| | import json |
| | import argparse |
| | import os |
| | import base64 |
| | import io |
| | import numpy as np |
| | import imageio.v3 as iio |
| | from PIL import Image |
| | from collections import defaultdict |
| | from concurrent.futures import ThreadPoolExecutor, as_completed |
| | from tqdm import tqdm |
| |
|
| | def parse_args(): |
| | parser = argparse.ArgumentParser(description="Build MS-Swift SFT Dataset (BBox + Start/End Points)") |
| | parser.add_argument("--traj_path", type=str, required=True, help="Path to processed_trajectories.jsonl") |
| | parser.add_argument("--task_path", type=str, required=True, help="Path to task.jsonl") |
| | parser.add_argument("--origin_bbox_path", type=str, required=True, help="Path to original bbox.jsonl") |
| | parser.add_argument("--output_path", type=str, required=True, help="Output path for swift_sft.jsonl") |
| | parser.add_argument("--workers", type=int, default=8, help="Number of worker threads") |
| | return parser.parse_args() |
| |
|
| | def load_auxiliary_data(task_path, bbox_path): |
| | print("Loading auxiliary metadata (Tasks & Labels)...") |
| | meta_map = defaultdict(dict) |
| | |
| | print(f"Reading tasks from {task_path}...") |
| | with open(task_path, 'r', encoding='utf-8') as f: |
| | for line in f: |
| | if not line.strip(): continue |
| | try: |
| | item = json.loads(line) |
| | v_path = item.get('video_path') |
| | task = item.get('object_bbox', {}).get('task', "") |
| | if v_path and task: |
| | meta_map[v_path]['task'] = task |
| | except Exception: |
| | continue |
| |
|
| | print(f"Reading labels from {bbox_path}...") |
| | count_labels = 0 |
| | with open(bbox_path, 'r', encoding='utf-8') as f: |
| | for line in f: |
| | if not line.strip(): continue |
| | try: |
| | item = json.loads(line) |
| | v_path = item.get('video_path') |
| | label = item.get('object_bbox', {}).get('label', "target_object") |
| | |
| | if v_path: |
| | meta_map[v_path]['label'] = label |
| | count_labels += 1 |
| | except Exception: |
| | continue |
| | |
| | print(f"Metadata loaded. Found labels for {count_labels} entries.") |
| | return meta_map |
| |
|
| | def numpy_to_base64(img_np): |
| | if img_np.dtype != np.uint8: |
| | img_np = (img_np).astype(np.uint8) |
| | |
| | img = Image.fromarray(img_np) |
| | buffer = io.BytesIO() |
| | img.save(buffer, format="JPEG", quality=90) |
| | img_bytes = buffer.getvalue() |
| | img_base64 = base64.b64encode(img_bytes).decode('utf-8') |
| | return f"data:image/jpeg;base64,{img_base64}" |
| |
|
| | def process_single_video(video_path, samples, meta_info): |
| | results = [] |
| | task_desc = meta_info.get('task', "") |
| | obj_label = meta_info.get('label', "target_object") |
| | |
| | if not task_desc: |
| | return [] |
| |
|
| | try: |
| | if not os.path.exists(video_path): |
| | return [] |
| |
|
| | samples.sort(key=lambda x: x['frame_idx']) |
| | |
| | for sample in samples: |
| | frame_idx = sample['frame_idx'] |
| | bbox = sample['bbox'] |
| | full_traj = sample['traj'] |
| | |
| | if not full_traj or len(full_traj) == 0: |
| | continue |
| |
|
| | start_point = [(bbox[0] + bbox[2]) // 2, (bbox[1] + bbox[3]) // 2] |
| | end_point = full_traj[-1] |
| | |
| | try: |
| | frame = iio.imread(video_path, index=frame_idx, plugin="pyav") |
| | image_base64 = numpy_to_base64(frame) |
| | |
| | response_data = { |
| | "box_2d": bbox, |
| | "label": obj_label, |
| | "start_point": start_point, |
| | "end_point": end_point |
| | } |
| | |
| | response_str = "```json\n" + json.dumps(response_data, ensure_ascii=False) + "\n```" |
| | |
| | user_content = ( |
| | f"<image>\n" |
| | f"Task: {task_desc}\n" |
| | f"Detect the target object and predict its trajectory start point and end point. " |
| | f"Output the result in JSON format with the following keys: \"box_2d\", \"label\", \"start_point\", and \"end_point\"." |
| | ) |
| | swift_item = { |
| | "messages": [ |
| | { |
| | "role": "user", |
| | "content": user_content |
| | }, |
| | { |
| | "role": "assistant", |
| | "content": response_str |
| | } |
| | ], |
| | "images": [image_base64] |
| | } |
| | results.append(swift_item) |
| | |
| | except IndexError as e: |
| | |
| | continue |
| | except Exception as e: |
| | |
| | continue |
| | |
| | except Exception as e: |
| | print(f"Error opening video {video_path}: {e}") |
| | return [] |
| | |
| | return results |
| |
|
| | def main(): |
| | args = parse_args() |
| | |
| | meta_map = load_auxiliary_data(args.task_path, args.origin_bbox_path) |
| | |
| | print(f"Loading trajectory data from {args.traj_path}...") |
| | video_groups = defaultdict(list) |
| | total_traj_samples = 0 |
| | |
| | with open(args.traj_path, 'r', encoding='utf-8') as f: |
| | for line in f: |
| | if not line.strip(): continue |
| | item = json.loads(line) |
| | v_path = item['video_path'] |
| | |
| | if v_path in meta_map and 'task' in meta_map[v_path]: |
| | video_groups[v_path].append(item) |
| | total_traj_samples += 1 |
| | |
| | print(f"Grouped {total_traj_samples} samples into {len(video_groups)} videos.") |
| | |
| | print(f"Starting processing with {args.workers} workers...") |
| | final_dataset = [] |
| | |
| | with ThreadPoolExecutor(max_workers=args.workers) as executor: |
| | future_to_video = {} |
| | for video_path, samples in video_groups.items(): |
| | meta_info = meta_map[video_path] |
| | future = executor.submit(process_single_video, video_path, samples, meta_info) |
| | future_to_video[future] = video_path |
| | |
| | for future in tqdm(as_completed(future_to_video), total=len(video_groups), desc="Processing Videos"): |
| | res = future.result() |
| | if res: |
| | final_dataset.extend(res) |
| | |
| | print(f"Saving {len(final_dataset)} samples to {args.output_path}...") |
| | os.makedirs(os.path.dirname(os.path.abspath(args.output_path)), exist_ok=True) |
| | |
| | with open(args.output_path, 'w', encoding='utf-8') as f: |
| | for item in final_dataset: |
| | f.write(json.dumps(item, ensure_ascii=False) + '\n') |
| | |
| | print("Done.") |
| |
|
| | if __name__ == "__main__": |
| | main() |
| |
|
| | ''' |
| | |
| | python sft_data/v3/build_v3_swift_data.py \ |
| | --traj_path /storage/v-xiangxizheng/dataset/InternData-M1/annotaions/traj_outputs/agilex_top_video_000k_034k_bbox_traj_results.jsonl \ |
| | --task_path /storage/v-xiangxizheng/dataset/InternData-M1/annotaions/agilex_top_video_000k_034k_task.jsonl \ |
| | --origin_bbox_path /storage/v-xiangxizheng/dataset/InternData-M1/annotaions/agilex_top_video_000k_034k_bbox.jsonl \ |
| | --output_path sft_data/v3/agilex_top_video_000k_034k_bbox_traj_sft_data_2pts.jsonl \ |
| | --workers 64 & |
| | |
| | python sft_data/v3/build_v3_swift_data.py \ |
| | --traj_path /storage/v-xiangxizheng/dataset/InternData-M1/annotaions/traj_outputs/franka_base_video_000k_100k_bbox_traj_results.jsonl \ |
| | --task_path /storage/v-xiangxizheng/dataset/InternData-M1/annotaions/franka_base_video_000k_100k_task.jsonl \ |
| | --origin_bbox_path /storage/v-xiangxizheng/dataset/InternData-M1/annotaions/franka_base_video_000k_100k_bbox.jsonl \ |
| | --output_path sft_data/v3/franka_base_video_000k_100k_bbox_traj_sft_data_2pts.jsonl \ |
| | --workers 64 & |
| | |
| | python sft_data/v3/build_v3_swift_data.py \ |
| | --traj_path /storage/v-xiangxizheng/dataset/InternData-M1/annotaions/traj_outputs/franka_base_video_100k_200k_bbox_traj_results.jsonl \ |
| | --task_path /storage/v-xiangxizheng/dataset/InternData-M1/annotaions/franka_base_video_100k_200k_task.jsonl \ |
| | --origin_bbox_path /storage/v-xiangxizheng/dataset/InternData-M1/annotaions/franka_base_video_100k_200k_bbox.jsonl \ |
| | --output_path sft_data/v3/franka_base_video_100k_200k_bbox_traj_sft_data_2pts.jsonl \ |
| | --workers 64 & |
| | |
| | |
| | hf upload sftv3 . --repo-type dataset --private & |
| | ''' |
