构建数据集

1# -*- coding: utf-8 -*-
2from mxnet import init
3from mxnet import ndarray as nd
4from mxnet.gluon import loss as gloss
5import gb
6
7n_train = 20
8n_test = 100
9
10num_inputs = 200
11true_w = nd.ones((num_inputs, 1)) * 0.01
12true_b = 0.05
13features = nd.random.normal(shape=(n_train+n_test, num_inputs))
14labels = nd.dot(features, true_w) + true_b
15labels += nd.random.normal(scale=0.01, shape=labels.shape)
16train_features, test_features = features[:n_train, :], features[n_train:, :]
17train_labels, test_labels = labels[:n_train], labels[n_train:]

数据迭代器

1from mxnet import autograd
2from mxnet.gluon import data as gdata
3
4batch_size = 1
5num_epochs = 10
6learning_rate = 0.003
7
8train_iter = gdata.DataLoader(gdata.ArrayDataset(
9    train_features, train_labels), batch_size, shuffle=True)
10loss = gloss.L2Loss()

训练并展示结果

gb.semilogy函数：绘制训练和测试数据的loss

1from mxnet import gluon
2from mxnet.gluon import nn
3
4def fit_and_plot(weight_decay):
5    net = nn.Sequential()
6    net.add(nn.Dense(1))
7    net.initialize(init.Normal(sigma=1))
8    # 对权重参数做 L2 范数正则化，即权重衰减。
9    trainer_w = gluon.Trainer(net.collect_params('.*weight'), 'sgd', {
10        'learning_rate': learning_rate, 'wd': weight_decay})
11    # 不对偏差参数做 L2 范数正则化。
12    trainer_b = gluon.Trainer(net.collect_params('.*bias'), 'sgd', {
13        'learning_rate': learning_rate})
14    train_ls = []
15    test_ls = []
16    for _ in range(num_epochs):
17        for X, y in train_iter:
18            with autograd.record():
19                l = loss(net(X), y)
20            l.backward()
21            # 对两个 Trainer 实例分别调用 step 函数。
22            trainer_w.step(batch_size)
23            trainer_b.step(batch_size)
24        train_ls.append(loss(net(train_features),
25                             train_labels).mean().asscalar())
26        test_ls.append(loss(net(test_features),
27                            test_labels).mean().asscalar())
28    gb.semilogy(range(1, num_epochs + 1), train_ls, 'epochs', 'loss',
29                range(1, num_epochs + 1), test_ls, ['train', 'test'])
30    return 'w[:10]:', net[0].weight.data()[:, :10], 'b:', net[0].bias.data()
31print fit_and_plot(5)

• 使用 Gluon 的 wd 超参数可以使用权重衰减来应对过拟合问题。
• 我们可以定义多个 Trainer 实例对不同的模型参数使用不同的迭代方法。