### 手写数字识别项目的完整实现 以下是基于 PyTorch 实现的一个完整的 Python 项目代码，用于手写数字识别（0-9）。该代码涵盖了自定义数据集创建、模型训练与测试、结果可视化以及手动输入数字的实时预测功能。 #### 自定义数据集生成 为了模拟 MNIST 数据集的效果，可以通过 `matplotlib` 绘制简单的图形并保存为图像文件作为自定义数据集的一部分。以下是一个简单示例： ```python import numpy as np from PIL import Image, ImageDraw import os def generate_custom_dataset(output_dir="custom_digits", num_samples=100): if not os.path.exists(output_dir): os.makedirs(output_dir) for digit in range(10): # Generate digits from 0 to 9 digit_dir = os.path.join(output_dir, str(digit)) if not os.path.exists(digit_dir): os.makedirs(digit_dir) for i in range(num_samples): img = Image.new('L', (28, 28), color='black') draw = ImageDraw.Draw(img) x_center, y_center = np.random.randint(5, 23, size=(2,)) radius = np.random.randint(4, 6) draw.ellipse((x_center - radius, y_center - radius, x_center + radius, y_center + radius), fill='white') file_path = os.path.join(digit_dir, f"{digit}_{i}.png") img.save(file_path) generate_custom_dataset() ``` 此部分代码会生成一个名为 `custom_digits` 的目录结构，其中包含从 0 到 9 的子文件夹，每个子文件夹存储对应的手写数字图片[^1]。 --- #### 加载数据集 通过 PyTorch 提供的 `ImageFolder` 功能加载自定义数据集，并将其转换为张量形式以便于后续操作。 ```python from torchvision import transforms, datasets from torch.utils.data import DataLoader transform = transforms.Compose([ transforms.Grayscale(), transforms.Resize((28, 28)), transforms.ToTensor(), ]) train_data = datasets.ImageFolder(root="./custom_digits/train", transform=transform) test_data = datasets.ImageFolder(root="./custom_digits/test", transform=transform) train_loader = DataLoader(train_data, batch_size=64, shuffle=True) test_loader = DataLoader(test_data, batch_size=64, shuffle=False) ``` 此处假设已将数据分为训练集和测试集两个子文件夹分别命名为 `./custom_digits/train` 和 `./custom_digits/test`[^1]。 --- #### 定义卷积神经网络模型 采用 CNN 结构构建分类器以提高准确性。 ```python import torch.nn as nn import torch.optim as optim class ConvNet(nn.Module): def __init__(self): super(ConvNet, self).__init__() self.conv_layer = nn.Sequential( nn.Conv2d(in_channels=1, out_channels=32, kernel_size=3, padding=1), nn.ReLU(), nn.MaxPool2d(kernel_size=2), nn.Conv2d(in_channels=32, out_channels=64, kernel_size=3, padding=1), nn.ReLU(), nn.MaxPool2d(kernel_size=2) ) self.fc_layer = nn.Sequential( nn.Linear(7 * 7 * 64, 128), nn.ReLU(), nn.Dropout(p=0.5), nn.Linear(128, 10) ) def forward(self, x): x = self.conv_layer(x) x = x.view(-1, 7 * 7 * 64) x = self.fc_layer(x) return x model = ConvNet().cuda() # Move the model to GPU if available. criterion = nn.CrossEntropyLoss() optimizer = optim.Adam(model.parameters(), lr=0.001) ``` 以上代码片段展示了如何设计适合本任务需求的小型卷积神经网络架构[^1]。 --- #### 模型训练过程 利用之前准备好的数据加载器完成整个训练流程。 ```python num_epochs = 10 for epoch in range(num_epochs): model.train() # Enable training mode which affects layers like BatchNorm and Dropout [^3]. running_loss = 0.0 correct_predictions = 0 for images, labels in train_loader: images, labels = images.cuda(), labels.cuda() optimizer.zero_grad() outputs = model(images) loss = criterion(outputs, labels) loss.backward() optimizer.step() _, predicted = torch.max(outputs.data, 1) correct_predictions += (predicted == labels).sum().item() running_loss += loss.item() accuracy = 100 * correct_predictions / len(train_data) print(f'Epoch [{epoch+1}/{num_epochs}], Loss: {running_loss:.4f}, Accuracy: {accuracy:.2f}%') ``` 在此阶段激活了批量标准化(Batch Normalization)与丢弃法(Dropout)，从而提升泛化能力[^3]。 --- #### 测试评估及随机样本展示 对未见过的数据执行推理，并挑选若干实例进行对比分析。 ```python import matplotlib.pyplot as plt def evaluate_model(): model.eval() # Switches off some specific layers/modes present only during training phase [^3]. with torch.no_grad(): total_correct = 0 total_images = 0 fig, axes = plt.subplots(nrows=2, ncols=5, figsize=(12, 6)) for idx, (images, labels) in enumerate(test_loader): images, labels = images.cuda(), labels.cuda() outputs = model(images) _, predictions = torch.max(outputs, dim=1) total_correct += (predictions == labels).sum().item() total_images += labels.size(0) if idx < 2*5: ax = axes[idx//5][idx%5] image_np = images[0].cpu().numpy()[0]*255 prediction_label = predictions[0].cpu().item() true_label = labels[0].cpu().item() ax.imshow(image_np, cmap='gray') ax.set_title(f'True:{true_label}\nPred:{prediction_label}') ax.axis('off') test_accuracy = 100 * total_correct / total_images print(f'\nTest Set Evaluation:\nAccuracy={test_accuracy:.2f}%\n') plt.show() evaluate_model() ``` 上述脚本实现了性能度量的同时还提供了直观的结果呈现方式[^1]。 --- #### 手动绘制数字并实时预测 最后一步允许用户直接在窗口内画出手写字母或符号交给算法判断其类别归属情况。 ```python from tkinter import * import cv2 root = Tk() canvas_width = canvas_height = 280 def paint(event): python_green = "#ffffff" x1, y1 = event.x - brush_size/2, event.y - brush_size/2 x2, y2 = event.x + brush_size/2, event.y + brush_size/2 w.create_oval(x1, y1, x2, y2, fill=python_green) brush_size = 15 w = Canvas(root, width=canvas_width, height=canvas_height, bg='black') w.pack(expand=YES, fill=BOTH) w.bind("<B1-Motion>", paint) def predict_digit(): filename = 'temp.png' ps = w.postscript(colormode='color') img = Image.open(io.BytesIO(ps.encode('utf-8'))) img_resized = img.resize((28, 28)).convert('L') array = np.array(img_resized)/255. tensor_input = torch.tensor(array.reshape(1, 1, 28, 28)).float().cuda() output = model(tensor_input) pred_class = int(torch.argmax(output, axis=-1)[0]) print(f"\nThe drawn number is most likely a '{pred_class}'.") Button(text="Predict!", command=predict_digit).pack(side=RIGHT) mainloop() ``` 这一模块借助 GUI 工具包让交互变得更加友好便捷。 ---