自訂損失函式

Python 金融 Machine Learning

Nathan George

Data Science Professor

方向不一致

含方向懲罰的 MSE

若預測與目標方向一致：

$ \sum(y - \hat{y})^2$

否則：

$\sum (y - \hat{y})^2 * \text{penalty} $

實作自訂損失函式

import tensorflow as tf

建立函式

import tensorflow as tf

# create loss function
def mean_squared_error(y_true, y_pred):

均方誤差損失

import tensorflow as tf

# create loss function
def mean_squared_error(y_true, y_pred):

    loss = tf.square(y_true - y_pred)
    return tf.reduce_mean(loss, axis=-1)

將自訂損失加到 keras

import tensorflow as tf

# create loss function
def mean_squared_error(y_true, y_pred):
    loss = tf.square(y_true - y_pred)
    return tf.reduce_mean(loss, axis=-1)

# enable use of loss with keras
import keras.losses
keras.losses.mean_squared_error = mean_squared_error

# fit the model with our mse loss function
model.compile(optimizer='adam', loss=mean_squared_error)
history = model.fit(scaled_train_features, train_targets, epochs=50)

檢查方向是否正確

tf.less(y_true * y_pred, 0)

正確方向：

neg * neg = pos
pos * pos = pos

錯誤方向：

neg * pos = neg
pos * neg = neg

使用 tf.where()

# create loss function
def sign_penalty(y_true, y_pred):
    penalty = 10.
    loss = tf.where(tf.less(y_true * y_pred, 0), 
                    penalty * tf.square(y_true - y_pred), 
                    tf.square(y_true - y_pred))

整合應用

# create loss function
def sign_penalty(y_true, y_pred):
    penalty = 100.
    loss = tf.where(tf.less(y_true * y_pred, 0),
                    penalty * tf.square(y_true - y_pred), 
                    tf.square(y_true - y_pred))

    return tf.reduce_mean(loss, axis=-1)

# enable use of loss with keras
keras.losses.sign_penalty = sign_penalty

使用自訂損失

# create the model
model = Sequential()
model.add(Dense(50,
                input_dim=scaled_train_features.shape[1],
                activation='relu'))
model.add(Dense(10, activation='relu'))
model.add(Dense(1, activation='linear'))

# fit the model with our custom 'sign_penalty' loss function
model.compile(optimizer='adam', loss=sign_penalty)
history = model.fit(scaled_train_features, train_targets, epochs=50)

領結形狀

train_preds = model.predict(scaled_train_features)
# scatter the predictions vs actual
plt.scatter(train_preds, train_targets)
plt.xlabel('predictions')
plt.ylabel('actual')
plt.show()

領結圖

自己動手做損失函式！

Python 金融 Machine Learning