File size: 16,109 Bytes
d388d33 | 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 | """Paired S3-vs-S2 statistics for each model under runs/, at TWO levels:
* per_image (Method 1) -- pool every test image across all phases, pair
S3 vs S2 by (phase, sample_id).
* phase_level (Method 2) -- one paired value per phase = the phase-mean of
the metric; pair S3-mean vs S2-mean across phases.
Same paired machinery as ablation_stats.py: paired t-test, Wilcoxon signed-rank
(zero_method="wilcox"), sign-flip permutation, bootstrap CI for the mean delta,
Hodges-Lehmann estimate + CI, Cohen's dz, rank-biserial r, Shapiro-Wilk on the
diffs, Holm correction. HD95 is the only lower-is-better metric.
Auto-discovers every model folder directly under runs/ (each distinct MODEL_NAME).
One .xlsx per model + a combined cross-model summary.
Usage:
python model_stats_s2_vs_s3.py # scans ./runs
python model_stats_s2_vs_s3.py --runs-root X
"""
from __future__ import annotations
import argparse
import json
import pathlib
import re
from collections import defaultdict
import numpy as np
import pandas as pd
from scipy import stats
# metric -> higher_is_better
METRICS = {
"biou_contour": True, # manuscript's reported "BIoU" (1-px contour)
"biou": True, # Cheng et al. CVPR 2021 d-pixel band Boundary IoU
"dice": True,
"iou": True,
"ppv": True,
"sen": True,
"hd95": False, # lower is better
}
PRIMARY = "biou_contour"
ALPHA = 0.05
N_BOOT = 20000
N_BOOT_HL = 4000
N_PERM = 20000
SEED = 20260709
# S3 is the first arm, S2 the second: Mean_delta = S3 - S2, positive favours S3
LABEL = {2: "S2 (baseline)", 3: "S3 (refinement)"}
# ---------------------------------------------------------------- stats helpers
def holm(pvals):
pvals = np.asarray(pvals, dtype=float)
order = np.argsort(pvals)
m = len(pvals)
adj = np.empty(m)
running = 0.0
for rank, idx in enumerate(order):
running = max(running, (m - rank) * pvals[idx])
adj[idx] = min(running, 1.0)
return adj
def boot_ci_mean(d, rng):
idx = rng.integers(0, len(d), size=(N_BOOT, len(d)))
return np.percentile(d[idx].mean(axis=1), [2.5, 97.5])
def hodges_lehmann(d):
i, j = np.triu_indices(len(d), k=0)
return float(np.median((d[i] + d[j]) / 2.0))
def boot_ci_hl(d, rng):
idx = rng.integers(0, len(d), size=(N_BOOT_HL, len(d)))
return np.percentile([hodges_lehmann(d[r]) for r in idx], [2.5, 97.5])
def sign_flip_perm_p(d, rng):
obs = abs(d.mean())
signs = rng.choice([-1.0, 1.0], size=(N_PERM, len(d)))
null = np.abs((signs * d).mean(axis=1))
return (np.sum(null >= obs - 1e-15) + 1) / (N_PERM + 1)
def rank_biserial(d):
nz = d[d != 0]
if len(nz) == 0:
return 0.0
r = stats.rankdata(np.abs(nz))
rp, rm = r[nz > 0].sum(), r[nz < 0].sum()
return float((rp - rm) / (rp + rm))
def analyse(x, y, metric, higher_better, level, rng):
"""x = S3 vector, y = S2 vector (paired). Returns (rows, delta)."""
d = np.asarray(x, float) - np.asarray(y, float)
n = len(d)
mean_d, sd_d = float(d.mean()), float(d.std(ddof=1)) if n > 1 else float("nan")
se = sd_d / np.sqrt(n) if n > 1 else float("nan")
if n >= 2:
t_stat, p_t = stats.ttest_rel(x, y)
crit = stats.t.ppf(1 - ALPHA / 2, df=n - 1)
t_ci = (mean_d - crit * se, mean_d + crit * se)
else:
t_stat, p_t, t_ci = np.nan, 1.0, (np.nan, np.nan)
if n >= 1 and not np.allclose(d, 0):
try:
w_stat, p_w = stats.wilcoxon(x, y, zero_method="wilcox", alternative="two-sided")
except ValueError:
w_stat, p_w = np.nan, 1.0
else:
w_stat, p_w = np.nan, 1.0
p_perm = sign_flip_perm_p(d, rng) if n >= 1 else 1.0
b_lo, b_hi = boot_ci_mean(d, rng) if n >= 2 else (np.nan, np.nan)
hl = hodges_lehmann(d) if n >= 1 else np.nan
hl_lo, hl_hi = boot_ci_hl(d, rng) if n >= 2 else (np.nan, np.nan)
shapiro_p = float(stats.shapiro(d).pvalue) if (n >= 3 and not np.allclose(d, 0)) else np.nan
p_holm_c = holm([p_t, p_w, p_perm])
better = (mean_d > 0) == higher_better
winner = LABEL[3] if better else LABEL[2]
rows = []
for name, stat, p_raw, p_adj in [
("Paired t-test", float(t_stat), float(p_t), p_holm_c[0]),
("Wilcoxon signed-rank", float(w_stat), float(p_w), p_holm_c[1]),
("Sign-flip permutation", mean_d, float(p_perm), p_holm_c[2]),
]:
rows.append({
"Level": level, "Metric": metric, "Higher_is_better": higher_better,
"Contrast": "S3 - S2", "Test": name, "n_pairs": n,
"n_zero_diff": int(np.sum(d == 0)), "Statistic": stat,
"Mean_delta": mean_d, "SD_delta": sd_d, "p_raw": p_raw,
"p_Holm_within_metric_level": p_adj,
"Sig_within_metric_level": "Yes" if p_adj < ALPHA else "No",
"Better_arm": winner if p_adj < ALPHA else "n.s.",
"t_CI_low": t_ci[0], "t_CI_high": t_ci[1],
"boot_CI_low": b_lo, "boot_CI_high": b_hi,
"CI_excludes_zero": "Yes" if (b_lo > 0) or (b_hi < 0) else "No",
"HodgesLehmann_delta": hl, "HL_CI_low": hl_lo, "HL_CI_high": hl_hi,
"Cohens_dz": mean_d / sd_d if (sd_d and sd_d > 0) else np.nan,
"Rank_biserial_r": rank_biserial(d),
"Shapiro_p_on_diffs": shapiro_p,
"Diffs_normal_at_0.05": "n/a" if np.isnan(shapiro_p) else ("No" if shapiro_p < ALPHA else "Yes"),
})
return rows, d
# ---------------------------------------------------------------- discovery
def read_threshold(final_dir: pathlib.Path):
rc = final_dir / "run_config.json"
if rc.exists():
try:
return json.loads(rc.read_text()).get("threshold")
except Exception:
return None
return None
def discover(runs_root: pathlib.Path):
"""model -> phase -> strategy -> {'per_sample': {sid: row}, 'threshold': float}"""
data: dict[str, dict[int, dict[int, dict]]] = defaultdict(lambda: defaultdict(dict))
for ev in runs_root.glob("*/**/strategy_*/final/evaluation.json"):
final_dir = ev.parent
ms = re.search(r"strategy_(\d+)", final_dir.parent.name)
if not ms:
continue
s = int(ms.group(1))
if s not in (2, 3):
continue
try:
model = ev.relative_to(runs_root).parts[0]
except ValueError:
continue
pm = re.search(r"phase_(\d+)", str(ev))
phase = int(pm.group(1)) if pm else 1
try:
payload = json.loads(ev.read_text())
except Exception:
continue
data[model][phase][s] = {
"per_sample": {row["sample_id"]: row for row in payload.get("per_sample", [])},
"threshold": read_threshold(final_dir),
}
return data
# ---------------------------------------------------------------- per model
def analyse_model(model, phase_map, rng):
phases = sorted(p for p in phase_map if 2 in phase_map[p] and 3 in phase_map[p])
if not phases:
return None
# per-image pooled arrays + per-phase means
img = {m: {2: [], 3: []} for m in METRICS}
phase_mean = {m: {2: [], 3: []} for m in METRICS}
phase_used, n_img_per_phase, split_warnings = [], [], []
thr = {2: None, 3: None}
for p in phases:
s2, s3 = phase_map[p][2], phase_map[p][3]
thr[2] = thr[2] or s2.get("threshold")
thr[3] = thr[3] or s3.get("threshold")
set2, set3 = set(s2["per_sample"]), set(s3["per_sample"])
if set2 != set3:
split_warnings.append(f"phase {p}: S2/S3 sample_id sets differ "
f"(|S2|={len(set2)}, |S3|={len(set3)}, common={len(set2 & set3)})")
ids = sorted(set2 & set3)
if not ids:
continue
phase_used.append(p)
n_img_per_phase.append(len(ids))
for m in METRICS:
v2 = np.array([s2["per_sample"][i][m] for i in ids], float)
v3 = np.array([s3["per_sample"][i][m] for i in ids], float)
img[m][2].append(v2)
img[m][3].append(v3)
phase_mean[m][2].append(float(v2.mean()))
phase_mean[m][3].append(float(v3.mean()))
test_rows, desc_rows = [], []
per_image = {"phase": np.concatenate([[p] * n for p, n in zip(phase_used, n_img_per_phase)]).astype(int)}
per_phase = {"phase": np.array(phase_used, int)}
for m, hib in METRICS.items():
x_img = np.concatenate(img[m][3]) if img[m][3] else np.array([])
y_img = np.concatenate(img[m][2]) if img[m][2] else np.array([])
pm2 = np.array(phase_mean[m][2], float)
pm3 = np.array(phase_mean[m][3], float)
# descriptives, both levels
for s, arr_img, arr_ph in [(2, y_img, pm2), (3, x_img, pm3)]:
desc_rows.append({
"Metric": m, "Higher_is_better": hib, "Arm": LABEL[s],
"n_images": len(arr_img), "img_Mean": arr_img.mean() if len(arr_img) else np.nan,
"img_SD": arr_img.std(ddof=1) if len(arr_img) > 1 else np.nan,
"img_Median": np.median(arr_img) if len(arr_img) else np.nan,
"n_phases": len(arr_ph), "phase_Mean": arr_ph.mean() if len(arr_ph) else np.nan,
"phase_SD": arr_ph.std(ddof=1) if len(arr_ph) > 1 else np.nan,
})
# per_image tests
rows, d_img = analyse(x_img, y_img, m, hib, "per_image", rng)
test_rows.extend(rows)
per_image[f"{m}__S2"] = y_img
per_image[f"{m}__S3"] = x_img
per_image[f"{m}__delta_S3_minus_S2"] = d_img
# phase_level tests
rows_ph, d_ph = analyse(pm3, pm2, m, hib, "phase_level", rng)
test_rows.extend(rows_ph)
per_phase[f"{m}__S2_phase_mean"] = pm2
per_phase[f"{m}__S3_phase_mean"] = pm3
per_phase[f"{m}__delta_S3_minus_S2"] = d_ph
tests_df = pd.DataFrame(test_rows)
# global Holm across this model's whole workbook (metrics x 2 levels x 3 tests)
tests_df["p_Holm_global"] = holm(tests_df["p_raw"].values)
tests_df["Sig_global"] = np.where(tests_df["p_Holm_global"] < ALPHA, "Yes", "No")
return {
"phases": phase_used,
"thr": thr,
"split_warnings": split_warnings,
"tests": tests_df,
"descriptives": pd.DataFrame(desc_rows),
"per_image": pd.DataFrame(per_image),
"per_phase": pd.DataFrame(per_phase),
}
def readme_frame(model, res):
return pd.DataFrame({"Field": [
"Model", "Contrast", "Phases used", "n phases", "S2 threshold", "S3 threshold",
"per_image level", "phase_level level", "Metrics", "PRIMARY", "hd95 direction",
"Wilcoxon zero handling", "Bootstrap", "Permutation", "Holm scope",
"Split check", "Split warnings", "Seed",
], "Value": [
model, "S3 (refinement) - S2 (baseline); positive delta favours S3",
", ".join(map(str, res["phases"])), len(res["phases"]),
str(res["thr"][2]), str(res["thr"][3]),
"pool all test images across phases, paired by (phase, sample_id) [manuscript Method 1]",
"one paired value per phase = phase-mean of the metric, paired across phases [manuscript Method 2]",
", ".join(METRICS), PRIMARY, "LOWER is better; direction handled in Better_arm",
"zero_method='wilcox' (zero diffs dropped)",
f"{N_BOOT} resamples for mean-delta CI; {N_BOOT_HL} for Hodges-Lehmann",
f"{N_PERM} sign flips; p floor ~= 1/{N_PERM + 1}",
"Holm across every test in this workbook (metrics x 2 levels x 3 tests) = p_Holm_global",
"asserted S2 and S3 test sample_id sets identical within each phase",
"; ".join(res["split_warnings"]) if res["split_warnings"] else "none",
f"numpy default_rng({SEED})",
]})
def write_workbook(model, res, out_dir):
out = out_dir / f"{model}__s2_vs_s3_stats.xlsx"
try:
with pd.ExcelWriter(out, engine="openpyxl") as xl:
readme_frame(model, res).to_excel(xl, sheet_name="README", index=False)
res["descriptives"].to_excel(xl, sheet_name="Descriptives", index=False)
res["tests"].to_excel(xl, sheet_name="Tests", index=False)
res["per_phase"].to_excel(xl, sheet_name="PerPhase", index=False)
res["per_image"].to_excel(xl, sheet_name="PerImage", index=False)
for sh in xl.book.worksheets:
for col in sh.columns:
w = max((len(str(c.value)) if c.value is not None else 0) for c in col)
sh.column_dimensions[col[0].column_letter].width = min(max(w + 2, 12), 62)
sh.freeze_panes = "A2"
return out
except Exception as exc: # openpyxl missing -> CSV fallback
print(f"[stats] xlsx failed for {model} ({exc}); writing CSVs instead")
res["tests"].to_csv(out_dir / f"{model}__tests.csv", index=False)
res["descriptives"].to_csv(out_dir / f"{model}__descriptives.csv", index=False)
res["per_phase"].to_csv(out_dir / f"{model}__per_phase.csv", index=False)
return None
def main() -> int:
ap = argparse.ArgumentParser()
ap.add_argument("--runs-root", default="runs")
args = ap.parse_args()
runs_root = pathlib.Path(args.runs_root).resolve()
if not runs_root.is_dir():
print(f"[stats] runs root not found: {runs_root}")
return 1
data = discover(runs_root)
if not data:
print(f"[stats] no evaluation.json found under {runs_root}")
return 1
out_dir = runs_root / "_stats"
out_dir.mkdir(parents=True, exist_ok=True)
rng = np.random.default_rng(SEED)
summary_rows = []
for model in sorted(data):
res = analyse_model(model, data[model], rng)
if res is None:
print(f"[stats] {model}: no phase has both S2 and S3 -- skipped")
continue
wb = write_workbook(model, res, out_dir)
print(f"[stats] {model:28s} phases={len(res['phases'])} -> {wb.name if wb else '(csv)'}"
+ (" [SPLIT WARNINGS]" if res["split_warnings"] else ""))
# headline: Wilcoxon row per metric per level for the cross-model summary
t = res["tests"]
for m in METRICS:
for level in ("per_image", "phase_level"):
r = t[(t.Metric == m) & (t.Level == level) & (t.Test == "Wilcoxon signed-rank")]
if r.empty:
continue
r = r.iloc[0]
summary_rows.append({
"Model": model, "Metric": m, "Level": level, "n_pairs": int(r.n_pairs),
"S3_minus_S2": r.Mean_delta, "Wilcoxon_p_raw": r.p_raw,
"p_Holm_global": r.p_Holm_global, "Sig_global": r.Sig_global,
"Better_arm": r.Better_arm, "Cohens_dz": r.Cohens_dz,
"boot_CI_low": r.boot_CI_low, "boot_CI_high": r.boot_CI_high,
})
summary_df = pd.DataFrame(summary_rows)
summary_df.to_csv(out_dir / "ALL_MODELS_summary.csv", index=False)
try:
with pd.ExcelWriter(out_dir / "ALL_MODELS_summary.xlsx", engine="openpyxl") as xl:
summary_df.to_excel(xl, sheet_name="Summary", index=False)
for sh in xl.book.worksheets:
for col in sh.columns:
w = max((len(str(c.value)) if c.value is not None else 0) for c in col)
sh.column_dimensions[col[0].column_letter].width = min(max(w + 2, 12), 40)
sh.freeze_panes = "A2"
except Exception:
pass
print(f"\n[stats] wrote per-model workbooks + ALL_MODELS_summary to {out_dir}")
if not summary_df.empty:
pd.set_option("display.width", 240, "display.max_columns", 40)
prim = summary_df[summary_df.Metric == PRIMARY]
print(f"\n=== PRIMARY ({PRIMARY}) S3-vs-S2, both levels ===")
print(prim.to_string(index=False, float_format=lambda v: f"{v:.4g}"))
return 0
if __name__ == "__main__":
raise SystemExit(main())
|