print('========== Part 2.4: Gradient Checking ==========')
sample_dims = [3, 5, 3]
sample_m = 5
X_s = nn.initialize_weights(sample_dims[0] - 1, sample_m, None, debug=True)
y_s = 1 + (np.arange(1, sample_m + 1) % sample_dims[-1])
Theta = [
nn.initialize_weights(sample_dims[i],
sample_dims[i + 1],
None,
debug=True) for i in range(len(sample_dims) - 1)
]
theta = np.block([t.reshape(t.size, order='F') for t in Theta])
sample_nn = NeuralNetwork(X_s, y_s, Theta, sample_dims)
grads = sample_nn.cost_grad(theta)
n_grads = sample_nn.cost_grad_numerical(theta)
print('----------------------')
print('Analytical | Numerical')
print('----------------------')
for g, n_g in zip(grads, n_grads):
print(' {0: .4f} | {1: .4f}'.format(g, n_g))
diff = np.linalg.norm(n_grads - grads) / np.linalg.norm(n_grads + grads)
print('Relative difference: {0}'.format(diff))
print('========== Part 2.4: Regularized Neural Networks ==========')
Theta = [data['Theta1'], data['Theta2']]
theta = np.block([t.reshape(t.size, order='F') for t in Theta])
nn.update_lambda(3)
J = nn.cost(theta)
print('Regularized cost: {0:0.6f} (expected: 0.576051)'.format(J))
