renpas22 commited on
Commit ·
41bcc92
1
Parent(s): 85ab8c2
Fix train_prm and train_rl signatures to accept max_steps and learning_rate
Browse files- src/reasoning/step_level_cot.py +63 -37
src/reasoning/step_level_cot.py
CHANGED
|
@@ -349,23 +349,30 @@ class StepLevelCoTTrainer:
|
|
| 349 |
|
| 350 |
def train_prm(
|
| 351 |
self,
|
| 352 |
-
|
| 353 |
-
|
| 354 |
-
|
|
|
|
| 355 |
) -> None:
|
| 356 |
"""
|
| 357 |
Train Process Reward Model.
|
| 358 |
|
| 359 |
Args:
|
| 360 |
-
|
| 361 |
-
|
| 362 |
-
|
|
|
|
| 363 |
"""
|
| 364 |
logger.info("Starting PRM training")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 365 |
|
| 366 |
prm_trainer = PRMTrainer(
|
| 367 |
model=self.prm,
|
| 368 |
-
learning_rate=
|
| 369 |
weight_decay=self.config.weight_decay,
|
| 370 |
warmup_steps=self.config.warmup_steps,
|
| 371 |
)
|
|
@@ -383,19 +390,22 @@ class StepLevelCoTTrainer:
|
|
| 383 |
)
|
| 384 |
|
| 385 |
best_val_loss = float('inf')
|
|
|
|
| 386 |
|
| 387 |
-
|
| 388 |
-
|
| 389 |
-
|
| 390 |
-
|
| 391 |
-
|
| 392 |
-
|
| 393 |
-
|
| 394 |
-
|
| 395 |
-
|
| 396 |
-
|
| 397 |
-
|
| 398 |
-
|
|
|
|
|
|
|
| 399 |
# Extract batch data
|
| 400 |
vision_features = batch['visual_features']
|
| 401 |
step_embeddings = batch['step_descriptions'] # Need to encode these
|
|
@@ -410,37 +420,53 @@ class StepLevelCoTTrainer:
|
|
| 410 |
target_rewards,
|
| 411 |
)
|
| 412 |
|
| 413 |
-
|
| 414 |
-
progress_bar.
|
| 415 |
-
|
| 416 |
-
avg_loss = total_loss / len(train_dataloader)
|
| 417 |
-
logger.info(f"Epoch {epoch + 1} - Avg Loss: {avg_loss:.4f}")
|
| 418 |
-
|
| 419 |
-
# Validation
|
| 420 |
-
if val_dataset is not None:
|
| 421 |
-
val_metrics = self._evaluate_prm(prm_trainer, val_dataset)
|
| 422 |
-
logger.info(f"Validation: {val_metrics}")
|
| 423 |
|
| 424 |
-
# Save
|
| 425 |
-
if
|
| 426 |
-
|
| 427 |
-
self._save_prm(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 428 |
|
| 429 |
logger.info("PRM training completed")
|
| 430 |
|
| 431 |
def train_rl(
|
| 432 |
self,
|
| 433 |
-
|
| 434 |
-
|
|
|
|
|
|
|
| 435 |
) -> None:
|
| 436 |
"""
|
| 437 |
Train policy with reinforcement learning.
|
| 438 |
|
| 439 |
Args:
|
| 440 |
-
|
| 441 |
-
|
|
|
|
|
|
|
| 442 |
"""
|
| 443 |
logger.info("Starting RL training")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 444 |
|
| 445 |
self.rl_trainer.train(
|
| 446 |
train_dataset=train_dataset,
|
|
|
|
| 349 |
|
| 350 |
def train_prm(
|
| 351 |
self,
|
| 352 |
+
max_steps: int = 5000,
|
| 353 |
+
learning_rate: float = 1e-5,
|
| 354 |
+
save_steps: int = 500,
|
| 355 |
+
eval_steps: int = 500,
|
| 356 |
) -> None:
|
| 357 |
"""
|
| 358 |
Train Process Reward Model.
|
| 359 |
|
| 360 |
Args:
|
| 361 |
+
max_steps: Maximum training steps
|
| 362 |
+
learning_rate: Learning rate for PRM training
|
| 363 |
+
save_steps: Save checkpoint every N steps
|
| 364 |
+
eval_steps: Evaluate every N steps
|
| 365 |
"""
|
| 366 |
logger.info("Starting PRM training")
|
| 367 |
+
logger.info(f"Max steps: {max_steps}, LR: {learning_rate}")
|
| 368 |
+
|
| 369 |
+
# Load datasets using the trainer's dataset loading methods
|
| 370 |
+
train_dataset = self.load_step_dataset(split='train')
|
| 371 |
+
val_dataset = self.load_step_dataset(split='validation') if eval_steps > 0 else None
|
| 372 |
|
| 373 |
prm_trainer = PRMTrainer(
|
| 374 |
model=self.prm,
|
| 375 |
+
learning_rate=learning_rate, # Use the passed parameter
|
| 376 |
weight_decay=self.config.weight_decay,
|
| 377 |
warmup_steps=self.config.warmup_steps,
|
| 378 |
)
|
|
|
|
| 390 |
)
|
| 391 |
|
| 392 |
best_val_loss = float('inf')
|
| 393 |
+
global_step = 0
|
| 394 |
|
| 395 |
+
# Training loop using steps instead of epochs
|
| 396 |
+
prm_trainer.model.train()
|
| 397 |
+
|
| 398 |
+
progress_bar = tqdm(
|
| 399 |
+
total=max_steps,
|
| 400 |
+
desc=f"PRM Training",
|
| 401 |
+
disable=not self.accelerator.is_local_main_process,
|
| 402 |
+
)
|
| 403 |
+
|
| 404 |
+
while global_step < max_steps:
|
| 405 |
+
for batch in train_dataloader:
|
| 406 |
+
if global_step >= max_steps:
|
| 407 |
+
break
|
| 408 |
+
|
| 409 |
# Extract batch data
|
| 410 |
vision_features = batch['visual_features']
|
| 411 |
step_embeddings = batch['step_descriptions'] # Need to encode these
|
|
|
|
| 420 |
target_rewards,
|
| 421 |
)
|
| 422 |
|
| 423 |
+
global_step += 1
|
| 424 |
+
progress_bar.update(1)
|
| 425 |
+
progress_bar.set_postfix({'loss': metrics['loss'], 'step': global_step})
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 426 |
|
| 427 |
+
# Save checkpoint
|
| 428 |
+
if global_step % save_steps == 0:
|
| 429 |
+
logger.info(f"Saving checkpoint at step {global_step}")
|
| 430 |
+
self._save_prm(global_step)
|
| 431 |
+
|
| 432 |
+
# Validation
|
| 433 |
+
if eval_steps > 0 and global_step % eval_steps == 0 and val_dataset is not None:
|
| 434 |
+
val_metrics = self._evaluate_prm(prm_trainer, val_dataset)
|
| 435 |
+
logger.info(f"Step {global_step} - Validation: {val_metrics}")
|
| 436 |
+
|
| 437 |
+
# Save best model
|
| 438 |
+
if val_metrics['mse'] < best_val_loss:
|
| 439 |
+
best_val_loss = val_metrics['mse']
|
| 440 |
+
self._save_prm(f"best_step_{global_step}")
|
| 441 |
+
|
| 442 |
+
progress_bar.close()
|
| 443 |
|
| 444 |
logger.info("PRM training completed")
|
| 445 |
|
| 446 |
def train_rl(
|
| 447 |
self,
|
| 448 |
+
max_steps: int = 8000,
|
| 449 |
+
learning_rate: float = 5e-6,
|
| 450 |
+
save_steps: int = 500,
|
| 451 |
+
eval_steps: int = 500,
|
| 452 |
) -> None:
|
| 453 |
"""
|
| 454 |
Train policy with reinforcement learning.
|
| 455 |
|
| 456 |
Args:
|
| 457 |
+
max_steps: Maximum training steps
|
| 458 |
+
learning_rate: Learning rate for RL training
|
| 459 |
+
save_steps: Save checkpoint every N steps
|
| 460 |
+
eval_steps: Evaluate every N steps
|
| 461 |
"""
|
| 462 |
logger.info("Starting RL training")
|
| 463 |
+
logger.info(f"Max steps: {max_steps}, LR: {learning_rate}")
|
| 464 |
+
|
| 465 |
+
# Load dataset using the trainer's dataset loading methods
|
| 466 |
+
train_dataset = self.load_step_dataset(split='train')
|
| 467 |
+
|
| 468 |
+
# Convert max_steps to iterations (approximate)
|
| 469 |
+
num_iterations = max_steps // self.config.train_batch_size
|
| 470 |
|
| 471 |
self.rl_trainer.train(
|
| 472 |
train_dataset=train_dataset,
|