| from __future__ import annotations |
| import sys,time,json |
| from pathlib import Path |
| import numpy as np |
| from numba import njit,prange,set_num_threads |
| sys.path.insert(0,'/mnt/data') |
| import msmarco_full_search_uniform1m as m |
| ROOT=m.ROOT; WORK=m.WORK; OUT=WORK/'structural_fusion'; OUT.mkdir(exist_ok=True); set_num_threads(5) |
| z=np.load(WORK/'amplitude_diag'/'fixed_eta1_pools.npz',allow_pickle=False); qids=[str(x) for x in z['qids'].tolist()]; valid=z['valid'].astype(np.int32); docs=z['docs']; idx=m.FullIndex(); M=m.M |
| texts=m.load_query_texts(qids) |
| @njit(parallel=True,cache=False) |
| def extras(dd,ip,ids,lexvec,dl,avgdl): |
| n=len(dd); cnt=np.zeros(n,np.float32); raw=np.zeros(n,np.float32); lf=np.zeros(n,np.float32) |
| for z in prange(n): |
| d=int(dd[z]); a=int(ip[d]); b=int(ip[d+1]); c=0.0; r=0.0 |
| for k in range(a,b): |
| t=int(ids[k]); v=lexvec[t] |
| if v>0: c+=1.; r+=v |
| cnt[z]=c; raw[z]=r; lf[z]=(1.0-m.LENGTH_B)+m.LENGTH_B*(float(dl[d])/avgdl) |
| return cnt,raw,lf |
| _=extras(np.array([0],np.uint32),idx.sup_ip,idx.sup_ids,np.zeros(M,np.float32),idx.dl,idx.avgdl) |
| COUNT=np.zeros_like(z['tail'],np.float32); RAW=np.zeros_like(COUNT); LF=np.zeros_like(COUNT); QTERMS=np.zeros(len(qids),np.int16); times=[] |
| for i,qid in enumerate(qids): |
| k=int(valid[i]); |
| if not k:continue |
| q=idx.query_vec(texts[qid]); QTERMS[i]=len(q.indices); lv=np.zeros(M,np.float32); lv[q.indices]=idx.idf[q.indices] |
| t=time.perf_counter(); c,r,l=extras(docs[i,:k],idx.sup_ip,idx.sup_ids,lv,idx.dl,idx.avgdl); times.append((time.perf_counter()-t)*1000); COUNT[i,:k]=c; RAW[i,:k]=r; LF[i,:k]=l |
| if (i+1)%200==0: print(i+1,float(np.median(times)),flush=True) |
| np.savez_compressed(OUT/'coordination_features.npz',qids=np.asarray(qids),valid=valid,qcount=COUNT,rawlex=RAW,lenfac=LF,qterms=QTERMS) |
| print('DONE',float(np.median(times)),float(np.percentile(times,95)),flush=True) |
|
|
