Compose sox_effects and torchaudio.transforms

[vincentqb]

I'd like to suggest a way of composing sox_effects with torchaudio.transforms. The latter is made of torch.nn.Module so they can be composed using torch.nn.Sequential. This would also follow the design pattern we established in torchaudio with transforms/functionals.

class SoxEffectsTransform(torch.nn.Module):

    def __init__(self, effects, sample_rate):
        self.effects = effects
        self.sample_rate = sample_rate

    def forward(self, tensor):
        data, _ = torchaudio.sox_effects.apply_effects_tensor(tensor, self.sample_rate, self.effects)
        return data


tensor, sample_rate = ...
effects = [
    ['gain', '-n', '-10'],
    ['speed', '0.5'],  # duration of data is 0.5 seconds
    ['rate', f'{sample_rate}'],
    ['pad', '0', '1.5'],  # add 1.5 seconds silence at the end
    ['trim', '0', '2'],  # get the first 2 seconds
]
sox_effects_transform = SoxEffectsTransform(effects, sample_rate)
transform = torchaudio.transforms.Spectrogram()

torch.nn.Sequential(sox_effects_transform, transform)

Would it be valuable to offer something like SoxEffectsTransform directly in torchaudio? It is currently in the docstring.

[lcances]

Hello,
That would be an elegant solution to mix both sox_effects and torchaudio.transforms. I am particularly interested since I deal with waveforms of different lengths and sampling rates. Until now, I was pre-processing my dataset by re-sampling all audio files to the desired waveform and padding / cropping them to the desired length.

To start, I just tried to make your SoxEffectsTransform works but the following error is raised.

RuntimeError: torchaudio.sox_effects.sox_effects.apply_effects_tensor requires module: torchaudio._torchaudio

I carefully read the source code and couldn't find the torchaudio._torchaudio module which is a requirement.
Deactivating this requirement obviously doesn't solve the problem but raise the following error (if it is of any interest)

RuntimeError: Tried to instantiate class 'torchaudio.TensorSignal', but it does not exist! Ensure that it is registered via torch::class_

[mthrok]

Hi @leocances

Which OS and version of torchaudio are you using? sox-related functionalities are only available in Linux/macOS distribution. Windows is not supported.

[mthrok]

I'd like to suggest a way of composing sox_effects with torchaudio.transforms. The latter is made of torch.nn.Module so they can be composed using torch.nn.Sequential.

I do not think adding it to transforms provide value for the maintenance cost. It is just a pass-through to the function form, which users can do as a part of their own transform easily.

[mthrok]

@leocances FYI: I am adding on-the-fly sample rate conversion here #816. Please give a +1 if you think this is useful.

[lcances]

Hi @mthrok
I'm using Ubuntu 20.04.1 LTS.
I have created a dedicated environment for testing the current version of torchaudio, cloned the repo, and build it locally. So I should have the latest version available of torchaudio.

I am new to the PyTorch source code in general, and I may have missed something.

[mthrok]

@leocances
Hmm, that should work.
Could you try BUILD_SOX=1 python setup.py clean install (or develop)
If that does not help can you create a separate issue and follow the instruction to report your environment, thanks?

[lcances]

BUILD_SOX=1 python setup.py clean install did not work at once.
I had to run the build command manually before

So, this did the trick.

python setup.py clean
BUILD_SOX=1 python setup.py build install

Thank a lot