Download sunstar_plus/model.py from summerAIO/sunstar-plus: direct link, hf CLI and curl.
- Browser
- Download file 6.45 kB
-
https://huggingface.co/summerAIO/sunstar-plus/resolve/main/sunstar_plus/model.py
- Command line
-
hf download hf://summerAIO/sunstar-plus/sunstar_plus/model.py
-
curl -L -o model.py https://huggingface.co/summerAIO/sunstar-plus/resolve/main/sunstar_plus/model.py
6.45 kB
| """Sun*+ 推論ラッパ。 | |
| SunstarPlus.downscale(...) : 観測なしダウンスケーリング (リード0, ERA5/再解析/NWP を地点へ) | |
| SunstarPlus.forecast(...) : 観測同化 (時刻 t0 までの AMeDAS 観測 + 近傍局, リード 1/3/6/12/24h) | |
| 入力 fields: ERA5 互換の 0.25° 格子場 dict {変数名: (T,H,W)}。単位 = tp [mm/h], 気温/露点/skin [°C], 気圧 [hPa], 日射 [J/m2/h] (ERA5 の生単位)。 | |
| 前後 >=24 h (観測特徴は >=200 h) の余白時刻を含めること。""" | |
| import gzip, json, pickle | |
| from pathlib import Path | |
| import numpy as np, pandas as pd, lightgbm as lgb | |
| from .features import * | |
| from .obsfeats import obsfeats, neighbor_table, nb_mean, NB_NAMES, DB_KEYS | |
| VARS = ["temp_c", "wind_u_ms", "wind_v_ms", "humidity_pct", "wind_speed_ms", "precip_mm"] | |
| LABEL = {"temp_c": "気温 [°C]", "wind_u_ms": "風 u [m/s]", "wind_v_ms": "風 v [m/s]", "humidity_pct": "相対湿度 [%]", "wind_speed_ms": "風速 [m/s]", "precip_mm": "降水量 [mm/h]"} | |
| def _post(var, x): | |
| if var == "humidity_pct": return np.clip(x, 0, 100) | |
| if var in ("precip_mm", "wind_speed_ms"): return np.maximum(x, 0) | |
| return x | |
| class SunstarPlus: | |
| def __init__(self, root="."): | |
| self.root = Path(root); mdir = self.root / "models" | |
| self.cfg = json.load(open(mdir / "config.json")) | |
| self.models = {t: {v: [self._load(mdir / f) for f in self.cfg["models"][t][v]] for v in VARS} for t in "AB"} | |
| self.sigma = {} | |
| for t in "AB": | |
| for v, d in self.cfg.get("sigma", {}).get(t, {}).items(): | |
| self.sigma[(t, v)] = dict(gbm=self._load(mdir / d["file"]), k=d["k"], ratio={int(k): x for k, x in d["ratio"].items()}) | |
| def _load(p): | |
| p = Path(p) | |
| return lgb.Booster(model_str=gzip.open(p, "rt").read()) if p.suffix == ".gz" else lgb.Booster(model_file=str(p)) | |
| # ------------------------------------------------------------------ helpers | |
| def _ens(self, models, X): return np.mean([m.predict(X) for m in models], 0) | |
| def _predict(self, task, X, base, lead, sigma, heldout): | |
| out = {} | |
| for v_i, var in enumerate(VARS): | |
| r = self._ens(self.models[task][var], X); mean = _post(var, base[:, v_i] + r) | |
| sg = None | |
| if sigma and (task, var) in self.sigma: | |
| d = self.sigma[(task, var)]; raw = np.maximum(d["gbm"].predict(np.concatenate([X, r[:, None]], 1)), 0.05) * np.sqrt(np.pi / 2) * d["k"] | |
| if heldout: raw = raw * d["ratio"].get(lead, 1.0) | |
| sg = np.maximum(raw, 0.05) | |
| out[var] = (mean, sg) | |
| return out | |
| # ------------------------------------------------------------------ public | |
| def downscale(self, times, fields, grid, stations, sigma=True, heldout=True): | |
| """times: DatetimeIndex (hourly UTC, 前後24h余白込み); fields: ERA5互換; grid: Grid; stations: DataFrame (stations_nationwide.csv 形式) | |
| returns {var: (mean (T,S), sigma (T,S) or None)} (余白 24h は NaN になりやすいので使わないこと) | |
| heldout=True: 学習に使っていない地点向けに σ を拡大 (既知地点では False)""" | |
| ss = StaticSet(stations, grid); F, names, base, n_v2 = build_features(times, fields, ss) | |
| T, S, nf = F.shape; flag = (np.asarray(times) >= np.datetime64("2024-10-01")).astype(np.float32) | |
| X = np.concatenate([F.reshape(T * S, nf), np.repeat(flag, S)[:, None]], 1) | |
| B = np.stack([base["t2m"], base["u10"], base["v10"], base["rh"], base["spd10"], np.zeros_like(base["tp"])], 2).reshape(T * S, 6) | |
| res = self._predict("A", X, B, 0, sigma, heldout) | |
| return {v: (m.reshape(T, S), None if s is None else s.reshape(T, S)) for v, (m, s) in res.items()} | |
| def forecast(self, times, fields, grid, stations, obs, lead, neighbors=True, sigma=True, heldout=True): | |
| """obs: (T,S,7) 時刻 times に揃えた AMeDAS 実測 [気温,u,v,相対湿度,降水,風速,気圧(NaN可)] (欠測は NaN)。 | |
| 初期時刻 t0=times[i] の観測から times[i+lead] を予測。returns {var: (mean (T,S) 有効時刻 index, sigma)}。 t0 以降の観測は使わない。 | |
| neighbors=True: 近傍局 (obs が有る他地点) の残差特徴を使う。stations に列 'has_obs' (bool) があればそれを近傍プールとする。""" | |
| ss = StaticSet(stations, grid); F, names, base, n_v2 = build_features(times, fields, ss) | |
| T, S, nf = F.shape; assert lead in (1, 3, 6, 12, 24) | |
| bb = dict(t2m=base["t2m"], d2m=base["d2m"], u10=base["u10"], v10=base["v10"], rh=base["rh"], spd10=base["spd10"]) | |
| G, gnames, res = obsfeats(np.asarray(obs, np.float32), bb) | |
| dbase = np.stack([bb[k] for k in DB_KEYS], 2) | |
| pool = stations["has_obs"].values.astype(bool) if "has_obs" in stations else np.isfinite(obs[:, :, 0]).any(0) | |
| if neighbors: | |
| idx, w = neighbor_table(stations, pool) | |
| pk = obs[:, :, 4]; rr = [res["T"], res["U"], res["V"], res["RH"], res["SPD"]] | |
| nbf = []; wsum = None | |
| for r_ in rr: | |
| m0, den = nb_mean(r_, idx, w); nbf.append(m0); wsum = den if wsum is None else wsum | |
| nb6 = [nb_mean(roll_mean(r_, -5, 0), idx, w)[0] for r_ in rr] | |
| P0, _ = nb_mean(pk, idx, w); P3, _ = nb_mean(roll_mean(pk, -2, 0), idx, w) | |
| NB = np.stack(nbf + nb6 + [P0, P3, P3, wsum], 2) | |
| else: NB = np.full((T, S, len(NB_NAMES)), np.nan, np.float32) | |
| ti = np.arange(lead, T); t0 = ti - lead | |
| flag = (np.asarray(times)[ti] >= np.datetime64("2024-10-01")).astype(np.float32) | |
| parts = [F[ti].reshape(-1, nf), np.repeat(flag, S)[:, None], G[t0].reshape(len(ti) * S, -1), | |
| (dbase[ti] - dbase[t0]).reshape(len(ti) * S, -1), np.full((len(ti) * S, 1), lead, np.float32), NB[t0].reshape(len(ti) * S, -1)] | |
| X = np.concatenate(parts, 1) | |
| B = np.stack([base["t2m"][ti], base["u10"][ti], base["v10"][ti], base["rh"][ti], base["spd10"][ti], np.zeros_like(base["tp"][ti])], 2).reshape(-1, 6) | |
| res_ = self._predict("B", X, B, lead, sigma, heldout) | |
| out = {} | |
| for v, (m, s) in res_.items(): | |
| full = np.full((T, S), np.nan, np.float32); full[ti] = m.reshape(len(ti), S) | |
| sf = None | |
| if s is not None: sf = np.full((T, S), np.nan, np.float32); sf[ti] = s.reshape(len(ti), S) | |
| out[v] = (full, sf) | |
| return out | |