Files
MeltMind/tools/plan_s1_queries.py

47 lines
3.6 KiB
Python

"""Metadata-only S1 cost accounting; never loads NPZ or calls a scorer."""
import argparse
import csv
import hashlib
import json
from pathlib import Path
def plan(summary,folds):
rows=list(csv.DictReader(Path(summary).open()));frows=list(csv.DictReader(Path(folds).open()))
if len(rows)!=24 or len(frows)!=24:raise ValueError('Expected frozen 24-part LOPO')
if len({r['source_key'] for r in rows})!=24:raise ValueError('Sources not distinct')
if [int(r['panel_index']) for r in rows]!=list(range(24)):raise ValueError('Unexpected panel order')
if sorted(int(f['test_panel_index']) for f in frows)!=list(range(24)):raise ValueError('Missing or repeated test part')
if any(r['status']!='PASS' for r in rows):raise ValueError('Incomplete O2 panel')
result=[]
for f in frows:
ti=int(f['test_panel_index']);train=list(map(int,f['train_panel_indices'].split(';')))
if train!=[i for i in range(24) if i!=ti]:raise ValueError('Not the frozen LOPO split')
test=rows[ti]
if f['test_case']!=test['case'] or f['test_source_key']!=test['source_key']:raise ValueError('Fold identity mismatch')
q=int(test['n_ref']);nr=sum(int(rows[i]['n_ref']) for i in train)
result.append(dict(fold_index=int(f['fold_index']),test_case=test['case'],test_source=test['source_key'],
eligible_unique_queries=q,training_distinct_references=nr,configurations=18,
query_2D=q,query_6D=6*q,query_10D=11*q,logical_8NN_queries=18*q,
tie_radius_queries=18*q,total_tree_search_calls=36*q,tree_builds=18,
extra_near_slice_queries=0,extra_query_bootstrap_queries=0,extra_source_bootstrap_queries=0,
worst_case_tie_candidate_distances=18*q*nr,worst_case_tie_distance_components=148*q*nr,
ranked_reference_matrix_bytes=nr*10*8,query_bootstrap_counts_bytes=200*int(test['unique_anchors'])*8))
return dict(scope='METADATA_ONLY; no real rank fitting, scoring, bootstrap evaluation or training',folds=result,
inputs={str(p):hashlib.sha256(Path(p).read_bytes()).hexdigest() for p in [summary,folds]},
numerical_parent='a86a74cbd943e0687ee69fb1cccdc003250113cc',
totals={k:sum(r[k] for r in result) for k in ['eligible_unique_queries','logical_8NN_queries','tie_radius_queries','total_tree_search_calls','tree_builds','worst_case_tie_candidate_distances','worst_case_tie_distance_components']},
configurations='A/B/C once each plus 3 shuffled controls times 5 seeds = 18; dimensions 2:1, 6:6, 10:11.',
bootstrap=dict(query_seed=2026111406,query_replicates=200,source_seed=2026111407,source_replicates=2000,source_index_bytes=2000*24*8),
runtime_and_rss='No real-score timing measured. Matrix sizes are component sizes, not a total RSS bound. Worst-case ties can scan all references.',
proposed_execution='After review and explicit execution scope: one supervised fold at a time, persist/release full fold arrays before next fold; agree wall/RSS limits before starting.')
if __name__=='__main__':
ap=argparse.ArgumentParser();ap.add_argument('--summary',type=Path,required=True);ap.add_argument('--folds',type=Path,required=True);ap.add_argument('--out',type=Path,required=True);args=ap.parse_args()
report=plan(args.summary,args.folds);args.out.mkdir(parents=True,exist_ok=True)
with (args.out/'query_cost.json').open('w') as f:json.dump(report,f,indent=2,allow_nan=False);f.write('\n')
with (args.out/'query_cost.csv').open('w') as f:
writer=csv.DictWriter(f,fieldnames=list(report['folds'][0]),lineterminator='\n');writer.writeheader();writer.writerows(report['folds'])
print(json.dumps(report['totals'],indent=2))