CNN 卷积神经网络

LeNet

由 Yann LeCun 发明的著名的 LeNet.

原版中激活函数用的都是 $\mathrm{sigmoid}$, 现在看来用 $\mathrm{ReLU}$ 或 $\mathrm{tanh}$ 也许会更合适.  

Fashion - mnist 数据集的识别

数据下载

def gener():  
    trans = transforms.ToTensor()     
    mnist_train = torchvision.datasets.FashionMNIST(root="./data", train=True, transform=trans, download=False)
    mnist_test = torchvision.datasets.FashionMNIST(root="./data", train=False, transform=trans, download=False)
    return mnist_train, mnist_test   

　　

$\mathrm{root}$ ：下载路径

$\mathrm{train}$ ：是否是训练数据

$\mathrm{transform}$： 将数据类型变为 $\mathrm{torch}$

$\mathrm{download}$ 表示是否需要重新下载（如果下载过就记为 $\mathrm{False}$ )

test_data.test_data 是一个 $(10000, 28, 28)$ 的 $\mathrm{tensor}$ 

test_data.test_labels 是测试标签.   

 

Dataloader 打包

train_dataloader = DataLoader(dataset = train_data, batch_size = 4, shuffle=True)  

打包训练数据，其中 $\mathrm{batch_size}$ 是每次训练时塞入的训练数据个数.  

默认一次遍历是要把全部数据遍历完的（即 15000 批次，每批次 3 个）  

 

使用 GPU 训练

device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
net.apply(init_weights) 
net = net.to(device) 

在训练时会使用的矩阵后面加上 $\mathrm{.to(device)}$ 即可.　　

 

完整代码

import torch  
from torch import nn   
import d2l  
import torchvision 
from torchvision import transforms
from torch.utils.data import DataLoader

def gener():  
    trans = transforms.ToTensor()     
    mnist_train = torchvision.datasets.FashionMNIST(root="./data", train=True, transform=trans, download=False)
    mnist_test = torchvision.datasets.FashionMNIST(root="./data", train=False, transform=trans, download=False)
    return mnist_train, mnist_test   

def init_weights(m):   
    if type(m) == nn.Linear or type(m) == nn.Conv2d:  
        nn.init.xavier_uniform_(m.weight)   

# LeNet  
net = nn.Sequential(
    nn.Conv2d(in_channels = 1, out_channels = 6, kernel_size = 5, padding = 2), 
    nn.ReLU(),  
    nn.MaxPool2d(kernel_size = 2, stride = 2),

    nn.Conv2d(in_channels = 6, out_channels = 16, kernel_size = 5), 
    nn.ReLU(), 
    nn.MaxPool2d(kernel_size = 2, stride = 2),  

    nn.Flatten(),  

    nn.Linear(16 * 5 * 5, 120),  
    nn.ReLU(),

    nn.Linear(120, 84),   
    nn.ReLU(), 

    nn.Linear(84, 10)  
)

# 获取 fashion - mnist 数据集    

# 训练数据已经自动归一到 (0,1) 区间了.    
train_data, test_data = gener()  
train_dataloader = DataLoader(dataset = train_data, batch_size = 4, shuffle=True)    

test_x = torch.unsqueeze(test_data.test_data, dim = 1).type(torch.FloatTensor)[:2000] / 255.0
test_y = test_data.test_labels[:2000]       

device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
net.apply(init_weights) 
net = net.to(device)   

loss_func = nn.CrossEntropyLoss()  
optimizer = torch.optim.SGD(net.parameters(), lr = 0.0075)  
epoch = 2

for st in range(epoch):  
    for i, batch_sample in enumerate(train_dataloader):  
        train_x = batch_sample[0].to(device) 
        label = batch_sample[1].to(device)   

        optimizer.zero_grad()   
        output = net(train_x)  
        loss = loss_func(output, label)     
        loss.backward()  
        optimizer.step()      
        if i % 2000 == 0:  
            print(epoch + 1, i + 1, loss.item())
        if i > 6000:   
            break  
  
test_x = test_x.to(device)  
test_y = test_y.to(device)  
y_hat = net(test_x)  
pred_y = torch.max(y_hat, 1)[1].data.squeeze()
  
print((1000 - (test_y[:1000] != pred_y[:1000]).int().sum()) / 10)  

　　

 

标签：nn,卷积,torch,test,神经网络,train,CNN,data,mathrm
来源： https://www.cnblogs.com/brady12/p/16110626.html