/FR-Loss-on-Mnist

Face Recognition Loss on Mnist, Pytorch implementation

Primary LanguagePythonMIT LicenseMIT

FR Loss on Mnist

Implement diffrent Face Recognition Loss with Pytorch. Testing and visualizing on Mnist. It's just a toy example but very useful to understand and compare those loss function.

Recent Updata

2020.1.14:

  • Fix some bugs in ArcFace
  • Visualize test data rather than training data
  • Add support for our QAMFace Loss

Example

Raw

epoch=79 2D Embedded Feature cos_epoch=79 Normalized Feature

CenterLoss

epoch=79 cos_epoch=79

SphereFace

epoch=79 cos_epoch=79

Quick Start

Dependencies

  • Pytorch >=1.0 (0.4 maybe work either)
  • tensorboardX >=1.4

I highly recommend you to use Anaconda.

How to run

  • Net.py include the implementation of network and loss functions
  • train.py contains the training and test process. Editing name to choose a proper loss function.

About the project

  • I try to use the same structure to implement different loss. If you have any questions or you find any mistakes, please submmit an issue. Thanks a lot!