Need help with Face_Pytorch?
Click the “chat” button below for chat support from the developer who created it, or find similar developers for support.

About the developer

617 Stars 142 Forks Apache License 2.0 188 Commits 51 Opened issues


face recognition algorithms in pytorch framework, including arcface, cosface, sphereface and so on

Services available


Need anything else?

Contributors list

# 72,774
188 commits

Currently, I have graduated from campus and doing another kind of job. So this project may not be updated again.


The implementation of popular face recognition algorithms in pytorch framework, including arcface, cosface and sphereface and so on.

All codes are evaluated on Pytorch 0.4.0 with Python 3.6, Ubuntu 16.04.10, CUDA 9.1 and CUDNN 7.1. Partially evaluated on Pytorch 1.0.

Data Preparation

For CNN training, I use CASIA-WebFace and Cleaned MS-Celeb-1M, aligned by MTCNN with the size of 112x112. For performance testing, I report the results on LFW, AgeDB-30, CFP-FP, MegaFace rank1 identification and verification.

For AgeDB-30 and CFP-FP, the aligned images and evaluation images pairs are restored from the mxnet binary file provided by insightface, tools are available in this repository. You should install a mxnet-cpu first for the image parsing, just do ' pip install mxnet ' is ok.
LFW @ BaiduNetdisk, AgeDB-30 @ BaiduNetdisk, CFP_FP @ BaiduNetdisk


MobileFaceNet: Struture described in MobileFaceNet
ResNet50: Original resnet structure
ResNet50-IR: CNN described in ArcFace paper
SEResNet50-IR: CNN described in ArcFace paper

Verification results on LFW, AgeDB-30 and CFP_FP

Small Protocol: trained with CASIA-WebFace of data size: 453580/10575
Large Protocol: trained with DeepGlint MS-Celeb-1M of data size: 3923399/86876

Model Type

Loss LFW AgeDB-30 CFP-FP Model Size protocol
MobileFaceNet ArcFace 99.23 93.26 94.34 4MB small
ResNet50-IR ArcFace 99.42 94.45 95.34 170MB small
SEResNet50-IR ArcFace 99.43 94.50 95.43 171MB small
MobileFaceNet ArcFace 99.58 96.57 92.90 4MB large
ResNet50-IR ArcFace 99.82 98.07 95.34 170MB large
SEResNet50-IR ArcFace 99.80 98.13 95.60 171MB large
ResNet100-IR ArcFace 99.83 98.28 96.41 256MB large

There exists an odd result fact that when training under small protocol, CFP-FP performances better than AgeDB-30, while when training with large scale dataset, CFP-FP performances worse than AgeDB-30.

MegaFace rank 1 identifiaction accuracy and [email protected]=1e-6 results

Model Type

Loss MF Acc. MF Ver. MF [email protected] MF [email protected] SIZE protocol
MobileFaceNet ArcFace 69.10 84.23 81.15 85.86 4MB small
ResNet50-IR ArcFace 74.31 88.23 87.44 89.56 170MB small
SEResNet50-IR ArcFace 74.37 88.32 88.30 89.65 171MB small
MobileFaceNet ArcFace 74.95 88.77 89.47 91.03 4MB large
ResNet50-IR ArcFace 79.61 96.02 96.58 96.78 170MB large
SEResNet50-IR ArcFace 79.91 96.10 97.01 97.60 171MB large
ResNet100-IR ArcFace 80.40 96.94 97.60 98.05 256MB large


  1. Download the source code to your machine.
  2. Prepare the train dataset and train list, test dataset and test verification pairs.
  3. Set your own dataset path and any other parameters in

4. Run file, test accuracy will print into log file during training process.

  1. Every evaluate file can work independently for the model test process. just set your own args in the file.


Visdom support for loss and accuracy during training process.

Softmax Loss vs SoftmaxCenter Loss. Left: softmax training set. Right: softmax + center loss training set.

train <img src="result/softmaxcenter.gif" alt="train" width="49.5%">




  1. Report the test results on DeepGlint Trillion Pairs Challenge.
  2. Add C++ api for fast deployment with pytorch 1.0.
  3. Train the ResNet100-based model.

We use cookies. If you continue to browse the site, you agree to the use of cookies. For more information on our use of cookies please see our Privacy Policy.