data_proccess.py

import json
import os
import random
from collections import defaultdict

# =================配置=================
INPUT_ROOT = '.'
OUTPUT_ROOT = './processed_data'
RATIO_BOX = 0.6  # 训练集：60% 带框，40% 盲测

DATASETS = {
    'data_sft': 'data_sft.jsonl',
    'data_grpo': 'data_grpo.jsonl',
    'data_test': 'data_test.jsonl'
}

def build_prompt_boxed(objects):
    """Type A: 带框识别 (多框合并或单框挑战)"""
    valid_objs = [o for o in objects if 'bbox' in o and len(o['bbox']) == 4]
    if not valid_objs: return None
    
    # 策略：如果物体 <= 8个，全部放入；否则只选最小的那个 (最难的小目标)
    if len(valid_objs) <= 8:
        boxes_str = "\n".join([f"<box>[{o['bbox'][0]},{o['bbox'][1]},{o['bbox'][2]},{o['bbox'][3]}]</box>" for o in valid_objs])
        labels = [o['category'] for o in valid_objs]
        user_msg = f"<image>请识别以下每个框内的物体，按顺序输出 JSON 列表：\n{boxes_str}"
        asst_msg = json.dumps(labels, ensure_ascii=False)
    else:
        # 选面积最小的
        target = min(valid_objs, key=lambda x: (x['bbox'][2]-x['bbox'][0]) * (x['bbox'][3]-x['bbox'][1]))
        b = target['bbox']
        user_msg = f"<image><box>[{b[0]},{b[1]},{b[2]},{b[3]}]</box>请识别框内物体，只输出类别名。"
        asst_msg = target['category']
        
    return [{"role": "user", "content": user_msg}, {"role": "assistant", "content": asst_msg}]

def build_prompt_blind(objects):
    """Type B: 全图盲测 (无框)"""
    cats = sorted(list(set([o['category'] for o in objects if 'category' in o])))
    if not cats: return None
    
    user_msg = "<image>请检测图中所有目标物体，特别是细小的物体。不要遗漏，以 JSON 列表输出类别。"
    asst_msg = json.dumps(cats, ensure_ascii=False)
    
    return [
        {"role": "user", "content": user_msg}, 
        {"role": "assistant", "content": asst_msg}
    ], cats # 返回 prompts 和 真值列表

def process_file(mode, out_name):
    in_path = os.path.join(INPUT_ROOT, mode, 'labels.jsonl')
    out_path = os.path.join(OUTPUT_ROOT, out_name)
    if not os.path.exists(in_path): return

    print(f"处理中: {mode} ...")
    
    # 1. 按类别分组
    cat_pool = defaultdict(list)
    with open(in_path, 'r', encoding='utf-8') as f:
        for line in f:
            data = json.loads(line)
            # 提取类别名 (从路径或第一个物体)
            c_name = data['image'].split('/')[1] if '/' in data['image'] else data['objects'][0]['category']
            cat_pool[c_name].append(data)

    final_data = []
    
    # 2. 遍历每个类别进行采样
    for c_name, items in cat_pool.items():
        random.shuffle(items)
        total = len(items)
        
        if mode == 'data_test':
            # 【测试集】100% 盲测
            for item in items:
                res = build_prompt_blind(item['objects'])
                if res:
                    msgs, gt_cats = res
                    final_data.append({
                        "image": item['image'],
                        "messages": msgs,
                        "ground_truth": gt_cats # 保留真值用于计算 Recall
                    })
        else:
            # 【训练集】60% 带框 + 40% 盲测
            split_idx = int(total * RATIO_BOX)
            
            # Type A (带框)
            for item in items[:split_idx]:
                msgs = build_prompt_boxed(item['objects'])
                if msgs:
                    final_data.append({"image": item['image'], "messages": msgs})
            
            # Type B (盲测)
            for item in items[split_idx:]:
                res = build_prompt_blind(item['objects'])
                if res:
                    msgs, _ = res
                    final_data.append({"image": item['image'], "messages": msgs})

    # 3. 全局打乱并保存
    random.shuffle(final_data)
    os.makedirs(OUTPUT_ROOT, exist_ok=True)
    
    with open(out_path, 'w', encoding='utf-8') as f:
        for item in final_data:
            f.write(json.dumps(item, ensure_ascii=False) + '\n')
            
    print(f"  -> 完成: {out_path} (共 {len(final_data)} 条)")

if __name__ == '__main__':
    random.seed(42)
    print("开始数据处理 (60/40 策略 & 测试集全盲测)...")
    for mode, name in DATASETS.items():
        process_file(mode, name)
    print("全部完成！")