| from __future__ import annotations |
| import sys,time,json |
| import numpy as np,pandas as pd |
| sys.path.insert(0,'/mnt/data') |
| import msmarco_best_tail_core as b |
| import msmarco_full_search_uniform1m as m |
| ROOT=m.ROOT; WORK=m.WORK |
| LEX=[0.0,0.5,1.0,1.5,2.0,2.5,3.0,4.0,5.0,7.5,10.0] |
| SEM=[0.0,0.05,0.1,0.25] |
| tr=pd.read_csv(ROOT/'train.tsv',sep='\t',usecols=['query-id']); uq=np.unique(tr['query-id'].to_numpy()); rng=np.random.default_rng(20260815); ids=[str(x) for x in rng.choice(uq,size=1000,replace=False)]; del tr |
| texts=m.load_query_texts(ids); qrels=m.qrels_from_tsv(ROOT/'train.tsv',ids,positive_only=True) |
| _=b.prepare(texts[ids[0]]) |
| runs={(ll,ss):{} for ll in LEX for ss in SEM}; times=[] |
| for z,qid in enumerate(ids): |
| t=time.perf_counter(); p=b.prepare(texts[qid]); times.append((time.perf_counter()-t)*1000) |
| if p is None: |
| for k in runs:runs[k][qid]=[] |
| continue |
| docs=p['cand_docs'][:b.P]; ts=p['cand_tail'][:b.P]; lx=p['lex'][:b.P]; sm=p['sem'][:b.P] |
| zt=m.zscore(ts); zl=m.zscore(lx); zs=m.zscore(sm) |
| for ll in LEX: |
| base=zt+np.float32(ll)*zl |
| for ss in SEM: |
| fin=base+np.float32(ss)*zs; oo=np.argsort(fin)[::-1][:100]; runs[(ll,ss)][qid]=[int(x) for x in docs[oo]] |
| if (z+1)%100==0: print('q',z+1,'median',float(np.median(times)),flush=True) |
| rows=[] |
| for ll in LEX: |
| for ss in SEM: |
| met=m.eval_run(runs[(ll,ss)],qrels); row={'lambda_lex':ll,'lambda_sem':ss,**met}; rows.append(row); print('W',ll,ss,met,flush=True) |
| rows.sort(key=lambda r:(r['nDCG@10'],r['MRR@10'],r['R@100']),reverse=True) |
| out={'protocol':'new tail fixed gamma=.25 lambda_M=.125 P=2000 h=0; final weights tuned on deterministic 1000 TRAIN validation','rows':rows,'best':rows[0],'median_prepare_ms':float(np.median(times))} |
| json.dump(out,open(WORK/'final_weight_sweep.json','w'),indent=2); print('BEST',rows[0],flush=True) |
|
|
