| """Like-for-like scoring of saved val predictions (v3 vs v2work) on identical rows. Read-only on v2work/v3.""" |
| import sys, json, numpy as np, polars as pl |
| sys.path.insert(0, '/workspace/huawei/v4') |
| from common import report |
| from scipy.stats import spearmanr |
| V2='/workspace/huawei/v2work'; V3='/workspace/huawei/v3' |
| out={} |
| for cut, v3pred, v2preds in [(25, 'full25.txt', {'v2work_notime_eft_ne (combo_val25_notime.npy, 300r)':'combo_val25_notime.npy', |
| 'v2work_notime_x2 (04:26)':'combo_val25_notime_x2.npy', |
| 'v2work_screen_base_25pct':'scr_base_c25.npy'}), |
| (26, 'full26_lr05.txt', {'v2work_screen_base_25pct (scr_base_c26.npy)':'scr_base_c26.npy'})]: |
| |
| v3 = pl.read_parquet(f'{V3}/feat/val{cut}.parquet', columns=['rowid','eid','clicked','tm_stale']) |
| v3 = v3.join(pl.read_parquet(f'/workspace/huawei/parallel/v3_val{cut}_pred.parquet'), on='rowid', how='left') |
| assert v3['p_v3'].null_count()==0 |
| v2 = pl.read_parquet(f'{V2}/feat_v3/val{cut}.parquet', columns=['rowid','eid','clicked','tm_stale']) |
| for nm,f in v2preds.items(): |
| p=np.load(f'{V2}/{f}'); assert len(p)==v2.height, (f,len(p),v2.height); v2=v2.with_columns(pl.Series(nm,p)) |
| rs3=set(v3['rowid'].to_list()); rs2=set(v2['rowid'].to_list()) |
| j = v3.join(v2.drop('eid','tm_stale').rename({'clicked':'clicked2'}), on='rowid', how='inner') |
| res={'rows_v3':v3.height,'rows_v2work':v2.height,'rows_common':j.height,'only_v3':len(rs3-rs2),'only_v2work':len(rs2-rs3), |
| 'label_mismatch':int((j['clicked']!=j['clicked2']).sum())} |
| res['v3 ('+v3pred+')']=report(j, j['p_v3'].to_numpy()) |
| for nm in v2preds: |
| res[nm]=report(j, j[nm].to_numpy()) |
| res[nm]['spearman_vs_v3']=round(float(spearmanr(j[nm].to_numpy(), j['p_v3'].to_numpy()).statistic),4) |
| out[f'val{cut}']=res |
| print(json.dumps({f'val{cut}':res},indent=1),flush=True) |
| json.dump(out,open('/workspace/huawei/parallel/score_val.json','w'),indent=1) |
|
|
