参考博文:深度学习——入门经典案例《波士顿房价预测》深度解析_什么是梯度下降-CSDN博客
tensorflow BP神经网络 波士顿房价预测_bp神经网络波士顿房价预测-CSDN博客
网络训练部分:
import numpy as np
import matplotlib.pyplot as plt
##使用了随机梯度下降法提高效率
def load_data():
# 读取以空格分开的文件,变成一个连续的数组
firstdata = np.fromfile('F:/boston-house-price-forecast-master/boston-house-price-forecast-master/housing.data', sep=' ')
# 添加属性
feature_names = ['CRIM', 'ZN', 'INDUS', 'CHAS', 'NOX', 'RM', 'AGE', 'DIS', 'RAD', 'TAX', 'PTRATIO', 'B', 'LSTAT',
'MEDV']
# 列的长度
feature_num = len(feature_names)
# print(firstdata.shape) 输出结果:(7084, )
# print(firstdata.shape[0] // feature_nums) 输出结果:506
# 构造506*14的二维数组
data = firstdata.reshape([firstdata.shape[0] // feature_num, feature_num])
# 训练集设置为总数据的80%
ratio = 0.8
offset = int(data.shape[0] * ratio)
training_data = data[:offset]
# print(training_data.shape)
# axis=0表示列
# axis=1表示行
maximums, minimums, avgs = training_data.max(axis=0), training_data.min(axis=0), training_data.sum(axis=0) / \
training_data.shape[0]
# 查看训练集每列的最大值、最小值、平均值
# print(maximums, minimums, avgs)
# 对所有数据进行归一化处理
for i in range(feature_num):
# print(maximums[i], minimums[i], avgs[i])
# 归一化,减去平均值是为了移除共同部分,凸显个体差异
data[:, i] = (data[:, i] - avgs[i]) / (maximums[i] - minimums[i])
# 覆盖上面的训练集
training_data = data[:offset]
# 剩下的20%为测试集
test_data = data[offset:]
return training_data, test_data
class Network(object):
def __init__(self, num_of_weights):
np.random.seed(0)
self.w = np.random.randn(num_of_weights, 1)
# self.w[5] = -100.
# self.w[9] = -100.
self.b = 0.
def forward(self, x):
z = np.dot(x, self.w) + self.b
return z
#添加损失函数
def loss(self, z, y):
# 根据下面y的取值可以确定y的结构
error = z - y
# num_samples为总行数404
num_samples = error.shape[0]
# cost为均方误差,用来评价模型的好坏
cost = error * error
# 计算损失时需要把每个样本的损失都考虑到
# 对单个样本的损失函数进行求和,并除以样本总数
cost = np.sum(cost) / num_samples
return cost
#训练过程,计算梯度的另一种方法,可以不修改
def gradient(self, x, y):
z = self.forward(x)
# 取数据的行数
N = x.shape[0]
# 计算w的梯度,总数相加再除以N
gradient_w = 1. / N * np.sum((z - y) * x, axis=0)
# 增加维度
gradient_w = gradient_w[:, np.newaxis]
# 计算b的梯度,同上
gradient_b = 1. / N * np.sum(z - y)
return gradient_w, gradient_b
#确定损失函数更小的点封装在train和update函数中,并在Network中添加
def update(self, gradient_w, gradient_b, eta=0.01):
self.w = self.w - eta * gradient_w
self.b = self.b - eta * gradient_b
#训练数据导入后,越接近模型训练结束,最后几个批次数据对模型参数的影响越大。为了避免模型记忆影响训练效果,需要进行样本乱序操作。
# num_epoches为训练的轮数,eta为步长
def train(self, training_data, num_epoches, batch_size=10, eta=0.01):
n = len(training_data)
losses = []
for epoch_id in range(num_epoches):
# 打乱样本顺序
np.random.shuffle(training_data)
# 将train_data分成多个mini_batch
# 循环取值,每次取出batch_size条数据
mini_batches = [training_data[k:k + batch_size] for k in range(0, n, batch_size)]
for iter_id, mini_batche in enumerate(mini_batches):
# 取mini_batch的前13列
x = mini_batche[:, :-1]
# 取mini_batch的最后1列
y = mini_batche[:, -1:]
# 前向计算
a = self.forward(x)
# 计算损失
loss = self.loss(a, y)
# 计算梯度
gradient_w, gradient_b = self.gradient(x, y)
# 更新参数
self.update(gradient_w, gradient_b, eta)
losses.append(loss)
print('Epoch {:3d} / iter {:3d}, loss = {:.4f}'.format(epoch_id, iter_id, loss))
return losses
def saveModel(self, fileName):
np.savez(fileName, self.w, self.b)
#开始训练以及作图
# 获取数据
training_data, test_data = load_data()
# 创建网络
net = Network(13)
# 启动训练,训练10000轮,每轮样本数目为100,步长为0.1
losses = net.train(training_data, num_epoches=10000, batch_size=100, eta=0.1)
# 画出损失函数的变化趋势
plot_x = np.arange(len(losses))
plot_y = np.array(losses)
plt.plot(plot_x, plot_y)
plt.show()
net.saveModel("model2")
# 打印输出查看模型
print(np.load('model2.npz')["arr_0"], np.load('model2.npz')["arr_1"])
预测部分:这部分是自己写的,有点小白了
import numpy as np
# 加载训练文件
datafile1 = 'F:/boston-house-price-forecast-master/boston-house-price-forecast-master/housing.data' # 原始文件,506行14列
data4 = np.fromfile(datafile1, sep=' ') # 7084 = 14 * 506,一行数据显示
# 每条包含14列
feature_names1 = ['CRIM', 'ZN', 'INDUS', 'CHAS', 'NOX', 'RM', 'AGE', 'DIS','RAD', 'TAX', 'PTRATIO', 'B', 'LSTAT',
'MEDV']
feature_num1 = len(feature_names1) # 14列
# print(data.shape[0] // feature_num) # 7084//14 向下取整保留整数
data4 = data4.reshape([data4.shape[0] // feature_num1, feature_num1]) # 506行,14列
Y=data4[:,-1]
# 最大值,最小值,平均值
maximums =data4.max(axis=0)
minimums = data4.min(axis=0)
avgs = data4.sum(axis=0) / data4.shape[0]
datafile = 'F:/boston-house-price-forecast-master/boston-house-price-forecast-master/yuce.data'
feature_names = ['CRIM', 'ZN', 'INDUS', 'CHAS', 'NOX', 'RM', 'AGE', 'DIS', 'RAD', 'TAX', 'PTRATIO', 'B', 'LSTAT']
data = np.fromfile(datafile, sep=' ')
feature_num = len(feature_names)
d = data.reshape([data.shape[0] // feature_num, feature_num])
# 按照训练集归一化的规律对所有的预测数据进行归一化处理
for i in range(feature_num): # 从0到13列
d[:, i] = (d[:, i] - avgs[i]) / (maximums[i] - minimums[i]) # 逐列的归一化
d[np.isnan(d)]=0
#数组转矩阵,不然之后的矩阵乘法会出错
data=np.matrix(d)
Y=np.matrix(Y)
Y=np.transpose(Y)
#数组
d1=np.load('model2.npz')["arr_0"]
#数组转矩阵
data1=np.matrix(d1)
data2=np.load('model2.npz')["arr_1"]
data3=data*data1+data2
#去标准化
firstdata2 = np.fromfile('F:/boston-house-price-forecast-master/boston-house-price-forecast-master/housing.data', sep=' ')
feature_names2 = ['CRIM', 'ZN', 'INDUS', 'CHAS', 'NOX', 'RM', 'AGE', 'DIS', 'RAD', 'TAX', 'PTRATIO', 'B', 'LSTAT',
'MEDV']
feature_num2 = len(feature_names2)
# print(firstdata.shape) 输出结果:(7084, )
# print(firstdata.shape[0] // feature_nums) 输出结果:506
# 构造506*14的二维数组
data4 = firstdata2.reshape([firstdata2.shape[0] // feature_num2, feature_num2])
p=data4[:,-1]
max = np.max(p)
min = np.min(p)
avg =np.mean(p)
#输出预测结果
w=data3*(max-min)+avg
np.savetxt('test.txt',w,fmt='%s')
import matplotlib.pyplot as plt
plt.title(f"BP network")
plt.plot(np.arange(200), Y[:200], "go-", label="True value")
plt.plot(np.arange(200), w[:200], "ro-", label="Predict value")
plt.legend(loc="best")
plt.show()
plt.figure(figsize=(10, 6))
plt.title("importance of feature in dateset", fontsize=18)
plt.ylabel("import level", fontsize=15, rotation=90)
运行结果:
训练loss图:
预测结果: