File size: 9,101 Bytes
077be11 | 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 | """
XGBoost Busy Detector - Hugging Face Inference Endpoint Handler
Custom handler for HF Inference Endpoints.
Loads XGBoost model, applies normalization, runs evidence accumulation scoring,
and returns busy_score + confidence + recommendation.
Derived from: src/normalization.py, src/scoring_engine.py, src/model.py
"""
from typing import Dict, Any, Tuple
import json
import math
import numpy as np
import pickle
from pathlib import Path
class EndpointHandler:
"""HF Inference Endpoint handler for XGBoost busy detection."""
def __init__(self, path: str = "."):
model_dir = Path(path)
# --- Load XGBoost model ---
model_path = None
for candidate in [
model_dir / "model.pkl",
model_dir / "busy_detector_v1.pkl",
model_dir / "busy_detector_5k.pkl",
]:
if candidate.exists():
model_path = candidate
break
if model_path is None:
raise FileNotFoundError(
f"No model file found in {model_dir}. "
"Expected model.pkl, busy_detector_v1.pkl, or busy_detector_5k.pkl"
)
with open(model_path, "rb") as f:
saved = pickle.load(f)
# Handle both raw model and dict-wrapped model
if isinstance(saved, dict):
self.model = saved.get("model") or saved.get("booster")
self.feature_names = saved.get("feature_names")
else:
self.model = saved
self.feature_names = None
print(f"✓ XGBoost model loaded from {model_path}")
# --- Load feature ranges ---
ranges_path = model_dir / "feature_ranges.json"
with open(ranges_path) as f:
ranges_data = json.load(f)
self.voice_ranges = ranges_data["voice_ranges"]
self.text_ranges = ranges_data["text_ranges"]
self.voice_order = ranges_data["voice_feature_order"]
self.text_order = ranges_data["text_feature_order"]
# --- Load scoring rules ---
rules_path = model_dir / "scoring_rules.json"
with open(rules_path) as f:
self.scoring = json.load(f)
self.weights = self.scoring["weights"]
self.thresholds = self.scoring["thresholds"]
print("✓ Feature ranges and scoring rules loaded")
# --------------------------------------------------------------------- #
# Public interface
# --------------------------------------------------------------------- #
def __call__(self, data: Dict[str, Any]) -> Dict[str, Any]:
"""
Entrypoint for HF Inference Endpoints.
Expected input (JSON):
{
"inputs": {
"audio_features": { "v1_snr": 15.0, ... },
"text_features": { "t1_explicit_busy": 0.8, ... }
}
}
Returns:
{
"busy_score": 0.72,
"confidence": 0.85,
"recommendation": "EXIT",
"ml_probability": 0.65,
"evidence_details": [...]
}
"""
# HF wraps payload in "inputs"
inputs = data.get("inputs", data)
audio_features = inputs.get("audio_features", {})
text_features = inputs.get("text_features", {})
# 1. Normalize
normalized = self._normalize_features(audio_features, text_features)
# 2. XGBoost inference
import xgboost as xgb
dmatrix = xgb.DMatrix(normalized.reshape(1, -1))
ml_prob = float(self.model.predict(dmatrix)[0])
# 3. Evidence accumulation scoring
final_score, confidence, details = self._score_with_evidence(
ml_prob, audio_features, text_features
)
# 4. Recommendation
recommendation = self._get_recommendation(final_score)
return {
"busy_score": round(final_score, 4),
"confidence": round(confidence, 4),
"recommendation": recommendation,
"ml_probability": round(ml_prob, 4),
"evidence_details": details,
}
# --------------------------------------------------------------------- #
# Normalization (mirrors src/normalization.py FeatureNormalizer)
# --------------------------------------------------------------------- #
def _normalize_value(self, value: float, min_val: float, max_val: float) -> float:
if max_val == min_val:
return 0.0
value = max(min_val, min(max_val, value))
return (value - min_val) / (max_val - min_val)
def _normalize_features(
self,
audio_features: Dict[str, float],
text_features: Dict[str, float],
) -> np.ndarray:
"""Min-max normalize all 26 features and concatenate."""
voice_norm = []
for feat in self.voice_order:
val = audio_features.get(feat, 0.0)
lo, hi = self.voice_ranges[feat]
voice_norm.append(self._normalize_value(val, lo, hi))
text_norm = []
for feat in self.text_order:
val = text_features.get(feat, 0.0)
lo, hi = self.text_ranges[feat]
text_norm.append(self._normalize_value(val, lo, hi))
return np.array(voice_norm + text_norm, dtype=np.float32)
# --------------------------------------------------------------------- #
# Evidence scoring (mirrors src/scoring_engine.py ScoringEngine)
# --------------------------------------------------------------------- #
@staticmethod
def _sigmoid(x: float) -> float:
return 1.0 / (1.0 + math.exp(-x))
@staticmethod
def _logit(p: float) -> float:
p = max(0.01, min(0.99, p))
return math.log(p / (1.0 - p))
def _score_with_evidence(
self,
ml_prob: float,
audio_features: Dict[str, float],
text_features: Dict[str, float],
) -> Tuple[float, float, list]:
"""Evidence accumulation scoring exactly matching ScoringEngine.calculate_score."""
evidence = 0.0
details = []
# --- Text evidence ---
explicit = text_features.get("t1_explicit_busy", 0.0)
if explicit > 0.5:
pts = self.weights["explicit_busy"] * explicit
evidence += pts
details.append(f"Explicit Intent (+{pts:.1f})")
explicit_free = text_features.get("t0_explicit_free", 0.0)
if explicit_free > 0.5:
pts = self.weights["explicit_free"] * explicit_free
evidence += pts
details.append(f"Explicit Free ({pts:.1f})")
short_ratio = text_features.get("t3_short_ratio", 0.0)
if short_ratio > 0.3:
pts = self.weights["short_answers"] * short_ratio
evidence += pts
details.append(f"Brief Responses (+{pts:.1f})")
deflection = text_features.get("t6_deflection", 0.0)
if deflection > 0.1:
pts = self.weights["deflection"] * deflection
evidence += pts
details.append(f"Deflection (+{pts:.1f})")
# --- Audio evidence ---
traffic = audio_features.get("v2_noise_traffic", 0.0)
if traffic > 0.5:
pts = self.weights["traffic_noise"] * traffic
evidence += pts
details.append(f"Traffic Context (+{pts:.1f})")
rate = audio_features.get("v3_speech_rate", 0.0)
if rate > 3.5:
pts = self.weights["rushed_speech"]
evidence += pts
details.append(f"Rushed Speech (+{pts:.1f})")
pitch_std = audio_features.get("v5_pitch_std", 0.0)
if pitch_std > 80.0:
evidence += 0.5
details.append("Voice Stress (+0.5)")
emotion_stress = audio_features.get("v11_emotion_stress", 0.0)
if emotion_stress > 0.6:
pts = self.weights["emotion_stress"] * emotion_stress
evidence += pts
details.append(f"Emotional Stress (+{pts:.1f})")
emotion_energy = audio_features.get("v12_emotion_energy", 0.0)
if emotion_energy > 0.7:
pts = self.weights["emotion_energy"] * emotion_energy
evidence += pts
details.append(f"High Energy (+{pts:.1f})")
# --- ML baseline ---
ml_evidence = self._logit(ml_prob) * self.weights["ml_model_factor"]
evidence += ml_evidence
details.append(f"ML Baseline ({ml_evidence:+.1f})")
# --- Final ---
final_score = self._sigmoid(evidence)
confidence = float(math.tanh(abs(evidence) / 2.0))
return final_score, confidence, details
def _get_recommendation(self, score: float) -> str:
if score < self.thresholds["continue"]:
return "CONTINUE"
elif score < self.thresholds["check_in"]:
return "CHECK_IN"
else:
return "EXIT"
|