How to label for this regression problem?

I am having trouble understanding how to label my data.

My inputs are spectrogram images of length 240, my desired output is a length 240 array of float values. Let’s assume I am trying to predict the amplitude of the spectrogram at each frame.

I have created a model with a custom head:

      (0): AdaptiveConcatPool2d(
    (ap): AdaptiveAvgPool2d(output_size=1)
    (mp): AdaptiveMaxPool2d(output_size=1)
      (1): Flatten()
      (2): BatchNorm1d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (3): Dropout(p=0.25, inplace=False)
      (4): Linear(in_features=1024, out_features=256, bias=True)
      (5): ReLU(inplace=True)
      (6): BatchNorm1d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (7): Dropout(p=0.5, inplace=False)
      (8): Linear(in_features=256, out_features=240, bias=True)

I have a df containing columns [Audio File, Amplitude]. Each cell in Amplitude contains a list of float values.

I have an AudioList:

    class AudioList(ItemList):
    _bunch = AudioDataBunch

    # TODO: __REPR__
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)

    def get(self, i):
        item = self.items[i]
        if isinstance(item, (Path, str)):
            return AudioItem(AudioData.load(str(item)))
        if isinstance(item, (tuple, np.ndarray)): #data,sr
            return AudioItem(AudioData(item[0],item[1]))
        print('Format not supported!', file=sys.stderr)

    def reconstruct(self, t:Tensor): return Image(t.transpose(1,2)) #FIXME!! No Image here

    def hear_xys(self, xs, ys, **kwargs):
        for x, y in zip(xs, ys): x.hear(title=y, **kwargs)

    # TODO: example with from_folder
    def from_folder(cls, path:PathOrStr='.', extensions:Collection[str]=None, **kwargs)->ItemList:
        extensions = ifnone(extensions, AUDIO_EXTENSIONS)
        return super().from_folder(path=path, extensions=extensions, **kwargs)

I am trying to create a Databunch with the Datablock API:

    def label_func(audio_file):
    filename = str(audio_file.split('/')[-1])
    y_vals = df[df['Audio File'].str.contains(filename)]['jawTrans_ty'].values[0]```

    data = (AudioList.from_df(df, path=audio_data, cols='Audio File')
        .label_from_func(label_func, label_cls=FloatList)

This does not work because FloatList expects a single Float value for each image to regress to.
How can I label my data to train against the length of the spectrogram?

Sorry, I am of no help but have you checked if anything is here Deep Learning with Audio Thread :slight_smile:

Thank you for the link. I’ve been leaning on it heavily.

My issue is more generic… I simply want to create labels with multiple float values for one item.