114 lines
7.9 KiB
Python
114 lines
7.9 KiB
Python
"""Post-hoc D1/D2/D4 from immutable existing outputs; no model or label loading."""
|
|
from collections import defaultdict
|
|
import json
|
|
from pathlib import Path
|
|
import time
|
|
import numpy as np
|
|
from freeze_g1_components import sha
|
|
from s2_recovery_supervisor import save
|
|
|
|
ROOT=Path('artifacts/s2-posthoc-diagnostics-2026-11-15')
|
|
EVAL=Path('artifacts/s2-custody-evaluation-2026-11-15')
|
|
|
|
|
|
def read(path):return json.loads(Path(path).read_text())
|
|
|
|
|
|
def grouped(values,ids):
|
|
groups=defaultdict(list)
|
|
for value,key in zip(values,ids):groups[key].append(float(value))
|
|
return float(np.mean([np.mean(v) for v in groups.values()]))
|
|
|
|
|
|
def main():
|
|
start=time.process_time();assert not ROOT.exists(),'Do not overwrite diagnostic evidence'
|
|
result=read(EVAL/'custody/result.json')
|
|
for name in ['component-scores.json','part-scores.json','tables.json']:
|
|
assert sha(EVAL/'custody'/name)==result['output_sha256'][name]
|
|
scores=read(EVAL/'custody/component-scores.json');curves={}
|
|
arms=['N','E','R2','B0'];pools=[25,50,100]
|
|
for pool in pools:
|
|
for arm in arms:
|
|
curves[pool,arm]={r['component_id']:r['value'] for r in scores if r['subset']=='all' and r['metric']=='relative_l2' and r['arm']==arm and r['pool']==pool}
|
|
common=set.intersection(*(set(v) for v in curves.values()))
|
|
all_comparison=set(curves[25,'N'])&set(curves[100,'N'])
|
|
d1=dict(posthoc=True,common_components=len(common),common_component_ids=sorted(common),common_folds=[0,2,3,4],
|
|
rows=[dict(pool_component_fraction=pool/100,arm=arm,original_components=len(curves[pool,arm]),
|
|
original_mean=float(np.mean(list(curves[pool,arm].values()))),
|
|
common_mean=float(np.mean([curves[pool,arm][k] for k in sorted(common)]))) for pool in pools for arm in arms],
|
|
N_25_to100_all5_relative_error_reduction=float(1-np.mean([curves[100,'N'][k] for k in all_comparison])/np.mean([curves[25,'N'][k] for k in all_comparison])),
|
|
note='Original50pctN denominator differs; common fourfold curve is a separate descriptive diagnostic, not replacement primary analysis')
|
|
records=[];histories={}
|
|
for fold in range(5):
|
|
for arm in ['N','E']:
|
|
for seed in [0,1,2]:
|
|
name=f'fold{fold}-pool100-{arm}128-seed{seed}';base=Path('artifacts/s2-main-matrix-2026-11-15/cells')/name/name
|
|
rr=read(base/'result.json');h=read(base/'history.json');histories[name]=h
|
|
y=np.array([v['validation_component_equal_relative_l2'] for v in h]);train=np.array([v['training_global_weighted_mse'] for v in h])
|
|
n=min(5,len(y));slope=float(np.polyfit(np.arange(n),y[-n:],1)[0])
|
|
records.append(dict(name=name,fold=fold,arm=arm,seed=seed,epochs=len(h),best_epoch_1based=rr['best_epoch']+1,
|
|
best_validation=rr['best_validation_score'],final_validation=float(y[-1]),
|
|
best_in_last5=rr['best_epoch']>=len(h)-5,best_at_final=rr['best_epoch']==len(h)-1,
|
|
final5_slope=slope,final5_endpoint_relative_change=float(y[-1]/y[-n]-1),
|
|
first_validation=float(y[0]),final_training_loss=float(train[-1]),final5_training_relative_change=float(train[-1]/train[-n]-1)))
|
|
summary=[]
|
|
for arm in ['N','E']:
|
|
rr=[r for r in records if r['arm']==arm]
|
|
summary.append(dict(arm=arm,runs=len(rr),all_reached42=all(r['epochs']==42 for r in rr),
|
|
median_best_epoch=float(np.median([r['best_epoch_1based'] for r in rr])),
|
|
best_in_last5=sum(r['best_in_last5'] for r in rr),best_at_final=sum(r['best_at_final'] for r in rr),
|
|
negative_final5_validation_slope=sum(r['final5_slope']<0 for r in rr),
|
|
median_final5_validation_change=float(np.median([r['final5_endpoint_relative_change'] for r in rr])),
|
|
median_final5_training_change=float(np.median([r['final5_training_relative_change'] for r in rr]))))
|
|
d2=dict(posthoc=True,epoch_display='one based',records=records,summary=summary,
|
|
limitation='Late best epoch/negative local slope does not prove a longer schedule will beat R2; cosine endpoint and fitting choices are confounded')
|
|
roster=read('artifacts/s2-main-preparation-2026-11-15/final-roster.json')['parts']
|
|
partrows=read(EVAL/'custody/part-scores.json');by=defaultdict(list)
|
|
for r in partrows:
|
|
if r['arm']=='B0' or r['pool']==100:by[r['case'],r['arm']].append(r['relative_l2']['value'])
|
|
geom=[]
|
|
for r in roster:
|
|
if not r['included']:continue
|
|
p=Path(r['preparation']['preparation_root'])/'result.json';assert sha(p)==r['preparation']['result_sha256'];s=read(p)['scales']
|
|
means={arm:float(np.mean(by[r['case'],arm])) for arm in arms}
|
|
assert len(by[r['case'],'N'])==len(by[r['case'],'E'])==3
|
|
geom.append(dict(case=r['case'],component_id=r['component_id'],nodes=r['nodes'],Rg=s['Rg'],t_part=s['t_part'],
|
|
errors=means,N_minus_B0=means['N']-means['B0'],N_loses_to_B0=means['N']>means['B0']))
|
|
bins=[]
|
|
for feature in ['Rg','t_part','nodes']:
|
|
x=np.array([r[feature] for r in geom]);cuts=np.quantile(x,[1/3,2/3],method='linear');assignment=np.searchsorted(cuts,x,side='right')
|
|
for b in range(3):
|
|
rr=[r for i,r in enumerate(geom) if assignment[i]==b];ids=[r['component_id'] for r in rr]
|
|
bins.append(dict(feature=feature,bin=b,tertile_cutpoints=cuts.tolist(),min=min(r[feature] for r in rr),max=max(r[feature] for r in rr),
|
|
parts=len(rr),components=len(set(ids)),N_loses_parts=sum(r['N_loses_to_B0'] for r in rr),
|
|
component_equal_loss_fraction=grouped([r['N_loses_to_B0'] for r in rr],ids),
|
|
component_equal_difference=grouped([r['N_minus_B0'] for r in rr],ids),
|
|
component_equal_N=grouped([r['errors']['N'] for r in rr],ids),
|
|
component_equal_B0=grouped([r['errors']['B0'] for r in rr],ids)))
|
|
d4=dict(posthoc=True,parts=617,geometry_units='original prepared geometry units; nodes count',
|
|
bins=bins,part_records=geom,N_loses_parts=sum(r['N_loses_to_B0'] for r in geom),
|
|
component_equal_loss_fraction=grouped([r['N_loses_to_B0'] for r in geom],[r['component_id'] for r in geom]),
|
|
limitation='Geometry-only empirical tertiles on617parts, posthoc descriptive; within-bin parts/component equal; no causal claim or multiplicity test')
|
|
ROOT.mkdir();save(ROOT/'D1-learning-curves.json',d1);save(ROOT/'D2-epoch-history.json',d2);save(ROOT/'D4-geometry-errors.json',d4)
|
|
# Standalone scientific figure from existing tabulations, no new fitted model.
|
|
import matplotlib
|
|
matplotlib.use('Agg')
|
|
import matplotlib.pyplot as plt
|
|
fig,axes=plt.subplots(1,3,figsize=(13,3.8))
|
|
for arm in arms:
|
|
axes[0].plot([.25,.5,1],[100*next(r['common_mean'] for r in d1['rows'] if r['arm']==arm and r['pool_component_fraction']==p) for p in [.25,.5,1]],'-o',label=arm)
|
|
axes[0].set(xlabel='Learning component fraction',ylabel='Component-equal relative L2 (%)',title='D1: common four-fold diagnostic');axes[0].legend()
|
|
for ax,arm in zip(axes[1:],['N','E']):
|
|
hs=[h for name,h in histories.items() if f'-{arm}128-' in name]
|
|
for h in hs:ax.plot([r['epoch']+1 for r in h],[r['validation_component_equal_relative_l2'] for r in h],alpha=.3,lw=.8)
|
|
ax.set(xlabel='Epoch',ylabel='Validation relative L2',title=f'D2: {arm}, 100% pool, 15 runs')
|
|
fig.tight_layout();fig.savefig(ROOT/'D1-D2-curves.png',dpi=200);plt.close(fig)
|
|
save(ROOT/'receipt.json',dict(status='D1_D2_D4_COMPLETE',cpu_seconds=time.process_time()-start,
|
|
new_labels_decoded=False,models_loaded=False,new_training=False,original_statistics_modified=False,
|
|
D3='WAITING_MATHEMATICAL_BOUNDARIES',ruling='1794731320.994789',code_sha256=sha(__file__),
|
|
output_sha256={p.name:sha(p) for p in ROOT.iterdir() if p.is_file()}))
|
|
print(json.dumps(dict(D1_reduction=d1['N_25_to100_all5_relative_error_reduction'],D2=summary,D4_loses=d4['N_loses_parts'],CPU=time.process_time()-start)))
|
|
|
|
|
|
if __name__=='__main__':main()
|