WIP: Use istft from torch

[mthrok]

In combination with pytorch/pytorch#34827, I was wrapping up istft function from pytorch package.
I thought all the tests would pass, but after moving istft to pytorch, the function is no longer jit-able as it does not have corresponding C++ implementation in ATen.

@vincentqb Should we disable the test until the C++ implmentation of istft becomes available?

$ python test/test_functional.py
/scratch/moto/torchaudio/torchaudio/functional.py:124: UserWarning: istft has been moved to PyTorch and will be deprecated, please use torch.istft instead.
  warnings.warn(
....../home/moto/conda/envs/PY3.8-cuda101/lib/python3.8/site-packages/librosa/filters.py:235: UserWarning: Empty filters detected in mel frequency basis. Some channels will produce empty responses. Try increasing your sampling rate (and fmax) or reducing n_mels.
  warnings.warn('Empty filters detected in mel frequency basis. '
....................../home/moto/conda/envs/PY3.8-cuda101/lib/python3.8/tempfile.py:819: ResourceWarning: Implicitly cleaning up <TemporaryDirectory '/tmp/tmpmphuzfhl'>
  _warnings.warn(warn_message, ResourceWarning)
.......E...gain: can't reclaim headroom
.
======================================================================
ERROR: test_torchscript_griffinlim (__main__.TestFunctional)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "test/test_functional.py", line 74, in test_torchscript_griffinlim
    _test_torchscript_functional(
  File "test/test_functional.py", line 31, in _test_torchscript_functional
    jit_out, py_out = _test_torchscript_functional_shape(py_method, *args, **kwargs)
  File "test/test_functional.py", line 21, in _test_torchscript_functional_shape
    jit_method = torch.jit.script(py_method)
  File "/scratch/moto/pytorch/torch/jit/__init__.py", line 1296, in script
    fn = torch._C._jit_script_compile(qualified_name, ast, _rcb, get_default_args(obj))
  File "/scratch/moto/pytorch/torch/jit/_recursive.py", line 568, in try_compile_fn
    return torch.jit.script(fn, _rcb=rcb)
  File "/scratch/moto/pytorch/torch/jit/__init__.py", line 1296, in script
    fn = torch._C._jit_script_compile(qualified_name, ast, _rcb, get_default_args(obj))
RuntimeError:
Unknown builtin op: aten::istft.
Here are some suggestions:
	aten::stft
	aten::ifft
	aten::irfft

The original call is:
  File "/scratch/moto/torchaudio/torchaudio/functional.py", line 126
    warnings.warn(
        'istft has been moved to PyTorch and will be deprecated, please use torch.istft instead.')
    return torch.istft(
           ~~~~~~~~~~~ <--- HERE
        input=stft_matrix, n_fft=n_fft, hop_length=hop_length, win_length=win_length, window=window,
        center=center, pad_mode=pad_mode, normalized=normalized, onesided=onesided, length=length)
'istft' is being compiled since it was called from 'griffinlim'
  File "/scratch/moto/torchaudio/torchaudio/functional.py", line 248

        # Invert with our current estimate of the phases
        inverse = istft(specgram * angles,
        ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~...  <--- HERE
                        n_fft=n_fft,
                        hop_length=hop_length,


----------------------------------------------------------------------
Ran 40 tests in 12.376s

FAILED (errors=1)

[vincentqb]

The test you show uses istft through the pytorch interface. There should be a way of invoking istft in a way that maintains jitability of istft. Have you tried jitting the istft function directly?

Let's make sure all the tests continue passing during the transition.

[mthrok]

import torch

kwargs = {
    'n_fft': 15,
    'hop_length': 3,
    'win_length': 11,
    'window': torch.hamming_window(11),
    'center': True,
    'pad_mode': 'constant',
    'normalized': True,
    'onesided': False,
}
input = torch.stft(torch.randn(3, 15), **kwargs)
jit_istft = torch.jit.script(torch.istft)

py_result = torch.istft(input, **kwargs)
jit_result = jit_istft(input, **kwargs)

print(py_result)
print(jit_result)

assert torch.allclose(py_result, jit_result, atol=1e-6)

I extracted the essence of the test and ran it separately, it turned out istft is jit-able after moving to pytorch.
So I guess there is something wrong with the test. I will take a look into it.