Gluon sgd

from mxnet import nd,autograd,init,gluon

from mxnet.gluon import data as gdata,loss as gloss,nn

num_inputs = 2

num_examples = 1000

true_w = [2,-3.4]

true_b = 4.2

features = nd.random.normal(scale=1,shape=(num_examples,num_inputs))

labels = true_w[0]*features[:,0] + true_w[1]*features[:,1] + true_b

labels += nd.random.normal(scale=0.01,shape=labels.shape)

# 造小批量数据集

dataset = gdata.ArrayDataset(features,labels)

batch_size = 10

data_iter = gdata.DataLoader(dataset,batch_size,shuffle=True)

# 定义网络

net = nn.Sequential()

net.add(nn.Dense(1))

net.initialize(init.Normal(sigma=0))

# 损失函数

loss = gloss.L2Loss()

# 优化算法

trainer = gluon.Trainer(net.collect_params(),'sgd',{'learning_rate':0.01})

num_epochs = 3

for epoch in range(1, num_epochs + 1):

    for X, y in data_iter:

        print(X)

        print(y)

        with autograd.record():

            l = loss(net(X), y)

        print(l)

        l.backward()

        trainer.step(batch_size)

    l = loss(net(features), labels)

    print('epoch %d, loss: %f' % (epoch, l.mean().asnumpy()))
    

从最简单的线性回归来说，小批量随机梯度下降的时候，X，y 从迭代器中取出，也是bach_size大小的数据集，那么网络的计算，同样也是小批量的。

即代码 l = loss(net(X),y) 包含了，小批量数据集，每一个数据丢到网络中，计算出返回值以后，和真实值得损失。