sample_box.py

import json
import random
from pathlib import Path
from collections import defaultdict

def sample_prompt_boxes(bboxes, k):
    """
    从 bboxes 中随机抽取 k 个作为 prompt
    如果 bboxes 数量 < k，则全部返回
    """
    if len(bboxes) <= k:
        return bboxes.copy(), []
    else:
        # 随机抽取 k 个作为 prompt
        prompt_indices = random.sample(range(len(bboxes)), k)
        prompt_boxes = [bboxes[i] for i in prompt_indices]
        # 剩余的作为 target（可选）
        target_boxes = [bboxes[i] for i in range(len(bboxes)) if i not in prompt_indices]
        return prompt_boxes, target_boxes

def format_bbox_for_prompt(bbox):
    """
    将 bbox 格式化为字符串，用于 prompt
    格式: [x1, y1, x2, y2]
    """
    return f"[{bbox[0]}, {bbox[1]}, {bbox[2]}, {bbox[3]}]"

def process_jsonl_for_fewshot(input_jsonl_path, output_dir, k_values=[1, 2, 4]):
    """
    处理原始 JSONL，生成 few-shot 格式的数据
    
    Args:
        input_jsonl_path: 输入的 JSONL 文件路径
        output_dir: 输出目录
        k_values: few-shot 的 k 值列表
    """
    output_dir = Path(output_dir)
    output_dir.mkdir(parents=True, exist_ok=True)
    
    # 为每个 k 值创建输出文件
    output_files = {}
    stats = {k: {'total_images': 0, 'total_prompts': 0, 'insufficient_count': 0} 
             for k in k_values}
    
    for k in k_values:
        output_files[k] = open(output_dir / f"fewshot_k{k}.jsonl", 'w', encoding='utf-8')
    
    # 读取原始 JSONL
    with open(input_jsonl_path, 'r', encoding='utf-8') as f:
        for line_num, line in enumerate(f, 1):
            data = json.loads(line.strip())
            
            image_path = data['image']
            image_rel_path = data.get('image_path', image_path)
            width = data['width']
            height = data['height']
            bboxes = data['bboxes']
            
            # 为每个 k 值生成数据
            for k in k_values:
                # 随机抽取 prompt boxes
                prompt_boxes, remaining_boxes = sample_prompt_boxes(bboxes, k)
                
                # 统计信息
                stats[k]['total_images'] += 1
                stats[k]['total_prompts'] += len(prompt_boxes)
                if len(bboxes) < k:
                    stats[k]['insufficient_count'] += 1
                
                # 构建 few-shot 数据结构
                fewshot_data = {
                    "image": image_path,
                    "image_path": image_rel_path,
                    "width": width,
                    "height": height,
                    "k": k,  # 记录 k 值
                    "total_gt_boxes": len(bboxes),
                    "prompt_boxes": prompt_boxes,  # 用于 few-shot 的框
                    "all_boxes": bboxes,  # 所有 GT 框（用于评估）
                    "prompt_boxes_count": len(prompt_boxes),
                    "remaining_boxes_count": len(remaining_boxes)
                }
                
                # 写入对应的文件
                output_files[k].write(json.dumps(fewshot_data, ensure_ascii=False) + '\n')
    
    # 关闭所有文件
    for f in output_files.values():
        f.close()
    
    # 打印统计信息
    print("\n" + "="*60)
    print("Few-shot 数据生成完成！")
    print("="*60)
    for k in k_values:
        print(f"\nK={k}:")
        print(f"  - 总图片数: {stats[k]['total_images']}")
        print(f"  - 总 prompt 框数: {stats[k]['total_prompts']}")
        print(f"  - 平均 prompt 框数: {stats[k]['total_prompts']/stats[k]['total_images']:.2f}")
        print(f"  - 框数不足 k 的图片数: {stats[k]['insufficient_count']}")
        print(f"  - 输出文件: {output_dir}/fewshot_k{k}.jsonl")
    
    return stats

def convert_to_conversation_format(fewshot_jsonl_path, output_conversation_path):
    """
    将 few-shot JSONL 转换为对话格式
    """
    with open(fewshot_jsonl_path, 'r', encoding='utf-8') as infile, \
         open(output_conversation_path, 'w', encoding='utf-8') as outfile:
        
        for line in infile:
            data = json.loads(line.strip())
            
            # 格式化 prompt boxes
            prompt_boxes_str = ", ".join([format_bbox_for_prompt(box['bbox']) for box in data['prompt_boxes']])
            
            # 构建对话
            conversation = {
                "image": data['image'],
                "conversations": [
                    {
                        "from": "human",
                        "value": f"Please detect all objects belonging to the same category as the boxes [{prompt_boxes_str}] in the image."
                    },
                    {
                        "from": "gpt",
                        "value": json.dumps({
                            "category": data['all_boxes'][0]['category'] if data['all_boxes'] else "unknown",
                            "bboxes": [box['bbox'] for box in data['all_boxes']]
                        })
                    }
                ],
                "metadata": {
                    "k": data['k'],
                    "total_gt_boxes": data['total_gt_boxes'],
                    "prompt_boxes_count": data['prompt_boxes_count']
                }
            }
            
            outfile.write(json.dumps(conversation, ensure_ascii=False) + '\n')
    
    print(f"对话格式已保存到: {output_conversation_path}")

def main():
    # 输入文件路径
    input_jsonl = "/home/disk2/hjl/ICL_QWEN/ICL_benchmark/dataset.jsonl"
    
    # 输出目录
    output_dir = "/home/disk2/hjl/ICL_QWEN/ICL_benchmark/fewshot_data"
    
    # 生成 few-shot JSONL 文件（k=1,2,4）
    print("步骤1: 生成 few-shot 结构化数据...")
    stats = process_jsonl_for_fewshot(input_jsonl, output_dir, k_values=[1, 2, 4])
    
    # 可选：转换为对话格式
    print("\n步骤2: 转换为对话格式...")
    for k in [1, 2, 4]:
        fewshot_file = output_dir / f"fewshot_k{k}.jsonl"
        conversation_file = output_dir / f"conversation_k{k}.jsonl"
        if fewshot_file.exists():
            convert_to_conversation_format(fewshot_file, conversation_file)
    
    print("\n所有处理完成！")

if __name__ == "__main__":
    # 设置随机种子，保证可重复性
    random.seed(42)
    main()