Pytorch implementation of CosFace
- Deep Learning Platform: PyTorch 0.4.1
- OS: CentOS Linux release 7.5
- Language: Python 2.7
- CUDA: 8.0
- Database:
WebFaceorVggFace2(You should first complete the data preprocessing section by following these steps https://github.com/wy1iu/sphereface#part-1-preprocessing) - Network:
sphere20,sphere64,LResnet50E-IR(In ArcFace paper)
Single model trained on CAISA-WebFace achieves ~99.2% accuracy on LFW (Link: https://pan.baidu.com/s/1uOBATynzBTzZwrIKC4kcAA Password: 69e6)
Note: Pytorch 0.4 seems to be very different from 0.3, which leads me to not fully reproduce the previous results. Currently still adjusting parameters....
The initialization of the fully connected layer does not use Xavier but is more conducive to model convergence.
| Network | Hyper-parameter | Accuracy on LFW |
|---|---|---|
| Sphere20 | s=30, m=0.35 | 99.08% |
| Sphere20 | s=30, m=0.40 | 99.23% |
| LResnet50E-IR(In ArcFace paper) | s=30, m=0.35 | 99.45% |