softmax回归的简洁实现

softmax回归的简洁实现

# 导包
%matplotlib inline
from mxnet import init
from mxnet.gluon import loss as gloss
from mxnet.gluon import nn, Trainer

1. 获取和读取数据

from utils import load_data_fashion_mnist
batch_size = 256
train_iter, test_iter = load_data_fashion_mnist(batch_size)

2. 定义和初始化模型

net = nn.Sequential()
net.add(nn.Dense(10))
net.initialize(init.Normal(sigma=0.01))

3. softmax和交叉熵损失函数

# Gluon提供了⼀个包括softmax运算和交叉熵损失计算的函数。它的数值稳定性更好。
loss = gloss.SoftmaxCrossEntropyLoss()

4. 定义优化算法

使用学习率为0.1的小批量随机梯度下降作为优化算法

trainer = Trainer(net.collect_params(), 'sgd', {'learning_rate': 0.1})

5. 训练模型

from utils import train_ch3
num_epochs = 5
train_ch3(net, train_iter, test_iter, loss, num_epochs, batch_size, None,
None, trainer)

epoch 1, loss 0.7897, train acc 0.745, test acc 0.806
epoch 2, loss 0.5734, train acc 0.811, test acc 0.817
epoch 3, loss 0.5286, train acc 0.823, test acc 0.825
epoch 4, loss 0.5047, train acc 0.830, test acc 0.835
epoch 5, loss 0.4899, train acc 0.834, test acc 0.839