Scikit-learn
human-activity-recognition
wearable
wrist
time-series
cpu
scikit-learn
WISP / tests /test_search_cache.py
Zipeng365's picture
wisp-har 0.2.0: accelerated implementation with speed (exact/fast) and HMM sequence (subject/recording) options
ac7f7cb verified
Raw History Blame Contribute Delete
4.27 kB
import json
from pathlib import Path
from tempfile import TemporaryDirectory
import unittest
import numpy as np
from sklearn.ensemble import ExtraTreesClassifier, RandomForestClassifier
from threadpoolctl import threadpool_limits
from wisp_fast import features as fx
from wisp_fast.search import fit_search
from wisp_fast.search_cache import SearchFeatures
from wisp_fast.search_forest import forest_fit_plan
from test_search_shared import splits
class SearchPlanTests(unittest.TestCase):
def test_forest_seed_and_tree_parity(self):
rng=np.random.default_rng(76)
x=rng.normal(size=(71,24))
y=np.resize([0,2,3,3],len(x))
for kind,weight in ((ExtraTreesClassifier,"balanced"),(RandomForestClassifier,"balanced_subsample")):
for jobs in (1,2):
kw=dict(random_state=7,n_estimators=11,n_jobs=jobs,class_weight=weight)
ref=kind(**kw).fit(x,y)
got=kind(**kw)
with forest_fit_plan(got):
got.fit(x,y)
self.assertNotIn("_make_estimator",got.__dict__)
for a,b in zip(ref.estimators_,got.estimators_):
self.assertEqual(a.get_params(),b.get_params())
for name in ("children_left","children_right","feature","threshold","value","weighted_n_node_samples"):
np.testing.assert_array_equal(getattr(a.tree_,name),getattr(b.tree_,name))
np.testing.assert_allclose(ref.predict_proba(x),got.predict_proba(x),atol=1e-15,rtol=0)
def test_interval_cache_normalization_and_seed(self):
x=np.random.default_rng(22).normal(size=(28,43,3))
mem=SearchFeatures(4,[x])
for fast in (True,False):
for seed in (3,11):
for factory in (fx.IntervalDistributionSketch,fx.InvariantPhysicsFeatureExtractor):
b=factory(fast_profile=fast,random_state=seed).fit(x)
np.testing.assert_array_equal(mem.transform([b],x)[0],b.transform(x))
if fx._cis_intervals is not None:
self.assertGreater(mem.rank_hits,1)
self.assertGreater(mem.dyad_hits,1)
self.assertLessEqual(mem.peak,mem.limit)
def test_certificates_against_complete_ledgers(self):
train,valid=splits()
with TemporaryDirectory() as tmp,threadpool_limits(1):
for method in ("wisp_random","wisp_evolution"):
arms={}
for policy in ("complete","certified"):
path=Path(tmp)/method/policy
_,s=fit_search(method,train,valid,output=path,mode="turbo",evaluations=25,
population=6,cost_weight=0,refit=False,stress_cache_mb=0,
stress_policy=policy,workers=2)
rows=[json.loads(l) for l in (path/"program_database.jsonl").read_text().splitlines()]
arms[policy]=s,rows
(s,full),(t,short)=arms.values()
self.assertEqual([r["id"] for r in full],[r["id"] for r in short])
self.assertEqual(s["best"]["id"],t["best"]["id"])
skipped=0
for row,original in zip(short,full):
if row["status"]=="pruned":
cert=row["result"]["prune_certificate"]
refscore=original["result"]["predictive_utility"]
idx=cert["incumbent_ledger_index"]
self.assertEqual(short[idx]["result"]["predictive_utility"],cert["incumbent_utility"])
self.assertLess(cert["upper_bound"],cert["incumbent_utility"])
self.assertLessEqual(refscore,cert["upper_bound"])
self.assertNotIn("predictive_utility",row["result"])
if method=="wisp_evolution":self.assertEqual(row["generation"],short[-1]["generation"])
skipped+=cert["skipped_views"]
else:
self.assertEqual(row["result"]["predictive_utility"],original["result"]["predictive_utility"])
self.assertEqual(skipped,t["skipped_stress_views"])
if method=="wisp_random":self.assertGreater(skipped,0)