I am wondering how zindi is evaluating the model. Given the example of zindi :
fn Pump Spinach abalimi
audio_files/009WL0S.wav 0.73 0.19 0.01
audio_files/00AH117.wav 0.03 0.45 0.99
Let's say the true label is Pump and abalimi for the 1st and 2nd item.
Are we computing the logloss for each class and then average all of them ? So we end up with 193 logloss then average them ?
I could be wrong but I think that if the evaluation was class wise, it needs to be declared in the evaluation part which is not the case.