def call(self, values):
"""Accumulates statistics for computing the reduction metric.
Arguments:
values: array-like
Per-example value.
Returns:
A scalar corresponding to the metric value.
"""
if self.reduction == SUM:
return M.sum(values)
if self.reduction in [WEIGHTED_MEAN, SUM_OVER_BATCH_SIZE]:
return M.sum(values) / values.size
raise NotImplementedError(f"{self.reduction} not implemented.")
def call(self, x):
"""Softmax activation function.
Arguments:
x: array-like
Input tensor.
Returns:
Tensor, output of softmax transformation.
"""
exps = M.exp(x - M.max(x, axis=-1, keepdims=True))
return exps / M.sum(exps, axis=-1, keepdims=True)
def __call__(self, y_true, y_pred):
"""Invokes the `Loss` instance.
Arguments:
y_true: array-like
Ground truth values.
y_pred: array-like
The predicted values.
Returns:
Weighted loss float tensor or scalar.
"""
losses = self.call(y_true, y_pred)
return M.sum(losses) / losses.size
def backward(self, loss):
_loss = M.dot(loss, M.transpose(self.kernel))
if not self.trainable:
return _loss
kernel_loss = M.dot(M.transpose(self.inputs), loss)
if self.kernel_regularizer:
kernel_loss += self.kernel_regularizer.gradient(kernel_loss)
bias_loss = M.sum(loss, axis=0, keepdims=True)
if self.bias_regularizer:
bias_loss += self.bias_regularizer.gradient(bias_loss)
self.kernel = self.kernel_optimizer.update(kernel_loss, self.kernel)
self.bias = self.bias_optimizer.update(bias_loss, self.bias)
# _loss = M.dot(loss, self.kernel)
return _loss
def call(self, x):
regularization = self.l2 * M.sum(M.square(x))
return regularization
def call(self, x):
regularization = 0.0
regularization += self.l1 * M.sum(M.abs(x))
regularization += self.l2 * M.sum(M.square(x))
return regularization