├── README.md
└── focal_loss.py


/README.md:
--------------------------------------------------------------------------------
 1 | # focal-loss
 2 | 
 3 | Tensorflow实现何凯明的Focal Loss, 该损失函数主要用于解决分类问题中的类别不平衡
 4 | 
 5 | focal_loss_sigmoid: 二分类loss
 6 | 
 7 | focal_loss_softmax: 多分类loss
 8 | 
 9 | Reference Paper : Focal Loss for Dense Object Detection
10 | 


--------------------------------------------------------------------------------
/focal_loss.py:
--------------------------------------------------------------------------------
 1 | # -*- coding: utf-8 -*-
 2 | import tensorflow as tf
 3 | 
 4 | """
 5 | Tensorflow实现何凯明的Focal Loss, 该损失函数主要用于解决分类问题中的类别不平衡
 6 | focal_loss_sigmoid: 二分类loss
 7 | focal_loss_softmax: 多分类loss
 8 | Reference Paper : Focal Loss for Dense Object Detection
 9 | """
10 | 
11 | def focal_loss_sigmoid(labels,logits,alpha=0.25,gamma=2):
12 |     """
13 |     Computer focal loss for binary classification
14 |     Args:
15 |       labels: A int32 tensor of shape [batch_size].
16 |       logits: A float32 tensor of shape [batch_size].
17 |       alpha: A scalar for focal loss alpha hyper-parameter. If positive samples number
18 |       > negtive samples number, alpha < 0.5 and vice versa.
19 |       gamma: A scalar for focal loss gamma hyper-parameter.
20 |     Returns:
21 |       A tensor of the same shape as `lables`
22 |     """
23 |     y_pred=tf.nn.sigmoid(logits)
24 |     labels=tf.to_float(labels)
25 |     L=-labels*(1-alpha)*((1-y_pred)*gamma)*tf.log(y_pred)-\
26 |       (1-labels)*alpha*(y_pred**gamma)*tf.log(1-y_pred)
27 |     return L
28 | 
29 | def focal_loss_softmax(labels,logits,gamma=2):
30 |     """
31 |     Computer focal loss for multi classification
32 |     Args:
33 |       labels: A int32 tensor of shape [batch_size].
34 |       logits: A float32 tensor of shape [batch_size,num_classes].
35 |       gamma: A scalar for focal loss gamma hyper-parameter.
36 |     Returns:
37 |       A tensor of the same shape as `lables`
38 |     """
39 |     y_pred=tf.nn.softmax(logits,dim=-1) # [batch_size,num_classes]
40 |     labels=tf.one_hot(labels,depth=y_pred.shape[1])
41 |     L=-labels*((1-y_pred)**gamma)*tf.log(y_pred)
42 |     L=tf.reduce_sum(L,axis=1)
43 |     return L
44 | 
45 | if __name__ == '__main__':
46 |     logits=tf.random_uniform(shape=[5],minval=-1,maxval=1,dtype=tf.float32)
47 |     labels=tf.Variable([0,1,0,0,1])
48 |     loss1=focal_loss_sigmoid(labels=labels,logits=logits)
49 | 
50 |     logits2=tf.random_uniform(shape=[5,4],minval=-1,maxval=1,dtype=tf.float32)
51 |     labels2=tf.Variable([1,0,2,3,1])
52 |     loss2=focal_loss_softmax(labels==labels2,logits=logits2)
53 | 
54 |     with tf.Session() as sess:
55 |         sess.run(tf.global_variables_initializer())
56 |         print sess.run(loss1)
57 |         print sess.run(loss2)
58 | 
59 | 


--------------------------------------------------------------------------------