一、库&数据集
- 数据集简介
- 样本数：60,000
- 样本(图片像素)：28*28
10个类别：
0 T-shirt/top
1 Trouser
2 Pullover
3 Dress 礼服
4 Coat 外套
5 Sandal 凉鞋
6 Shirt 衬衫
7 Sneaker 运动鞋
8 Bag
9 Ankle boot 高跟鞋
样本
import torch
from torchvision import datasets, transforms
#Transform for the data
transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5, ), (0.5, ))])
#training data
trainset = datasets.FashionMNIST('~/.pytorch/F_MNIST_data', download=True, train=True, transform=transform)
trainloader = torch.utils.data.DataLoader(trainset, batch_size=64, shuffle=True)
#test data
testset= datasets.FashionMNIST('~/.pytorch/F_MNIST_data', download=True, train=False, transform=transform)
testloader = torch.utils.data.DataLoader(trainset, batch_size=64, shuffle=True)
- 目标
指标：训练损失低于某一阈值
二、模型的基本框架
- 搭建网络结构
#方法1）nn.Sequential()
model = nn.Sequential(nn.Linear(784, 196)
nn.ReLU(),
nn.Linear(196, 49),
nn.ReLU(),
nn.Linear(49, 10))
#方法2）封装成类(继承nn.Module)
class Classifier(nn.Module):
def __init__(self):
super().__init__()
self.fc1 = nn.Linear(784, 256)
self.fc2 = nn.Linear(256, 128)
self.fc3 = nn.Linear(128, 64)
self.fc4 = nn.Linear(64, 10)
def forward(self, x):
x = x.view(x.shape[0], -1)
x = F.relu(self.fc1(x))
x = F.relu(self.fc2(x))
x = F.relu(self.fc3(x))
x = F.log_softmax(self.fc4(x), dim=1)
return x
- 训练+验证测试
目标函数(损失函数)
nn.CrossEntropyLoss()优化器(权值更新)
optim.SGD()
optim.Adam()
#目标函数(损失函数)
criterion = nn.CrossEntropyLoss() #相当于 nn.LogSoftmax(dim=1) & nn.NLLLoss()
#优化器(权值更新)
optimizer = optim.SGD(model.parameters(), lr=0.03)
#模型训练
epoch = 5
for e in range(epoch):
running_loss = 0
for images, labels in trainloader:
images = images.view(images.shape[0], -1)
optimizer.zero_grad()
logits = model(images)
loss = criterion(logits, labels)
loss.backward()
optimizer.step()
running_loss += loss.item()
print("Training loss: {} \n" .format(running_loss/len(trainloader)))
#验证测试
%matplotlib inline
%config InlineBackend.figure_format = 'retina'
import helper #自定义模块
dataiter = iter(testloader)
images, labels = dataiter.next()
img = images[0]
img = img.resize_(1, 784) #行向量
ps = F.softmax(model(img), dim=1) #model 是nn.Sequantial()网络 缺少输出层
#可视化
helper.view_classify(img.resize_(1, 28, 28), ps, version='Fashion')
测试结果
三、模型优化
- 针对过拟合
方法1）早期停止early-stopping@复杂度图
比如8-10次迭代即可方法2）dropout
随机关闭输入(针对中间隐藏层)，防止一部分权值“拥兵自重”
#dropout完整代码
class Network(nn.Module):
def __init__(self):
super().__init__()
self.fc1 = nn.Linear(784, 196)
self.fc2 = nn.Linear(196, 49)
self.fc3 = nn.Linear(49, 10)
self.dropout = nn.Dropout(p=0.25)
def forward(self, x):
x = x.view(x.shape[0], -1)
x = self.dropout(F.relu(self.fc1(x)))
x = self.dropout(F.relu(self.fc2(x)))
x = F.log_softmax(self.fc3(x), dim=1)
return x
#训练模型
model = Network()
criterion = nn.NLLLoss()
optimizer = optim.Adam(model.parameters(), lr=0.002)
epochs = 6
train_losses, test_losses = [],[]
for e in range(epochs):
running_loss = 0
for images, labels in trainloader:
optimizer.zero_grad()
log_ps = model(images)
loss = criterion(log_ps, labels)
loss.backward()
optimizer.step()
running_loss += loss.item()
test_loss = 0
accuracy = 0
with torch.no_grad():
model.eval() #设置dropout
for images, labels in testloader:
log_ps = model(images) #结果维度 64*10
test_loss += criterion(log_ps, labels)
ps = torch.exp(log_ps)
tp_ps, tp_class = ps.topk(1, dim=1) #结果维度 64*1
equals = (tp_class == labels.view(*tp_class.shape))
accuracy += torch.mean(equals.type(torch.FloatTensor))
train_losses.append(running_loss/len(trainloader))
test_losses.append(test_loss/len(testloader))
model.train()
print("{}" .format())
#验证测试
%matplotlib inline
%config InlineBackend.figure_format = 'retina'
import helper #自定义模块
model.eval() #测试环节，关闭dropout
dataiter = iter(testloader)
images, labels = dataiter.next()
img = images[0]
img = img.view(1, 784) #行向量
with torch.no_grad():
output = model(img)
ps = torch.exp(output)
#可视化
helper.view_classify(img.view(1, 28, 28), ps, version='Fashion')
dropout后的训练结果
References
nn.Module类详解