tools/data_gen/update_split.py

#!/usr/bin/env python3
"""
更新 zones_data_mixed.json 的 split 信息:
1. 从 zones_data_coarse.json 中读取已有的 split 映射
2. matterport3d 数据全部设为 train
3. 从 3d-front 中额外随机选择 500 条作为 test
"""

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


def main():
    parser = argparse.ArgumentParser(description="Update split in zones_data_mixed.json")
    parser.add_argument("--input", type=str, 
                       default="/home/v-meiszhang/amlt-project/InternScenes/data/zones_data_mixed.json",
                       help="Input zones_data_mixed.json file")
    parser.add_argument("--reference", type=str,
                       default="/home/v-meiszhang/amlt-project/InternScenes/tools/data_gen/zones_data_coarse.json",
                       help="Reference file for split mapping")
    parser.add_argument("--output", type=str,
                       default="/home/v-meiszhang/amlt-project/InternScenes/data/zones_data_mixed.json",
                       help="Output file path")
    parser.add_argument("--3dfront-test-size", type=int, default=500,
                       help="Number of 3d-front samples to mark as test")
    parser.add_argument("--seed", type=int, default=42,
                       help="Random seed")
    
    args = parser.parse_args()
    
    # 读取参考文件获取 split 映射
    print(f"Loading reference file: {args.reference}")
    with open(args.reference, 'r') as f:
        ref_data = json.load(f)
    
    # 构建 scene_path -> split 的映射
    split_mapping = {}
    for entry in ref_data.get('data', []):
        scene_path = entry.get('scene_path', '')
        split = entry.get('split', 'train')
        if scene_path:
            split_mapping[scene_path] = split
    
    print(f"Loaded {len(split_mapping)} split mappings from reference")
    
    # 统计参考文件的 split 分布
    ref_split_counts = defaultdict(int)
    for split in split_mapping.values():
        ref_split_counts[split] += 1
    print(f"Reference split distribution: {dict(ref_split_counts)}")
    
    # 读取要更新的文件
    print(f"\nLoading input file: {args.input}")
    with open(args.input, 'r') as f:
        mixed_data = json.load(f)
    
    # 统计数据集分布
    dataset_counts = defaultdict(int)
    for entry in mixed_data.get('data', []):
        scene_path = entry.get('scene_path', '')
        if scene_path:
            dataset = scene_path.split('/')[0]
            dataset_counts[dataset] += 1
    print(f"Dataset distribution: {dict(dataset_counts)}")
    
    # 收集 3d-front 的 train 数据索引（用于后续选择 test）
    front3d_train_indices = []
    
    # 更新 split
    updated_count = 0
    matterport_count = 0
    
    for i, entry in enumerate(mixed_data.get('data', [])):
        scene_path = entry.get('scene_path', '')
        if not scene_path:
            continue
        
        dataset = scene_path.split('/')[0]
        
        # matterport3d 全部设为 train
        if dataset == 'matterport3d':
            entry['split'] = 'train'
            matterport_count += 1
            continue
        
        # 其他数据集使用参考文件的 split
        if scene_path in split_mapping:
            old_split = entry.get('split', 'train')
            new_split = split_mapping[scene_path]
            if old_split != new_split:
                entry['split'] = new_split
                updated_count += 1
        
        # 收集 3d-front 的 train 索引
        if dataset == '3d-front' and entry.get('split') == 'train':
            front3d_train_indices.append(i)
    
    print(f"\nUpdated {updated_count} entries based on reference split")
    print(f"Set {matterport_count} matterport3d entries to train")
    
    # 从 3d-front 中随机选择额外的 test
    random.seed(args.seed)
    test_size = getattr(args, '3dfront_test_size', 500)
    
    if len(front3d_train_indices) >= test_size:
        selected_test_indices = random.sample(front3d_train_indices, test_size)
        for idx in selected_test_indices:
            mixed_data['data'][idx]['split'] = 'test'
        print(f"Selected {test_size} 3d-front entries as additional test")
    else:
        print(f"Warning: Only {len(front3d_train_indices)} 3d-front train entries available, selecting all as test")
        for idx in front3d_train_indices:
            mixed_data['data'][idx]['split'] = 'test'
    
    # 统计最终的 split 分布
    final_split_counts = defaultdict(int)
    dataset_split_counts = defaultdict(lambda: defaultdict(int))
    
    for entry in mixed_data.get('data', []):
        split = entry.get('split', 'train')
        scene_path = entry.get('scene_path', '')
        dataset = scene_path.split('/')[0] if scene_path else 'unknown'
        
        final_split_counts[split] += 1
        dataset_split_counts[dataset][split] += 1
    
    print(f"\nFinal split distribution: {dict(final_split_counts)}")
    print("\nPer-dataset split distribution:")
    for dataset, splits in sorted(dataset_split_counts.items()):
        print(f"  {dataset}: {dict(splits)}")
    
    # 更新 metadata
    if 'metadata' in mixed_data:
        mixed_data['metadata']['split_statistics'] = {
            'total': dict(final_split_counts),
            'per_dataset': {k: dict(v) for k, v in dataset_split_counts.items()}
        }
    
    # 保存结果
    print(f"\nSaving to: {args.output}")
    with open(args.output, 'w') as f:
        json.dump(mixed_data, f, ensure_ascii=False, indent=2)
    
    print("Done!")


if __name__ == "__main__":
    main()