def custom_cross_entropy(self, y_true, y_pred):
# y_true has the payoffs in the last dimension
y_true, payoffs = splitter(y_true)
if self.method == 'lay':
tp_weight = K.abs(payoffs)
fp_weight = K.abs(payoffs)
tn_weight = 1
fn_weight = 0.95
elif self.method == 'back':
tp_weight = K.abs(payoffs) # opportunity cost
tn_weight = 0 # opportunity cost
fp_weight = 1 # cost
fn_weight = K.abs(payoffs) # cost
loss = -K.mean(
fn_weight * y_true * K.log(y_pred + _EPSILON)
+ # fn cost (not backing if it should)
fp_weight * (1 - y_true) *
K.log(1 - y_pred + _EPSILON) # fp cost (backing the wrong one)
# + tp_weight * y_true * K.log(1 - y_pred + _EPSILON) # tp (correctly backing)
# + tn_weight * (1 - y_true) * K.log(y_pred + _EPSILON) # tn (correctly not backing)
)
return loss
def triplet_loss_v2(y_true, y_pred, N=512, beta=512, epsilon=1e-8):
"""
参考：https://towardsdatascience.com/lossless-triplet-loss-7e932f990b24
！！！需要输出层使用sigmoid激活函数！
Implementation of the triplet loss function
Arguments:
y_true -- true labels, required when you define a loss in Keras, you don't need it in this function.
y_pred -- python list containing three objects:
anchor -- the encodings for the anchor data
positive -- the encodings for the positive data (similar to anchor)
negative -- the encodings for the negative data (different from anchor)
N -- The number of dimension
beta -- The scaling factor, N is recommended
epsilon -- The Epsilon value to prevent ln(0)
Returns:
loss -- real number, value of the loss
"""
anc, pos, neg = y_pred[:, 0:N], y_pred[:, N:N * 2], y_pred[:, N * 2:]
# 欧式距离
pos_dist = K.sum(K.square(anc - pos), axis=-1, keepdims=True)
neg_dist = K.sum(K.square(anc - neg), axis=-1, keepdims=True)
pos_dist = -K.log(-(pos_dist / beta) + 1 + epsilon)
neg_dist = -K.log(-((N - neg_dist) / beta) + 1 + epsilon)
# compute loss
loss = neg_dist + pos_dist
return loss
def reed_hard_loss(y_true, y_pred):
'''Expects a binary class matrix instead of a vector of scalar classes.
'''
return -K.log(K.max(y_pred, axis=1, keepdims=True) + 1e-8)
def reed_soft_loss(y_true, y_pred):
'''Expects a binary class matrix instead of a vector of scalar classes.
'''
return -K.batch_dot(y_pred, K.log(y_pred + 1e-8), axes=(1, 1))
def jaccard_coef_loss(y_true, y_pred):
# __author__ = Vladimir Iglovikov
return -K.log(jaccard_coef(y_true, y_pred)) + binary_crossentropy(y_pred, y_true)