data_split.py

import json
import os
import shutil
import random
from collections import defaultdict

# =================配置=================
COCO_ROOT = '.'
OUT_ROOT = '.'  # 输出目录

# 数量配置
NUM_SFT = 100
NUM_GRPO = 200
NUM_TEST = 200

# =================工具函数：坐标转换=================
def convert_coco_to_qwen_bbox(coco_bbox, img_width, img_height):
    """
    将 COCO 格式 [x, y, w, h] 转换为 Qwen-VL 格式 [x1, y1, x2, y2] (归一化到 0-1000)
    """
    x, y, w, h = coco_bbox
    
    # 1. 转为左上角 + 右下角 (像素)
    x1 = x
    y1 = y
    x2 = x + w
    y2 = y + h
    
    # 2. 归一化到 0-1000
    # 使用 max/min 防止因浮点误差或标注超出边界导致数值越界
    norm_x1 = max(0, min(1000, int((x1 / img_width) * 1000)))
    norm_y1 = max(0, min(1000, int((y1 / img_height) * 1000)))
    norm_x2 = max(0, min(1000, int((x2 / img_width) * 1000)))
    norm_y2 = max(0, min(1000, int((y2 / img_height) * 1000)))
    
    # 3. 返回格式化列表 [x1, y1, x2, y2]
    # 注意：这里直接返回整数列表，JSON 序列化后就是 [530,562,576,650] 这种格式
    return [norm_x1, norm_y1, norm_x2, norm_y2]

# =================1. 中英文映射表 (COCO 80类)=================
COCO_CN_MAP = {
    "person": "人", "bicycle": "自行车", "car": "汽车", "motorcycle": "摩托车",
    "airplane": "飞机", "bus": "公交车", "train": "火车", "truck": "卡车",
    "boat": "船", "traffic light": "交通灯", "fire hydrant": "消防栓",
    "stop sign": "停车标志", "parking meter": "停车计时器", "bench": "长椅",
    "bird": "鸟", "cat": "猫", "dog": "狗", "horse": "马", "sheep": "羊",
    "cow": "牛", "elephant": "大象", "bear": "熊", "zebra": "斑马",
    "giraffe": "长颈鹿", "backpack": "背包", "umbrella": "雨伞",
    "handbag": "手提包", "tie": "领带", "suitcase": "行李箱",
    "frisbee": "飞盘", "skis": "滑雪板", "snowboard": "单板滑雪",
    "sports ball": "运动球", "kite": "风筝", "baseball bat": "棒球棒",
    "baseball glove": "棒球手套", "skateboard": "滑板", "surfboard": "冲浪板",
    "tennis racket": "网球拍", "bottle": "瓶子", "wine glass": "酒杯",
    "cup": "杯子", "fork": "叉子", "knife": "刀", "spoon": "勺子",
    "bowl": "碗", "banana": "香蕉", "apple": "苹果", "sandwich": "三明治",
    "orange": "橘子", "broccoli": "西兰花", "carrot": "胡萝卜",
    "hot dog": "热狗", "pizza": "披萨", "donut": "甜甜圈", "cake": "蛋糕",
    "chair": "椅子", "couch": "沙发", "potted plant": "盆栽", "bed": "床",
    "dining table": "餐桌", "toilet": "马桶", "tv": "电视", "laptop": "笔记本电脑",
    "mouse": "鼠标", "remote": "遥控器", "keyboard": "键盘", "cell phone": "手机",
    "microwave": "微波炉", "oven": "烤箱", "toaster": "烤面包机", "sink": "水槽",
    "refrigerator": "冰箱", "book": "书", "clock": "时钟", "vase": "花瓶",
    "scissors": "剪刀", "teddy bear": "泰迪熊", "hair drier": "吹风机",
    "toothbrush": "牙刷"
}

def get_cn_name(en_name):
    return COCO_CN_MAP.get(en_name, en_name)

# =================2. 确定选中的 50 个类别=================
def get_selected_ids():
    stats = defaultdict(list)
    cat_map = {} 
    
    for split in ['train', 'val']:
        path = os.path.join(COCO_ROOT, 'annotations', f'instances_{split}2017.json')
        with open(path) as f: 
            data = json.load(f)
        if split == 'train':
            cat_map = {c['id']: c['name'] for c in data['categories']}
        
        for ann in data['annotations']:
            w, h = ann['bbox'][2], ann['bbox'][3]
            stats[ann['category_id']].append(w * h)
            
    avg_areas = [(cid, sum(areas)/len(areas)) for cid, areas in stats.items()]
    avg_areas.sort(key=lambda x: x[1])
    
    selected = [x[0] for x in (avg_areas[:25] + avg_areas[-25:])]
    return selected, cat_map

selected_ids, id_to_en_name = get_selected_ids()
id_to_cn_name = {cid: get_cn_name(name) for cid, name in id_to_en_name.items()}

print(f"已选定 {len(selected_ids)} 个类别。")

output_file = os.path.join(OUT_ROOT, 'selected_categories.json') 
fine_names = [id_to_cn_name[cid] for cid in selected_ids[:25]]
reg_names = [id_to_cn_name[cid] for cid in selected_ids[25:]]

result_data = {"fine": fine_names, "reg": reg_names}
with open(output_file, 'w', encoding='utf-8') as f:
    json.dump(result_data, f, ensure_ascii=False, indent=2)

# =================3. 加载全量数据索引 (包含图片宽高)=================
img_db = defaultdict(list)  
img_info = {}               # img_id -> {file_name, split, width, height}

for split in ['train', 'val']:
    path = os.path.join(COCO_ROOT, 'annotations', f'instances_{split}2017.json')
    with open(path) as f: 
        data = json.load(f)
    
    # 关键修改：保存图片的 width 和 height
    for img in data['images']:
        img_info[img['id']] = {
            'file_name': img['file_name'], 
            'split': split,
            'width': img['width'],
            'height': img['height']
        }
        
    for ann in data['annotations']:
        if ann['category_id'] in selected_ids:
            img_db[ann['image_id']].append(ann)

# 按类别构建初始图片池
class_pool = defaultdict(list)
for img_id, anns in img_db.items():
    cids = set(ann['category_id'] for ann in anns)
    for cid in cids:
        class_pool[cid].append(img_id)

random.seed(42)

# =================4. 流式构建三个数据集=================
tasks = [
    {'mode': 'data_sft',  'count': NUM_SFT},
    {'mode': 'data_grpo', 'count': NUM_GRPO},
    {'mode': 'data_test', 'count': NUM_TEST}
]

used_images_global = set()

for task in tasks:
    mode = task['mode']
    count_per_cat = task['count']
    
    print(f"\n开始构建 {mode} (每类目标: {count_per_cat})...")
    
    mode_img_root = os.path.join(OUT_ROOT, mode, 'images')
    os.makedirs(mode_img_root, exist_ok=True)
    jsonl_path = os.path.join(OUT_ROOT, mode, 'labels.jsonl')
    
    total_count = 0
    
    with open(jsonl_path, 'w', encoding='utf-8') as jf:
        for cid in selected_ids:
            cat_name = id_to_cn_name[cid] 
            
            full_pool = class_pool[cid]
            available_pool = [img_id for img_id in full_pool if img_id not in used_images_global]
            
            num_needed = min(len(available_pool), count_per_cat)
            if len(available_pool) < count_per_cat and len(available_pool) > 0:
                print(f"  [提示] 类别 '{cat_name}' 剩余可用图片仅 {len(available_pool)} 张。")
            elif len(available_pool) == 0:
                print(f"  [警告] 类别 '{cat_name}' 无可用图片！")
            
            sampled_ids = random.sample(available_pool, num_needed) if num_needed > 0 else []
            
            cat_dir = os.path.join(mode_img_root, cat_name)
            os.makedirs(cat_dir, exist_ok=True)
            
            for img_id in sampled_ids:
                info = img_info[img_id]
                fname = info['file_name']
                split = info['split']
                
                # 获取图片真实宽高用于转换
                w_orig = info['width']
                h_orig = info['height']
                
                src = os.path.join(COCO_ROOT, f'{split}2017', fname)
                dst = os.path.join(cat_dir, fname)
                if not os.path.exists(dst):
                    shutil.copy2(src, dst)
                
                used_images_global.add(img_id)
                
                # 【核心修改】构建 objects 时转换 bbox
                objects = []
                for a in img_db[img_id]:
                    # 1. 转换坐标
                    new_bbox = convert_coco_to_qwen_bbox(a['bbox'], w_orig, h_orig)
                    # 2. 构建对象
                    objects.append({
                        "category": id_to_cn_name[a['category_id']], 
                        "bbox": new_bbox  # 现在是 [x1, y1, x2, y2] (0-1000)
                    })
                
                rel_path = f"images/{cat_name}/{fname}"
                
                # 写入 JSONL
                jf.write(json.dumps({"image": rel_path, "objects": objects}, ensure_ascii=False) + '\n')
                total_count += 1
    
    print(f"  -> {mode} 完成，共生成 {total_count} 张图片。")

print("\n================ 全部完成 ================")
print(f"输出目录: {OUT_ROOT}")