#!/usr/bin/env python3 """ OOD no_input heatmap: scene × distractor coloured by KEPT (+1) vs REMOVED (-1). Filters command_type == "no_input" from each OOD eval set. Logic (GT is always speech-only in OOD): success_clap = 1 → output closer to speech-only GT → model REMOVED → score = -1 success_clap = 0 → output closer to mixture → model KEPT → score = +1 Cell value = mean score across all samples in that scene × distractor cell. +1.0 = always kept -1.0 = always removed 0.0 = 50 / 50 Three heatmaps — one per OOD split — two subplots per figure (one per model). Usage: python plot_ood_no_input_heatmap.py """ from pathlib import Path import numpy as np import pandas as pd import matplotlib matplotlib.use("Agg") import matplotlib.pyplot as plt import seaborn as sns BASE_DIR = Path(__file__).parent MODELS = { "combined_v1": "experiments_final/combined_v1", "no_TSDL_old_mixtures": "experiments_final/no_TSDL_old_mixtures", } OOD_SPLITS = { "OOD_distractors": "eval_outputs_OOD_distractors", # unseen distractor classes "OOD_backgrounds": "eval_outputs_OOD_backgrounds", # unseen background scenes "OOD_both": "eval_outputs_OOD_both", # both unseen } CSV_NAME = "event_detection_scores_gt_relative.csv" CMD_TYPE = "no_input" SUCCESS_COL = "success_clap" # ═══════════════════════════════════════════════════════════════════════════════ def load(path: Path) -> pd.DataFrame: df = pd.read_csv(path) df = df[df["error"].isna() | (df["error"] == "")] df = df[df["command_type"] == CMD_TYPE].copy() df[SUCCESS_COL] = pd.to_numeric(df[SUCCESS_COL], errors="coerce") # kept_score: +1 = kept, -1 = removed df["kept_score"] = 1 - 2 * df[SUCCESS_COL] # success=1→-1, success=0→+1 df["scene"] = df["mixture_id"].str.extract(r"^(.+?)_\d+dist_") return df def build_pivot(df: pd.DataFrame, scene_order: list, dist_order: list) -> pd.DataFrame: pivot = ( df.groupby(["scene", "distractor_name"])["kept_score"] .mean() .unstack("distractor_name") .reindex(index=scene_order, columns=dist_order) ) return pivot def annot_array(pivot: pd.DataFrame) -> np.ndarray: """Return string array: +value with KEPT/REM label.""" arr = np.empty(pivot.shape, dtype=object) for i, row in enumerate(pivot.index): for j, col in enumerate(pivot.columns): v = pivot.iloc[i, j] if np.isnan(v): arr[i, j] = "" else: pct = abs(v) * 100 label = "KEPT" if v > 0 else "REM" arr[i, j] = f"{label}\n{pct:.0f}%" return arr def plot_split(split_name: str, dfs: dict, out_path: Path): """ dfs: {model_name: DataFrame} — already filtered to no_input for this split. """ # Shared distractor + scene ordering by pooled kept_score (most-kept first) all_df = pd.concat(dfs.values(), ignore_index=True) if all_df.empty: print(f" [SKIP] No no_input data for {split_name}") return dist_order = ( all_df.groupby("distractor_name")["kept_score"] .mean() .sort_values(ascending=False) # most-kept → most-removed .index.tolist() ) scene_order = ( all_df.groupby("scene")["kept_score"] .mean() .sort_values(ascending=False) .index.tolist() ) if not scene_order: scene_order = ["(all)"] n_models = len(dfs) n_dist = max(len(dist_order), 1) n_scene = max(len(scene_order), 1) fig, axes = plt.subplots( 1, n_models, figsize=(max(10, n_dist * 0.55) * n_models, max(4, n_scene * 0.55 + 3)), squeeze=False, ) fig.suptitle( f"OOD no_input | {split_name}\n" f"KEPT (+1) vs REMOVED (−1) — model default behaviour on unseen sounds", fontsize=12, fontweight="bold", y=1.02, ) for ax, (model_name, df) in zip(axes[0], dfs.items()): if df.empty: ax.set_title(f"{model_name}\n(no data)") ax.axis("off") continue pivot = build_pivot(df, scene_order, dist_order) annots = annot_array(pivot) mask = pivot.isna() sns.heatmap( pivot, ax=ax, mask=mask, annot=annots, fmt="", annot_kws={"size": 7, "weight": "bold"}, vmin=-1, vmax=1, center=0, cmap="RdYlGn", # green = kept, red = removed linewidths=0.4, linecolor="#cccccc", cbar_kws={ "label": "← REMOVED | KEPT →", "shrink": 0.6, "ticks": [-1, -0.5, 0, 0.5, 1], }, ) # Overall kept % per distractor (column header) dist_overall = df.groupby("distractor_name")["kept_score"].mean() xlabels = [] for d in dist_order: v = dist_overall.get(d, float("nan")) if np.isnan(v): xlabels.append(d) else: pct = abs(v) * 100 tag = "K" if v > 0 else "R" xlabels.append(f"{d}\n({tag}{pct:.0f}%)") ax.set_xticklabels(xlabels, fontsize=7.5, rotation=45, ha="right") # Overall kept % per scene (row header) scene_overall = df.groupby("scene")["kept_score"].mean() ylabels = [] for s in scene_order: v = scene_overall.get(s, float("nan")) if np.isnan(v): ylabels.append(s) else: pct = abs(v) * 100 tag = "K" if v > 0 else "R" ylabels.append(f"{s} ({tag}{pct:.0f}%)") ax.set_yticklabels(ylabels, fontsize=9, rotation=0) ax.set_title(model_name, fontsize=10, fontweight="bold", pad=8) ax.set_xlabel("Distractor (OOD)", fontsize=10) ax.set_ylabel("Scene", fontsize=10) plt.tight_layout() fig.savefig(out_path, dpi=150, bbox_inches="tight") plt.close(fig) print(f"Saved: {out_path}") # ═══════════════════════════════════════════════════════════════════════════════ def main(): for split_name, split_dir in OOD_SPLITS.items(): dfs = {} for model_name, model_dir in MODELS.items(): path = BASE_DIR / model_dir / split_dir / CSV_NAME if not path.exists(): print(f"[WARN] missing: {path}") continue df = load(path) print(f" {model_name} / {split_name}: {len(df)} no_input rows, " f"{df['distractor_name'].nunique()} distractors, " f"{df['scene'].nunique()} scenes") dfs[model_name] = df if not dfs: print(f"[SKIP] {split_name} — no data found") continue out = BASE_DIR / f"heatmap_ood_no_input_{split_name}.png" plot_split(split_name, dfs, out) # ── Print summary table ──────────────────────────────────────────────────── print(f"\n{'═'*70}") print(" OOD no_input summary: % KEPT per model per split (CLAP)") print(f"{'═'*70}") print(f" {'split':<25} {'combined_v1':>20} {'no_TSDL_old_mixtures':>22}") print(" " + "─" * 70) for split_name, split_dir in OOD_SPLITS.items(): row = f" {split_name:<25}" for model_name, model_dir in MODELS.items(): path = BASE_DIR / model_dir / split_dir / CSV_NAME if not path.exists(): row += f" {'N/A':>20}" continue df = load(path) if df.empty: row += f" {'N/A':>20}" else: kept_pct = (df["kept_score"] > 0).mean() * 100 row += f" {kept_pct:>19.1f}%" print(row) # ── Per-distractor breakdown (OOD_distractors only, combined_v1) ────────── path = BASE_DIR / "experiments_final/combined_v1" / OOD_SPLITS["OOD_distractors"] / CSV_NAME if path.exists(): df = load(path) if not df.empty: print(f"\n{'═'*70}") print(" Per-distractor breakdown (OOD_distractors, combined_v1)") print(" sorted by kept_score (green = model tends to keep, red = tends to remove)") print(f"{'═'*70}") print(f" {'distractor':<36} {'N':>4} {'kept_score':>10} {'behaviour'}") print(" " + "─" * 65) stats = ( df.groupby("distractor_name")["kept_score"] .agg(["mean", "count"]) .sort_values("mean", ascending=False) ) for name, row in stats.iterrows(): v = row["mean"] n = int(row["count"]) pct = abs(v) * 100 beh = f"KEPT ({pct:.0f}%)" if v > 0 else f"REMOVED ({pct:.0f}%)" bar = ("█" * int(pct / 5)).ljust(20) print(f" {name:<36} {n:>4} {v:>+10.3f} {beh} {bar}") if __name__ == "__main__": main()