def predict(self, image_matrix):
X = image_matrix.reshape((1, image_matrix.size))
mu = np.loadtxt('mu.csv')
sigma = np.loadtxt('sigma.csv')
X, mu, sigma = normalize(X, mu, sigma)
hidden_size = self.config['hidden_size']
input_size = self.config['input_size']
num_labels = self.config['num_labels']
theta1 = self.nn_params[:((hidden_size) * (input_size + 1))].reshape(
(hidden_size, input_size + 1))
theta2 = self.nn_params[((hidden_size) * (input_size + 1)):].reshape(
(num_labels, hidden_size + 1))
a1 = insert_bias(X)
z2 = theta1.dot(a1.T)
a2 = sigmoid(z2)
a2 = insert_bias(a2.T)
print theta2.shape
z3 = theta2.dot(a2.T)
h = sigmoid(z3)
print h
return h
def nn_cfx(self, X, y, nn_params):
input_size = self.config['input_size']
num_labels = self.config['num_labels']
hidden_size = self.config['hidden_size']
lambda_ = self.config['lambda']
theta1 = nn_params[:((hidden_size) * (input_size + 1))].reshape(
(hidden_size, input_size + 1))
theta2 = nn_params[((hidden_size) * (input_size + 1)):].reshape(
(num_labels, hidden_size + 1))
m = X.shape[0]
J = 0
theta1_grad = np.zeros(theta1.shape)
theta2_grad = np.zeros(theta2.shape)
a1 = insert_bias(X)
z2 = theta1.dot(a1.T)
a2 = sigmoid(z2)
a2 = insert_bias(a2.T)
z3 = theta2.dot(a2.T)
h = sigmoid(z3)
yk = np.zeros((num_labels, m))
#back propagation
for i in range(m):
yk[int(y[i])-1, i] = 1.0
error = (-yk) * np.log(h) - (1 - yk) * np.log(1 - h)
J = (1.0/m)*sum(sum(error))
t1 = np.array(theta1[:,1:])
t2 = np.array(theta2[:,1:])
sum1 = sum(sum(np.power(t1,2)))
sum2 = sum(sum(np.power(t2,2)))
r = (lambda_/(2.0*m))*(sum1 + sum2)
J += r
for t in range(m):
z2 = np.matrix(theta1.dot(a1[t,:].T)).T #change to t later
a2 = sigmoid(z2)
a2 = insert_bias_row(a2)
z3 = theta2.dot(a2)
h = sigmoid(z3)
z2 = insert_bias_row(z2)
output = np.matrix(yk[:,t]).T #change to t later
d3 = np.matrix(h - output)
sg = np.matrix(sigmoid_gradient(z2))
d2 = np.multiply(theta2.T.dot(d3),sg)
d2 = d2[1:,:]
theta2_grad += d3.dot(a2.T)
theta1_grad += d2.dot(np.matrix(a1[t,:])) #change to t later
# regularization
theta1_grad[:,0] = np.matrix(theta1_grad[:,0]/(m*1.0))
theta1_grad[:,1:] = (theta1_grad[:,1:]*(1/(m*1.0)) + ((lambda_/(m*1.0)*theta1[:,1:])))
theta2_grad[:,0] = np.matrix(theta2_grad[:,0]/(m*1.0))
theta2_grad[:,1:] = (theta2_grad[:,1:]*(1/(m*1.0)) + ((lambda_/(m*1.0)*theta2[:,1:])))
#print accuracy(predict1(theta1_grad, theta2_grad, X), y)
return J, wrap(theta1_grad, theta2_grad)
