experiments/reviewer_analyses/interface_ablation.py

from __future__ import annotations

import argparse
from pathlib import Path

from common import group_by, markdown_table, read_csv, safe_mean, safe_std, write_csv


VSTYLE_DIMS = ["overall", "ESMOS", "IEMOS", "TTMOS", "PVMOS", "NCMOS"]
RENDER_DIMS = ["autopcp_ref", "autopcp_ref_var", "emo2vec_ref", "audonnx_ref", "emo2vec_ref_unsmooth", "audonnx_ref_unsmooth", "wer"]
THINKER_DIMS = ["emotion_f1", "intensity_acc", "transition_f1", "vad_bin_f1", "global_style_f1"]


def main() -> None:
    parser = argparse.ArgumentParser(description="Build the explicit interface ablation table.")
    parser.add_argument("--vstyle", required=True, help="CSV with system, seed, and VStyle/MOS dimensions.")
    parser.add_argument("--render", required=True, help="CSV with system, seed, and rendering metrics.")
    parser.add_argument("--thinker", default="", help="Optional CSV with system, seed, and per-field Thinker metrics.")
    parser.add_argument("--out-dir", required=True)
    args = parser.parse_args()

    vstyle = read_csv(Path(args.vstyle))
    render = read_csv(Path(args.render))
    thinker = read_csv(Path(args.thinker)) if args.thinker else []

    systems = sorted(set(r["system"] for r in vstyle) | set(r["system"] for r in render) | set(r["system"] for r in thinker))
    v_by_system = group_by(vstyle, "system")
    r_by_system = group_by(render, "system")
    t_by_system = group_by(thinker, "system")

    rows = []
    for system in systems:
        row = {"system": system}
        for dim in VSTYLE_DIMS:
            row[dim] = _mean_pm(v_by_system.get(system, []), dim)
        for dim in RENDER_DIMS:
            row[dim] = _mean_pm(r_by_system.get(system, []), dim)
        for dim in THINKER_DIMS:
            row[dim] = _mean_pm(t_by_system.get(system, []), dim)
        rows.append(row)

    out_dir = Path(args.out_dir)
    write_csv(out_dir / "tab_interface.csv", rows)
    (out_dir / "tab_interface.md").write_text(markdown_table(rows), encoding="utf-8")


def _mean_pm(rows, key):
    vals = [r.get(key, "") for r in rows if r.get(key, "") != ""]
    if not vals:
        return "--"
    return f"{safe_mean(vals):.3f} +/- {safe_std(vals):.3f}"


if __name__ == "__main__":
    main()