i think we fixed mexc interface at the end!!!
This commit is contained in:
Dobromir Popov
@ -133,19 +133,37 @@ class CheckpointManager:
|
||||
|
||||
def load_best_checkpoint(self, model_name: str) -> Optional[Tuple[str, CheckpointMetadata]]:
|
||||
try:
|
||||
if model_name not in self.checkpoints or not self.checkpoints[model_name]:
|
||||
logger.warning(f"No checkpoints found for model: {model_name}")
|
||||
return None
|
||||
# First, try the standard checkpoint system
|
||||
if model_name in self.checkpoints and self.checkpoints[model_name]:
|
||||
# Filter out checkpoints with non-existent files
|
||||
valid_checkpoints = [
|
||||
cp for cp in self.checkpoints[model_name]
|
||||
if Path(cp.file_path).exists()
|
||||
]
|
||||
|
||||
if valid_checkpoints:
|
||||
best_checkpoint = max(valid_checkpoints, key=lambda x: x.performance_score)
|
||||
logger.debug(f"Loading best checkpoint for {model_name}: {best_checkpoint.checkpoint_id}")
|
||||
return best_checkpoint.file_path, best_checkpoint
|
||||
else:
|
||||
# Clean up invalid metadata entries
|
||||
invalid_count = len(self.checkpoints[model_name])
|
||||
logger.warning(f"Found {invalid_count} invalid checkpoint entries for {model_name}, cleaning up metadata")
|
||||
self.checkpoints[model_name] = []
|
||||
self._save_metadata()
|
||||
|
||||
best_checkpoint = max(self.checkpoints[model_name], key=lambda x: x.performance_score)
|
||||
# Fallback: Look for existing saved models in the legacy format
|
||||
logger.debug(f"No valid checkpoints found for model: {model_name}, attempting to find legacy saved models")
|
||||
legacy_model_path = self._find_legacy_model(model_name)
|
||||
|
||||
if not Path(best_checkpoint.file_path).exists():
|
||||
# temporary disable logging to avoid spam
|
||||
# logger.error(f"Best checkpoint file not found: {best_checkpoint.file_path}")
|
||||
return None
|
||||
if legacy_model_path:
|
||||
# Create checkpoint metadata for the legacy model using actual file data
|
||||
legacy_metadata = self._create_legacy_metadata(model_name, legacy_model_path)
|
||||
logger.debug(f"Found legacy model for {model_name}: {legacy_model_path}")
|
||||
return str(legacy_model_path), legacy_metadata
|
||||
|
||||
logger.debug(f"Loading best checkpoint for {model_name}: {best_checkpoint.checkpoint_id}")
|
||||
return best_checkpoint.file_path, best_checkpoint
|
||||
logger.warning(f"No checkpoints or legacy models found for: {model_name}")
|
||||
return None
|
||||
|
||||
except Exception as e:
|
||||
logger.error(f"Error loading best checkpoint for {model_name}: {e}")
|
||||
@ -181,16 +199,8 @@ class CheckpointManager:
|
||||
# Bonus for processing more training samples
|
||||
score += min(10, metrics['training_samples'] / 10)
|
||||
|
||||
# Ensure minimum score for any training activity
|
||||
if score == 0.0 and metrics:
|
||||
# Use the first available metric with better scaling
|
||||
first_metric = next(iter(metrics.values()))
|
||||
if first_metric > 0:
|
||||
score = max(0.1, min(10, first_metric))
|
||||
else:
|
||||
score = 0.1
|
||||
|
||||
return max(score, 0.1)
|
||||
# Return actual calculated score - NO SYNTHETIC MINIMUM
|
||||
return score
|
||||
|
||||
def _should_save_checkpoint(self, model_name: str, performance_score: float) -> bool:
|
||||
"""Improved checkpoint saving logic with more frequent saves during training"""
|
||||
@ -331,6 +341,110 @@ class CheckpointManager:
|
||||
stats['total_size_mb'] += model_size
|
||||
|
||||
return stats
|
||||
|
||||
def _find_legacy_model(self, model_name: str) -> Optional[Path]:
|
||||
"""Find legacy saved models based on model name patterns"""
|
||||
base_dir = Path(self.base_dir)
|
||||
|
||||
# Define model name mappings and patterns for legacy files
|
||||
legacy_patterns = {
|
||||
'dqn_agent': [
|
||||
'dqn_agent_best_policy.pt',
|
||||
'enhanced_dqn_best_policy.pt',
|
||||
'improved_dqn_agent_best_policy.pt',
|
||||
'dqn_agent_final_policy.pt'
|
||||
],
|
||||
'enhanced_cnn': [
|
||||
'cnn_model_best.pt',
|
||||
'optimized_short_term_model_best.pt',
|
||||
'optimized_short_term_model_realtime_best.pt',
|
||||
'optimized_short_term_model_ticks_best.pt'
|
||||
],
|
||||
'extrema_trainer': [
|
||||
'supervised_model_best.pt'
|
||||
],
|
||||
'cob_rl': [
|
||||
'best_rl_model.pth_policy.pt',
|
||||
'rl_agent_best_policy.pt'
|
||||
],
|
||||
'decision': [
|
||||
# Decision models might be in subdirectories, but let's check main dir too
|
||||
'decision_best.pt',
|
||||
'decision_model_best.pt',
|
||||
# Check for transformer models which might be used as decision models
|
||||
'enhanced_dqn_best_policy.pt',
|
||||
'improved_dqn_agent_best_policy.pt'
|
||||
]
|
||||
}
|
||||
|
||||
# Get patterns for this model name
|
||||
patterns = legacy_patterns.get(model_name, [])
|
||||
|
||||
# Also try generic patterns based on model name
|
||||
patterns.extend([
|
||||
f'{model_name}_best.pt',
|
||||
f'{model_name}_best_policy.pt',
|
||||
f'{model_name}_final.pt',
|
||||
f'{model_name}_final_policy.pt'
|
||||
])
|
||||
|
||||
# Search for the model files
|
||||
for pattern in patterns:
|
||||
candidate_path = base_dir / pattern
|
||||
if candidate_path.exists():
|
||||
logger.debug(f"Found legacy model file: {candidate_path}")
|
||||
return candidate_path
|
||||
|
||||
# Also check subdirectories
|
||||
for subdir in base_dir.iterdir():
|
||||
if subdir.is_dir() and subdir.name == model_name:
|
||||
for pattern in patterns:
|
||||
candidate_path = subdir / pattern
|
||||
if candidate_path.exists():
|
||||
logger.debug(f"Found legacy model file in subdirectory: {candidate_path}")
|
||||
return candidate_path
|
||||
|
||||
return None
|
||||
|
||||
def _create_legacy_metadata(self, model_name: str, file_path: Path) -> CheckpointMetadata:
|
||||
"""Create metadata for legacy model files using only actual file information"""
|
||||
try:
|
||||
file_size_mb = file_path.stat().st_size / (1024 * 1024)
|
||||
created_time = datetime.fromtimestamp(file_path.stat().st_mtime)
|
||||
|
||||
# NO SYNTHETIC DATA - use only actual file information
|
||||
return CheckpointMetadata(
|
||||
checkpoint_id=f"legacy_{model_name}_{int(created_time.timestamp())}",
|
||||
model_name=model_name,
|
||||
model_type=model_name,
|
||||
file_path=str(file_path),
|
||||
created_at=created_time,
|
||||
file_size_mb=file_size_mb,
|
||||
performance_score=0.0, # Unknown performance - use 0, not synthetic values
|
||||
accuracy=None,
|
||||
loss=None,
|
||||
val_accuracy=None,
|
||||
val_loss=None,
|
||||
reward=None,
|
||||
pnl=None,
|
||||
epoch=None,
|
||||
training_time_hours=None,
|
||||
total_parameters=None,
|
||||
wandb_run_id=None,
|
||||
wandb_artifact_name=None
|
||||
)
|
||||
except Exception as e:
|
||||
logger.error(f"Error creating legacy metadata for {model_name}: {e}")
|
||||
# Return a basic metadata with minimal info - NO SYNTHETIC VALUES
|
||||
return CheckpointMetadata(
|
||||
checkpoint_id=f"legacy_{model_name}",
|
||||
model_name=model_name,
|
||||
model_type=model_name,
|
||||
file_path=str(file_path),
|
||||
created_at=datetime.now(),
|
||||
file_size_mb=0.0,
|
||||
performance_score=0.0 # Unknown - use 0, not synthetic
|
||||
)
|
||||
|
||||
_checkpoint_manager = None
|
||||
|
||||
|
||||
Reference in New Issue
Block a user