fix resample type hint

[faroit]

As indicated in the docstrings, the sample rates would need to be float. I also discovered that using int can result in an error

  File "/Users/faro/repositories/open-unmix-pytorch-dev/env/lib/python3.7/site-packages/torch/nn/modules/module.py", line 722, in _call_impl
    result = self.forward(*input, **kwargs)
  File "/Users/faro/repositories/open-unmix-pytorch-dev/env/lib/python3.7/site-packages/torchaudio/transforms.py", line 596, in forward
    waveform = kaldi.resample_waveform(waveform, self.orig_freq, self.new_freq)
  File "/Users/faro/repositories/open-unmix-pytorch-dev/env/lib/python3.7/site-packages/torchaudio/compliance/kaldi.py", line 930, in resample_waveform
    window_width, lowpass_cutoff, lowpass_filter_width, device, dtype)
  File "/Users/faro/repositories/open-unmix-pytorch-dev/env/lib/python3.7/site-packages/torchaudio/compliance/kaldi.py", line 807, in _get_LR_indices_and_weights
    assert lowpass_cutoff < min(orig_freq, new_freq) / 2
RuntimeError: Integer division of tensors using div or / is no longer supported, and in a future release div will perform true division as in Python 3. Use true_divide or floor_divide (// in Python) instead.

[vincentqb]

The convention we've been using so far as been int as you can see for instance in MFCC transform. To maintain the same behavior, we should replace the division by floor_divide. Do you want to open a pull request for this? Do you have a code to reproduce the error you mention?

#890

[faroit]

The convention we've been using so far as been int as you can see for instance in MFCC transform

I actually used int for decades but recently learned that float is more appropriate and mathematically correct. It is also often used in resamplers e.g, in sox.

To maintain the same behavior, we should replace the division by floor_divide. Do you want to open a pull request for this?

yes, but in any case, either the docstring or the type hint of thins function would need to changed as well.

[vincentqb]

Thanks for the feedback. I'd be happy to update the code to floats, #891. However, this may interfere with C++ extensions when saving.

The current behavior of this particular assertion (in torchaudio) is to do integer division, so I'd expect to use floor_division. That being said, this code is ported from Kaldi: do you know if the original intent was floor or true division?

[mthrok]

This has been resolved in #1499 (part of #1487)