Distributed Arcface Training in Pytorch
This is a deep learning library that makes face recognition efficient, and effective, which can train tens of millions identity on a single server.
Requirements
- Install pytorch (torch>=1.6.0), our doc for install.md.
pip install -r requirements.txt
.- Download the dataset from https://github.com/deepinsight/insightface/tree/master/recognition/datasets .
How to Training
To train a model, run train.py
with the path to the configs:
1. 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 configs/ms1mv3_r50
2. 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 train.py configs/ms1mv3_r50
Node 1:
python -m torch.distributed.launch --nproc_per_node=8 --nnodes=2 --node_rank=1 --master_addr="ip1" --master_port=1234 train.py train.py configs/ms1mv3_r50
3.Training resnet2060 with 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 configs/ms1mv3_r2060.py
Model Zoo
- The models are available for non-commercial research purposes only.
- All models can be found in here.
- Baidu Yun Pan: e8pw
- onedrive
ICCV2021-MFR
Performance onICCV2021-MFR testset consists of non-celebrities so we can ensure that it has very few overlap with public available face recognition training set, such as MS1M and CASIA as they mostly collected from online celebrities. As the result, we can evaluate the FAIR performance for different algorithms.
For ICCV2021-MFR-ALL set, TAR is measured on all-to-all 1:1 protocal, with FAR less than 0.000001(e-6). The globalised multi-racial testset contains 242,143 identities and 1,624,305 images.
For ICCV2021-MFR-MASK set, TAR is measured on mask-to-nonmask 1:1 protocal, with FAR less than 0.0001(e-4). Mask testset contains 6,964 identities, 6,964 masked images and 13,928 non-masked images. There are totally 13,928 positive pairs and 96,983,824 negative pairs.
Datasets | backbone | Training throughout | Size / MB | ICCV2021-MFR-MASK | ICCV2021-MFR-ALL |
---|---|---|---|---|---|
MS1MV3 | r18 | - | 91 | 47.85 | 68.33 |
Glint360k | r18 | 8536 | 91 | 53.32 | 72.07 |
MS1MV3 | r34 | - | 130 | 58.72 | 77.36 |
Glint360k | r34 | 6344 | 130 | 65.10 | 83.02 |
MS1MV3 | r50 | 5500 | 166 | 63.85 | 80.53 |
Glint360k | r50 | 5136 | 166 | 70.23 | 87.08 |
MS1MV3 | r100 | - | 248 | 69.09 | 84.31 |
Glint360k | r100 | 3332 | 248 | 75.57 | 90.66 |
MS1MV3 | mobilefacenet | 12185 | 7.8 | 41.52 | 65.26 |
Glint360k | mobilefacenet | 11197 | 7.8 | 44.52 | 66.48 |
Performance on IJB-C and Verification Datasets
Datasets | backbone | IJBC(1e-05) | IJBC(1e-04) | agedb30 | cfp_fp | lfw | log |
---|---|---|---|---|---|---|---|
MS1MV3 | r18 | 92.07 | 94.66 | 97.77 | 97.73 | 99.77 | log |
MS1MV3 | r34 | 94.10 | 95.90 | 98.10 | 98.67 | 99.80 | log |
MS1MV3 | r50 | 94.79 | 96.46 | 98.35 | 98.96 | 99.83 | log |
MS1MV3 | r100 | 95.31 | 96.81 | 98.48 | 99.06 | 99.85 | log |
MS1MV3 | r2060 | 95.34 | 97.11 | 98.67 | 99.24 | 99.87 | log |
Glint360k | r18-0.1 | 93.16 | 95.33 | 97.72 | 97.73 | 99.77 | log |
Glint360k | r34-0.1 | 95.16 | 96.56 | 98.33 | 98.78 | 99.82 | log |
Glint360k | r50-0.1 | 95.61 | 96.97 | 98.38 | 99.20 | 99.83 | log |
Glint360k | r100-0.1 | 95.88 | 97.32 | 98.48 | 99.29 | 99.82 | log |
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. Partial FC is a sparse variant of the model parallel architecture for large sacle face recognition. Partial FC use a sparse softmax, where each batch dynamicly sample a subset of class centers for training. In each iteration, only a sparse part of the parameters will be updated, which can reduce a lot of GPU memory and calculations. With Partial FC, we can scale trainset of 29 millions identities, the largest to date. Partial FC also supports multi-machine distributed training and mixed precision training.
More details see speed_benchmark.md in docs.
1. Training speed of different parallel methods (samples / second), Tesla V100 32GB * 8. (Larger is better)
-
means training failed because of gpu memory limitations.
Number of Identities in Dataset | Data Parallel | Model Parallel | Partial FC 0.1 |
---|---|---|---|
125000 | 4681 | 4824 | 5004 |
1400000 | 1672 | 3043 | 4738 |
5500000 | - | 1389 | 3975 |
8000000 | - | - | 3565 |
16000000 | - | - | 2679 |
29000000 | - | - | 1855 |
2. GPU memory cost of different parallel methods (MB per GPU), Tesla V100 32GB * 8. (Smaller is better)
Number of Identities in Dataset | Data Parallel | Model Parallel | Partial FC 0.1 |
---|---|---|---|
125000 | 7358 | 5306 | 4868 |
1400000 | 32252 | 11178 | 6056 |
5500000 | - | 32188 | 9854 |
8000000 | - | - | 12310 |
16000000 | - | - | 19950 |
29000000 | - | - | 32324 |
Evaluation ICCV2021-MFR and IJB-C
More details see eval.md in docs.
Test
We tested many versions of PyTorch. Please create an issue if you are having trouble.
- torch 1.6.0
- torch 1.7.1
- torch 1.8.0
- torch 1.9.0
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}
}