Files
gogo2/core/cnn_training_pipeline.py
Dobromir Popov 12865fd3ef replay system
2025-07-20 12:37:02 +03:00

785 lines
30 KiB
Python

"""
CNN Training Pipeline with Comprehensive Data Storage and Replay
This module implements a robust CNN training pipeline that:
1. Integrates with the comprehensive training data collection system
2. Stores all backpropagation data for gradient replay
3. Enables retraining on most profitable setups
4. Maintains training episode profitability tracking
5. Supports both real-time and batch training modes
Key Features:
- Integration with TrainingDataCollector for data validation
- Gradient and loss storage for each training step
- Profitable episode prioritization and replay
- Comprehensive training metrics and validation
- Real-time pivot point prediction with outcome tracking
"""
import asyncio
import logging
import numpy as np
import pandas as pd
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import Dataset, DataLoader
from datetime import datetime, timedelta
from pathlib import Path
from typing import Dict, List, Optional, Tuple, Any, Callable
from dataclasses import dataclass, field
import json
import pickle
from collections import deque, defaultdict
import threading
from concurrent.futures import ThreadPoolExecutor
from .training_data_collector import (
TrainingDataCollector,
TrainingEpisode,
ModelInputPackage,
get_training_data_collector
)
logger = logging.getLogger(__name__)
@dataclass
class CNNTrainingStep:
"""Single CNN training step with complete backpropagation data"""
step_id: str
timestamp: datetime
episode_id: str
# Input data
input_features: torch.Tensor
target_labels: torch.Tensor
# Forward pass results
model_outputs: Dict[str, torch.Tensor]
predictions: Dict[str, Any]
confidence_scores: torch.Tensor
# Loss components
total_loss: float
pivot_prediction_loss: float
confidence_loss: float
regularization_loss: float
# Backpropagation data
gradients: Dict[str, torch.Tensor] # Gradients for each parameter
gradient_norms: Dict[str, float] # Gradient norms for monitoring
# Model state
model_state_dict: Optional[Dict[str, torch.Tensor]] = None
optimizer_state: Optional[Dict[str, Any]] = None
# Training metadata
learning_rate: float = 0.001
batch_size: int = 32
epoch: int = 0
# Profitability tracking
actual_profitability: Optional[float] = None
prediction_accuracy: Optional[float] = None
training_value: float = 0.0 # Value of this training step for replay
@dataclass
class CNNTrainingSession:
"""Complete CNN training session with multiple steps"""
session_id: str
start_timestamp: datetime
end_timestamp: Optional[datetime] = None
# Session configuration
training_mode: str = 'real_time' # 'real_time', 'batch', 'replay'
symbol: str = ''
# Training steps
training_steps: List[CNNTrainingStep] = field(default_factory=list)
# Session metrics
total_steps: int = 0
average_loss: float = 0.0
best_loss: float = float('inf')
convergence_achieved: bool = False
# Profitability metrics
profitable_predictions: int = 0
total_predictions: int = 0
profitability_rate: float = 0.0
# Session value for replay prioritization
session_value: float = 0.0
class CNNPivotPredictor(nn.Module):
"""CNN model for pivot point prediction with comprehensive output"""
def __init__(self,
input_channels: int = 10, # Multiple timeframes
sequence_length: int = 300, # 300 bars
hidden_dim: int = 256,
num_pivot_classes: int = 3, # high, low, none
dropout_rate: float = 0.2):
super(CNNPivotPredictor, self).__init__()
self.input_channels = input_channels
self.sequence_length = sequence_length
self.hidden_dim = hidden_dim
# Convolutional layers for pattern extraction
self.conv_layers = nn.Sequential(
# First conv block
nn.Conv1d(input_channels, 64, kernel_size=7, padding=3),
nn.BatchNorm1d(64),
nn.ReLU(),
nn.Dropout(dropout_rate),
# Second conv block
nn.Conv1d(64, 128, kernel_size=5, padding=2),
nn.BatchNorm1d(128),
nn.ReLU(),
nn.Dropout(dropout_rate),
# Third conv block
nn.Conv1d(128, 256, kernel_size=3, padding=1),
nn.BatchNorm1d(256),
nn.ReLU(),
nn.Dropout(dropout_rate),
)
# LSTM for temporal dependencies
self.lstm = nn.LSTM(
input_size=256,
hidden_size=hidden_dim,
num_layers=2,
batch_first=True,
dropout=dropout_rate,
bidirectional=True
)
# Attention mechanism
self.attention = nn.MultiheadAttention(
embed_dim=hidden_dim * 2, # Bidirectional LSTM
num_heads=8,
dropout=dropout_rate,
batch_first=True
)
# Output heads
self.pivot_classifier = nn.Sequential(
nn.Linear(hidden_dim * 2, hidden_dim),
nn.ReLU(),
nn.Dropout(dropout_rate),
nn.Linear(hidden_dim, num_pivot_classes)
)
self.pivot_price_regressor = nn.Sequential(
nn.Linear(hidden_dim * 2, hidden_dim),
nn.ReLU(),
nn.Dropout(dropout_rate),
nn.Linear(hidden_dim, 1)
)
self.confidence_head = nn.Sequential(
nn.Linear(hidden_dim * 2, hidden_dim // 2),
nn.ReLU(),
nn.Linear(hidden_dim // 2, 1),
nn.Sigmoid()
)
# Initialize weights
self.apply(self._init_weights)
def _init_weights(self, module):
"""Initialize weights with proper scaling"""
if isinstance(module, nn.Linear):
torch.nn.init.xavier_uniform_(module.weight)
if module.bias is not None:
torch.nn.init.zeros_(module.bias)
elif isinstance(module, nn.Conv1d):
torch.nn.init.kaiming_normal_(module.weight, mode='fan_out', nonlinearity='relu')
def forward(self, x):
"""
Forward pass through CNN pivot predictor
Args:
x: Input tensor [batch_size, input_channels, sequence_length]
Returns:
Dict containing predictions and hidden states
"""
batch_size = x.size(0)
# Convolutional feature extraction
conv_features = self.conv_layers(x) # [batch, 256, sequence_length]
# Prepare for LSTM (transpose to [batch, sequence, features])
lstm_input = conv_features.transpose(1, 2) # [batch, sequence_length, 256]
# LSTM processing
lstm_output, (hidden, cell) = self.lstm(lstm_input) # [batch, sequence_length, hidden_dim*2]
# Attention mechanism
attended_output, attention_weights = self.attention(
lstm_output, lstm_output, lstm_output
)
# Use the last timestep for predictions
final_features = attended_output[:, -1, :] # [batch, hidden_dim*2]
# Generate predictions
pivot_logits = self.pivot_classifier(final_features)
pivot_price = self.pivot_price_regressor(final_features)
confidence = self.confidence_head(final_features)
return {
'pivot_logits': pivot_logits,
'pivot_price': pivot_price,
'confidence': confidence,
'hidden_states': final_features,
'attention_weights': attention_weights,
'conv_features': conv_features,
'lstm_output': lstm_output
}
class CNNTrainingDataset(Dataset):
"""Dataset for CNN training with training episodes"""
def __init__(self, training_episodes: List[TrainingEpisode]):
self.episodes = training_episodes
self.valid_episodes = self._validate_episodes()
def _validate_episodes(self) -> List[TrainingEpisode]:
"""Validate and filter episodes for training"""
valid = []
for episode in self.episodes:
try:
# Check if episode has required data
if (episode.input_package.cnn_features is not None and
episode.actual_outcome.outcome_validated):
valid.append(episode)
except Exception as e:
logger.warning(f"Invalid episode {episode.episode_id}: {e}")
logger.info(f"Validated {len(valid)}/{len(self.episodes)} episodes for training")
return valid
def __len__(self):
return len(self.valid_episodes)
def __getitem__(self, idx):
episode = self.valid_episodes[idx]
# Extract features
features = torch.from_numpy(episode.input_package.cnn_features).float()
# Create labels from actual outcomes
pivot_class = self._determine_pivot_class(episode.actual_outcome)
pivot_price = episode.actual_outcome.optimal_exit_price
confidence_target = episode.actual_outcome.profitability_score
return {
'features': features,
'pivot_class': torch.tensor(pivot_class, dtype=torch.long),
'pivot_price': torch.tensor(pivot_price, dtype=torch.float),
'confidence_target': torch.tensor(confidence_target, dtype=torch.float),
'episode_id': episode.episode_id,
'profitability': episode.actual_outcome.profitability_score
}
def _determine_pivot_class(self, outcome) -> int:
"""Determine pivot class from outcome"""
if outcome.price_change_15m > 0.5: # Significant upward movement
return 0 # High pivot
elif outcome.price_change_15m < -0.5: # Significant downward movement
return 1 # Low pivot
else:
return 2 # No significant pivot
class CNNTrainer:
"""CNN trainer with comprehensive data storage and replay capabilities"""
def __init__(self,
model: CNNPivotPredictor,
device: str = 'cuda',
learning_rate: float = 0.001,
storage_dir: str = "cnn_training_storage"):
self.model = model.to(device)
self.device = device
self.learning_rate = learning_rate
# Storage
self.storage_dir = Path(storage_dir)
self.storage_dir.mkdir(parents=True, exist_ok=True)
# Optimizer
self.optimizer = torch.optim.AdamW(
self.model.parameters(),
lr=learning_rate,
weight_decay=1e-5
)
# Learning rate scheduler
self.scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
self.optimizer, mode='min', patience=10, factor=0.5
)
# Training data collector
self.data_collector = get_training_data_collector()
# Training sessions storage
self.training_sessions: List[CNNTrainingSession] = []
self.current_session: Optional[CNNTrainingSession] = None
# Training statistics
self.training_stats = {
'total_sessions': 0,
'total_steps': 0,
'best_validation_loss': float('inf'),
'profitable_predictions': 0,
'total_predictions': 0,
'replay_sessions': 0
}
# Background training
self.is_training = False
self.training_thread = None
logger.info(f"CNN Trainer initialized")
logger.info(f"Model parameters: {sum(p.numel() for p in self.model.parameters()):,}")
logger.info(f"Storage directory: {self.storage_dir}")
def start_real_time_training(self, symbol: str):
"""Start real-time training for a symbol"""
if self.is_training:
logger.warning("CNN training already running")
return
self.is_training = True
self.training_thread = threading.Thread(
target=self._real_time_training_worker,
args=(symbol,),
daemon=True
)
self.training_thread.start()
logger.info(f"Started real-time CNN training for {symbol}")
def stop_training(self):
"""Stop training"""
self.is_training = False
if self.training_thread:
self.training_thread.join(timeout=10)
if self.current_session:
self._finalize_training_session()
logger.info("CNN training stopped")
def _real_time_training_worker(self, symbol: str):
"""Real-time training worker"""
logger.info(f"Real-time CNN training worker started for {symbol}")
while self.is_training:
try:
# Get high-priority episodes for training
episodes = self.data_collector.get_high_priority_episodes(
symbol=symbol,
limit=100,
min_priority=0.3
)
if len(episodes) >= 32: # Minimum batch size
self._train_on_episodes(episodes, training_mode='real_time')
# Wait before next training cycle
threading.Event().wait(300) # Train every 5 minutes
except Exception as e:
logger.error(f"Error in real-time training worker: {e}")
threading.Event().wait(60) # Wait before retrying
logger.info(f"Real-time CNN training worker stopped for {symbol}")
def train_on_profitable_episodes(self,
symbol: str,
min_profitability: float = 0.7,
max_episodes: int = 500) -> Dict[str, Any]:
"""Train specifically on most profitable episodes"""
try:
# Get all episodes for symbol
all_episodes = self.data_collector.training_episodes.get(symbol, [])
# Filter for profitable episodes
profitable_episodes = [
ep for ep in all_episodes
if (ep.actual_outcome.is_profitable and
ep.actual_outcome.profitability_score >= min_profitability)
]
# Sort by profitability and limit
profitable_episodes.sort(
key=lambda x: x.actual_outcome.profitability_score,
reverse=True
)
profitable_episodes = profitable_episodes[:max_episodes]
if len(profitable_episodes) < 10:
logger.warning(f"Insufficient profitable episodes for {symbol}: {len(profitable_episodes)}")
return {'status': 'insufficient_data', 'episodes_found': len(profitable_episodes)}
# Train on profitable episodes
results = self._train_on_episodes(
profitable_episodes,
training_mode='profitable_replay'
)
logger.info(f"Trained on {len(profitable_episodes)} profitable episodes for {symbol}")
return results
except Exception as e:
logger.error(f"Error training on profitable episodes: {e}")
return {'status': 'error', 'error': str(e)}
def _train_on_episodes(self,
episodes: List[TrainingEpisode],
training_mode: str = 'batch') -> Dict[str, Any]:
"""Train on a batch of episodes with comprehensive data storage"""
try:
# Start new training session
session = CNNTrainingSession(
session_id=f"{training_mode}_{datetime.now().strftime('%Y%m%d_%H%M%S')}",
start_timestamp=datetime.now(),
training_mode=training_mode,
symbol=episodes[0].input_package.symbol if episodes else 'unknown'
)
self.current_session = session
# Create dataset and dataloader
dataset = CNNTrainingDataset(episodes)
dataloader = DataLoader(
dataset,
batch_size=32,
shuffle=True,
num_workers=2
)
# Training loop
self.model.train()
total_loss = 0.0
num_batches = 0
for batch_idx, batch in enumerate(dataloader):
# Move to device
features = batch['features'].to(self.device)
pivot_class = batch['pivot_class'].to(self.device)
pivot_price = batch['pivot_price'].to(self.device)
confidence_target = batch['confidence_target'].to(self.device)
# Forward pass
self.optimizer.zero_grad()
outputs = self.model(features)
# Calculate losses
classification_loss = F.cross_entropy(outputs['pivot_logits'], pivot_class)
regression_loss = F.mse_loss(outputs['pivot_price'].squeeze(), pivot_price)
confidence_loss = F.binary_cross_entropy(
outputs['confidence'].squeeze(),
confidence_target
)
# Combined loss
total_batch_loss = classification_loss + 0.5 * regression_loss + 0.3 * confidence_loss
# Backward pass
total_batch_loss.backward()
# Gradient clipping
torch.nn.utils.clip_grad_norm_(self.model.parameters(), max_norm=1.0)
# Store gradients before optimizer step
gradients = {}
gradient_norms = {}
for name, param in self.model.named_parameters():
if param.grad is not None:
gradients[name] = param.grad.clone().detach()
gradient_norms[name] = param.grad.norm().item()
# Optimizer step
self.optimizer.step()
# Create training step record
step = CNNTrainingStep(
step_id=f"{session.session_id}_step_{batch_idx}",
timestamp=datetime.now(),
episode_id=f"batch_{batch_idx}",
input_features=features.detach().cpu(),
target_labels=pivot_class.detach().cpu(),
model_outputs={k: v.detach().cpu() for k, v in outputs.items()},
predictions=self._extract_predictions(outputs),
confidence_scores=outputs['confidence'].detach().cpu(),
total_loss=total_batch_loss.item(),
pivot_prediction_loss=classification_loss.item(),
confidence_loss=confidence_loss.item(),
regularization_loss=0.0,
gradients=gradients,
gradient_norms=gradient_norms,
learning_rate=self.optimizer.param_groups[0]['lr'],
batch_size=features.size(0)
)
# Calculate training value for this step
step.training_value = self._calculate_step_training_value(step, batch)
# Add to session
session.training_steps.append(step)
total_loss += total_batch_loss.item()
num_batches += 1
# Log progress
if batch_idx % 10 == 0:
logger.debug(f"Batch {batch_idx}: Loss = {total_batch_loss.item():.4f}")
# Finalize session
session.end_timestamp = datetime.now()
session.total_steps = num_batches
session.average_loss = total_loss / num_batches if num_batches > 0 else 0.0
session.best_loss = min(step.total_loss for step in session.training_steps)
# Calculate session value
session.session_value = self._calculate_session_value(session)
# Update scheduler
self.scheduler.step(session.average_loss)
# Save session
self._save_training_session(session)
# Update statistics
self.training_stats['total_sessions'] += 1
self.training_stats['total_steps'] += session.total_steps
if training_mode == 'profitable_replay':
self.training_stats['replay_sessions'] += 1
logger.info(f"Training session completed: {session.session_id}")
logger.info(f"Average loss: {session.average_loss:.4f}")
logger.info(f"Session value: {session.session_value:.3f}")
return {
'status': 'success',
'session_id': session.session_id,
'average_loss': session.average_loss,
'total_steps': session.total_steps,
'session_value': session.session_value
}
except Exception as e:
logger.error(f"Error in training session: {e}")
return {'status': 'error', 'error': str(e)}
finally:
self.current_session = None
def _extract_predictions(self, outputs: Dict[str, torch.Tensor]) -> Dict[str, Any]:
"""Extract human-readable predictions from model outputs"""
try:
pivot_probs = F.softmax(outputs['pivot_logits'], dim=1)
predicted_class = torch.argmax(pivot_probs, dim=1)
return {
'pivot_class': predicted_class.cpu().numpy().tolist(),
'pivot_probabilities': pivot_probs.cpu().numpy().tolist(),
'pivot_price': outputs['pivot_price'].cpu().numpy().tolist(),
'confidence': outputs['confidence'].cpu().numpy().tolist()
}
except Exception as e:
logger.warning(f"Error extracting predictions: {e}")
return {}
def _calculate_step_training_value(self,
step: CNNTrainingStep,
batch: Dict[str, Any]) -> float:
"""Calculate the training value of a step for replay prioritization"""
try:
value = 0.0
# Base value from loss (lower loss = higher value)
if step.total_loss > 0:
value += 1.0 / (1.0 + step.total_loss)
# Bonus for high profitability episodes in batch
avg_profitability = torch.mean(batch['profitability']).item()
value += avg_profitability * 0.3
# Bonus for gradient magnitude (indicates learning)
avg_grad_norm = np.mean(list(step.gradient_norms.values()))
value += min(avg_grad_norm / 10.0, 0.2) # Cap at 0.2
return min(value, 1.0)
except Exception as e:
logger.warning(f"Error calculating step training value: {e}")
return 0.0
def _calculate_session_value(self, session: CNNTrainingSession) -> float:
"""Calculate overall session value for replay prioritization"""
try:
if not session.training_steps:
return 0.0
# Average step values
avg_step_value = np.mean([step.training_value for step in session.training_steps])
# Bonus for convergence
convergence_bonus = 0.0
if len(session.training_steps) > 10:
early_loss = np.mean([s.total_loss for s in session.training_steps[:5]])
late_loss = np.mean([s.total_loss for s in session.training_steps[-5:]])
if early_loss > late_loss:
convergence_bonus = min((early_loss - late_loss) / early_loss, 0.3)
# Bonus for profitable replay sessions
mode_bonus = 0.2 if session.training_mode == 'profitable_replay' else 0.0
return min(avg_step_value + convergence_bonus + mode_bonus, 1.0)
except Exception as e:
logger.warning(f"Error calculating session value: {e}")
return 0.0
def _save_training_session(self, session: CNNTrainingSession):
"""Save training session to disk"""
try:
session_dir = self.storage_dir / session.symbol / 'sessions'
session_dir.mkdir(parents=True, exist_ok=True)
# Save full session data
session_file = session_dir / f"{session.session_id}.pkl"
with open(session_file, 'wb') as f:
pickle.dump(session, f)
# Save session metadata
metadata = {
'session_id': session.session_id,
'start_timestamp': session.start_timestamp.isoformat(),
'end_timestamp': session.end_timestamp.isoformat() if session.end_timestamp else None,
'training_mode': session.training_mode,
'symbol': session.symbol,
'total_steps': session.total_steps,
'average_loss': session.average_loss,
'best_loss': session.best_loss,
'session_value': session.session_value
}
metadata_file = session_dir / f"{session.session_id}_metadata.json"
with open(metadata_file, 'w') as f:
json.dump(metadata, f, indent=2)
logger.debug(f"Saved training session: {session.session_id}")
except Exception as e:
logger.error(f"Error saving training session: {e}")
def _finalize_training_session(self):
"""Finalize current training session"""
if self.current_session:
self.current_session.end_timestamp = datetime.now()
self._save_training_session(self.current_session)
self.training_sessions.append(self.current_session)
self.current_session = None
def get_training_statistics(self) -> Dict[str, Any]:
"""Get comprehensive training statistics"""
stats = self.training_stats.copy()
# Add recent session information
if self.training_sessions:
recent_sessions = sorted(
self.training_sessions,
key=lambda x: x.start_timestamp,
reverse=True
)[:10]
stats['recent_sessions'] = [
{
'session_id': s.session_id,
'timestamp': s.start_timestamp.isoformat(),
'mode': s.training_mode,
'average_loss': s.average_loss,
'session_value': s.session_value
}
for s in recent_sessions
]
# Calculate profitability rate
if stats['total_predictions'] > 0:
stats['profitability_rate'] = stats['profitable_predictions'] / stats['total_predictions']
else:
stats['profitability_rate'] = 0.0
return stats
def replay_high_value_sessions(self,
symbol: str,
min_session_value: float = 0.7,
max_sessions: int = 10) -> Dict[str, Any]:
"""Replay high-value training sessions"""
try:
# Find high-value sessions
high_value_sessions = [
s for s in self.training_sessions
if (s.symbol == symbol and
s.session_value >= min_session_value)
]
# Sort by value and limit
high_value_sessions.sort(key=lambda x: x.session_value, reverse=True)
high_value_sessions = high_value_sessions[:max_sessions]
if not high_value_sessions:
return {'status': 'no_high_value_sessions', 'sessions_found': 0}
# Replay sessions
total_replayed = 0
for session in high_value_sessions:
# Extract episodes from session steps
episode_ids = list(set(step.episode_id for step in session.training_steps))
# Get corresponding episodes
episodes = []
for episode_id in episode_ids:
# Find episode in data collector
for ep in self.data_collector.training_episodes.get(symbol, []):
if ep.episode_id == episode_id:
episodes.append(ep)
break
if episodes:
self._train_on_episodes(episodes, training_mode='high_value_replay')
total_replayed += 1
logger.info(f"Replayed {total_replayed} high-value sessions for {symbol}")
return {
'status': 'success',
'sessions_replayed': total_replayed,
'sessions_found': len(high_value_sessions)
}
except Exception as e:
logger.error(f"Error replaying high-value sessions: {e}")
return {'status': 'error', 'error': str(e)}
# Global instance
cnn_trainer = None
def get_cnn_trainer(model: CNNPivotPredictor = None) -> CNNTrainer:
"""Get global CNN trainer instance"""
global cnn_trainer
if cnn_trainer is None:
if model is None:
model = CNNPivotPredictor()
cnn_trainer = CNNTrainer(model)
return cnn_trainer