def test_nnCostFunction(self):
cost = nnCostFunction(unrolltheta(self.theta1, self.theta2),
self.s_1, self.s_2, self.K,
self.X, self.y, lamda=0)
self.assertAlmostEqual(cost, 0.287629, places=6)
cost_reg = nnCostFunction(unrolltheta(self.theta1, self.theta2),
self.s_1, self.s_2, self.K,
self.X, self.y, lamda=1.0)
self.assertAlmostEqual(cost_reg, 0.383770, places=6)
def computeNumericalGradient( theta, input_layer_size, hidden_layer_size, num_labels, X, y, lamda ):
numgrad = np.zeros( np.shape(theta) )
perturb = np.zeros( np.shape(theta) ) #38 x 1
e = 1e-4
num_elements = np.shape(theta)[0]
for p in range(0, num_elements) :
perturb[p] = e
loss1 = nnCostFunction( theta - perturb, input_layer_size, hidden_layer_size, num_labels, X, y, lamda)
loss2 = nnCostFunction( theta + perturb, input_layer_size, hidden_layer_size, num_labels, X, y, lamda)
numgrad[p] = (loss2 - loss1) / (2 * e)
perturb[p] = 0
return numgrad
def computeNumericalGradient(theta, input_layer_size, hidden_layer_size,
num_labels, X, y, lamda):
numgrad = np.zeros(np.shape(theta))
perturb = np.zeros(np.shape(theta)) #38 x 1
e = 1e-4
num_elements = np.shape(theta)[0]
for p in range(0, num_elements):
perturb[p] = e
loss1 = nnCostFunction(theta - perturb, input_layer_size,
hidden_layer_size, num_labels, X, y, lamda)
loss2 = nnCostFunction(theta + perturb, input_layer_size,
hidden_layer_size, num_labels, X, y, lamda)
numgrad[p] = (loss2 - loss1) / (2 * e)
perturb[p] = 0
return numgrad