insightface/recognition/partial_fc/pytorch

Parital FC

TODO

• No BUG Sampling
• Pytorch Experiments (Glint360k, 1.0/0.1)
• 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

There is a loss of accuracy in MS1MV2 with sampling, but it has the same good accuracy in Glint360k.

This is what we found.

The more identities in the training set, the closer the results from the sampling training are to those from Full Softmax training!

IJBC-Glint360K (mxnet)

+--------------+-------+--------+-------+-------+-------+
|   Methods    | 1e-05 | 0.0001 | 0.001 |  0.01 |  0.1  |
+--------------+-------+--------+-------+-------+-------+
| IJBC-1.0     | 95.92 | 97.30  | 98.13 | 98.78 | 99.28 |
+--------------+-------+--------+-------+-------+-------+
| IJBC-0.1     | 95.89 | 97.27  | 98.12 | 98.77 | 99.28 |
+--------------+-------+--------+-------+-------+-------+

IJBC-Glint360K (pytorch)

+--------------+-------+--------+-------+-------+-------+
|   Methods    | 1e-05 | 0.0001 | 0.001 |  0.01 |  0.1  |
+--------------+-------+--------+-------+-------+-------+
| IJBC-1.0     |       |        |       |       |       |
+--------------+-------+--------+-------+-------+-------+
| IJBC-0.1     |       |        |       |       |       |
+--------------+-------+--------+-------+-------+-------+

IJBC-MS1MV2 (pytorch)

+--------------+-------+--------+-------+-------+-------+
|   Methods    | 1e-05 | 0.0001 | 0.001 |  0.01 |  0.1  |
+--------------+-------+--------+-------+-------+-------+
| IJBC-1.0     | 94.22 | 96.27  | 97.61 | 98.34 | 99.08 |
+--------------+-------+--------+-------+-------+-------+
| IJBC-0.1     | 92.34 | 96.24  | 97.61 | 98.51 | 99.16 |
+--------------+-------+--------+-------+-------+-------+

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}
}