insightface/recognition/partial_fc/README.md

Partial-FC

Partial FC is a distributed deep learning training framework for face recognition. The goal of Partial FC is to facilitate large-scale classification task (e.g. 10 or 100 million identities). It is much faster than the model parallel solution and there is no performance drop.

Contents

Partial FC

• Largest Face Recognition Dataset: Glint360k
• Distributed Training Performance
• FAQ
• Citation

Glint360K

We clean, merge, and release the largest and cleanest face recognition dataset Glint360K, which contains 17091657 images of 360232 individuals. By employing the Patial FC training strategy, baseline models trained on Glint360K can easily achieve state-of-the-art performance. Detailed evaluation results on the large-scale test set (e.g. IFRT, IJB-C and Megaface) are as follows:

Evaluation on IFRT

r denotes the sampling rate of negative class centers.

Backbone	Dataset	African	Caucasian	Indian	Asian	ALL
R50	MS1M-V3	76.24	86.21	84.44	37.43	71.02
R124	MS1M-V3	81.08	89.06	87.53	38.40	74.76
R100	Glint360k(r=1.0)	89.50	94.23	93.54	65.07	88.67
R100	Glint360k(r=0.1)	90.45	94.60	93.96	63.91	88.23

Evaluation on IJB-C and Megaface

We employ ResNet100 as the backbone and CosFace (m=0.4) as the loss function. TAR@FAR=1e-4 is reported on the IJB-C datasets, and TAR@FAR=1e-6 is reported on the Megaface dataset.

Test Dataset	IJB-C	Megaface_Id	Megaface_Ver
MS1MV2	96.4	98.3	98.6
Glint360k	97.3	99.1	99.1

License

The Glint360K dataset (and the models trained with this dataset) are available for non-commercial research purposes only.

Download

• Baidu Drive (code:o3az)
• Magnet URI: magnet:?xt=urn:btih:E5F46EE502B9E76DA8CC3A0E4F7C17E4000C7B1E&dn=glint360k

Refer to the following command to unzip.

cat glint360k_* | tar -xzvf -

# Don't forget the last '-'!

# cf7433cbb915ac422230ba33176f4625  glint360k_00
# 589a5ea3ab59f283d2b5dd3242bc027a  glint360k_01
# 8d54fdd5b1e4cd55e1b9a714d76d1075  glint360k_02
# cd7f008579dbed9c5af4d1275915d95e  glint360k_03
# 64666b324911b47334cc824f5f836d4c  glint360k_04
# a318e4d32493dd5be6b94dd48f9943ac  glint360k_05
# c3ae1dcbecea360d2ec2a43a7b6f1d94  glint360k_06
# md5sum:
# 5d9cd9f262ec87a5ca2eac5e703f7cdf train.idx
# 8483be5af6f9906e19f85dee49132f8e train.rec

Use unpack_glint360k.py to unpack.

Pretrain models

• Baidu Drive (code:befi)
• Google Drive

Docker

Coming soon.

Training Speed Benchmark

1. Train Glint360K Using MXNET

Backbone	GPU	FP16	BatchSize / it	Throughput img / sec
R100	8 * Tesla V100-SXM2-32GB	False	64	1748
R100	8 * Tesla V100-SXM2-32GB	True	64	3357
R100	8 * Tesla V100-SXM2-32GB	False	128	1847
R100	8 * Tesla V100-SXM2-32GB	True	128	3867
R50	8 * Tesla V100-SXM2-32GB	False	64	2921
R50	8 * Tesla V100-SXM2-32GB	True	64	5428
R50	8 * Tesla V100-SXM2-32GB	False	128	3045
R50	8 * Tesla V100-SXM2-32GB	True	128	6112

Performance On Million Identities

We neglect the influence of IO. All experiments use mixed-precision training, and the backbone is ResNet50.

1 Million Identities On 8 RTX2080Ti

Method	GPUs	BatchSize	Memory/M	Throughput img/sec	W
Model Parallel	8	1024	10408	2390	GPU
Partial FC(Ours)	8	1024	8100	2780	GPU

10 Million Identities On 64 RTX2080Ti

Method	GPUs	BatchSize	Memory/M	Throughput img/sec	W
Model Parallel	64	2048	9684	4483	GPU
Partial FC(Ours)	64	4096	6722	12600	GPU

TODO

• Mixed precision training (pytorch)
• Pipeline Parallel (pytorch)
• Docker (include mxnet and pytorch)
• A Wonderful Documents

FAQ

Glint360K's Face Alignment Settings?

We use a same alignment setting with MS1MV2, code is here.

Why update Glint360K, is there a bug in the previous version?

In the previous version of Glint360K, there is no bug when using softmax training, but there is a bug in triplet training. In the latest Glint360k, this bug has been fixed.

Dataset in Google Drive or Dropbox?

The torrent has been released.

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}
}