BCELoss and BCEWithLogits stop decreasing

I have multi label classification task, i have 90 labels, and every data have 1.2 label on average. I try to use BCELoss and BCEWithLogits and it’s stuck on early epoch and the model will give score 0.5 on every single label prediction.

So if i round the score, it means every single data doesnt have label cause the prediction is 0 (zero). It means my micro-f1 score will be 0 too.

Do tou have any suggestion to this problem?

But on several data point, the data has multiple label. I want my model can predict multiple label

I’m sorry, I was wrong about BCELoss, I’ll withdraw my post.

Would you be able to show us your code, so we might be able to spot any issues ?