def _get_gradient_hess(self, y, y_pred):
"""
获取一阶、二阶导数信息
:param y:真实值
:param y_pred:预测值
:return:
"""
if self.loss == 'squarederror':
return y_pred - y, np.ones_like(y)
elif self.loss == 'logistic':
return utils.sigmoid(y_pred) - utils.sigmoid(
y), utils.sigmoid(y_pred) * (1 - utils.sigmoid(y_pred))
elif self.loss == 'poisson':
return np.exp(y_pred) - y, np.exp(y_pred)
elif self.loss == 'gamma':
return 1.0 - y * np.exp(-1.0 * y_pred), y * np.exp(-1.0 * y_pred)
elif self.loss == 'tweedie':
if self.p == 1:
return np.exp(y_pred) - y, np.exp(y_pred)
elif self.p == 2:
return 1.0 - y * np.exp(-1.0 * y_pred), y * np.exp(
-1.0 * y_pred)
else:
return np.exp(y_pred * (2.0 - self.p)) - y * np.exp(
y_pred * (1.0 - self.p)), (2.0 - self.p) * np.exp(
y_pred * (2.0 - self.p)) - (1.0 - self.p) * y * np.exp(
y_pred * (1.0 - self.p))
def predict(self, X):
"""
:param X:
:return:
"""
# 归一化
if self.normal:
X = (X - self.xmin) / self.xmax
# reshape
X = X[:, self.replace_ind]
# 去掉第一列bias以及非组合特征
X_ = X[:, self.positive_ind:]
n_sample, n_feature = X_.shape
pol = np.zeros(n_sample)
for i in range(0, n_feature - 1):
for j in range(i + 1, n_feature):
pol += X_[:, i] * X_[:, j] * np.dot(self.V[i, self.fields[self.positive_ind + j]],
self.V[j, self.fields[self.positive_ind + i]])
linear_rst = np.c_[np.ones(n_sample), X] @ self.w.reshape(-1) + pol
if self.objective == "squarederror":
return linear_rst
elif self.objective in ["poisson", "gamma", "tweedie"]:
return np.exp(linear_rst)
else:
return utils.sigmoid(linear_rst) > 0.5
def predict_proba(self, X):
"""
:param X:
:return:
"""
if self.normal:
X = (X - self.xmin) / self.xmax
n_sample, n_feature = X.shape
X_V = X @ self.V
X_V_2 = X_V * X_V
X_2_V_2 = (X * X) @ (self.V * self.V)
pol = 0.5 * np.sum(X_V_2 - X_2_V_2, axis=1)
linear_rst = np.c_[np.ones(n_sample), X] @ self.w.reshape(-1) + pol
pos_proba = utils.sigmoid(linear_rst)
return np.c_[1.0 - pos_proba, pos_proba]
def predict(self, X):
"""
:param X:
:return:
"""
if self.normal:
X = (X - self.xmin) / self.xmax
n_sample, n_feature = X.shape
X_V = X @ self.V
X_V_2 = X_V * X_V
X_2_V_2 = (X * X) @ (self.V * self.V)
pol = 0.5 * np.sum(X_V_2 - X_2_V_2, axis=1)
linear_rst = np.c_[np.ones(n_sample), X] @ self.w.reshape(-1) + pol
if self.objective == "squarederror":
return linear_rst
elif self.objective in ["poisson", "gamma", "tweedie"]:
return np.exp(linear_rst)
else:
return utils.sigmoid(linear_rst) > 0.5
def predict_proba(self, X):
"""
logistic regression用
:param X:
:return:
"""
# 归一化
if self.normal:
X = (X - self.xmin) / self.xmax
# reshape
X = X[:, self.replace_ind]
# 去掉第一列bias以及非组合特征
X_ = X[:, self.positive_ind:]
n_sample, n_feature = X_.shape
pol = np.zeros(n_sample)
for i in range(0, n_feature - 1):
for j in range(i + 1, n_feature):
pol += X_[:, i] * X_[:, j] * np.dot(self.V[i, self.fields[self.positive_ind + j]],
self.V[j, self.fields[self.positive_ind + i]])
pos_proba = utils.sigmoid(np.c_[np.ones(n_sample), X] @ self.w.reshape(-1) + pol)
return np.c_[1.0 - pos_proba, pos_proba]
def fit(self, X, y, eval_set=None, show_log=True):
X_o = X.copy()
if self.normal:
self.xmin = X.min(axis=0)
self.xmax = X.max(axis=0) + 1e-8
X = (X - self.xmin) / self.xmax
n_sample, n_feature = X.shape
x_y = np.c_[np.ones(n_sample), X, y]
# 记录loss
train_losses = []
eval_losses = []
# 调整一下学习率
if self.adjust_lr:
self.lr = max(self.lr, 1 / n_feature)
# 初始化参数
self.w = np.random.random((n_feature + 1, 1)) * 1e-3
self.V = np.random.random((n_feature, self.hidden_dim)) * 1e-3
if self.solver == 'adam':
# 缓存梯度一阶，二阶估计
w_1 = np.zeros_like(self.w)
V_1 = np.zeros_like(self.V)
w_2 = np.zeros_like(self.w)
V_2 = np.zeros_like(self.V)
# 更新参数
count = 0
for epoch in range(self.epochs):
# 验证集记录
best_eval_value = np.power(2., 1023)
eval_count = 0
np.random.shuffle(x_y)
for index in range(x_y.shape[0] // self.batch_size):
count += 1
batch_x_y = x_y[self.batch_size * index:self.batch_size *
(index + 1)]
batch_x = batch_x_y[:, :-1]
batch_y = batch_x_y[:, -1:]
# 计算链式求导第一层梯度
if self.objective == "squarederror":
y_x_t = self._y(batch_x).reshape((-1, 1)) - batch_y
elif self.objective == "poisson":
y_x_t = np.exp(self._y(batch_x).reshape((-1, 1))) - batch_y
elif self.objective == "gamma":
y_x_t = 1.0 - batch_y * np.exp(
-1.0 * self._y(batch_x).reshape((-1, 1)))
elif self.objective == 'tweedie':
if self.tweedie_p == 1:
y_x_t = np.exp(self._y(batch_x).reshape(
(-1, 1))) - batch_y
elif self.tweedie_p == 2:
y_x_t = 1.0 - batch_y * np.exp(
-1.0 * self._y(batch_x).reshape((-1, 1)))
else:
y_x_t = np.exp(self._y(batch_x).reshape((-1, 1)) * (2.0 - self.tweedie_p)) \
- batch_y * np.exp(self._y(batch_x).reshape((-1, 1)) * (1.0 - self.tweedie_p))
else:
# 二分类
y_x_t = utils.sigmoid(self._y(batch_x).reshape(
(-1, 1))) - batch_y
# 更新w
w_reg = self.lamb * self.w + self.alpha * np.where(
self.w > 0, 1, 0)
w_reg[0, 0] = 0.0
w_grad = (np.sum(y_x_t * batch_x, axis=0) /
self.batch_size).reshape((-1, 1)) + w_reg
if self.solver == 'sgd':
self.w = self.w - self.lr * w_grad
elif self.solver == 'adam':
w_1 = self.rho_1 * w_1 + (1 - self.rho_1) * w_grad
w_2 = self.rho_2 * w_2 + (1 - self.rho_2) * w_grad * w_grad
w_1_ = w_1 / (1 - np.power(self.rho_1, count))
w_2_ = w_2 / (1 - np.power(self.rho_2, count))
self.w = self.w - (self.lr * w_1_) / (np.sqrt(w_2_) + 1e-8)
# 更新 V
batch_x_ = batch_x[:, 1:]
V_X = batch_x_ @ self.V
X_2 = batch_x_ * batch_x_
# 从i,f单个元素逐步更新有点慢
# for i in range(self.V.shape[0]):
# for f in range(self.V.shape[1]):
# if self.solver == "sgd":
# self.V[i, f] -= self.lr * (
# np.sum(y_x_t.reshape(-1) * (batch_x_[:, i] * V_X[:, f] - self.V[i, f] * X_2[:, i]))
# / self.batch_size + self.lamb * self.V[i, f] + self.alpha * (self.V[i, f] > 0))
# elif self.solver == "adam":
# v_reg = self.lamb * self.V[i, f] + self.alpha * (self.V[i, f] > 0)
# v_grad = np.sum(y_x_t.reshape(-1) * (
# batch_x_[:, i] * V_X[:, f] - self.V[i, f] * X_2[:, i])) / self.batch_size + v_reg
# V_1[i, f] = self.rho_1 * V_1[i, f] + (1 - self.rho_1) * v_grad
# V_2[i, f] = self.rho_2 * V_2[i, f] + (1 - self.rho_2) * v_grad * v_grad
# v_1_ = V_1[i, f] / (1 - np.power(self.rho_1, count))
# v_2_ = V_2[i, f] / (1 - np.power(self.rho_2, count))
# self.V[i, f] = self.V[i, f] - (self.lr * v_1_) / (np.sqrt(v_2_) + 1e-8)
# 从隐变量的维度进行更新
for f in range(self.V.shape[1]):
V_reg = self.lamb * self.V[:, f] + self.alpha * (
self.V[:, f] > 0)
V_grad = np.sum(y_x_t * (batch_x_ * V_X[:, f].reshape(
(-1, 1)) - X_2 * self.V[:, f]),
axis=0) + V_reg
if self.solver == 'sgd':
self.V[:, f] = self.V[:, f] - self.lr * V_grad
elif self.solver == 'adam':
V_1[:,
f] = self.rho_1 * V_1[:, f] + (1 -
self.rho_1) * V_grad
V_2[:, f] = self.rho_2 * V_2[:, f] + (
1 - self.rho_2) * V_grad * V_grad
V_1_ = V_1[:, f] / (1 - np.power(self.rho_1, count))
V_2_ = V_2[:, f] / (1 - np.power(self.rho_2, count))
self.V[:, f] = self.V[:, f] - (self.lr * V_1_) / (
np.sqrt(V_2_) + 1e-8)
# 计算eval loss
eval_loss = None
if eval_set is not None:
eval_x, eval_y = eval_set
eval_loss = np.std(eval_y - self.predict(eval_x))
eval_losses.append(eval_loss)
# 是否显示
if show_log:
train_loss = np.std(y - self.predict(X_o))
print("epoch:", epoch + 1, "/", self.epochs, ",samples:",
(index + 1) * self.batch_size, "/", n_sample,
",train loss:", train_loss, ",eval loss:", eval_loss)
train_losses.append(train_loss)
# 是否早停
if eval_loss is not None and self.early_stopping_rounds is not None:
if eval_loss < best_eval_value:
eval_count = 0
best_eval_value = eval_loss
else:
eval_count += 1
if eval_count >= self.early_stopping_rounds:
print(
"---------------early_stopping-----------------------------"
)
break
return train_losses, eval_losses
def fit(self, X, y, eval_set=None, show_log=False, fields=None):
"""
:param X:
:param y:
:param eval_set:
:param show_log:
:param fields: 为None时，退化为FM
:return:
"""
X_o = X.copy()
# 归一化
if self.normal:
self.xmin = X.min(axis=0)
self.xmax = X.max(axis=0) + 1e-7
X = (X - self.xmin) / self.xmax
n_sample, n_feature = X.shape
# 处理fields
if fields is None:
self.replace_ind = list(range(0, n_feature))
self.positive_ind = 0
self.fields = [0] * n_feature
self.field_num = 1
else:
self.replace_ind = np.argsort(fields).tolist()
self.positive_ind = np.sum([1 if item < 0 else 0 for item in fields])
self.fields = sorted(fields)
self.field_num = len(set(self.fields[self.positive_ind:]))
# reshape X
X = X[:, self.replace_ind]
x_y = np.c_[np.ones(n_sample), X, y]
# 记录loss
train_losses = []
eval_losses = []
# 调整一下学习率
if self.adjust_lr:
self.lr = max(self.lr, 1 / n_feature)
# 初始化参数
self.w = np.random.random((n_feature + 1, 1)) * 1e-3
self.V = np.random.random((n_feature - self.positive_ind, self.field_num, self.hidden_dim)) * 1e-3
if self.solver == 'adam':
# 缓存梯度一阶，二阶估计
w_1 = np.zeros_like(self.w)
V_1 = np.zeros_like(self.V)
w_2 = np.zeros_like(self.w)
V_2 = np.zeros_like(self.V)
# 更新参数
count = 0
for epoch in range(self.epochs):
# 验证集记录
best_eval_value = np.power(2., 1023)
eval_count = 0
np.random.shuffle(x_y)
for index in range(x_y.shape[0] // self.batch_size):
count += 1
batch_x_y = x_y[self.batch_size * index:self.batch_size * (index + 1)]
batch_x = batch_x_y[:, :-1]
batch_y = batch_x_y[:, -1:]
# 计算链式求导第一层梯度
if self.objective == "squarederror":
y_x_t = self._y(batch_x).reshape((-1, 1)) - batch_y
elif self.objective == "poisson":
y_x_t = np.exp(self._y(batch_x).reshape((-1, 1))) - batch_y
elif self.objective == "gamma":
y_x_t = 1.0 - batch_y * np.exp(-1.0 * self._y(batch_x).reshape((-1, 1)))
elif self.objective == 'tweedie':
if self.tweedie_p == 1:
y_x_t = np.exp(self._y(batch_x).reshape((-1, 1))) - batch_y
elif self.tweedie_p == 2:
y_x_t = 1.0 - batch_y * np.exp(-1.0 * self._y(batch_x).reshape((-1, 1)))
else:
y_x_t = np.exp(self._y(batch_x).reshape((-1, 1)) * (2.0 - self.tweedie_p)) \
- batch_y * np.exp(self._y(batch_x).reshape((-1, 1)) * (1.0 - self.tweedie_p))
else:
# 二分类
y_x_t = utils.sigmoid(self._y(batch_x).reshape((-1, 1))) - batch_y
# 更新w
if self.solver == 'sgd':
self.w = self.w - (self.lr * (np.sum(y_x_t * batch_x, axis=0) / self.batch_size).reshape(
(-1, 1)) + self.lamb * self.w + self.alpha * np.where(self.w > 0, 1, 0))
elif self.solver == 'adam':
w_reg = self.lamb * self.w + self.alpha * np.where(self.w > 0, 1, 0)
w_grad = (np.sum(y_x_t * batch_x, axis=0) / self.batch_size).reshape(
(-1, 1)) + w_reg
w_1 = self.rho_1 * w_1 + (1 - self.rho_1) * w_grad
w_2 = self.rho_2 * w_2 + (1 - self.rho_2) * w_grad * w_grad
w_1_ = w_1 / (1 - np.power(self.rho_1, count))
w_2_ = w_2 / (1 - np.power(self.rho_2, count))
self.w = self.w - (self.lr * w_1_) / (np.sqrt(w_2_) + 1e-8)
# 更新 V
batch_x_ = batch_x[:, 1 + self.positive_ind:]
# 逐元素更新
for i in range(0, batch_x_.shape[1] - 1):
for j in range(i + 1, batch_x_.shape[1]):
for k in range(0, self.hidden_dim):
v_reg_l = self.lamb * self.V[i, self.fields[self.positive_ind + j], k] + \
self.alpha * (self.V[i, self.fields[self.positive_ind + j], k] > 0)
v_grad_l = np.sum(y_x_t.reshape(-1) * batch_x_[:, i] * batch_x_[:, j] *
self.V[
j, self.fields[self.positive_ind + i], k]) / self.batch_size + v_reg_l
v_reg_r = self.lamb * self.V[j, self.fields[self.positive_ind + i], k] + \
self.alpha * (self.V[j, self.fields[self.positive_ind + i], k] > 0)
v_grad_r = np.sum(y_x_t.reshape(-1) * batch_x_[:, i] * batch_x_[:, j] *
self.V[
i, self.fields[self.positive_ind + j], k]) / self.batch_size + v_reg_r
if self.solver == "sgd":
self.V[i, self.fields[self.positive_ind + j], k] -= self.lr * v_grad_l
self.V[j, self.fields[self.positive_ind + i], k] -= self.lr * v_grad_r
elif self.solver == "adam":
V_1[i, self.fields[self.positive_ind + j], k] = self.rho_1 * V_1[
i, self.fields[self.positive_ind + j], k] + (1 - self.rho_1) * v_grad_l
V_2[i, self.fields[self.positive_ind + j], k] = self.rho_2 * V_2[
i, self.fields[self.positive_ind + j], k] + (1 - self.rho_2) * v_grad_l * v_grad_l
v_1_l = V_1[i, self.fields[self.positive_ind + j], k] / (
1 - np.power(self.rho_1, count))
v_2_l = V_2[i, self.fields[self.positive_ind + j], k] / (
1 - np.power(self.rho_2, count))
V_1[j, self.fields[self.positive_ind + i], k] = self.rho_1 * V_1[
j, self.fields[self.positive_ind + i], k] + (1 - self.rho_1) * v_grad_r
V_2[j, self.fields[self.positive_ind + i], k] = self.rho_2 * V_2[
j, self.fields[self.positive_ind + i], k] + (1 - self.rho_2) * v_grad_r * v_grad_r
v_1_r = V_1[j, self.fields[self.positive_ind + i], k] / (
1 - np.power(self.rho_1, count))
v_2_r = V_2[j, self.fields[self.positive_ind + i], k] / (
1 - np.power(self.rho_2, count))
self.V[i, self.fields[self.positive_ind + j], k] -= (self.lr * v_1_l) / (
np.sqrt(v_2_l) + 1e-8)
self.V[j, self.fields[self.positive_ind + i], k] -= (self.lr * v_1_r) / (
np.sqrt(v_2_r) + 1e-8)
# 计算eval loss
eval_loss = None
if eval_set is not None:
eval_x, eval_y = eval_set
if self.objective == 'logistic':
eval_loss = np.mean(eval_y != self.predict(eval_x))
else:
eval_loss = np.std(eval_y - self.predict(eval_x))
eval_losses.append(eval_loss)
# 是否显示
if show_log:
if self.objective == 'logistic':
train_loss = np.mean(y != self.predict(X_o))
else:
train_loss = np.std(y - self.predict(X_o))
print("epoch:", epoch + 1, "/", self.epochs, ",samples:", (index + 1) * self.batch_size, "/",
n_sample,
",train loss:",
train_loss, ",eval loss:", eval_loss)
train_losses.append(train_loss)
# 是否早停
if eval_loss is not None and self.early_stopping_rounds is not None:
if eval_loss < best_eval_value:
eval_count = 0
best_eval_value = eval_loss
else:
eval_count += 1
if eval_count >= self.early_stopping_rounds:
print("---------------early_stopping-----------------------------")
break
return train_losses, eval_losses