(原)InsightFace及其mxnet代码
转载请注明出处:
http://www.cnblogs.com/darkknightzh/p/8525287.html
论文
InsightFace : Additive Angular Margin Loss for Deep Face Recognition
https://arxiv.org/abs/1801.07698
官方mxnet代码:
https://github.com/deepinsight/insightface
说明:没用过mxnet,下面的代码注释只是纯粹从代码的角度来分析并进行注释,如有错误之处,敬请谅解,并欢迎指出。
先查看sphereface,查看$\psi (\theta )$的介绍:http://www.cnblogs.com/darkknightzh/p/8524937.html
论文arcface中,定义$\psi (\theta )$为:
$\psi (\theta )=\cos ({{\theta }_{yi}}+m)$
同时对w及x均进行了归一化,为了使得训练能收敛,增加了一个参数s=64,最终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,
${{W}_{j}}=\frac{{{W}_{j}}}{\left\| {{W}_{j}} \right\|}$,${{x}_{i}}=\frac{{{x}_{i}}}{\left\| {{x}_{i}} \right\|}$,$\cos {{\theta }_{j}}=W_{j}^{T}{{x}_{i}}$
程序中先对w及x归一化,然后通过全连接层得到cosθ,再扩大s倍,得到scosθ。
对于yi处,由于
$\cos (\theta +m)=\cos \theta \cos m-\sin \theta \sin m$
以及
$\sin \theta =\sqrt{1-{{\cos }^{2}}\theta }$
得到sinθ。
由于$\cos (\theta +m)$非单调,设置了easy_margin标志,当其为真时,使用0作为阈值,当特征和权重的cos值小于0,直接截断;当其为假时,使用cos(pi-m)=-cos(m)作为阈值。该阈值小于0。
之后判断时,当easy_margin为真时,若s*cos(θ+m)小于0,直接使用s*cos(θ);当easy_margin为假时,若s*cos(θ+m)小于0,使用s*cos(θ)-s*m*sin(m)。
具体的代码如下(完整代码见参考网址):
1 s = args.margin_s # 参数s 2 m = args.margin_m # 参数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') # 对w进行归一化 6 nembedding = mx.symbol.L2Normalization(embedding, mode='instance', name='fc1n')*s # 对x进行归一化,并得到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),'为转置,此处得到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),即当前batch中yi处的scos(theta) 10 cos_t = zy/s # 由于fc7及zy均为cos的s倍,此处除以s,得到实际的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) # 阈值,避免theta + m >= pi,实际上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作为阈值,得到超过阈值的索引 18 else: 19 cond_v = cos_t - threshold #easy_margin=False,使用threshold(负数)作为阈值。 20 cond = mx.symbol.Activation(data=cond_v, act_type='relu') # 得到超过阈值的索引 21 body = cos_t*cos_t # 通过cos*cos + sin * sin = 1, 来得到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),此处为cos(theta)*cos(m) 25 b = sin_t*sin_m # 此处为sin(theta)*sin(m) 26 new_zy = new_zy - b # 此处为cos(theta)*cos(m)-sin(theta)*sin(m)=cos(theta+m) 27 new_zy = new_zy*s # 此处为s*cos(theta+m),扩充了s倍 28 if args.easy_margin: 29 zy_keep = zy # zy_keep为zy,即s*cos(theta) 30 else: 31 zy_keep = zy - s*mm # zy_keep为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)不变,<0的裁剪为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) # 对应yi处为new_zy - zy 38 fc7 = fc7+body # 对应yi处,fc7=zy + (new_zy - zy) = new_zy,即cond中>0的为s*cos(theta+m),<0的裁剪为s*cos(theta) or s*cos(theta)-s*m*sin(m)
posted on 2018-03-07 20:55 darkknightzh 阅读(13080) 评论(0) 编辑 收藏 举报