Datasets:
File size: 2,824 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 | """Paired warmed latency, same images/GPU, balanced plan order, no GT use."""
from pathlib import Path
import sys,json,time,os
import numpy as np
from PIL import Image
ROOT=Path(__file__).resolve().parents[1];from project_paths import legacy_root;OLD=legacy_root(ROOT);sys.path.insert(0,str(OLD/'code'))
from infer import ClipSeg,GroundedSam
import phase1_repair as p
from r2_repair import primary_plan,primary_mask,repair
import torch
torch.set_num_threads(4)
rows=[json.loads(s) for s in (ROOT/'data/rich_synthetic.jsonl').read_text().splitlines()]
chosen=[r for r in rows if r['render']=='rich' and r['mode']=='relational' and r['endpoint']==0][:12]
output=[]
for model,cls in [('clipseg',ClipSeg),('groundedsam',GroundedSam)]:
engine=cls()
image=Image.open(ROOT/chosen[0]['image_path']).convert('RGB');q=p.parse_query(chosen[0]['query']);engine.predict(image,p.query_plan(chosen[0]['query'])[1]);torch.cuda.synchronize()
for i,r in enumerate(chosen):
image=Image.open(ROOT/r['image_path']).convert('RGB');q=p.parse_query(r['query'])
plans={'frozen':[q.original],'sfap':primary_plan(r['query']),'cacp':p.query_plan(r['query'])[1]}
for rep in range(3):
order=['frozen','sfap','cacp'];shift=(i+rep)%3;order=order[shift:]+order[:shift]
for variant in order:
torch.cuda.synchronize();start=time.perf_counter();maps,stats=engine.predict(image,plans[variant]);torch.cuda.synchronize();forward=time.perf_counter()-start
maps={k:v.astype(np.float16).astype(np.float32) for k,v in maps.items()}
start=time.perf_counter()
if variant=='frozen':mask=p.native(maps[q.original],engine.threshold)
elif variant=='sfap':mask=primary_mask(r['query'],maps,engine.threshold)
else:mask=repair(r['query'],maps,engine.threshold)[0]['cacp']
cpu=time.perf_counter()-start
if variant=='cacp':assert np.array_equal(primary_mask(r['query'],maps,engine.threshold),repair(r['query'],maps,engine.threshold)[0]['safe_anchor'])
output.append(dict(model=model,image=r['image_path'],scene_id=r['scene_id'],repeat=rep,order=order,variant=variant,queries=len(plans[variant]),forward_seconds=forward,postprocess_seconds=cpu,total_seconds=forward+cpu))
print('TIMING',model,i+1,len(chosen),flush=True)
del engine;torch.cuda.empty_cache()
dest=ROOT/'results/paired_timing.json';dest.write_text(json.dumps(dict(gpu=torch.cuda.get_device_name(0),slurm_job_id=os.getenv('SLURM_JOB_ID'),records=output,warmup='One complete four-query plan per model before timing',cache='Disabled',repeats=3,selection='First twelve rich relational scene records, text/ID selection; no label or score selection'),indent=2))
print('PAIRED_TIMING_COMPLETE',flush=True)
|