用PyTorch实现MNIST手写体识别

实验环境

• win10 + anaconda + jupyter notebook
• Pytorch1.1.0
• Python3.7
• gpu环境（可选）

MNIST数据集介绍

MNIST 包括6万张28x28的训练样本，1万张测试样本，可以说是CV里的“Hello Word”。本文使用的CNN网络将MNIST数据的识别率提高到了99%。下面我们就开始进行实战。

导入包

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torchvision import datasets, transforms
torch.__version__

定义超参数

BATCH_SIZE=512
EPOCHS=20 
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu") 

数据集

我们直接使用PyTorch中自带的dataset，并使用DataLoader对训练数据和测试数据分别进行读取。如果下载过数据集这里download可选择False

train_loader = torch.utils.data.DataLoader(
        datasets.MNIST('data', train=True, download=True, 
                       transform=transforms.Compose([
                           transforms.ToTensor(),
                           transforms.Normalize((0.1307,), (0.3081,))
                       ])),
        batch_size=BATCH_SIZE, shuffle=True)

test_loader = torch.utils.data.DataLoader(
        datasets.MNIST('data', train=False, transform=transforms.Compose([
                           transforms.ToTensor(),
                           transforms.Normalize((0.1307,), (0.3081,))
                       ])),
        batch_size=BATCH_SIZE, shuffle=True)

定义网络

该网络包括两个卷积层和两个线性层，最后输出10个维度，即代表0-9十个数字。

class ConvNet(nn.Module):
    def __init__(self):
        super().__init__()
        self.conv1=nn.Conv2d(1,10,5) # input:(1,28,28) output:(10,24,24) 
        self.conv2=nn.Conv2d(10,20,3) # input:(10,12,12) output:(20,10,10)
        self.fc1 = nn.Linear(20*10*10,500)
        self.fc2 = nn.Linear(500,10)
    def forward(self,x):
        in_size = x.size(0)
        out = self.conv1(x)
        out = F.relu(out)
        out = F.max_pool2d(out, 2, 2)  
        out = self.conv2(out)
        out = F.relu(out)
        out = out.view(in_size,-1)
        out = self.fc1(out)
        out = F.relu(out)
        out = self.fc2(out)
        out = F.log_softmax(out,dim=1)
        return out

实例化网络

model = ConvNet().to(DEVICE) # 将网络移动到gpu上
optimizer = optim.Adam(model.parameters()) # 使用Adam优化器

定义训练函数

def train(model, device, train_loader, optimizer, epoch):
    model.train()
    for batch_idx, (data, target) in enumerate(train_loader):
        data, target = data.to(device), target.to(device)
        optimizer.zero_grad()
        output = model(data)
        loss = F.nll_loss(output, target)
        loss.backward()
        optimizer.step()
        if(batch_idx+1)%30 == 0: 
            print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
                epoch, batch_idx * len(data), len(train_loader.dataset),
                100. * batch_idx / len(train_loader), loss.item()))

定义测试函数

def test(model, device, test_loader):
    model.eval()
    test_loss = 0
    correct = 0
    with torch.no_grad():
        for data, target in test_loader:
            data, target = data.to(device), target.to(device)
            output = model(data)
            test_loss += F.nll_loss(output, target, reduction='sum').item() # 将一批的损失相加
            pred = output.max(1, keepdim=True)[1] # 找到概率最大的下标
            correct += pred.eq(target.view_as(pred)).sum().item()

    test_loss /= len(test_loader.dataset)
    print('\nTest set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format(
        test_loss, correct, len(test_loader.dataset),
        100. * correct / len(test_loader.dataset)))

开始训练

for epoch in range(1, EPOCHS + 1):
    train(model, DEVICE, train_loader, optimizer, epoch)
    test(model, DEVICE, test_loader)

实验结果

Train Epoch: 1 [14848/60000 (25%)]	Loss: 0.375058
Train Epoch: 1 [30208/60000 (50%)]	Loss: 0.255248
Train Epoch: 1 [45568/60000 (75%)]	Loss: 0.128060

Test set: Average loss: 0.0992, Accuracy: 9690/10000 (97%)

Train Epoch: 2 [14848/60000 (25%)]	Loss: 0.093066
Train Epoch: 2 [30208/60000 (50%)]	Loss: 0.087888
Train Epoch: 2 [45568/60000 (75%)]	Loss: 0.068078

Test set: Average loss: 0.0599, Accuracy: 9816/10000 (98%)

Train Epoch: 3 [14848/60000 (25%)]	Loss: 0.043926
Train Epoch: 3 [30208/60000 (50%)]	Loss: 0.037321
Train Epoch: 3 [45568/60000 (75%)]	Loss: 0.068404

Test set: Average loss: 0.0416, Accuracy: 9859/10000 (99%)

Train Epoch: 4 [14848/60000 (25%)]	Loss: 0.031654
Train Epoch: 4 [30208/60000 (50%)]	Loss: 0.041341
Train Epoch: 4 [45568/60000 (75%)]	Loss: 0.036493

Test set: Average loss: 0.0361, Accuracy: 9873/10000 (99%)

Train Epoch: 5 [14848/60000 (25%)]	Loss: 0.027688
Train Epoch: 5 [30208/60000 (50%)]	Loss: 0.019488
Train Epoch: 5 [45568/60000 (75%)]	Loss: 0.018023

Test set: Average loss: 0.0344, Accuracy: 9875/10000 (99%)

Train Epoch: 6 [14848/60000 (25%)]	Loss: 0.024212
Train Epoch: 6 [30208/60000 (50%)]	Loss: 0.018689
Train Epoch: 6 [45568/60000 (75%)]	Loss: 0.040412

Test set: Average loss: 0.0350, Accuracy: 9879/10000 (99%)

Train Epoch: 7 [14848/60000 (25%)]	Loss: 0.030426
Train Epoch: 7 [30208/60000 (50%)]	Loss: 0.026939
Train Epoch: 7 [45568/60000 (75%)]	Loss: 0.010722

Test set: Average loss: 0.0287, Accuracy: 9892/10000 (99%)

Train Epoch: 8 [14848/60000 (25%)]	Loss: 0.021109
Train Epoch: 8 [30208/60000 (50%)]	Loss: 0.034845
Train Epoch: 8 [45568/60000 (75%)]	Loss: 0.011223

Test set: Average loss: 0.0299, Accuracy: 9904/10000 (99%)

Train Epoch: 9 [14848/60000 (25%)]	Loss: 0.011391
Train Epoch: 9 [30208/60000 (50%)]	Loss: 0.008091
Train Epoch: 9 [45568/60000 (75%)]	Loss: 0.039870

Test set: Average loss: 0.0341, Accuracy: 9890/10000 (99%)

Train Epoch: 10 [14848/60000 (25%)]	Loss: 0.026813
Train Epoch: 10 [30208/60000 (50%)]	Loss: 0.011159
Train Epoch: 10 [45568/60000 (75%)]	Loss: 0.024884

Test set: Average loss: 0.0286, Accuracy: 9901/10000 (99%)

Train Epoch: 11 [14848/60000 (25%)]	Loss: 0.006420
Train Epoch: 11 [30208/60000 (50%)]	Loss: 0.003641
Train Epoch: 11 [45568/60000 (75%)]	Loss: 0.003402

Test set: Average loss: 0.0377, Accuracy: 9894/10000 (99%)

Train Epoch: 12 [14848/60000 (25%)]	Loss: 0.006866
Train Epoch: 12 [30208/60000 (50%)]	Loss: 0.012617
Train Epoch: 12 [45568/60000 (75%)]	Loss: 0.008548

Test set: Average loss: 0.0311, Accuracy: 9908/10000 (99%)

Train Epoch: 13 [14848/60000 (25%)]	Loss: 0.010539
Train Epoch: 13 [30208/60000 (50%)]	Loss: 0.002952
Train Epoch: 13 [45568/60000 (75%)]	Loss: 0.002313

Test set: Average loss: 0.0293, Accuracy: 9905/10000 (99%)

Train Epoch: 14 [14848/60000 (25%)]	Loss: 0.002100
Train Epoch: 14 [30208/60000 (50%)]	Loss: 0.000779
Train Epoch: 14 [45568/60000 (75%)]	Loss: 0.005952

Test set: Average loss: 0.0335, Accuracy: 9897/10000 (99%)

Train Epoch: 15 [14848/60000 (25%)]	Loss: 0.006053
Train Epoch: 15 [30208/60000 (50%)]	Loss: 0.002559
Train Epoch: 15 [45568/60000 (75%)]	Loss: 0.002555

Test set: Average loss: 0.0357, Accuracy: 9894/10000 (99%)

Train Epoch: 16 [14848/60000 (25%)]	Loss: 0.000895
Train Epoch: 16 [30208/60000 (50%)]	Loss: 0.004923
Train Epoch: 16 [45568/60000 (75%)]	Loss: 0.002339

Test set: Average loss: 0.0400, Accuracy: 9893/10000 (99%)

Train Epoch: 17 [14848/60000 (25%)]	Loss: 0.004136
Train Epoch: 17 [30208/60000 (50%)]	Loss: 0.000927
Train Epoch: 17 [45568/60000 (75%)]	Loss: 0.002084

Test set: Average loss: 0.0353, Accuracy: 9895/10000 (99%)

Train Epoch: 18 [14848/60000 (25%)]	Loss: 0.004508
Train Epoch: 18 [30208/60000 (50%)]	Loss: 0.001272
Train Epoch: 18 [45568/60000 (75%)]	Loss: 0.000543

Test set: Average loss: 0.0380, Accuracy: 9894/10000 (99%)

Train Epoch: 19 [14848/60000 (25%)]	Loss: 0.001699
Train Epoch: 19 [30208/60000 (50%)]	Loss: 0.000661
Train Epoch: 19 [45568/60000 (75%)]	Loss: 0.000275

Test set: Average loss: 0.0339, Accuracy: 9905/10000 (99%)

Train Epoch: 20 [14848/60000 (25%)]	Loss: 0.000441
Train Epoch: 20 [30208/60000 (50%)]	Loss: 0.000695
Train Epoch: 20 [45568/60000 (75%)]	Loss: 0.000467

Test set: Average loss: 0.0396, Accuracy: 9894/10000 (99%)

总结

一个实际项目的工作流程：找到数据集，对数据做预处理，定义我们的模型，调整超参数，测试训练，再通过训练结果对超参数进行调整或者对模型进行调整。