File size: 17,304 Bytes
a5899e5 | 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 | """Train/evaluate expected greenness v2-ca exactly as locked in configs/model_v2_ca.yaml.
python ca/30_train_ca.py --git-commit <hash>
Training/evaluation on a 150k-cell sample; out-of-fold predictions for every eligible cell.
Writes runs/v2_ca/: metrics.json, model_ca.ubj, feature_schema.json, conformal.json, climatology / weather
climatology parquet, summer_residual.npy (cells x years, residual vs all-years climatology), expected_block.parquet.
"""
from __future__ import annotations
import argparse
import json
import time
from pathlib import Path
import numpy as np
import pandas as pd
import pyarrow.parquet as pq
import xgboost as xgb
from sklearn.linear_model import Ridge
from sklearn.model_selection import GroupKFold
CUBE = Path("/root/trendex-ca/data/cube")
RUN = Path("/root/trendex-ca/runs/v2_ca")
YEARS = np.arange(2001, 2026)
MONTHS = np.arange(5, 10)
CUBE_MONTHS = list(range(4, 11))
STATIC = ["elev_mean", "elev_std", "slope_mean", "steep_share", "aspect_sin", "aspect_cos", "tpi",
"grass_frac", "barren_frac", "crop_frac", "lc_type1", "lat"]
WEATHER = ["precip_m", "precip_2m", "precip_winter", "tair_m", "tair_prev", "soil_m", "swe_apr"]
FEATURES = STATIC + ["month", "clim_ndvi"] + WEATHER
FIXED = {"tree_method": "hist", "device": "cuda", "objective": "reg:squarederror"}
SEEDS = [0, 1, 2]
N_SAMPLE = 150_000
def log(m):
print(f"[{time.strftime('%H:%M:%S')}] {m}", flush=True)
def load_arrays(cells_all: pd.DataFrame, elig: np.ndarray) -> dict:
"""(n_elig, 25 years, 7 months Apr-Oct) arrays straight from the yearly files (rows = month blocks x all cells)."""
n_all, idx = len(cells_all), np.nonzero(elig)[0]
shape = (len(idx), len(YEARS), len(CUBE_MONTHS))
arr = {k: np.full(shape, np.nan, "float32") for k in ["ndvi", "precip", "precip_winter", "tair", "soil_moist", "swe"]}
for j, yr in enumerate(YEARS):
t = pq.read_table(CUBE / "monthly" / f"year={yr}" / "part.parquet",
columns=["month", "ndvi", "ndvi_ok", "precip", "precip_winter", "tair", "soil_moist", "swe"])
months = t.column("month").to_numpy()
assert len(months) % n_all == 0, "monthly file is not month-blocks x all cells"
ok = t.column("ndvi_ok").to_numpy(zero_copy_only=False)
cols = {k: t.column(k).to_numpy(zero_copy_only=False).astype("float32") for k in arr}
cols["ndvi"] = np.where(ok, cols["ndvi"], np.nan)
for b in range(len(months) // n_all):
mo = int(months[b * n_all])
mi = CUBE_MONTHS.index(mo)
for k in arr:
arr[k][:, j, mi] = cols[k][b * n_all:(b + 1) * n_all][idx]
log(f"loaded {yr}")
return arr
def build_features(cells: pd.DataFrame, arr: dict, run: Path) -> dict:
mi = [CUBE_MONTHS.index(m) for m in MONTHS]
prev = [CUBE_MONTHS.index(m - 1) for m in MONTHS]
take = lambda a, ix: a[:, :, ix] # noqa: E731
anom = lambda x: x - np.nanmean(x, axis=1, keepdims=True) # noqa: E731
with np.errstate(all="ignore"):
raw_m = {"precip_m": take(arr["precip"], mi), "precip_2m": take(arr["precip"], mi) + take(arr["precip"], prev),
"tair_m": take(arr["tair"], mi), "tair_prev": take(arr["tair"], prev), "soil_m": take(arr["soil_moist"], mi)}
raw_y = {"precip_winter": arr["precip_winter"][:, :, CUBE_MONTHS.index(5)],
"swe_apr": arr["swe"][:, :, CUBE_MONTHS.index(4)]}
f = {k: anom(v) for k, v in raw_m.items()}
f |= {k: np.repeat(anom(v)[:, :, None], len(MONTHS), axis=2) for k, v in raw_y.items()}
wc = pd.concat([pd.DataFrame({"cell_id": cells["cell_id"].values, "month": int(mo)} |
{f"{k}_mean": np.nanmean(v[:, :, j], axis=1).astype("float32") for k, v in raw_m.items()} |
{f"{k}_mean": np.nanmean(v, axis=1).astype("float32") for k, v in raw_y.items()})
for j, mo in enumerate(MONTHS)], ignore_index=True)
wc.to_parquet(run / "weather_climatology.parquet", index=False)
f["ndvi"] = take(arr["ndvi"], mi)
return f
def clim_fold(ndvi, train_years):
tm = np.isin(YEARS, train_years)[None, :, None]
fin = np.isfinite(ndvi)
s = np.nansum(np.where(tm, ndvi, np.nan), axis=1, keepdims=True)
n = np.sum(fin & tm, axis=1, keepdims=True).astype("float32")
own = np.where(fin & tm, ndvi, 0.0)
with np.errstate(all="ignore"):
return ((s - own) / (n - (fin & tm))).astype("float32")
def matrix(f: dict, cells: pd.DataFrame, rows: np.ndarray, years: np.ndarray, clim: np.ndarray):
"""Feature matrix for cell positions `rows`, the given years and May-Sep. Returns X, y, meta."""
yi = np.nonzero(np.isin(YEARS, years))[0]
sub = lambda a: a[rows][:, yi].reshape(-1) # noqa: E731
nm = len(MONTHS)
shape = (len(rows), len(yi), nm)
cols = {}
for k in STATIC:
cols[k] = np.broadcast_to(cells[k].values.astype("float32")[rows][:, None, None], shape).reshape(-1)
cols["month"] = np.broadcast_to(MONTHS.astype("float32")[None, None, :], shape).reshape(-1)
cols["clim_ndvi"] = sub(clim)
for k in WEATHER:
cols[k] = sub(f[k])
X = np.column_stack([cols[k] for k in FEATURES]).astype("float32")
ndvi = sub(f["ndvi"])
y = ndvi - cols["clim_ndvi"]
meta = {"row": np.broadcast_to(rows[:, None, None], shape).reshape(-1),
"year": np.broadcast_to(YEARS[yi][None, :, None], shape).reshape(-1),
"month": cols["month"].astype("int8"), "ndvi": ndvi}
return X, y.astype("float32"), meta
class RidgeModel:
def __init__(self, alpha):
self.alpha = alpha
def fit(self, X, y):
self.med = np.nanmedian(X, axis=0)
Z = np.where(np.isnan(X), self.med, X)
self.mu, self.sd = Z.mean(0), Z.std(0)
self.sd[self.sd == 0] = 1
self.r = Ridge(alpha=self.alpha).fit((Z - self.mu) / self.sd, y)
return self
def predict(self, X):
Z = np.where(np.isnan(X), self.med, X)
return self.r.predict((Z - self.mu) / self.sd)
def fit_xgb(X, y, params, n_est, seed, Xv=None, yv=None):
m = xgb.XGBRegressor(**FIXED, **params, n_estimators=n_est, random_state=seed,
early_stopping_rounds=50 if Xv is not None else None)
m.fit(X, y, eval_set=[(Xv, yv)] if Xv is not None else None, verbose=False)
return m
def boot(sse_a, sse_b, rng, n=2000):
idx = rng.integers(0, len(sse_a), size=(n, len(sse_a)))
dist = 1 - sse_a[idx].sum(1) / sse_b[idx].sum(1)
return {"point": float(1 - sse_a.sum() / sse_b.sum()), "ci_lo": float(np.quantile(dist, 0.025)),
"ci_hi": float(np.quantile(dist, 0.975)), "units": int(len(sse_a))}
def score(y, preds, units, cuts, rng):
out = {"n_rows": int(len(y))}
sse = {}
for k, p in preds.items():
e = y - p
t, q = np.digitize(y, cuts), np.digitize(p, cuts)
out[k] = {"rmse": float(np.sqrt(np.mean(e ** 2))), "mae": float(np.mean(np.abs(e))),
"anomaly_corr": float(np.corrcoef(y, p)[0, 1]) if np.std(p) > 0 else None,
"tercile_bacc": float(np.mean([np.mean(q[t == c] == c) for c in range(3) if np.any(t == c)]))}
sse[k] = pd.Series(e ** 2).groupby(units).sum().values
out["skill_xgb_vs_ridge"] = boot(sse["xgb"], sse["ridge"], rng)
out["skill_xgb_vs_clim"] = boot(sse["xgb"], sse["clim"], rng)
out["skill_ridge_vs_clim"] = boot(sse["ridge"], sse["clim"], rng)
return out
def finite(X, y, *extra):
ok = np.isfinite(y)
return (X[ok], y[ok]) + tuple(e[ok] for e in extra)
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--git-commit", required=True)
a = ap.parse_args()
RUN.mkdir(parents=True, exist_ok=True)
rng = np.random.default_rng(0)
metrics = {"protocol": "configs/model_v2_ca.yaml", "git_commit": a.git_commit, "features": FEATURES}
cells_all = pd.read_parquet(CUBE / "cells.parquet")
elig = cells_all["mask_reason"].isna().values
cells = cells_all[elig].reset_index(drop=True)
arr = load_arrays(cells_all, elig)
f = build_features(cells, arr, RUN)
del arr
n = len(cells)
sample = np.sort(np.random.default_rng(0).choice(n, size=min(N_SAMPLE, n), replace=False))
sb = ((cells["block_row"].values // 3).astype(np.int64) * 100000 + cells["block_col"].values // 3)
metrics["data"] = {"eligible_cells": int(n), "sample_cells": int(len(sample)),
"by_country_eligible": cells["country"].value_counts().astype(int).to_dict()}
log(f"eligible cells {n}, sample {len(sample)}")
# --- tuning on the inner split (sample only)
c_in = clim_fold(f["ndvi"], np.arange(2001, 2016))
Xi, yi_ = finite(*matrix(f, cells, sample, np.arange(2001, 2016), c_in)[:2])
Xv, yv, mv = matrix(f, cells, sample, np.arange(2016, 2020), c_in)
Xv, yv, mv_month = finite(Xv, yv, mv["month"])
ridge_rmse = {al: float(np.sqrt(np.mean((yv - RidgeModel(al).fit(Xi, yi_).predict(Xv)) ** 2)))
for al in [1, 10, 100, 1000, 10000]}
alpha = min(ridge_rmse, key=ridge_rmse.get)
import optuna
optuna.logging.set_verbosity(optuna.logging.WARNING)
def objective(tr):
p = {"max_depth": tr.suggest_int("max_depth", 4, 10),
"learning_rate": tr.suggest_float("learning_rate", 0.03, 0.2, log=True),
"min_child_weight": tr.suggest_float("min_child_weight", 1, 200, log=True),
"subsample": tr.suggest_float("subsample", 0.5, 1.0),
"colsample_bytree": tr.suggest_float("colsample_bytree", 0.5, 1.0),
"reg_lambda": tr.suggest_float("reg_lambda", 1e-3, 30, log=True)}
m = fit_xgb(Xi, yi_, p, 2000, 0, Xv, yv)
tr.set_user_attr("n_est", int(m.best_iteration) + 1)
return float(np.sqrt(np.mean((yv - m.predict(Xv)) ** 2)))
st = optuna.create_study(direction="minimize", sampler=optuna.samplers.TPESampler(seed=0))
st.optimize(objective, n_trials=20, timeout=600)
best, n_est = st.best_params, st.best_trial.user_attrs["n_est"]
metrics["tuning"] = {"trials": len(st.trials), "best_params": best, "n_estimators": n_est,
"inner_rmse_xgb": st.best_value, "ridge_alpha": alpha, "inner_rmse_ridge": ridge_rmse}
log(f"tuning {len(st.trials)} trials, xgb {st.best_value:.4f} vs ridge {ridge_rmse[alpha]:.4f}, n_est {n_est}")
cuts = np.quantile(yi_, [1 / 3, 2 / 3])
res_in = yv - fit_xgb(Xi, yi_, best, n_est, 0).predict(Xv)
conf = {int(mo): [float(np.quantile(res_in[mv_month == mo], q)) for q in (0.1, 0.9)] for mo in MONTHS}
del Xi, yi_, Xv, yv
# --- temporal test
c_t = clim_fold(f["ndvi"], np.arange(2001, 2020))
Xtr, ytr = finite(*matrix(f, cells, sample, np.arange(2001, 2020), c_t)[:2])
Xte, yte, mte = matrix(f, cells, sample, np.arange(2020, 2026), c_t)
Xte, yte, te_year, te_month = finite(Xte, yte, mte["year"], mte["month"])
ridge = RidgeModel(alpha).fit(Xtr, ytr)
p_r = ridge.predict(Xte)
ps, seed_skill = [], []
for s in SEEDS:
p = fit_xgb(Xtr, ytr, best, n_est, s).predict(Xte)
ps.append(p)
seed_skill.append(float(1 - np.mean((yte - p) ** 2) / np.mean((yte - p_r) ** 2)))
p_x = np.mean(ps, axis=0)
temporal = score(yte, {"xgb": p_x, "ridge": p_r, "clim": np.zeros_like(yte)}, te_year, cuts, rng)
temporal["seed_skill_vs_ridge"] = {"values": seed_skill, "mean": float(np.mean(seed_skill)), "sd": float(np.std(seed_skill))}
lo = np.array([conf[int(m)][0] for m in te_month]); hi = np.array([conf[int(m)][1] for m in te_month])
temporal["conformal_coverage_p10_p90"] = float(np.mean(((yte - p_x) >= lo) & ((yte - p_x) <= hi)))
metrics["temporal_test"] = temporal
log(f"temporal skill vs ridge {temporal['skill_xgb_vs_ridge']}")
del Xtr, ytr, Xte, yte
# --- spatial test (sample, all years, superblock groups)
c_all = clim_fold(f["ndvi"], YEARS)
Xa, ya, ma = matrix(f, cells, sample, YEARS, c_all)
Xa, ya, rows_a = finite(Xa, ya, ma["row"])
groups = sb[rows_a]
preds = {"xgb": np.zeros_like(ya), "ridge": np.zeros_like(ya)}
for itr, ite in GroupKFold(5).split(Xa, groups=groups):
preds["xgb"][ite] = fit_xgb(Xa[itr], ya[itr], best, n_est, 0).predict(Xa[ite])
preds["ridge"][ite] = RidgeModel(alpha).fit(Xa[itr], ya[itr]).predict(Xa[ite])
preds["clim"] = np.zeros_like(ya)
spatial = score(ya, preds, groups, cuts, rng)
metrics["spatial_test"] = spatial
log(f"spatial skill vs ridge {spatial['skill_xgb_vs_ridge']}")
passed = temporal["skill_xgb_vs_ridge"]["ci_lo"] > 0 and spatial["skill_xgb_vs_ridge"]["ci_lo"] > 0
chosen = "xgb" if passed else "ridge"
metrics["gate"] = {"rule": "CI lower bound of skill_xgb_vs_ridge > 0 on temporal AND spatial tests",
"passed": bool(passed), "shipped_model": chosen}
log(f"GATE passed={passed} -> {chosen}")
# --- production model (sample, all years) + importance
prod = fit_xgb(Xa, ya, best, n_est, 0) if chosen == "xgb" else RidgeModel(alpha).fit(Xa, ya)
if chosen == "xgb":
booster = prod.get_booster()
booster.feature_names = FEATURES
booster.save_model(RUN / "model_ca.ubj")
sel = np.random.default_rng(1).choice(len(Xa), size=min(200_000, len(Xa)), replace=False)
contrib = booster.predict(xgb.DMatrix(Xa[sel], feature_names=FEATURES), pred_contribs=True)
metrics["feature_importance_mean_abs"] = dict(sorted(zip(FEATURES, np.abs(contrib[:, :-1]).mean(0).round(6).tolist()),
key=lambda kv: -kv[1]))
del Xa, ya
# --- out-of-fold for ALL eligible cells (5 year blocks), residual vs all-years climatology
clim_all = np.nanmean(f["ndvi"], axis=1, keepdims=True) # (n, 1, 5)
summer_res = np.full((n, len(YEARS)), np.nan, "float32")
bcode, buniq = pd.factorize(cells["block_id"])
nb = len(buniq)
acc = {k: np.zeros((nb, len(YEARS), len(MONTHS))) for k in ("observed", "expected", "residual", "n")}
all_rows = np.arange(n)
summer_idx = [list(MONTHS).index(m) for m in (6, 7, 8)]
for blk in [np.arange(s, s + 5) for s in (2001, 2006, 2011, 2016, 2021)]:
trn = np.setdiff1d(YEARS, blk)
cf = clim_fold(f["ndvi"], trn)
Xs, ys = finite(*matrix(f, cells, sample, trn, cf)[:2])
model = fit_xgb(Xs, ys, best, n_est, 0) if chosen == "xgb" else RidgeModel(alpha).fit(Xs, ys)
del Xs, ys
for chunk in np.array_split(all_rows, 10):
X, _, meta = matrix(f, cells, chunk, blk, cf)
pa_ = model.predict(X).astype("float32").reshape(len(chunk), len(blk), len(MONTHS))
yi = np.nonzero(np.isin(YEARS, blk))[0]
expected = clim_all[chunk] + pa_
obs = f["ndvi"][chunk][:, yi]
res = obs - expected
with np.errstate(all="ignore"):
summer_res[chunk[:, None], yi[None, :]] = np.nanmean(res[:, :, summer_idx], axis=2)
bc = bcode[chunk]
for j, yy in enumerate(yi):
for k in range(len(MONTHS)):
v = np.isfinite(res[:, j, k])
acc["n"][:, yy, k] += np.bincount(bc[v], minlength=nb)
for name, val in (("observed", obs), ("expected", expected), ("residual", res)):
acc[name][:, yy, k] += np.bincount(bc[v], weights=val[v, j, k], minlength=nb)
log(f"oof block {blk[0]}-{blk[-1]} done")
np.save(RUN / "summer_residual.npy", summer_res)
cells[["cell_id", "block_id", "region", "country"]].to_parquet(RUN / "eligible_cells.parquet", index=False)
B, Y, M = np.meshgrid(np.arange(nb), YEARS, MONTHS, indexing="ij")
with np.errstate(all="ignore"):
eb = pd.DataFrame({"block_id": np.asarray(buniq)[B.ravel()], "year": Y.ravel(), "month": M.ravel(),
"observed": (acc["observed"] / acc["n"]).ravel(), "expected": (acc["expected"] / acc["n"]).ravel(),
"residual": (acc["residual"] / acc["n"]).ravel(), "n_cells": acc["n"].ravel().astype(int)})
eb = eb[eb["n_cells"] > 0].reset_index(drop=True)
eb["p10"] = eb["expected"] + eb["month"].map(lambda m: conf[int(m)][0])
eb["p90"] = eb["expected"] + eb["month"].map(lambda m: conf[int(m)][1])
eb.to_parquet(RUN / "expected_block.parquet", index=False)
pd.concat([pd.DataFrame({"cell_id": cells["cell_id"].values, "month": int(mo),
"clim_ndvi": clim_all[:, 0, j]}) for j, mo in enumerate(MONTHS)]) \
.to_parquet(RUN / "climatology.parquet", index=False)
(RUN / "feature_schema.json").write_text(json.dumps({"features": FEATURES, "static": STATIC, "weather": WEATHER,
"months": MONTHS.tolist(), "shipped_model": chosen,
"target": "ndvi_anomaly = ndvi - clim_ndvi"}, indent=2))
(RUN / "conformal.json").write_text(json.dumps(conf, indent=2))
(RUN / "metrics.json").write_text(json.dumps(metrics, indent=2, default=float))
log("TRAIN_CA_DONE")
if __name__ == "__main__":
main()
|