Datasets:
Formats:
parquet
Size:
1M - 10M
Tags:
gaussian-splatting
fault-tolerance
single-event-upset
reliability
radiance-fields
computer-graphics
License:
| """E14: predicting criticality without fault injection (runs on CPU, e.g. M4). | |
| Using only static features of a fault site (which field, which bit, its class, | |
| and the stored value), we train a classifier to predict whether the upset is | |
| catastrophic, without rendering. A high area under the ROC curve means the | |
| critical bits can be identified from parameter statistics alone, which turns the | |
| expensive injection campaign into a cheap static screen for future models. | |
| """ | |
| import argparse | |
| import glob | |
| import os | |
| import numpy as np | |
| import matplotlib | |
| matplotlib.use("Agg") | |
| import matplotlib.pyplot as plt | |
| from sklearn.ensemble import RandomForestClassifier | |
| from sklearn.linear_model import LogisticRegression | |
| from sklearn.model_selection import cross_val_score, train_test_split | |
| from sklearn.metrics import roc_auc_score, roc_curve | |
| from sklearn.preprocessing import StandardScaler | |
| FIELDS = ["means", "scales", "quats", "opacities", "sh0", "shN"] | |
| def load(root): | |
| X, y = [], [] | |
| for s in sorted(glob.glob(os.path.join(root, "campaign", "shard_*_fp32.npz"))): | |
| if s.endswith("_guard.npz"): | |
| continue | |
| d = np.load(s, allow_pickle=True); a = d["data"]; cols = list(d["cols"]); ci = {c: i for i, c in enumerate(cols)} | |
| field = a[:, ci["field_id"]]; bit = a[:, ci["bit"]]; bc = a[:, ci["bitclass"]] | |
| cv = a[:, ci["clean_val"]]; fr = a[:, ci["fracchg"]]; cat = a[:, ci["cat"]] | |
| fonehot = np.eye(6)[field.astype(int)] | |
| bconehot = np.eye(3)[bc.astype(int)] | |
| absv = np.abs(cv); logabs = np.log10(absv + 1e-12); sgn = np.sign(cv) | |
| feat = np.column_stack([fonehot, bconehot, bit, absv, logabs, sgn]) | |
| lab = ((cat > 0.5) | (fr > 0.01)).astype(int) | |
| X.append(feat); y.append(lab) | |
| return np.concatenate(X), np.concatenate(y) | |
| FEATNAMES = (["field=" + f for f in FIELDS] + ["sign", "exp", "mantissa", "bit", "absval", "log|val|", "signval"]) | |
| def main(): | |
| ap = argparse.ArgumentParser() | |
| ap.add_argument("--root", default="data_local") | |
| ap.add_argument("--out", default="../generated") | |
| args = ap.parse_args() | |
| os.makedirs(args.out, exist_ok=True) | |
| X, y = load(args.root) | |
| print(f"samples={len(y)} positives={y.sum()} ({100*y.mean():.2f}%)") | |
| macros = {} | |
| Xtr, Xte, ytr, yte = train_test_split(X, y, test_size=0.3, random_state=0, stratify=y) | |
| rf = RandomForestClassifier(n_estimators=200, max_depth=10, n_jobs=-1, random_state=0) | |
| rf.fit(Xtr, ytr) | |
| p = rf.predict_proba(Xte)[:, 1] | |
| auc_rf = roc_auc_score(yte, p) | |
| sc = StandardScaler().fit(Xtr) | |
| lr = LogisticRegression(max_iter=2000) | |
| lr.fit(sc.transform(Xtr), ytr) | |
| auc_lr = roc_auc_score(yte, lr.predict_proba(sc.transform(Xte))[:, 1]) | |
| # field+bit-only model (no value information at all) | |
| fb_cols = list(range(0, 9)) + [9] # field one-hot + bitclass + bit | |
| rf2 = RandomForestClassifier(n_estimators=200, max_depth=8, n_jobs=-1, random_state=0) | |
| rf2.fit(Xtr[:, fb_cols], ytr) | |
| auc_fb = roc_auc_score(yte, rf2.predict_proba(Xte[:, fb_cols])[:, 1]) | |
| imp = rf.feature_importances_ | |
| top = FEATNAMES[int(np.argmax(imp))] | |
| macros["predAUC"] = f"{auc_rf:.3f}" | |
| macros["predAUClr"] = f"{auc_lr:.3f}" | |
| macros["predAUCfieldbit"] = f"{auc_fb:.3f}" | |
| macros["predTopFeat"] = top.replace("=", " ").replace("_", " ") | |
| fpr, tpr, _ = roc_curve(yte, p) | |
| plt.figure(figsize=(5.4, 4.2)) | |
| plt.plot(fpr, tpr, label=f"random forest (AUC {auc_rf:.3f})") | |
| fpr2, tpr2, _ = roc_curve(yte, rf2.predict_proba(Xte[:, fb_cols])[:, 1]) | |
| plt.plot(fpr2, tpr2, "--", label=f"field+bit only (AUC {auc_fb:.3f})") | |
| plt.plot([0, 1], [0, 1], ":", color="gray") | |
| plt.xlabel("false positive rate"); plt.ylabel("true positive rate") | |
| plt.legend(loc="lower right"); plt.grid(alpha=0.3) | |
| plt.savefig(os.path.join(args.out, "fig_predictor.pdf"), bbox_inches="tight"); plt.close() | |
| with open(os.path.join(args.out, "predictor_numbers.tex"), "w") as f: | |
| defaults = {"predAUC": "0.0", "predAUClr": "0.0", "predAUCfieldbit": "0.0", "predTopFeat": "n/a"} | |
| for k, v in defaults.items(): | |
| macros.setdefault(k, v) | |
| for k, v in macros.items(): | |
| f.write(f"\\newcommand{{\\{k}}}{{{v}}}\n") | |
| print("PREDICTOR macros:", macros) | |
| if __name__ == "__main__": | |
| main() | |