We use arcface* to evaluate SN in face recognition task. We used pytorch to reproduce the insightface.
- Arcface: Additive angular margin loss for deep face recognition J Deng, J Guo, N Xue, S Zafeiriou - arXiv preprint arXiv:1801.07698, 2018
./train.sh configs/config_resnet50bn.yaml
Download the pretrained models from Model Zoo and put them into the {repo_root}/face_recognition/data/pretrained_model
./test.sh configs/config_resnet50bn.yaml
We provide models pretrained with SN on MS1M-ArcFace, and compare to those pretrained with BN as reference. If you use these models in research, please cite the SN paper. The configuration of SN is denoted as (#GPUs, #images per GPU).
| Model | MegaFace(%) | Epochs | LR Scheduler | Weight Decay | Download |
|---|---|---|---|---|---|
| ResNet100+SYNCSN (16,32) | 98.51% | 20 | Initial lr=0.1 decay=0.1 steps[12,15,18] | 5e-4 | [Google Drive] [Baidu Pan] |
| ResNet100+SN (10,52) | 98.10% | 20 | Initial lr=0.1 decay=0.1 steps[12,15,18] | 5e-4 | [Google Drive] [Baidu Pan] |
| ResNet100+BN (8,64) | 98.29% | 20 | Initial lr=0.1 decay=0.1 steps[12,15,18] | 5e-4 | [Google Drive] [Baidu Pan] |
| ResNet50+SYNCSN (16,32) | 97.94% | 20 | Initial lr=0.1 decay=0.1 steps[12,15,18] | 5e-4 | [Google Drive] [Baidu Pan] |
| ResNet50+SYNCSN+ARGMAX (8,64) | 98.26% | 10 | Initial lr=0.001 decay=0.1 steps[5,] | 5e-4 | [Google Drive] [Baidu Pan] |
| ResNet50+SN (8,64) | 97.84% | 20 | Initial lr=0.1 decay=0.1 steps[12,15,18] | 5e-4 | [Google Drive] [Baidu Pan] |
| ResNet50+BN (8,64) | 97.59% | 20 | Initial lr=0.1 decay=0.1 steps[12,15,18] | 5e-4 | [Google Drive] [Baidu Pan] |