Weighted binary cross entropy loss using pos_weight

1

Hi,
I have a unbalanced dataset, so i tried to use pos_weight in BCEwithlogit loss:

torch.nn.BCEWithLogitsLoss(pos_weight=weights)(outputs,targets)
But I observed the loss is fluctuating very badly and results are also bad. shouldn’t they atleast be on par with the results i got without using class weights. Can anyone tell why this is happening and suggest another way for using class weights.

2

What are the metrics that you’re monitoring?
If you set weights equal to \#mojority/\#minority, then you should be checking average recall.

3

I’m using F1 score (both micro and macro)

4

image

F1 score is calculated from the precision and recall .
by change p_c you’re changing these two metric in opposite way.
and in practice you’re gaining some recall percentage for way worse precision.