Function main

codeclash/analysis/viz/round_score_distribution.py:173–224

Calculate normalized scores and plot histograms stratified by game and by model.

(log_dir: Path)

def main(log_dir: Path) -> None:
    """Calculate normalized scores and plot histograms stratified by game and by model."""
    logger.info(f"Processing tournaments from {log_dir}")

    scores_by_game = {}
    scores_by_model = {}

    for metadata_path in tqdm(list(log_dir.rglob("metadata.json"))):
        try:
            game_name, game_scores, model_scores = get_normalized_scores(metadata_path)
            if game_name:
                # Collect by game
                for game, scores in game_scores.items():
                    if game not in scores_by_game:
                        scores_by_game[game] = []
                    scores_by_game[game].extend(scores)

                # Collect by model
                for model, scores in model_scores.items():
                    if model not in scores_by_model:
                        scores_by_model[model] = []
                    scores_by_model[model].extend(scores)
        except Exception as e:
            logger.error(f"Error processing {metadata_path}: {e}", exc_info=True)
            continue

    if not scores_by_game:
        logger.warning("No scores collected")
        return

    all_scores = [s for scores in scores_by_game.values() for s in scores]
    logger.info(
        f"Collected {len(all_scores)} normalized scores across {len(scores_by_game)} games and {len(scores_by_model)} models"
    )

    # Plot by game
    output_by_game = ASSETS_DIR / "round_score_distribution_by_game.pdf"
    plot_stratified(
        scores_by_game,
        output_by_game,
        title="Distribution of Normalized Player Scores by Game (Valid Rounds Only)",
        by_model=False,
    )

    # Plot by model
    output_by_model = ASSETS_DIR / "round_score_distribution_by_model.pdf"
    plot_stratified(
        scores_by_model,
        output_by_model,
        title="Distribution of Normalized Player Scores by Model (Valid Rounds Only)",
        by_model=True,
    )


if __name__ == "__main__":