Datasets:
File size: 2,864 Bytes
9126e0d | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 | """Reconstruct extremal controlled examples from stored predictions, without GPU.
Selection is deliberately outcome-based and is disclosed in the figure caption.
It never alters evaluation metrics. RGB and masks are actual experiment files.
"""
from pathlib import Path
import json,gzip,zipfile,io,base64
import numpy as np
from PIL import Image
from repair import repaired_maps,score_mask
ROOT=Path(__file__).resolve().parents[1]
def main():
manifest={r['id']:r for r in map(json.loads,(ROOT/'data/manifest.jsonl').read_text().splitlines())}
out=ROOT/'results/qualitative';out.mkdir(parents=True,exist_ok=True);examples=[]
for model in ['clipseg','groundedsam']:
rows=[json.loads(l) for l in gzip.open(ROOT/'results/analysis'/(model+'_evaluated.jsonl.gz'),'rt')]
candidates=[r for r in rows if r['domain']=='controlled' and r['split']=='test' and r['corruption']=='clean' and r['mode']=='relation' and r['target_count']>0]
candidates.sort(key=lambda r:(r['scores']['cacp']['iou']-r['scores']['frozen']['iou'],r['id']))
for kind,r in [('gain',candidates[-1]),('loss',candidates[0])]:
d=manifest[r['id']];z=np.load(ROOT/r['cache_path']);maps={q:z['p%d'%i].astype(np.float32) for i,q in enumerate(z['texts'].tolist())}
pred,flags=repaired_maps(d['query'],maps,.5 if model=='clipseg' else .25)
gt=np.asarray(Image.open(ROOT/d['gt_path']))>0;key=model+'_'+kind;assets={}
for method in ['frozen','cacp']:
assert score_mask(pred[method],gt)==r['scores'][method]
Image.open(ROOT/d['image_path']).save(out/(key+'_rgb.png'));assets['rgb']=key+'_rgb.png'
for method,mask in [('gt',gt),('frozen',pred['frozen']),('cacp',pred['cacp'])]:
Image.fromarray(mask.astype(np.uint8)*255).save(out/(key+'_'+method+'.png'));assets[method]=key+'_'+method+'.png'
examples.append(dict(model=model,kind=kind,record=d,flags=flags,scores={m:r['scores'][m] for m in ['frozen','cacp']},assets=assets))
meta={'selection':'Largest gain and largest loss in mask IoU per pipeline among clean, held-out, positive controlled relation records; ties broken by record ID. These are extremal illustrations, not representative samples.','examples':examples}
(out/'qualitative_examples.json').write_text(json.dumps(meta,indent=2))
buffer=io.BytesIO()
with zipfile.ZipFile(buffer,'w',zipfile.ZIP_DEFLATED) as z:
for f in sorted(out.iterdir()):
if f.suffix in ['.json','.png']:z.write(f,f.name)
(ROOT/'results/qualitative_bundle.b64').write_text(base64.b64encode(buffer.getvalue()).decode())
print(json.dumps({'bytes':len(buffer.getvalue()),'examples':[(r['model'],r['kind'],r['record']['id'],r['scores']['frozen']['iou'],r['scores']['cacp']['iou']) for r in examples]}))
if __name__=='__main__':main()
|