trading_intelligence/training.py

"""
Training Pipeline
==================
End-to-end training with multi-task learning,
data loading, and proper financial time-series splits.
"""

import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import Dataset, DataLoader
import numpy as np
import pandas as pd
from typing import Dict, List, Tuple, Optional
import os
import json
import time

from trading_intelligence.feature_engine import FeatureEngine
from trading_intelligence.prediction_model import TradingTransformer, MultiTaskLoss


class FinancialTimeSeriesDataset(Dataset):
    """
    PyTorch Dataset for financial time series.
    
    Uses walk-forward split (no random shuffling to preserve temporal order).
    """
    
    def __init__(self, X: np.ndarray, y: np.ndarray):
        """
        Args:
            X: (N, num_features, lookback_window) feature sequences
            y: (N, num_targets) target values
        """
        self.X = torch.FloatTensor(X)
        self.y = torch.FloatTensor(y)
    
    def __len__(self):
        return len(self.X)
    
    def __getitem__(self, idx):
        return self.X[idx], self.y[idx]


class TrainingPipeline:
    """
    Complete training pipeline for the trading intelligence system.
    
    Features:
    1. Data loading and feature engineering
    2. Walk-forward temporal splits
    3. Multi-task training (direction + return + risk)
    4. Learning rate scheduling
    5. Early stopping with patience
    6. Comprehensive logging
    """
    
    def __init__(
        self,
        lookback_window: int = 60,
        prediction_horizons: List[int] = [1, 5, 20],
        d_model: int = 128,
        n_heads: int = 8,
        n_layers: int = 3,
        d_ff: int = 256,
        patch_len: int = 8,
        stride: int = 4,
        dropout: float = 0.1,
        learning_rate: float = 1e-3,
        batch_size: int = 64,
        max_epochs: int = 100,
        patience: int = 10,
        device: str = 'auto',
    ):
        self.lookback_window = lookback_window
        self.prediction_horizons = prediction_horizons
        self.d_model = d_model
        self.n_heads = n_heads
        self.n_layers = n_layers
        self.d_ff = d_ff
        self.patch_len = patch_len
        self.stride = stride
        self.dropout = dropout
        self.learning_rate = learning_rate
        self.batch_size = batch_size
        self.max_epochs = max_epochs
        self.patience = patience
        
        if device == 'auto':
            self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
        else:
            self.device = torch.device(device)
        
        self.feature_engine = FeatureEngine(lookback_window, prediction_horizons)
        self.model = None
        self.loss_fn = None
        self.optimizer = None
        self.scheduler = None
        self.training_history = []
    
    def prepare_data(self, df: pd.DataFrame, 
                     train_ratio: float = 0.7,
                     val_ratio: float = 0.15) -> Tuple[DataLoader, DataLoader, DataLoader]:
        """
        Prepare data with walk-forward temporal splits.
        
        Args:
            df: Raw OHLCV DataFrame
            train_ratio: Fraction for training (earliest data)
            val_ratio: Fraction for validation (middle)
            
        Returns:
            train_loader, val_loader, test_loader
        """
        # Feature engineering
        features_df = self.feature_engine.compute_all_features(df)
        
        # Normalize features
        features_df, self.norm_params = self.feature_engine.normalize_features(features_df)
        
        # Create target columns
        target_cols = []
        for h in self.prediction_horizons:
            target_cols.extend([f'target_direction_{h}', f'target_return_{h}'])
        
        # Create sequences
        X, y = self.feature_engine.create_sequences(features_df, target_cols=target_cols)
        
        # Remove any NaN/Inf
        valid_mask = np.isfinite(X).all(axis=(1, 2)) & np.isfinite(y).all(axis=1)
        X = X[valid_mask]
        y = y[valid_mask]
        
        print(f"Total valid samples: {len(X)}")
        print(f"Features per sample: {X.shape[1]} channels x {X.shape[2]} timesteps")
        print(f"Targets per sample: {y.shape[1]}")
        
        # Temporal split (NO shuffling - preserves time order)
        n = len(X)
        train_end = int(n * train_ratio)
        val_end = int(n * (train_ratio + val_ratio))
        
        X_train, y_train = X[:train_end], y[:train_end]
        X_val, y_val = X[train_end:val_end], y[train_end:val_end]
        X_test, y_test = X[val_end:], y[val_end:]
        
        print(f"Train: {len(X_train)}, Val: {len(X_val)}, Test: {len(X_test)}")
        
        # Store test data for evaluation
        self.X_test = X_test
        self.y_test = y_test
        
        # Create DataLoaders
        train_dataset = FinancialTimeSeriesDataset(X_train, y_train)
        val_dataset = FinancialTimeSeriesDataset(X_val, y_val)
        test_dataset = FinancialTimeSeriesDataset(X_test, y_test)
        
        train_loader = DataLoader(train_dataset, batch_size=self.batch_size, shuffle=True)
        val_loader = DataLoader(val_dataset, batch_size=self.batch_size, shuffle=False)
        test_loader = DataLoader(test_dataset, batch_size=self.batch_size, shuffle=False)
        
        # Initialize model with correct number of channels
        self.num_channels = X.shape[1]
        self._init_model()
        
        return train_loader, val_loader, test_loader
    
    def _init_model(self):
        """Initialize model, loss, optimizer, and scheduler."""
        self.model = TradingTransformer(
            num_channels=self.num_channels,
            seq_len=self.lookback_window,
            patch_len=self.patch_len,
            stride=self.stride,
            d_model=self.d_model,
            n_heads=self.n_heads,
            n_layers=self.n_layers,
            d_ff=self.d_ff,
            num_horizons=len(self.prediction_horizons),
            dropout=self.dropout,
        ).to(self.device)
        
        self.loss_fn = MultiTaskLoss(
            num_horizons=len(self.prediction_horizons)
        ).to(self.device)
        
        total_params = sum(p.numel() for p in self.model.parameters())
        print(f"Model initialized: {total_params:,} parameters")
        print(f"Device: {self.device}")
        
        self.optimizer = optim.AdamW(
            list(self.model.parameters()) + list(self.loss_fn.parameters()),
            lr=self.learning_rate,
            weight_decay=1e-4
        )
        
        self.scheduler = optim.lr_scheduler.CosineAnnealingWarmRestarts(
            self.optimizer, T_0=10, T_mult=2
        )
    
    def _parse_targets(self, y_batch: torch.Tensor) -> Dict[str, torch.Tensor]:
        """Parse target tensor into direction and return components."""
        num_horizons = len(self.prediction_horizons)
        
        # y layout: [dir_1, ret_1, dir_5, ret_5, dir_20, ret_20]
        directions = torch.stack([y_batch[:, i*2] for i in range(num_horizons)], dim=1)
        returns = torch.stack([y_batch[:, i*2+1] for i in range(num_horizons)], dim=1)
        
        return {
            'direction': directions,
            'returns': returns,
        }
    
    def train_epoch(self, train_loader: DataLoader) -> Dict[str, float]:
        """Train for one epoch."""
        self.model.train()
        epoch_losses = {'total': 0, 'direction': 0, 'return': 0, 'risk': 0}
        num_batches = 0
        
        for X_batch, y_batch in train_loader:
            X_batch = X_batch.to(self.device)
            y_batch = y_batch.to(self.device)
            
            # Forward pass
            predictions = self.model(X_batch)
            targets = self._parse_targets(y_batch)
            
            # Compute loss
            losses = self.loss_fn(predictions, targets)
            
            # Backward pass
            self.optimizer.zero_grad()
            losses['total_loss'].backward()
            
            # Gradient clipping
            torch.nn.utils.clip_grad_norm_(self.model.parameters(), max_norm=1.0)
            
            self.optimizer.step()
            
            # Accumulate
            epoch_losses['total'] += losses['total_loss'].item()
            epoch_losses['direction'] += losses['direction_loss'].item()
            epoch_losses['return'] += losses['return_loss'].item()
            epoch_losses['risk'] += losses['risk_loss'].item()
            num_batches += 1
        
        return {k: v / max(num_batches, 1) for k, v in epoch_losses.items()}
    
    @torch.no_grad()
    def validate(self, val_loader: DataLoader) -> Dict[str, float]:
        """Validate model."""
        self.model.eval()
        epoch_losses = {'total': 0, 'direction': 0, 'return': 0, 'risk': 0}
        all_direction_preds = []
        all_direction_targets = []
        num_batches = 0
        
        for X_batch, y_batch in val_loader:
            X_batch = X_batch.to(self.device)
            y_batch = y_batch.to(self.device)
            
            predictions = self.model(X_batch)
            targets = self._parse_targets(y_batch)
            
            losses = self.loss_fn(predictions, targets)
            
            epoch_losses['total'] += losses['total_loss'].item()
            epoch_losses['direction'] += losses['direction_loss'].item()
            epoch_losses['return'] += losses['return_loss'].item()
            epoch_losses['risk'] += losses['risk_loss'].item()
            
            # Track direction accuracy
            dir_preds = (torch.sigmoid(predictions['direction_logits']) > 0.5).float()
            all_direction_preds.append(dir_preds.cpu())
            all_direction_targets.append(targets['direction'].cpu())
            num_batches += 1
        
        avg_losses = {k: v / max(num_batches, 1) for k, v in epoch_losses.items()}
        
        # Direction accuracy per horizon
        if all_direction_preds:
            all_preds = torch.cat(all_direction_preds, dim=0)
            all_targets = torch.cat(all_direction_targets, dim=0)
            for i, h in enumerate(self.prediction_horizons):
                acc = (all_preds[:, i] == all_targets[:, i]).float().mean().item()
                avg_losses[f'direction_acc_{h}'] = acc
        
        return avg_losses
    
    def train(self, train_loader: DataLoader, val_loader: DataLoader) -> Dict:
        """
        Full training loop with early stopping.
        
        Returns training history.
        """
        best_val_loss = float('inf')
        patience_counter = 0
        best_model_state = None
        
        print(f"\n{'='*60}")
        print(f"Starting Training ({self.max_epochs} max epochs)")
        print(f"{'='*60}")
        
        for epoch in range(self.max_epochs):
            start = time.time()
            
            # Train
            train_losses = self.train_epoch(train_loader)
            
            # Validate
            val_metrics = self.validate(val_loader)
            
            # Update scheduler
            self.scheduler.step()
            
            elapsed = time.time() - start
            
            # Log
            epoch_record = {
                'epoch': epoch + 1,
                'train_loss': train_losses['total'],
                'val_loss': val_metrics['total'],
                'train_dir_loss': train_losses['direction'],
                'val_dir_loss': val_metrics['direction'],
                'train_ret_loss': train_losses['return'],
                'val_ret_loss': val_metrics['return'],
                'lr': self.optimizer.param_groups[0]['lr'],
                'elapsed': elapsed,
            }
            for h in self.prediction_horizons:
                key = f'direction_acc_{h}'
                if key in val_metrics:
                    epoch_record[key] = val_metrics[key]
            
            self.training_history.append(epoch_record)
            
            # Print progress
            acc_str = " | ".join([
                f"DA-{h}d: {val_metrics.get(f'direction_acc_{h}', 0):.1%}" 
                for h in self.prediction_horizons
            ])
            print(
                f"Epoch {epoch+1:3d}/{self.max_epochs} | "
                f"Train: {train_losses['total']:.4f} | "
                f"Val: {val_metrics['total']:.4f} | "
                f"{acc_str} | "
                f"LR: {self.optimizer.param_groups[0]['lr']:.6f} | "
                f"{elapsed:.1f}s"
            )
            
            # Early stopping
            if val_metrics['total'] < best_val_loss:
                best_val_loss = val_metrics['total']
                patience_counter = 0
                best_model_state = {k: v.cpu().clone() for k, v in self.model.state_dict().items()}
            else:
                patience_counter += 1
                if patience_counter >= self.patience:
                    print(f"\nEarly stopping at epoch {epoch+1} (patience={self.patience})")
                    break
        
        # Restore best model
        if best_model_state:
            self.model.load_state_dict(best_model_state)
            self.model.to(self.device)
            print(f"Restored best model (val_loss={best_val_loss:.4f})")
        
        return {
            'best_val_loss': best_val_loss,
            'total_epochs': len(self.training_history),
            'history': self.training_history,
        }
    
    def save_model(self, path: str):
        """Save model and training artifacts."""
        os.makedirs(os.path.dirname(path) if os.path.dirname(path) else '.', exist_ok=True)
        
        save_dict = {
            'model_state': self.model.state_dict(),
            'loss_fn_state': self.loss_fn.state_dict(),
            'norm_params': self.norm_params if hasattr(self, 'norm_params') else {},
            'feature_names': self.feature_engine.feature_names,
            'config': {
                'lookback_window': self.lookback_window,
                'prediction_horizons': self.prediction_horizons,
                'num_channels': self.num_channels,
                'd_model': self.d_model,
                'n_heads': self.n_heads,
                'n_layers': self.n_layers,
                'd_ff': self.d_ff,
                'patch_len': self.patch_len,
                'stride': self.stride,
                'dropout': self.dropout,
            },
            'training_history': self.training_history,
        }
        
        torch.save(save_dict, path)
        print(f"Model saved to {path}")
    
    def load_model(self, path: str):
        """Load model from checkpoint."""
        checkpoint = torch.load(path, map_location=self.device, weights_only=False)
        config = checkpoint['config']
        
        # Restore all architecture params from checkpoint
        self.num_channels = config['num_channels']
        self.d_model = config.get('d_model', self.d_model)
        self.n_heads = config.get('n_heads', self.n_heads)
        self.n_layers = config.get('n_layers', self.n_layers)
        self.d_ff = config.get('d_ff', self.d_ff)
        self.patch_len = config.get('patch_len', self.patch_len)
        self.stride = config.get('stride', self.stride)
        self.dropout = config.get('dropout', self.dropout)
        self.lookback_window = config.get('lookback_window', self.lookback_window)
        if 'prediction_horizons' in config:
            self.prediction_horizons = config['prediction_horizons']
        
        self._init_model()
        
        self.model.load_state_dict(checkpoint['model_state'])
        self.loss_fn.load_state_dict(checkpoint['loss_fn_state'])
        self.norm_params = checkpoint.get('norm_params', {})
        self.feature_engine.feature_names = checkpoint.get('feature_names', [])
        self.training_history = checkpoint.get('training_history', [])
        
        print(f"Model loaded from {path}")