tests/evaluate_release.py

"""Additional same-information baseline and exact-solver timing."""
from pathlib import Path
import json,sys,time
import numpy as np
from scipy.spatial import cKDTree
sys.path.insert(0,str(Path(__file__).resolve().parents[1]/'src'))
from train import features,reconstruct,network
from rod import Rod,Controls
from safetensors.torch import load_file
import torch
R=Path(__file__).resolve().parents[1]
rows=[json.loads(l) for l in (R/'data/equilibria.jsonl').open()]
train=[v for v in rows if v['split']=='train'];test=[v for v in rows if v['split']=='test']
x=np.array([features(v['controls']) for v in train]);mean=x.mean(0);std=x.std(0);std[std<1e-6]=1
pred=[]
for v in test:
    allowed=np.array([t['branch']==v['branch'] for t in train])
    subset=np.flatnonzero(allowed)
    distances=np.sum(((x[subset]-features(v['controls']))/std)**2,axis=1)
    pred.append(train[subset[np.argmin(distances)]]['z'])
p=reconstruct(np.array(pred),test);true=reconstruct(np.array([v['z'] for v in test]),test)
report={'branch_conditioned_1nn':{'centerline_rmse_over_L':float(np.sqrt(np.mean(np.sum((p-true)**2,axis=2)))),'tip_mean_error_over_L':float(np.mean(np.linalg.norm(p[:,-1]-true[:,-1],axis=1)))}}
config=json.loads((R/'models/branch_conditioned.json').read_text());net=network(11);net.load_state_dict(load_file(str(R/'models/branch_conditioned.safetensors')));net.eval()
sample=test[::20];xt=torch.tensor((np.array([features(v['controls'],v['branch']) for v in sample])-config['mean'])/config['std'],dtype=torch.float32)
with torch.no_grad():
    for _ in range(10):net(xt)
    start=time.perf_counter()
    for _ in range(100):z=net(xt)
    report['neural_batched_ms_per_state']=(time.perf_counter()-start)*1000/(100*len(sample))
r=Rod();times=[];corr=[]
for i,v in enumerate(sample):
    start=time.perf_counter();result=r.solve(Controls(**v['controls']),z[i].numpy());times.append(time.perf_counter()-start)
    corr.append(dict(id=v['id'],accepted=result['accepted'],branch_preserved=bool((v['branch']=='straight' and np.max(np.abs(result['z']))<1e-5) or (v['branch']=='up' and result['tip'][1]>1e-5) or (v['branch']=='down' and result['tip'][1]<-1e-5)),tip_error=float(np.linalg.norm(result['tip']-v['tip']))))
report['corrector_median_ms']=float(1000*np.median(times));report['corrector_checks']=corr
assert all(c['accepted'] and c['branch_preserved'] for c in corr)
# Exact duplicate geometries can occur at zero load and on reversible sweeps;
# scene-disjoint split is guaranteed, shape-disjointness is NOT claimed.
report['number_scenes']=len({r['scene_id'] for r in rows})
report['number_paths']=len({r['path_id'] for r in rows})
report['sampled_stable_pairs']=sum(r['continuous_stable_pair'] for r in rows)
(R/'reports/additional_baselines.json').write_text(json.dumps(report,indent=2))
print({k:v for k,v in report.items() if k!='corrector_checks'})