Spaces:
Sleeping
Sleeping
File size: 16,409 Bytes
df31aa1 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 | """
Active Learning Module for Cognexa ML Service
Implements uncertainty-based active learning to identify the most informative
samples for human review, minimizing labeling cost while maximizing model improvement.
Strategies:
- Least confidence sampling: pick samples where model is least certain
- Margin sampling: smallest gap between top-2 class probabilities
- Entropy sampling: highest Shannon entropy across class probabilities
- Query-by-committee (QBC): disagreement between ensemble members
"""
from __future__ import annotations
import json
import logging
import uuid
from dataclasses import dataclass, asdict, field
from datetime import datetime, timedelta
from pathlib import Path
from typing import Dict, List, Optional, Tuple, Any
import numpy as np
logger = logging.getLogger(__name__)
# ---------------------------------------------------------------------------
# Data Structures
# ---------------------------------------------------------------------------
@dataclass
class UncertainSample:
"""A sample flagged by active learning as needing human review."""
sample_id: str
user_id: str
task_id: Optional[str]
features: Dict[str, float]
current_prediction: Dict[str, Any]
uncertainty_score: float # 0-1; higher = more uncertain
uncertainty_method: str # 'least_confidence' | 'margin' | 'entropy' | 'qbc'
query_priority: int # 1 (highest) to 5 (lowest)
created_at: str
reviewed: bool = False
review_label: Optional[Any] = None
reviewer_id: Optional[str] = None
reviewed_at: Optional[str] = None
@dataclass
class ActiveLearningBatch:
"""A batch of uncertain samples to present to reviewers."""
batch_id: str
strategy: str
samples: List[UncertainSample]
total_pool_size: int
batch_size: int
created_at: str
expected_model_gain: float # Estimated accuracy improvement if all labeled
@dataclass
class ActiveLearningStats:
"""Aggregate statistics for active learning progress."""
total_queried: int
total_reviewed: int
total_pending: int
review_rate: float
avg_uncertainty: float
coverage_by_strategy: Dict[str, int]
model_improvement_estimate: float
# ---------------------------------------------------------------------------
# Uncertainty Estimators
# ---------------------------------------------------------------------------
class UncertaintyEstimator:
"""Computes uncertainty scores from model predictions."""
@staticmethod
def least_confidence(probabilities: List[float]) -> float:
"""
1 - max(p): how far the most-confident class is from certainty.
Range [0, 1); higher means more uncertain.
"""
if not probabilities:
return 0.5
return 1.0 - max(probabilities)
@staticmethod
def margin_confidence(probabilities: List[float]) -> float:
"""
Margin between two highest probabilities.
Range [0, 1); lower margin = higher uncertainty.
Returns 1 - margin so higher = more uncertain.
"""
if len(probabilities) < 2:
return 0.5
sorted_probs = sorted(probabilities, reverse=True)
margin = sorted_probs[0] - sorted_probs[1]
return 1.0 - margin
@staticmethod
def entropy(probabilities: List[float]) -> float:
"""
Shannon entropy: H = -sum(p * log2(p)).
Normalised to [0, 1] by dividing by log2(n_classes).
"""
if not probabilities:
return 0.5
n = len(probabilities)
if n == 1:
return 0.0
eps = 1e-10
h = -sum(p * np.log2(p + eps) for p in probabilities if p > 0)
return h / np.log2(n)
@staticmethod
def query_by_committee(predictions_list: List[List[float]]) -> float:
"""
Vote entropy: how much committee members disagree.
predictions_list: list of probability vectors from each committee member.
Returns 0 if all agree, 1 if maximally disagree.
"""
if not predictions_list or len(predictions_list) < 2:
return 0.0
n_members = len(predictions_list)
n_classes = len(predictions_list[0])
# Count votes for each class (argmax)
votes = [int(np.argmax(p)) for p in predictions_list]
vote_counts = np.bincount(votes, minlength=n_classes)
vote_probs = vote_counts / n_members
eps = 1e-10
h = -sum(v * np.log2(v + eps) for v in vote_probs if v > 0)
return h / np.log2(n_members) if n_members > 1 else 0.0
# ---------------------------------------------------------------------------
# Active Learning Selector
# ---------------------------------------------------------------------------
class ActiveLearningSelector:
"""
Selects the most informative unlabeled samples from a candidate pool.
"""
def __init__(self, strategy: str = "entropy", threshold: float = 0.65):
"""
Args:
strategy: 'least_confidence' | 'margin' | 'entropy' | 'qbc'
threshold: minimum uncertainty score to include in batch (0-1)
"""
self.strategy = strategy
self.threshold = threshold
self.estimator = UncertaintyEstimator()
def score_sample(
self,
prediction: Dict[str, Any],
committee_predictions: Optional[List[List[float]]] = None,
) -> float:
"""Compute an uncertainty score for a single prediction."""
completion_prob = float(prediction.get("completion_probability", 0.5))
delay_prob = float(prediction.get("delay_probability", 1.0 - completion_prob))
probs = [completion_prob, delay_prob]
if self.strategy == "least_confidence":
return self.estimator.least_confidence(probs)
elif self.strategy == "margin":
return self.estimator.margin_confidence(probs)
elif self.strategy == "entropy":
return self.estimator.entropy(probs)
elif self.strategy == "qbc" and committee_predictions:
return self.estimator.query_by_committee(committee_predictions)
else:
# Default: entropy
return self.estimator.entropy(probs)
def select_batch(
self,
candidate_pool: List[Dict[str, Any]],
batch_size: int = 20,
user_id: Optional[str] = None,
) -> ActiveLearningBatch:
"""
Select the top-k uncertain samples from candidate_pool.
Args:
candidate_pool: list of dicts with keys:
- task_id (str)
- features (Dict[str, float])
- prediction (Dict[str, Any])
- committee_predictions (optional List[List[float]])
batch_size: how many samples to include
user_id: optional user constraint
Returns:
ActiveLearningBatch with ranked uncertain samples.
"""
scored: List[Tuple[float, Dict[str, Any]]] = []
for candidate in candidate_pool:
prediction = candidate.get("prediction", {})
committee = candidate.get("committee_predictions")
score = self.score_sample(prediction, committee)
if score >= self.threshold:
scored.append((score, candidate))
# Sort descending by uncertainty
scored.sort(key=lambda x: x[0], reverse=True)
top_k = scored[:batch_size]
samples = []
for rank, (score, candidate) in enumerate(top_k, 1):
priority = min(5, max(1, int((1.0 - score) * 5) + 1))
sample = UncertainSample(
sample_id=str(uuid.uuid4()),
user_id=user_id or candidate.get("user_id", "unknown"),
task_id=candidate.get("task_id"),
features=candidate.get("features", {}),
current_prediction=candidate.get("prediction", {}),
uncertainty_score=round(score, 4),
uncertainty_method=self.strategy,
query_priority=priority,
created_at=datetime.utcnow().isoformat(),
)
samples.append(sample)
# Estimate model gain (heuristic: based on avg uncertainty of selected batch)
avg_uncertainty = np.mean([s.uncertainty_score for s in samples]) if samples else 0.0
model_gain = avg_uncertainty * 0.05 # ~5% improvement per 1.0 of uncertainty
return ActiveLearningBatch(
batch_id=str(uuid.uuid4()),
strategy=self.strategy,
samples=samples,
total_pool_size=len(candidate_pool),
batch_size=len(samples),
created_at=datetime.utcnow().isoformat(),
expected_model_gain=round(model_gain, 4),
)
# ---------------------------------------------------------------------------
# Active Learning Manager (persistence + orchestration)
# ---------------------------------------------------------------------------
class ActiveLearningManager:
"""
Manages the active learning pipeline:
- Stores uncertain sample batches
- Tracks which have been reviewed
- Provides stats for the dashboard
"""
def __init__(self, data_dir: str = "data/active_learning"):
self.data_dir = Path(data_dir)
self.data_dir.mkdir(parents=True, exist_ok=True)
self.pending_file = self.data_dir / "pending_samples.json"
self.reviewed_file = self.data_dir / "reviewed_samples.json"
self._pending: List[UncertainSample] = self._load(self.pending_file)
self._reviewed: List[UncertainSample] = self._load(self.reviewed_file)
# -- Persistence ----------------------------------------------------------
def _load(self, path: Path) -> List[UncertainSample]:
if not path.exists():
return []
try:
with open(path) as f:
raw = json.load(f)
return [UncertainSample(**item) for item in raw]
except Exception as e:
logger.warning("Could not load %s: %s", path, e)
return []
def _save_pending(self):
with open(self.pending_file, "w") as f:
json.dump([asdict(s) for s in self._pending], f, indent=2)
def _save_reviewed(self):
with open(self.reviewed_file, "w") as f:
json.dump([asdict(s) for s in self._reviewed], f, indent=2)
# -- Public API -----------------------------------------------------------
def add_batch(self, batch: ActiveLearningBatch):
"""Persist a new batch of uncertain samples."""
self._pending.extend(batch.samples)
self._save_pending()
logger.info(
"Active learning: %d samples added (strategy=%s)", len(batch.samples), batch.strategy
)
def get_pending_samples(
self,
user_id: Optional[str] = None,
limit: int = 20,
) -> List[UncertainSample]:
"""Retrieve pending (unreviewed) samples for review."""
samples = [s for s in self._pending if not s.reviewed]
if user_id:
samples = [s for s in samples if s.user_id == user_id]
# Prioritise by uncertainty score descending
samples.sort(key=lambda s: s.uncertainty_score, reverse=True)
return samples[:limit]
def submit_review(
self,
sample_id: str,
label: Any,
reviewer_id: Optional[str] = None,
) -> bool:
"""Mark a sample as reviewed with a human-provided label."""
for sample in self._pending:
if sample.sample_id == sample_id and not sample.reviewed:
sample.reviewed = True
sample.review_label = label
sample.reviewer_id = reviewer_id
sample.reviewed_at = datetime.utcnow().isoformat()
self._reviewed.append(sample)
self._pending = [s for s in self._pending if s.sample_id != sample_id]
self._save_pending()
self._save_reviewed()
logger.info("Sample %s reviewed with label=%s", sample_id, label)
return True
return False
def get_reviewed_samples(
self,
since_hours: int = 168, # default: last 7 days
limit: int = 200,
) -> List[UncertainSample]:
"""Retrieve recently reviewed samples (used to trigger retraining)."""
cutoff = datetime.utcnow() - timedelta(hours=since_hours)
results = [
s for s in self._reviewed
if s.reviewed_at and datetime.fromisoformat(s.reviewed_at) > cutoff
]
return results[:limit]
def get_training_data(self) -> List[Dict[str, Any]]:
"""Export reviewed samples as training records."""
records = []
for s in self._reviewed:
if s.review_label is not None:
record = {
**s.features,
"label": s.review_label,
"task_id": s.task_id,
"user_id": s.user_id,
"reviewed_at": s.reviewed_at,
}
records.append(record)
return records
def get_stats(self) -> ActiveLearningStats:
"""Aggregate stats for the active learning dashboard."""
all_samples = self._pending + self._reviewed
reviewed = [s for s in all_samples if s.reviewed]
pending = [s for s in self._pending if not s.reviewed]
strategy_counts: Dict[str, int] = {}
for s in all_samples:
strategy_counts[s.uncertainty_method] = (
strategy_counts.get(s.uncertainty_method, 0) + 1
)
avg_unc = (
float(np.mean([s.uncertainty_score for s in pending])) if pending else 0.0
)
# Rough model gain estimate: each reviewed sample reduces uncertainty by ~0.1%
model_gain = min(0.20, len(reviewed) * 0.001)
return ActiveLearningStats(
total_queried=len(all_samples),
total_reviewed=len(reviewed),
total_pending=len(pending),
review_rate=len(reviewed) / max(1, len(all_samples)),
avg_uncertainty=round(avg_unc, 4),
coverage_by_strategy=strategy_counts,
model_improvement_estimate=round(model_gain, 4),
)
def should_retrain(self, min_new_samples: int = 50) -> bool:
"""Return True if enough new reviewed samples warrant model retraining."""
new_samples = self.get_reviewed_samples(since_hours=24)
return len(new_samples) >= min_new_samples
def export_for_retraining(self) -> Dict[str, Any]:
"""Export all reviewed data ready for model retraining."""
training_data = self.get_training_data()
return {
"records": training_data,
"count": len(training_data),
"exported_at": datetime.utcnow().isoformat(),
"ready_for_training": len(training_data) >= 10,
}
# ---------------------------------------------------------------------------
# Convenience factory (singleton)
# ---------------------------------------------------------------------------
_manager_instance: Optional[ActiveLearningManager] = None
def get_active_learning_manager() -> ActiveLearningManager:
global _manager_instance
if _manager_instance is None:
_manager_instance = ActiveLearningManager()
return _manager_instance
def run_active_learning_query(
candidate_pool: List[Dict[str, Any]],
strategy: str = "entropy",
batch_size: int = 20,
user_id: Optional[str] = None,
threshold: float = 0.55,
) -> Dict[str, Any]:
"""
High-level entrypoint: score a pool of prediction candidates and return
the most uncertain subset formatted for the REST API.
"""
selector = ActiveLearningSelector(strategy=strategy, threshold=threshold)
batch = selector.select_batch(candidate_pool, batch_size=batch_size, user_id=user_id)
manager = get_active_learning_manager()
manager.add_batch(batch)
return {
"batch_id": batch.batch_id,
"strategy": batch.strategy,
"samples_selected": batch.batch_size,
"pool_size": batch.total_pool_size,
"expected_model_gain": batch.expected_model_gain,
"samples": [asdict(s) for s in batch.samples],
"created_at": batch.created_at,
}
|