tools/cap_process.py

# import os
# import torch
# import json
# import re
# from PIL import Image
# import torch.nn.functional as F
# from safetensors.torch import load_file
# from huggingface_hub import hf_hub_download

# import sys
# sys.path.append("/gpfs/home/ym621/UniPointMap")
# import open_clip

# # ---------------------------
# # Helpers
# # ---------------------------
# def load_safetensor_from_hf(repo_id, filename, repo_type="dataset"):
#     cached_path = hf_hub_download(
#         repo_id=repo_id,
#         filename=filename,
#         revision='7bb7c7f3d379c5145bb06d2cf0949c66ac9a2c4e',
#         repo_type=repo_type,
#         local_files_only=True
#     )
#     return load_file(cached_path)

# def load_json(data_path: str):
#     with open(data_path, "r", encoding="utf-8") as f:
#         return json.load(f)

# def load_jsonl(path):
#     data = []
#     with open(path, "r", encoding="utf-8") as f:
#         for line in f:
#             if line.strip():
#                 data.append(json.loads(line))
#     return data

# # ---------------------------
# # Load CLIP model
# # ---------------------------
# device = "cuda" if torch.cuda.is_available() else "cpu"
# model, _, preprocess = open_clip.create_model_and_transforms(
#     'ViT-B-16', pretrained='datacomp_xl_s13b_b90k'
# )
# tokenizer = open_clip.get_tokenizer('ViT-B-16')
# model = model.to(device).eval()

# # ---------------------------
# # Preload reference captions
# # ---------------------------
# scannet_data = load_jsonl('/gpfs/home/ym621/UniPointMap/PointMapVerse/existing_datasets/ScanNet/annotations/scannet_caption_per_view.jsonl')
# arkitscenes_data = load_jsonl('/gpfs/home/ym621/UniPointMap/PointMapVerse/existing_datasets/Arkitscenes/annotations/arkitscenes_caption_per_view.jsonl')
# rscan_data = load_jsonl('/gpfs/home/ym621/UniPointMap/PointMapVerse/existing_datasets/3RScan/annotations/3rscan_caption_per_view.jsonl')

# org_data = {}
# cur_scan_id = ''
# for idx, data in enumerate([scannet_data, arkitscenes_data, rscan_data]):
#     if idx == 0:
#         root = 'light_scannet'
#     elif idx == 1:
#         root = 'light_arkitscenes'
#     else:
#         root = 'light_3rscan'

#     local_idx = 0
#     for item in data:
#         if item['scan_id'] != cur_scan_id:
#             cur_scan_id = item['scan_id']
#             local_idx = 0
#         scan_id = f"{root}/{item['scan_id']}_{local_idx}"
#         org_data[scan_id] = item['utterance'].split('.')
#         local_idx += 1

# # ---------------------------
# # Caching safetensors
# # ---------------------------
# safetensor_cache = {}
# def get_image_from_safetensor(image_path, idx):
#     if image_path not in safetensor_cache:
#         safetensor_cache[image_path] = load_safetensor_from_hf(
#             'MatchLab/PointMapVerse', image_path
#         )
#     return safetensor_cache[image_path]['color_images'][idx]

# # ---------------------------
# # Process captions
# # ---------------------------
# caption_dir = "../captions"
# captions = [f for f in os.listdir(caption_dir) if f.endswith('.json')]

# filtered_captions = {}
# count, total_count = 0, 0

# for cap in captions:
#     cap_path = os.path.join(caption_dir, cap)
#     caption_data = load_json(cap_path)

#     for k, v in caption_data.items():
#         image_path = f"{'_'.join(k.split('_')[:-1])}.safetensors"
#         idx = int(k.split('_')[-1])

#         # --- load + preprocess image ---
#         img_tensor = get_image_from_safetensor(image_path, idx)
#         img_tensor = img_tensor.cpu().numpy()
#         pil_img = Image.fromarray(img_tensor.astype("uint8")).convert("RGB")
#         image = preprocess(pil_img).unsqueeze(0).to(device)

#         with torch.no_grad():
#             image_features = model.encode_image(image)
#             image_features = F.normalize(image_features, dim=-1)

#         # --- clean captions ---
#         if "1." in v:
#             v = v.split("1.", 1)[-1].strip()
#             v = "1." + v
#         if not v.startswith('1.'):
#             v = ["An image showing an indoor scene."]
#             count += 1
#         else:
#             v = re.split(r'\s*\d+\.\s*', v)
#             v = [c.strip().replace('*', '') for c in v if c.strip()]
#             if len(v) < 4:
#                 v = ["An image showing an indoor scene."]
#                 count += 1

#         # --- combine old + new captions ---
#         old_v = org_data.get(k, [])
#         all_v = old_v + v

#         # --- encode captions ---
#         with torch.no_grad():
#             text_tokens = tokenizer(all_v).to(device)
#             text_features = model.encode_text(text_tokens)
#             text_features = F.normalize(text_features, dim=-1)

#             sims = (image_features @ text_features.T).squeeze(0)  # [num_caps]

#         # --- sort captions (fast torch.topk instead of sorted) ---
#         topk_vals, topk_idx = torch.topk(sims, k=len(all_v))
#         # print(topk_vals)
#         sorted_captions = [all_v[i] for i in topk_idx.tolist()]
#         # print(sorted_captions)

#         filtered_captions[k] = sorted_captions
#         total_count += 1
#         if total_count % 50 == 0:
#             print(f"Processed {total_count} files...")

# # ---------------------------
# # Save results
# # ---------------------------
# output_path = os.path.join(caption_dir, "filtered_captions_sorted.json")
# with open(output_path, "w", encoding="utf-8") as f:
#     json.dump(filtered_captions, f, indent=4)

# print(f'Total captions not starting with "1.": {count} out of {total_count} captions.')
# print(f"Sorted captions saved to {output_path}")


# ---------------------------
# Save results as JSONL
# ---------------------------
import os
import json
from transformers import AutoTokenizer

# ---------------------------
# Paths
# ---------------------------
caption_dir = "../captions"
filtered_json_path = os.path.join(caption_dir, "filtered_captions_sorted.json")

# Output files for each dataset
output_paths = {
    "scannet": os.path.join(caption_dir, "filtered_captions_scannet.jsonl"),
    "arkitscenes": os.path.join(caption_dir, "filtered_captions_arkitscenes.jsonl"),
    "3rscan": os.path.join(caption_dir, "filtered_captions_3rscan.jsonl"),
}

# ---------------------------
# Load filtered captions
# ---------------------------
with open(filtered_json_path, "r", encoding="utf-8") as f:
    filtered_captions = json.load(f)

print(f"Loaded {len(filtered_captions)} scan entries.")

# ---------------------------
# Setup tokenizer (bert-base-uncased)
# ---------------------------
tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")

# ---------------------------
# Open three output files
# ---------------------------
files = {k: open(path, "w", encoding="utf-8") for k, path in output_paths.items()}
line_ids = {"scannet": 1, "arkitscenes": 1, "3rscan": 1}

# ---------------------------
# Convert and write entries
# ---------------------------
for k, sorted_captions in filtered_captions.items():
    # Determine dataset type
    if k.startswith("light_scannet"):
        dataset = "scannet"
    elif k.startswith("light_arkitscenes"):
        dataset = "arkitscenes"
    elif k.startswith("light_3rscan"):
        dataset = "3rscan"
    else:
        continue  # skip unknown dataset keys

    image_path = f"{'_'.join(k.split('_')[:-1])}.safetensors"
    scan_id = "_".join(k.split("_")[:-1]).split("/")[-1]  # e.g. scene0000_00

    # Clean and join top-5 captions
    sorted_captions = [cap.replace('.', '').strip() for cap in sorted_captions]

    entry = {
        "item_id": f"{dataset}_train_{line_ids[dataset]:06d}",
        "scan_id": scan_id,
        "utterance": sorted_captions,
        "safetensors_path": image_path,
    }

    # Write to the correct file
    files[dataset].write(json.dumps(entry) + "\n")
    line_ids[dataset] += 1

# ---------------------------
# Close files
# ---------------------------
for f in files.values():
    f.close()

print(f"✅ Saved entries to:")
for k, path in output_paths.items():
    print(f"  {k}: {path} ({line_ids[k]-1} entries)")