1.混淆矩阵（Confusion Matrix）

1.1基础理论

（1）在机器学习、深度学习领域中，混淆矩阵常用于监督学习，匹配矩阵常用于无监督学习。主要用来比较分类结果和实际预测值。

（2）图中表达的含义：混淆矩阵的每一列代表了预测类别，每一行代表了数据的真实类别。

1.2 实现代码

import torch
import os
from PIL import Image
import numpy as np
import matplotlib.pyplot as plt
from sklearn.metrics import confusion_matrix
import seaborn as sns
from torchvision import transforms

classes = ['bai_xing_hua_jin_gui', 'beetle', 'chui_mian_jie', 'ci_e_ke', 'da_qing_ye_chan','dou_yuan_jing','fan_qie_qian_ye_ying_larva','fan_qie_qian_ye_ying_mature','hong_zhi_zhu','huang_zong_ke']

# classes = ['白星化金龟', '甲虫', '吹绵蚧', '刺蛾科', '大青叶蝉','豆芫菁','番茄潜叶蛾幼虫','番茄潜叶蛾成虫','红蜘蛛','蝗总科']

def predict_image(model, image_path, true_label):
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
    img = Image.open(image_path)
    val_transform = transforms.Compose([
        transforms.Resize((224, 224)),
        transforms.ToTensor(),
        transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
    ])
    tensor_img = val_transform(img)
    tensor_img = tensor_img.to(device)
    tensor_img = tensor_img.unsqueeze(0)
    output = model(tensor_img)
    _, pred = output.max(1)
    pred_label = classes[pred.item()]
    return pred_label, true_label

if __name__ == '__main__':
    # 1. 加载模型
    model_path = r"/kaggle/input/mymodel3/resnet101_final.pth"
    model = torch.load(model_path)
    model.eval()

    # 2. 预测多张图片并记录真实标签和预测结果
    true_labels = []
    pred_labels = []
    images_dir = r"/kaggle/input/insects-new/validation"
    for label in os.listdir(images_dir):
        label_dir = os.path.join(images_dir, label)
        if not os.path.isdir(label_dir):
            continue
        for img_name in os.listdir(label_dir):
            img_path = os.path.join(label_dir, img_name)
            true_labels.append(label)
            pred_label, _ = predict_image(model, img_path, label)
            pred_labels.append(pred_label)

    # 3. 计算混淆矩阵
    cm = confusion_matrix(true_labels, pred_labels, labels=classes)

    # 4. 计算归一化的混淆矩阵
    cm_normalized = cm.astype('float') / cm.sum(axis=1)[:, np.newaxis]
    
    # 5. 绘制混淆矩阵
    save_path = "/kaggle/working/confusion_matrix.png"
    plt.figure(figsize=(8, 6))
    sns.heatmap(cm_normalized, annot=True, cmap='Blues', xticklabels=classes, yticklabels=classes, fmt='.2f')
    plt.xlabel('预测标签')
    plt.ylabel('真实标签')
    plt.tight_layout()  # 自动调整子图参数
    plt.savefig(save_path)
    plt.show()

注意：以下数值需要和训练时的数值一样！

---

2.热力图

import matplotlib.pyplot as plt
import numpy as np
from PIL import Image
import torch
from torchcam.methods import GradCAMpp
# CAM GradCAM GradCAMpp ISCAM LayerCAM SSCAM ScoreCAM SmoothGradCAMpp XGradCAM
from torchvision import transforms
from torchcam.utils import overlay_mask


# 有 GPU 就用 GPU，没有就用 CPU
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
print('device', device)


model = torch.load('/kaggle/input/mymodel3/resnet101_final.pth')
model = model.eval().to(device)

cam_extractor = GradCAMpp(model)

# 要与训练集保持一致
test_transform = transforms.Compose([transforms.RandomResizedCrop(224),
                             transforms.RandomHorizontalFlip(),
                             transforms.RandomVerticalFlip(),
                             transforms.RandomGrayscale(),
                             transforms.ToTensor(),
                             transforms.RandomErasing(),
                             transforms.Normalize([0.460, 0.483, 0.396], [0.171, 0.150, 0.190])])

# 载入目标图像
img_path = '/kaggle/input/insects-new/train/hong_zhi_zhu/13845.jpg'
img_pil = Image.open(img_path)
input_tensor = test_transform(img_pil).unsqueeze(0).to(device) # 预处理
# 预测标签
pred_logits = model(input_tensor)
pred_id = torch.topk(pred_logits, 1)[1].detach().cpu().numpy().squeeze().item()

activation_map = cam_extractor(pred_id, pred_logits)
activation_map = activation_map[0][0].detach().cpu().numpy()
# 矩阵热力图
plt.imshow(activation_map)
plt.show()
plt.savefig('/kaggle/working/activation_map.png')

# 将原图重合
result = overlay_mask(img_pil, Image.fromarray(activation_map), alpha=0.7)
result.save('/kaggle/working/result.png')

result