Multi-label classification pipeline

Just to confirm the current best approach of multilabel classification with DALI.

I'm following the instruction from #874 to create a multi-label loader and using the file_list to ingest an ID that points to the multi-hot encoding vector.

It is necessary to implement an ExternalSource to accomplish this?
Because, when I try to use the dictionary of <indices, labels> inside define_graph directly, I'm getting the following error:

File "train.py", line 64, in define_graph
    print((images, self.combinations(labels)))
TypeError: 'numpy.ndarray' object is not callable

What am I doing wrong?

My define_graph looks like this:

def define_graph(self):
        inputs, labels = self.reader(name="Reader")
        images = self.decode(inputs)
        if self.device is 'gpu':
            labels = labels.gpu()
        images = self.cmn(images)

        return (
            images,
            self.combinations(labels)
        )

Start with train.py at line 64 and the guidance in issue #874; inspect how labels returned by the reader and self.combinations are represented inside define_graph. Reproduce the shown TypeError with the pipeline and determine whether the multi-label loader path requires ExternalSource; done when the current approach and corrected usage are verified.