文章作者：Tyan
博客：noahsnail.com | CSDN | 简书
注：本文为李沐大神的《动手学深度学习》的课程笔记！
# 导入mxnet
import mxnet as mx
# 设置随机种子
mx.random.seed(2)
from mxnet import gluon
from mxnet import ndarray as nd
from mxnet import autograd
from mxnet import image
辅助函数
from utils import load_data_fashion_mnist, accuracy, evaluate_accuracy
获取和读取数据
# 批数据大小
batch_size = 256
# 获取训练数据和测试数据
train_data, test_data = load_data_fashion_mnist(batch_size)
定义和初始化模型
# 定义一个空的模型
net = gluon.nn.Sequential()
# name_scope作用, 方便管理参数命名
with net.name_scope():
# 加入一个平铺层, 其会将输入数据平铺为batch_size*?维
net.add(gluon.nn.Flatten())
# 加入一个全连接层, 输出为10类
net.add(gluon.nn.Dense(10))
# 参数初始化
net.initialize()
Softmax和交叉熵损失函数
# 定义交叉熵损失
softmax_cross_entropy = gluon.loss.SoftmaxCrossEntropyLoss()
优化
# 定义训练器和优化方法
trainer = gluon.Trainer(net.collect_params(), 'sgd', {'learning_rate': 0.1})
训练
# 定义迭代周期
epochs = 5
# 训练
for epoch in range(epochs):
# 训练损失
train_loss = 0.0
# 训练集准确率
train_acc = 0.0
# 迭代训练
for data, label in train_data:
# 记录梯度
with autograd.record():
# 计算输出
output = net(data)
# 计算损失
loss = softmax_cross_entropy(output, label)
# 反向传播求梯度
loss.backward()
# 梯度下降
trainer.step(batch_size)
# 总的训练损失
train_loss += nd.mean(loss).asscalar()
# 总的训练准确率
train_acc += accuracy(output, label)
# 测试集的准确率
test_acc = evaluate_accuracy(test_data, net)
print("Epoch %d. Loss: %f, Train acc %f, Test acc %f" % (
epoch, train_loss / len(train_data), train_acc / len(train_data), test_acc))
Epoch 0. Loss: 0.793821, Train acc 0.744107, Test acc 0.786659
Epoch 1. Loss: 0.575076, Train acc 0.809879, Test acc 0.820112
Epoch 2. Loss: 0.530560, Train acc 0.822583, Test acc 0.831731
Epoch 3. Loss: 0.506161, Train acc 0.829728, Test acc 0.835837
Epoch 4. Loss: 0.488752, Train acc 0.834769, Test acc 0.834135
