linear_regression.py

# coding: utf8
# https://www.yewen.us/blog/2018/10/append-machine-learning-3-linear-regression/

from mxnet import autograd, nd

num_inputs = 9          # 特征数，当前问题里的变量数 1-9
num_examples = 1000     # 样例数，我们会随机生成多少份样例来学习
true_w = nd.array([0, 0, 0, 0, 0, 1, 0, 2, 1])  # 真实值
features = nd.random.normal(scale=1, shape=(num_examples, num_inputs))  # 随机生成数据集
labels = nd.dot(features, true_w)                                       # 数据集对应的结果

w = nd.random.normal(scale=0.01, shape=(9, 1))
w.attach_grad()


def linreg(X, w):
    return nd.dot(X, w)


def squared_loss(y_hat, y):
    return (y_hat - y.reshape(y_hat.shape)) ** 2 / 2


def sgd(param, lr, batch_size):
    param[:] = param - lr * param.grad / batch_size


def train():
    lr = 0.01
    num_epochs = 1000
    net = linreg
    loss = squared_loss

    for epoch in range(num_epochs):
        with autograd.record():
            l = loss(net(features, w), labels)
        l.backward()
        sgd(w, lr, labels.size)
        train_l = loss(net(features, w), labels)
        if epoch % 100 == 99:
            print('epoch {}, loss {}, w {}'.format(epoch + 1, train_l.mean().asnumpy(), w))


if __name__ == "__main__":
    train()
    test = nd.array([1, 0, 0, 2, 2, 1, 0, 0, 0])    # 测试集，565441
    print(nd.dot(test, w))