Scikit-learn
human-activity-recognition
wearable
wrist
time-series
cpu
scikit-learn
WISP / src /wisp_fast /funnel /controller.py
Zipeng365's picture
wisp-har 0.4.0: WISP-Funnel search and the adaptive HMM switch (default decoder)
93a9282 verified
Raw History Blame Contribute Delete
18.6 kB
# The search controller of wisp.search.controller with one hook, ``_fingerprint``, so the funnel
# can compute the dataset fingerprint once and share it between its two controllers. Everything
# else is unchanged; see SOURCE_PROVENANCE.json (version_0_4_0).
from __future__ import annotations
import hashlib
import json
import os
import pickle
import time
from concurrent.futures import ThreadPoolExecutor
from dataclasses import dataclass
from pathlib import Path
from typing import Any
import numpy as np
from wisp.core.data import HARData, concatenate_har
from wisp.core.metrics import classification_report_dict
from wisp.cpu.algorithms import build_cpu_estimator, estimate_pickle_size_mb
from wisp.search.cascade import evaluate_candidate_cascade
from .fingerprint import DatasetFingerprint, compute_dataset_fingerprint
from wisp.search.grammar import CandidateSpec, initial_candidates, mutate_candidate, normalize_spec
from wisp.search.operator_program import candidate_certificate
from wisp.search.posterior import MotifPosterior, island_key, prepare_parents
from wisp.search.profiles import SearchProfile, get_profile
from wisp.utils.jsonl import append_jsonl, read_jsonl
def _positive_int_env(name: str, default: int) -> int:
try:
return max(int(os.environ.get(name, default)), 1)
except (TypeError, ValueError):
return max(int(default), 1)
def _stable_id(*, seed: int, generation: int, index: int, spec: CandidateSpec) -> str:
payload = json.dumps(spec.to_dict(), sort_keys=True, separators=(',', ':'), default=str)
digest = hashlib.sha1(f'{seed}:{generation}:{index}:{payload}'.encode()).hexdigest()[:12]
return f'g{generation:03d}c{index:04d}_{digest}'
def _spec_key(spec: CandidateSpec) -> str:
return json.dumps(normalize_spec(spec).to_dict(), sort_keys=True, separators=(',', ':'), default=str)
@dataclass
class WISPSearchController:
train: HARData
valid: HARData
output: str | Path
profile: str | SearchProfile = 'wisp_evolution'
test: HARData | None = None
population: int = 24
generations: int = 20
seed: int = 42
cost_weight: float = 0.01
robustness_weight: float = 0.1
candidate_workers: int | None = None
n_jobs_per_candidate: int | None = None
def __post_init__(self) -> None:
if isinstance(self.profile, str):
self.profile = get_profile(self.profile)
if self.train.n_classes != self.valid.n_classes:
raise ValueError('train and validation must share the same global label vocabulary')
if self.test is not None and self.test.n_classes != self.train.n_classes:
raise ValueError('test must share the same global label vocabulary')
def _fingerprint(self):
return compute_dataset_fingerprint(self.train,seed=self.seed)
def _worker_count(self) -> int:
if self.candidate_workers is not None:
return max(int(self.candidate_workers), 1)
return _positive_int_env('WISP_CANDIDATE_WORKERS', 1)
def _per_candidate_jobs(self) -> int:
if self.n_jobs_per_candidate is not None:
return max(int(self.n_jobs_per_candidate), 1)
explicit = os.environ.get('WISP_N_JOBS_PER_CANDIDATE')
if explicit is not None:
return _positive_int_env('WISP_N_JOBS_PER_CANDIDATE', 1)
total = _positive_int_env('WISP_N_JOBS', 1)
return max(total // self._worker_count(), 1)
def _runtime_spec(self, spec: CandidateSpec) -> CandidateSpec:
spec = normalize_spec(spec)
if spec.route != 'cpu':
raise ValueError('WISPSearchController is CPU-only; GPU search is not released')
params = dict(spec.params)
params.setdefault('n_jobs', self._per_candidate_jobs())
return CandidateSpec(name=spec.name, params=params, operators=list(spec.operators), route='cpu', parent_id=spec.parent_id, mutation=spec.mutation)
def _constrain_spec(self, spec: CandidateSpec) -> CandidateSpec | None:
spec = self._runtime_spec(spec)
allowed = self.profile.allowed_cpu_families
if allowed is not None and spec.name not in set(allowed):
return None
if self.profile.force_no_hmm:
params = dict(spec.params)
params['use_hmm'] = False
operators = [op for op in spec.operators if op != 'hmm']
if 'argmax' not in operators:
operators.append('argmax')
spec = CandidateSpec(name=spec.name, params=params, operators=operators, route='cpu', parent_id=spec.parent_id, mutation=spec.mutation)
if spec.mutation in set(self.profile.disabled_mutations):
return None
return spec
def _evaluate_one(self, spec: CandidateSpec) -> tuple[str, dict[str, Any]]:
try:
cascade = evaluate_candidate_cascade(spec, self.train, self.valid, cost_weight=self.cost_weight, robustness_weight=self.robustness_weight if self.profile.use_stress else 0.0, max_stress_transforms=4 if self.profile.use_stress else 0)
result = cascade.to_dict()
predictive = float(result.get('cascade_score', result.get('score', -1000000000.0)))
result['predictive_utility'] = predictive
return ('ok' if cascade.accepted else 'rejected', result)
except Exception as exc:
return ('failed', {'predictive_utility': -1000000000.0, 'breeding_score': -1000000000.0, 'error': repr(exc), 'candidate': spec.to_dict()})
def _evaluate_many(self, specs: list[CandidateSpec]) -> list[tuple[str, dict[str, Any]]]:
if not specs:
return []
workers = min(self._worker_count(), len(specs))
if workers <= 1:
return [self._evaluate_one(spec) for spec in specs]
with ThreadPoolExecutor(max_workers=workers, thread_name_prefix='wisp') as pool:
return list(pool.map(self._evaluate_one, specs))
def _initial_specs(self, posterior: MotifPosterior, fingerprint: DatasetFingerprint | None, family_transfer_bias: dict[str, float], rng: np.random.Generator, count: int) -> list[CandidateSpec]:
specs = posterior.rank_initial_candidates(fingerprint=fingerprint, seed=self.seed, route='cpu', include_v2=self.profile.include_v2_families, motif_mode=self.profile.motif_mode, family_transfer_bias=family_transfer_bias)
specs = [constrained for spec in specs if (constrained := self._constrain_spec(spec)) is not None]
if not specs:
raise RuntimeError(f'Profile {self.profile.name} removed every initial CPU family')
base = list(specs)
i = 0
while len(specs) < count:
parent = base[i % len(base)]
clone = CandidateSpec.from_dict(parent.to_dict())
clone.params = dict(clone.params)
clone.params['seed'] = int(clone.params.get('seed', self.seed)) + 1009 * (i + 1)
clone.mutation = 'initial_seed_variant'
specs.append(clone)
i += 1
return [spec for spec in specs[:count]]
def _random_spec(self, rng: np.random.Generator, fingerprint: DatasetFingerprint | None) -> CandidateSpec:
template_seed=int(rng.integers(0, 2 ** 31 - 1))
template=self._templates[int(rng.integers(0,len(self._templates)))]
params=dict(template.params);params['seed']+=template_seed
spec=CandidateSpec(template.name,params,list(template.operators),template.route,template.parent_id,template.mutation)
depth = int(rng.integers(1, 5))
for _ in range(depth):
spec = mutate_candidate(spec, rng=rng, mutation_bias=None, motif_weights=self._motif_weights)
spec.parent_id = None
spec.mutation = f'random_depth_{depth}'
constrained = self._constrain_spec(spec)
if constrained is None:
return self._random_spec(rng, fingerprint)
return constrained
def _record(self, *, db_path: Path, candidate_id: str, generation: int, spec: CandidateSpec, status: str, result: dict[str, Any], fingerprint: DatasetFingerprint | None, parent_id: str | None) -> dict[str, Any]:
row = {'id': candidate_id, 'generation': int(generation), 'parent_id': parent_id, 'island': island_key(spec), 'status': status, 'profile': self._profile_dict, 'dataset_fingerprint': self._fingerprint_dict, 'motif_weights': self._motif_weights, 'operator_certificate': candidate_certificate(spec), 'result': result, 'spec': spec.to_dict()}
if self._ledger is None:
append_jsonl(db_path,row)
else:
self._ledger.write(json.dumps(row,ensure_ascii=False,sort_keys=True)+'\n')
return row
def _final_refit(self, best: dict[str, Any], out_dir: Path) -> dict[str, Any] | None:
if self.test is None:
return None
spec = CandidateSpec.from_dict(best['spec'])
model = build_cpu_estimator(spec.to_dict())
refit = concatenate_har(self.train, self.valid, source='train+validation')
t0 = time.perf_counter()
model.fit(refit.X, refit.y, groups=refit.subject, time_index=refit.time_index, n_classes=refit.n_classes)
fit_s = time.perf_counter() - t0
t1 = time.perf_counter()
pred = model.predict(self.test.X, groups=self.test.subject, time_index=self.test.time_index)
pred_s = time.perf_counter() - t1
metrics = classification_report_dict(self.test.y, pred, fit_seconds=fit_s, predict_seconds=pred_s, model_size_mb=estimate_pickle_size_mb(model), label_names=self.test.label_names)
np.savez_compressed(out_dir / 'locked_test_predictions.npz', y_true=self.test.y, y_pred=pred, subject=self.test.subject, time_index=self.test.time_index)
with (out_dir / 'selected_model.pkl').open('wb') as f:
pickle.dump(model, f)
(out_dir / 'locked_test_metrics.json').write_text(json.dumps(metrics, indent=2, ensure_ascii=False, default=str), encoding='utf-8')
return metrics
def run(self) -> dict[str, Any]:
self._ledger=None
try:
return self._run()
finally:
if self._ledger is not None:
self._ledger.close()
self._ledger=None
def _run(self) -> dict[str, Any]:
out_dir = Path(self.output)
out_dir.mkdir(parents=True, exist_ok=True)
db_path = out_dir / 'program_database.jsonl'
summary_path = out_dir / 'summary.json'
config = {'profile': self.profile.to_dict(), 'population': int(self.population), 'generations': int(self.generations), 'seed': int(self.seed), 'cost_weight': float(self.cost_weight), 'robustness_weight': float(self.robustness_weight), 'candidate_workers': self._worker_count(), 'n_jobs_per_candidate': self._per_candidate_jobs()}
if summary_path.exists():
old = json.loads(summary_path.read_text(encoding='utf-8'))
if old.get('status') == 'completed' and old.get('search_config') == config:
return old
if db_path.exists() and db_path.stat().st_size:
raise RuntimeError(f'Refusing to append to non-empty search database {db_path}')
self._ledger=db_path.open('a',encoding='utf-8')
rng = np.random.default_rng(self.seed)
fingerprint = self._fingerprint() if self.profile.use_fingerprint else None
family_transfer, mutation_transfer = ({}, {})
self._templates=initial_candidates(seed=0,include_v2=self.profile.include_v2_families,route='cpu')
self._motif_weights=None if fingerprint is None else fingerprint.motif_weights()
self._fingerprint_dict=None if fingerprint is None else fingerprint.to_dict()
self._profile_dict=self.profile.to_dict()
posterior = MotifPosterior()
archive: list[CandidateSpec] = []
population: list[tuple[str, CandidateSpec, dict[str, Any]]] = []
rows: list[dict[str, Any]] = []
seen: set[str] = set()
budget = int(self.profile.max_evaluations)
initial_count = min(9, budget)
seeds = self._initial_specs(posterior, fingerprint, family_transfer, rng, initial_count)
seed_results = self._evaluate_many(seeds)
for i, (spec, (status, result)) in enumerate(zip(seeds, seed_results)):
novelty = 0.0
result['novelty_bonus'] = novelty
result['breeding_score'] = float(result.get('predictive_utility', -1000000000.0)) + novelty
cid = _stable_id(seed=self.seed, generation=0, index=i, spec=spec)
row = self._record(db_path=db_path, candidate_id=cid, generation=0, spec=spec, status=status, result=result, fingerprint=fingerprint, parent_id=None)
rows.append(row)
seen.add(_spec_key(spec))
archive.append(spec)
if status == 'ok':
population.append((cid, spec, result))
if self.profile.use_posterior:
posterior.update(spec, result, score_key='predictive_utility')
if not population:
raise RuntimeError('No accepted initial candidate')
evaluated = len(seeds)
generation = 0
while evaluated < budget and generation < int(self.generations):
generation += 1
remaining = budget - evaluated
width = max(self.population - max(2, self.population // 3), 1)
coalesce = getattr(self, "coalesce_random", False)
proposal_count = min((self.generations-generation+1)*width if coalesce else width, remaining)
proposals: list[tuple[str, str | None, CandidateSpec]] = []
attempts = 0
mutation_bias = posterior.mutation_bias(fingerprint, motif_mode=self.profile.motif_mode, transfer_bias=mutation_transfer) if self.profile.use_posterior else mutation_transfer if self.profile.transfer_prior else {}
for disabled in self.profile.disabled_mutations:
mutation_bias.pop(disabled, None)
draw=prepare_parents(population,use_islands=self.profile.use_islands) if self.profile.search_strategy=='evolution' else None
while len(proposals) < proposal_count:
attempts += 1
if attempts > proposal_count * 100:
raise RuntimeError('Unable to generate enough unique candidate specifications')
if self.profile.search_strategy == 'random':
parent_id = None
child = self._random_spec(rng, fingerprint)
else:
parent_id, parent, parent_result = draw(rng)
child = None
if child is None:
child = mutate_candidate(parent, rng=rng, parent_id=parent_id, mutation_bias=mutation_bias, motif_weights=self._motif_weights)
child = self._constrain_spec(child)
if child is None:
continue
key = _spec_key(child)
if key in seen:
continue
seen.add(key)
index = evaluated + len(proposals)
logical_generation = generation + len(proposals)//width if coalesce else generation
proposals.append((_stable_id(seed=self.seed, generation=logical_generation, index=index, spec=child), parent_id, child))
evaluated_results = self._evaluate_many([spec for _, _, spec in proposals])
new_items: list[tuple[str, CandidateSpec, dict[str, Any]]] = []
for position, ((cid, parent_id, spec), (status, result)) in enumerate(zip(proposals, evaluated_results)):
novelty = posterior.novelty_bonus(spec, archive) if self.profile.use_novelty and status == 'ok' else 0.0
result['novelty_bonus'] = float(novelty)
result['breeding_score'] = float(result.get('predictive_utility', -1000000000.0)) + float(novelty)
row = self._record(db_path=db_path, candidate_id=cid, generation=generation+position//width if coalesce else generation, spec=spec, status=status, result=result, fingerprint=fingerprint, parent_id=parent_id)
rows.append(row)
archive.append(spec)
if status == 'ok':
new_items.append((cid, spec, result))
if self.profile.use_posterior:
posterior.update(spec, result, score_key='predictive_utility')
evaluated += len(proposals)
if self.profile.search_strategy == 'evolution':
combined = population + new_items
combined.sort(key=lambda item: float(item[2].get('breeding_score', -1000000000.0)), reverse=True)
population = combined[:int(self.population)]
else:
population = (population + new_items)[-int(self.population):]
if not population:
raise RuntimeError(f'No accepted candidate after generation {generation}')
self._ledger.flush()
if getattr(self,'retain_rows',False):
self.evaluated_rows=rows
ok_rows = [row for row in rows if row['status'] == 'ok']
if not ok_rows:
raise RuntimeError('Search completed without an accepted candidate')
final_key = self.profile.final_selection
best = max(ok_rows, key=lambda row: float(row['result'].get(final_key, -1000000000.0)))
best_breeding = max(ok_rows, key=lambda row: float(row['result'].get('breeding_score', -1000000000.0)))
best_predictive = max(ok_rows, key=lambda row: float(row['result'].get('predictive_utility', -1000000000.0)))
(out_dir / 'selected_candidate.json').write_text(json.dumps(best, indent=2, ensure_ascii=False, default=str), encoding='utf-8')
test_metrics = self._final_refit(best, out_dir)
summary = {'status': 'completed', 'version': 'wisp_search_controller', 'profile': self.profile.name, 'n_rows': len(rows), 'n_ok': len(ok_rows), 'n_evaluated': evaluated, 'best': best, 'best_predictive': best_predictive, 'best_breeding': best_breeding, 'locked_test_metrics': test_metrics, 'search_config': config, 'dataset_fingerprint': None if fingerprint is None else {**fingerprint.to_dict(), 'motifs': fingerprint.motifs(), 'motif_weights': fingerprint.motif_weights()}, 'posterior_summary': posterior.summary(), 'cross_task_family_bias': family_transfer, 'cross_task_mutation_bias': mutation_transfer}
summary_path.write_text(json.dumps(summary, indent=2, ensure_ascii=False, default=str), encoding='utf-8')
return summary
__all__ = ['WISPSearchController']