[Reproduction] NaN AUC and Abnormal Metrics observed on PadChest Unseen/Rare Classes

[OPMZZZ]

Hello,

Thank you for your great work. I am currently trying to reproduce the zero-shot unseen classification results reported in your paper.

Description of the Issue: I have successfully reproduced the reported results on most of the disease test sets. However, when evaluating on the PadChest dataset (specifically the unseen/rare subclasses), I observed anomalous behavior in the metric report.

Specifically:

NaN AUC values: Several classes (e.g., round atelectasis, surgery humeral, empyema) show nan for AUC ROC.

Abnormal Recall/Precision: For these specific classes, the Precision is 0 while Recall is 1, which seems mathematically conflicting (unless handled by specific edge-case logic).

Average Metrics: Consequently, the averaged AUC is nan, and the average Accuracy seems lower than expected.

Logs / Results Table: Here is the output table I obtained:

Class Name	Accuracy	Max F1	AUC ROC	Precision	Recall
round atelectasis	0.000128	0	nan	0	1
surgery humeral	0.000128	0	nan	0	1
empyema	0.000128	0	nan	0	1
pulmonary artery hypertension	0.000128	0	nan	0	1
aortic aneurysm	0.000128	0	nan	0	1
... (omitted other classes)	...	...	...	...	...
Average	0.65647	0.0465	nan	0.0288	0.5982