Pytorch實現MNIST手寫數字識別

Pytorch是熱門的深度學習框架之一，通過經典的MNIST 數據集進行快速的pytorch入門。

導入庫

from torchvision.datasets import MNIST
from torchvision.transforms import ToTensor, Compose, Normalize
from torch.utils.data import DataLoader
import torch
import torch.nn.functional as F
import torch.nn as nn
import os
import numpy as np

准備數據集

path = './data'

# 使用Compose 將tensor化和正則化操作打包
transform_fn = Compose([
    ToTensor(),
    Normalize(mean=(0.1307,), std=(0.3081,))
])
mnist_dataset = MNIST(root=path, train=True, transform=transform_fn)

data_loader = torch.utils.data.DataLoader(dataset=mnist_dataset, batch_size=2, shuffle=True)

# 1. 構建函數,數據集預處理
BATCH_SIZE = 128
TEST_BATCH_SIZE = 1000
def get_dataloader(train=True, batch_size=BATCH_SIZE):
    '''
    train=True, 獲取訓練集
    train=False 獲取測試集
    '''
    transform_fn = Compose([
        ToTensor(),
        Normalize(mean=(0.1307,), std=(0.3081,))
    ])
    dataset = MNIST(root='./data', train=train, transform=transform_fn)
    data_loader = DataLoader(dataset=dataset, batch_size=BATCH_SIZE, shuffle=True)
    return data_loader

構建模型

class MnistModel(nn.Module):
    def __init__(self):
        super().__init__()  # 繼承父類
        self.fc1 = nn.Linear(1*28*28, 28)  # 添加全連接層
        self.fc2 = nn.Linear(28, 10)
        
    def forward(self, input):
        x = input.view(-1, 1*28*28)
        x = self.fc1(x)
        x = F.relu(x)
        out = self.fc2(x)
        return F.log_softmax(out, dim=-1)  # log_softmax 與 nll_loss合用，計算交叉熵
        

模型訓練

mnist_model = MnistModel()
optimizer = torch.optim.Adam(params=mnist_model.parameters(), lr=0.001)

# 如果有模型則加載
if os.path.exists('./model'):
    mnist_model.load_state_dict(torch.load('model/mnist_model.pkl'))
    optimizer.load_state_dict(torch.load('model/optimizer.pkl'))

def train(epoch):
    data_loader = get_dataloader()
    
    for index, (data, target) in enumerate(data_loader):
        optimizer.zero_grad()  # 梯度先清零
        output = mnist_model(data)
        loss = F.nll_loss(output, target)
        loss.backward()  # 誤差反向傳播計算
        optimizer.step()  # 更新梯度
        
        if index % 100 == 0:
            # 保存訓練模型
            torch.save(mnist_model.state_dict(), 'model/mnist_model.pkl')
            torch.save(optimizer.state_dict(), 'model/optimizer.pkl')
            print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
                epoch, index * len(data), len(data_loader.dataset),
                       100. * index / len(data_loader), loss.item()))

for i in range(epoch=5):
    train(i)

Train Epoch: 0 [0/60000 (0%)]	Loss: 0.023078
Train Epoch: 0 [12800/60000 (21%)]	Loss: 0.019347
Train Epoch: 0 [25600/60000 (43%)]	Loss: 0.105870
Train Epoch: 0 [38400/60000 (64%)]	Loss: 0.050866
Train Epoch: 0 [51200/60000 (85%)]	Loss: 0.097995
Train Epoch: 1 [0/60000 (0%)]	Loss: 0.108337
Train Epoch: 1 [12800/60000 (21%)]	Loss: 0.071196
Train Epoch: 1 [25600/60000 (43%)]	Loss: 0.022856
Train Epoch: 1 [38400/60000 (64%)]	Loss: 0.028392
Train Epoch: 1 [51200/60000 (85%)]	Loss: 0.070508
Train Epoch: 2 [0/60000 (0%)]	Loss: 0.037416
Train Epoch: 2 [12800/60000 (21%)]	Loss: 0.075977
Train Epoch: 2 [25600/60000 (43%)]	Loss: 0.024356
Train Epoch: 2 [38400/60000 (64%)]	Loss: 0.042203
Train Epoch: 2 [51200/60000 (85%)]	Loss: 0.020883
Train Epoch: 3 [0/60000 (0%)]	Loss: 0.023487
Train Epoch: 3 [12800/60000 (21%)]	Loss: 0.024403
Train Epoch: 3 [25600/60000 (43%)]	Loss: 0.073619
Train Epoch: 3 [38400/60000 (64%)]	Loss: 0.074042
Train Epoch: 3 [51200/60000 (85%)]	Loss: 0.036283
Train Epoch: 4 [0/60000 (0%)]	Loss: 0.021305
Train Epoch: 4 [12800/60000 (21%)]	Loss: 0.062750
Train Epoch: 4 [25600/60000 (43%)]	Loss: 0.016911
Train Epoch: 4 [38400/60000 (64%)]	Loss: 0.039599
Train Epoch: 4 [51200/60000 (85%)]	Loss: 0.026689

模型測試

def test():
    loss_list = []
    acc_list = []
    
    test_loader = get_dataloader(train=False, batch_size = TEST_BATCH_SIZE)
    mnist_model.eval()  # 設為評估模式
    
    for index, (data, target) in enumerate(test_loader):
        with torch.no_grad():
            out = mnist_model(data)
            loss = F.nll_loss(out, target)
            loss_list.append(loss)
            
            pred = out.data.max(1)[1]
            acc = pred.eq(target).float().mean()  # eq()函數用於將兩個tensor中的元素對比，返回布爾值
            acc_list.append(acc)
           
        
    print('平均准確率, 平均損失', np.mean(acc_list), np.mean(loss_list))

test()

平均准確率, 平均損失 0.9662777 0.12309619