def grad(
self,
parameters,
activations,
x,
y_true=None,
dy_true=None,
lambd=0.0,
gamma=0.0,
):
"""
Gradient of cost function for training
:param x:
:param parameters:
:param activations:
:param y_true:
:param dy_true:
:param lambd:
:param gamma:
:return:
"""
y_pred, caches = L_model_forward(x, parameters, activations)
dy_pred, dy_caches = L_grads_forward(x, parameters, activations)
grad = L_model_backward(y_pred, y_true, dy_pred, dy_true, caches,
dy_caches, lambd, gamma)
return grad
def cost(
self,
parameters,
activations,
x,
y_true=None,
dy_true=None,
lambd=0.0,
gamma=0.0,
):
"""
Cost function for training
:param x:
:param parameters:
:param activations:
:param y_true:
:param dy_true:
:param lambd:
:param gamma:
:return:
"""
y_pred, caches = L_model_forward(x, parameters, activations)
dy_pred, dy_caches = L_grads_forward(x, parameters, activations)
w = [value for name, value in parameters.items() if "W" in name]
cost = lse(y_true, y_pred, lambd, w, dy_true, dy_pred, gamma)
return cost
def evaluate(self, X):
"""
Predict output(s) given inputs X.
:param X: inputs to neural network, np array of shape (n_x, m) where n_x = no. inputs, m = no. training examples
:return: Y: prediction, Y = np array of shape (n_y, m) where n_y = no. of outputs and m = no. of examples
"""
assert X.shape[0] == self.number_of_inputs
number_of_examples = X.shape[1]
mu_x, sigma_x = self._scale_factors['x']
mu_y, sigma_y = self._scale_factors['y']
X_norm = (X - mu_x) / sigma_x
Y_norm, _ = L_model_forward(X_norm, self.parameters, self.activations)
Y = (Y_norm * sigma_y + mu_y).reshape(self.number_of_outputs, number_of_examples)
return Y
def gradient(self, X):
"""
Predict output(s) given inputs X.
:param X: inputs to neural network, np array of shape (n_x, m) where n_x = no. inputs, m = no. training examples
:return: J: prediction, J = np array of shape (n_y, n_x, m) = Jacobian
"""
assert X.shape[0] == self.number_of_inputs
number_of_examples = X.shape[1]
mu_x, sigma_x = self._scale_factors['x']
mu_y, sigma_y = self._scale_factors['y']
X_norm = (X - mu_x) / sigma_x
Y_norm, _ = L_model_forward(X_norm, self.parameters, self.activations)
J_norm, _ = L_grads_forward(X_norm, self.parameters, self.activations)
J = (J_norm * sigma_y / sigma_x).reshape(self.number_of_outputs, self.number_of_inputs, number_of_examples)
return J