Non-integer categorical labels produce error.

[jcausey-astate]

When using the image_dataset_from_dataframe() function, which calls the image_dataset_from_paths_and_labels() function, given a dataframe whose label column contains string categorical labels (e.g. ['apple', 'orange', 'kiwi']), using the "categorical" label_mode, the following error results:

UFuncTypeError: ufunc 'maximum' did not contain a loop with signature matching types (dtype('<U25'), dtype('<U25')) -> None

This traces back to:

File working_dir.venv/lib/python3.10/site-packages/imflow/imflow.py:549, in image_dataset_from_paths_and_labels(image_paths, labels, label_mode, color_mode, batch_size, image_size, shuffle, seed, validation_split, subset, interpolation, resize_with_pad)
    547    num_classes = 2
    548  if label_mode in ('int', 'categorical'):
--> 549    num_classes = np.max(labels) + 1
    550  if label_mode == 'multi_class':
    551    num_classes = labels.shape[1]

where np.max(labels) is being called, but labels contains strings.

Maybe doing a separate case for string 'categorical' labels, with e.g. num_classes = np.unique(labels).shape[0] would solve this?

P.S. It also looks like you have an assertion that would not allow multi_class as a label type (line 491):

if label_mode not in {'int', 'categorical', 'multi_label', 'binary', None}:

but parts of the code (line 542 if label_mode in ('multi_class', 'multi_label'):) seem to allow for it.
(If you like, I could make this a separate issue for tracking.)