File size: 21,117 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 |
"""
Gradient Descent Training Loop
=============================
This module implements the main training loop that orchestrates gradient descent
optimization with backpropagation for the MangoMAS multi-agent system.
The training loop includes:
- Forward and backward passes
- Gradient computation and optimization
- Learning rate scheduling
- Comprehensive monitoring and logging
- Model checkpointing and validation
"""
import logging
import time
import math
from typing import Dict, List, Optional, Tuple, Any
import torch
import torch.nn as nn
from pathlib import Path
from .optimizers import OptimizerFactory
from .backpropagation import BackpropagationEngine, LoRABackpropagationEngine
from .loss_functions import LossFunctionFactory
from .schedulers import SchedulerFactory
from .monitoring import GradientMonitor, TrainingMonitor, PerformanceMonitor
logger = logging.getLogger(__name__)
class GradientDescentTrainer:
"""
Main training class that orchestrates gradient descent optimization
This class provides a complete training pipeline with:
- Real gradient descent and backpropagation
- Comprehensive monitoring and logging
- Model checkpointing and validation
- Integration with MangoMAS agent system
"""
def __init__(self,
optimizer_type: str = 'adam',
learning_rate: float = 1e-3,
scheduler_type: str = 'cosine',
loss_function_type: str = 'cross_entropy',
device: torch.device = None,
max_grad_norm: float = 1.0,
gradient_accumulation_steps: int = 1,
mixed_precision: bool = False,
**kwargs):
self.optimizer_type = optimizer_type
self.learning_rate = learning_rate
self.scheduler_type = scheduler_type
self.loss_function_type = loss_function_type
self.device = device or torch.device('cuda' if torch.cuda.is_available() else 'cpu')
self.max_grad_norm = max_grad_norm
self.gradient_accumulation_steps = gradient_accumulation_steps
self.mixed_precision = mixed_precision
# Initialize components
self.optimizer = None
self.scheduler = None
self.loss_function = None
self.backprop_engine = None
# Monitoring
self.gradient_monitor = GradientMonitor()
self.training_monitor = TrainingMonitor()
self.performance_monitor = PerformanceMonitor()
# Training state
self.current_epoch = 0
self.current_step = 0
self.best_loss = float('inf')
self.training_start_time = None
# Configuration
self.config = {
'optimizer_type': optimizer_type,
'learning_rate': learning_rate,
'scheduler_type': scheduler_type,
'loss_function_type': loss_function_type,
'max_grad_norm': max_grad_norm,
'gradient_accumulation_steps': gradient_accumulation_steps,
'mixed_precision': mixed_precision,
**kwargs
}
logger.info(f"Initialized GradientDescentTrainer with config: {self.config}")
def setup_training(self, model: nn.Module, training_data: List[Dict[str, Any]]):
"""
Setup training components
Args:
model: The neural network model to train
training_data: Training dataset
"""
logger.info("Setting up training components...")
# Move model to device
model.to(self.device)
# Get trainable parameters
trainable_params = [p for p in model.parameters() if p.requires_grad]
logger.info(f"Found {len(trainable_params)} trainable parameters")
# Initialize optimizer
optimizer_config = OptimizerFactory.get_default_config(self.optimizer_type)
optimizer_config.update({'lr': self.learning_rate})
self.optimizer = OptimizerFactory.create_optimizer(
self.optimizer_type, trainable_params, **optimizer_config
)
# Initialize scheduler
scheduler_config = SchedulerFactory.get_default_config(self.scheduler_type)
scheduler_config.update({'total_steps': len(training_data)})
self.scheduler = SchedulerFactory.create_scheduler(
self.scheduler_type, self.optimizer, **scheduler_config
)
# Initialize loss function
loss_config = LossFunctionFactory.get_default_config(self.loss_function_type)
self.loss_function = LossFunctionFactory.create_loss_function(
self.loss_function_type, **loss_config
)
# Initialize backpropagation engine
if hasattr(model, 'lora_params'):
# LoRA model
self.backprop_engine = LoRABackpropagationEngine(
model, model.lora_params, self.device
)
else:
# Standard model
self.backprop_engine = BackpropagationEngine(model, self.device)
logger.info("Training setup complete")
def train_epoch(self, model: nn.Module, training_data: List[Dict[str, Any]],
epoch: int) -> Dict[str, float]:
"""
Train for one epoch using gradient descent and backpropagation
Args:
model: The neural network model
training_data: Training dataset
epoch: Current epoch number
Returns:
Dictionary of training metrics
"""
logger.info(f"Starting epoch {epoch}")
model.train()
epoch_loss = 0.0
epoch_accuracy = 0.0
num_batches = 0
# Process training data in batches
batch_size = 32 # Default batch size
num_batches = math.ceil(len(training_data) / batch_size)
for batch_idx in range(num_batches):
start_idx = batch_idx * batch_size
end_idx = min(start_idx + batch_size, len(training_data))
batch_data = training_data[start_idx:end_idx]
# Process batch
batch_metrics = self.train_batch(model, batch_data, epoch, batch_idx)
epoch_loss += batch_metrics['loss']
epoch_accuracy += batch_metrics.get('accuracy', 0.0)
# Update step counter
self.current_step += 1
# Log progress
if batch_idx % 10 == 0:
logger.info(f"Epoch {epoch}, Batch {batch_idx}/{num_batches}, "
f"Loss: {batch_metrics['loss']:.4f}")
# Compute epoch averages
avg_loss = epoch_loss / num_batches
avg_accuracy = epoch_accuracy / num_batches if num_batches > 0 else 0.0
# Update monitors
self.training_monitor.update(
loss=avg_loss,
accuracy=avg_accuracy,
learning_rate=self.optimizer.lr,
epoch=epoch
)
# Update scheduler
self.scheduler.step(epoch=epoch, metrics={'loss': avg_loss})
logger.info(f"Epoch {epoch} complete - Loss: {avg_loss:.4f}, "
f"Accuracy: {avg_accuracy:.4f}, LR: {self.optimizer.lr:.6f}")
return {
'loss': avg_loss,
'accuracy': avg_accuracy,
'learning_rate': self.optimizer.lr,
'num_batches': num_batches
}
def train_batch(self, model: nn.Module, batch_data: List[Dict[str, Any]],
epoch: int, batch_idx: int) -> Dict[str, float]:
"""
Train on a single batch using gradient descent and backpropagation
Args:
model: The neural network model
batch_data: Batch of training data
epoch: Current epoch number
batch_idx: Current batch index
Returns:
Dictionary of batch metrics
"""
# Prepare batch data
inputs, targets = self._prepare_batch(batch_data)
# Forward pass
with self.performance_monitor.time_step('forward'):
outputs = model(inputs)
# Compute loss
loss = self.loss_function(outputs, targets)
# Scale loss for gradient accumulation
if self.gradient_accumulation_steps > 1:
loss = loss / self.gradient_accumulation_steps
# Backward pass
with self.performance_monitor.time_step('backward'):
loss.backward()
# Gradient accumulation
if (batch_idx + 1) % self.gradient_accumulation_steps == 0:
# Apply gradient clipping
grad_norm = self.backprop_engine.apply_gradient_clipping(self.max_grad_norm)
# Get gradients for monitoring
gradients = self.backprop_engine.compute_gradients(loss, retain_graph=False)
self.gradient_monitor.update(gradients)
# Optimizer step
with self.performance_monitor.time_step('optimizer'):
self.optimizer.step()
# Zero gradients
self.optimizer.zero_grad()
# Update performance monitoring
self.performance_monitor.update_compute_time(time.time() - self.training_start_time)
# Compute accuracy (if applicable)
accuracy = self._compute_accuracy(outputs, targets)
return {
'loss': loss.item() * self.gradient_accumulation_steps,
'accuracy': accuracy,
'grad_norm': grad_norm if 'grad_norm' in locals() else 0.0
}
def _prepare_batch(self, batch_data: List[Dict[str, Any]]) -> Tuple[torch.Tensor, torch.Tensor]:
"""
Prepare batch data for training
Args:
batch_data: Raw batch data
Returns:
Tuple of (inputs, targets) tensors
"""
# Extract inputs and targets
inputs = []
targets = []
for item in batch_data:
# Convert text to tokens (simplified)
if 'instruction' in item and 'response' in item:
# For text generation tasks
input_text = item['instruction']
target_text = item['response']
# Simple tokenization (in practice, use proper tokenizer)
input_tokens = self._simple_tokenize(input_text)
target_tokens = self._simple_tokenize(target_text)
inputs.append(input_tokens)
targets.append(target_tokens)
# Convert to tensors
if inputs and targets:
# Pad sequences to same length
max_len = max(len(seq) for seq in inputs + targets)
inputs = [seq + [0] * (max_len - len(seq)) for seq in inputs]
targets = [seq + [0] * (max_len - len(seq)) for seq in targets]
inputs_tensor = torch.tensor(inputs, dtype=torch.long, device=self.device)
targets_tensor = torch.tensor(targets, dtype=torch.long, device=self.device)
else:
# Fallback: create dummy data
batch_size = len(batch_data)
seq_len = 128
inputs_tensor = torch.randint(0, 1000, (batch_size, seq_len), device=self.device)
targets_tensor = torch.randint(0, 1000, (batch_size, seq_len), device=self.device)
return inputs_tensor, targets_tensor
def _simple_tokenize(self, text: str) -> List[int]:
"""
Simple tokenization for demonstration
Args:
text: Input text
Returns:
List of token IDs
"""
# Simple character-based tokenization
tokens = []
for char in text[:100]: # Limit length
tokens.append(ord(char) % 1000) # Map to vocabulary
return tokens
def _compute_accuracy(self, outputs: torch.Tensor, targets: torch.Tensor) -> float:
"""
Compute accuracy for the batch
Args:
outputs: Model outputs
targets: Target values
Returns:
Accuracy score
"""
if outputs.dim() > 1 and outputs.size(1) > 1:
# Classification task
predictions = torch.argmax(outputs, dim=1)
if targets.dim() == 1:
correct = (predictions == targets).float().sum()
accuracy = correct / targets.size(0)
else:
# Multi-label case
accuracy = 0.0
else:
# Regression task - use a simple threshold
accuracy = 0.0
return accuracy.item() if isinstance(accuracy, torch.Tensor) else accuracy
def validate(self, model: nn.Module, validation_data: List[Dict[str, Any]]) -> Dict[str, float]:
"""
Validate the model
Args:
model: The neural network model
validation_data: Validation dataset
Returns:
Dictionary of validation metrics
"""
logger.info("Running validation...")
model.eval()
total_loss = 0.0
total_accuracy = 0.0
num_batches = 0
with torch.no_grad():
batch_size = 32
num_batches = math.ceil(len(validation_data) / batch_size)
for batch_idx in range(num_batches):
start_idx = batch_idx * batch_size
end_idx = min(start_idx + batch_size, len(validation_data))
batch_data = validation_data[start_idx:end_idx]
# Prepare batch
inputs, targets = self._prepare_batch(batch_data)
# Forward pass
outputs = model(inputs)
# Compute loss
loss = self.loss_function(outputs, targets)
total_loss += loss.item()
# Compute accuracy
accuracy = self._compute_accuracy(outputs, targets)
total_accuracy += accuracy
# Compute averages
avg_loss = total_loss / num_batches if num_batches > 0 else 0.0
avg_accuracy = total_accuracy / num_batches if num_batches > 0 else 0.0
logger.info(f"Validation - Loss: {avg_loss:.4f}, Accuracy: {avg_accuracy:.4f}")
return {
'val_loss': avg_loss,
'val_accuracy': avg_accuracy
}
def train(self, model: nn.Module, training_data: List[Dict[str, Any]],
validation_data: Optional[List[Dict[str, Any]]] = None,
num_epochs: int = 10, save_dir: Optional[str] = None) -> Dict[str, Any]:
"""
Complete training loop with gradient descent and backpropagation
Args:
model: The neural network model to train
training_data: Training dataset
validation_data: Validation dataset (optional)
num_epochs: Number of training epochs
save_dir: Directory to save checkpoints
Returns:
Dictionary of training results
"""
logger.info(f"Starting training for {num_epochs} epochs")
# Setup training
self.setup_training(model, training_data)
# Initialize training state
self.training_start_time = time.time()
self.current_epoch = 0
self.current_step = 0
# Training history
training_history = []
validation_history = []
# Main training loop
for epoch in range(num_epochs):
self.current_epoch = epoch
# Train epoch
epoch_metrics = self.train_epoch(model, training_data, epoch)
training_history.append(epoch_metrics)
# Validation
if validation_data:
val_metrics = self.validate(model, validation_data)
validation_history.append(val_metrics)
# Update best loss
if val_metrics['val_loss'] < self.best_loss:
self.best_loss = val_metrics['val_loss']
# Save best model
if save_dir:
self.save_checkpoint(model, save_dir, epoch, val_metrics)
# Check for convergence
if self.training_monitor.detect_convergence():
logger.info("Training converged, stopping early")
break
# Log epoch summary
logger.info(f"Epoch {epoch} Summary:")
logger.info(f" Training Loss: {epoch_metrics['loss']:.4f}")
logger.info(f" Training Accuracy: {epoch_metrics['accuracy']:.4f}")
if validation_data:
logger.info(f" Validation Loss: {val_metrics['val_loss']:.4f}")
logger.info(f" Validation Accuracy: {val_metrics['val_accuracy']:.4f}")
logger.info(f" Learning Rate: {self.optimizer.lr:.6f}")
# Training complete
training_time = time.time() - self.training_start_time
# Get final statistics
gradient_stats = self.gradient_monitor.get_statistics()
training_stats = self.training_monitor.get_statistics()
performance_stats = self.performance_monitor.get_statistics()
results = {
'training_history': training_history,
'validation_history': validation_history,
'final_metrics': {
'best_loss': self.best_loss,
'final_loss': training_history[-1]['loss'] if training_history else 0.0,
'final_accuracy': training_history[-1]['accuracy'] if training_history else 0.0,
'training_time': training_time,
'total_steps': self.current_step,
'total_epochs': self.current_epoch + 1
},
'gradient_stats': gradient_stats,
'training_stats': training_stats,
'performance_stats': performance_stats,
'config': self.config
}
logger.info("Training complete!")
logger.info(f"Final Loss: {results['final_metrics']['final_loss']:.4f}")
logger.info(f"Best Loss: {results['final_metrics']['best_loss']:.4f}")
logger.info(f"Training Time: {training_time:.2f} seconds")
return results
def save_checkpoint(self, model: nn.Module, save_dir: str, epoch: int,
metrics: Dict[str, float]):
"""
Save model checkpoint
Args:
model: The neural network model
save_dir: Directory to save checkpoint
epoch: Current epoch
metrics: Training metrics
"""
save_path = Path(save_dir)
save_path.mkdir(parents=True, exist_ok=True)
checkpoint = {
'epoch': epoch,
'model_state_dict': model.state_dict(),
'optimizer_state_dict': self.optimizer.state_dict(),
'scheduler_state_dict': self.scheduler.state_dict(),
'best_loss': self.best_loss,
'metrics': metrics,
'config': self.config
}
checkpoint_path = save_path / f'checkpoint_epoch_{epoch}.pt'
torch.save(checkpoint, checkpoint_path)
logger.info(f"Checkpoint saved to {checkpoint_path}")
def load_checkpoint(self, model: nn.Module, checkpoint_path: str):
"""
Load model checkpoint
Args:
model: The neural network model
checkpoint_path: Path to checkpoint file
"""
checkpoint = torch.load(checkpoint_path, map_location=self.device)
model.load_state_dict(checkpoint['model_state_dict'])
self.optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
self.scheduler.load_state_dict(checkpoint['scheduler_state_dict'])
self.best_loss = checkpoint['best_loss']
logger.info(f"Checkpoint loaded from {checkpoint_path}")
def get_training_summary(self) -> Dict[str, Any]:
"""
Get comprehensive training summary
Returns:
Dictionary of training summary
"""
return {
'gradient_stats': self.gradient_monitor.get_statistics(),
'training_stats': self.training_monitor.get_statistics(),
'performance_stats': self.performance_monitor.get_statistics(),
'anomalies': self.gradient_monitor.detect_anomalies(),
'convergence': self.training_monitor.detect_convergence()
}
|