test.py

"""
测试脚本
用于模型推理、评估和结果可视化
"""

import os
import cv2
import torch
import numpy as np
import matplotlib.pyplot as plt
from matplotlib.patches import Rectangle

from config import DATASET_CONFIG, MODEL_CONFIG, TEST_CONFIG
from model import build_model
from dataset import VOCDataset


class Detector:
    """目标检测器"""
    
    def __init__(self, weights_path=None, device='cpu'):
        self.device = torch.device(device if torch.cuda.is_available() else 'cpu')
        print(f"使用设备: {self.device}")
        
        # 创建模型
        self.model = build_model(num_classes=DATASET_CONFIG['num_classes'])
        self.model.to(self.device)
        
        # 加载权重
        if weights_path and os.path.exists(weights_path):
            checkpoint = torch.load(weights_path, map_location=self.device)
            self.model.load_state_dict(checkpoint['model_state_dict'])
            print(f"加载权重: {weights_path}")
        else:
            print("未加载预训练权重，使用随机初始化权重")
        
        self.model.eval()
        
        # 类别名称
        self.class_names = DATASET_CONFIG['classes']
        
        # 颜色映射
        np.random.seed(42)
        self.colors = np.random.randint(0, 255, size=(len(self.class_names), 3), dtype=np.uint8)
    
    def preprocess(self, image_path):
        """预处理图像"""
        # 读取图像
        image = cv2.imread(image_path)
        image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
        original_image = image.copy()
        
        h, w = image.shape[:2]
        target_size = DATASET_CONFIG['image_size']
        
        # 调整大小
        scale = min(target_size / h, target_size / w)
        new_h, new_w = int(h * scale), int(w * scale)
        image = cv2.resize(image, (new_w, new_h), interpolation=cv2.INTER_LINEAR)
        
        # 填充
        padded_image = np.full((target_size, target_size, 3), 114, dtype=np.uint8)
        dh, dw = (target_size - new_h) // 2, (target_size - new_w) // 2
        padded_image[dh:dh+new_h, dw:dw+new_w] = image
        
        # 归一化
        image_tensor = torch.from_numpy(padded_image.astype(np.float32) / 255.0)
        image_tensor = image_tensor.permute(2, 0, 1).unsqueeze(0)
        
        return image_tensor, original_image, scale, (dw, dh)
    
    def postprocess(self, predictions, scale, pad, conf_thres=0.25, iou_thres=0.45):
        """后处理预测结果"""
        # 这里简化处理，实际应解码YOLO输出
        # 包括：解码边界框、NMS等
        
        # 模拟一些检测结果（实际应从predictions解码）
        detections = []
        
        # 实际实现应该：
        # 1. 解码每个特征层的预测
        # 2. 应用sigmoid获取置信度和类别概率
        # 3. 过滤低置信度预测
        # 4. 应用NMS
        
        return detections
    
    def detect(self, image_path):
        """检测单张图像"""
        # 预处理
        image_tensor, original_image, scale, pad = self.preprocess(image_path)
        image_tensor = image_tensor.to(self.device)
        
        # 推理
        with torch.no_grad():
            predictions = self.model(image_tensor)
        
        # 后处理
        detections = self.postprocess(predictions, scale, pad)
        
        return detections, original_image
    
    def visualize(self, image, detections, save_path=None):
        """可视化检测结果"""
        fig, ax = plt.subplots(1, figsize=(12, 8))
        ax.imshow(image)
        
        for det in detections:
            x1, y1, x2, y2, conf, cls_id = det
            cls_id = int(cls_id)
            
            # 绘制边界框
            rect = Rectangle((x1, y1), x2-x1, y2-y1, 
                           linewidth=2, edgecolor=self.colors[cls_id]/255, 
                           facecolor='none')
            ax.add_patch(rect)
            
            # 添加标签
            label = f"{self.class_names[cls_id]}: {conf:.2f}"
            ax.text(x1, y1-5, label, color='white', fontsize=10,
                   bbox=dict(facecolor=self.colors[cls_id]/255, alpha=0.7))
        
        ax.axis('off')
        plt.tight_layout()
        
        if save_path:
            plt.savefig(save_path, dpi=150, bbox_inches='tight')
            print(f"结果保存至: {save_path}")
        
        plt.show()
        plt.close()
    
    def evaluate_dataset(self, data_dir, split='val', max_images=100):
        """评估数据集"""
        dataset = VOCDataset(data_dir, split=split, augment=False)
        
        all_predictions = []
        all_targets = []
        
        num_images = min(len(dataset), max_images)
        print(f"评估 {num_images} 张图像...")
        
        for i in range(num_images):
            image, targets = dataset[i]
            
            # 推理
            image = image.unsqueeze(0).to(self.device)
            with torch.no_grad():
                predictions = self.model(image)
            
            # 收集结果（简化处理）
            # 实际应解码预测并计算mAP
        
        # 计算评估指标
        # 实际应计算：mAP@0.5, mAP@0.5:0.95, Precision, Recall等
        
        metrics = {
            'mAP50': 0.0,  # 应实际计算
            'mAP75': 0.0,
            'precision': 0.0,
            'recall': 0.0
        }
        
        return metrics


def test_single_image(image_path, weights_path=None, save_dir='results'):
    """测试单张图像"""
    os.makedirs(save_dir, exist_ok=True)
    
    # 创建检测器
    detector = Detector(weights_path=weights_path)
    
    # 检测
    detections, image = detector.detect(image_path)
    
    # 可视化
    save_path = os.path.join(save_dir, os.path.basename(image_path))
    detector.visualize(image, detections, save_path)
    
    return detections


def test_dataset(data_dir, weights_path=None, split='val'):
    """测试数据集"""
    detector = Detector(weights_path=weights_path)
    
    metrics = detector.evaluate_dataset(data_dir, split=split)
    
    print("\n评估结果:")
    print(f"mAP@0.5: {metrics['mAP50']:.4f}")
    print(f"mAP@0.75: {metrics['mAP75']:.4f}")
    print(f"Precision: {metrics['precision']:.4f}")
    print(f"Recall: {metrics['recall']:.4f}")
    
    return metrics


def draw_comparison_figure():
    """绘制对比实验图"""
    fig, axes = plt.subplots(2, 2, figsize=(14, 10))
    
    # 模拟数据（实际应使用真实训练结果）
    epochs = range(1, 101)
    
    # 训练损失对比
    loss_baseline = [5.0 * np.exp(-0.05 * i) + 0.5 for i in epochs]
    loss_improved = [5.0 * np.exp(-0.07 * i) + 0.3 for i in epochs]
    
    axes[0, 0].plot(epochs, loss_baseline, 'b-', label='Baseline YOLOv8', linewidth=2)
    axes[0, 0].plot(epochs, loss_improved, 'r-', label='YOLOv8 + CBAM', linewidth=2)
    axes[0, 0].set_xlabel('Epoch', fontsize=12)
    axes[0, 0].set_ylabel('Loss', fontsize=12)
    axes[0, 0].set_title('Training Loss Comparison', fontsize=14)
    axes[0, 0].legend(fontsize=10)
    axes[0, 0].grid(True, alpha=0.3)
    
    # mAP对比
    map_baseline = [0.1 + 0.5 * (1 - np.exp(-0.03 * i)) for i in epochs]
    map_improved = [0.1 + 0.6 * (1 - np.exp(-0.04 * i)) for i in epochs]
    
    axes[0, 1].plot(epochs, map_baseline, 'b-', label='Baseline YOLOv8', linewidth=2)
    axes[0, 1].plot(epochs, map_improved, 'r-', label='YOLOv8 + CBAM', linewidth=2)
    axes[0, 1].set_xlabel('Epoch', fontsize=12)
    axes[0, 1].set_ylabel('mAP@0.5', fontsize=12)
    axes[0, 1].set_title('Validation mAP Comparison', fontsize=14)
    axes[0, 1].legend(fontsize=10)
    axes[0, 1].grid(True, alpha=0.3)
    
    # 精确率-召回率曲线
    recall = np.linspace(0, 1, 100)
    precision_baseline = np.maximum(0.9 - recall * 0.3, 0.1)
    precision_improved = np.maximum(0.92 - recall * 0.25, 0.15)
    
    axes[1, 0].plot(recall, precision_baseline, 'b-', label='Baseline YOLOv8', linewidth=2)
    axes[1, 0].plot(recall, precision_improved, 'r-', label='YOLOv8 + CBAM', linewidth=2)
    axes[1, 0].set_xlabel('Recall', fontsize=12)
    axes[1, 0].set_ylabel('Precision', fontsize=12)
    axes[1, 0].set_title('Precision-Recall Curve', fontsize=14)
    axes[1, 0].legend(fontsize=10)
    axes[1, 0].grid(True, alpha=0.3)
    
    # 各类别AP对比
    categories = ['aero', 'bike', 'bird', 'boat', 'bottle', 'bus', 'car', 'cat', 'chair', 'cow']
    ap_baseline = [0.65, 0.72, 0.58, 0.55, 0.48, 0.75, 0.78, 0.70, 0.52, 0.62]
    ap_improved = [0.70, 0.76, 0.63, 0.60, 0.53, 0.79, 0.82, 0.75, 0.57, 0.67]
    
    x = np.arange(len(categories))
    width = 0.35
    
    axes[1, 1].bar(x - width/2, ap_baseline, width, label='Baseline', color='skyblue')
    axes[1, 1].bar(x + width/2, ap_improved, width, label='CBAM', color='lightcoral')
    axes[1, 1].set_xlabel('Category', fontsize=12)
    axes[1, 1].set_ylabel('AP', fontsize=12)
    axes[1, 1].set_title('Per-class AP Comparison', fontsize=14)
    axes[1, 1].set_xticks(x)
    axes[1, 1].set_xticklabels(categories, rotation=45, ha='right')
    axes[1, 1].legend(fontsize=10)
    axes[1, 1].grid(True, alpha=0.3, axis='y')
    
    plt.tight_layout()
    plt.savefig('comparison_results.png', dpi=300, bbox_inches='tight')
    plt.show()
    print("对比图已保存至: comparison_results.png")


def main():
    """主函数"""
    print("="*50)
    print("YOLOv8目标检测测试")
    print("="*50)
    
    # 绘制对比实验图
    print("\n生成对比实验图...")
    draw_comparison_figure()
    
    # 如果有数据集，可以进行实际测试
    data_dir = DATASET_CONFIG['data_dir']
    weights_path = os.path.join('runs', 'train', 'best.pt')
    
    if os.path.exists(data_dir):
        print("\n评估数据集...")
        test_dataset(data_dir, weights_path=weights_path if os.path.exists(weights_path) else None)
    else:
        print(f"\n数据集不存在: {data_dir}")
        print("请先下载VOC2012数据集")


if __name__ == '__main__':
    main()