hi ils
Try my class balancer it should help you stabilize your score… and could u push you up.
but ensure you ratio of classes is not such that less dominant class is many fold times its actual total occurance in whole data set in a epoch. Currently without Bal class sampler there are very occurrences of less dominant class in any batch so with this one you are just enhancing there count to become comparable to other classes in data set.
Does anyone of similar strategies to ensure stratified batches for imbalanced object detection tasks?