Spaces:
Sleeping
Sleeping
| """Split evaluation shared by Phase 2+ (same metric definitions as Phase 1): | |
| pooled overall metrics, per class, per true-height band, per city, and per tile.""" | |
| from __future__ import annotations | |
| import numpy as np | |
| import pandas as pd | |
| from .metrics import HeightMetrics, _Acc, label_stats, pixel_stats | |
| HEIGHT_EDGES = [0, 2, 5, 10, 20, 1000] | |
| BIN_NAMES = {i: (f"{HEIGHT_EDGES[i]}-{HEIGHT_EDGES[i + 1]}m" if HEIGHT_EDGES[i + 1] < 1000 | |
| else f">{HEIGHT_EDGES[i]}m") for i in range(len(HEIGHT_EDGES) - 1)} | |
| class SplitEvaluator: | |
| def __init__(self): | |
| self.cls = HeightMetrics() | |
| self.bin = HeightMetrics(label_names=BIN_NAMES) | |
| self.city: dict[str, HeightMetrics] = {} | |
| self.rows = [] | |
| def add(self, tid: str, pred: np.ndarray, gt: np.ndarray, valid: np.ndarray, cls: np.ndarray) -> dict: | |
| nb = len(HEIGHT_EDGES) - 1 | |
| hb = np.digitize(gt, HEIGHT_EDGES[1:-1]) | |
| joint = cls[valid].astype(np.int64) * nb + hb[valid] | |
| T = label_stats(pixel_stats(pred[valid], gt[valid]), joint, 7 * nb).reshape(7, nb, 8) | |
| self.cls.add_table(T.sum(1)) | |
| self.bin.add_table(T.sum(0)) | |
| tot = T.sum((0, 1)) | |
| city = tid.split("_")[0] | |
| self.city.setdefault(city, HeightMetrics()).overall.add_vec(tot) | |
| a = _Acc(); a.add_vec(tot); tm = a.result() | |
| row = {"id": tid, "city": city, "valid_frac": float(valid.mean()), "n_valid": int(valid.sum()), | |
| "rmse": tm["rmse"], "mae": tm["mae"], "r": tm["pearson_r"], "bias": tm["bias"]} | |
| self.rows.append(row) | |
| return row | |
| def result(self) -> dict: | |
| res = self.cls.result() | |
| res["per_height_bin"] = self.bin.result()["per_class"] | |
| res["per_city"] = {c: h.result()["overall"] for c, h in sorted(self.city.items())} | |
| return res | |
| def per_tile(self) -> pd.DataFrame: | |
| return pd.DataFrame(self.rows) | |