從最初的簡單實現(xiàn)菱皆，到后面一步步的整合代碼塊，終于達(dá)到了可讀挨稿、便于調(diào)試的程度仇轻。代碼雖然清晰了，但是問題依然存在奶甘。目前主要的問題便是權(quán)重學(xué)習(xí)不到東西篷店，loss總是不下降。
目前的版本loss可以下降疲陕，很快下降到0，但是查看生成的y值钉赁，與真實值差距很大蹄殃，故推斷l(xiāng)oss存在問題。
import numpy as np
import pandas as pd
import tensorflow as tf
#轉(zhuǎn)為onehot編碼
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
#獲取一批次的數(shù)據(jù)
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激活函數(shù)你踩，備用函數(shù)還有relu
? ? ? ? tf.summary.histogram('layer'+layer+'/l'+layer, l)
? ? return l
hiddenDim = 200 #隱藏層神經(jīng)元數(shù)
save_file = './train_model.ckpt'
istrain = True
istensorborad = False
pointer = 0
if istrain:
? ? samples = 400
? ? batch = 5 #每批次的數(shù)據(jù)輸入數(shù)量
else:
? ? samples = 550
? ? batch = 1 #每批次的數(shù)據(jù)輸入數(shù)量
with tf.name_scope('inputdate-x-y'):
? ? #導(dǎo)入
? ? df = pd.DataFrame(pd.read_csv('GHMX.CSV',header=0))
? ? #產(chǎn)生 y_data 值 (1, n)
? ? y_date = df['number'].values
? ? y_date = y_date.reshape((-1,1))
? ? #產(chǎn)生 x_data 值 (n, 4+12+31+24)
? ? df = df.drop('number', axis=1)
? ? df = turn_onehot(df)
? ? x_data = df.values
###生成神經(jīng)網(wǎng)絡(luò)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')
#生成神經(jīng)網(wǎng)絡(luò)
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)
#計算loss
with tf.name_scope('loss'):
? ? #loss = tf.reduce_mean(tf.reduce_sum(tf.square(y - y_), name='square'), name='loss')? #損失函數(shù)，損失不下降，換用別的函數(shù)
? ? #loss = -tf.reduce_sum(y_*tf.log(y))? #損失仍然不下降
? ? 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/讀取模型
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)
