| import json, hashlib, numpy as np, polars as pl |
| from scipy.stats import spearmanr |
| D='/workspace/huawei/v2work' |
| s=pl.read_csv(f'{D}/out/submission_v6.csv',schema_overrides={'rowid':pl.Utf8}) |
| e=pl.read_parquet(f'{D}/eval_pred_v6.parquet').with_columns(pl.col('rowid').cast(pl.Utf8)) |
| c=pl.read_parquet(f'{D}/cache_mx/eval_bc6b041e949f.parquet',columns=['rowid']) |
| b=np.mean([np.load(f'{D}/ens/v6base_c27_s{k}.npy')[:,1] for k in (0,1)],0); sl=np.load(f'{D}/ens/v6slow_c27_s0.npy')[:,2] |
| p=1/(1+np.exp(-(b+sl)/2)); r=pl.DataFrame({'rowid':c['rowid'].cast(pl.Utf8),'pre':p}) |
| j=s.join(e,on='rowid',how='left').join(r,on='rowid',how='left') |
| v3=pl.read_csv(f'{D}/out/submission_v3.csv',schema_overrides={'rowid':pl.Utf8}).rename({'predicted_probability':'v3'}) |
| j=j.join(v3,on='rowid',how='left') |
| out={'sha256':hashlib.sha256(open(f'{D}/out/submission_v6.csv','rb').read()).hexdigest(),'rows':s.height,'unique':s['rowid'].n_unique(), |
| 'nulls_eval_pred':int(j['p'].null_count()),'maxabs_sub_vs_eval_pred_v6':float((j['predicted_probability']-j['p']).abs().max()), |
| 'maxabs_sub_vs_recomputed_blend_c27':float((j['predicted_probability']-j['pre']).abs().max()), |
| 'min':float(s['predicted_probability'].min()),'max':float(s['predicted_probability'].max()),'mean':float(s['predicted_probability'].mean()), |
| 'spearman_vs_v2work_submission_v3':float(spearmanr(j['predicted_probability'].to_numpy(),j['v3'].to_numpy()).statistic)} |
| print(json.dumps(out,indent=1)); json.dump(out,open('subcheck.json','w'),indent=1) |
|
|
