Pytorch_Face_Recognition

Pytorch implementation of mainstream face recognition algorithms(ArcFace, CosFace).

1. Introduction

(1) Pytorch implementation of ArcFace and CosFace.
(2) Cleaned datasets are provided, including WebFace, MS-Celeb-1M, LFW, AgeDB-30, CFP-FP and MegaFace.
(3) Pretrained models are provided. See 3. Results and Pretrained Models for further details
(4) Automatic Mixed Precision(AMP) Training is supported to accelerate training process.
(5) Visdom is supported to visualize the changes of loss and accuracy during training process.

2. Usage

(1) Environment Preparation

Python 3.7
Pytorch 1.4
torchvision 0.5
cudatoolkit 10.0
apex 0.1 (optional)
visdom 0.1.8.9 (optional)

(2) Download this project to your machine.

(3) Datasets Preparation

Training Datasets :

CASIA-WebFace (453580/10575)
MS-Celeb-1M (3923399/86876)
Test Datasets :

LFW
CFP-FP
AgeDB-30
MegaFace

(4) For training datasets, use Utils/Datasets_Utils/generate_dataset_list.py to generate dataset files list.

(5) Set training hyperparameters like batch size, backbone, initial learning rate in Config/config.py

(6) Run Train.py to start training process and training information will be saved in the log file.

(7) Use LFW_Evaluation.py, AgeDB-30_Evaluation.py, CFP-FP_Evaluation.py and MegaFace_Evaluation to run individual evaluation.

Tips :
(1) In config.py, you can choose to open parameter adjustment mode. In this mode, training information and model will not be saved. It is convenient when you adjust the training hyperparameters.

(2) When you want to visualize the training process, you can turn the option 'use_visdom' in config.py to True. Before use it, make sure that you have installed the visdom and opened the server.

(3) You can use Automatic Mixed Precision(AMP) to accelerate training process. It allows you to use the bigger batch size and AMP can also avoid gradient explosion problem. To use it, just set option 'use_amp' in config.py as True. Remember to make sure that you have installed the apex and your GPU has TensorCore before use AMP.

(4) If you have multiple GPU devices and want to run parallel training, just set option 'use_multi_gpus' in config.py as True.

3. Results and Pretrained Models

(1) LFW, AgeDB-30 and CFP-FP Evaluation Results

Training Dataset	Backbone	Model Size	Loss	LFW	AgeDB-30	CFP-FP	Pretrained Models
CASIA-WebFace	MobileFaceNet	4MB	ArcFace	99.3333	92.5833	94.0143	BaiduNetDisk Extraction code:e3qm
CASIA-WebFace	ResNet50-IR	170MB	ArcFace	99.4667	93.9333	95.5571	BaiduNetDisk Extraction code:byqs
CASIA-WebFace	SEResNet50-IR	171MB	ArcFace	99.3833	93.9333	95.5857	BaiduNetDisk Extraction code:c355
CASIA-WebFace	ResNet100-IR	256MB	ArcFace	99.5833	94.3500	96.0429	BaiduNetDisk Extraction code:kqsi

(2) MegaFace Rank 1 Identifiaction Accuracy and Verfication TPR@FPR=1e-6 Results

Training Dataset	Backbone	Model Size	Loss	Identification Rank1 Acc	Verfication TPR@FPR=1e-6
CASIA-WebFace	MobileFaceNet	4MB	ArcFace	68.46	83.49
CASIA-WebFace	ResNet50-IR	170MB	ArcFace	74.50	89.89
CASIA-WebFace	SEResNet50-IR	171MB	ArcFace	74.72	89.41
CASIA-WebFace	ResNet100-IR	256MB	ArcFace	74.39	90.86

(3) The experimental condition :

CPU : Intel(R) Xeon(R) Silver 4114 CPU @ 2.20GHz 10 cores 20 threads × 2
Memory : 128GB
GPU : RTX2080ti × 2

(4) Training hyperparameters :

CASIA-WebFace: Batch size: 256
Initial learning rate: 0.05
Total epoch = 36
Learning rate scheduler = [22, 31]
S = 32

MS-Celeb-1M: Batch size: 256
Initial learning rate: 0.05
Total epoch = 25
Learning rate scheduler = [13, 21]
S = 32

4. References

wujiyang/Face_Pytorch
deepinsight/insightface
Xiaoccer/MobileFaceNet_Pytorch
TreB1eN/InsightFace_Pytorch

XinWangg/Pytorch_Face_Recognition-