Fixing the number of per class ground truths calculation in compute_macro_f_score()

[astrocyted]

Hi,

The number of ground truths per each class seem to be done in a degenerate way in the following line in case of zero tp_ratios leading to nan values of num_gt

        num_gts = per_class_tp / tp_ratios

https://github.com/DCASE-REPO/psds_eval/blob/41cb9639633271fd676a6a7f26dbff36fbc0c9ba/src/psds_eval/psds.py#L755C41-L755C41

Since there is already a relevant method implemented to obtain gts in the class, the above line seems unnecessary and can be replaced by :

        num_gts = self._get_dataset_counts().values[:-1].astype(float)

the [:-1] slicing is to remove the dummy class injected at the end of the counts tensor, confusion matrix, etc. by the original developers referred to as "injected_psds_world_label". similar slicing for example is perfomed in the line above of the above mentioned line to obtain the tp counts:

per_class_tp = np.diag(counts)[:-1]

[aFewThings]

Event label "injected_psds_world_label" is not added always as a last element, and it seems that self._get_dataset_counts() returns ordered data by event label alphabetically.

This may address the issue temporarily.

num_gts = self._get_dataset_counts().drop(WORLD).values.astype(float)