ResizeSignal Always takes Random Chunk

[kevinbird15]

I was looking at using ResizeSignal to cut an input and output down to the same chunk of audio and realized that's not currently an easy thing to do. I was thinking that adding an option that would allow for crop_start to be set during initialization would be helpful. I haven't written anything yet, but wanted to see if that's something that would be helpful for anybody else.

Code in question:

class ResizeSignal(Transform):
    """Crops signal to be length specified in ms by duration, padding if needed"""

    def __init__(self, duration, pad_mode=AudioPadType.Zeros):
        self.duration = duration
        self.pad_mode = pad_mode
        if pad_mode not in [
            AudioPadType.Zeros,
            AudioPadType.Zeros_After,
            AudioPadType.Repeat,
        ]:
            raise ValueError(
                f"""pad_mode {pad_mode} not currently supported,
                only AudioPadType.Zeros, AudioPadType.Zeros_After,
                or AudioPadType.Repeat"""
            )

    def encodes(self, ai: AudioTensor) -> AudioTensor:
        sig = ai.data
        orig_samples = ai.nsamples
        crop_samples = int((self.duration / 1000) * ai.sr)
        if orig_samples == crop_samples:
            return ai
        elif orig_samples < crop_samples:
            ai.data = _tfm_pad_signal(sig, crop_samples, pad_mode=self.pad_mode)
        else:
            crop_start = random.randint(0, int(orig_samples - crop_samples)) ########################Always Random at the moment
            ai.data = sig[:, crop_start : crop_start + crop_samples]
        return ai

The other thing I though is that you could potentially set a random.seed but I wasn't able to make that solution to work in my code.

[scart97]

It makes sense, and would be a good addition to the library. Also I just realized that this code is not using pytorch to create random numbers, that is a major silent bug:
https://tanelp.github.io/posts/a-bug-that-plagues-thousands-of-open-source-ml-projects/

[kevinbird15]

That bug is super interesting, thanks for sharing!