Face Recognition in PyTorch

This project mainly contains various loss function and light neural network.

Light Neural Network

MobileNet-v3

EfficientNet

Loss function:

These functions addresses deep face recognition (FR) problem under open-set protocol, where ideal face features are expected to have smaller maximal intra-class distance than minimal inter-class distance under a suitably chosen met- ric space.

Softmax:

CenterLoss:

L-Softmax:

where

Calculate phi(theta) and k:

Calculate cos(m * theta):

Angular-Softmax:

Cosine-Softmax:

where,

ArcFace:

Reference

https://github.com/MuggleWang/CosFace_pytorch/blob/master/layer.py

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

SphereFace: Deep Hypersphere Embedding for Face Recognition

A Discriminative Feature Learning Approach for Deep Face Recognition

CosFace: Large Margin Cosine Loss for Deep Face Recognition

Large-Margin Softmax Loss for Convolutional Neural Networks

EfficientNet

MobileNet-v3