Tensorflow学习笔记：正则化线性回归

最新推荐文章于 2024-04-12 12:34:01 发布

转载最新推荐文章于 2024-04-12 12:34:01 发布 · 117 阅读

0 ·

CC 4.0 BY-SA版权

原文链接：https://my.oschina.net/wujux/blog/1801242

文章标签：

#人工智能 #数据结构与算法 #python

本文通过一个具体的多项式回归实例介绍了如何使用正则化技术来防止模型过拟合。文中详细展示了利用Python及其相关库如NumPy、TensorFlow等进行数据集划分、模型构建与训练的过程，并对不同正则化参数下的模型性能进行了对比。

为什么80%的码农都做不了架构师？>>>

#coding:utf-8
'''
正则化
'''

import numpy as np
import tensorflow as tf
import matplotlib.pyplot as plt

# 按照比例分割测试和训练数据值
def split_dataset(x_dataset,y_dataset,ratio):
    x_arr = np.arange(x_dataset.size)
    y_arr = np.arange(y_dataset.size)
    np.random.shuffle(x_arr)
    np.random.shuffle(y_arr)

    x_num_train = int(ratio * x_dataset.size)
    y_num_train = int(ratio * y_dataset.size)

    x_train = x_dataset[x_arr[0:x_num_train]]
    y_train = y_dataset[y_arr[0:y_num_train]]
    x_test = x_dataset[x_arr[x_num_train:x_dataset.size]]
    y_test = y_dataset[y_arr[y_num_train:y_dataset.size]]

    return x_train,x_test,y_train,y_test

learning_rate = 0.001
training_epochs = 1000
reg_lambda = 0.0

x_dataset = np.linspace(-1,1,101)

# 多项式系数个数
num_coeffs = 9
y_dataset_params = [0.0] * num_coeffs
y_dataset_params[2] = 1
y_dataset = 0
for i in range(num_coeffs):
    y_dataset += y_dataset_params[i] * np.power(x_dataset,i)

y_dataset += np.random.randn(*x_dataset.shape) * 0.33

(x_train,x_test,y_train,y_test) = split_dataset(x_dataset,y_dataset,0.7)

X = tf.placeholder(tf.float32)
Y = tf.placeholder(tf.float32)

def model(X,w):
    terms = []

    for i in range(num_coeffs):
        term = tf.multiply(w[i],tf.pow(X,i))
        terms.append(term)
    
    return tf.add_n(terms)

w = tf.Variable([0.] * num_coeffs,name='parameters',dtype=tf.float32)
y_model = model(X,w)

# 定义正则化损失函数
cost = tf.div(tf.add(tf.reduce_sum(tf.square(Y-y_model)),
        tf.multiply(reg_lambda, tf.reduce_sum(tf.square(w)))),
        2*x_train.size)

train_op = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost)

init_op = tf.global_variables_initializer()

sess = tf.Session()

sess.run(init_op)

for reg_lambda in np.linspace(0,1,100):
    for epoch in range(training_epochs):
        sess.run(train_op,feed_dict={X:x_train,Y:y_train})
    final_cost = sess.run(cost,feed_dict={X:x_test,Y:y_test})
    print('reg lambda:',reg_lambda)
    print('final cost:',final_cost)

sess.close()

转载于:https://my.oschina.net/wujux/blog/1801242