TypeError When Using TensorAudio Batch Transforms

[vishalbakshi]

First off, thanks for such an awesome library, I'm using it for the first time and really enjoying it.

I am running into a TypeError when using one of the TensorAudio batch transforms (VolumeBatch, PitchShift, TimeStretch, PitchShiftOrTimeStretch) when calling auds.summary where auds is the DataBlock.

I am using fastxtend version 0.1.7 and fastai version 2.7.14 on Kaggle.

I have created a reproducible example in this Kaggle Notebook. I chose to use Kaggle so that I can easily incorporate an audio classification dataset in the example.

In short, when I run the following code in that notebook:

auds = DataBlock(blocks = (AudioBlock, CategoryBlock),  
                 get_x = ColReader("filename", pref=path/"audio"/"audio"), 
                 splitter = ColSplitter(),
                 get_y = ColReader("category"),
                 batch_tfms = PitchShift())

auds.summary(df)

I get the error:

TypeError: BatchRandTransform.__call__() missing 1 required positional argument: 'split_idx'

I don't get this error when using the different item transforms (Flip, Roll, Volume, etc.) I also tried using fastai's built-in aug_transforms and RandomResizedCropGPU as batch_tfms and while it obviously doesn't do anything to the TensorAudio, auds.summary runs without error.

I looked at the source code for PitchShift and split_idx is defined at the top.

I also saw that there is an audio_fixes branch open so perhaps this is already a known issue. I'm not proficient at the internals of fastai, but am happy to help troubleshoot this issue if you can point me in a general direction of where to start looking. Or, perhaps I'm implementing it wrong and that's why I'm getting this error.

[warner-benjamin]

I think the issue is probably due to BatchRandTransform.__call__ not defaulting split_idx=None

Can you add this code right after the imports and see if it resolves the issue?

from fastxtend.transform import BatchRandTransform

def call_fix(self,
    b:Tensor|tuple[Tensor,...], # Batch item(s)
    split_idx:int|None = None, # Train (0) or valid (1) index
    **kwargs
) -> Tensor|tuple[Tensor,...]:
    "Call `super().__call__` if `self.do`"
    self.before_call(b, split_idx=split_idx)
    return super().__call__(b, split_idx=split_idx, **kwargs) if self.do else b
    
BatchRandTransform.__call__ = call_fix

The audio_fixes branch was merged in #18. I forgot to delete it.

[vishalbakshi]

Thanks that works!

This code:

auds = DataBlock(blocks = (AudioBlock, CategoryBlock),  
                 get_x = ColReader("filename", pref=path/"audio"/"audio"), 
                 splitter = ColSplitter(),
                 get_y = ColReader("category"),
                 batch_tfms = PitchShift())

auds.summary(df)

Now outputs the expected result:

Note that I was running into the following error when using your fix:

when running auds.summary(df):

So I modified the return statement as follows (ChatGPT assist) and it worked:

def call_fix(self,
    b:Tensor|tuple[Tensor,...], # Batch item(s)
    split_idx:int|None = None, # Train (0) or valid (1) index
    **kwargs
) -> Tensor|tuple[Tensor,...]:
    "Call `super().__call__` if `self.do`"
    self.before_call(b, split_idx=split_idx)
    return super(BatchRandTransform, self).__call__(b, split_idx=split_idx, **kwargs) if self.do else b
    
BatchRandTransform.__call__ = call_fix

[warner-benjamin]

@vishalbakshi can you check that #24 resolves this issue? If so, I will merge it and cut a release.