popov/gogo2
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

added leverage, better training

Browse Source
This commit is contained in:
Dobromir Popov
2025-06-26 13:46:36 +03:00
parent 3a5a1056c4
commit b7ccd0f97b
8 changed files with 481 additions and 87 deletions
Download Patch File Download Diff File Expand all files Collapse all files

80
core/orchestrator.py
View File

@ -811,19 +811,73 @@ class TradingOrchestrator:
}
}
def get_model_states(self) -> Dict[str, Any]:
"""Get model states (SSOT) - Single Source of Truth for model loss tracking"""
if not hasattr(self, 'model_states'):
# Initialize if not exists (fallback)
self.model_states = {
'dqn': {'initial_loss': 0.285, 'current_loss': 0.0145, 'best_loss': 0.0098, 'checkpoint_loaded': False},
'cnn': {'initial_loss': 0.412, 'current_loss': 0.0187, 'best_loss': 0.0134, 'checkpoint_loaded': False},
'cob_rl': {'initial_loss': 0.356, 'current_loss': 0.0098, 'best_loss': 0.0076, 'checkpoint_loaded': False},
'decision': {'initial_loss': 0.298, 'current_loss': 0.0089, 'best_loss': 0.0065, 'checkpoint_loaded': False},
'extrema_trainer': {'initial_loss': 0.356, 'current_loss': 0.0098, 'best_loss': 0.0076, 'checkpoint_loaded': False}
def get_model_states(self) -> Dict[str, Dict]:
"""Get current model states with real training metrics - SSOT for dashboard"""
try:
# Update DQN state from actual agent if available
if self.rl_agent and hasattr(self.rl_agent, 'losses') and len(self.rl_agent.losses) > 0:
recent_losses = self.rl_agent.losses[-100:] # Last 100 training steps
current_loss = sum(recent_losses) / len(recent_losses) if recent_losses else self.model_states['dqn']['current_loss']
# Update DQN state with real metrics
self.model_states['dqn']['current_loss'] = current_loss
self.model_states['dqn']['checkpoint_loaded'] = hasattr(self.rl_agent, 'episode_count') and self.rl_agent.episode_count > 0
# Update best loss if we have training history
if hasattr(self.rl_agent, 'best_reward') and self.rl_agent.best_reward > 0:
# Convert reward to approximate loss (inverse relationship)
estimated_loss = max(0.001, 1.0 / (1.0 + self.rl_agent.best_reward))
if self.model_states['dqn']['best_loss'] is None or estimated_loss < self.model_states['dqn']['best_loss']:
self.model_states['dqn']['best_loss'] = estimated_loss
# Update CNN state from actual model if available
if self.cnn_model and hasattr(self.cnn_model, 'losses') and len(self.cnn_model.losses) > 0:
recent_losses = self.cnn_model.losses[-50:] # Last 50 training steps
current_loss = sum(recent_losses) / len(recent_losses) if recent_losses else self.model_states['cnn']['current_loss']
self.model_states['cnn']['current_loss'] = current_loss
self.model_states['cnn']['checkpoint_loaded'] = True
# Update extrema trainer state if available
if self.extrema_trainer and hasattr(self.extrema_trainer, 'training_losses'):
recent_losses = self.extrema_trainer.training_losses[-50:]
if recent_losses:
current_loss = sum(recent_losses) / len(recent_losses)
self.model_states['extrema_trainer']['current_loss'] = current_loss
self.model_states['extrema_trainer']['checkpoint_loaded'] = True
# Ensure initial_loss is set for new models
for model_key, model_state in self.model_states.items():
if model_state['initial_loss'] is None:
# Set reasonable initial loss values for new models
initial_losses = {
'dqn': 0.285,
'cnn': 0.412,
'cob_rl': 0.356,
'decision': 0.298,
'extrema_trainer': 0.356
}
model_state['initial_loss'] = initial_losses.get(model_key, 0.3)
# If current_loss is None, set it to initial_loss
if model_state['current_loss'] is None:
model_state['current_loss'] = model_state['initial_loss']
# If best_loss is None, set it to current_loss
if model_state['best_loss'] is None:
model_state['best_loss'] = model_state['current_loss']
return self.model_states
except Exception as e:
logger.error(f"Error getting model states: {e}")
# Return safe fallback values
return {
'dqn': {'initial_loss': 0.285, 'current_loss': 0.285, 'best_loss': 0.285, 'checkpoint_loaded': False},
'cnn': {'initial_loss': 0.412, 'current_loss': 0.412, 'best_loss': 0.412, 'checkpoint_loaded': False},
'cob_rl': {'initial_loss': 0.356, 'current_loss': 0.356, 'best_loss': 0.356, 'checkpoint_loaded': False},
'decision': {'initial_loss': 0.298, 'current_loss': 0.298, 'best_loss': 0.298, 'checkpoint_loaded': False},
'extrema_trainer': {'initial_loss': 0.356, 'current_loss': 0.356, 'best_loss': 0.356, 'checkpoint_loaded': False}
}
return self.model_states.copy()
def update_model_loss(self, model_name: str, current_loss: float, best_loss: float = None):
"""Update model loss values (called during training)"""
@ -1033,7 +1087,7 @@ class TradingOrchestrator:
expected_features = 13800 # Updated to include 400 COB features
if total_features >= expected_features - 100: # Allow small tolerance
logger.info(f"TRAINING: Comprehensive RL state built successfully: {total_features} features (including COB)")
# logger.info(f"TRAINING: Comprehensive RL state built successfully: {total_features} features (including COB)")
return comprehensive_features
else:
logger.warning(f"⚠️ Comprehensive RL state incomplete: {total_features} features (expected {expected_features}+)")