MXNet动手学深度学习笔记：Gluon实现正则化

1#coding:utf-8
2'''
3正则化
4'''
5import mxnet as mx
6from mxnet import gluon
7from mxnet import ndarray
8from mxnet import autograd
9import numpy as np
10import matplotlib.pyplot as plt
11from mxnet import nd
12import random
13
14num_train = 20
15num_test = 100
16num_inputs = 200
17#模型真实参数
18true_w = nd.ones((num_inputs,1))
19true_b = 0.05
20
21#生成测试和训练数据
22X = nd.random.normal(shape=(num_train+num_test,num_inputs))
23y = nd.dot(X,true_w) + true_b
24y += 0.01 * nd.random.normal(shape=y.shape)
25
26X_train,X_test = X[:num_train,:],X[num_train:,:]
27y_train,y_test = y[:num_train],y[num_train:]
28
29batch_size = 1
30dataset_train = gluon.data.ArrayDataset(X_train, y_train)
31data_iter_train = gluon.data.DataLoader(dataset_train, batch_size,shuffle=True)
32
33# 定义损失函数
34square_loss = gluon.loss.L2Loss()
35
36# 定义测试函数
37def test(net,X,y):
38    return square_loss(net(X),y).mean().asscalar()
39
40# 定义训练函数
41def train(weight_decay):
42    epochs = 10
43    learning_rate = 0.005
44    net = gluon.nn.Sequential()
45    with net.name_scope():
46        net.add(gluon.nn.Dense(1))
47    
48    net.collect_params().initialize()
49    trainer = gluon.Trainer(net.collect_params(),'sgd',
50            {'learning_rate':learning_rate,'wd':weight_decay})
51
52    train_loss = []
53    test_loss = []
54
55    for e in range(epochs):
56        for data,label in data_iter_train:
57            with autograd.record():
58                output = net(data)
59                loss = square_loss(output,label)
60            loss.backward()
61            trainer.step(batch_size)
62
63        train_loss.append(test(net,X_train,y_train))
64        test_loss.append(test(net,X_test,y_test))
65
66    
67    plt.plot(train_loss)
68    plt.plot(test_loss)
69    plt.legend(['train','test'])
70    plt.show()
71
72# 未使用正则化
73# train(0)
74
75# 使用正则化
76train(5)