code-HFD / benchmark /review.py
Tinghong-Ye22's picture
Add files using upload-large-folder tool
fb0074a verified
Raw History Blame Contribute Delete
3.07 kB
"""Compact review of frozen visual prefixes and exact task/command ledgers."""
import argparse
from collections import defaultdict
import json
from pathlib import Path
import sys
import textwrap
import numpy as np
from PIL import Image, ImageDraw, ImageFont
from common import ROOT, WORK, write_json
sys.path[:0]=[str(ROOT),str(ROOT/'src')]
def main():
p=argparse.ArgumentParser()
p.add_argument('--cohort',type=Path,default=WORK/'outputs/sessions2048')
p.add_argument('--per-task',type=int,default=4)
a=p.parse_args()
path=a.cohort/'manifest.json'
if not path.exists():path=a.cohort/'selection_progress.json'
rows=json.loads(path.read_text())['rows']
groups=defaultdict(list)
for r in rows:groups[r['task']].append(r)
dest=a.cohort/'review';dest.mkdir(exist_ok=True)
fonts=list(Path('/usr/share/fonts').rglob('*CJK*.ttc'))
font=ImageFont.truetype(str(fonts[0]) if fonts else '/usr/share/fonts/truetype/dejavu/DejaVuSans.ttf',18)
manifest=[]
from psi.data.seed.video_overlay import decode_video
for task, candidates in groups.items():
# Spread cases by duration, rather than always the first/easiest ones.
candidates.sort(key=lambda r:r['reference_frames'])
choices=np.unique(np.linspace(0,len(candidates)-1,min(a.per_task,len(candidates))).astype(int))
for i in choices:
r=candidates[i];o=r['overlay'];qend=o['question_end']
if r['visual_kind']=='video':
times=np.linspace(0,qend,6)
pixels,pts=decode_video(o['video_path'],times,end_seconds=o['camera_duration_seconds'],side=360)
frames=[Image.fromarray(v) for v in pixels]
else:
frames=[Image.open(o['image_path']).convert('RGB').resize((360,360))];pts=[0.]
board=Image.new('RGB',(1080,1100),'white');draw=ImageDraw.Draw(board)
line=10
lines=[f"{r['sample_id']} | {task} | {r['turn_count']} turns | {r['voice']}",
f"Question ends {qend:.2f}s; session {r['reference_frames']/50:.2f}s. All frames below precede question end.",
'Q: '+o['question'],'A: '+o['answer'],
'Motion: '+' -> '.join(s['caption'] for s in r['segments'])]
for text in lines:
for part in textwrap.wrap(text,95):
draw.text((10,line),part,fill='black',font=font);line+=24
top=max(260,line+5)
for k,(frame,t) in enumerate(zip(frames,pts)):
x=(k%3)*360;y=top+(k//3)*390
board.paste(frame,(x,y));draw.text((x+8,y+363),f'PTS {t:.2f}s',fill='black',font=font)
out=dest/(r['sample_id']+'.jpg');board.save(out,quality=92)
manifest.append(dict(sample_id=r['sample_id'],task=task,contact_sheet=str(out),
question=o['question'],answer=o['answer'],question_end=qend,pts=list(map(float,pts))))
write_json(dest/'cases.json',manifest)
print(json.dumps(manifest,ensure_ascii=False,indent=2))
if __name__=='__main__':main()