在tensorflow里RNN才是做回归计算的正规军,其中LSTM更是让人工智能有了记忆,如果cnn最适合做的是图像识别,那么LSTM就是视频识别。网上的教程多是用正余弦数据在做预测,输入输出都是一维,我这用波士顿房价,输入是13个特征!
注意与前面两个模型不同的是,没有用train_test_split把训练数据分割,而是用的时序数据。
代码中注释比较少,不明白的可以看周莫烦的视频!
https://morvanzhou.github.io/tutorials/machine-learning/tensorflow/5-09-RNN3/
# 参考https://morvanzhou.github.io/tutorials/machine-learning/tensorflow/5-09-RNN3/
from sklearn.datasets import load_boston
from sklearn import preprocessing
import tensorflow as tf
import numpy as np
# 波士顿房价数据
boston = load_boston()
x = boston.data
y = boston.target
print('波士顿数据X:',x.shape)# (506, 13)
# print(x[::100])
print('波士顿房价Y:',y.shape)
# print(y[::100])
# 数据标准化
ss_x = preprocessing.StandardScaler()
train_x = ss_x.fit_transform(x)
ss_y = preprocessing.StandardScaler()
train_y = ss_y.fit_transform(y.reshape(-1, 1))
BATCH_START = 0 # 建立 batch data 时候的 index
TIME_STEPS = 10 # backpropagation through time 的 time_steps
BATCH_SIZE = 30
INPUT_SIZE = 13 # sin 数据输入 size
OUTPUT_SIZE = 1 # cos 数据输出 size
CELL_SIZE = 10 # RNN 的 hidden unit size
LR = 0.006 # learning rate
def get_batch_boston():
global train_x, train_y,BATCH_START, TIME_STEPS
x_part1 = train_x[BATCH_START : BATCH_START+TIME_STEPS*BATCH_SIZE]
y_part1 = train_y[BATCH_START : BATCH_START+TIME_STEPS*BATCH_SIZE]
print('时间段=', BATCH_START, BATCH_START + TIME_STEPS * BATCH_SIZE)
seq =x_part1.reshape((BATCH_SIZE, TIME_STEPS ,INPUT_SIZE))
res =y_part1.reshape((BATCH_SIZE, TIME_STEPS ,1))
BATCH_START += TIME_STEPS
# returned seq, res and xs: shape (batch, step, input)
#np.newaxis 用来增加一个维度 变为三个维度,第三个维度将用来存上一批样本