def thetas(data, size, lambda_c, labels):
l_in = line_len(data)
theta1 = initialize_weights(l_in, size)
theta2 = initialize_weights(size, 2)
cost = lambda p: cost_function(get_x(data), get_y(data), lambda_c, p, l_in, size, labels)
init_p = np.concatenate((make1d(theta1), make1d(theta2)), axis=1)
params = fmin(cost, init_p.reshape((len(init_p), 1)))
return reshape(params, l_in, size, labels)
def thetas(data, size, lambda_c, labels):
l_in = line_len(data)
theta1 = initialize_weights(l_in, size)
theta2 = initialize_weights(size, 2)
cost = lambda p: cost_function(get_x(data), get_y(data), lambda_c, p, l_in, size, labels)
init_p = np.concatenate((make1d(theta1), make1d(theta2)), axis=1)
params = fmin(cost, init_p.reshape((len(init_p), 1)))
return reshape(params, l_in, size, labels)
def scale(data):
m = line_len(data)
x_min, x_max = numpy.array([1e10] * m), zeros(m)
for (x, _) in data:
for i in range(m):
x_min[i] = min(x_min[i], x[i])
x_max[i] = max(x_max[i], x[i])
for (x, _) in data:
for i in range(m):
x[i] = (x[i] - x_min[i]) / (x_max[i] - x_min[i]) if x_max[i] != x_min[i] else 1
return data
def optimize_size_lambda(data, labels, size_max=20, base=2.0, degree_max=10, coefficient=100):
result, error = (line_len(data), 0), 2.0
for size in range(1, size_max):
for degree in range(1, degree_max):
lambda_c = (base ** degree) / coefficient
average_error = cross_validate(data, size, lambda_c, labels)
print("%d %f %f\n" % (size, lambda_c, average_error))
if average_error < error:
error = average_error
result = (size, lambda_c)
return result
def optimize_size_lambda(data, labels, size_max=20, base=2.0, degree_max=10, coefficient=100):
result, error = (line_len(data), 0), 2.0
for size in range(1, size_max):
for degree in range(1, degree_max):
lambda_c = (base ** degree) / coefficient
average_error = cross_validate(data, size, lambda_c, labels)
print('%d %f %f\n' % (size, lambda_c, average_error))
if average_error < error:
error = average_error
result = (size, lambda_c)
return result
def scale(data):
m = line_len(data)
x_min, x_max = numpy.array([1e10] * m), zeros(m)
for (x, _) in data:
for i in range(m):
x_min[i] = min(x_min[i], x[i])
x_max[i] = max(x_max[i], x[i])
for (x, _) in data:
for i in range(m):
x[i] = (x[i] - x_min[i]) / (
x_max[i] - x_min[i]) if x_max[i] != x_min[i] else 1
return data