KaiyangZhou/pytorch-center-loss

Combines cross-entropy loss with a learnable center loss component that minimizes intra-class feature variance while maximizing inter-class separation, enabling tighter feature clustering for face recognition and person re-identification tasks. The implementation maintains per-class feature centers updated via separate SGD optimization, with configurable weight balancing between classification and center loss terms. Includes a complete MNIST demonstration pipeline with visualization of feature space evolution, showing significant accuracy improvements (10% → 98%+) when center loss supplements softmax training.

995 stars. No commits in the last 6 months.