"""Cached frozen DINO, original images, and new image-resolution expert gaze.""" from pathlib import Path import csv,hashlib import numpy as np import torch from torch.utils.data import Dataset from image_data import load_image from heatmap import make_image_heatmap def load_torch(path):return torch.load(path,map_location='cpu',weights_only=False,mmap=True) def fixation_records(manifest): path=Path(manifest['root'])/manifest['fixation_csv'];gaze={} with path.open(encoding='utf-8-sig') as stream: for row in csv.DictReader(stream):gaze.setdefault(Path(row['IMAGE']).stem,[]).append(row) return gaze,hashlib.sha256(path.read_bytes()).hexdigest() class FeatureImageDataset(Dataset): def __init__(self,cache,images,ids,manifest,sigma_px,augment=False): self.ids=ids;self.augment=augment self.rows=[load_torch(Path(cache)/(name+'.pt')) for name in ids] self.images=[load_image(images[name],512,row['native'].shape[-2:]) for name,row in zip(ids,self.rows)] gaze,self.fixation_csv_sha256=fixation_records(manifest);self.heatmaps=[];counts=[] for name in ids: records=sorted(gaze[name],key=lambda row:float(row['CURRENT_FIX_INDEX'])) xy=np.asarray([[float(row['CURRENT_FIX_X']),float(row['CURRENT_FIX_Y'])] for row in records],np.float32) duration=np.asarray([float(row['CURRENT_FIX_DURATION']) for row in records],np.float32) assert len(records)>0 and np.isfinite(xy).all() and np.isfinite(duration).all() assert ((xy>=0)&(xy<=1)).all() and (duration>=0).all(),name # Same per-image duration transform as the trusted prior cache producer. duration=(duration-duration.min())/(duration.max()-duration.min()+1e-8) if duration.max()>duration.min() else np.ones_like(duration) h=make_image_heatmap(xy,duration,image_size=(512,512),sigma_px=sigma_px) assert h.shape==(1,512,512) and h.max()>0 self.heatmaps.append(h);counts.append(len(records)) self.gaze_audit={'n':len(ids),'sigma_px':sigma_px,'shape':[1,512,512], 'min_fixations':min(counts),'max_fixations':max(counts),'mean_fixations':float(np.mean(counts)), 'source':'original CSV coordinates, not upsampled cached heatmaps', 'duration_transform':'per-image min-max; constant durations -> all ones', 'fixation_csv_sha256':self.fixation_csv_sha256} def __len__(self):return len(self.ids) def __getitem__(self,index): source=self.rows[index] row={k:source[k] for k in ['id','features','mask','native']} row.update(image=self.images[index],gaze=self.heatmaps[index]) if self.augment: dims=[] if torch.rand(())<.5:dims.append(-1) if torch.rand(())<.5:dims.append(-2) if dims: row['features']=[torch.flip(f,dims) for f in row['features']] for key in ['image','gaze','mask']:row[key]=torch.flip(row[key],dims) return row def collate(rows): return {'features':[torch.stack([r['features'][i] for r in rows]) for i in range(4)], **{k:torch.stack([r[k] for r in rows]) for k in ['image','gaze','mask']}, 'id':[r['id'] for r in rows],'native':[r['native'] for r in rows]}