完成作业3.3.2

digit_mlp_class/visualize.py

@@ -0,0 +1,350 @@
+# -*- coding: utf-8 -*-
+"""
+可视化工具 - 展示神经网络各层的输出
+
+用于课堂教学,让学生直观理解:
+1. 输入图像长什么样
+2. 第一层隐藏层学到了什么特征
+3. 各层激活值的变化
+
+使用方法:
+    python visualize.py              # 可视化测试集前5张
+    python visualize.py --single    # 可视化单张图片
+"""
+
+import numpy as np
+from PIL import Image
+import os
+import sys
+import matplotlib
+matplotlib.use('Agg')  # 无头模式,不显示图形
+import matplotlib.pyplot as plt
+
+
+def visualize_input_image(img_vector, save_path='visualizations/input.png'):
+    """把784维向量还原成28x28图像并保存"""
+    img = img_vector.reshape(28, 28) * 255
+    img = img.astype(np.uint8)
+    Image.fromarray(img).save(save_path)
+    return save_path
+
+
+def visualize_activations(model, img_vector, save_dir='visualizations'):
+    """
+    可视化网络各层的激活值
+    """
+    os.makedirs(save_dir, exist_ok=True)
+
+    # 前向传播获取各层激活值
+    model.forward(img_vector.reshape(1, -1))
+
+    # 1. 保存输入图像
+    visualize_input_image(img_vector, os.path.join(save_dir, '01_input.png'))
+
+    # 2. 可视化第一层激活(隐藏层)
+    hidden_activations = model.a1[0]  # (128,)
+    visualize_hidden_layer(hidden_activations, os.path.join(save_dir, '02_hidden.png'))
+
+    # 3. 可视化输出层概率
+    output_probs = model.probs[0]  # (10,)
+    visualize_output_prob(output_probs, os.path.join(save_dir, '03_output_prob.png'))
+
+    # 4. 生成汇总图
+    create_summary_image(img_vector, hidden_activations, output_probs, save_dir)
+
+    return save_dir
+
+
+def visualize_hidden_layer(activations, save_path):
+    """
+    可视化隐藏层激活值
+    把128个神经元的激活值排成8x16网格显示
+    """
+    grid_cols = 16
+    grid_rows = 8
+    cell_size = 24
+
+    img_h = grid_rows * cell_size
+    img_w = grid_cols * cell_size
+    grid = np.ones((img_h, img_w)) * 255
+
+    for i, act in enumerate(activations):
+        row = i // grid_cols
+        col = i % grid_cols
+        intensity = max(0, min(1, act * 2))
+        color = int(255 * (1 - intensity * 0.7))
+        grid[row*cell_size:(row+1)*cell_size-1, col*cell_size:(col+1)*cell_size-1] = color
+
+    Image.fromarray(grid.astype(np.uint8)).save(save_path)
+
+
+def visualize_output_prob(probs, save_path):
+    """可视化输出层概率分布"""
+    fig, ax = plt.subplots(figsize=(8, 4))
+
+    digits = list(range(10))
+    colors = ['#3498db' if i != np.argmax(probs) else '#e74c3c' for i in digits]
+
+    bars = ax.bar(digits, probs, color=colors)
+    ax.set_xlabel('数字', fontsize=12)
+    ax.set_ylabel('概率', fontsize=12)
+    ax.set_title('输出层:各数字的预测概率', fontsize=14)
+    ax.set_xticks(digits)
+    ax.set_ylim(0, 1)
+
+    max_idx = np.argmax(probs)
+    ax.annotate(f'{probs[max_idx]:.1%}',
+                xy=(max_idx, probs[max_idx]),
+                ha='center', va='bottom', fontsize=10, color='#e74c3c', fontweight='bold')
+
+    plt.tight_layout()
+    plt.savefig(save_path, dpi=100, bbox_inches='tight')
+    plt.close()
+
+
+def create_summary_image(img_vector, hidden_activations, output_probs, save_dir):
+    """创建汇总图"""
+    fig = plt.figure(figsize=(14, 6))
+
+    # 1. 输入图像
+    ax1 = fig.add_subplot(2, 4, 1)
+    ax1.imshow(img_vector.reshape(28, 28), cmap='gray')
+    ax1.set_title('(1) Input Image\n(28x28 pixels)', fontsize=11)
+    ax1.axis('off')
+
+    # 2. 像素值分布
+    ax2 = fig.add_subplot(2, 4, 2)
+    ax2.hist(img_vector, bins=30, color='#3498db', alpha=0.7, edgecolor='white')
+    ax2.set_title('(2) Pixel Value Distribution\n(normalized 0~1)', fontsize=11)
+    ax2.set_xlabel('像素值')
+    ax2.set_ylabel('频数')
+
+    # 3. 隐藏层激活(热力图)
+    ax3 = fig.add_subplot(2, 4, 3)
+    # 128 = 8 × 16
+    act_2d = hidden_activations.reshape(8, 16)
+    im = ax3.imshow(act_2d, cmap='Blues', aspect='auto')
+    ax3.set_title(f'(3) Hidden Layer\n(128 neurons)', fontsize=11)
+    ax3.axis('off')
+    plt.colorbar(im, ax=ax3, shrink=0.6)
+
+    # 4. 隐藏层激活(条形图)
+    ax4 = fig.add_subplot(2, 4, 4)
+    ax4.bar(range(len(hidden_activations)), hidden_activations, color='#3498db', alpha=0.7)
+    ax4.set_title('(4) Neuron Activations', fontsize=11)
+    ax4.set_xlabel('神经元编号')
+    ax4.set_ylabel('激活强度')
+
+    # 5. 输出概率
+    ax5 = fig.add_subplot(2, 4, 5)
+    digits = list(range(10))
+    colors = ['#3498db' if i != np.argmax(output_probs) else '#e74c3c' for i in digits]
+    ax5.bar(digits, output_probs, color=colors)
+    ax5.set_title('(5) Output Probabilities', fontsize=11)
+    ax5.set_xlabel('数字')
+    ax5.set_ylabel('概率')
+    ax5.set_ylim(0, 1)
+    ax5.set_xticks(digits)
+
+    # 6. 最大概率
+    ax6 = fig.add_subplot(2, 4, 6)
+    ax6.axis('off')
+    predicted = np.argmax(output_probs)
+    confidence = output_probs[predicted]
+    result_text = f'预测: {predicted}\n置信度: {confidence:.1%}'
+    ax6.text(0.5, 0.5, result_text, fontsize=24, ha='center', va='center',
+             bbox=dict(boxstyle='round', facecolor='#2ecc71', alpha=0.9),
+             transform=ax6.transAxes, color='white', fontweight='bold')
+    ax6.set_title('(6) Recognition Result', fontsize=11)
+
+    # 7. 网络结构
+    ax7 = fig.add_subplot(2, 4, 7)
+    ax7.axis('off')
+    structure_text = (
+        '┌─────────────────┐\n'
+        '│   输入层 784     │\n'
+        '│  (28×28展平)    │\n'
+        '└────────┬────────┘\n'
+        '         │\n'
+        '    线性变换+ReLU\n'
+        '         │\n'
+        '┌────────┴────────┐\n'
+        '│  隐藏层 128     │\n'
+        '│  (特征提取)    │\n'
+        '└────────┬────────┘\n'
+        '         │\n'
+        '    线性变换+Softmax\n'
+        '         │\n'
+        '┌────────┴────────┐\n'
+        '│   输出层 10      │\n'
+        '│  (数字0~9概率)  │\n'
+        '└─────────────────┘'
+    )
+    ax7.text(0.1, 0.95, structure_text, fontsize=9, va='top',
+             family='monospace', transform=ax7.transAxes,
+             bbox=dict(boxstyle='round', facecolor='#f8f9fa', alpha=0.9))
+    ax7.set_title('(7) Network Structure', fontsize=11)
+
+    # 8. 参数量说明
+    ax8 = fig.add_subplot(2, 4, 8)
+    ax8.axis('off')
+    params_text = (
+        'MLP 参数量计算:\n\n'
+        'W1: 784 × 128 = 100,352\n'
+        'b1: 128\n\n'
+        'W2: 128 × 10 = 1,280\n'
+        'b2: 10\n\n'
+        '─────────────────\n'
+        '总计: 101,770 参数\n\n'
+        '全部用 NumPy 实现\n'
+        '无需任何深度学习框架!'
+    )
+    ax8.text(0.1, 0.95, params_text, fontsize=10, va='top',
+             family='monospace', transform=ax8.transAxes)
+    ax8.set_title('(8) Parameters', fontsize=11)
+
+    plt.suptitle('MLP Feature Maps Visualization - Handwritten Digits', fontsize=16, fontweight='bold', y=1.02)
+    plt.tight_layout()
+    plt.savefig(os.path.join(save_dir, 'summary.png'), dpi=120, bbox_inches='tight')
+    plt.close()
+
+
+def load_model_or_train():
+    """加载已训练的模型,如果没有则训练一个"""
+    import glob
+    model_files = glob.glob('mnist_mlp_*.npy')
+
+    if model_files:
+        timestamps = sorted(set(
+            f.replace('mnist_mlp_', '').replace('_W1.npy', '')
+            for f in model_files if '_W1.npy' in f
+        ))
+        if timestamps:
+            model_path = 'mnist_mlp_' + timestamps[-1]
+            print(f"加载模型: {model_path}")
+
+            from model_numpy import MLP
+            model = MLP(input_size=784, hidden_size=128, num_classes=10, learning_rate=0.1)
+            model.W1 = np.load(f'{model_path}_W1.npy')
+            model.b1 = np.load(f'{model_path}_b1.npy')
+            model.W2 = np.load(f'{model_path}_W2.npy')
+            model.b2 = np.load(f'{model_path}_b2.npy')
+            return model
+
+    # 没有模型,用sklearn快速训练一个用于演示
+    print("未找到已训练模型,使用sklearn数据快速训练演示模型...")
+    from sklearn.datasets import fetch_openml
+    from sklearn.model_selection import train_test_split
+
+    mnist = fetch_openml('mnist_784', version=1, as_frame=False, parser='auto')
+    X = mnist.data[:10000].astype(np.float32) / 255.0
+    y = mnist.target[:10000].astype(int)
+
+    # 用sklearn的MLP代替
+    from sklearn.neural_network import MLPClassifier
+    model = MLPClassifier(
+        hidden_layer_sizes=(128,),
+        max_iter=20,
+        alpha=1e-4,
+        solver='sgd',
+        learning_rate_init=0.1,
+        random_state=42,
+        verbose=False
+    )
+    model.fit(X, y)
+    print("sklearn模型训练完成(仅用于可视化演示)")
+    return model
+
+
+def main():
+    """主函数"""
+    os.makedirs('visualizations', exist_ok=True)
+
+    # 加载数据
+    from dataset import load_data
+    print("加载MNIST数据集...")
+    X_train, y_train, X_test, y_test = load_data()
+
+    # 加载模型
+    model = load_model_or_train()
+
+    # 获取真实标签
+    if len(y_test.shape) > 1:
+        y_test_labels = np.argmax(y_test, axis=1)
+    else:
+        y_test_labels = y_test
+
+    # 可视化测试集前5张
+    print("\n可视化测试集前5张图片...")
+    for i in range(5):
+        img = X_test[i]
+        true_label = y_test_labels[i]
+
+        sub_dir = f'visualizations/sample_{i}_true{true_label}'
+        os.makedirs(sub_dir, exist_ok=True)
+
+        if hasattr(model, 'predict_proba'):
+            probs = model.predict_proba(img.reshape(1, -1))[0]
+            predicted = np.argmax(probs)
+        else:
+            predicted = model.predict(img.reshape(1, -1))[0]
+
+        print(f"  样本{i}: 真实={true_label}, 预测={predicted}")
+        visualize_activations(model, img, sub_dir)
+
+    # 创建对比汇总
+    create_comparison_summary(X_test[:5], y_test_labels[:5], model, 'visualizations')
+
+    print("\n✅ 可视化完成!")
+    print("   查看 visualizations/ 目录下的图片和汇总图")
+
+
+def create_comparison_summary(X_samples, y_true, model, save_dir):
+    """创建多个样本的对比汇总图"""
+    n_samples = len(X_samples)
+
+    fig = plt.figure(figsize=(4 * n_samples, 8))
+
+    for i in range(n_samples):
+        img = X_samples[i]
+        true_label = y_true[i]
+
+        if hasattr(model, 'predict_proba'):
+            probs = model.predict_proba(img.reshape(1, -1))[0]
+            predicted = np.argmax(probs)
+        else:
+            predicted = model.predict(img.reshape(1, -1))[0]
+
+        # 输入图像
+        ax = fig.add_subplot(3, n_samples, i + 1)
+        ax.imshow(img.reshape(28, 28), cmap='gray')
+        ax.set_title(f'真实: {true_label}', fontsize=12)
+        ax.axis('off')
+
+        # 隐藏层激活
+        ax = fig.add_subplot(3, n_samples, i + 1 + n_samples)
+        model.forward(img.reshape(1, -1))
+        # 128 = 8 × 16
+        hidden = model.a1[0].reshape(8, 16)
+        ax.imshow(hidden, cmap='Blues', aspect='auto')
+        ax.set_title(f'隐藏层激活', fontsize=10)
+        ax.axis('off')
+
+        # 输出概率
+        ax = fig.add_subplot(3, n_samples, i + 1 + 2*n_samples)
+        digits = list(range(10))
+        colors = ['#e74c3c' if d == predicted else '#3498db' for d in digits]
+        ax.bar(digits, probs, color=colors)
+        ax.set_title(f'预测: {predicted}', fontsize=12)
+        ax.set_ylim(0, 1)
+        ax.set_xticks(digits)
+        ax.tick_params(labelsize=8)
+
+    plt.suptitle('多样本特征图对比', fontsize=16, fontweight='bold')
+    plt.tight_layout()
+    plt.savefig(os.path.join(save_dir, 'comparison.png'), dpi=120, bbox_inches='tight')
+    plt.close()
+
+
+if __name__ == '__main__':
+    main()