Files
MeltMind/tools/phase3_item_c_data.py
T

58 lines
2.8 KiB
Python

"""Learning-only query cache. No validation/test labels or fitted selection."""
import json
from pathlib import Path
import numpy as np
from phase2_h1_worker_v2 import checked, sha, save
from phase2_h2_data import cached
from s2_custody_common import prediction_inputs
from phase3_item_c_loss import candidates
def require_learning(view, case):
f = view['folds'][0]
if case not in f['learning'] or case in f['validation'] or case in f['test']:
raise PermissionError('Location truth restricted to learning role')
def prepare_queries(view, rows, output):
output = Path(output)
output.mkdir(exist_ok=False)
manifest = {}
by_case = {r['case']: r for r in rows}
if len(by_case) != len(rows):
raise ValueError('Repeated roster case')
for case in view['folds'][0]['learning']:
require_learning(view, case) # before ANY case array access
row = by_case[case]
q = prediction_inputs(row) # existing label-free query recipe
full = cached(view, case, 'raw_ids')
idx = np.searchsorted(full, q['raw_ids'])
if np.any(idx >= len(full)) or not np.array_equal(full[idx], q['raw_ids']):
raise ValueError('Original query/full-node identity')
if q['t_set'] != view['parts'][case]['t_set']:
raise ValueError('Original process-time identity')
# The original scorer uses physical seconds with exact equality ties.
# FP32 residual reconstruction is not an acceptable true-peak source.
digest = view['parts'][case]['output_sha256']['seconds.npy']
with (output/'learning-location-target-access.jsonl').open('a') as f:
f.write(json.dumps(dict(case=case, role='learning', seconds_sha256=digest))+'\n')
seconds = cached(view, case, 'seconds')
d = candidates(full, q['raw_ids'], q['counts'], q['coordinates_over_Rg'], seconds[idx])
p = output/(case+'.npz')
np.savez(p, **d, original_raw_ids=q['raw_ids'], original_counts=q['counts'],
original_anchor_ids=q['ids'], original_near_wall=q['near_wall'],
original_coordinates_over_Rg=q['coordinates_over_Rg'])
manifest[case] = dict(path=str(p), sha256=sha(p), seconds_sha256=digest,
queries_sha256=row['preparation']['output_sha256']['queries.npz'],
unique_candidates=len(d['indices']), true_peak_raw_id=d['true_peak_raw_id'],
true_peak_ties=d['true_peak_ties'], undefined=False)
save(output/'manifest.json', manifest)
return manifest
def query_data(view, case, manifest):
require_learning(view, case)
r = manifest[case]
with np.load(checked(r['path'], r['sha256']), allow_pickle=False) as z:
return {k:z[k].copy() for k in ('indices', 'distances')}