compare_eval_results.py

#!/usr/bin/env python3
"""Compare evaluation results across multiple models, split by command type.

Usage:
    python compare_eval_results.py \
        experiments_v2_location/combined_v1/eval_outputs_old_no_speech/results.eval.csv \
        experiments_v2_location/combined_v1_large/eval_outputs_old_no_speech/results.eval.csv \
        experiments_v2_location/no_TSDL_old_mixtures/eval_outputs_old_no_speech/results.eval.csv \
        experiments_v2_location/no_TSDL_old_mixtures_large/eval_outputs_old_no_speech/results.eval.csv

    # Save summary to CSV:
    python compare_eval_results.py --out summary.csv ...
"""
import argparse
import os
import pandas as pd
import numpy as np


SHORT_NAMES = {
    "scale_invariant_signal_noise_ratio": "SI-SNR-i",
    "signal_noise_ratio": "SNR-i",
    "si_snr": "SI-SNR-abs",
    "si_snr_improvement": "SI-SNR-i",
    "snr_improvement": "SNR-i",
    "si_snr_absolute": "SI-SNR-abs",
    "td_loss": "TD Loss",
    "td_freq_weighted_score": "TD FreqW",
    "td_multi_scale_score": "TD MultiS",
    "td_combined_score": "TD Combined",
    "delta_ITD": "dITD",
    "delta_ITD_gcc": "dITD_gcc",
    "delta_ILD": "dILD",
    "msclap_score": "CLAP",
    "spatial_clap_score": "Spatial CLAP",
}

# Non-metric columns to exclude from aggregation
META_COLS = {"mixture_id", "mixture_file", "command_type", "user_input", "target_sources"}


def extract_model_name(path):
    parts = path.replace("\\", "/").split("/")
    for i, p in enumerate(parts):
        if p.startswith("eval_outputs"):
            return parts[i - 1]
    return os.path.basename(os.path.dirname(path))


def print_table(rows, metric_cols, title=None):
    """Print a formatted table given a list of (model_name, metrics_dict) rows."""
    if title:
        print(f"\n{'=' * 40}")
        print(f"  {title}")
        print(f"{'=' * 40}")

    if not rows:
        print("  (no data)")
        return

    model_names = [r[0] for r in rows]
    max_name_len = max(len(n) for n in model_names)
    col_width = 14

    header = f"{'Model':<{max_name_len}}"
    for m in metric_cols:
        display = SHORT_NAMES.get(m, m)
        header += f"  {display:>{col_width}}"
    print(header)
    print("-" * len(header))

    for name, metrics in rows:
        row = f"{name:<{max_name_len}}"
        for m in metric_cols:
            val = metrics.get(m, float("nan"))
            row += f"  {val:>{col_width}.4f}"
        print(row)


def main():
    parser = argparse.ArgumentParser(description="Compare eval results across models")
    parser.add_argument("paths", nargs="+", help="Paths to results.eval.csv files")
    parser.add_argument("--out", type=str, default=None, help="Save summary table to CSV")
    args = parser.parse_args()

    # Load all CSVs
    all_dfs = {}
    for path in args.paths:
        name = extract_model_name(path)
        df = pd.read_csv(path)
        df['model'] = name
        all_dfs[name] = df

    combined = pd.concat(all_dfs.values(), ignore_index=True)

    # Identify metric columns (numeric, non-meta)
    metric_cols = [c for c in combined.columns
                   if c not in META_COLS and c != 'model'
                   and pd.api.types.is_numeric_dtype(combined[c])]

    model_names = list(all_dfs.keys())

    def safe_means(df, metric_cols):
        return {m: df[m].mean() if m in df.columns else float("nan") for m in metric_cols}

    # --- Overall ---
    overall_rows = []
    for name in model_names:
        overall_rows.append((name, safe_means(all_dfs[name], metric_cols)))
    print_table(overall_rows, metric_cols, "OVERALL")

    # --- By command_type ---
    cmd_types = sorted(combined['command_type'].dropna().unique())
    for cmd in cmd_types:
        cmd_rows = []
        for name in model_names:
            subset = all_dfs[name][all_dfs[name]['command_type'] == cmd]
            if len(subset) == 0:
                continue
            cmd_rows.append((name, safe_means(subset, metric_cols)))
        n_samples = len(combined[(combined['command_type'] == cmd) & (combined['model'] == model_names[0])])
        print_table(cmd_rows, metric_cols, f"command_type = {cmd}  (n={n_samples})")

    # --- Save summary CSV ---
    if args.out:
        summary_rows = []
        for cmd in ['overall'] + cmd_types:
            for name in model_names:
                df = all_dfs[name]
                subset = df if cmd == 'overall' else df[df['command_type'] == cmd]
                if len(subset) == 0:
                    continue
                row = {'model': name, 'command_type': cmd, 'n_samples': len(subset)}
                for m in metric_cols:
                    row[SHORT_NAMES.get(m, m)] = subset[m].mean() if m in subset.columns else float("nan")
                summary_rows.append(row)
        pd.DataFrame(summary_rows).to_csv(args.out, index=False)
        print(f"\nSaved to {args.out}")


if __name__ == "__main__":
    main()