轉載請注明出處:
http://www.cnblogs.com/darkknightzh/p/8525287.html
論文
InsightFace : Additive Angular Margin Loss for Deep Face Recognition
https://arxiv.org/abs/1801.07698
官方mxnet代碼:
https://github.com/deepinsight/insightface
說明:沒用過mxnet,下面的代碼注釋只是純粹從代碼的角度來分析並進行注釋,如有錯誤之處,敬請諒解,並歡迎指出。
先查看sphereface,查看$\psi (\theta )$的介紹:http://www.cnblogs.com/darkknightzh/p/8524937.html
論文arcface中,定義$\psi (\theta )$為:
$\psi (\theta )=\cos ({{\theta }_{yi}}+m)$
同時對w及x均進行了歸一化,為了使得訓練能收斂,增加了一個參數s=64,最終loss如下:
$L=-\frac{1}{m}\sum\limits_{i=1}^{m}{\log \frac{{{e}^{s(\cos ({{\theta }_{yi}}+m))}}}{{{e}^{s(\cos ({{\theta }_{yi}}+m))}}+\sum\nolimits_{j=n,j\ne yi}^{n}{{{e}^{s\cos {{\theta }_{j}}}}}}}$
其中,
${{W}_{j}}=\frac{{{W}_{j}}}{\left\| {{W}_{j}} \right\|}$,${{x}_{i}}=\frac{{{x}_{i}}}{\left\| {{x}_{i}} \right\|}$,$\cos {{\theta }_{j}}=W_{j}^{T}{{x}_{i}}$
程序中先對w及x歸一化,然后通過全連接層得到cosθ,再擴大s倍,得到scosθ。
對於yi處,由於
$\cos (\theta +m)=\cos \theta \cos m-\sin \theta \sin m$
以及
$\sin \theta =\sqrt{1-{{\cos }^{2}}\theta }$
得到sinθ。
由於$\cos (\theta +m)$非單調,設置了easy_margin標志,當其為真時,使用0作為閾值,當特征和權重的cos值小於0,直接截斷;當其為假時,使用cos(pi-m)=-cos(m)作為閾值。該閾值小於0。
之后判斷時,當easy_margin為真時,若s*cos(θ+m)小於0,直接使用s*cos(θ);當easy_margin為假時,若s*cos(θ+m)小於0,使用s*cos(θ)-s*m*sin(m)。
具體的代碼如下(完整代碼見參考網址):
1 s = args.margin_s # 參數s 2 m = args.margin_m # 參數m 3 4 _weight = mx.symbol.Variable("fc7_weight", shape=(args.num_classes, args.emb_size), lr_mult=1.0) # (C,F) 5 _weight = mx.symbol.L2Normalization(_weight, mode='instance') # 對w進行歸一化 6 nembedding = mx.symbol.L2Normalization(embedding, mode='instance', name='fc1n')*s # 對x進行歸一化,並得到s*x,(B,F) 7 fc7 = mx.sym.FullyConnected(data=nembedding, weight = _weight, no_bias = True, num_hidden=args.num_classes, name='fc7') # Y=XW'+b,(B,F)*(C,F)'=(B,C),'為轉置,此處得到scos(theta) 8 9 zy = mx.sym.pick(fc7, gt_label, axis=1) # 得到fc7中gt_label位置的值。(B,1)或者(B),即當前batch中yi處的scos(theta) 10 cos_t = zy/s # 由於fc7及zy均為cos的s倍,此處除以s,得到實際的cos值。(B,1)或者(B) 11 12 cos_m = math.cos(m) 13 sin_m = math.sin(m) 14 mm = math.sin(math.pi-m)*m # sin(pi-m)*m = sin(m)*m 15 threshold = math.cos(math.pi-m) # 閾值,避免theta + m >= pi,實際上threshold < 0 16 if args.easy_margin: 17 cond = mx.symbol.Activation(data=cos_t, act_type='relu') #easy_margin=True,直接使用0作為閾值,得到超過閾值的索引 18 else: 19 cond_v = cos_t - threshold #easy_margin=False,使用threshold(負數)作為閾值。 20 cond = mx.symbol.Activation(data=cond_v, act_type='relu') # 得到超過閾值的索引 21 body = cos_t*cos_t # 通過cos*cos + sin * sin = 1, 來得到sin_theta 22 body = 1.0-body 23 sin_t = mx.sym.sqrt(body) # sin_theta 24 new_zy = cos_t*cos_m # cos(theta+m)=cos(theta)*cos(m)-sin(theta)*sin(m),此處為cos(theta)*cos(m) 25 b = sin_t*sin_m # 此處為sin(theta)*sin(m) 26 new_zy = new_zy - b # 此處為cos(theta)*cos(m)-sin(theta)*sin(m)=cos(theta+m) 27 new_zy = new_zy*s # 此處為s*cos(theta+m),擴充了s倍 28 if args.easy_margin: 29 zy_keep = zy # zy_keep為zy,即s*cos(theta) 30 else: 31 zy_keep = zy - s*mm # zy_keep為zy-s*sin(m)*m=s*cos(theta)-s*m*sin(m) 32 new_zy = mx.sym.where(cond, new_zy, zy_keep) # cond中>0的保持new_zy=s*cos(theta+m)不變,<0的裁剪為zy_keep= s*cos(theta) or s*cos(theta)-s*m*sin(m) 33 34 diff = new_zy - zy # 35 diff = mx.sym.expand_dims(diff, 1) 36 gt_one_hot = mx.sym.one_hot(gt_label, depth = args.num_classes, on_value = 1.0, off_value = 0.0) 37 body = mx.sym.broadcast_mul(gt_one_hot, diff) # 對應yi處為new_zy - zy 38 fc7 = fc7+body # 對應yi處,fc7=zy + (new_zy - zy) = new_zy,即cond中>0的為s*cos(theta+m),<0的裁剪為s*cos(theta) or s*cos(theta)-s*m*sin(m)
