| """Doğruluk ve kalibrasyon ölçüleri (yalnızca numpy).""" |
| import numpy as np |
|
|
|
|
| def ece(conf, correct, bins=15): |
| conf, correct = np.asarray(conf, float), np.asarray(correct, float) |
| edges = np.linspace(0, 1, bins + 1) |
| out = 0.0 |
| for lo, hi in zip(edges[:-1], edges[1:]): |
| m = (conf > lo) & (conf <= hi) if lo > 0 else (conf >= lo) & (conf <= hi) |
| if m.any(): |
| out += m.mean() * abs(conf[m].mean() - correct[m].mean()) |
| return float(out) |
|
|
|
|
| def brier(conf, correct): |
| conf, correct = np.asarray(conf, float), np.asarray(correct, float) |
| return float(np.mean((conf - correct) ** 2)) |
|
|
|
|
| def auroc(conf, correct): |
| """Güvenin doğru/yanlışı ayırma gücü (Mann-Whitney).""" |
| conf, correct = np.asarray(conf, float), np.asarray(correct, bool) |
| pos, neg = conf[correct], conf[~correct] |
| if len(pos) == 0 or len(neg) == 0: |
| return float("nan") |
| order = np.argsort(np.concatenate([pos, neg]), kind="mergesort") |
| ranks = np.empty(len(order)) |
| allv = np.concatenate([pos, neg])[order] |
| i = 0 |
| r = np.empty(len(order)) |
| while i < len(allv): |
| j = i |
| while j + 1 < len(allv) and allv[j + 1] == allv[i]: |
| j += 1 |
| r[i:j + 1] = (i + j) / 2 + 1 |
| i = j + 1 |
| ranks[order] = r |
| return float((ranks[:len(pos)].sum() - len(pos) * (len(pos) + 1) / 2) / (len(pos) * len(neg))) |
|
|
|
|
| def risk_coverage(conf, correct, coverages=(1.0, 0.8, 0.6, 0.4, 0.2)): |
| """En emin olduğu adımların %c'sinde doğruluk.""" |
| conf, correct = np.asarray(conf, float), np.asarray(correct, float) |
| order = np.argsort(-conf, kind="stable") |
| out = {} |
| for c in coverages: |
| k = max(1, int(round(c * len(order)))) |
| out[f"acc@{int(c * 100)}%"] = float(correct[order[:k]].mean()) |
| return out |
|
|
|
|
| def bootstrap_ci(correct, n=1000, seed=0): |
| correct = np.asarray(correct, float) |
| rng = np.random.default_rng(seed) |
| means = [correct[rng.integers(0, len(correct), len(correct))].mean() for _ in range(n)] |
| return float(np.percentile(means, 2.5)), float(np.percentile(means, 97.5)) |
|
|
|
|
| def summarize(conf, correct): |
| d = {"n": int(len(correct)), "acc": float(np.mean(correct)), |
| "acc_ci95": bootstrap_ci(correct), "ece": ece(conf, correct), |
| "brier": brier(conf, correct), "auroc": auroc(conf, correct)} |
| d.update(risk_coverage(conf, correct)) |
| return d |
|
|
