Spaces:
Runtime error
Runtime error
File size: 8,956 Bytes
c3efd49 |
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 |
"""Anomaly detection for training monitoring."""
import numpy as np
from typing import List, Dict, Optional, Callable
from collections import deque
import logging
logger = logging.getLogger(__name__)
class AnomalyDetector:
"""
Detects anomalies during training.
Monitors for reward collapse, gradient explosion, and other issues.
"""
def __init__(
self,
window_size: int = 10,
alert_callback: Optional[Callable] = None
):
"""
Initialize anomaly detector.
Args:
window_size: Size of sliding window for detection
alert_callback: Optional callback function for alerts
"""
self.window_size = window_size
self.alert_callback = alert_callback or self._default_alert
# Sliding windows for metrics
self.reward_window = deque(maxlen=window_size)
self.loss_window = deque(maxlen=window_size)
self.gradient_window = deque(maxlen=window_size)
# Alert history
self.alerts = []
logger.info(f"AnomalyDetector initialized: window_size={window_size}")
def _default_alert(self, alert_type: str, message: str, severity: str) -> None:
"""
Default alert handler.
Args:
alert_type: Type of alert
message: Alert message
severity: Severity level
"""
log_func = {
'critical': logger.critical,
'warning': logger.warning,
'info': logger.info
}.get(severity, logger.warning)
log_func(f"[{alert_type}] {message}")
def update(
self,
reward: Optional[float] = None,
loss: Optional[float] = None,
gradient_norm: Optional[float] = None
) -> List[Dict[str, str]]:
"""
Update detector with new metrics and check for anomalies.
Args:
reward: Current reward value
loss: Current loss value
gradient_norm: Current gradient norm
Returns:
List of detected anomalies
"""
anomalies = []
# Update windows
if reward is not None:
self.reward_window.append(reward)
if loss is not None:
self.loss_window.append(loss)
if gradient_norm is not None:
self.gradient_window.append(gradient_norm)
# Check for anomalies
if len(self.reward_window) >= self.window_size:
reward_anomaly = self.detect_reward_collapse()
if reward_anomaly:
anomalies.append(reward_anomaly)
if len(self.gradient_window) >= 3: # Need fewer samples for gradient check
gradient_anomaly = self.detect_gradient_explosion()
if gradient_anomaly:
anomalies.append(gradient_anomaly)
if len(self.loss_window) >= self.window_size:
loss_anomaly = self.detect_loss_divergence()
if loss_anomaly:
anomalies.append(loss_anomaly)
# Store and alert
for anomaly in anomalies:
self.alerts.append(anomaly)
self.alert_callback(
anomaly['type'],
anomaly['message'],
anomaly['severity']
)
return anomalies
def detect_reward_collapse(self) -> Optional[Dict[str, str]]:
"""
Detect reward collapse (rewards stop changing).
Returns:
Anomaly dictionary if detected, None otherwise
"""
if len(self.reward_window) < self.window_size:
return None
rewards = list(self.reward_window)
# Check if variance is very low
variance = np.var(rewards)
if variance < 1e-6:
return {
'type': 'reward_collapse',
'message': f'Reward collapse detected: variance={variance:.2e}',
'severity': 'critical',
'details': {
'variance': variance,
'mean_reward': np.mean(rewards)
}
}
# Check if rewards are consistently decreasing
if len(rewards) >= 5:
recent_trend = np.polyfit(range(len(rewards)), rewards, 1)[0]
if recent_trend < -0.01: # Significant negative trend
return {
'type': 'reward_decline',
'message': f'Reward declining: trend={recent_trend:.4f}',
'severity': 'warning',
'details': {
'trend': recent_trend,
'mean_reward': np.mean(rewards)
}
}
return None
def detect_gradient_explosion(self) -> Optional[Dict[str, str]]:
"""
Detect gradient explosion (very large gradients).
Returns:
Anomaly dictionary if detected, None otherwise
"""
if len(self.gradient_window) < 3:
return None
gradients = list(self.gradient_window)
latest_gradient = gradients[-1]
# Check for very large gradient
if latest_gradient > 100.0:
return {
'type': 'gradient_explosion',
'message': f'Gradient explosion detected: norm={latest_gradient:.2f}',
'severity': 'critical',
'details': {
'gradient_norm': latest_gradient,
'mean_gradient': np.mean(gradients)
}
}
# Check for rapidly increasing gradients
if len(gradients) >= 3:
gradient_growth = gradients[-1] / (gradients[-3] + 1e-8)
if gradient_growth > 10.0:
return {
'type': 'gradient_growth',
'message': f'Rapid gradient growth: {gradient_growth:.2f}x',
'severity': 'warning',
'details': {
'growth_factor': gradient_growth,
'current_gradient': latest_gradient
}
}
return None
def detect_loss_divergence(self) -> Optional[Dict[str, str]]:
"""
Detect loss divergence (loss increasing or becoming NaN/Inf).
Returns:
Anomaly dictionary if detected, None otherwise
"""
if len(self.loss_window) < self.window_size:
return None
losses = list(self.loss_window)
latest_loss = losses[-1]
# Check for NaN or Inf
if np.isnan(latest_loss) or np.isinf(latest_loss):
return {
'type': 'loss_invalid',
'message': f'Invalid loss detected: {latest_loss}',
'severity': 'critical',
'details': {
'loss_value': str(latest_loss)
}
}
# Check for consistently increasing loss
if len(losses) >= 5:
loss_trend = np.polyfit(range(len(losses)), losses, 1)[0]
if loss_trend > 0.1: # Significant positive trend
return {
'type': 'loss_divergence',
'message': f'Loss diverging: trend={loss_trend:.4f}',
'severity': 'warning',
'details': {
'trend': loss_trend,
'current_loss': latest_loss,
'mean_loss': np.mean(losses)
}
}
return None
def get_alerts(self) -> List[Dict[str, str]]:
"""
Get all alerts.
Returns:
List of alert dictionaries
"""
return self.alerts
def get_recent_alerts(self, n: int = 10) -> List[Dict[str, str]]:
"""
Get most recent alerts.
Args:
n: Number of recent alerts to return
Returns:
List of recent alert dictionaries
"""
return self.alerts[-n:]
def clear_alerts(self) -> None:
"""Clear all alerts."""
self.alerts.clear()
logger.info("Alerts cleared")
def get_summary(self) -> Dict[str, any]:
"""
Get summary of detected anomalies.
Returns:
Summary dictionary
"""
alert_types = {}
for alert in self.alerts:
alert_type = alert['type']
alert_types[alert_type] = alert_types.get(alert_type, 0) + 1
return {
'total_alerts': len(self.alerts),
'alert_types': alert_types,
'recent_alerts': self.get_recent_alerts(5)
}
|