now pls move to the new mobilefacenet-V2(http://github.com/qidiso/mobilefacenet-V2) i modified the network ,and get 99.683% in LFW and 96.717 in agedb30.

trainning for a week long time .now i get better result.

lr-batch-epoch: 0.001 7999 0 testing verification.. (12000, 128) infer time 12.323731 [lfw][8000]XNorm: 11.118196 [lfw][8000]Accuracy-Flip: 0.99583+-0.00375 testing verification.. (14000, 128) infer time 14.580451 [cfp_fp][8000]XNorm: 9.335661 [cfp_fp][8000]Accuracy-Flip: 0.88786+-0.01615 testing verification.. (12000, 128) infer time 12.362448 [agedb_30][8000]XNorm: 11.044563 [agedb_30][8000]Accuracy-Flip: 0.96083+-0.00827 saving 4 INFO:root:Saved checkpoint to "../models/MobileFaceNet/model-y1-arcface-0004.params" [8000]Accuracy-Highest: 0.96133

or

infer time 12.754381 [lfw][46000]XNorm: 11.113467 [lfw][46000]Accuracy-Flip: 0.99550+-0.00395 testing verification.. (14000, 128) infer time 14.228613 [cfp_fp][46000]XNorm: 9.320184 [cfp_fp][46000]Accuracy-Flip: 0.89257+-0.01589 testing verification.. (12000, 128) infer time 12.060387 [agedb_30][46000]XNorm: 11.055198 [agedb_30][46000]Accuracy-Flip: 0.96117+-0.00746 saving 23

前言

本文主要记录下复现mobilefacenet的流程，参考mobilefacenet作者月生给的基本流程，基于insightface的4月27日

4bc813215a4603474c840c85fa2113f5354c7180

版本代码在2块1080单显卡训练调试。

训练步骤

1.拉取配置insightface工程的基础环境；

2.softmax loss初调：lr0.1，softmax的fc7配置wd_mult=10.0和no_bias=True,训练4万步;

切换到src目录下，修改train_softmax.py： 179-182行：

  if args.loss_type==0: #softmax
    _bias = mx.symbol.Variable('fc7_bias', lr_mult=2.0, wd_mult=0.0)
    fc7 = mx.sym.FullyConnected(data=embedding, weight = _weight, bias = _bias, num_hidden=args.num_classes, name='fc7')

修改为：

  if args.loss_type==0: #softmax
    #_bias = mx.symbol.Variable('fc7_bias', lr_mult=2.0, wd_mult=0.0)
    # fc7 = mx.sym.FullyConnected(data=embedding, weight = _weight, bias = _bias, num_hidden=args.num_classes, name='fc7')
    fc7 = mx.sym.FullyConnected(data=embedding, weight = _weight, no_bias=True, num_hidden=args.num_classes, name='fc7')

363行：

 if args.network[0]=='r' or args.network[0]=='y':

修改为：

 if args.network[0]=='r' :

这样保证uniform初始化；

运行：

CUDA_VISIBLE_DEVICES='0' python -u train_softmax.py --network y1 --ckpt 2 --loss-type 0 --per-batch-size 512 --emb-size 128 --fc7-wd-mult 10  --data-dir  ../datasets/faces_ms1m_112x112  --prefix ../models/MobileFaceNet/model-y1-softmax

3.agedb（先从s=128开始训练）：从2步训练好的模型继续用arcface loss训练，s=128, m=0.5，起始lr=0.1，在[20000, 30000, 40000]步降低lr，这时能得到lfw acc 0.9955左右，agedb-30 acc 0.96.250左右的最终模型。

CUDA_VISIBLE_DEVICES='0' python -u train_softmax.py --network y1 --ckpt 2 --loss-type 4 --lr 0.001 --lr-steps 20000,30000,40000 --emb-size 128 --per-batch-size 512 --margin-s 128 --data-dir ../datasets/faces_ms1m_112x112 --pretrained ../models/MobileFaceNet/model-y1-arcface,80 --prefix ../models/MobileFaceNet/model-y1-arcface

4.arcface loss调试：s=64, m=0.5, 起始lr=0.001，在[80000, 120000, 140000, 160000]步处降低lr，总共训练16万步。这时，LFW acc能到0.99583左右，agedb-30 acc能到0.96167以上。

切换到src目录下：

CUDA_VISIBLE_DEVICES='0' python -u train_softmax.py --network y1 --ckpt 2 --loss-type 4 --lr-steps 80000,120000,140000,160000 --emb-size 128 --per-batch-size 512 --data-dir ../datasets/faces_ms1m_112x112 --pretrained ../models/MobileFaceNet/model-y1-softmax,20 --prefix ../models/MobileFaceNet/model-y1-arcface

##转为caffe 后的精度测试