks like the data output by inference contains lots of NaNs so pyannote.core cannot display it.
inference should be fixed first :)

[hbredin]

I reviewed everything but inference.

Why? It is late and I need to sleep but also I would like to know your thoughts on updating the shape of model output from (num_frames, num_classes) to (num_classes, num_frames) to be coherent with the new (num_channels, num_samples) input shape.

[hbredin]

Actually, this is not the same. normalize is not meant as standardize (mean substraction and putting variance to one).

It is meant as normalizing the power of the signal so that we can later control the signal-to-noise ratio when summing two chunks.

[mogwai]

I see, I think torchaudio does that out of the box? https://github.com/pytorch/audio/blob/3b9e93372dd48649624ac2bbf660bb2e3384820e/torchaudio/backend/_soundfile_backend.py#L93

[hbredin]

This is not the same thing. My understanding is that torchaudio divides waveform by max(abs(waveform)) so that it lies in [-1., 1.] interval. My normalize computes a different thing.

[hbredin]

🍾 🥳 Thanks!

[hbredin]

FYI, there is a out-of-bound bug in Inference due to some kind of rounding errors.
Working on it: I am not sure whether this has been introduced by this PR.