Partial FC: Training 10 Million Identities on a Single Machine

Abstract

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.