【theano】04-regression回归例子(简易神经网络)

1.导入包

2.定义Layer

 1 class Layer(object):
 2 
 3 def __init__(self, inputs, in_size, out_size, activation_function=None):
 4     self.W = theano.shared(np.random.normal(0, 1, (in_size, out_size)))
 5     self.b = theano.shared(np.zeros((out_size, )) + 0.1)
 6     self.Wx_plus_b = T.dot(inputs, self.W) + self.b
 7     self.activation_function = activation_function
 8     if activation_function is None:
 9         self.outputs = self.Wx_plus_b
10     else:
11         self.outputs = self.activation_function(self.Wx_plus_b)

3.Make up some fake data虚拟数据

1 x_data = np.linspace(-1, 1, 300)[:, np.newaxis]
2 noise = np.random.normal(0, 0.05, x_data.shape)
3 y_data = np.square(x_data) - 0.5 + noise # y = x^2 - 0.5

4.show the fake data可视化虚拟数据

1 plt.scatter(x_data, y_data)
2 plt.show()

5.determind the inputs dtype

1 x = T.dmatrix("x")
2 y = T.dmatrix("y")

6.网络搭建

1 l1 = Layer(x, 1, 10, T.nnet.relu)
2 l2 = Layer(l1.outputs, 10, 1, None)

7.计算Loss

1 # compute the loss
2 cost = T.mean(T.square(l2.outputs - y))
3 # compute the gradients
4 gW1, gb1, gW2, gb2 = T.grad(cost, [l1.W, l1.b, l2.W, l2.b])
5 #grad是梯度计算

8.梯度下降

 1 learning_rate = 0.05
 2 #学习率<1
 3 train = theano.function(
 4     inputs=[x, y],
 5     outputs=[cost],
 6     updates=[(l1.W, l1.W - learning_rate * gW1),
 7     (l1.b, l1.b - learning_rate * gb1),
 8     (l2.W, l2.W - learning_rate * gW2),
 9     (l2.b, l2.b - learning_rate * gb2)])
10 #updates更新参数

9.预测

1 predict = theano.function(inputs=[x], outputs=l2.outputs)
2 #输入x预测y

10.训练

1 for i in range(1000):
2     # training
3     err = train(x_data, y_data)    #放入数据
4     if i % 50 == 0:
5             print(err)

11.可视化结果

1 # plot the real data
2 fig = plt.figure()
3 ax.scatter(x_data, y_data)    #scatter散点形式
4 plt.show()
5 
6 ax = fig.add_subplot(1,1,1)    #从画布的1行1列第1点开始
7 plt.ion()        #显示画布后不停止运行程序（终止block）
8 plt.show()

 1 for i in range(1000):
 2     # training
 3     err = train(x_data, y_data)    #放入数据
 4     if i % 50 == 0:
 5     # to visualize the result and improvement
 6         try:    #在第一次时跳过remove命令
 7             ax.lines.remove(lines[0])       #清除上一次的线
 8         except Exception:
 9             pass
10     prediction_value = predict(x_data)
11     # plot the prediction
12     lines = ax.plot(x_data, prediction_value, 'r-', lw=5)    
13     #r red - 线 lw 粗细
14     plt.pause(.5)
15     #暂停0.5s

posted on 2022-08-02 15:22 Jolyne123 阅读(36) 评论(0) 收藏举报