| import os |
| os.environ.setdefault('MPLCONFIGDIR',os.path.abspath('.cache/matplotlib')) |
| import json,hashlib,numpy as np |
| from pathlib import Path |
| from bci_data import cached |
| assert Path('reports/frozen_selection.json').exists(),'Audit test data only after freeze' |
| rows=[] |
| global_hashes=set() |
| for subject in range(1,16): |
| for split in ['train','validation']: |
| xx,_,_=cached(subject,split);global_hashes.update(hashlib.sha256(q.tobytes()).hexdigest() for q in xx) |
| for s in range(1,16): |
| train,_,_=cached(s,'train');val,_,_=cached(s,'validation');test,_,_=cached(s,'test',evaluation=True) |
| a=np.concatenate([train,val]);ha={hashlib.sha256(q.tobytes()).hexdigest() for q in a};duplicates=sum(hashlib.sha256(q.tobytes()).hexdigest() in ha for q in test) |
| |
| af=a[:,:,::4].reshape(len(a),-1).astype(float);tf=test[:,:,::4].reshape(len(test),-1).astype(float);af-=af.mean(1,keepdims=True);tf-=tf.mean(1,keepdims=True);af/=np.linalg.norm(af,axis=1,keepdims=True);tf/=np.linalg.norm(tf,axis=1,keepdims=True);mx=(tf@af.T).max(1) |
| rows.append({'subject':s,'exact_test_duplicates_with_train_validation':duplicates,'exact_duplicates_any_training_subject':sum(hashlib.sha256(q.tobytes()).hexdigest() in global_hashes for q in test),'max_test_to_training_raw_correlation':float(mx.max()),'n_test_correlation_gt_0_99':int((mx>.99).sum())}) |
| json.dump(rows,open('reports/final_leakage_audit.json','w'),indent=2);print('Exact duplicate count:',sum(r['exact_test_duplicates_with_train_validation'] for r in rows)) |
|
|
