Upsampling ratio doesn't seem to behave as the documentation suggests it would for custom model training with multiple small classes

[wcornwell]

Describe the bug

Upsampling ratio doesn't seem to behave as the documentation suggests it would. For examples with large number of classes, the number of upsamples doesn't go up as fast as the clip distribution + upsampling ratio would suggest.

I think after upsampling the first small class, the counter (y_temp) inflates globally, causing subsequent classes to skip upsampling even if they're underrepresented.

I think the issue is here:

BirdNET-Analyzer/birdnet_analyzer/model.py

Line 351 in 59ddce6

while y[:, i].sum() + len(y_temp) < min_samples:

This doesn't really have any effect for binary custom classifiers, but for a multi-class, custom model with multiple small classes, then this can make it hard to upsample them properly.

[max-mauermann]

Hi,
thanks for reporting this, it is indeed an error and not working as intended.