| from __future__ import annotations |
|
|
| import json |
| import random |
| from dataclasses import dataclass |
| from pathlib import Path |
|
|
| from PIL import Image |
| from torch.utils.data import Dataset |
|
|
|
|
| @dataclass(frozen=True) |
| class ImageRecord: |
| image_id: str |
| path: Path |
|
|
|
|
| def load_manifest(path: str | Path, data_root: str | Path | None = None) -> list[ImageRecord]: |
| manifest = Path(path) |
| root = Path(data_root) if data_root else manifest.parent |
| rows = json.loads(manifest.read_text(encoding="utf-8")) |
| if not isinstance(rows, list): |
| raise ValueError(f"{manifest}: expected a JSON list") |
| records = [] |
| for index, row in enumerate(rows): |
| if not isinstance(row, dict) or not {"image_id", "image_path"} <= row.keys(): |
| raise ValueError(f"{manifest}: invalid row {index}") |
| relative = Path(row["image_path"]) |
| candidates = (relative,) if relative.is_absolute() else ( |
| root / relative, |
| manifest.parent / relative, |
| manifest.parent.parent / relative, |
| manifest.parent.parent.parent / relative, |
| ) |
| resolved = next((candidate for candidate in candidates if candidate.is_file()), None) |
| if resolved is None: |
| raise FileNotFoundError(f"Cannot resolve {relative} from {manifest}") |
| records.append(ImageRecord(str(row["image_id"]), resolved.resolve())) |
| if len({record.image_id for record in records}) != len(records): |
| raise ValueError(f"{manifest}: duplicate image IDs") |
| return records |
|
|
|
|
| def pseudo_split(records: list[ImageRecord], seed: int = 42, gallery_size: int = 1000, query_size: int = 200): |
| if len(records) < gallery_size: |
| raise ValueError(f"Need at least {gallery_size} records") |
| shuffled = records.copy() |
| random.Random(seed).shuffle(shuffled) |
| gallery = shuffled[:gallery_size] |
| query_indices = random.Random(seed + 1).sample(range(gallery_size), query_size) |
| queries = [(gallery[index], seed * 100_000 + order) for order, index in enumerate(query_indices)] |
| return shuffled[gallery_size:], gallery, queries |
|
|
|
|
| class ImageDataset(Dataset): |
| def __init__(self, records: list[ImageRecord], transform): |
| self.records, self.transform = records, transform |
|
|
| def __len__(self): |
| return len(self.records) |
|
|
| def __getitem__(self, index): |
| record = self.records[index] |
| with Image.open(record.path) as image: |
| return record.image_id, self.transform(image.convert("RGB")) |
|
|
|
|
| class TrainingDataset(Dataset): |
| def __init__(self, records: list[ImageRecord], method: str, transforms): |
| self.records, self.method, self.transforms = records, method, transforms |
|
|
| def __len__(self): |
| return len(self.records) |
|
|
| def __getitem__(self, index): |
| with Image.open(self.records[index].path) as source: |
| image = source.convert("RGB") |
| clean = self.transforms.clean(image) |
| sample = {"clean": clean} |
| if self.method in {"simclr", "joint"}: |
| sample["view1"] = self.transforms.contrastive(image) |
| sample["view2"] = self.transforms.contrastive(image) |
| if self.method in {"autoencoder", "joint"}: |
| sample["degraded"] = self.transforms.degraded(image) |
| if self.method in {"rotation", "joint"}: |
| rotated, label = self.transforms.rotation(image) |
| sample["rotated"], sample["rotation_label"] = rotated, label |
| return sample |
|
|
|
|
