Files
MeltMind/tools/test_gre24_models.py
T

66 lines
3.7 KiB
Python

"""Role isolation, full-row learning, early stopping and real guard receipts."""
import json
from pathlib import Path
import resource
import sys
import tempfile
import unittest
import numpy as np
from gre24_models import xgb_model,fnn_model,eligible_roles,require_training_role,early_stop_update
from gre24_job_guard import supervise
class ModelChecks(unittest.TestCase):
def test_role_isolation_and_missing_no_replacement(self):
s=dict(train=['a','b'],test=['c','d'],fnn_learning=['a'],fnn_validation=['b'])
roles,missing=eligible_roles(s,['a','c','d'])
self.assertEqual(roles['train'],['a']);self.assertEqual(missing['train'],['b'])
with self.assertRaises(PermissionError):require_training_role('c',s,'train')
with self.assertRaises(PermissionError):require_training_role('a',s,'test')
def test_earlystop_strict_improvement_patience(self):
best=1.;stale=0
for i in range(5):
best,stale,improved,stop=early_stop_update(best,stale,1.)
self.assertFalse(improved);self.assertEqual(stop,i==4)
best,stale,improved,stop=early_stop_update(best,stale,.9)
self.assertTrue(improved);self.assertEqual(stale,0);self.assertFalse(stop)
with self.assertRaises(ValueError):early_stop_update(best,stale,float('nan'))
def test_xgboost_fixed_recipe_learns_all_rows(self):
rng=np.random.default_rng(42);x=rng.normal(size=(512,19)).astype(np.float32);y=x[:,0]+.5*x[:,1]
m=xgb_model(42,threads=1);m.fit(x,y)
self.assertLess(np.mean((m.predict(x)-y)**2),.02)
self.assertEqual(m.n_estimators,100);self.assertEqual(m.max_depth,8)
def test_fnn_architecture_gradients_and_checkpoint_restore(self):
import torch
torch.set_num_threads(1);torch.manual_seed(42);m=fnn_model()
linear=[a for a in m if isinstance(a,torch.nn.Linear)]
self.assertEqual([(a.in_features,a.out_features) for a in linear],[(19,100),(100,100),(100,100),(100,100),(100,1)])
x=torch.randn(64,19);y=torch.ones(64,1);o=torch.optim.Adam(m.parameters(),lr=.001)
before=float(((m(x)-y)**2).mean())
for _ in range(15):
o.zero_grad();loss=((m(x)-y)**2).mean();loss.backward();o.step()
self.assertLess(float(((m(x)-y)**2).mean()),before)
with tempfile.TemporaryDirectory() as tmp:
p=Path(tmp)/'m.pt';torch.save(m.state_dict(),p);n=fnn_model();n.load_state_dict(torch.load(p,weights_only=True))
torch.testing.assert_close(m(x),n(x))
def test_guard_records_failure_cpu_and_refuses_repeat(self):
limits=dict(total_cpu_seconds=10,job_cpu_seconds=5,job_wall_seconds=5,tree_rss_bytes=256*2**20,disk_bytes=2**20,max_processes=1)
with tempfile.TemporaryDirectory() as tmp:
r=supervise([sys.executable,'-c','raise SystemExit(7)'],Path(tmp),'fail',limits,poll=.02)
self.assertEqual(r['returncode'],7);self.assertGreater(r['cpu_seconds'],0)
self.assertEqual(r['status'],'STOPPED_OR_FAILED')
with self.assertRaises(FileExistsError):supervise([sys.executable,'-c','pass'],Path(tmp),'fail',limits)
if __name__=='__main__':
result=unittest.TextTestRunner(verbosity=2).run(unittest.defaultTestLoader.loadTestsFromTestCase(ModelChecks))
u=resource.getrusage(resource.RUSAGE_SELF);ch=resource.getrusage(resource.RUSAGE_CHILDREN)
Path('artifacts/gre24-reproduction-2026-11-15/model-synthetic-receipt.json').write_text(json.dumps(dict(
tests=result.testsRun,passed=result.wasSuccessful(),failures=len(result.failures),errors=len(result.errors),
CPU_s=u.ru_utime+u.ru_stime+ch.ru_utime+ch.ru_stime,GPU_used=False),indent=2))
raise SystemExit(0 if result.wasSuccessful() else 1)