pytorch-center-loss

Why we included this project

Center loss is a simple but effective idea for recognition problems: you add a term that pulls every example toward its class center, so features from the same identity cluster together while staying separable from other classes. That is the objective behind this repo, a compact PyTorch implementation of Wen et al.'s 2016 paper, and the same loss powers the deep-person-reid project. The useful code sits in a single small file you can copy straight into an existing pipeline without adopting a bigger library. The repo also ships a runnable MNIST demo that logs softmax and center loss separately, so you can watch the two terms trade off as training progresses. It reads more like a clean template than standalone software, which is exactly what you want when implementing discriminative feature learning in your own model.