File size: 20,217 Bytes
d187b57 |
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 |
# training_config.py
from asyncio.log import logger
from dataclasses import dataclass
from typing import Dict, List
import json
import torch
import numpy as np
from pathlib import Path
from contextlib import nullcontext
from dataclasses import asdict
import os
@dataclass
class DynamicClassWeights:
"""Handles class weights per language using dynamic batch statistics"""
weights_file: str = 'weights/language_class_weights.json'
def __init__(self, weights_file: str = 'weights/language_class_weights.json'):
self.weights_file = weights_file
self.toxicity_labels = ['toxic', 'severe_toxic', 'obscene', 'threat', 'insult', 'identity_hate']
self.language_columns = ['en', 'es', 'fr', 'it', 'tr', 'pt', 'ru']
# Initialize base scaling factors from file if available
try:
with open(self.weights_file, 'r') as f:
data = json.load(f)
self.lang_scaling = {}
for lang in self.language_columns:
if lang in data['weights']:
# Calculate average scaling per language
scales = [float(data['weights'][lang][label]['1'])
for label in self.toxicity_labels]
self.lang_scaling[lang] = sum(scales) / len(scales)
else:
self.lang_scaling[lang] = 1.0
except Exception as e:
logger.warning(f"Could not load weights from {self.weights_file}: {str(e)}")
self._initialize_defaults()
# Initialize running statistics for each language
self.running_stats = {lang: {
'pos_counts': torch.zeros(len(self.toxicity_labels)),
'total_counts': torch.zeros(len(self.toxicity_labels)),
'smoothing_factor': 0.95 # EMA smoothing factor
} for lang in self.language_columns}
def _initialize_defaults(self):
"""Initialize safe default scaling factors"""
self.lang_scaling = {lang: 1.0 for lang in self.language_columns}
def _update_running_stats(self, langs, labels):
"""Update running statistics for each language"""
unique_langs = set(langs)
for lang in unique_langs:
if lang not in self.running_stats:
continue
lang_mask = torch.tensor([l == lang for l in langs], dtype=torch.bool)
lang_labels = labels[lang_mask]
if len(lang_labels) == 0:
continue
# Calculate current batch statistics
pos_count = lang_labels.sum(dim=0).float()
total_count = torch.full_like(pos_count, len(lang_labels))
# Update running statistics with EMA
alpha = self.running_stats[lang]['smoothing_factor']
self.running_stats[lang]['pos_counts'] = (
alpha * self.running_stats[lang]['pos_counts'] +
(1 - alpha) * pos_count
)
self.running_stats[lang]['total_counts'] = (
alpha * self.running_stats[lang]['total_counts'] +
(1 - alpha) * total_count
)
def get_weights_for_batch(self, langs: List[str], labels: torch.Tensor, device: torch.device) -> Dict[str, torch.Tensor]:
"""
Calculate dynamic weights and focal parameters based on batch and historical statistics
Args:
langs: List of language codes
labels: Binary labels tensor [batch_size, num_labels]
device: Target device for tensors
Returns:
Dict with weights, alpha, and gamma tensors
"""
try:
batch_size = len(langs)
num_labels = labels.size(1)
# Update running statistics
self._update_running_stats(langs, labels)
# Calculate positive ratio per language in current batch
lang_pos_ratios = {}
batch_pos_ratios = torch.zeros(num_labels, device=device)
lang_counts = {}
for lang in set(langs):
lang_mask = torch.tensor([l == lang for l in langs], dtype=torch.bool, device=device)
if not lang_mask.any():
continue
# Calculate language-specific positive ratio
lang_labels = labels[lang_mask]
lang_pos_ratio = lang_labels.float().mean(dim=0)
lang_pos_ratios[lang] = lang_pos_ratio
# Weighted contribution to batch statistics
lang_count = lang_mask.sum()
lang_counts[lang] = lang_count
batch_pos_ratios += lang_pos_ratio * (lang_count / batch_size)
# Combine batch and historical statistics
weights = torch.ones(batch_size, num_labels, device=device)
alpha = torch.zeros(num_labels, device=device)
gamma = torch.zeros(num_labels, device=device)
for i, (lang, label_vec) in enumerate(zip(langs, labels)):
if lang not in self.running_stats:
continue
# Get historical statistics for this language
hist_pos_ratio = (
self.running_stats[lang]['pos_counts'] /
(self.running_stats[lang]['total_counts'] + 1e-7)
).to(device)
# Combine historical and current batch statistics
current_pos_ratio = lang_pos_ratios.get(lang, batch_pos_ratios)
combined_pos_ratio = 0.7 * hist_pos_ratio + 0.3 * current_pos_ratio
# Calculate stable weights using log-space
log_ratio = torch.log1p(1.0 / (combined_pos_ratio + 1e-7))
class_weights = torch.exp(log_ratio.clamp(-2, 2))
# Apply language-specific scaling
weights[i] = class_weights * self.lang_scaling.get(lang, 1.0)
# Update focal parameters
alpha_contrib = 1.0 / (combined_pos_ratio + 1e-7).clamp(0.05, 0.95)
gamma_contrib = log_ratio.clamp(1.0, 4.0)
# Accumulate weighted contributions
weight = lang_counts.get(lang, 1) / batch_size
alpha += alpha_contrib * weight
gamma += gamma_contrib * weight
# Apply class-specific adjustments based on statistical analysis
# Order: toxic, severe_toxic, obscene, threat, insult, identity_hate
class_adjustments = {
'en': [1.0, 1.0, 0.9, 0.85, 1.1, 1.0], # English has more obscene/threat
'ru': [1.0, 1.0, 1.0, 1.0, 0.9, 1.0], # Russian has more insults
'tr': [1.0, 1.0, 1.0, 1.0, 0.9, 0.95], # Turkish pattern
'es': [1.0, 1.0, 1.0, 1.0, 0.9, 1.0], # Spanish pattern
'fr': [1.0, 1.0, 1.0, 1.0, 0.9, 1.0], # French pattern
'it': [1.0, 1.0, 1.0, 1.0, 0.9, 1.0], # Italian pattern
'pt': [1.0, 1.0, 1.0, 1.0, 0.9, 1.0] # Portuguese pattern
}
# Apply adjustments to weights
for i, lang in enumerate(langs):
if lang in class_adjustments:
# Multiply weights by language-specific class adjustments
weights[i] *= torch.tensor(class_adjustments[lang], device=device)
# Normalize weights to prevent extreme values
weights = weights / weights.mean()
return {
'weights': weights.clamp(0.1, 10.0), # Prevent extreme values
'alpha': alpha.clamp(0.1, 5.0), # [num_labels]
'gamma': gamma.clamp(1.0, 4.0) # [num_labels]
}
except Exception as e:
logger.error(f"Error computing batch weights: {str(e)}")
# Fallback to safe default values
return {
'weights': torch.ones((batch_size, num_labels), device=device),
'alpha': torch.full((num_labels,), 0.25, device=device),
'gamma': torch.full((num_labels,), 2.0, device=device)
}
@dataclass
class MetricsTracker:
"""Tracks training and validation metrics with error handling"""
best_auc: float = 0.0
train_losses: List[float] = None
val_losses: List[float] = None
val_aucs: List[float] = None
epoch_times: List[float] = None
def __post_init__(self):
self.train_losses = []
self.val_losses = []
self.val_aucs = []
self.epoch_times = []
def update_train(self, loss: float):
"""Update training metrics with validation"""
try:
if not isinstance(loss, (int, float)) or np.isnan(loss) or np.isinf(loss):
print(f"Warning: Invalid loss value: {loss}")
return
self.train_losses.append(float(loss))
except Exception as e:
print(f"Warning: Could not update training metrics: {str(e)}")
def update_validation(self, metrics: Dict) -> bool:
"""Update validation metrics with error handling"""
try:
if not isinstance(metrics, dict):
raise ValueError("Metrics must be a dictionary")
loss = metrics.get('loss', float('inf'))
auc = metrics.get('auc', 0.0)
if np.isnan(loss) or np.isinf(loss):
print(f"Warning: Invalid loss value: {loss}")
loss = float('inf')
if np.isnan(auc) or np.isinf(auc):
print(f"Warning: Invalid AUC value: {auc}")
auc = 0.0
self.val_losses.append(float(loss))
self.val_aucs.append(float(auc))
# Update best AUC if needed
if auc > self.best_auc:
self.best_auc = auc
return True
return False
except Exception as e:
print(f"Warning: Could not update validation metrics: {str(e)}")
return False
def update_time(self, epoch_time: float):
"""Update timing metrics with validation"""
try:
if not isinstance(epoch_time, (int, float)) or epoch_time <= 0:
print(f"Warning: Invalid epoch time: {epoch_time}")
return
self.epoch_times.append(float(epoch_time))
except Exception as e:
print(f"Warning: Could not update timing metrics: {str(e)}")
def get_eta(self, current_epoch: int, total_epochs: int) -> str:
"""Calculate ETA based on average epoch time with error handling"""
try:
if not self.epoch_times:
return "Calculating..."
if current_epoch >= total_epochs:
return "Complete"
avg_epoch_time = sum(self.epoch_times) / len(self.epoch_times)
remaining_epochs = total_epochs - current_epoch
eta_seconds = avg_epoch_time * remaining_epochs
hours = int(eta_seconds // 3600)
minutes = int((eta_seconds % 3600) // 60)
return f"{hours:02d}:{minutes:02d}:00"
except Exception as e:
print(f"Warning: Could not calculate ETA: {str(e)}")
return "Unknown"
@dataclass
class TrainingConfig:
"""Basic training configuration with consolidated default values"""
# Model parameters
model_name: str = "xlm-roberta-large"
max_length: int = 512
hidden_size: int = 1024
num_attention_heads: int = 16
model_dropout: float = 0.0
freeze_layers: int = 8
# Dataset parameters
cache_dir: str = 'cached_dataset'
label_columns: List[str] = None # Will be initialized in __post_init__
# Training parameters
batch_size: int = 128
grad_accum_steps: int = 1
epochs: int = 6
lr: float = 2e-5
num_cycles: int = 2
weight_decay: float = 2e-7
max_grad_norm: float = 1.0
warmup_ratio: float = 0.1
label_smoothing: float = 0.01
min_lr_ratio: float = 0.01
# Memory optimization
activation_checkpointing: bool = True
mixed_precision: str = "fp16"
_num_workers: int = None # Private storage for num_workers
gc_frequency: int = 500
tensor_float_32: bool = True
# Cosine scheduler parameters
num_cycles: int = 2
def __post_init__(self):
"""Initialize and validate configuration"""
# Initialize label columns
self.label_columns = [
'toxic', 'severe_toxic', 'obscene',
'threat', 'insult', 'identity_hate'
]
# Set environment variables for memory optimization
os.environ['PYTORCH_CUDA_ALLOC_CONF'] = 'max_split_size_mb:128,expandable_segments:True'
# Rest of the initialization code...
if self.lr <= 0:
raise ValueError(f"Learning rate must be positive, got {self.lr}")
if self.lr < 1e-7:
raise ValueError(f"Learning rate too small: {self.lr}")
if self.lr > 1.0:
raise ValueError(f"Learning rate too large: {self.lr}")
# Validate weight decay and learning rate combination
if self.weight_decay > 0:
wd_to_lr_ratio = self.weight_decay / self.lr
if wd_to_lr_ratio > 0.1:
logger.warning(
"Weight decay too high: %.2e (%.2fx learning rate). "
"Should be 0.01-0.1x learning rate.",
self.weight_decay, wd_to_lr_ratio
)
effective_lr = self.lr * (1 - self.weight_decay)
if effective_lr < self.lr * 0.9:
logger.warning(
"Weight decay %.2e reduces effective learning rate to %.2e (%.1f%% reduction)",
self.weight_decay, effective_lr, (1 - effective_lr/self.lr) * 100
)
# Set device with memory optimization
if torch.cuda.is_available():
try:
torch.cuda.init()
# Set memory allocation strategy
torch.cuda.set_per_process_memory_fraction(0.95) # Leave some GPU memory free
self.device = torch.device('cuda')
if self.mixed_precision == "bf16":
if not torch.cuda.is_bf16_supported():
print("Warning: BF16 not supported on this GPU. Falling back to FP16")
self.mixed_precision = "fp16"
if self.tensor_float_32:
if torch.cuda.get_device_capability()[0] >= 8:
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
else:
print("Warning: TF32 not supported on this GPU. Disabling.")
self.tensor_float_32 = False
except Exception as e:
print(f"Warning: CUDA initialization failed: {str(e)}")
self.device = torch.device('cpu')
self.mixed_precision = "no"
else:
self.device = torch.device('cpu')
if self.mixed_precision != "no":
print("Warning: Mixed precision not supported on CPU. Disabling.")
self.mixed_precision = "no"
# Create directories with error handling
try:
for directory in ["weights", "logs"]:
dir_path = Path(directory)
if not dir_path.exists():
dir_path.mkdir(parents=True)
elif not dir_path.is_dir():
raise NotADirectoryError(f"{directory} exists but is not a directory")
except Exception as e:
print(f"Error creating directories: {str(e)}")
raise
# Initialize toxicity labels
self.toxicity_labels = ['toxic', 'severe_toxic', 'obscene', 'threat', 'insult', 'identity_hate']
self.num_labels = len(self.toxicity_labels)
# Set use_mixed_precision flag
self.use_mixed_precision = self.mixed_precision != "no"
def validate_model_config(self, model):
"""Validate configuration against model architecture"""
try:
# Validate layer freezing
if self.freeze_layers > 0:
total_layers = len(list(model.base_model.encoder.layer))
if self.freeze_layers > total_layers:
raise ValueError(f"Can't freeze {self.freeze_layers} layers in {total_layers}-layer model")
logger.info(f"Freezing {self.freeze_layers} out of {total_layers} layers")
# Validate parameter groups and weight decay
param_groups = self.get_param_groups(model)
if self.weight_decay > 0:
low_lr_groups = [g for g in param_groups if g['lr'] < 0.01]
if low_lr_groups:
logger.warning("Found parameter groups with low learning rates (< 0.01) and non-zero weight decay:")
for group in low_lr_groups:
logger.warning(f"Group with lr={group['lr']:.4f}")
return True
except Exception as e:
logger.error(f"Model configuration validation failed: {str(e)}")
raise
@property
def dtype(self) -> torch.dtype:
"""Get the appropriate dtype based on mixed precision settings"""
if self.mixed_precision == "bf16":
return torch.bfloat16
elif self.mixed_precision == "fp16":
return torch.float16
return torch.float32
def get_autocast_context(self):
"""Get the appropriate autocast context based on configuration."""
if not self.use_mixed_precision:
return nullcontext()
dtype = torch.bfloat16 if self.mixed_precision == "bf16" else torch.float16
return torch.autocast(device_type=self.device.type, dtype=dtype)
def to_serializable_dict(self):
"""Convert config to a dictionary for saving."""
config_dict = asdict(self)
return config_dict
def get_param_groups(self, model):
"""Get parameter groups with base learning rate"""
return [{'params': model.parameters(), 'lr': self.lr}]
@property
def use_amp(self):
"""Check if AMP should be used based on device and mixed precision setting"""
return self.device.type == 'cuda' and self.mixed_precision != "no"
@property
def grad_norm_clip(self):
"""Adaptive gradient clipping based on precision"""
if self.mixed_precision == "bf16":
return 1.5 # BF16 can handle slightly higher gradients than FP16
if self.mixed_precision == "fp16":
return 1.0 # Most conservative for FP16 due to lower precision
return 5.0 # Full precision can handle larger gradients
@property
def num_workers(self):
"""Dynamically adjust workers based on system resources"""
if self._num_workers is None:
cpu_count = os.cpu_count()
if cpu_count is None:
self._num_workers = 0
else:
# Leave at least 2 CPUs free, max 4 workers
self._num_workers = min(4, max(0, cpu_count - 2))
logger.info(f"Dynamically set num_workers to {self._num_workers} (CPU count: {cpu_count})")
return self._num_workers
@num_workers.setter
def num_workers(self, value):
"""Allow manual override of num_workers"""
self._num_workers = value
logger.info(f"Manually set num_workers to {value}") |