[人工智能]keras计算Generalized Dice Loss（GDL)的代码解析

这个代码在很多地方都能找到，贴出其中一个链接https://zhuanlan.zhihu.com/p/103426335

其中的代码增加上我个人理解后的注释为：

# keras在安装tensorflow后还需要单独安装
from keras import backend as K	
# y_pred表示预测的结果（Batch, Height, Width, Num_Class)
# y_true表示真实结果 （Batch, Height, Width, Num_Class)
def generalized_dice_coeff(y_true, y_pred):
    Ncl = y_pred.shape[-1]  # N classes 类别总数
    w = K.zeros(shape=(Ncl,))  # （0，0, ... ) (一共Ncl个0）
    w = K.sum(y_true, axis=(0,1,2)) # 在类别这一维度方向上进行求和，求和的结果是
     （Num_Class,), 即计算该类别上所有图片对应的标签之和，换句话说也就是统计图片上属于该类别的像素点的总数。
    w = 1/(w**2+0.000001) # 利用类别对应的像素点的总数来计算权重，总数越多反而权重越小
    # Compute gen dice coef:
    numerator = y_true*y_pred  # （Batch, Height, Width, Num_Class)
    numerator = w*K.sum(numerator,(0,1,2,3)) # （Num_Class,) * （,) = （Num_Class,),  每个类别的权重 *  True Positive总数（或者说是所有像素点对应的预测概率和真实概率之积）
    numerator = K.sum(numerator)  # 加权求和
    denominator = y_true+y_pred # （Batch, Height, Width, Num_Class)
    denominator = w*K.sum(denominator,(0,1,2,3)) # 每个类别的权重 * （所有像素点对应的预测概率和真实概率之和）
    denominator = K.sum(denominator) # 加权求和
    gen_dice_coef = 2*numerator/denominator
    return gen_dice_coef
    
def generalized_dice_loss(y_true, y_pred):
    return 1 - generalized_dice_coeff(y_true, y_pred)

自己更改为tf版本为

下面展示一些 内联代码片。

# logits  (B,H,W, C）
# labels (B, H, W,1)
def generalized_dice_coeff(logits, labels, classes):
    Ncl = classes             # 计算类别总数
    w = tf.zeros(shape=(Ncl,))
    label = labels[..., 0]
    onehot_labels = tf.one_hot(label,Ncl)
    w = tf.reduce_sum(onehot_labels, axis=(0,1,2))     # 计算所有的和，而不考虑计算某个维度上的和
    eps = 0.000001
    w = 1/(w**2+eps)

    # Compute gen dice coef:
    numerator = logits * onehot_labels
    numerator = w * tf.reduce_sum(numerator)
    numerator = tf.reduce_sum(numerator)
    denominator = logits + onehot_labels
    denominator = w*tf.reduce_sum(denominator)
    denominator = tf.reduce_sum(denominator)
    gen_dice_coef = 2*numerator/denominator
    return gen_dice_coef

def generalized_dice_loss(logits, labels, classes):
    return 1 - generalized_dice_coeff(logits, labels, classes)