Evgeny Smirnov, Senior Researcher of S peech Technology Center
SphereFace, ArcFace, CosFace, D- Softmax, CurricularFace, etc.
Problem : Linear increase of memory and computation consumption with an increasing number of classes Classifier weights for millions of classes do not fit in ordinal GPU memory, and the loss function computation time is too large for practical use.
Zhu, Zheng, et al. "Webface260m: A benchmark unveiling the power of million-scale deep face recognition", CVPR 2021
Zhu, Zheng, et al. "Webface260m: A benchmark unveiling the power of million-scale deep face recognition", CVPR 2021
Zhu, Zheng, et al. "Webface260m: A benchmark unveiling the power of million-scale deep face recognition", CVPR 2021
Sampled Softmax variants : D- Softmax -K ( He et al, 2020) PPRN ( An et al, 2020) Problems : Memory consumption GPU- To -CPU transfer Prototype obsolescence He, Lanqing, et al. " Softmax dissection : Towards understanding intra-and inter-class objective for embedding learning ", AAAI 2020 An, Xiang, et al. " Partial FC: Training 10 Million Identities on a Single Machine ", arXiv:2010.05222
GPU memory is fixed, but we still need to keep classifier weights ( prototypes ) for all classes in the dataset in the ( non -GPU) memory. We also need to transfer sampled classifier weights to GPU and back at each training iteration.
Only sampled classifier weights are updated Others remain fixed between training iterations When the number of classes is large, individual classifier weights are sampled and updated too rarely, so they become obsolete and do not represent their classes correctly anymore
Smirnov et al. " Prototype Memory for Large-scale Face Representation Learning ", IEEE Access 2022
Smirnov et al. " Prototype Memory for Large-scale Face Representation Learning ", IEEE Access 2022
Smirnov et al. " Prototype Memory for Large-scale Face Representation Learning ", IEEE Access 2022
Smirnov et al. " Prototype Memory for Large-scale Face Representation Learning ", IEEE Access 2022
Smirnov et al. " Prototype Memory for Large-scale Face Representation Learning ", IEEE Access 2022
Smirnov et al. " Prototype Memory for Large-scale Face Representation Learning ", IEEE Access 2022
Smirnov et al. " Prototype Memory for Large-scale Face Representation Learning ", IEEE Access 2022
Sampled Softmax Prototype Memory Smirnov et al. " Prototype Memory for Large-scale Face Representation Learning ", IEEE Access 2022
Smirnov et al. " Prototype Memory for Large-scale Face Representation Learning ", IEEE Access 2022
Smirnov et al. " Prototype Memory for Large-scale Face Representation Learning ", IEEE Access 2022
Prototype Memory is a novel method for training face recognition models on large datasets. It is fast, memory-efficient, and more accurate than other similar methods. It is useful for preventing the problem of “ prototype obsolescence ”, which emerges in large-scale datasets.
Moscow 59/2 Zemlyanoy Val St. 109004 +7 495 669 7440 stc-int@speechpro.com St. Petersburg Vyborgskaya Embankment 45, Bldg. E 194044 +7 812 325 8848 stc-int@speechpro.com Evgeny Smirnov, Senior R esearcher of Speech Technology Center e- mail : smirnov-e@speechpro.com
