#!/usr/bin/env python3 """Extract resumable multi-view DINOv3 descriptors for the v801 probe. The legacy extractor immediately reduces four views to an element-wise mean and maximum. v801 keeps each view and adds patch maximum/std descriptors so that model selection can decide how much spatially local information to keep. """ from __future__ import annotations import argparse import fcntl import hashlib import json import os import platform import tempfile import time from pathlib import Path import numpy as np import torch from PIL import Image from transformers import AutoImageProcessor, AutoModel from dataset_utils import load_dataset from extract_features import file_sha256, four_views FEATURE_SCHEMA = "dinov3_four_view_cls_patchmean_patchmax_patchstd_v801" DESCRIPTORS = ("cls", "patch_mean", "patch_max", "patch_std") def atomic_savez(path: Path, **arrays: np.ndarray) -> None: path.parent.mkdir(parents=True, exist_ok=True) descriptor, temporary = tempfile.mkstemp( prefix=f".{path.name}.", suffix=".tmp.npz", dir=path.parent ) os.close(descriptor) temporary_path = Path(temporary) try: np.savez_compressed(temporary_path, **arrays) os.replace(temporary_path, path) finally: temporary_path.unlink(missing_ok=True) def atomic_write_json(path: Path, value: dict[str, object]) -> None: path.parent.mkdir(parents=True, exist_ok=True) descriptor, temporary = tempfile.mkstemp( prefix=f".{path.name}.", suffix=".tmp", dir=path.parent ) temporary_path = Path(temporary) try: with os.fdopen(descriptor, "w", encoding="utf-8") as handle: json.dump(value, handle, ensure_ascii=False, indent=2) handle.write("\n") handle.flush() os.fsync(handle.fileno()) os.replace(temporary_path, path) finally: temporary_path.unlink(missing_ok=True) def cache_identity(dataset: object, model_path: Path) -> tuple[str, dict[str, str], str]: model_digests = { name: file_sha256(model_path / name) for name in ("model.safetensors", "config.json", "preprocessor_config.json") if (model_path / name).is_file() } if "model.safetensors" not in model_digests: raise FileNotFoundError(f"missing model weights under {model_path}") source = hashlib.sha256() source.update(FEATURE_SCHEMA.encode()) source.update(json.dumps(dataset.classes, ensure_ascii=False).encode()) source.update(np.asarray(dataset.labels, dtype=np.uint8).tobytes()) for name, path in zip(dataset.file_names, dataset.image_paths): source.update(str(name).encode()) source.update(bytes.fromhex(file_sha256(path))) source_digest = source.hexdigest() identity = hashlib.sha256() identity.update(FEATURE_SCHEMA.encode()) identity.update(json.dumps(model_digests, sort_keys=True).encode()) identity.update(source_digest.encode()) return identity.hexdigest(), model_digests, source_digest def legacy_features(descriptors: np.ndarray) -> np.ndarray: """Reconstruct the v800 4096-D representation from richer descriptors.""" cls_patch = descriptors[:, :, :2].reshape(len(descriptors), 4, -1) vectors = np.concatenate((cls_patch.mean(axis=1), cls_patch.max(axis=1)), axis=1) norms = np.linalg.norm(vectors, axis=1, keepdims=True) return (vectors / np.maximum(norms, 1e-12)).astype(np.float32) def main() -> None: parser = argparse.ArgumentParser() parser.add_argument("--dataset", type=Path, required=True) parser.add_argument("--model", type=Path, required=True) parser.add_argument("--output", type=Path, required=True) parser.add_argument("--batch-size", type=int, default=2) parser.add_argument("--threads", type=int, default=4) parser.add_argument("--num-shards", type=int, default=1) parser.add_argument("--shard-index", type=int, default=0) args = parser.parse_args() if args.num_shards < 1 or not 0 <= args.shard_index < args.num_shards: raise ValueError("shard-index must be in [0, num-shards)") dataset = load_dataset(args.dataset) torch.set_num_threads(args.threads) processor = AutoImageProcessor.from_pretrained(args.model, local_files_only=True) model = AutoModel.from_pretrained(args.model, local_files_only=True).eval() register_tokens = int(model.config.num_register_tokens) hidden_size = int(model.config.hidden_size) started = time.time() key, model_digests, source_digest = cache_identity(dataset, args.model) parts_dir = args.output.parent / f"{args.output.stem}.parts.{key[:16]}" parts_dir.mkdir(parents=True, exist_ok=True) starts = list(range(0, len(dataset.image_paths), args.batch_size)) assigned_starts = [ start for batch_index, start in enumerate(starts) if batch_index % args.num_shards == args.shard_index ] for start in assigned_starts: paths = dataset.image_paths[start : start + args.batch_size] end = start + len(paths) part_path = parts_dir / f"{start:04d}-{end:04d}.npz" if part_path.is_file(): with np.load(part_path) as part: if str(part["cache_key"]) != key: raise ValueError(f"cached provenance mismatch: {part_path}") if not np.array_equal(part["file_names"], dataset.file_names[start:end]): raise ValueError(f"cached filename mismatch: {part_path}") expected = (len(paths), 4, len(DESCRIPTORS), hidden_size) if part["view_descriptors"].shape != expected: raise ValueError(f"cached descriptor shape mismatch: {part_path}") print(f"DINOv3 v801 {end}/{len(dataset.image_paths)} (cached)", flush=True) continue views: list[Image.Image] = [] for path in paths: with Image.open(path) as image: views.extend(four_views(image)) inputs = processor( images=views, return_tensors="pt", do_resize=False, do_center_crop=False ) with torch.inference_mode(): hidden = model(**inputs).last_hidden_state patches = hidden[:, 1 + register_tokens :] descriptors = torch.stack( ( hidden[:, 0], patches.mean(dim=1), patches.amax(dim=1), patches.std(dim=1, unbiased=False), ), dim=1, ).reshape(len(paths), 4, len(DESCRIPTORS), hidden_size) atomic_savez( part_path, cache_key=np.asarray(key), file_names=dataset.file_names[start:end], view_descriptors=descriptors.cpu().numpy().astype(np.float32), ) print(f"DINOv3 v801 {end}/{len(dataset.image_paths)}", flush=True) missing = [] for start in starts: end = min(start + args.batch_size, len(dataset.image_paths)) if not (parts_dir / f"{start:04d}-{end:04d}.npz").is_file(): missing.append(start) if missing: print( f"Shard {args.shard_index}/{args.num_shards} complete; " f"waiting for {len(missing)} parts.", flush=True, ) return lock_path = args.output.with_suffix(args.output.suffix + ".lock") lock_path.parent.mkdir(parents=True, exist_ok=True) with lock_path.open("w") as lock: fcntl.flock(lock, fcntl.LOCK_EX) blocks = [] for start in starts: end = min(start + args.batch_size, len(dataset.image_paths)) part_path = parts_dir / f"{start:04d}-{end:04d}.npz" with np.load(part_path) as part: if str(part["cache_key"]) != key: raise ValueError(f"cached provenance mismatch: {part_path}") if not np.array_equal(part["file_names"], dataset.file_names[start:end]): raise ValueError(f"cached filename mismatch: {part_path}") blocks.append(part["view_descriptors"].astype(np.float32)) descriptors = np.concatenate(blocks) expected = (len(dataset.image_paths), 4, len(DESCRIPTORS), hidden_size) if descriptors.shape != expected or not np.isfinite(descriptors).all(): raise ValueError(f"unexpected or non-finite descriptor tensor: {descriptors.shape}") base = legacy_features(descriptors) atomic_savez( args.output, file_names=dataset.file_names, features=base, view_descriptors=descriptors, labels=dataset.labels, classes=np.asarray(dataset.classes), cache_key=np.asarray(key), feature_schema=np.asarray(FEATURE_SCHEMA), descriptor_names=np.asarray(DESCRIPTORS), source_digest=np.asarray(source_digest), model_file_sha256_json=np.asarray(json.dumps(model_digests, sort_keys=True)), ) metadata = { "feature_type": FEATURE_SCHEMA, "cache_key": key, "source_digest": source_digest, "model_path": str(args.model.resolve()), "model_sha256": model_digests["model.safetensors"], "model_file_sha256": model_digests, "samples": len(dataset.image_paths), "legacy_feature_shape": list(base.shape), "view_descriptor_shape": list(descriptors.shape), "descriptor_names": list(DESCRIPTORS), "hidden_size": hidden_size, "num_register_tokens": register_tokens, "patch_size": int(model.config.patch_size), "batch_size": args.batch_size, "threads": args.threads, "num_shards": args.num_shards, "seconds": round(time.time() - started, 3), "python": platform.python_version(), "torch": torch.__version__, } atomic_write_json(args.output.with_suffix(".json"), metadata) print(json.dumps(metadata, ensure_ascii=False, indent=2), flush=True) if __name__ == "__main__": main()