File size: 18,896 Bytes
c8b77b5 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 |
"""
Training Monitoring
==================
This module provides comprehensive monitoring capabilities for gradient descent
training, including gradient tracking, loss monitoring, and performance metrics.
"""
import logging
import math
import time
from typing import Dict, Optional, Any
import torch
import numpy as np
import matplotlib.pyplot as plt
from collections import defaultdict, deque
import json
logger = logging.getLogger(__name__)
class GradientMonitor:
"""
Monitor gradient statistics during training
Tracks gradient norms, distributions, and anomalies to ensure
stable training and detect potential issues.
"""
def __init__(self, max_history: int = 1000):
self.max_history = max_history
self.gradient_norms = deque(maxlen=max_history)
self.gradient_means = deque(maxlen=max_history)
self.gradient_stds = deque(maxlen=max_history)
self.gradient_maxs = deque(maxlen=max_history)
self.gradient_mins = deque(maxlen=max_history)
self.parameter_stats = defaultdict(lambda: {
'norms': deque(maxlen=max_history),
'means': deque(maxlen=max_history),
'stds': deque(maxlen=max_history)
})
self.anomaly_count = 0
self.last_anomaly_time = None
logger.info("Initialized GradientMonitor")
def update(self, gradients: Dict[str, torch.Tensor]):
"""
Update gradient statistics
Args:
gradients: Dictionary of parameter gradients
"""
total_norm = 0.0
total_mean = 0.0
total_std = 0.0
total_max = float('-inf')
total_min = float('inf')
param_count = 0
for name, grad in gradients.items():
if grad is not None:
# Compute statistics for this parameter
param_norm = grad.data.norm(2).item()
param_mean = grad.data.mean().item()
param_std = grad.data.std().item()
param_max = grad.data.max().item()
param_min = grad.data.min().item()
# Update parameter-specific stats
self.parameter_stats[name]['norms'].append(param_norm)
self.parameter_stats[name]['means'].append(param_mean)
self.parameter_stats[name]['stds'].append(param_std)
# Accumulate global stats
total_norm += param_norm ** 2
total_mean += param_mean
total_std += param_std ** 2
total_max = max(total_max, param_max)
total_min = min(total_min, param_min)
param_count += 1
# Compute global statistics
if param_count > 0:
total_norm = math.sqrt(total_norm)
total_mean /= param_count
total_std = math.sqrt(total_std / param_count)
# Store global stats
self.gradient_norms.append(total_norm)
self.gradient_means.append(total_mean)
self.gradient_stds.append(total_std)
self.gradient_maxs.append(total_max)
self.gradient_mins.append(total_min)
def detect_anomalies(self) -> Dict[str, Any]:
"""
Detect gradient anomalies
Returns:
Dictionary of detected anomalies
"""
anomalies = {
'exploding_gradients': False,
'vanishing_gradients': False,
'gradient_imbalance': False,
'nan_gradients': False,
'gradient_spikes': False
}
if len(self.gradient_norms) < 2:
return anomalies
current_norm = self.gradient_norms[-1]
# Check for exploding gradients
if current_norm > 10.0:
anomalies['exploding_gradients'] = True
self.anomaly_count += 1
self.last_anomaly_time = time.time()
logger.warning(f"Exploding gradients detected: norm={current_norm:.6f}")
# Check for vanishing gradients
if current_norm < 1e-6:
anomalies['vanishing_gradients'] = True
self.anomaly_count += 1
self.last_anomaly_time = time.time()
logger.warning(f"Vanishing gradients detected: norm={current_norm:.6f}")
# Check for gradient spikes
if len(self.gradient_norms) >= 10:
recent_norms = list(self.gradient_norms)[-10:]
avg_norm = np.mean(recent_norms[:-1])
if current_norm > 3 * avg_norm:
anomalies['gradient_spikes'] = True
logger.warning(f"Gradient spike detected: {current_norm:.6f} vs avg {avg_norm:.6f}")
# Check for NaN gradients
if math.isnan(current_norm) or math.isnan(self.gradient_means[-1]):
anomalies['nan_gradients'] = True
self.anomaly_count += 1
self.last_anomaly_time = time.time()
logger.warning("NaN gradients detected")
# Check for gradient imbalance between parameters
if len(self.parameter_stats) > 1:
param_norms = [stats['norms'][-1] for stats in self.parameter_stats.values()
if len(stats['norms']) > 0]
if param_norms and max(param_norms) / min(param_norms) > 1000:
anomalies['gradient_imbalance'] = True
logger.warning("Gradient imbalance detected between parameters")
return anomalies
def get_statistics(self) -> Dict[str, Any]:
"""
Get comprehensive gradient statistics
Returns:
Dictionary of gradient statistics
"""
if not self.gradient_norms:
return {}
stats = {
'current_norm': self.gradient_norms[-1],
'mean_norm': np.mean(self.gradient_norms),
'std_norm': np.std(self.gradient_norms),
'min_norm': min(self.gradient_norms),
'max_norm': max(self.gradient_norms),
'current_mean': self.gradient_means[-1],
'current_std': self.gradient_stds[-1],
'current_max': self.gradient_maxs[-1],
'current_min': self.gradient_mins[-1],
'anomaly_count': self.anomaly_count,
'parameter_count': len(self.parameter_stats)
}
# Add parameter-specific statistics
param_stats = {}
for name, stats_dict in self.parameter_stats.items():
if stats_dict['norms']:
param_stats[name] = {
'current_norm': stats_dict['norms'][-1],
'mean_norm': np.mean(stats_dict['norms']),
'std_norm': np.std(stats_dict['norms'])
}
stats['parameter_stats'] = param_stats
return stats
def plot_gradients(self, save_path: Optional[str] = None):
"""
Plot gradient statistics
Args:
save_path: Path to save the plot
"""
if not self.gradient_norms:
logger.warning("No gradient data to plot")
return
fig, axes = plt.subplots(2, 2, figsize=(12, 8))
# Plot gradient norms
axes[0, 0].plot(self.gradient_norms)
axes[0, 0].set_title('Gradient Norms')
axes[0, 0].set_xlabel('Step')
axes[0, 0].set_ylabel('Norm')
axes[0, 0].grid(True)
# Plot gradient means
axes[0, 1].plot(self.gradient_means)
axes[0, 1].set_title('Gradient Means')
axes[0, 1].set_xlabel('Step')
axes[0, 1].set_ylabel('Mean')
axes[0, 1].grid(True)
# Plot gradient stds
axes[1, 0].plot(self.gradient_stds)
axes[1, 0].set_title('Gradient Standard Deviations')
axes[1, 0].set_xlabel('Step')
axes[1, 0].set_ylabel('Std')
axes[1, 0].grid(True)
# Plot gradient range
axes[1, 1].plot(self.gradient_maxs, label='Max')
axes[1, 1].plot(self.gradient_mins, label='Min')
axes[1, 1].set_title('Gradient Range')
axes[1, 1].set_xlabel('Step')
axes[1, 1].set_ylabel('Value')
axes[1, 1].legend()
axes[1, 1].grid(True)
plt.tight_layout()
if save_path:
plt.savefig(save_path, dpi=300, bbox_inches='tight')
logger.info(f"Gradient plots saved to {save_path}")
plt.show()
class TrainingMonitor:
"""
Monitor training progress and performance
Tracks loss, accuracy, learning rates, and other training metrics
to provide comprehensive training insights.
"""
def __init__(self, max_history: int = 1000):
self.max_history = max_history
self.losses = deque(maxlen=max_history)
self.accuracies = deque(maxlen=max_history)
self.learning_rates = deque(maxlen=max_history)
self.training_times = deque(maxlen=max_history)
self.epoch_metrics = defaultdict(lambda: {
'loss': deque(maxlen=max_history),
'accuracy': deque(maxlen=max_history),
'learning_rate': deque(maxlen=max_history)
})
self.best_loss = float('inf')
self.best_accuracy = 0.0
self.training_start_time = time.time()
logger.info("Initialized TrainingMonitor")
def update(self, loss: float, accuracy: Optional[float] = None,
learning_rate: Optional[float] = None, epoch: Optional[int] = None):
"""
Update training metrics
Args:
loss: Current loss value
accuracy: Current accuracy (optional)
learning_rate: Current learning rate (optional)
epoch: Current epoch (optional)
"""
current_time = time.time()
# Update global metrics
self.losses.append(loss)
if accuracy is not None:
self.accuracies.append(accuracy)
if learning_rate is not None:
self.learning_rates.append(learning_rate)
self.training_times.append(current_time - self.training_start_time)
# Update epoch-specific metrics
if epoch is not None:
self.epoch_metrics[epoch]['loss'].append(loss)
if accuracy is not None:
self.epoch_metrics[epoch]['accuracy'].append(accuracy)
if learning_rate is not None:
self.epoch_metrics[epoch]['learning_rate'].append(learning_rate)
# Update best metrics
if loss < self.best_loss:
self.best_loss = loss
if accuracy is not None and accuracy > self.best_accuracy:
self.best_accuracy = accuracy
def get_statistics(self) -> Dict[str, Any]:
"""
Get comprehensive training statistics
Returns:
Dictionary of training statistics
"""
if not self.losses:
return {}
stats = {
'current_loss': self.losses[-1],
'best_loss': self.best_loss,
'mean_loss': np.mean(self.losses),
'std_loss': np.std(self.losses),
'min_loss': min(self.losses),
'max_loss': max(self.losses),
'best_accuracy': self.best_accuracy,
'total_steps': len(self.losses),
'training_time': self.training_times[-1] if self.training_times else 0
}
if self.accuracies:
stats.update({
'current_accuracy': self.accuracies[-1],
'mean_accuracy': np.mean(self.accuracies),
'std_accuracy': np.std(self.accuracies)
})
if self.learning_rates:
stats.update({
'current_learning_rate': self.learning_rates[-1],
'mean_learning_rate': np.mean(self.learning_rates),
'min_learning_rate': min(self.learning_rates),
'max_learning_rate': max(self.learning_rates)
})
return stats
def detect_convergence(self, patience: int = 10, threshold: float = 1e-4) -> bool:
"""
Detect if training has converged
Args:
patience: Number of steps to wait for improvement
threshold: Minimum improvement threshold
Returns:
True if training has converged
"""
if len(self.losses) < patience:
return False
recent_losses = list(self.losses)[-patience:]
best_recent = min(recent_losses)
improvement = self.best_loss - best_recent
return improvement < threshold
def plot_training_curves(self, save_path: Optional[str] = None):
"""
Plot training curves
Args:
save_path: Path to save the plot
"""
if not self.losses:
logger.warning("No training data to plot")
return
fig, axes = plt.subplots(2, 2, figsize=(12, 8))
# Plot loss curve
axes[0, 0].plot(self.losses)
axes[0, 0].set_title('Training Loss')
axes[0, 0].set_xlabel('Step')
axes[0, 0].set_ylabel('Loss')
axes[0, 0].grid(True)
# Plot accuracy curve
if self.accuracies:
axes[0, 1].plot(self.accuracies)
axes[0, 1].set_title('Training Accuracy')
axes[0, 1].set_xlabel('Step')
axes[0, 1].set_ylabel('Accuracy')
axes[0, 1].grid(True)
# Plot learning rate curve
if self.learning_rates:
axes[1, 0].plot(self.learning_rates)
axes[1, 0].set_title('Learning Rate')
axes[1, 0].set_xlabel('Step')
axes[1, 0].set_ylabel('Learning Rate')
axes[1, 0].grid(True)
# Plot training time
if self.training_times:
axes[1, 1].plot(self.training_times)
axes[1, 1].set_title('Training Time')
axes[1, 1].set_xlabel('Step')
axes[1, 1].set_ylabel('Time (seconds)')
axes[1, 1].grid(True)
plt.tight_layout()
if save_path:
plt.savefig(save_path, dpi=300, bbox_inches='tight')
logger.info(f"Training curves saved to {save_path}")
plt.show()
def save_metrics(self, file_path: str):
"""
Save training metrics to file
Args:
file_path: Path to save the metrics
"""
metrics = {
'losses': list(self.losses),
'accuracies': list(self.accuracies),
'learning_rates': list(self.learning_rates),
'training_times': list(self.training_times),
'best_loss': self.best_loss,
'best_accuracy': self.best_accuracy,
'statistics': self.get_statistics()
}
with open(file_path, 'w') as f:
json.dump(metrics, f, indent=2)
logger.info(f"Training metrics saved to {file_path}")
def load_metrics(self, file_path: str):
"""
Load training metrics from file
Args:
file_path: Path to load the metrics from
"""
with open(file_path, 'r') as f:
metrics = json.load(f)
self.losses = deque(metrics['losses'], maxlen=self.max_history)
self.accuracies = deque(metrics['accuracies'], maxlen=self.max_history)
self.learning_rates = deque(metrics['learning_rates'], maxlen=self.max_history)
self.training_times = deque(metrics['training_times'], maxlen=self.max_history)
self.best_loss = metrics['best_loss']
self.best_accuracy = metrics['best_accuracy']
logger.info(f"Training metrics loaded from {file_path}")
class PerformanceMonitor:
"""
Monitor system performance during training
Tracks memory usage, compute time, and other system metrics
to optimize training efficiency.
"""
def __init__(self):
self.memory_usage = deque(maxlen=1000)
self.compute_times = deque(maxlen=1000)
self.gpu_usage = deque(maxlen=1000)
self.step_times = []
self.forward_times = []
self.backward_times = []
self.optimizer_times = []
logger.info("Initialized PerformanceMonitor")
def update_memory(self, memory_mb: float):
"""Update memory usage"""
self.memory_usage.append(memory_mb)
def update_compute_time(self, time_seconds: float):
"""Update compute time"""
self.compute_times.append(time_seconds)
def update_gpu_usage(self, gpu_percent: float):
"""Update GPU usage"""
self.gpu_usage.append(gpu_percent)
def time_step(self, step_name: str):
"""Context manager for timing steps"""
return StepTimer(self, step_name)
def get_statistics(self) -> Dict[str, Any]:
"""Get performance statistics"""
stats = {}
if self.memory_usage:
stats['memory'] = {
'current_mb': self.memory_usage[-1],
'mean_mb': np.mean(self.memory_usage),
'max_mb': max(self.memory_usage)
}
if self.compute_times:
stats['compute'] = {
'current_seconds': self.compute_times[-1],
'mean_seconds': np.mean(self.compute_times),
'total_seconds': sum(self.compute_times)
}
if self.gpu_usage:
stats['gpu'] = {
'current_percent': self.gpu_usage[-1],
'mean_percent': np.mean(self.gpu_usage),
'max_percent': max(self.gpu_usage)
}
return stats
class StepTimer:
"""Context manager for timing training steps"""
def __init__(self, monitor: PerformanceMonitor, step_name: str):
self.monitor = monitor
self.step_name = step_name
self.start_time = None
def __enter__(self):
self.start_time = time.time()
return self
def __exit__(self, exc_type, exc_val, exc_tb):
elapsed = time.time() - self.start_time
if self.step_name == 'forward':
self.monitor.forward_times.append(elapsed)
elif self.step_name == 'backward':
self.monitor.backward_times.append(elapsed)
elif self.step_name == 'optimizer':
self.monitor.optimizer_times.append(elapsed)
else:
self.monitor.step_times.append(elapsed)
|