UNet Model not learning for multiclass image segmentation task

[ptrblck]

@ankita.buntolia

I tried with 30 epochs just to cross-check, but in the second epoch, loss dropped from 36% to 5%, which does not seem right.

The loss is not given in a percentage and an initial drop is common, so could you explain your concern a bit?

@ankita.buntolia

Is this the correct way to calculate class weights?

No, since you are now weighting the majority class higher and are thus adding more penalty to a wrong classification of class0. Divide by the class frequencies as explained in this post.