[Paper Review] Partial FC: Training 10 Million Identities on a Single Machine

Xiang An, Xuhan Zhu+7 more|arXiv (Cornell University)|Oct 11, 2020

Face recognition and analysisComputer Science23 references32 citations

TL;DR

The paper introduces PPRN-based softmax approximation and a distributed training strategy to train face recognition models with tens of millions of identities on limited hardware, using only 10% of class centers per iteration while achieving state-of-the-art results.

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 memories is gradually becoming irreconcilable. In this paper, we thoroughly analyze the optimization goal of softmax-based loss functions and the difficulty of training massive identities. We find that the importance of negative classes in softmax function in face representation learning is not as high as we previously thought. The experiment demonstrates no loss of accuracy when training with only 10\% randomly sampled classes for the softmax-based loss functions, compared with training with full classes using state-of-the-art models on mainstream benchmarks. We also implement a very efficient distributed sampling algorithm, taking into account model accuracy and training efficiency, which uses only eight NVIDIA RTX2080Ti to complete classification tasks with tens of millions of identities. The code of this paper has been made available https://github.com/deepinsight/insightface/tree/master/recognition/partial_fc.

Motivation & Objective

• Motivate the challenge of training softmax-based losses with massively large identity counts under GPU memory limits.
• Propose a sampling-based softmax approximation that preserves accuracy while using only a subset of class centers.
• Develop a distributed training strategy that reduces communication and memory overhead for ultra-large-scale classification.
• Introduce and release Glint360K, a large clean face recognition dataset to support large-scale experiments.

Proposed method

• Formulate softmax loss with fixed normed features and weights, focusing on angular separation between features and class centers.
• Propose Positive Plus Randomly Negative (PPRN) sampling: always include positive class centers and randomly sample negatives, showing robustness to sampling rate.
• Distribute the approximation by storing non-overlapping subsets of W on each GPU and aggregating local logits to approximate the full softmax with reduced communication.
• Derive memory and throughput analyses comparing model parallel and the proposed approach and show scalability to 10M+ identities on multi-GPU setups.
• Provide training settings (ResNet backbones, CosFace/ArcFace losses, batch size 512, learning rates) and report performance across multiple benchmarks and large-scale datasets.

Experimental results

Research questions

• RQ1 Can a softmax-based loss for face recognition preserve accuracy when only a subset of class centers is used per iteration?
• RQ2 How can we efficiently train extremely large-scale softmax classifications (millions of identities) without exhausting GPU memory?
• RQ3 Does a distributed approximation strategy with partial class centers maintain state-of-the-art performance on standard benchmarks?
• RQ4 What is the impact of sampling strategies (PPRN vs random) on model accuracy and convergence?
• RQ5 Can a newly released large-scale dataset (Glint360K) support competitive results with minimal class-center usage?

Key findings

• Training with only 10% of class centers can achieve accuracy on par with full-class softmax on mainstream benchmarks.
• PPRN sampling (include positives and randomly sampled negatives) yields robust performance across sampling rates 0.1, 0.5, and 1.0.
• The proposed distributed approximation reduces communication and memory overhead and scales to 10M identities on eight GPUs and 100M+ identities with more GPUs.
• When trained on Glint360K, the method achieves state-of-the-art verification results on MegaFace and competitive or better results on IJB-B and IJB-C with 10% sampling.
• The approach enables practical training of ultra-large-scale face recognition models on commodity hardware (e.g., 8–64 GPUs) with substantial speedups over traditional model-parallel softmax.
• Glint360K is released as a large, cleaned dataset to support future large-scale face recognition research.