Class-balanced-loss-pytorch

Why we included this project

Class imbalance is a common headache when training classifiers, and this library tackles it at the loss level. It implements the CVPR 2019 class-balanced loss, which reweights each class by its effective number of samples instead of the raw count, so rare classes get a fairer share of the training signal. The whole thing is one self-contained PyTorch module with a small API, so you can drop it into an existing training loop without restructuring anything. That makes it a practical starting point for long-tail recognition, medical imaging, or any dataset where some classes are scarce. The author's Medium article walks through the math behind the loss, which helps before you commit to it.