初始版本代码（伪）

从最初的简单实现，到后面一步步的整合代码块，终于达到了可读、便于调试的程度。代码虽然清晰了，但是问题依然存在。目前主要的问题便是权重学习不到东西，loss总是不下降。

代码成长历程，保存了每个版本

---

目前的版本loss可以下降，很快下降到0，但是查看生成的y值，与真实值差距很大，故推断loss存在问题。

import numpy as np

import pandas as pd

import tensorflow as tf

#转为onehot编码

def turn_onehot(df):

    for key in df.columns:

        oneHot = pd.get_dummies(df[key])

        for oneHotKey in oneHot.columns: #防止重名

            oneHot = oneHot.rename(columns={oneHotKey : key+'_'+str(oneHotKey)})

        df = df.drop(key, axis=1)

        df = df.join(oneHot)

    return df

#获取一批次的数据

def get_batch(x_date, y_date, batch):

    global pointer

    x_date_batch = x_date[pointer:pointer+batch]

    y_date_batch = y_date[pointer:pointer+batch]

    pointer = pointer + batch

    return x_date_batch, y_date_batch

#生成layer

def add_layer(input_num, output_num, x, layer, active=None):

    with tf.name_scope('layer'+layer+'/W'+layer):

        W = tf.Variable(tf.random_normal([input_num, output_num]), name='W'+layer)

        tf.summary.histogram('layer'+layer+'/W'+layer, W)

    with tf.name_scope('layer'+layer+'/b'+layer):

        b = tf.Variable(tf.zeros([1, output_num])+0.1, name='b'+layer)

        tf.summary.histogram('layer'+layer+'/b'+layer, b)

    with tf.name_scope('layer'+layer+'/l'+layer):

        l = active(tf.matmul(x, W)+b) #使用sigmoid激活函数，备用函数还有relu

        tf.summary.histogram('layer'+layer+'/l'+layer, l)

    return l

hiddenDim = 200 #隐藏层神经元数

save_file = './train_model.ckpt'

istrain = True

istensorborad = False

pointer = 0

if istrain:

    samples = 400

    batch = 5 #每批次的数据输入数量

else:

    samples = 550

    batch = 1 #每批次的数据输入数量

with tf.name_scope('inputdate-x-y'):

    #导入

    df = pd.DataFrame(pd.read_csv('GHMX.CSV',header=0))

    #产生 y_data 值 (1, n)

    y_date = df['number'].values

    y_date = y_date.reshape((-1,1))

    #产生 x_data 值 (n, 4+12+31+24)

    df = df.drop('number', axis=1)

    df = turn_onehot(df)

    x_data = df.values

###生成神经网络模型

#占位符

with tf.name_scope('inputs'):

    x = tf.placeholder("float", shape=[None, 71], name='x_input')

    y_ = tf.placeholder("float", shape=[None, 1], name='y_input')

#生成神经网络

l1 = add_layer(71, hiddenDim, x, '1', tf.nn.relu)

l2 = add_layer(hiddenDim, hiddenDim, l1, '2', tf.nn.relu)

#l3 = add_layer(hiddenDim, hiddenDim, l2, '3', tf.nn.relu)

#l4 = add_layer(hiddenDim, hiddenDim, l3, '4', tf.nn.relu)

#l5 = add_layer(hiddenDim, hiddenDim, l4, '5', tf.nn.relu)

#l6 = add_layer(hiddenDim, hiddenDim, l5, '6', tf.nn.relu)

#l7 = add_layer(hiddenDim, hiddenDim, l6, '7', tf.nn.relu)

#l8 = add_layer(hiddenDim, hiddenDim, l7, '8', tf.nn.relu)

#l9 = add_layer(hiddenDim, hiddenDim, l8, '9', tf.nn.relu)

y = add_layer(hiddenDim, 1, l2, '10', tf.nn.relu)

#计算loss

with tf.name_scope('loss'):

    #loss = tf.reduce_mean(tf.reduce_sum(tf.square(y - y_), name='square'), name='loss')  #损失函数，损失不下降，换用别的函数

    #loss = -tf.reduce_sum(y_*tf.log(y))  #损失仍然不下降

    loss = -tf.reduce_sum(y_*tf.log(tf.clip_by_value(y,1e-10,1.0)) , name='loss')

    tf.summary.scalar('loss', loss)

#梯度下降

with tf.name_scope('train_step'):

    train_step = tf.train.GradientDescentOptimizer(0.0005).minimize(loss)

#初始化

sess = tf.Session()

if istensorborad:

    merged = tf.summary.merge_all()

    writer = tf.summary.FileWriter('logs/', sess.graph)

sess.run(tf.initialize_all_variables())

#保存/读取模型

saver = tf.train.Saver()

if not istrain:

    saver.restore(sess, save_file)

for i in range(samples):

    x_date_batch, y_date_batch = get_batch(x_data, y_date, batch)

    feed_dict = {x: x_date_batch, y_: y_date_batch}

    if istrain:

        sess.run(train_step, feed_dict=feed_dict)

        print(y.eval(feed_dict, sess))

    else:

        sess.run(loss, feed_dict=feed_dict)

        print(test_assess.eval(feed_dict, sess))

    if istensorborad:

        result = sess.run(merged, feed_dict=feed_dict)

        writer.add_summary(result,i)

#保存模型

if istrain:

    saver.save(sess, save_file)