analysis/ablation_experiments/summarize_ablation.py

#!/usr/bin/env python3
"""
汇总 DeCLIP 消融实验结果

生成包含以下内容的 Excel 表格:
1. 各消融实验的最佳结果 (按 novel_ap50)
2. 与基线 (CLIPSelf, ClearCLIP) 的对比
3. 各 epoch 的详细结果
"""

import os
import pandas as pd

# ===== 消融实验结果 =====

# JEPA-GSC 消融实验 (Ablation_JEPA-GSC_EVA-B_DINOv2-B_csa_560)
JEPA_GSC_RESULTS = {
    'experiment': 'JEPA-GSC',
    'description': 'Using I-JEPA as Global Scene Context instead of DINOv2',
    'backbone': 'EVA02-CLIP-B-16',
    'epochs': {
        1: {'base_ap50': 50.953, 'novel_ap50': 37.065, 'all_ap50': 47.321, 'bbox_mAP': 0.230},
        2: {'base_ap50': 54.300, 'novel_ap50': 39.589, 'all_ap50': 50.453, 'bbox_mAP': 0.270},
        3: {'base_ap50': 56.357, 'novel_ap50': 38.061, 'all_ap50': 51.572, 'bbox_mAP': 0.286},
    }
}

# SAM-GSC 消融实验 (Ablation_SAM-GSC_EVA-B_DINOv2-B_csa_560)
SAM_GSC_RESULTS = {
    'experiment': 'SAM-GSC',
    'description': 'Using SAM as Global Scene Context instead of DINOv2',
    'backbone': 'EVA02-CLIP-B-16',
    'epochs': {
        1: {'base_ap50': 50.557, 'novel_ap50': 36.800, 'all_ap50': 46.959, 'bbox_mAP': 0.233},
        2: {'base_ap50': 54.251, 'novel_ap50': 37.820, 'all_ap50': 49.954, 'bbox_mAP': 0.267},
        3: {'base_ap50': 56.274, 'novel_ap50': 37.691, 'all_ap50': 51.414, 'bbox_mAP': 0.283},
    }
}

# ===== 基线结果 (仅供参考，不输出到汇总) =====
BASELINE_RESULTS = {
    'CLIPSelf': {
        'description': 'CLIPSelf with DINOv2-GSC (Base, Ensemble)',
        'base_ap50': 54.94,
        'novel_ap50': 37.51,
        'all_ap50': 50.38,
        'bbox_mAP': 0.277,
    },
}


def get_best_epoch(results, metric='novel_ap50'):
    """获取指定指标最高的 epoch"""
    best_epoch = None
    best_value = -1
    for epoch, metrics in results['epochs'].items():
        if metrics[metric] > best_value:
            best_value = metrics[metric]
            best_epoch = epoch
    return best_epoch, results['epochs'][best_epoch]


def generate_summary_excel(output_path):
    """生成汇总 Excel 表格"""
    
    with pd.ExcelWriter(output_path, engine='openpyxl') as writer:
        
        # ===== Sheet 1: Best Results (只包含消融实验) =====
        rows = []
        
        # 添加消融实验 (最佳 epoch)
        for exp_results in [JEPA_GSC_RESULTS, SAM_GSC_RESULTS]:
            best_epoch, best_metrics = get_best_epoch(exp_results, 'novel_ap50')
            rows.append({
                'Model': exp_results['experiment'],
                'Description': exp_results['description'],
                'Best Epoch': best_epoch,
                'base_ap50': best_metrics['base_ap50'],
                'novel_ap50': best_metrics['novel_ap50'],
                'all_ap50': best_metrics['all_ap50'],
                'bbox_mAP': best_metrics['bbox_mAP'] * 100,
            })
        
        df_summary = pd.DataFrame(rows)
        df_summary.to_excel(writer, sheet_name='Best Results', index=False)
        
        # ===== Sheet 2: All Epochs - JEPA-GSC =====
        jepa_rows = []
        for epoch, metrics in JEPA_GSC_RESULTS['epochs'].items():
            jepa_rows.append({
                'Epoch': epoch,
                'base_ap50': metrics['base_ap50'],
                'novel_ap50': metrics['novel_ap50'],
                'all_ap50': metrics['all_ap50'],
                'bbox_mAP': metrics['bbox_mAP'] * 100,
            })
        df_jepa = pd.DataFrame(jepa_rows)
        df_jepa.to_excel(writer, sheet_name='JEPA-GSC All Epochs', index=False)
        
        # ===== Sheet 3: All Epochs - SAM-GSC =====
        sam_rows = []
        for epoch, metrics in SAM_GSC_RESULTS['epochs'].items():
            sam_rows.append({
                'Epoch': epoch,
                'base_ap50': metrics['base_ap50'],
                'novel_ap50': metrics['novel_ap50'],
                'all_ap50': metrics['all_ap50'],
                'bbox_mAP': metrics['bbox_mAP'] * 100,
            })
        df_sam = pd.DataFrame(sam_rows)
        df_sam.to_excel(writer, sheet_name='SAM-GSC All Epochs', index=False)
        
    print(f"Excel saved to: {output_path}")


def print_summary():
    """打印汇总信息"""
    print("=" * 80)
    print("DeCLIP Ablation Experiments Summary")
    print("=" * 80)
    
    print("\n--- Ablation Experiments (Best by novel_ap50) ---")
    for exp_results in [JEPA_GSC_RESULTS, SAM_GSC_RESULTS]:
        best_epoch, best_metrics = get_best_epoch(exp_results, 'novel_ap50')
        print(f"{exp_results['experiment']} (Epoch {best_epoch}): "
              f"base_ap50={best_metrics['base_ap50']:.2f}, "
              f"novel_ap50={best_metrics['novel_ap50']:.2f}, "
              f"all_ap50={best_metrics['all_ap50']:.2f}")
    
    print("=" * 80)


if __name__ == '__main__':
    import argparse
    parser = argparse.ArgumentParser(description='Summarize DeCLIP ablation experiments')
    parser.add_argument('--output', type=str, 
                        default='/mnt/bn/strategy-mllm-train/user/wangjunjie/code/xiaomoguhzz/DeCLIP_private/ablation_experiments/ablation_summary.xlsx',
                        help='Output Excel file path')
    args = parser.parse_args()
    
    print_summary()
    generate_summary_excel(args.output)