File size: 16,600 Bytes
c22b544 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 | #!/usr/bin/env python3
"""
OOD stem-based CLAP heatmaps — all 3 splits × 2 models.
Metric: CLAP cosine similarity between model output crop and isolated
distractor stem crop (from evaluate_event_detection.py).
High CLAP sim → distractor present in output → KEPT (green)
Low CLAP sim → distractor removed → REMOVED (red)
Only no_input command type (unambiguous: no user command given).
Outputs (one figure per OOD split, two subplots per figure):
heatmap_ood_stem_OOD_backgrounds.png
heatmap_ood_stem_OOD_distractors.png
heatmap_ood_stem_OOD_both.png
"""
import re
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
# ── Config ────────────────────────────────────────────────────────────────────
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_backgrounds": ("eval_outputs_OOD_backgrounds",
"Background scene (OOD)",
"Distractor class (in-distribution)"),
"OOD_distractors": ("eval_outputs_OOD_distractors",
"Background scene (in-distribution)",
"Distractor class (OOD)"),
"OOD_both": ("eval_outputs_OOD_both",
"Background scene (OOD)",
"Distractor class (OOD)"),
}
CMD_TYPE = "no_input"
CSV_NAME = "event_detection_scores.csv"
# ── Distractor ordering + display names (matches reference figure) ─────────────
# Order: left-to-right as in the reference heatmap image
DISTRACTOR_CANONICAL_ORDER = [
"computer_typing", "drill", "cricket", "sneeze", "cough",
"jackhammer", "hammer", "engine", "applause", "birds_chirping",
"fireworks", "helicopter", "dog", "car_horn", "footsteps",
"alarm_clock", "slam", "baby_cry", "train_horn", "cat",
"ringtone", "cellphone_buzz_vibrating_alert", "car_alarm",
"glass_breaking", "door_knock", "siren", "boom",
"doorbell", "fire_alarm", "gunshot",
]
DISPLAY_NAMES = {
"computer_typing": "Computer Typing",
"drill": "Drill",
"cricket": "Cricket",
"sneeze": "Sneeze",
"cough": "Cough",
"jackhammer": "Jackhammer",
"hammer": "Hammer",
"engine": "Engine",
"applause": "Applause",
"birds_chirping": "Birds Chirping",
"fireworks": "Fireworks",
"helicopter": "Helicopter",
"dog": "Dog",
"car_horn": "Car Horn",
"footsteps": "Footsteps",
"alarm_clock": "Alarm Clock",
"slam": "Slam",
"baby_cry": "Baby Cry",
"train_horn": "Train Horn",
"cat": "Cat",
"ringtone": "Ringtone",
"cellphone_buzz_vibrating_alert":"Smartphone Vibration",
"car_alarm": "Car Alarm",
"glass_breaking": "Glass Breaking",
"door_knock": "Door Knock",
"siren": "Siren",
"boom": "Boom",
"doorbell": "Doorbell",
"fire_alarm": "Fire Alarm",
"gunshot": "Gunshot",
}
# Scene display names (title-case with spaces)
SCENE_DISPLAY_NAMES = {
"airplane": "Airplane",
"bus_station": "Bus Station",
"gym": "Gym",
"harbour": "Harbour",
"library": "Library",
"museum": "Museum",
"office": "Office",
"park": "Park",
"shopping_mall": "Shopping Mall",
"train_station": "Train Station",
"airport": "Airport",
"beach": "Beach",
"bus": "Bus",
"cafe": "Cafe",
"coffee_shop": "Coffee Shop",
"city": "City",
"forest": "Forest",
"home": "Home",
"kitchen": "Kitchen",
"restaurant": "Restaurant",
"supermarket": "Supermarket",
"train": "Train",
}
def canonical_dist_order(present: list) -> list:
"""Return distractors in canonical order; extras appended at end."""
ordered = [d for d in DISTRACTOR_CANONICAL_ORDER if d in present]
extras = [d for d in present if d not in DISTRACTOR_CANONICAL_ORDER]
return ordered + extras
# ── Helpers ───────────────────────────────────────────────────────────────────
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["clap_sim"] = pd.to_numeric(df["clap_sim"], errors="coerce")
df["scene"] = df["mixture_id"].str.extract(r"^(.+?)_\d+dist_")
return df.dropna(subset=["clap_sim", "scene"])
def build_pivot(df: pd.DataFrame, scene_order, dist_order) -> pd.DataFrame:
return (
df.groupby(["scene", "distractor_name"])["clap_sim"]
.mean()
.unstack("distractor_name")
.reindex(index=scene_order, columns=dist_order)
)
def annot_array(pivot: pd.DataFrame) -> np.ndarray:
"""Single-line value annotation — colour already conveys K/R."""
arr = np.empty(pivot.shape, dtype=object)
for i in range(pivot.shape[0]):
for j in range(pivot.shape[1]):
v = pivot.iloc[i, j]
arr[i, j] = "" if np.isnan(v) else f"{v:.2f}"
return arr
def dist_xlabel(name, mean_val, threshold):
tag = "K" if mean_val >= threshold else "R"
disp = DISPLAY_NAMES.get(name, name)
return f"{disp}\n({tag} {mean_val:.2f})"
def scene_ylabel(name, mean_val, threshold):
tag = "K" if mean_val >= threshold else "R"
disp = SCENE_DISPLAY_NAMES.get(name, name.replace("_", " ").title())
return f"{disp} ({tag} {mean_val:.2f})"
# ── Main ──────────────────────────────────────────────────────────────────────
def plot_split(split_key: str, dist_order_override=None):
split_dir, ylabel, xlabel = OOD_SPLITS[split_key]
# Load both models
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_key}: {len(df)} rows, "
f"{df['scene'].nunique()} scenes, "
f"{df['distractor_name'].nunique()} distractors")
dfs[model_name] = df
if not dfs:
print(f" [SKIP] {split_key} — no data")
return
# Shared ordering + threshold (pooled across both models)
all_df = pd.concat(dfs.values(), ignore_index=True)
threshold = all_df["clap_sim"].median()
print(f" Threshold (pooled median): {threshold:.3f}")
# Fixed canonical distractor order; scenes sorted by mean CLAP sim
present_dists = all_df["distractor_name"].unique().tolist()
if dist_order_override is not None:
# Use shared override; keep only those actually present in this split
dist_order = [d for d in dist_order_override if d in present_dists]
# Append any extras not in the canonical override (safety net)
extras = [d for d in present_dists if d not in dist_order_override]
dist_order = dist_order + canonical_dist_order(extras)
else:
dist_order = canonical_dist_order(present_dists)
scene_order = (
all_df.groupby("scene")["clap_sim"]
.mean().sort_values(ascending=False).index.tolist()
)
n_dist = len(dist_order)
n_scene = len(scene_order)
ood_tag = {
"OOD_backgrounds": "OOD: background scenes unseen, distractors known",
"OOD_distractors": "OOD: distractor classes unseen, backgrounds known",
"OOD_both": "OOD: both background scenes and distractor classes unseen",
}[split_key]
# ── One figure per model ──────────────────────────────────────────────────
for model_name, df in dfs.items():
# Layout
cell_w = 0.62
cell_h = 0.28
margin_w = 3.5
hm_title = 0.8
bp_h = 2.8
xlbl_h = 2.5
suptitle_h = 0.9
fw = n_dist * cell_w + margin_w
heatmap_h = n_scene * cell_h + hm_title
fh = suptitle_h + heatmap_h + bp_h + xlbl_h
fig, axes = plt.subplots(
2, 1,
figsize=(fw, fh),
gridspec_kw={"height_ratios": [heatmap_h, bp_h + xlbl_h],
"hspace": 0.02},
squeeze=False,
)
ax_hm = axes[0][0]
ax_box = axes[1][0]
fig.suptitle(
f"{split_key} | {model_name} | no_input | stem-based CLAP similarity\n"
f"{ood_tag}\n"
f"green = KEPT (distractor present), red = REMOVED "
f"| K/R boundary = pooled median ({threshold:.3f})",
fontsize=11, fontweight="bold", y=1.01,
)
# ── Heatmap ───────────────────────────────────────────────────────────
pivot = build_pivot(df, scene_order, dist_order)
annots = annot_array(pivot)
sns.heatmap(
pivot, ax=ax_hm, mask=pivot.isna(),
annot=annots, fmt="",
annot_kws={"size": 7, "weight": "bold"},
vmin=0.20, vmax=0.70, center=threshold,
cmap="RdYlGn", linewidths=0.4, linecolor="#cccccc",
cbar_kws={
"label": "CLAP sim ← REMOVED | KEPT →",
"shrink": 0.7,
"ticks": [0.20, 0.30, threshold, 0.50, 0.60, 0.70],
},
)
scene_means = df.groupby("scene")["clap_sim"].mean()
ax_hm.set_xticklabels([])
ax_hm.set_xlabel("")
ax_hm.set_yticklabels(
[scene_ylabel(s, scene_means.get(s, np.nan), threshold)
for s in scene_order],
fontsize=9, rotation=0,
)
kept_pct = (df["clap_sim"] >= threshold).mean() * 100
ax_hm.set_title(f"KEPT {kept_pct:.0f}% overall",
fontsize=10, fontweight="bold", pad=6)
ax_hm.set_ylabel(ylabel, fontsize=9)
# ── Bar chart: mean kept rate (%) per distractor ──────────────────────
# Use only distractors present in this model — same column order as heatmap
model_dists = df["distractor_name"].unique()
model_dist_order = [d for d in dist_order if d in model_dists]
bp_df = df[["distractor_name", "clap_sim"]].copy()
bp_df["kept"] = (bp_df["clap_sim"] >= threshold).astype(float)
mean_kept = bp_df.groupby("distractor_name")["kept"].mean() * 100 # percentage
# Map each distractor to its column centre in dist_order
col_idx = {d: i for i, d in enumerate(dist_order)}
bar_positions = np.array([col_idx[d] + 0.5 for d in model_dist_order])
bar_values = [mean_kept.get(d, np.nan) for d in model_dist_order]
bar_colors = [
"#2ca02c" if (not np.isnan(v) and v >= 50) else "#d62728"
for v in bar_values
]
ax_box.bar(
bar_positions, bar_values,
width=0.6, color=bar_colors, alpha=0.85, zorder=3,
edgecolor="white", linewidth=0.4,
)
# Annotate each bar with its value
for pos, val, color in zip(bar_positions, bar_values, bar_colors):
if not np.isnan(val):
ax_box.text(
pos, val + 1.5, f"{val:.0f}%",
ha="center", va="bottom", fontsize=5.5,
fontweight="bold", color="black",
)
ax_box.axhline(50, color="gray", linestyle="--",
linewidth=1.0, alpha=0.8, label="50% boundary (K/R)")
# Both axes share [0, n_dist] so column i is at [i, i+1]
ax_hm.set_xlim(0, n_dist)
ax_box.set_xlim(0, n_dist)
ax_box.set_ylim(0, 115)
ax_box.set_yticks([0, 25, 50, 75, 100])
ax_box.set_yticklabels(["0%", "25%", "50%", "75%", "100%"], fontsize=8)
# Tick at each heatmap column centre
ax_box.set_xticks(np.arange(n_dist) + 0.5)
ax_box.set_xticklabels(
[DISPLAY_NAMES.get(d, d) for d in dist_order],
fontsize=7.5, rotation=45, ha="right",
)
ax_box.set_xlabel(xlabel, fontsize=10)
ax_box.set_ylabel("Mean Kept Rate (%)", fontsize=9)
ax_box.set_title("Mean kept rate per distractor across scenes "
"(green ≥ 50%, red < 50%)",
fontsize=10, fontweight="bold", pad=6)
ax_box.yaxis.grid(True, linestyle="--", linewidth=0.5, alpha=0.5, zorder=0)
ax_box.set_axisbelow(True)
ax_box.legend(fontsize=8, loc="upper left", framealpha=0.85)
# ── Force exact horizontal alignment: boxplot width = heatmap width ───
# tight_layout makes the colorbar steal space from ax_hm; ax_box ends up
# wider. After layout is resolved, snap ax_box to ax_hm's x-extent.
plt.tight_layout()
fig.canvas.draw()
pos_hm = ax_hm.get_position()
pos_box = ax_box.get_position()
ax_box.set_position([pos_hm.x0, pos_box.y0,
pos_hm.width, pos_box.height])
out = BASE_DIR / f"heatmap_ood_stem_{split_key}_{model_name}.png"
fig.savefig(out, dpi=150, bbox_inches="tight")
plt.close(fig)
print(f" Saved: {out}")
def print_summary():
print(f"\n{'═'*72}")
print(" OOD no_input summary — % KEPT (CLAP sim ≥ pooled median)")
print(f"{'═'*72}")
print(f" {'split':<20} {'combined_v1':>20} {'no_TSDL_old_mixtures':>22}")
print(" " + "─" * 65)
for split_key, (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 path.exists():
dfs[model_name] = load(path)
if not dfs:
continue
threshold = pd.concat(dfs.values())["clap_sim"].median()
row = f" {split_key:<20}"
for model_name in MODELS:
if model_name in dfs:
pct = (dfs[model_name]["clap_sim"] >= threshold).mean() * 100
row += f" {pct:>19.1f}%"
else:
row += f" {'N/A':>20}"
print(row)
if __name__ == "__main__":
# Pre-compute a shared distractor ordering for OOD_distractors + OOD_both
# so that the same distractor always appears in the same column in both figures.
SHARED_SPLITS = {"OOD_distractors", "OOD_both"}
shared_all_dists = set()
for sk in SHARED_SPLITS:
split_dir = OOD_SPLITS[sk][0]
for model_dir in MODELS.values():
path = BASE_DIR / model_dir / split_dir / CSV_NAME
if path.exists():
df_tmp = load(path)
shared_all_dists.update(df_tmp["distractor_name"].unique())
shared_dist_order = canonical_dist_order(list(shared_all_dists))
print(f"\nShared distractor order (OOD_distractors + OOD_both): "
f"{len(shared_dist_order)} classes")
for split_key in OOD_SPLITS:
print(f"\n{'─'*60}")
print(f" Plotting {split_key} ...")
override = shared_dist_order if split_key in SHARED_SPLITS else None
plot_split(split_key, dist_order_override=override)
print_summary()
|