SMART/test/test_hardness_methods.py

import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import torchvision
import torchvision.transforms as transforms
from utils.carlini_wagner_l2 import carlini_wagner_l2
from tqdm import tqdm
import os
import json
import time
from sklearn.calibration import calibration_curve

# 设置随机种子以便结果可重现
def set_seed(seed):
    torch.manual_seed(seed)
    torch.cuda.manual_seed(seed)
    torch.cuda.manual_seed_all(seed)
    np.random.seed(seed)
    torch.backends.cudnn.deterministic = True
    torch.backends.cudnn.benchmark = False

set_seed(42)

# 设置设备
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"Using device: {device}")

# 创建结果目录
TEST_RESULTS_DIR = "test_results"
os.makedirs(TEST_RESULTS_DIR, exist_ok=True)

# 加载CIFAR-10数据集
def load_cifar10(batch_size=1):
    transform_test = transforms.Compose([
        transforms.ToTensor(),
    ])
    testset = torchvision.datasets.CIFAR10(root='./data', train=False,
                                          download=True, transform=transform_test)
    testloader = torch.utils.data.DataLoader(testset, batch_size=batch_size,
                                          shuffle=False, num_workers=2)
    return testloader, testset

# 加载ResNet50模型
def load_model():
    model = torchvision.models.resnet50(pretrained=True)
    
    # 修改最后的全连接层以适应CIFAR-10的10个类别
    num_ftrs = model.fc.in_features
    model.fc = nn.Linear(num_ftrs, 10)
    
    # 尝试加载预训练的CIFAR-10权重
    try:
        weight_path = "/share/pretrained_weights/cifar10_resnet50_cross_entropy.model"
        if os.path.exists(weight_path):
            state_dict = torch.load(weight_path)
            model.load_state_dict(state_dict)
            print(f"Loaded pretrained weights from {weight_path}")
        else:
            print(f"Warning: Pretrained weights not found at {weight_path}")
            print("Using ImageNet pretrained weights with modified final layer")
    except Exception as e:
        print(f"Error loading pretrained weights: {e}")
    
    model = model.to(device)
    model.eval()
    return model

# 获取模型特征提取器(返回倒数第二层特征)
def get_feature_extractor(model):
    class FeatureExtractor(nn.Module):
        def __init__(self, model):
            super().__init__()
            self.model = model
            self.features = None
            
            # 注册钩子来获取特征
            def hook_fn(module, input, output):
                self.features = output
                
            # 注册到倒数第二层
            model.avgpool.register_forward_hook(hook_fn)
            
        def forward(self, x):
            logits = self.model(x)
            return self.features, logits
            
    return FeatureExtractor(model).to(device)

# 方法1: 在logits上使用CW攻击计算hardness
def compute_hardness_cw_logits(model, inputs, logits, debug=True):
    """
    通过在logits上使用CW L2攻击来计算样本的硬度
    
    Args:
        model: 模型
        inputs: 输入样本
        logits: 样本的logits
        debug: 是否打印调试信息
        
    Returns:
        hardness: 样本硬度值
        success: 攻击是否成功
    """
    model.eval()
    inputs = inputs.to(device)
    
    # 获取原始预测
    with torch.no_grad():
        if logits is None:
            outputs = model(inputs)
        else:
            outputs = logits
            
        original_pred = torch.argmax(outputs, 1)
        probs = F.softmax(outputs, dim=1)
        confidence = probs.max(dim=1)[0].item()
    
    if debug:
        print(f"Original prediction: {original_pred.item()}, confidence: {confidence:.4f}")
    
    # 创建一个简单的logits到输出的模型
    class LogitsModel(nn.Module):
        def __init__(self):
            super().__init__()
        
        def forward(self, x):
            return x
    
    logits_model = LogitsModel().to(device)
    
    # 准备函数包装器
    def model_fn(x):
        return logits_model(x)
    
    # 将logits转换为可攻击的输入格式
    logits_input = outputs.detach().clone()
    
    # 应用CW L2攻击到logits
    try:
        if debug:
            print("Running CW attack on logits with parameters:")
            print(f"- Binary search steps: 5")
            print(f"- Max iterations: 300")
            print(f"- Learning rate: 0.01")
            print(f"- Early stop: True")
        
        # 计算logits的最小值和最大值作为clip范围
        logits_min = float(logits_input.min())
        logits_max = float(logits_input.max())
        
        adv_logits, l2_norms, success_mask = carlini_wagner_l2(
            model_fn=model_fn,
            x=logits_input,
            n_classes=logits_input.shape[1],
            y=original_pred,
            targeted=False,
            lr=0.01,  # logits空间可能需要更高的学习率
            binary_search_steps=5,
            max_iterations=300,
            confidence=0,
            clip_min=logits_min,
            clip_max=logits_max,
            debug=debug,
            early_stop=True
        )
        
        # 获取对抗logits的预测
        adv_pred = torch.argmax(adv_logits, 1)
        adv_probs = F.softmax(adv_logits, dim=1)
        adv_confidence = adv_probs.max(dim=1)[0].item()
        
        # 攻击成功与否直接从success_mask获取
        attack_success = success_mask.any().item()
        
        # 获取L2范数
        l2_val = l2_norms[0].item() if l2_norms.numel() > 0 else 0.0
        
        if debug:
            print(f"Attack result: {'SUCCESS' if attack_success else 'FAILURE'}")
            print(f"Adversarial prediction: {adv_pred.item()}, confidence: {adv_confidence:.4f}")
            print(f"L2 perturbation norm in logits space: {l2_val:.4f}")
        
        if attack_success:
            # 根据logits空间的L2值范围来调整映射
            # 对于logits空间，范围可能与图像空间不同
            # 这里假设logits空间的L2范围为[0,10]
            logits_l2_max = 10.0
            mapped_hardness = 10.0 * min(l2_val, logits_l2_max) / logits_l2_max
            if debug:
                print(f"Mapped hardness (CW logits): {mapped_hardness:.4f}")
            
            return mapped_hardness, True
        else:
            # 攻击失败，表示样本非常稳定
            if debug:
                print("Attack failed in logits space - using high hardness value 9.0")
            return 9.0, False
            
    except Exception as e:
        if debug:
            print(f"Error in logits CW attack: {e}")
        # 使用备用方法(基于logit gap)
        logits_tensor = torch.tensor(outputs.cpu().numpy(), dtype=torch.float32)
        original_pred = torch.argmax(logits_tensor).item()
        
        # Calculate logit gap
        sorted_logits, _ = torch.sort(logits_tensor, descending=True)
        logit_gap = (sorted_logits[0][0] - sorted_logits[0][1]).item()
        
        # 如果攻击失败，使用logit gap作为硬度值的指标
        logit_gap_score = 5.0 + 5.0 * (logit_gap / (logit_gap + 1.0))
        if debug:
            print(f"Using fallback logit gap hardness due to error: {logit_gap_score:.4f}")
        return logit_gap_score, False

# 方法2: 在logits上添加高斯噪声计算hardness
def compute_hardness_gaussian_logits(model, inputs, num_trials=100, noise_std_range=(0.1, 5.0), debug=True):
    """
    通过在logits上添加高斯噪声来计算样本的硬度
    
    Args:
        model: 模型
        inputs: 输入样本
        num_trials: 每个噪声级别的试验次数
        noise_std_range: 噪声标准差范围
        debug: 是否打印调试信息
        
    Returns:
        hardness: 样本硬度值
    """
    model.eval()
    inputs = inputs.to(device)
    
    # 获取原始预测和logits
    with torch.no_grad():
        logits = model(inputs)
        original_pred = torch.argmax(logits, 1).item()
        
    # 噪声标准差列表
    noise_stds = np.linspace(noise_std_range[0], noise_std_range[1], 10)
    
    # 对每个噪声级别，尝试多次并记录预测改变的频率
    pred_changes = []
    
    for noise_std in noise_stds:
        changes = 0
        for _ in range(num_trials):
            # 添加高斯噪声到logits
            noise = torch.randn_like(logits) * noise_std
            noisy_logits = logits + noise
            noisy_pred = torch.argmax(noisy_logits, 1).item()
            
            # 检查预测是否改变
            if noisy_pred != original_pred:
                changes += 1
        
        # 计算该噪声级别下的预测改变比例
        change_rate = changes / num_trials
        pred_changes.append(change_rate)
        
        # 如果预测改变率超过50%，停止尝试更大的噪声
        if change_rate > 0.5:
            break
    
    # 寻找首次使预测改变率超过20%的噪声级别
    threshold_idx = next((i for i, rate in enumerate(pred_changes) if rate >= 0.2), len(noise_stds) - 1)
    threshold_std = noise_stds[threshold_idx]
    
    # 将噪声阈值映射到硬度值
    # 噪声越小，样本越容易被扰动，硬度越低
    # 我们将噪声范围映射到硬度范围[0,10]
    max_std = noise_std_range[1]
    mapped_hardness = 10.0 * min(threshold_std, max_std) / max_std
    
    if debug:
        print(f"Original prediction: {original_pred}")
        print(f"Threshold noise std: {threshold_std:.4f}")
        print(f"Mapped hardness (gaussian logits): {mapped_hardness:.4f}")
    
    return mapped_hardness, True

# 方法3: 在特征上添加高斯噪声计算hardness
def compute_hardness_gaussian_features(model, inputs, feature_extractor, num_trials=100, noise_std_range=(0.1, 1.0), debug=True):
    """
    通过在特征上添加高斯噪声来计算样本的硬度
    
    Args:
        model: 模型
        inputs: 输入样本
        feature_extractor: 特征提取器
        num_trials: 每个噪声级别的试验次数
        noise_std_range: 噪声标准差范围
        debug: 是否打印调试信息
        
    Returns:
        hardness: 样本硬度值
    """
    model.eval()
    feature_extractor.eval()
    inputs = inputs.to(device)
    
    # 获取原始特征和预测
    with torch.no_grad():
        features, logits = feature_extractor(inputs)
        original_pred = torch.argmax(logits, 1).item()
        features = features.squeeze()
    
    # 噪声标准差列表
    noise_stds = np.linspace(noise_std_range[0], noise_std_range[1], 10)
    
    # 对每个噪声级别，尝试多次并记录预测改变的频率
    pred_changes = []
    
    # 创建一个函数来从特征到logits
    def features_to_logits(feats):
        # 这里取决于你的模型结构，需要相应调整
        feats = feats.view(feats.size(0), -1)
        return model.fc(feats)
    
    for noise_std in noise_stds:
        changes = 0
        for _ in range(num_trials):
            # 添加高斯噪声到特征
            noise = torch.randn_like(features) * noise_std
            noisy_features = features + noise
            
            # 从特征计算logits
            noisy_logits = features_to_logits(noisy_features.unsqueeze(0))
            noisy_pred = torch.argmax(noisy_logits, 1).item()
            
            # 检查预测是否改变
            if noisy_pred != original_pred:
                changes += 1
        
        # 计算该噪声级别下的预测改变比例
        change_rate = changes / num_trials
        pred_changes.append(change_rate)
        
        # 如果预测改变率超过50%，停止尝试更大的噪声
        if change_rate > 0.5:
            break
    
    # 寻找首次使预测改变率超过20%的噪声级别
    threshold_idx = next((i for i, rate in enumerate(pred_changes) if rate >= 0.2), len(noise_stds) - 1)
    threshold_std = noise_stds[threshold_idx]
    
    # 将噪声阈值映射到硬度值
    # 噪声越小，样本越容易被扰动，硬度越低
    # 我们将噪声范围映射到硬度范围[0,10]
    max_std = noise_std_range[1]
    mapped_hardness = 10.0 * min(threshold_std, max_std) / max_std
    
    if debug:
        print(f"Original prediction: {original_pred}")
        print(f"Threshold noise std: {threshold_std:.4f}")
        print(f"Mapped hardness (gaussian features): {mapped_hardness:.4f}")
    
    return mapped_hardness, True

# 方法4: 在特征上使用CW攻击计算hardness
def compute_hardness_cw_features(model, inputs, feature_extractor, debug=True):
    """
    通过在特征空间上使用CW攻击来计算样本的硬度
    
    Args:
        model: 模型
        inputs: 输入样本
        feature_extractor: 特征提取器
        debug: 是否打印调试信息
        
    Returns:
        hardness: 样本硬度值
    """
    model.eval()
    feature_extractor.eval()
    inputs = inputs.to(device)
    
    # 获取原始特征和预测
    with torch.no_grad():
        features, logits = feature_extractor(inputs)
        original_pred = torch.argmax(logits, 1)
        features = features.squeeze()
    
    # 创建一个特征到logits的模型包装器
    class FeatureWrapper(nn.Module):
        def __init__(self, model):
            super().__init__()
            self.fc = model.fc
        
        def forward(self, x):
            x = x.view(x.size(0), -1)
            return self.fc(x)
    
    feature_model = FeatureWrapper(model).to(device)
    
    # 准备函数包装器
    def model_fn(x):
        return feature_model(x)
    
    # 应用CW攻击到特征
    try:
        if debug:
            print("Running CW attack on features with parameters:")
            print(f"- Binary search steps: 5")
            print(f"- Max iterations: 300")
            print(f"- Learning rate: 0.01")
            print(f"- Early stop: True")
        
        # 重塑特征形状以适应CW攻击函数
        feature_input = features.unsqueeze(0)
        
        # 可能需要调整这些参数以适应特征空间
        adv_features, l2_norms, success_mask = carlini_wagner_l2(
            model_fn=model_fn,
            x=feature_input,
            n_classes=logits.shape[1],
            y=original_pred,
            targeted=False,
            lr=0.01,  # 特征空间可能需要更高的学习率
            binary_search_steps=5,
            max_iterations=300,
            confidence=0,
            clip_min=float(feature_input.min()),
            clip_max=float(feature_input.max()),
            debug=debug,
            early_stop=True
        )
        
        # 获取对抗特征的预测
        with torch.no_grad():
            adv_logits = feature_model(adv_features)
            adv_pred = torch.argmax(adv_logits, 1)
        
        # 攻击成功与否直接从success_mask获取
        attack_success = success_mask.any().item()
        
        # 获取L2范数
        l2_val = l2_norms[0].item() if l2_norms.numel() > 0 else 0.0
        
        if debug:
            print(f"Attack result: {'SUCCESS' if attack_success else 'FAILURE'}")
            print(f"Original prediction: {original_pred.item()}")
            print(f"Adversarial prediction: {adv_pred.item()}")
            print(f"L2 perturbation norm in feature space: {l2_val:.4f}")
        
        if attack_success:
            # 根据特征空间的L2值范围来调整映射
            # 对于特征空间，范围可能与图像空间不同
            # 这里假设特征空间的L2范围为[0,2]
            feature_l2_max = 2.0
            mapped_hardness = 10.0 * min(l2_val, feature_l2_max) / feature_l2_max
            if debug:
                print(f"Mapped hardness (CW features): {mapped_hardness:.4f}")
            
            return mapped_hardness, True
        else:
            # 攻击失败，表示样本非常稳定
            if debug:
                print("Attack failed in feature space - using high hardness value 9.0")
            return 9.0, False
            
    except Exception as e:
        if debug:
            print(f"Error in feature CW attack: {e}")
        # 使用备用方法(基于logit gap)
        logits_tensor = torch.tensor(logits.cpu().numpy(), dtype=torch.float32)
        original_pred = torch.argmax(logits_tensor).item()
        
        # Calculate logit gap
        sorted_logits, _ = torch.sort(logits_tensor, descending=True)
        logit_gap = (sorted_logits[0][0] - sorted_logits[0][1]).item()
        
        # 如果攻击失败，使用logit gap作为硬度值的指标
        logit_gap_score = 5.0 + 5.0 * (logit_gap / (logit_gap + 1.0))
        if debug:
            print(f"Using fallback logit gap hardness due to error: {logit_gap_score:.4f}")
        return logit_gap_score, False

# 计算ECE (Expected Calibration Error)
def compute_ece(probs, labels, n_bins=15):
    """计算ECE (Expected Calibration Error)"""
    bin_boundaries = np.linspace(0, 1, n_bins + 1)
    bin_lowers = bin_boundaries[:-1]
    bin_uppers = bin_boundaries[1:]
    
    confidences = np.max(probs, axis=1)
    predictions = np.argmax(probs, axis=1)
    accuracies = (predictions == labels)
    
    ece = 0.0
    for bin_lower, bin_upper in zip(bin_lowers, bin_uppers):
        in_bin = np.logical_and(confidences > bin_lower, confidences <= bin_upper)
        prop_in_bin = np.mean(in_bin)
        if prop_in_bin > 0:
            accuracy_in_bin = np.mean(accuracies[in_bin])
            avg_confidence_in_bin = np.mean(confidences[in_bin])
            ece += np.abs(avg_confidence_in_bin - accuracy_in_bin) * prop_in_bin
    
    return ece

# Temperature Scaling
class TemperatureScaling:
    """标准温度缩放校准方法"""
    def __init__(self, temp=1.0):
        self.temp = temp
    
    def fit(self, logits, labels):
        """通过最小化验证集上的NLL找到最优温度"""
        from scipy.optimize import minimize
        
        def objective(temp, logits, labels):
            scaled_logits = logits / temp
            probs = F.softmax(torch.tensor(scaled_logits), dim=1).numpy()
            nll = -np.mean(np.log(probs[np.arange(len(labels)), labels] + 1e-10))
            return nll
        
        # 简单的参数寻优
        res = minimize(lambda t: objective(t, logits, labels), 
                      x0=np.array([self.temp]), method='BFGS')
        self.temp = res.x[0]
        
    def calibrate(self, logits):
        """应用温度缩放到logits"""
        scaled_logits = logits / self.temp
        return F.softmax(torch.tensor(scaled_logits), dim=1).numpy()

# 比较不同硬度计算方法的校准效果
def compare_hardness_methods(model, testloader, num_samples=50, run_methods=None):
    """
    比较不同硬度计算方法的校准效果
    
    Args:
        model: 要评估的模型
        testloader: 测试数据加载器
        num_samples: 要评估的样本数量
        run_methods: 要运行的方法列表，可选值：['cw_logits', 'gaussian_logits', 'gaussian_features', 'cw_features']
                     如果为None，则运行所有方法
    """
    feature_extractor = get_feature_extractor(model)
    
    # 所有可用的方法
    all_methods = {
        'cw_logits': {'hardness_values': [], 'temps': [], 'func': compute_hardness_cw_logits},
        'gaussian_logits': {'hardness_values': [], 'temps': [], 'func': compute_hardness_gaussian_logits},
        'gaussian_features': {'hardness_values': [], 'temps': [], 'func': compute_hardness_gaussian_features},
        'cw_features': {'hardness_values': [], 'temps': [], 'func': compute_hardness_cw_features}
    }
    
    # 确定要运行的方法
    if run_methods is None:
        # 如果未指定，运行所有方法
        methods = all_methods
        method_names = list(all_methods.keys())
    else:
        # 只保留指定的方法
        methods = {method: all_methods[method] for method in run_methods if method in all_methods}
        method_names = run_methods
        
        if not methods:
            print("警告: 未指定有效的方法，将运行所有方法")
            methods = all_methods
            method_names = list(all_methods.keys())
        else:
            print(f"将运行以下方法: {', '.join(method_names)}")
    
    all_logits = []
    all_labels = []
    
    # 收集样本和计算不同方法的硬度
    print(f"Calculating hardness using different methods for {num_samples} samples...")
    
    for i, (inputs, labels) in enumerate(tqdm(testloader)):
        if i >= num_samples:
            break
            
        inputs = inputs.to(device)
        labels = labels.to(device)
        
        # 获取logits
        with torch.no_grad():
            logits = model(inputs)
        
        all_logits.append(logits.cpu().numpy())
        all_labels.append(labels.cpu().numpy())
        
        print(f"\n------ Sample {i+1}/{num_samples} ------")
        
        # 运行每种指定的方法
        for idx, method_name in enumerate(method_names):
            method_info = methods[method_name]
            func = method_info['func']
            
            print(f"\nMethod {idx+1}: {method_name}")
            
            # 根据方法类型调用相应的函数
            if method_name == 'cw_logits':
                hardness, _ = func(model, inputs, logits, debug=True)
            elif method_name == 'gaussian_logits':
                hardness, _ = func(model, inputs, debug=True)
            elif method_name in ['gaussian_features', 'cw_features']:
                hardness, _ = func(model, inputs, feature_extractor, debug=True)
                
            methods[method_name]['hardness_values'].append(hardness)
        
        print("----------------------------------------\n")
    
    # 整理收集的数据
    all_logits = np.vstack(all_logits)
    all_labels = np.hstack(all_labels)
    
    # 计算硬度到温度的映射
    for method_name, method_data in methods.items():
        # 使用简单的线性映射：硬度越高，温度越高
        hardness_temps = [0.5 + 1.5 * (h / 10.0) for h in method_data['hardness_values']]
        methods[method_name]['temps'] = hardness_temps
    
    # 评估校准效果
    print("\nEvaluating calibration performance...")
    
    # 1. 未校准
    uncal_probs = F.softmax(torch.tensor(all_logits), dim=1).numpy()
    uncal_ece = compute_ece(uncal_probs, all_labels)
    uncal_acc = np.mean(np.argmax(uncal_probs, axis=1) == all_labels)
    
    # 2. 标准温度缩放
    ts = TemperatureScaling()
    ts.fit(all_logits, all_labels)
    ts_probs = ts.calibrate(all_logits)
    ts_ece = compute_ece(ts_probs, all_labels)
    ts_acc = np.mean(np.argmax(ts_probs, axis=1) == all_labels)
    
    # 3. 各种硬度方法的样本级温度缩放
    results = {
        'uncal': {'ece': uncal_ece, 'acc': uncal_acc},
        'ts': {'ece': ts_ece, 'acc': ts_acc, 'temp': ts.temp}
    }
    
    for method_name, method_data in methods.items():
        hardness_temps = method_data['temps']
        shats_probs = []
        
        for i, logits in enumerate(all_logits):
            temp = hardness_temps[i] if i < len(hardness_temps) else 1.0
            scaled_logits = logits / temp
            probs = F.softmax(torch.tensor(scaled_logits), dim=1).numpy()
            shats_probs.append(probs)
        
        shats_probs = np.vstack(shats_probs)
        shats_ece = compute_ece(shats_probs, all_labels)
        shats_acc = np.mean(np.argmax(shats_probs, axis=1) == all_labels)
        
        results[method_name] = {
            'ece': shats_ece, 
            'acc': shats_acc,
            'hardness_values': method_data['hardness_values'],
            'temps': hardness_temps,
            'avg_hardness': np.mean(method_data['hardness_values']),
            'avg_temp': np.mean(hardness_temps)
        }
    
    # 打印结果
    print("\nCalibration Results:")
    print(f"Number of samples: {len(all_labels)}")
    print(f"TS optimal temperature: {ts.temp:.4f}")
    print("\nMethod     Avg Hardness   Avg Temp    Accuracy    ECE")
    print("-" * 60)
    print(f"Uncal      -              -           {uncal_acc:.4f}      {uncal_ece:.4f}")
    print(f"TS         -              {ts.temp:.4f}       {ts_acc:.4f}      {ts_ece:.4f}")
    
    for method_name, method_results in results.items():
        if method_name not in ['uncal', 'ts']:
            print(f"{method_name:<10} {method_results['avg_hardness']:.4f}         {method_results['avg_temp']:.4f}       {method_results['acc']:.4f}      {method_results['ece']:.4f}")
    
    # 保存结果
    save_path = os.path.join(TEST_RESULTS_DIR, "hardness_methods_comparison.json")
    with open(save_path, 'w') as f:
        json.dump(results, f, indent=4, default=float)
    
    print(f"\nResults saved to {save_path}")
    
    # 计算不同硬度方法之间的相关性
    print("\nCalculating correlations between hardness methods...")
    correlations = {}
    
    if len(methods) > 1:  # 只有当有多个方法时才计算相关性
        method_names = list(methods.keys())
        for i, method1 in enumerate(method_names):
            for method2 in method_names[i+1:]:
                h1 = methods[method1]['hardness_values']
                h2 = methods[method2]['hardness_values']
                # 确保长度一致
                min_len = min(len(h1), len(h2))
                
                if min_len > 0:
                    corr = np.corrcoef(h1[:min_len], h2[:min_len])[0, 1]
                    correlations[f"{method1}_vs_{method2}"] = corr
                    print(f"Correlation between {method1} and {method2}: {corr:.4f}")
    else:
        print("只有一种方法被运行，无法计算相关性")
    
    # 将相关性也保存到结果中
    results['correlations'] = correlations
    
    # 更新保存的结果
    with open(save_path, 'w') as f:
        json.dump(results, f, indent=4, default=float)
    
    return results

if __name__ == "__main__":
    print("Loading CIFAR-10 dataset...")
    testloader, _ = load_cifar10(batch_size=1)
    
    print("Loading model...")
    model = load_model()
    
    # 指定要运行的方法 (可选值: 'cw_logits', 'gaussian_logits', 'gaussian_features', 'cw_features')
    # 设置为None将运行所有方法
    run_methods = ['gaussian_logits', 'gaussian_features']  # 示例：只运行基于高斯噪声的方法
    
    print("\n--- Comparing different hardness calculation methods ---")
    start_time = time.time()
    results = compare_hardness_methods(model, testloader, num_samples=20, run_methods=run_methods)
    end_time = time.time()
    
    print(f"Total execution time: {(end_time - start_time) / 60:.2f} minutes")
    print(f"Results saved to {TEST_RESULTS_DIR}/hardness_methods_comparison.json")