def run_winnow_margin():
alpha = [1.1, 1.01, 1.005, 1.0005, 1.0001]
gamma = [2.0, 0.3, 0.04, 0.006, 0.001]
best = [0, 0, 0, 0]
for c in range(len(alpha)):
for g in range(len(gamma)):
w_500, theta_500, _ = algorithms.winnow_m(D1_500_y, D1_500_x,
alpha[c], gamma[g], 500)
w_1000, theta_1000, _ = algorithms.winnow_m(
D1_1000_y, D1_1000_x, alpha[c], gamma[g], 1000)
correct_500 = verify(w_500, theta_500, D2_500_x, D2_500_y)
print("winnow(margin) with n = 500, alpha = " + str(alpha[c]) +
" gamma = " + str(gamma[g]))
print(correct_500)
correct_1000 = verify(w_1000, theta_1000, D2_1000_x, D2_1000_y)
print("winnow(margin) with n = 1000, alpha = " + str(alpha[c]) +
" gamma = " + str(gamma[g]))
print(correct_1000)
if (correct_500 + correct_1000) / 2.0 > (best[0] + best[1]) / 2.0:
best[0] = correct_500
best[1] = correct_1000
best[2] = alpha[c]
best[3] = gamma[g]
print("bestresult: correct_500 = " + str(best[0]) + " correct_1000 = " +
str(best[1]) + " alpha = " + str(best[2]) + " gamma = " +
str(best[3]))
def test_winnow_margin():
alpha = [1.01, 1.1, 1.1]
gamma = [2.0, 0.006, 0.001]
for n in range(3):
w1, theta1, _ = algorithms.winnow_m(np.tile(d_y[n], 20),
np.tile(d_x[n], (20, 1)), alpha[n],
gamma[n], m_val[n])
correct1 = verify(w1, theta1, dxx[n], dyy[n])
print("TEST winnow_margin with m = " + str(m_val[n]) + ", alpha = " +
str(alpha[n]) + " gamma = " + str(gamma[n]))
print(correct1)
def plot_mistake_1000():
_, __, error1 = algorithms.perceptron(D_1000_y, D_1000_x)
_, __, error2 = algorithms.perceptron_m(D_1000_y, D_1000_x, 0.005)
_, __, error3 = algorithms.winnow(D_1000_y, D_1000_x, 1.1, 1000)
_, __, error4 = algorithms.winnow_m(D_1000_y, D_1000_x, 1.1, 2.0, 1000)
_, __, error5 = algorithms.adagrad(D_1000_y, D_1000_x, 0.25)
p1, = plt.plot(error1, color="blue", label="perceptron")
p2, = plt.plot(error2, color="red", label="perceptron with margin")
p3, = plt.plot(error3, color="orange", label="winnow")
p4, = plt.plot(error4, color="green", label="winnow with margin")
p5, = plt.plot(error5, color="black", label="adagrad")
plt.legend(handles=[p1, p2, p3, p4, p5], loc=2)
plt.title("mistake bound n=1000")
plt.show()
def run_winnow_margin():
alpha = [1.1, 1.01, 1.005, 1.0005, 1.0001]
gamma = [2.0, 0.3, 0.04, 0.006, 0.001]
for n in range(3):
best = [0, 0, 0, 0]
for c in range(len(alpha)):
for g in range(len(gamma)):
w1, theta1, _ = algorithms.winnow_m(d1y[n], d1x[n], alpha[c],
gamma[g], m_val[n])
correct1 = verify(w1, theta1, d2x[n], d2y[n])
print("winnow_margin with m = " + str(m_val[n]) +
", alpha = " + str(alpha[c]) + " gamma = " +
str(gamma[g]))
print(correct1)
if correct1 > best[0]:
best[0] = correct1
best[1] = alpha[c]
best[2] = gamma[g]
print("bestresult for m = " + str(m_val[n]) + ": correct1 = " +
str(best[0]) + " alpha = " + str(best[1]) + " gamma = " +
str(best[2]))