Tensorflow:深度神经网络DNN预测波士顿房价（boston house price）【一】

刚学习TF，尝试拿sklearn的datasets来做下练习。波士顿房价数据是个不错的选择。
python3.5
tensorflow 0.12

把所有的包import进来

# coding: utf-8
import tensorflow as tf
from sklearn.datasets import load_boston
import matplotlib.pyplot as plt
from sklearn.preprocessing import scale
from sklearn.model_selection import train_test_split

获取数据

boston = load_boston()
# X = scale(boston.data)
# y = scale(boston.target.reshape((-1,1)))


X_train,X_test,y_train,y_test = train_test_split(boston.data,boston.target,test_size=0.1,random_state=0)
X_train = scale(X_train)
X_test = scale(X_test)
y_train = scale(y_train.reshape((-1,1)))
y_test = scale(y_test.reshape((-1,1)))

定义每一层网络结构，写了个add_layer，让添加网络更加灵活。
完全参照莫烦老师：https://morvanzhou.github.io/tutorials/machine-learning/tensorflow/3-1-add-layer/

def add_layer(inputs,input_size,output_size,activation_function=None):
    with tf.variable_scope("Weights"):
        Weights = tf.Variable(tf.random_normal(shape=[input_size,output_size]),name="weights")
    with tf.variable_scope("biases"):
        biases = tf.Variable(tf.zeros(shape=[1,output_size]) + 0.1,name="biases")
    with tf.name_scope("Wx_plus_b"):
        Wx_plus_b = tf.matmul(inputs,Weights) + biases
    with tf.name_scope("dropout"):
        Wx_plus_b = tf.nn.dropout(Wx_plus_b,keep_prob=keep_prob_s)
    if activation_function is None:
        return Wx_plus_b
    else:
        with tf.name_scope("activation_function"):
            return activation_function(Wx_plus_b)

定义占位符和网络层数

xs = tf.placeholder(shape=[None,X_train.shape[1]],dtype=tf.float32,name="inputs")
ys = tf.placeholder(shape=[None,1],dtype=tf.float32,name="y_true")
keep_prob_s = tf.placeholder(dtype=tf.float32)

with tf.name_scope("layer_1"):
    l1 = add_layer(xs,13,10,activation_function=tf.nn.relu)
# with tf.name_scope("layer_2"):
#     l2 = add_layer(l1,6,10,activation_function=tf.nn.relu)
with tf.name_scope("y_pred"):
    pred = add_layer(l1,10,1)

# 这里多于的操作，是为了保存pred的操作，做恢复用。我只知道这个笨方法。
pred = tf.add(pred,0,name='pred')

with tf.name_scope("loss"):
    loss = tf.reduce_mean(tf.reduce_sum(tf.square(ys - pred),reduction_indices=[1]))  # mse
    tf.summary.scalar("loss",tensor=loss)
with tf.name_scope("train"):
    # train_op =tf.train.GradientDescentOptimizer(learning_rate=0.01).minimize(loss)
    train_op = tf.train.AdamOptimizer(learning_rate=0.01).minimize(loss)

数据可视化，训练参数的定义

# draw pics
fig = plt.figure()
ax = fig.add_subplot(1,1,1)
ax.plot(range(50),y_train[0:50],'b')  #展示前50个数据
ax.set_ylim([-2,5])
plt.ion()
plt.show()

# parameters
keep_prob=1  # 防止过拟合，取值一般在0.5到0.8。我这里是1，没有做过拟合处理
ITER =5000  # 训练次数

定义训练过程

def fit(X, y, ax, n, keep_prob):
    init = tf.global_variables_initializer()
    feed_dict_train = {ys: y, xs: X, keep_prob_s: keep_prob}
    with tf.Session() as sess:
        saver = tf.train.Saver(tf.global_variables(), max_to_keep=15)
        merged = tf.summary.merge_all()
        writer = tf.summary.FileWriter(logdir="nn_boston_log", graph=sess.graph)  #写tensorbord
        sess.run(init)
        for i in range(n):
            _loss, _ = sess.run([loss, train_op], feed_dict=feed_dict_train)

            if i % 100 == 0:
                print("epoch:%d\tloss:%.5f" % (i, _loss))
                y_pred = sess.run(pred, feed_dict=feed_dict_train)
                rs = sess.run(merged, feed_dict=feed_dict_train)
                writer.add_summary(summary=rs, global_step=i)  #写tensorbord
                saver.save(sess=sess, save_path="nn_boston_model/nn_boston.model", global_step=i) # 保存模型
                try:
                    ax.lines.remove(lines[0])
                except:
                    pass
                lines = ax.plot(range(50), y_pred[0:50], 'r--')
                plt.pause(1)

        saver.save(sess=sess, save_path="nn_boston_model/nn_boston.model", global_step=n)  # 保存模型

训练网络

fit(X=X_train,y=y_train,n=ITER,keep_prob=keep_prob,ax=ax)

epoch:100 loss:0.76055
epoch:200 loss:0.31463
epoch:300 loss:0.19911
epoch:400 loss:0.14731
epoch:500 loss:0.12500
……..
epoch:4700 loss:0.04773
epoch:4800 loss:0.04756
epoch:4900 loss:0.04745

蓝色为实际数据，红色虚线为拟合数据。

代码里面有写tensorbord，开启终端：
(我这里写到了nn_boston_log)

tensorbord --logdir='nn_boston_log'

可以查看loss的变化和网络结构图