Problem with EON Tuner

Report bugs
#1

Question/Issue:
Trying to use EON Tuner for the last few days, and the confusion matrices are not appearing as expected - i.e. the categories all have dashes within instead of numbers. Worked really well previously with the same dataset.

Project ID:
144609

Context/Use case:
Training audio classification model

#2

Hello @LydiaK,

I just had a quick look at your project. I think the issue comes from your strongly imbalanced dataset and from the way we “round” the results.

I can see,

  • class 1 you have something like 4500 samples that last 4 seconds
  • class 2 you have 800 samples that last 1 second
  • class 3 you have 650 samples that last 1 second

So your class 1 is over represented so all your results will fall into the class 1.

I hope that helps,

Best,

Louis

#3

Hi Louis - OK thanks for having a look. Yes I know my classes are imbalanced - I am trying to train a model for a rare event. But I will balance the training data and see if this helps. I’m just surprised as it seems to have worked previously.

I’ll see if that fixes things.

Thanks,

Lydia

#4

Hi Louis - in deed that was the issue, thanks so much for your help!

Lydia

1 Like