IT工程師數位筆記本

 1 from mxnet import ndarray as nd
 2 from mxnet import autograd
 3 from mxnet import gluon
 4 
 5 num_inputs = 2
 6 num_examples = 1000
 7 
 8 true_w = [2, -3.4]
 9 true_b = 4.2
10 
11 X = nd.random_normal(shape=(num_examples, num_inputs)) #1000行，2列的數據集
12 y = true_w[0] * X[:, 0] + true_w[1] * X[:, 1] + true_b #已知答案的結果
13 y += .01 * nd.random_normal(shape=y.shape) #加入噪音
14 
15 #1 隨機讀取10行數據
16 batch_size = 10
17 dataset = gluon.data.ArrayDataset(X, y)
18 data_iter = gluon.data.DataLoader(dataset, batch_size, shuffle=True)
19 
20 #2 定義回歸模型
21 net = gluon.nn.Sequential()
22 net.add(gluon.nn.Dense(1))
23 
24 #3 參數初始化
25 net.initialize()
26 
27 #4 損失函數
28 square_loss = gluon.loss.L2Loss()
29 
30 #5 指定訓練方法
31 trainer = gluon.Trainer(net.collect_params(), 'sgd', {'learning_rate': 0.1})
32     
33 #6 訓練
34 epochs = 5
35 batch_size = 10
36 for e in range(epochs):
37     total_loss = 0
38     for data, label in data_iter:
39         with autograd.record():
40             output = net(data)
41             loss = square_loss(output, label)
42         loss.backward()
43         trainer.step(batch_size)
44         total_loss += nd.sum(loss).asscalar()
45     print("Epoch %d, average loss: %f" % (e, total_loss/num_examples))
46 
47 #7 輸出結果
48 dense = net[0]
49 print(true_w)
50 print(dense.weight.data())
51 print(true_b)
52 print(dense.bias.data())