insightface/recognition/partial_fc/pytorch

Parital FC

The sampled version of pytorch is still being improved, and there is a bug which the accuracy can not reach mxnet version, but you can use the unsampled version of pytorch first.

If you want to reproduce the accuracy in the paper, it is strongly recommended to use mxnet first. All experiments in the paper are done by mxnet.

Pytorch 目前是还是预览版本，模型并行是没问题的，但是0.1的采样暂时无法使用，
如果要使用采样(复现论文中的精度)，强烈建议优先使用mxnet, 所有论文的实验均是mxnet完成的。
我们会马上修复这个bug。

Insightface 社区需要大家一起贡献才会变得更好，欢迎大家提交Pull Request.

TODO

• No BUG Sampling
• Mixed precision training
• Pipeline Parallel
• Checkpoint
• Docker
• A Wonderful Documents

How to run

cuda=10.1
pytorch==1.6.0
pip install -r requirement.txt

bash run.sh

使用 bash run.sh 这个命令运行。

Results

MS1MV2-IJBC

+--------------+-------+-------+--------+-------+-------+-------+
|   Methods    | 1e-06 | 1e-05 | 0.0001 | 0.001 |  0.01 |  0.1  |
+--------------+-------+-------+--------+-------+-------+-------+
| cosface-IJBC | 86.63 | 94.22 | 96.37  | 97.61 | 98.34 | 99.08 |
+--------------+-------+-------+--------+-------+-------+-------+

Citation

If you find Partial-FC or Glint360K useful in your research, please consider to cite the following related paper:

Partial FC

@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}
}