Fine-Tuning Lightweight Embedding Models for Custom Facial Recognition

Standard pre-trained models like ArcFace or MobileFaceNet work remarkably well out of the box, but when dealing with restricted lighting conditions or specific camera angles in controlled environments, fine-tuning the embedding space with triplet loss yields much cleaner cluster separation.

Curious if anyone has experimented with contrastive learning frameworks in PyTorch for biometric identification datasets. What margin values are you finding optimal for cosine distance separation?