def compute_loss(self, inputs, mask=None):
pred, ytrue = inputs
acc = keras.metrics.sparse_categorical_accuracy(ytrue, pred)
self.add_metric(acc, name='clf_acc')
ytrue = K.cast(ytrue, 'int32')
ytrue = K.one_hot(ytrue, num_classes=num_classes)
ytrue = K.reshape(ytrue, (-1, num_classes))
loss = ytrue * K.log(pred + K.epsilon()) + (1 - ytrue) * K.log(1 - pred + K.epsilon())
loss = -K.mean(loss)
loss = loss * self.alpha
self.add_metric(loss, name='clf_loss')
return loss
def mlm_loss(inputs):
"""计算loss的函数，需要封装为一个层
"""
y_true, y_pred, mask = inputs
loss = K.sparse_categorical_crossentropy(y_true, y_pred, from_logits=True)
loss = K.sum(loss * mask) / (K.sum(mask) + K.epsilon())
return loss
def mlm_acc(inputs):
"""计算准确率的函数，需要封装为一个层
"""
y_true, y_pred, mask = inputs
y_true = K.cast(y_true, floatx)
acc = keras.metrics.sparse_categorical_accuracy(y_true, y_pred)
acc = K.sum(acc * mask) / (K.sum(mask) + K.epsilon())
return acc
def __init__(self,
learning_rate=0.001,
beta_1=0.9,
beta_2=0.99,
epsilon=1e-6,
bias_correct=True,
**kwargs):
kwargs['name'] = kwargs.get('name', 'Adam')
super(Adam, self).__init__(**kwargs)
self._set_hyper('learning_rate', learning_rate)
self._set_hyper('beta_1', beta_1)
self._set_hyper('beta_2', beta_2)
self.epsilon = epsilon or K.epsilon()
self.bias_correct = bias_correct
def compute_loss_of_scl(self, inputs, mask=None):
y_pred, y_true = inputs
label_mask = self.get_label_mask(y_true)
y_pred = K.l2_normalize(y_pred, axis=1) # 特征向量归一化
similarities = K.dot(y_pred, K.transpose(y_pred)) # 相似矩阵
similarities = similarities - K.eye(K.shape(y_pred)[0]) * 1e12 # 排除对角线,即 i == j
similarities = similarities / self.T # Temperature scale
similarities = K.exp(similarities) # exp
sum_similarities = K.sum(similarities, axis=-1, keepdims=True) # sum i != k
scl = similarities / sum_similarities
scl = K.log(scl + K.epsilon()) # sum log
scl = -K.sum(scl * label_mask, axis=1, keepdims=True) / (K.sum(label_mask, axis=1, keepdims=True) + K.epsilon())
return K.mean(scl)
def __init__(self,
learning_rate=0.001,
beta_1=0.9,
beta_2=0.999,
amsgrad=False,
**kwargs):
self.initial_decay = kwargs.pop('decay', 0.0)
self.epsilon = kwargs.pop('epsilon', K.epsilon())
learning_rate = kwargs.pop('lr', learning_rate)
super(AdaBelief, self).__init__(**kwargs)
with K.name_scope(self.__class__.__name__):
self.iterations = K.variable(0, dtype='int64', name='iterations')
self.learning_rate = K.variable(learning_rate,
name='learning_rate')
self.beta_1 = K.variable(beta_1, name='beta_1')
self.beta_2 = K.variable(beta_2, name='beta_2')
self.decay = K.variable(self.initial_decay, name='decay')
self.amsgrad = amsgrad