TY - GEN
T1 - Partial FC
T2 - 18th IEEE/CVF International Conference on Computer Vision Workshops, ICCVW 2021
AU - An, Xiang
AU - Zhu, Xuhan
AU - Gao, Yuan
AU - Xiao, Yang
AU - Zhao, Yongle
AU - Feng, Ziyong
AU - Wu, Lan
AU - Qin, Bin
AU - Zhang, Ming
AU - Zhang, Debing
AU - Fu, Ying
N1 - Publisher Copyright:
© 2021 IEEE.
PY - 2021
Y1 - 2021
N2 - Face recognition has been an active and vital topic among computer vision community for a long time. Previous researches mainly focus on loss functions used for facial feature extraction network, among which the improvements of softmax-based loss functions greatly promote the performance of face recognition. However, the contradiction between the drastically increasing number of face identities and the shortage of GPU memory is gradually becoming irreconcilable. In this work, we theoretically analyze the upper limit of model parallelism in face recognition in the first place. Then we propose a load-balanced sparse distributed classification training method, Partial FC, which is capable of using a machine with only 8 Nvidia Tesla V100 GPUs to implement training on a face recognition data set with up to 29 million IDs. Furthermore, we are able to train on data set with 100 million IDs in 64 RTX2080Ti GPUs. We have verified the effectiveness of Partial FC in 8 mainstream face recognition trainsets, and find that Partial FC is effective in all face recognition training sets. The code of this paper has been made available at https://github.com/deepinsight/insightface/tree/master/recognition/partial_fc.
AB - Face recognition has been an active and vital topic among computer vision community for a long time. Previous researches mainly focus on loss functions used for facial feature extraction network, among which the improvements of softmax-based loss functions greatly promote the performance of face recognition. However, the contradiction between the drastically increasing number of face identities and the shortage of GPU memory is gradually becoming irreconcilable. In this work, we theoretically analyze the upper limit of model parallelism in face recognition in the first place. Then we propose a load-balanced sparse distributed classification training method, Partial FC, which is capable of using a machine with only 8 Nvidia Tesla V100 GPUs to implement training on a face recognition data set with up to 29 million IDs. Furthermore, we are able to train on data set with 100 million IDs in 64 RTX2080Ti GPUs. We have verified the effectiveness of Partial FC in 8 mainstream face recognition trainsets, and find that Partial FC is effective in all face recognition training sets. The code of this paper has been made available at https://github.com/deepinsight/insightface/tree/master/recognition/partial_fc.
UR - http://www.scopus.com/inward/record.url?scp=85118715880&partnerID=8YFLogxK
U2 - 10.1109/ICCVW54120.2021.00166
DO - 10.1109/ICCVW54120.2021.00166
M3 - Conference contribution
AN - SCOPUS:85118715880
T3 - Proceedings of the IEEE International Conference on Computer Vision
SP - 1445
EP - 1449
BT - Proceedings - 2021 IEEE/CVF International Conference on Computer Vision Workshops, ICCVW 2021
PB - Institute of Electrical and Electronics Engineers Inc.
Y2 - 11 October 2021 through 17 October 2021
ER -