diff --git a/GRID_SEARCH_README.md b/GRID_SEARCH_README.md
new file mode 100644
index 0000000..3f00253
--- /dev/null
+++ b/GRID_SEARCH_README.md
@@ -0,0 +1,149 @@
+# CosFace超参数网格搜索使用指南
+
+## 概述
+
+已完成CosFace超参数网格搜索功能的实施，用于优化`scale (s)`和`margin (m)`两个核心超参数。
+
+## 修改文件清单
+
+### 1. `requirements.txt`
+- ✅ 添加依赖：`pandas>=2.0.0` 和 `seaborn>=0.12.0`
+
+### 2. `train/train_cosface_embedding.py`
+- ✅ `TaskConfig`添加`test_dir`字段
+- ✅ `TASKS`字典中所有任务添加测试集路径
+
+### 3. `train/grid_search_cosface.py` (新建)
+- ✅ 完整的网格搜索脚本 (~350行)
+
+## 搜索空间配置
+
+当前为dish任务配置的搜索空间：
+```python
+GRID_PARAMS = {
+    'dish': {
+        's': [56.0, 60.0, 64.0, 68.0],    # 4个scale值
+        'm': [0.32, 0.35, 0.38, 0.40],    # 4个margin值
+    }
+}
+```
+
+**总配置数**: 4 × 4 = **16组实验**
+
+## 使用方法
+
+### 1. 安装依赖
+```bash
+pip install pandas>=2.0.0 seaborn>=0.12.0
+```
+
+### 2. 运行网格搜索
+```bash
+# 完整搜索（16组配置）
+python train/grid_search_cosface.py --task dish
+
+# 测试运行（限制配置数）
+python train/grid_search_cosface.py --task dish --max_configs 2
+
+# 自定义参数
+python train/grid_search_cosface.py --task dish --epochs 100 --patience 10 --min_epochs 25
+```
+
+### 3. 参数说明
+- `--task`: 任务名称 (`dish` / `whole_ingredient` / `processed_ingredient`)
+- `--max_configs`: 限制最大配置数（用于测试，可选）
+- `--epochs`: 每个配置的最大训练轮数（默认100）
+- `--patience`: 早停容忍轮数（默认10）
+- `--min_epochs`: 最小训练轮数（默认25）
+
+## 输出文件
+
+运行后会在`model/DishClassification/grid_search_YYYYMMDD_HHMMSS/`目录下生成：
+
+1. **`grid_search_results.csv`** - 所有配置的详细结果表格
+2. **`heatmap.png`** - 参数热力图（测试集准确率）
+3. **`summary.txt`** - 搜索总结报告（含最佳配置）
+4. **`model_s{s}_m{m}.pth`** - 每个配置的模型权重
+5. **`grid_search.log`** - 完整训练日志
+
+## 评估指标
+
+- **主要指标**: 测试集准确率 (`test_acc`) - 用于选择最佳配置
+- **辅助指标**: 验证集准确率、训练轮数、训练时间
+
+## 应用最佳配置
+
+网格搜索完成后：
+
+1. 查看`summary.txt`找到最佳配置
+2. 手动更新`train/train_cosface_embedding.py`中的`TASKS`字典：
+```python
+'dish': TaskConfig(
+    # ... 其他配置保持不变 ...
+    cosface_s=64.0,  # 更新为最佳s值
+    cosface_m=0.38,  # 更新为最佳m值
+)
+```
+3. 后续训练将自动使用最佳配置
+
+## 预计耗时
+
+基于以下假设：
+- 每个配置平均训练25-40个epoch（早停机制）
+- 每个epoch约1-2分钟
+- **单个配置**: ~30-60分钟
+- **16组配置总耗时**: ~8-16小时
+
+**建议**: 使用GPU运行，可在夜间或周末执行完整搜索。
+
+## 注意事项
+
+1. **数据集要求**: 确保`dataset/DishClassification/test/`目录存在且有数据
+2. **GPU推荐**: 网格搜索计算量大，强烈建议使用GPU
+3. **磁盘空间**: 每个配置约占用500MB，16组需8GB空间
+4. **中断恢复**: 当前版本不支持断点续训，建议一次性完成
+
+## 高级用法
+
+### 修改搜索空间
+编辑`train/grid_search_cosface.py`中的`GRID_PARAMS`字典：
+```python
+GRID_PARAMS = {
+    'dish': {
+        's': [60.0, 64.0, 68.0, 72.0],     # 自定义scale范围
+        'm': [0.30, 0.35, 0.40, 0.45],     # 自定义margin范围
+    }
+}
+```
+
+### 调整早停策略
+通过命令行参数调整：
+```bash
+python train/grid_search_cosface.py --task dish --patience 15 --min_epochs 30
+```
+
+## 故障排除
+
+**问题1**: `ModuleNotFoundError: No module named 'pandas'`
+- 解决: `pip install pandas seaborn`
+
+**问题2**: 测试集目录不存在
+- 解决: 确认`dataset/DishClassification/test/`路径正确且有数据
+
+**问题3**: CUDA out of memory
+- 解决: 减小`batch_size`或在CPU上运行（速度较慢）
+
+## 示例结果解读
+
+`summary.txt`示例：
+```
+最佳配置:
+  s (scale) = 64.0
+  m (margin) = 0.38
+  测试集准确率 = 98.50%
+  验证集准确率 = 100.00%
+  训练轮数 = 32
+  训练时间 = 45.3 分钟
+```
+
+这表示s=64.0, m=0.38是最优组合，在测试集上达到98.50%准确率。
diff --git a/requirements.txt b/requirements.txt
index 00f3c31..b8a8ba0 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -28,3 +28,5 @@ typing_extensions==4.15.0
 faiss-cpu>=1.7.0
 requests>=2.28.0
 openai>=1.0.0
+pandas>=2.0.0
+seaborn>=0.12.0
diff --git a/train/grid_search.log b/train/grid_search.log
new file mode 100644
index 0000000..e69de29
diff --git a/train/grid_search_cosface.py b/train/grid_search_cosface.py
new file mode 100644
index 0000000..89819a6
--- /dev/null
+++ b/train/grid_search_cosface.py
@@ -0,0 +1,371 @@
+"""
+CosFace超参数网格搜索脚本
+用法: python train/grid_search_cosface.py --task dish
+"""
+import os
+import sys
+import time
+import json
+import logging
+import itertools
+from datetime import datetime
+from typing import List, Dict, Tuple
+
+import torch
+import torch.nn as nn
+import torch.optim as optim
+from torch.utils.data import DataLoader
+from torchvision import datasets, transforms
+from tqdm import tqdm
+import pandas as pd
+import seaborn as sns
+import matplotlib.pyplot as plt
+
+# 导入训练脚本中的组件
+sys.path.append(os.path.join(os.path.dirname(__file__), '..'))
+from train.train_cosface_embedding import (
+    TASKS, build_transforms, CosFaceHead, run_epoch, device
+)
+from net.resnet_embedding import create_resnet50_embedding
+from settings import settings
+
+# 日志配置
+logging.basicConfig(
+    level=logging.INFO,
+    format='%(asctime)s - %(levelname)s - %(message)s',
+    handlers=[
+        logging.FileHandler('grid_search.log', encoding='utf-8'),
+        logging.StreamHandler()
+    ]
+)
+logger = logging.getLogger(__name__)
+
+# ==================== 网格搜索空间定义 ====================
+GRID_PARAMS = {
+    'dish': {
+        's': [56.0, 60.0, 64.0, 68.0],  # scale参数
+        'm': [0.32, 0.35, 0.38, 0.40],  # margin参数
+    },
+    'whole_ingredient': {
+        's': [56.0, 60.0, 64.0, 68.0],
+        'm': [0.32, 0.35, 0.38, 0.40],
+    },
+    'processed_ingredient': {
+        's': [56.0, 60.0, 64.0, 68.0],
+        'm': [0.30, 0.33, 0.36, 0.39],
+    },
+}
+
+
+def evaluate_on_test(model: nn.Module, head: nn.Module, test_loader: DataLoader) -> Tuple[float, float]:
+    """在测试集上评估模型"""
+    model.eval()
+    head.eval()
+    criterion = nn.CrossEntropyLoss()
+    
+    total_loss = 0.0
+    total_samples = 0
+    correct_samples = 0
+    
+    with torch.no_grad():
+        for images, labels in tqdm(test_loader, desc='测试集评估', leave=False):
+            images = images.to(device)
+            labels = labels.to(device)
+            
+            feats = model(images)
+            logits = head(feats)  # 测试时不加margin
+            loss = criterion(logits, labels)
+            
+            pred = logits.argmax(dim=1)
+            correct = (pred == labels).sum().item()
+            
+            total_loss += loss.item() * labels.size(0)
+            total_samples += labels.size(0)
+            correct_samples += correct
+    
+    avg_loss = total_loss / max(total_samples, 1)
+    avg_acc = 100.0 * correct_samples / max(total_samples, 1)
+    
+    return avg_acc, avg_loss
+
+
+def train_single_config(
+    cfg,
+    s: float,
+    m: float,
+    train_loader: DataLoader,
+    val_loader: DataLoader,
+    test_loader: DataLoader,
+    num_classes: int,
+    max_epochs: int = 100,
+    patience: int = 10,
+    min_epochs: int = 25,
+    min_delta: float = 0.001
+) -> Tuple[nn.Module, nn.Module, Dict]:
+    """训练单个超参数配置"""
+    logger.info(f'开始训练配置: s={s}, m={m}')
+    
+    # 创建模型
+    model = create_resnet50_embedding(
+        embedding_dim=cfg.embedding_dim,
+        pretrained=True,
+        use_internal_preprocess=False
+    ).to(device)
+    
+    head = CosFaceHead(
+        in_features=cfg.embedding_dim,
+        num_classes=num_classes,
+        s=s,
+        m=m
+    ).to(device)
+    
+    # 优化器和学习率调度
+    criterion = nn.CrossEntropyLoss()
+    optimizer = optim.AdamW(
+        list(model.parameters()) + list(head.parameters()),
+        lr=cfg.lr,
+        weight_decay=1e-4
+    )
+    scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=max_epochs)
+    
+    # 早停变量
+    best_val_loss = float('inf')
+    early_stop_counter = 0
+    best_epoch = 0
+    
+    # 训练历史
+    train_losses = []
+    val_losses = []
+    train_accs = []
+    val_accs = []
+    
+    start_time = time.time()
+    
+    # 训练循环
+    for epoch in range(max_epochs):
+        # 训练一个epoch
+        tr_loss, tr_acc = run_epoch(model, head, train_loader, criterion, optimizer, train=True)
+        va_loss, va_acc = run_epoch(model, head, val_loader, criterion, optimizer=None, train=False)
+        scheduler.step()
+        
+        train_losses.append(tr_loss)
+        val_losses.append(va_loss)
+        train_accs.append(tr_acc)
+        val_accs.append(va_acc)
+        
+        logger.info(f'Epoch {epoch+1}/{max_epochs} | Train Loss: {tr_loss:.4f}, Acc: {tr_acc:.2f}% | Val Loss: {va_loss:.4f}, Acc: {va_acc:.2f}%')
+        
+        # 早停逻辑
+        if va_loss < best_val_loss - min_delta:
+            best_val_loss = va_loss
+            early_stop_counter = 0
+            best_epoch = epoch
+        else:
+            early_stop_counter += 1
+        
+        # 早停判断
+        if epoch >= min_epochs and early_stop_counter >= patience:
+            logger.info(f'早停触发于Epoch {epoch+1}, 最佳Epoch: {best_epoch+1}')
+            break
+    
+    elapsed_time = time.time() - start_time
+    actual_epochs = len(train_losses)
+    
+    # 测试集评估
+    test_acc, test_loss = evaluate_on_test(model, head, test_loader)
+    
+    logger.info(f'✓ 训练完成 | 耗时: {elapsed_time/60:.1f}分钟 | 实际Epoch: {actual_epochs} | Val Acc={val_accs[-1]:.2f}% | Test Acc={test_acc:.2f}%')
+    
+    train_info = {
+        's': s,
+        'm': m,
+        'actual_epochs': actual_epochs,
+        'early_stopped': actual_epochs < max_epochs,
+        'training_time_minutes': elapsed_time / 60,
+        'best_val_loss': best_val_loss,
+        'final_train_loss': train_losses[-1],
+        'final_val_loss': val_losses[-1],
+        'final_train_acc': train_accs[-1],
+        'final_val_acc': val_accs[-1],
+        'test_acc': test_acc,
+        'test_loss': test_loss,
+    }
+    
+    return model, head, train_info
+
+
+def save_results_csv(results: List[Dict], save_path: str):
+    """保存结果到CSV"""
+    df = pd.DataFrame(results)
+    df = df.sort_values('test_acc', ascending=False)
+    df.to_csv(save_path, index=False, encoding='utf-8-sig')
+    logger.info(f'✓ 结果CSV已保存: {save_path}')
+    
+    # 打印前5名
+    logger.info('\n' + '='*60)
+    logger.info('测试集准确率TOP5配置:')
+    logger.info('='*60)
+    top5 = df.head(5)
+    for idx, row in top5.iterrows():
+        logger.info(f'Rank {idx+1}: s={row["s"]}, m={row["m"]} | Test Acc={row["test_acc"]:.2f}% | '
+                   f'Val Acc={row["final_val_acc"]:.2f}% | Epochs={int(row["actual_epochs"])}')
+    logger.info('='*60 + '\n')
+
+
+def visualize_heatmap(results: List[Dict], save_path: str):
+    """生成参数热力图"""
+    df = pd.DataFrame(results)
+    
+    # 创建透视表
+    pivot = df.pivot_table(values='test_acc', index='m', columns='s', aggfunc='mean')
+    
+    plt.figure(figsize=(10, 8))
+    sns.heatmap(pivot, annot=True, fmt='.2f', cmap='YlGnBu', cbar_kws={'label': 'Test Accuracy (%)'})
+    plt.title('CosFace超参数网格搜索热力图 (测试集准确率)', fontsize=14, fontweight='bold')
+    plt.xlabel('Scale (s)', fontsize=12)
+    plt.ylabel('Margin (m)', fontsize=12)
+    plt.tight_layout()
+    plt.savefig(save_path, dpi=300, bbox_inches='tight')
+    plt.close()
+    logger.info(f'✓ 热力图已保存: {save_path}')
+
+
+def grid_search_main(
+    task_key: str = 'dish',
+    max_configs: int = None,
+    max_epochs: int = 100,
+    patience: int = 10,
+    min_epochs: int = 25
+):
+    """网格搜索主函数"""
+    logger.info('='*80)
+    logger.info(f'开始CosFace超参数网格搜索 - 任务: {task_key}')
+    logger.info('='*80)
+    
+    cfg = TASKS[task_key]
+    params = GRID_PARAMS[task_key]
+    
+    # 生成所有参数组合
+    param_combinations = list(itertools.product(params['s'], params['m']))
+    total_configs = len(param_combinations)
+    
+    if max_configs is not None and max_configs < total_configs:
+        param_combinations = param_combinations[:max_configs]
+        logger.info(f'⚠ 限制最大配置数为 {max_configs}')
+    
+    logger.info(f'任务配置: {cfg.name}')
+    logger.info(f'搜索空间: s={params["s"]}, m={params["m"]}')
+    logger.info(f'总配置数: {len(param_combinations)}')
+    logger.info(f'早停参数: patience={patience}, min_epochs={min_epochs}')
+    logger.info(f'评估指标: 测试集准确率 (test_acc)')
+    logger.info('='*80 + '\n')
+    
+    # 创建保存目录
+    timestamp = datetime.now().strftime('%Y%m%d_%H%M%S')
+    save_dir = os.path.join(settings.BASE_DIR, 'model', cfg.name, f'grid_search_{timestamp}')
+    os.makedirs(save_dir, exist_ok=True)
+    
+    # 准备数据
+    transform_train, transform_val = build_transforms(cfg.aug_strength)
+    
+    train_ds = datasets.ImageFolder(cfg.train_dir, transform=transform_train)
+    val_ds = datasets.ImageFolder(cfg.val_dir, transform=transform_val)
+    test_ds = datasets.ImageFolder(cfg.test_dir, transform=transform_val)
+    
+    train_loader = DataLoader(train_ds, batch_size=cfg.batch_size, shuffle=True, num_workers=0, drop_last=True)
+    val_loader = DataLoader(val_ds, batch_size=cfg.batch_size, shuffle=False, num_workers=0)
+    test_loader = DataLoader(test_ds, batch_size=cfg.batch_size, shuffle=False, num_workers=0)
+    
+    num_classes = len(train_ds.classes)
+    logger.info(f'数据集: 训练={len(train_ds)}, 验证={len(val_ds)}, 测试={len(test_ds)}, 类别数={num_classes}\n')
+    
+    # 网格搜索
+    results = []
+    total_start = time.time()
+    
+    for i, (s, m) in enumerate(param_combinations, 1):
+        logger.info(f'\n[配置 {i}/{len(param_combinations)}] s={s}, m={m}')
+        logger.info('-'*80)
+        
+        try:
+            model, head, train_info = train_single_config(
+                cfg, s, m,
+                train_loader, val_loader, test_loader,
+                num_classes,
+                max_epochs=max_epochs,
+                patience=patience,
+                min_epochs=min_epochs
+            )
+            results.append(train_info)
+            
+            # 保存模型
+            model_path = os.path.join(save_dir, f'model_s{s}_m{m}.pth')
+            torch.save({
+                'backbone_state_dict': model.state_dict(),
+                'head_state_dict': head.state_dict(),
+                's': s,
+                'm': m,
+                'test_acc': train_info['test_acc'],
+                'class_names': train_ds.classes,
+            }, model_path)
+            
+        except Exception as e:
+            logger.error(f'配置 s={s}, m={m} 训练失败: {str(e)}')
+            continue
+    
+    total_elapsed = time.time() - total_start
+    
+    # 保存结果
+    csv_path = os.path.join(save_dir, 'grid_search_results.csv')
+    save_results_csv(results, csv_path)
+    
+    heatmap_path = os.path.join(save_dir, 'heatmap.png')
+    visualize_heatmap(results, heatmap_path)
+    
+    # 找到最佳配置
+    best_result = max(results, key=lambda x: x['test_acc'])
+    
+    # 保存总结报告
+    summary_path = os.path.join(save_dir, 'summary.txt')
+    with open(summary_path, 'w', encoding='utf-8') as f:
+        f.write('='*80 + '\n')
+        f.write(f'CosFace超参数网格搜索总结 - {cfg.name}\n')
+        f.write('='*80 + '\n\n')
+        f.write(f'总耗时: {total_elapsed/3600:.2f} 小时\n')
+        f.write(f'搜索配置数: {len(results)}/{len(param_combinations)}\n')
+        f.write(f'评估指标: 测试集准确率\n\n')
+        f.write('最佳配置:\n')
+        f.write(f'  s (scale) = {best_result["s"]}\n')
+        f.write(f'  m (margin) = {best_result["m"]}\n')
+        f.write(f'  测试集准确率 = {best_result["test_acc"]:.2f}%\n')
+        f.write(f'  验证集准确率 = {best_result["final_val_acc"]:.2f}%\n')
+        f.write(f'  训练轮数 = {best_result["actual_epochs"]}\n')
+        f.write(f'  训练时间 = {best_result["training_time_minutes"]:.1f} 分钟\n\n')
+        f.write('使用建议:\n')
+        f.write(f'在 train/train_cosface_embedding.py 的 TASKS["{task_key}"] 中更新:\n')
+        f.write(f'  cosface_s={best_result["s"]}\n')
+        f.write(f'  cosface_m={best_result["m"]}\n')
+    
+    logger.info(f'\n✓ 网格搜索完成! 总耗时: {total_elapsed/3600:.2f} 小时')
+    logger.info(f'✓ 最佳配置: s={best_result["s"]}, m={best_result["m"]}, Test Acc={best_result["test_acc"]:.2f}%')
+    logger.info(f'✓ 结果已保存到: {save_dir}')
+
+
+if __name__ == '__main__':
+    import argparse
+    parser = argparse.ArgumentParser(description='CosFace超参数网格搜索')
+    # parser.add_argument('--task', choices=list(TASKS.keys()), default='dish', help='任务名称')
+    parser.add_argument('--task', choices=list(TASKS.keys()), default='whole_ingredient', help='任务名称')
+    parser.add_argument('--max_configs', type=int, default=None, help='最大配置数（用于测试）')
+    parser.add_argument('--epochs', type=int, default=100, help='每个配置的最大训练轮数')
+    parser.add_argument('--patience', type=int, default=10, help='早停容忍轮数')
+    parser.add_argument('--min_epochs', type=int, default=25, help='最小训练轮数')
+    args = parser.parse_args()
+    
+    grid_search_main(
+        task_key=args.task,
+        max_configs=args.max_configs,
+        max_epochs=args.epochs,
+        patience=args.patience,
+        min_epochs=args.min_epochs
+    )
diff --git a/train/train_cosface_embedding.py b/train/train_cosface_embedding.py
index 65d3918..f15611f 100644
--- a/train/train_cosface_embedding.py
+++ b/train/train_cosface_embedding.py
@@ -35,6 +35,7 @@ class TaskConfig:
     name: str
     train_dir: str
     val_dir: str
+    test_dir: str  # 测试集目录（用于网格搜索评估）
     embedding_dim: int
     batch_size: int
     lr: float
@@ -48,6 +49,7 @@ TASKS = {
         name='DishClassification',
         train_dir=os.path.join(settings.BASE_DIR, 'dataset', 'DishClassification', 'train'),
         val_dir=os.path.join(settings.BASE_DIR, 'dataset', 'DishClassification', 'val'),
+        test_dir=os.path.join(settings.BASE_DIR, 'dataset', 'DishClassification', 'test'),
         embedding_dim=512,
         batch_size=32,
         lr=5e-4,
@@ -61,6 +63,7 @@ TASKS = {
         name='WholeIngredientRecognition',
         train_dir=os.path.join(settings.BASE_DIR, 'dataset', 'WholeIngredientRecognition', 'train'),
         val_dir=os.path.join(settings.BASE_DIR, 'dataset', 'WholeIngredientRecognition', 'val'),
+        test_dir=os.path.join(settings.BASE_DIR, 'dataset', 'WholeIngredientRecognition', 'test'),
         embedding_dim=512,
         batch_size=64,
         lr=8e-4,
@@ -72,6 +75,7 @@ TASKS = {
         name='ProcessedIngredientRecognition',
         train_dir=os.path.join(settings.BASE_DIR, 'dataset', 'ProcessedIngredientRecognition', 'train'),
         val_dir=os.path.join(settings.BASE_DIR, 'dataset', 'ProcessedIngredientRecognition', 'val'),
+        test_dir=os.path.join(settings.BASE_DIR, 'dataset', 'ProcessedIngredientRecognition', 'test'),
         embedding_dim=512,
         batch_size=32,
         lr=5e-4,