Arcface Pytorch (Distributed Version of ArcFace)

Set Up

torch >= 1.6.0

More details see eval.md in docs.

Training

1. Single node, 1 GPUs:

python -m torch.distributed.launch --nproc_per_node=1 --nnodes=1 --node_rank=0 --master_addr="127.0.0.1" --master_port=1234 train.py

2. Single node, 8 GPUs:

python -m torch.distributed.launch --nproc_per_node=8 --nnodes=1 --node_rank=0 --master_addr="127.0.0.1" --master_port=1234 train.py

3. Multiple nodes, each node 8 GPUs:

Node 0:

python -m torch.distributed.launch --nproc_per_node=8 --nnodes=2 --node_rank=0 --master_addr="ip1" --master_port=1234 train.py

Node 1:

python -m torch.distributed.launch --nproc_per_node=8 --nnodes=2 --node_rank=1 --master_addr="ip1" --master_port=1234 train.py

Evaluation IJBC

More details see eval.md in docs.

Speed Benchmark

ArcFace_torch can train large-scale face recognition training set efficiently and quickly.
When the number of classes in training sets is greater than 300K and the training is sufficient, partial fc sampling strategy will get same accuracy with several times faster training performance and smaller GPU memory.

Training speed of different parallel methods (samples/second), Tesla V100 32GB * 8. (Larger is better)

Method	Bs128-R100-2 Million Identities	Bs128-R50-4 Million Identities	Bs64-R50-8 Million Identities
Data Parallel	1	1	1
Model Parallel	1362	1600	482
Fp16 + Model Parallel	2006	2165	767
Fp16 + Partial Fc 0.1	3247	4385	3001

GPU memory cost of different parallel methods (GB per GPU), Tesla V100 32GB * 8. (Smaller is better)

Method	Bs128-R100-2 Million Identities	Bs128-R50-4 Million Identities	Bs64-R50-8 Million Identities
Data Parallel	OOM	OOM	OOM
Model Parallel	27.3	30.3	32.1
Fp16 + Model Parallel	20.3	26.6	32.1
Fp16 + Partial Fc 0.1	11.9	10.8	11.1

Model Zoo

The models are available for non-commercial research purposes only.

All Model Can be found in here.
Baidu Yun Pan: e8pw

MS1MV3

Datasets	log	backbone	IJBC(1e-05)	IJBC(1e-04)	agedb30	cfp_fp	lfw
MS1MV3-Arcface	log	r18-fp16	92.07	94.66	97.77	97.73	99.77
MS1MV3-Arcface	log	r34-fp16	94.10	95.90	98.10	98.67	99.80
MS1MV3-Arcface	log	r50-fp16	94.79	96.46	98.35	98.96	99.83
MS1MV3-Arcface	log	r100-fp16	95.31	96.81	98.48	99.06	99.85

Glint360k

Datasets	log	backbone	IJBC(1e-05)	IJBC(1e-04)	agedb30	cfp_fp	lfw
Glint360k-Cosface	log	r18-fp16-0.1	93.16	95.33	97.72	97.73	99.77
Glint360k-Cosface	log	r34-fp16-0.1	95.16	96.56	98.33	98.78	99.82
Glint360k-Cosface	log	r50-fp16-0.1	95.61	96.97	98.38	99.20	99.83
Glint360k-Cosface	log	r100-fp16-0.1	95.88	97.32	98.48	99.29	99.82

0.1 means sample rate is 0.1.

More details see eval.md in docs.

Test

We test on PyTorch versions 1.6.0, 1.7.1, and 1.8.0. Please create an issue if you are having trouble.

Citation

@inproceedings{deng2019arcface,
  title={Arcface: Additive angular margin loss for deep face recognition},
  author={Deng, Jiankang and Guo, Jia and Xue, Niannan and Zafeiriou, Stefanos},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  pages={4690--4699},
  year={2019}
}
@inproceedings{an2020partical_fc,
  title={Partial FC: Training 10 Million Identities on a Single Machine},
  author={An, Xiang and Zhu, Xuhan and Xiao, Yang and Wu, Lan and Zhang, Ming and Gao, Yuan and Qin, Bin and
  Zhang, Debing and Fu Ying},
  booktitle={Arxiv 2010.05222},
  year={2020}
}

ichenning / arcface-pytorch Goto Github PK

arcface-pytorch's Introduction

Arcface Pytorch (Distributed Version of ArcFace)

Contents

Set Up

Training

1. Single node, 1 GPUs:

2. Single node, 8 GPUs:

3. Multiple nodes, each node 8 GPUs:

Evaluation IJBC

Speed Benchmark

Model Zoo

MS1MV3

Glint360k

Test

Citation

arcface-pytorch's People

Contributors

Stargazers

Watchers

Recommend Projects

Recommend Topics

Recommend Org