def e_val_versus_gamma():
x, y = load_samples("./features.train")
gamma_list = [1e-1, 1e0, 1e1, 1e2, 1e3]
gamma_winner = np.zeros(5)
for index in range(100):
print("test", index)
e_val_list = np.zeros(5)
val_index = np.random.choice(np.arange(len(y)), 1000, replace=False)
val_x = x[val_index]
val_y = y[val_index]
train_x = np.delete(x, val_index, axis=0)
train_y = np.delete(y, val_index)
val_y_ova_0 = ova(val_y, 0)
train_y_ova_0 = ova(train_y, 0)
for i, g in enumerate(gamma_list):
arg_option = "-q -t 2 -c 0.1 -g " + str(g)
model = svm_train(train_y_ova_0, train_x, arg_option)
_, p_acc, _ = svm_predict(val_y_ova_0, val_x, model)
e_val_list[i] = (100 - p_acc[0]) / 100
gamma_winner[np.argmin(e_val_list)] += 1
print(gamma_winner)
plt.bar(np.log10(gamma_list), gamma_winner)
plt.show()
def e_out_versus_gamma():
train_x, train_y = load_samples("./features.train")
test_x, test_y = load_samples("./features.test")
train_y_ova_0 = ova(train_y, 0)
test_y_ova_0 = ova(test_y, 0)
gamma_list = [1e0, 1e1, 1e2, 1e3, 1e4]
e_out_list = []
for g in gamma_list:
arg_option = "-t 2 -c 0.1 -g " + str(g)
model = svm_train(train_y_ova_0, train_x, arg_option)
_, p_acc, _ = svm_predict(test_y_ova_0, test_x, model)
e_out_list.append((100 - p_acc[0]) / 100)
print(e_out_list)
plt.scatter(np.log10(gamma_list), e_out_list)
plt.show()
def sv_versus_c():
x, y = load_samples("./features.train")
y_ova_8 = ova(y, 8)
c_list = [1e-5, 1e-3, 1e-1, 1e1, 1e3]
sv_list = []
for c in c_list:
arg_option = "-t 1 -d 2 -r 1 -g 1 -c " + str(c)
model = svm_train(y_ova_8, x, arg_option)
nr_sv = model.get_nr_sv()
sv_list.append(nr_sv)
print(sv_list)
plt.scatter(np.log10(c_list), sv_list)
plt.show()
def e_in_versus_c():
x, y = load_samples("./features.train")
y_ova_8 = ova(y, 8)
c_list = [1e-5, 1e-3, 1e-1, 1e1, 1e3]
e_in_list = []
for c in c_list:
arg_option = "-t 1 -d 2 -r 1 -g 1 -c " + str(c)
model = svm_train(y_ova_8, x, arg_option)
_, p_acc, _ = svm_predict(y_ova_8, x, model)
e_in_list.append((100 - p_acc[0]) / 100)
print(e_in_list)
plt.scatter(np.log10(c_list), e_in_list)
plt.show()
def sv_versus_gamma():
x, y = load_samples("./features.train")
y_ova_8 = ova(y, 8)
g_list = list(range(-10, 10))
sv_list = []
for g in g_list:
gamma = pow(10, g)
arg_option = "-t 2 -c 1 -g " + str(gamma)
model = svm_train(y_ova_8, x, arg_option)
nr_sv = model.get_nr_sv()
sv_list.append(nr_sv)
print(sv_list)
plt.scatter(g_list, sv_list)
plt.show()
def weight_versus_c():
x, y = load_samples("./features.train")
y_ova_0 = ova(y, 0)
c_list = [1e-5, 1e-3, 1e-1, 1e1, 1e3]
w_norm_list = []
for c in c_list:
arg_option = "-t 0 -c " + str(c)
model = svm_train(y_ova_0, x, arg_option)
sv_coef = np.squeeze(np.array(model.get_sv_coef()))
sv = pd.DataFrame(model.get_SV()).values[:, 1:]
w = np.dot(sv_coef, sv)
w_norm_list.append(np.linalg.norm(w))
print(w_norm_list)
plt.scatter(np.log10(c_list), w_norm_list)
plt.show()
def margin_versus_c():
x, y = load_samples("./features.train")
y_ova_0 = ova(y, 0)
c_list = [1e-3, 1e-2, 1e-1, 1e0, 1e1]
margin_list = []
for c in c_list:
arg_option = "-q -t 2 -g 80 -c " + str(c)
model = svm_train(y_ova_0, x, arg_option)
sv_coef = np.squeeze(np.array(model.get_sv_coef()))
sv = pd.DataFrame(model.get_SV()).values[:, 1:]
sv_indices = np.add(model.get_sv_indices(), -1)
coef_y = sv_coef * y[sv_indices]
w_square = np.sum(rbf_kernel(sv, 80) * np.dot(coef_y, coef_y.T))
w_norm = sqrt(w_square)
margin_list.append(1 / w_norm)
print("margin in Z space:")
print(margin_list)
plt.scatter(np.log10(c_list), margin_list)
plt.show()
def margin_versus_gamma():
x, y = load_samples("./features.train")
y_ova_0 = ova(y, 0)
g_list = list(range(-10, 10))
margin_list = []
for g in g_list:
gamma = pow(10, g)
arg_option = "-t 2 -c 1 -g " + str(gamma)
model = svm_train(y_ova_0, x, arg_option)
sv_coef = np.squeeze(np.array(model.get_sv_coef()))
sv = pd.DataFrame(model.get_SV()).values[:, 1:]
sv_indices = np.add(model.get_sv_indices(), -1)
coef_y = sv_coef * y[sv_indices]
w_square = np.sum(rbf_kernel(sv, gamma) * np.dot(coef_y, coef_y.T))
w_norm = sqrt(w_square)
margin_list.append(1 / w_norm)
print(margin_list)
plt.ylim(0, 1e-3)
plt.scatter(g_list, margin_list)
plt.show()