Confusing IID behavior in masking transforms.

T.TimeMasking and T. FrequencyMasking have a parameter for changing the way mask is applied IID or not.

https://github.com/pytorch/audio/blob/1f136671b84071a2fe1d5b762df64f3a76310c31/torchaudio/transforms.py#L890-L899

https://github.com/pytorch/audio/blob/1f136671b84071a2fe1d5b762df64f3a76310c31/torchaudio/transforms.py#L905-L914

However this behavior is not automatically enforced, users have to make sure that the input Tensor is 4D by themselves manually, because of if ~ else clause here.

https://github.com/pytorch/audio/blob/1f136671b84071a2fe1d5b762df64f3a76310c31/torchaudio/transforms.py#L884-L887

This logic is error prune and can introduce subtle error, which is hard to detect. Instead it should be throwing an error if the input Tensor is not 4D.

        if self.iid_masks:
            if specgram.dim() != 4:
                throw ValueError(...)
            return F.mask_along_axis_iid(specgram, self.mask_param, mask_value, self.axis + 1)
        return F.mask_along_axis(specgram, self.mask_param, mask_value, self.axis)

Start in torchaudio/transforms.py at the TimeMasking and FrequencyMasking implementations and the linked conditional around lines 884-887. Check how iid_masks selects the masking function and how tensor dimensionality is handled. Done means invalid input dimensions for IID masking are rejected explicitly rather than silently taking the other path.