why not mean of class-wise accuracy ?

[liyunlongaaa]

# best model on the validation set
stats, _ = validate(audio_model, val_loader, args, 'valid_set')
# note it is NOT mean of class-wise accuracy
val_acc = stats[0]['acc']                                           
val_mAUC = np.mean([stat['auc'] for stat in stats])

for val_acc = stats[0]['acc'] ，I can't understand why only the accuracy of the first class is taken, can this also explain the problem, will it be more convincing to choose the mean of all classes?

[YuanGongND]

Hi there,

As commented above the line, accuracy is not a class-wise metric (i.e., it is not a one vs all accuracy). All classes have the same accuracy value, so taking the mean should lead to the same value.

https://github.com/YuanGongND/ast/blob/87a80043154eb4bb34ebceb4dc3e2d91a99235f4/src/utilities/stats.py#L26

-Yuan

[liyunlongaaa]

oh I see, thank you!