Trains all players and pulls data for one player and outputs the pred…

[LalitYemireddy]

What I changed:

• Added src/training_player_data/train_one_player.py
• Implemented a global training pipeline that:
• Trains on all NBA players instead of a single player
• Uses rolling window feature engineering
• Applies log-transform (log1p) to stabilize regression targets
• Uses mini-batch training via DataLoader
• Implements early stopping with configurable patience
• Outputs:
• Test MAE (in points)
• Last N-game average baseline
• Predicted next-game points for a given person_id
• Added safe handling for early termination (Ctrl+C) to preserve the best model weights

Why I changed it:
The model now improves:

• Model stability
• Training efficiency (mini-batching)
• Generalization (training on 1.27M+ rows)
• Prediction realism (log-transformed targets + non-negative outputs)

How I tested it:

• Trained on 1,278,222 rows with 11 rolling features.
• Verified validation loss decreases during training.
• Confirmed early stopping works correctly.
• Verified that stopping training with Ctrl+C still loads the best model and prints results.
• Use this command to test it out:
  py src\training_player_data\train_one_player.py --person_id 1627739 --window 5