"""One record is one image/query condition; a unit is one native item or pair.""" from pathlib import Path import json import math import numpy as np from PIL import Image def decode_rle(rle): """Decode COCO compressed or uncompressed RLE to a bool [height, width] mask.""" counts = rle['counts'] if isinstance(counts, str): decoded = [] pos = 0 while pos < len(counts): value = 0 shift = 0 while True: if pos >= len(counts): raise ValueError('Truncated COCO RLE') code = ord(counts[pos]) - 48 pos += 1 if not 0 <= code <= 63: raise ValueError('Invalid COCO RLE character') value |= (code & 31) << shift shift += 5 if not code & 32: if code & 16: value |= -1 << shift break if len(decoded) > 2: value += decoded[-2] decoded.append(value) counts = decoded height, width = map(int, rle['size']) counts = [int(v) for v in counts] if height <= 0 or width <= 0 or any(v < 0 for v in counts) or sum(counts) != height * width: raise ValueError('RLE runs do not match image dimensions') flat = np.zeros(height * width, dtype=bool) offset = 0 for idx, length in enumerate(counts): if idx % 2: flat[offset:offset + length] = True offset += length return flat.reshape((height, width), order='F') class Dataset: """Read normalized catalogs and frozen selections without dataset-specific code. iter_samples() defaults to selected primary conditions. For HalluSegBench, get_pair() exposes the four image/query combinations. References are kept separate from the image/question returned by model_input(). """ def __init__(self, root='prepared_data'): self.root = Path(root).resolve() self.report = json.loads((self.root / 'build_report.json').read_text()) if self.report['schema_version'] != 1 or self.report['status'] != 'complete': raise ValueError('Unsupported or incomplete prepared dataset') self._records = {} self._datasets = {} self._pairs = {} self._selections = {} for name in self.report['datasets']: path = self.root / 'manifests' / (name + '.jsonl') samples = [json.loads(line) for line in path.read_text().splitlines() if line] self._datasets[name] = samples self._selections[name] = json.loads((self.root / 'selections' / (name + '.json')).read_text()) for sample in samples: if sample['sample_id'] in self._records: raise ValueError('Duplicate sample ID: ' + sample['sample_id']) self._records[sample['sample_id']] = sample if sample['pair_id']: self._pairs.setdefault(sample['pair_id'], {})[sample['condition']] = sample def get(self, sample_id): return self._records[sample_id] def get_pair(self, sample_or_pair_id): pair_id = (sample_or_pair_id.get('pair_id') if isinstance(sample_or_pair_id, dict) else sample_or_pair_id) if pair_id in self._records: pair_id = self._records[pair_id]['pair_id'] if not pair_id: raise ValueError('This sample has no counterfactual pair') return dict(self._pairs[pair_id]) def iter_samples(self, dataset=None, subset='selected', condition='primary'): """subset: selected/test/development/all; condition: primary/all/exact name.""" if subset not in {'selected', 'test', 'development', 'all'}: raise ValueError('Unknown subset: ' + subset) names = [dataset] if dataset else list(self._datasets) for name in names: selection = self._selections[name] allowed = None if subset == 'all' else set(selection[subset + '_unit_ids']) for sample in self._datasets[name]: if allowed is not None and sample['unit_id'] not in allowed: continue if condition == 'primary' and not sample['is_primary']: continue if condition not in {'primary', 'all'} and sample['condition'] != condition: continue yield sample def _path(self, relative): path = (self.root / relative).resolve() if not path.is_relative_to(self.root): raise ValueError('Asset path escapes prepared dataset') return path def load_image(self, sample): with Image.open(self._path(sample['image']['path'])) as image: return image.convert('RGB') def model_input(self, sample): """Only these fields go to a model; no answers, target labels, or boxes.""" return {'image': self.load_image(sample), 'question': sample['question']} def reference(self, sample): return sample['reference'] def boxes(self, sample): return [r['bbox_xyxy'] for r in sample['reference']['regions']] def mask(self, sample, region_index=None): """True target masks; returns None for bbox-only supervision. Empty annotated regions return an all-false mask; inspect target_status to distinguish absence, no match, and no annotated target. Bbox-only positives and unavailable annotations return None. """ regions = sample['reference']['regions'] height, width = sample['image']['height'], sample['image']['width'] if sample['reference']['region_kind'] == 'unavailable': return None if not regions: return np.zeros((height, width), dtype=bool) if any(r['mask'] is None for r in regions): return None use = regions if region_index is None else [regions[region_index]] merged = np.zeros((height, width), dtype=bool) for region in use: spec = region['mask'] if spec['encoding'] == 'coco_rle': mask = decode_rle(spec) elif spec['encoding'] == 'binary_png': with Image.open(self._path(spec['path'])) as image: mask = np.asarray(image.convert('L')) > 0 else: raise ValueError('Unknown mask encoding') if mask.shape != merged.shape: raise ValueError('Mask size does not match this image') merged |= mask return merged def region_mask(self, sample): """Rasterized annotated regions for attention: mask if present, else bbox. Read reference.region_kind when aggregating scores; bbox rasterization is not a segmentation annotation. Floating boxes use pixel-center tests. """ mask = self.mask(sample) if mask is not None: return mask if sample['reference']['region_kind'] == 'unavailable': return None h, w = sample['image']['height'], sample['image']['width'] out = np.zeros((h, w), dtype=bool) for x1, y1, x2, y2 in self.boxes(sample): left, right = max(0, math.ceil(x1 - .5)), min(w, math.ceil(x2 - .5)) top, bottom = max(0, math.ceil(y1 - .5)), min(h, math.ceil(y2 - .5)) out[top:bottom, left:right] = True return out