Labeled data ingestion for train/eval

[sammlapp]

Simplify user experience by allowing user to pass a table specifying various annotation files, corresponding audio files/root paths, the annotation column name, the data split (eg train/val/test), annotated class, and the annotation type (eg multi hot csv, categorical csv, raven, audacity, whombat, dipper single target, dipper multi target)

If annotated class is empty assumes all classes annotated

Returns a dictionary with labels for each split

CatLabels_dict= ingest_labels(labels_summary.csv or df, class_list,upsample=500, downsample=None)
catlabels["train"].label_counts()
CatLabels.convert()
catlabels.trim_whitespace()
CatLabels.dropclips(labels="uncertain")

Optionally Upsamples and or downsamples training split to N samples per class

Consider also supporting xeno canto ingestion via britekit