Focal loss Tensorflow实现

最新推荐文章于 2022-10-20 22:41:14 发布

chengjinpei

最新推荐文章于 2022-10-20 22:41:14 发布

阅读量697

点赞数

分类专栏：数据预处理文章标签：自然语言处理

本文链接：https://blog.csdn.net/chengjinpei/article/details/107915312

版权

数据预处理专栏收录该内容

7 篇文章 1 订阅

订阅专栏

数据不均衡--Focal loss Tensorflow实现

def focal_loss(prediction_tesnsor,target_tensor,weights = None,alpha = 0.25,gamma=2):
	"""
	FL = -alpha *(z-p)^gamma*log(p) - (1-alpha)*p^gamma *log(1-p)
	which alpha = 0.25,gamma = 2,p = sigmoid(x),z = target_tensor
	"""
	sigmoid_p = tf.nn.sigmoid(prediction_tesnsor)
	zeros = array_ops.zeros.like(sigmoid_p,dtype = sigmoid_p.type)

	#对比正标签
	pos_p_sub = array_ops.where(target_tensor > zeros,target_tensor - sigmoid_p,zeros)

	#对于负标签
	neg_p_sub = array_ops.where(target_tensor > zeros,zeros,sigmoid_p)

	fl_loss = -alpha * (pos_p_sub**gamma)*tf.log(tf.clip_by_value(sigmoid_p,1e-8,1.0))\
			- (1-alpha) * (neg_p_sub**gamma)*tf.log(tf.clip_by_value(1-sigmoid_p,1e-8,1.0))
	return tf.reduce_sum(fl_loss)