（Tensorflow之三)神經網絡程序一

import tensorflow as tf
from numpy.random import RandomState

#輸入層(設有3個輸入變量1x3)
rdm = RandomState(1)
IN = rdm.rand(10000,3) #隨機生層10個3維數組，取值範圍在0～1之間；
x =tf.placeholder(tf.float32,shape=(None,3),name='x_input')

#設定輸入層到隱含層隨機權重（設隱含層有3個神經元3x3）
in_to_hd_w1 = tf.Variable(tf.random_normal([3,3], stddev=1,seed=1))

#隱含層3個神經元的值(1x3)
hd_layer = tf.matmul(x,in_to_hd_w1)

#隱含層到輸出層隨機權重(3x1)
hd_to_out_w2 = tf.Variable(tf.random_normal([3,1], stddev=1,seed=1))

#訓練輸出值,輸出值Y非零即1
OUT = [[int(x1+x2+x3<1)] for (x1,x2,x3) in IN]
y = tf.placeholder(tf.float32,shape=(None,1),name='y_output')

#僅有一個輸出值(1x1)
yy = tf.matmul(hd_layer,hd_to_out_w2)

sess = tf.Session()

#注：初始話全部函數
sess.run(tf.global_variables_initializer());

#訓練
#損失函數
learning_rate = 0.001    
data_size = 10
loss = -tf.reduce_mean(y*tf.log(tf.clip_by_value(yy,1e-10,1.0)))
training = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss);

#設定訓練的輪數
time_s = 1000
#設定每次訓練的個數
batch = 10
for i in range(0,10000,1):
    sess.run(training,feed_dict={x:IN[i:(i+1)*data_size],y:OUT[i:(i+1)*data_size]})
    if(i%1000==0):
        print("loss is:")
        print(sess.run(loss,feed_dict={x:IN,y:OUT}))
        #print("in value is:")
        #print(IN[i])
        #print("out value is:")
        #print(OUT[i])
        #print("predict out is:")  
        #print(sess.run(yy,feed_dict={x:IN[i:i+1]}))

運行結果：

loss is:
1.44677
loss is:
1.02309
loss is:
0.737368
loss is:
0.49516
loss is:
0.037638
loss is:
0.0149277
loss is:
0.0121234
loss is:
0.010189
loss is:
0.00865641
loss is:
0.00752897   

可以看到：損失值越來越小，學習趨於收斂；