analyze_detection_scores_gt_relative_complement.py

#!/usr/bin/env python3
"""
Complement-based GT event detection analysis.

Metric: success rate = fraction of samples where
    score(model output, GT) > score(model output, complement_GT)

GT and complement are constructed from spatially-rendered stems:
  - REMOVED: GT = speech only,  complement = speech + distractor
  - PRESENT: GT = speech + dist, complement = speech only

This removes the background-noise bias of the old mixture-based metric.
Three metrics reported: SI-SNR, NXCorr, CLAP similarity.

Only eval_outputs_test_3k (non-OOD) is used.

Usage:
    python analyze_detection_scores_gt_relative_complement.py
"""

import json
from pathlib import Path

import numpy as np
import pandas as pd

# ── Paths ─────────────────────────────────────────────────────────────────────
BASE_DIR    = Path(__file__).parent
JSON_DIR    = BASE_DIR / "data/audio_mixtures_old/test"

# ── SNR bins ──────────────────────────────────────────────────────────────────
SNR_BINS   = [-np.inf, -5, 0, 5, 10, 15, np.inf]
SNR_LABELS = ["< -5 dB", "-5 to 0 dB", "0 to 5 dB", "5 to 10 dB", "10 to 15 dB", "≥ 15 dB"]

MODELS_FINAL = {
    "combined_v1_broken_left_ch":   BASE_DIR / "experiments_v2/combined_v1_broken_left_ch",
    "combined_v1_large":            BASE_DIR / "experiments_v2/combined_v1_large",
    "no_TSDL_old_mixtures":         BASE_DIR / "experiments_v2/no_TSDL_old_mixtures",
    "no_TSDL_old_mixtures_large":   BASE_DIR / "experiments_v2/no_TSDL_old_mixtures_large",
}

TEST_3K_CSV = "eval_outputs_test_3k/event_detection_scores_complement.csv"

METRICS = [
    ("SI-SNR",   "success_sisnr"),
    ("NXCorr",   "success_nxcorr"),
    ("CLAP sim", "success_clap"),
    ("Pooled",   "success_pooled"),
]


# ═══════════════════════════════════════════════════════════════════════════════

def _load_snr_lookup() -> dict:
    """Build {(mixture_id, distractor_name): target_snr_db} from JSON files."""
    lookup = {}
    for jpath in JSON_DIR.glob("*.json"):
        try:
            meta = json.loads(jpath.read_text())
            mid  = meta.get("mixture_id", jpath.stem)
            for dist_name, info in meta.get("snr_info", {}).items():
                if dist_name in ("speech", "background"):
                    continue
                snr = info.get("target_snr_db")
                if snr is not None:
                    lookup[(mid, dist_name)] = snr
        except Exception:
            pass
    return lookup


_SNR_LOOKUP = None  # type: dict


def load_csv(path: Path) -> pd.DataFrame:
    global _SNR_LOOKUP
    if _SNR_LOOKUP is None:
        _SNR_LOOKUP = _load_snr_lookup()

    df = pd.read_csv(path)
    df = df[df["error"].isna() | (df["error"] == "")]
    for _, col in METRICS:
        if col != "success_pooled":
            df[col] = pd.to_numeric(df[col], errors="coerce")
    # Pooled: per-sample mean of the 3 binary success flags
    df["success_pooled"] = df[["success_sisnr", "success_nxcorr", "success_clap"]].mean(axis=1)
    # Parse number of distractors from mixture_id  (e.g. "airport_1dist_005_rep1_v0" → 1)
    df["num_distractors"] = (
        df["mixture_id"].str.extract(r"(\d+)dist", expand=False)
                        .astype(float).astype("Int64")
    )
    # Join per-distractor SNR from JSON metadata
    df["distractor_snr_db"] = df.apply(
        lambda r: _SNR_LOOKUP.get((r["mixture_id"], r["distractor_name"]), np.nan),
        axis=1,
    )
    df["snr_bin"] = pd.cut(
        df["distractor_snr_db"], bins=SNR_BINS, labels=SNR_LABELS, right=True
    )
    return df


def print_section(title: str):
    print(f"\n{'═'*70}")
    print(f"  {title}")
    print(f"{'═'*70}")


def success_rate(df: pd.DataFrame, col: str) -> float:
    """Fraction of rows where success == 1 (output closer to GT than complement)."""
    valid = df[col].notna()
    if valid.sum() == 0:
        return float("nan")
    return df.loc[valid, col].mean() * 100.0


# ═══════════════════════════════════════════════════════════════════════════════
# Score statistics  (out→GT vs out→complement absolute values)
# ═══════════════════════════════════════════════════════════════════════════════

def print_score_stats(dfs: dict):
    print_section("Score statistics  (output→GT  and  output→complement)")
    score_cols = [
        ("out_si_snr_db",   "out→GT SI-SNR"),
        ("comp_si_snr_db",  "out→comp SI-SNR"),
        ("out_nxcorr",      "out→GT NXCorr"),
        ("comp_nxcorr",     "out→comp NXCorr"),
        ("out_clap_sim",    "out→GT CLAP"),
        ("comp_clap_sim",   "out→comp CLAP"),
        ("success_sisnr",   "success SI-SNR"),
        ("success_nxcorr",  "success NXCorr"),
        ("success_clap",    "success CLAP"),
        ("success_pooled",  "success Pooled"),
    ]
    model_names = list(dfs.keys())
    header = f"  {'column':<16} {'stat':<8}"
    for name in model_names:
        header += f"  {name:>26}"
    print(header)
    print("  " + "─" * (26 + len(model_names) * 28))

    for col, label in score_cols:
        for stat, fn in [("mean", np.nanmean), ("median", np.nanmedian)]:
            row = f"  {label:<16} {stat:<8}"
            for df in dfs.values():
                if col in df.columns:
                    df[col] = pd.to_numeric(df[col], errors="coerce")
                    row += f"  {fn(df[col].dropna().values):>26.4f}"
                else:
                    row += f"  {'N/A':>26}"
            print(row)
        print()


# ═══════════════════════════════════════════════════════════════════════════════
# Overall success rate
# ═══════════════════════════════════════════════════════════════════════════════

def print_overall(dfs: dict):
    print_section(
        "Overall success rate  (% samples where output closer to GT than complement)\n"
        "  Threshold-free — no operating-point tuning required"
    )
    model_names = list(dfs.keys())
    print(f"\n  {'Metric':<12} {'N':>6}", end="")
    for name in model_names:
        print(f"  {name:>26}", end="")
    print()
    print("  " + "─" * (20 + len(model_names) * 28))

    for m_label, col in METRICS:
        first_df = list(dfs.values())[0]
        n = first_df[col].notna().sum()
        print(f"  {m_label:<12} {n:>6}", end="")
        for df in dfs.values():
            sr = success_rate(df, col)
            print(f"  {'%.2f%%' % sr:>26}", end="")
        print()


# ═══════════════════════════════════════════════════════════════════════════════
# By gt_label (REMOVED / PRESENT)
# ═══════════════════════════════════════════════════════════════════════════════

def print_by_gt_label(dfs: dict):
    print_section("Success rate by gt_label  (REMOVED = model should remove; PRESENT = keep)")
    model_names = list(dfs.keys())

    for m_label, col in METRICS:
        print(f"\n  [ {m_label} ]")
        print(f"  {'gt_label':<12} {'N':>6}", end="")
        for name in model_names:
            print(f"  {name:>26}", end="")
        print()
        print("  " + "─" * (20 + len(model_names) * 28))

        for lbl in ["REMOVED", "PRESENT"]:
            n_shown = False
            row_str = f"  {lbl:<12}"
            for df in dfs.values():
                sub   = df[df["gt_label"] == lbl]
                valid = sub[col].notna()
                if not n_shown:
                    row_str += f" {valid.sum():>6}"
                    n_shown = True
                sr = sub.loc[valid, col].mean() * 100.0 if valid.sum() > 0 else float("nan")
                row_str += f"  {'%.2f%%' % sr if not np.isnan(sr) else 'N/A':>26}"
            print(row_str)


# ═══════════════════════════════════════════════════════════════════════════════
# By command_type
# ═══════════════════════════════════════════════════════════════════════════════

def print_by_command_type(dfs: dict):
    print_section("Success rate by command_type")
    model_names = list(dfs.keys())
    command_types = sorted(
        set().union(*[set(df["command_type"].dropna()) for df in dfs.values()])
    )

    for m_label, col in METRICS:
        print(f"\n  [ {m_label} ]\n")
        print(f"  {'command_type':<22} {'N':>6}", end="")
        for name in model_names:
            print(f"  {name:>26}", end="")
        print()
        print("  " + "─" * (30 + len(model_names) * 28))

        for ct in command_types:
            n_shown = False
            row_str = f"  {ct:<22}"
            for df in dfs.values():
                sub   = df[df["command_type"] == ct]
                valid = sub[col].notna()
                if not n_shown:
                    row_str += f" {valid.sum():>6}"
                    n_shown = True
                sr = sub.loc[valid, col].mean() * 100.0 if valid.sum() > 0 else float("nan")
                row_str += f"  {'%.2f%%' % sr if not np.isnan(sr) else 'N/A':>26}"
            print(row_str)


# ═══════════════════════════════════════════════════════════════════════════════
# By number of distractors in mixture
# ═══════════════════════════════════════════════════════════════════════════════

def print_by_num_distractors(dfs: dict):
    print_section(
        "Success rate by number of distractors in mixture\n"
        "  (does the model struggle more with more distractors?)"
    )
    model_names = list(dfs.keys())
    all_counts = sorted(
        set().union(*[set(df["num_distractors"].dropna().unique()) for df in dfs.values()])
    )

    for m_label, col in METRICS:
        print(f"\n  [ {m_label} ]\n")
        print(f"  {'num_distractors':<18} {'N':>6}", end="")
        for name in model_names:
            print(f"  {name:>26}", end="")
        print()
        print("  " + "─" * (26 + len(model_names) * 28))

        for nd in all_counts:
            n_shown = False
            row_str = f"  {str(int(nd)) + ' distractor(s)':<18}"
            for df in dfs.values():
                sub   = df[df["num_distractors"] == nd]
                valid = sub[col].notna()
                if not n_shown:
                    row_str += f" {valid.sum():>6}"
                    n_shown = True
                sr = sub.loc[valid, col].mean() * 100.0 if valid.sum() > 0 else float("nan")
                row_str += f"  {'%.2f%%' % sr if not np.isnan(sr) else 'N/A':>26}"
            print(row_str)

    # Also show the cross-tab: num_distractors × command_type for CLAP
    print_section("Success rate: num_distractors × command_type  (CLAP sim, combined_v1)")
    col = "success_clap"
    first_df = list(dfs.values())[0]
    command_types = sorted(first_df["command_type"].dropna().unique())

    header_label = "num_dist / cmd_type"
    print(f"\n  {header_label:<20}", end="")
    for ct in command_types:
        print(f"  {ct:>18}", end="")
    print(f"  {'ALL':>18}")
    print("  " + "─" * (22 + (len(command_types) + 1) * 20))

    for nd in all_counts:
        row_str = f"  {str(int(nd)) + ' dist':<20}"
        sub_nd = first_df[first_df["num_distractors"] == nd]
        for ct in command_types:
            sub = sub_nd[sub_nd["command_type"] == ct]
            valid = sub[col].notna()
            sr = sub.loc[valid, col].mean() * 100.0 if valid.sum() > 0 else float("nan")
            row_str += f"  {'%.1f%%' % sr if not np.isnan(sr) else 'N/A':>18}"
        valid_all = sub_nd[col].notna()
        sr_all = sub_nd.loc[valid_all, col].mean() * 100.0 if valid_all.sum() > 0 else float("nan")
        row_str += f"  {'%.1f%%' % sr_all:>18}"
        print(row_str)


# ═══════════════════════════════════════════════════════════════════════════════
# By distractor  (CLAP — sorted by first model)
# ═══════════════════════════════════════════════════════════════════════════════

def print_by_distractor(dfs: dict):
    model_names = list(dfs.keys())
    print_section(
        f"Success rate by distractor_name  (CLAP)  sorted by {model_names[0]}"
    )
    col = "success_clap"
    dist_names = sorted(list(dfs.values())[0]["distractor_name"].dropna().unique())

    rows = []
    for dname in dist_names:
        row = {"distractor": dname}
        for name, df in dfs.items():
            sub   = df[df["distractor_name"] == dname]
            valid = sub[col].notna()
            n     = valid.sum()
            sr    = sub.loc[valid, col].mean() * 100.0 if n > 0 else float("nan")
            row[name]        = sr
            row[f"{name}_n"] = n
        rows.append(row)

    dist_df = pd.DataFrame(rows).sort_values(model_names[0], ascending=False)

    print(f"\n  {'distractor':<32} {'N':>5}", end="")
    for name in model_names:
        print(f"  {name:>26}", end="")
    print()
    print("  " + "─" * (39 + len(model_names) * 28))

    for _, r in dist_df.iterrows():
        n = int(r[f"{model_names[0]}_n"])
        print(f"  {r['distractor']:<32} {n:>5}", end="")
        for name in model_names:
            v = r[name]
            print(f"  {'%.2f%%' % v if not np.isnan(v) else 'N/A':>26}", end="")
        print()


# ═══════════════════════════════════════════════════════════════════════════════
# By distractor SNR bin
# ═══════════════════════════════════════════════════════════════════════════════

def print_by_snr(dfs: dict):
    print_section(
        "Success rate by distractor SNR bin  (target_snr_db from JSON metadata)\n"
        "  SNR = distractor level relative to speech at mixing time"
    )
    model_names = list(dfs.keys())

    # Coverage report
    first_df = list(dfs.values())[0]
    n_total   = len(first_df)
    n_matched = first_df["distractor_snr_db"].notna().sum()
    print(f"\n  SNR lookup coverage: {n_matched}/{n_total} rows "
          f"({100*n_matched/n_total:.1f}%)")

    for m_label, col in METRICS:
        print(f"\n  [ {m_label} ]\n")
        print(f"  {'SNR bin':<16} {'N':>6}", end="")
        for name in model_names:
            print(f"  {name:>26}", end="")
        print()
        print("  " + "─" * (24 + len(model_names) * 28))

        for lbl in SNR_LABELS:
            n_shown = False
            row_str = f"  {lbl:<16}"
            for df in dfs.values():
                sub   = df[df["snr_bin"] == lbl]
                valid = sub[col].notna()
                if not n_shown:
                    row_str += f" {valid.sum():>6}"
                    n_shown = True
                sr = sub.loc[valid, col].mean() * 100.0 if valid.sum() > 0 else float("nan")
                row_str += f"  {'%.2f%%' % sr if not np.isnan(sr) else 'N/A':>26}"
            print(row_str)


# ═══════════════════════════════════════════════════════════════════════════════

def main():
    dfs = {}
    for name, model_path in MODELS_FINAL.items():
        p = model_path / TEST_3K_CSV
        if not p.exists():
            print(f"[WARN] CSV not found: {p}")
            continue
        df = load_csv(p)
        print(f"Loaded {len(df):>5} rows  ← {name}")
        dfs[name] = df

    if not dfs:
        print("No CSVs found. Run the GT-relative eval jobs first.")
        return

    print_score_stats(dfs)
    print_overall(dfs)
    print_by_gt_label(dfs)
    print_by_command_type(dfs)
    print_by_num_distractors(dfs)
    print_by_distractor(dfs)
    print_by_snr(dfs)


if __name__ == "__main__":
    main()