你要的答案或许都在这里:小鹏的博客目录
paper:
CosFace: Large Margin Cosine Loss(MLCL) for Deep Face Recognition
下载地址: https://arxiv.org/pdf/1801.09414.pdf
论文中的cos loss:
cos loss 的 TF 实现:
# coding=utf-8 import tensorflow as tf import numpy as np def py_func(func, inp, Tout, stateful = True, name=None, grad_func=None): rand_name = 'PyFuncGrad' + str(np.random.randint(0,1E+8)) tf.RegisterGradient(rand_name)(grad_func) g = tf.get_default_graph() with g.gradient_override_map({'PyFunc':rand_name}): return tf.py_func(func,inp,Tout,stateful=stateful, name=name) def coco_forward(xw, y, m, name=None): #pdb.set_trace() xw_copy = xw.copy() num = len(y) orig_ind = range(num) xw_copy[orig_ind,y] -= m return xw_copy def coco_help(grad,y): grad_copy = grad.copy() return grad_copy def coco_backward(op, grad): y = op.inputs[1] m = op.inputs[2] grad_copy = tf.py_func(coco_help,[grad,y],tf.float32) return grad_copy,y,m def coco_func(xw,y,m, name=None): with tf.op_scope([xw,y,m],name,"Coco_func") as name: coco_out = py_func(coco_forward,[xw,y,m],tf.float32,name=name,grad_func=coco_backward) return coco_out def cos_loss(x, y, num_cls, reuse=False, alpha=0.25, scale=64,name = 'cos_loss'): ''''' x: B x D - features y: B x 1 - labels num_cls: 1 - total class number alpah: 1 - margin scale: 1 - scaling paramter ''' # define the classifier weights xs = x.get_shape() y = tf.reshape(tf.cast(y, dtype = tf.int32),[-1]) with tf.variable_scope('centers_var',reuse=reuse) as center_scope: w = tf.get_variable("centers", [xs[1], num_cls], dtype=tf.float32, initializer=tf.contrib.layers.xavier_initializer(),trainable=True) #normalize the feature and weight #(N,D) x_feat_norm = tf.nn.l2_normalize(x,1,1e-10) #(D,C) w_feat_norm = tf.nn.l2_normalize(w,0,1e-10) # get the scores after normalization #(N,C) xw_norm = tf.matmul(x_feat_norm, w_feat_norm) #value = tf.identity(xw) #substract the marigin and scale it value = coco_func(xw_norm,y,alpha) * scale # compute the loss as softmax loss cos_loss = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(labels=y, logits=value)) return cos_loss