Spaces:
Sleeping
Sleeping
File size: 18,859 Bytes
b2d89cf |
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 |
"""
Model trainer module for the Gmail Unsubscriber AI suite.
This module is responsible for:
- Loading prepared email data
- Splitting data into training and evaluation sets
- Loading pre-trained transformer model
- Fine-tuning the model on the email dataset
- Evaluating model performance
- Saving the fine-tuned model for prediction
The trained model is optimized for classifying emails as "important" or "unsubscribable".
"""
import os
import time
import pandas as pd
import numpy as np
import torch
from typing import Dict, List, Tuple, Optional, Any, Union, Callable
from sklearn.model_selection import train_test_split
from sklearn.metrics import accuracy_score, precision_recall_fscore_support, roc_auc_score, confusion_matrix
from sklearn.utils.class_weight import compute_class_weight
# Hugging Face imports
from transformers import (
AutoTokenizer,
AutoModelForSequenceClassification,
Trainer,
TrainingArguments,
EarlyStoppingCallback,
IntervalStrategy,
PreTrainedTokenizer,
PreTrainedModel
)
from datasets import Dataset
import torch.nn as nn
# Local imports
from . import config
from . import utils
from .task_utils import AiTaskLogger
def compute_metrics_for_classification(eval_pred) -> Dict[str, float]:
"""
Compute evaluation metrics for the model.
Args:
eval_pred: Tuple of predictions and labels from the trainer
Returns:
Dictionary of metrics including accuracy, precision, recall, F1, etc.
"""
predictions, labels = eval_pred
# For classification, take the argmax to get predicted classes
preds = np.argmax(predictions, axis=1)
# Calculate overall metrics
precision, recall, f1, _ = precision_recall_fscore_support(
labels, preds, average='weighted'
)
acc = accuracy_score(labels, preds)
# Calculate metrics specific to the "unsubscribable" class
# Get label positions: 0 for important, 1 for unsubscribable
unsub_class_idx = config.LABEL_UNSUBSCRIBABLE_ID
# Calculate class-specific precision, recall, F1
precision_unsub, recall_unsub, f1_unsub, _ = precision_recall_fscore_support(
labels, preds, average=None, labels=[unsub_class_idx]
)
# Compute ROC AUC if possible (requires probability scores)
try:
# Use the probability score for the unsubscribable class
probs_unsub = predictions[:, unsub_class_idx]
# Convert labels to binary (1 for unsubscribable, 0 for others)
binary_labels = (labels == unsub_class_idx).astype(int)
# Calculate ROC AUC
roc_auc = roc_auc_score(binary_labels, probs_unsub)
except (ValueError, IndexError):
roc_auc = 0.0
# Return all metrics in a dictionary
metrics = {
'accuracy': acc,
'precision': precision,
'recall': recall,
'f1': f1,
'precision_unsub': float(precision_unsub[0]) if len(precision_unsub) > 0 else 0.0,
'recall_unsub': float(recall_unsub[0]) if len(recall_unsub) > 0 else 0.0,
'f1_unsub': float(f1_unsub[0]) if len(f1_unsub) > 0 else 0.0,
'roc_auc': roc_auc
}
return metrics
class WeightedLossTrainer(Trainer):
"""Custom trainer that supports class weights for imbalanced datasets."""
def __init__(self, class_weights=None, label_smoothing=0.0, *args, **kwargs):
super().__init__(*args, **kwargs)
self.class_weights = class_weights
self.label_smoothing = label_smoothing
def compute_loss(self, model, inputs, return_outputs=False):
"""Compute loss with class weights and optional label smoothing."""
labels = inputs.pop("labels")
outputs = model(**inputs)
logits = outputs.get('logits')
# Create loss function with class weights
if self.class_weights is not None:
weight = torch.tensor(self.class_weights, dtype=torch.float32).to(logits.device)
loss_fct = nn.CrossEntropyLoss(weight=weight, label_smoothing=self.label_smoothing)
else:
loss_fct = nn.CrossEntropyLoss(label_smoothing=self.label_smoothing)
loss = loss_fct(logits.view(-1, self.model.config.num_labels), labels.view(-1))
return (loss, outputs) if return_outputs else loss
def tokenize_dataset(
examples: Dict[str, List],
tokenizer: PreTrainedTokenizer,
max_length: int
) -> Dict[str, List]:
"""
Tokenize a batch of examples.
Args:
examples: Dictionary of example lists
tokenizer: Hugging Face tokenizer
max_length: Maximum sequence length
Returns:
Dictionary of tokenized examples
"""
return tokenizer(
examples['text'],
padding='max_length',
truncation=True,
max_length=max_length,
return_tensors='pt'
)
def prepare_datasets_for_training(
data_file: str,
test_size: float,
tokenizer: PreTrainedTokenizer,
max_length: int,
task_logger: AiTaskLogger,
compute_weights: bool = True
) -> Tuple[Dataset, Dataset, Optional[np.ndarray]]:
"""
Load and prepare datasets for training.
Args:
data_file: Path to the prepared data file (CSV)
test_size: Proportion of data to use for evaluation
tokenizer: Hugging Face tokenizer
max_length: Maximum sequence length
task_logger: Logger for tracking progress
compute_weights: Whether to compute class weights for imbalanced data
Returns:
Tuple of (train_dataset, eval_dataset, class_weights)
"""
task_logger.info(f"Loading data from {data_file}")
try:
# Load the dataset
df = pd.read_csv(data_file)
task_logger.info(f"Loaded {len(df)} examples from {data_file}")
# Check for required columns
if 'text' not in df.columns or 'label' not in df.columns:
task_logger.error("Data file is missing required columns 'text' and/or 'label'")
raise ValueError("Data file has wrong format")
# Split into training and evaluation sets
train_df, eval_df = train_test_split(
df, test_size=test_size, stratify=df['label'], random_state=42
)
task_logger.info(f"Split into {len(train_df)} training examples and {len(eval_df)} evaluation examples")
# Compute class weights for imbalanced data
class_weights = None
if compute_weights:
class_weights = compute_class_weight(
'balanced',
classes=np.array([0, 1]),
y=train_df['label'].values
)
task_logger.info(f"Computed class weights: {class_weights}")
# Create HF datasets
train_dataset = Dataset.from_pandas(train_df)
eval_dataset = Dataset.from_pandas(eval_df)
# Create a tokenization function that uses our tokenizer
def tokenize_function(examples):
return tokenize_dataset(examples, tokenizer, max_length)
# Apply tokenization
task_logger.info("Tokenizing datasets")
train_dataset = train_dataset.map(
tokenize_function,
batched=True,
desc="Tokenizing training dataset"
)
eval_dataset = eval_dataset.map(
tokenize_function,
batched=True,
desc="Tokenizing evaluation dataset"
)
# Set format for PyTorch
train_dataset.set_format(
type='torch',
columns=['input_ids', 'attention_mask', 'label']
)
eval_dataset.set_format(
type='torch',
columns=['input_ids', 'attention_mask', 'label']
)
task_logger.info("Datasets prepared successfully")
return train_dataset, eval_dataset, class_weights
except Exception as e:
task_logger.error(f"Error preparing datasets: {str(e)}", e)
raise
def train_unsubscriber_model(task_logger: AiTaskLogger) -> Dict[str, Any]:
"""
Train the unsubscriber model on the prepared dataset.
This function:
1. Loads prepared data
2. Initializes a pre-trained transformer model
3. Fine-tunes it on the email classification task
4. Evaluates performance
5. Saves the model for later use
Args:
task_logger: Logger for tracking task status
Returns:
Dictionary with training results and metrics
"""
# Start timing
start_time = time.time()
# Start the task
task_logger.start_task("Starting AI model training")
try:
# Check if prepared data exists
if not os.path.exists(config.PREPARED_DATA_FILE):
task_logger.error(f"Prepared data file not found at {config.PREPARED_DATA_FILE}.")
task_logger.fail_task("Training failed: No prepared data available. Please run data preparation first.")
return {"success": False, "error": "No prepared data available"}
# 1. Load and initialize the tokenizer
task_logger.info(f"Loading tokenizer for model: {config.PRE_TRAINED_MODEL_NAME}")
tokenizer = AutoTokenizer.from_pretrained(config.PRE_TRAINED_MODEL_NAME)
# 2. Prepare datasets
task_logger.update_progress(0.1, "Preparing datasets")
train_dataset, eval_dataset, class_weights = prepare_datasets_for_training(
config.PREPARED_DATA_FILE,
config.TEST_SPLIT_SIZE,
tokenizer,
config.MAX_SEQ_LENGTH,
task_logger,
compute_weights=True
)
# 3. Initialize model
task_logger.update_progress(0.2, f"Initializing model: {config.PRE_TRAINED_MODEL_NAME}")
model = AutoModelForSequenceClassification.from_pretrained(
config.PRE_TRAINED_MODEL_NAME,
num_labels=config.NUM_LABELS
)
# Check device availability
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
fp16_enabled = config.FP16_TRAINING and torch.cuda.is_available()
# Force GPU check and provide detailed info
if torch.cuda.is_available():
gpu_name = torch.cuda.get_device_name(0)
gpu_memory = torch.cuda.get_device_properties(0).total_memory / (1024 ** 3) # Convert to GB
task_logger.info(f"GPU detected: {gpu_name} with {gpu_memory:.2f} GB memory")
task_logger.info(f"CUDA version: {torch.version.cuda}")
# Set CUDA device explicitly
torch.cuda.set_device(0)
# Enable cuDNN benchmarking for better performance
torch.backends.cudnn.benchmark = True
torch.backends.cudnn.enabled = True
else:
task_logger.warning("No GPU detected! Training will be slow on CPU.")
task_logger.warning("Make sure you have the CUDA version of PyTorch installed.")
task_logger.warning("To install PyTorch with CUDA support, run:")
task_logger.warning("pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118")
task_logger.info(f"Training on device: {device}")
if fp16_enabled:
task_logger.info("FP16 mixed precision training enabled")
# Move model to device
model.to(device)
# 4. Set up training arguments
task_logger.update_progress(0.3, "Setting up training configuration")
# Calculate number of steps
num_train_examples = len(train_dataset)
num_train_steps = (num_train_examples // config.TRAIN_BATCH_SIZE) * config.NUM_TRAIN_EPOCHS
warmup_steps = int(num_train_steps * config.WARMUP_STEPS_RATIO)
# Convert evaluation and save strategies to enum values
eval_strategy = (IntervalStrategy.EPOCH
if config.EVALUATION_STRATEGY == "epoch"
else IntervalStrategy.STEPS)
save_strategy = (IntervalStrategy.EPOCH
if config.SAVE_STRATEGY == "epoch"
else IntervalStrategy.STEPS)
# Create output directory if it doesn't exist
os.makedirs(config.FINE_TUNED_MODEL_DIR, exist_ok=True)
# Define training arguments (with compatibility fixes for older transformers versions)
training_args_dict = {
"output_dir": config.FINE_TUNED_MODEL_DIR,
"per_device_train_batch_size": config.TRAIN_BATCH_SIZE,
"per_device_eval_batch_size": config.EVAL_BATCH_SIZE,
"num_train_epochs": config.NUM_TRAIN_EPOCHS,
"learning_rate": config.LEARNING_RATE,
"weight_decay": config.WEIGHT_DECAY,
"warmup_steps": warmup_steps,
"logging_steps": 50,
"save_total_limit": 2,
"load_best_model_at_end": config.LOAD_BEST_MODEL_AT_END,
"metric_for_best_model": config.METRIC_FOR_BEST_MODEL,
"greater_is_better": True,
"report_to": "none",
"disable_tqdm": False,
"no_cuda": False, # Explicitly enable CUDA
"use_cpu": False, # Explicitly disable CPU-only mode
}
# Add evaluation and save strategies
try:
training_args_dict["evaluation_strategy"] = eval_strategy
training_args_dict["save_strategy"] = save_strategy
except:
# Fallback for older versions
training_args_dict["evaluation_strategy"] = "epoch"
training_args_dict["save_strategy"] = "epoch"
# Add logging dir if supported
try:
training_args_dict["logging_dir"] = os.path.join(config.FINE_TUNED_MODEL_DIR, "logs")
except:
pass
# Add FP16 if supported and available
if fp16_enabled:
try:
training_args_dict["fp16"] = True
except:
task_logger.warning("FP16 training not supported on this setup")
# Add dataloader workers if supported
try:
training_args_dict["dataloader_num_workers"] = 2
except:
pass
# Add gradient accumulation if configured
if hasattr(config, 'GRADIENT_ACCUMULATION_STEPS'):
training_args_dict["gradient_accumulation_steps"] = config.GRADIENT_ACCUMULATION_STEPS
# Add label smoothing if configured
if hasattr(config, 'LABEL_SMOOTHING_FACTOR'):
training_args_dict["label_smoothing_factor"] = config.LABEL_SMOOTHING_FACTOR
training_args = TrainingArguments(**training_args_dict)
# 5. Initialize the trainer with class weights
task_logger.update_progress(0.4, "Initializing trainer with class balancing")
# Use weighted loss trainer if we have class weights
if class_weights is not None:
trainer = WeightedLossTrainer(
model=model,
args=training_args,
train_dataset=train_dataset,
eval_dataset=eval_dataset,
compute_metrics=compute_metrics_for_classification,
class_weights=class_weights,
label_smoothing=config.LABEL_SMOOTHING_FACTOR if hasattr(config, 'LABEL_SMOOTHING_FACTOR') else 0.0,
callbacks=[
EarlyStoppingCallback(
early_stopping_patience=config.EARLY_STOPPING_PATIENCE,
early_stopping_threshold=config.EARLY_STOPPING_THRESHOLD
)
]
)
task_logger.info("Using weighted loss function for class imbalance")
else:
trainer = Trainer(
model=model,
args=training_args,
train_dataset=train_dataset,
eval_dataset=eval_dataset,
compute_metrics=compute_metrics_for_classification,
callbacks=[
EarlyStoppingCallback(
early_stopping_patience=config.EARLY_STOPPING_PATIENCE,
early_stopping_threshold=config.EARLY_STOPPING_THRESHOLD
)
]
)
# 6. Train the model
task_logger.update_progress(0.5, "Starting model training")
trainer.train()
# 7. Evaluate the model
task_logger.update_progress(0.9, "Evaluating model")
eval_results = trainer.evaluate()
# Format evaluations nicely for logging
metrics_str = "\n".join([f" {k}: {v:.4f}" for k, v in eval_results.items()])
task_logger.info(f"Evaluation results:\n{metrics_str}")
# 8. Save the final model
task_logger.update_progress(0.95, "Saving fine-tuned model")
trainer.save_model(config.FINE_TUNED_MODEL_DIR)
tokenizer.save_pretrained(config.FINE_TUNED_MODEL_DIR)
# Save a human-readable summary of model info
with open(os.path.join(config.FINE_TUNED_MODEL_DIR, "model_info.txt"), "w") as f:
f.write(f"Base model: {config.PRE_TRAINED_MODEL_NAME}\n")
f.write(f"Training completed: {utils.get_current_timestamp()}\n")
f.write(f"Training examples: {len(train_dataset)}\n")
f.write(f"Evaluation examples: {len(eval_dataset)}\n")
f.write(f"Evaluation metrics:\n")
for k, v in eval_results.items():
f.write(f" {k}: {v:.4f}\n")
# Calculate elapsed time
elapsed_time = time.time() - start_time
hours, remainder = divmod(elapsed_time, 3600)
minutes, seconds = divmod(remainder, 60)
time_str = f"{int(hours)}h {int(minutes)}m {int(seconds)}s"
# Complete the task
result = {
"success": True,
"metrics": eval_results,
"model_dir": config.FINE_TUNED_MODEL_DIR,
"training_time": time_str,
"base_model": config.PRE_TRAINED_MODEL_NAME,
"num_train_examples": len(train_dataset),
"num_eval_examples": len(eval_dataset)
}
task_logger.complete_task(
f"Model training completed successfully in {time_str}",
result
)
return result
except Exception as e:
task_logger.error("Error during model training", e)
task_logger.fail_task(f"Model training failed: {str(e)}")
return {
"success": False,
"error": str(e)
} |