def train(self, max_depth=3):
for i in range(self.T):
pred_y = np.array([]) # k*m
# Train weak learner using distribution D_t
for j in range(self.k):
self.h_t = self.model(max_depth=max_depth, presort=True)
self.h_t.fit(self.X,
self.y[j * self.m:(j + 1) * self.m],
sample_weight=self.D_t[j * self.m:(j + 1) *
self.m])
# Get weak hypothesis h_t
self.h.append(self.h_t)
# Choose alpha_t
pred_y = np.append(pred_y, self.h_t.predict(self.X))
r_t = np.dot(self.D_t, np.multiply(self.y, pred_y))
if abs(r_t - 1) < 0.00000001:
self.alpha_t = 0.5 * log(
(1 + r_t + 0.000001) / (1 - r_t + 0.000001))
else:
self.alpha_t = 0.5 * log((1 + r_t) / (1 - r_t))
self.alphas = np.append(self.alphas, self.alpha_t)
# Update
self.D_t = np.multiply(
self.D_t,
list(map(exp, -self.alpha_t * np.multiply(self.y, pred_y))))
self.D_t /= np.sum(self.D_t)
# self.D = np.append(self.D, self.D_t)
ret_index = self.predict(self.test_X, i + 1)
# for i in range(len(X_test)):
# print(ret_index[i])
# print(y_test[i])
scores = one_error(ret_index, self.test_y)
y_train = []
at_n = 3
for j in range(len(ret_index)):
tmp = [0] * self.k
for ll in ret_index[j]:
if ret_index[j].index(ll) < self.k - at_n:
continue
tmp[self.class_list.index(ll)] = 1
ret_index[j] = tmp
tmp = [0] * self.k
for ll in self.test_y[j]:
tmp[self.class_list.index(ll)] = 1
y_train.append(tmp)
precision, recall, error = evaluate(np.array(ret_index),
np.array(y_train))
print(i, precision, recall, error, scores)
def SDSS():
path = 'data/SDSS.csv'
x, y, class_list = load_data(path)
X_train, X_test, y_train, y_test = train_test_split(x,
y,
test_size=0.4,
random_state=42)
clf = discrete_Adaboost_MH(X_train, y_train, class_list, T=50)
ret_index = clf.predict(X_test)
for i in range(len(X_test)):
print(ret_index[i])
print(y_test[i])
scores = one_error(ret_index, y_test, all_=True)
print('----------SDSS one error-----------\n', scores)
def yeast():
path = 'data/yeast'
X_train, y_train, X_test, y_test, class_list = load_data(path)
clf = discrete_Adaboost_MR(X_train,
y_train,
X_test,
y_test,
class_list,
T=100)
ret_index, pred = clf.predict(X_test, clf.T)
# for i in range(len(X_test)):
# print(ret_index[i])
# print(y_test[i])
scores = one_error(ret_index, y_test)
print('----------yeast one error on test-----------\n', scores)
def yeast():
path = 'data/yeast'
X_train, y_train, X_test, y_test, class_list = load_data(path)
clf = discrete_Adaboost_MH(X_train,
y_train,
X_test,
y_test,
class_list,
T=200)
ret_index = clf.predict(X_test, clf.T)
for i in range(len(X_test)):
print(ret_index[i])
print(y_test[i])
scores = one_error(ret_index, y_test, all_=True)
print('----------yeast one error-----------\n', scores)
def train(self, max_depth=3):
for i in range(self.T):
pred_y = np.array([]) # k*m
# Train weak learner using distribution D_t
for j in range(self.k):
self.h_t = self.model(max_depth=max_depth, presort=True)
self.h_t.fit(self.X,
self.y[j * self.m:(j + 1) * self.m],
sample_weight=np.sum(self.D_t, axis=1)[:, j])
# Get weak hypothesis h_t
self.h.append(self.h_t)
# Choose alpha_t
pred_y = np.append(pred_y, self.h_t.predict(self.X))
tmp_sum = 0
for (x_i, l0, l1) in self.pairs:
tmp_sum += self.D_t[x_i][l0][l1] * (pred_y[l1 * self.m + x_i] -
pred_y[l0 * self.m + x_i])
# for i in range(self.m):
# for l0 in range(self.k):
# for l1 in range(self.k):
# if abs(self.D_t[i][l0][l1]) < 0.0000001:
# continue
# tmp_sum += self.D_t[i][l0][l1]*(pred_y[i*self.k+l1]-pred_y[i*self.k+l0])
r_t = tmp_sum * 0.5
if abs(r_t - 1) < 0.00000001:
self.alpha_t = 0.5 * log(
(1 + r_t + 0.000001) / (1 - r_t + 0.000001))
else:
self.alpha_t = 0.5 * log((1 + r_t) / (1 - r_t))
self.alphas = np.append(self.alphas, self.alpha_t)
# Update
for (x_i, l0, l1) in self.pairs:
self.D_t[x_i][l0][l1] = self.D_t[x_i][l0][l1] * exp(
0.5 * self.alpha_t *
(pred_y[l0 * self.m + x_i] - pred_y[l1 * self.m + x_i]))
self.D_t /= np.sum(self.D_t)
# self.D = np.append(self.D, self.D_t)
ret_index, pred = self.predict(self.test_X, i + 1)
# for i in range(len(X_test)):
# print(ret_index[i])
# print(y_test[i])
scores = one_error(ret_index, self.test_y)
y_train = []
at_n = 3
for j in range(len(ret_index)):
tmp = [0] * self.k
for ll in ret_index[j]:
if ret_index[j].index(ll) < self.k - at_n:
continue
tmp[self.class_list.index(ll)] = 1
ret_index[j] = tmp
tmp = [0] * self.k
for ll in self.test_y[j]:
tmp[self.class_list.index(ll)] = 1
y_train.append(tmp)
precision, recall, error = evaluate(np.array(ret_index),
np.array(y_train))
print(i, precision, recall, error, scores)