定义add_layer添加层
1 def add_layer(inputs, in_size, activation_funcation=None): 2 Weights = tf.Variable(tf.random_normal([in_size, out_size])) 3 #Weights用随机数比全为0时的效果更好；定义其格式为：行是in_size行，列是out_size列 4 biases = tf.Variable(tf.zeros[1,out_size]+0.1) 5 #biases不允许为0，所以+0.1 6 Wx_plus_b = tf.matmul(inputs, Weights) + biases #y=Wx+b 7 8 if activation_function is None: 9 outputs = Wx_plus_b #如果无激励函数，就是线性的 10 else: 11 outputs = activation_function(Wx_plus_b) #调用激励函数 12 return outputs
数据
1 x_data = np.linspace(-1,1,300)[:,np.newaxis] #x_data有300个(-1,1)的例子，最后增加一个新维度 2 noise = np.random.normal(0,0.05,x_data.shape) #均值为0.方差为0.05，noise(噪点)的格式和x_data一样 3 y_data = np.square(x_data)-0.5+noise #np.square平方
方法
1 xs = tf.placeholder(tf.float32,[None,1]) #一次一个x_data;None-无论给多少个值都ok 2 ys = tf.placeholder(tf.float32,[None,1])
1 1l=add_layers(xs,1,10,activation_function=tf.nn.relu) 2 #输入层，输入的是xs,in_size=1,out_size=10,activation_function激励函数是relu 3 prediction = add_layer(1l,10,1,activation_function=None) #隐藏层 4 5 loss = tf.reduce_mean(tf.reduce_sum(tf.square(ys-prediction), 6 reduction_indices=[1])) 7 8 #reduce_sum对所有例子的loss求和再reduce_mean求平均 9 10 train_step = tf.train.GradientDescentOptimizer(0.1).minimize(loss) #学习效率要<1;Optimizer是优化器
1 init = tf.initialize_all_variables() #初始 2 sess = tf.Session() 3 sess.run(init)
执行
1 sess.run(init) 2 for i in range(1000): 3 sess.run(train_step,feed_dict={xs:x_data,ys:y_data}) 4 if i%50 ==0: #每50显示一下loss 5 print(sess.run(loss,feed_dict={xs:x_data,ys:y_data}))
