File size: 7,725 Bytes
a5899e5 ce7a57d 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 | """Package v2-ca for HF (runs/v2_ca/hf_model) + what-if grid by region and country (runs/v2_ca/products/scenarios).
Golden test: 100 rows rebuilt with the training code; predict_ca.py must reproduce the training model exactly.
"""
from __future__ import annotations
import importlib.util
import json
import shutil
import sys
from pathlib import Path
import numpy as np
import pandas as pd
import xgboost as xgb
ROOT = Path("/root/trendex-ca")
RUN = ROOT / "runs" / "v2_ca"
HF = RUN / "hf_model"
MONTHS, PRECIP, TEMP = [5, 6, 7, 8, 9], [-30, -20, -10, 0, 10, 20, 30], [-2, -1, 0, 1, 2]
def load_module(path, name):
spec = importlib.util.spec_from_file_location(name, path)
mod = importlib.util.module_from_spec(spec)
sys.modules[name] = mod
spec.loader.exec_module(mod)
return mod
def main():
HF.mkdir(parents=True, exist_ok=True)
for f in ["model_ca.ubj", "feature_schema.json", "climatology.parquet", "weather_climatology.parquet",
"conformal.json", "metrics.json"]:
shutil.copy(RUN / f, HF / f)
shutil.copy(ROOT / "scripts" / "predict_ca.py", HF / "predict_ca.py")
shutil.copy(ROOT / "configs" / "model_v2_ca.yaml", HF / "model_v2_ca.yaml")
tr = load_module(ROOT / "scripts" / "30_train_ca.py", "train_ca")
cells = pd.read_parquet(tr.CUBE / "cells.parquet")
cells_e = cells[cells["mask_reason"].isna()].reset_index(drop=True)
pm = load_module(HF / "predict_ca.py", "predict_ca")
model = pm.load(HF)
# golden: features for 100 random eligible cells in 2024 July via what_if inputs at zero deltas is too
# synthetic, so rebuild real rows from the cube for 300 cells and keep 100 finite ones
rng = np.random.default_rng(0)
pick = np.sort(rng.choice(len(cells_e), size=300, replace=False))
elig = cells["mask_reason"].isna().values
sub_mask = np.zeros(len(cells), bool)
sub_mask[np.nonzero(elig)[0][pick]] = True
arr = tr.load_arrays(cells, sub_mask)
tmp = RUN / "_golden_tmp"
tmp.mkdir(exist_ok=True)
f = tr.build_features(cells_e.iloc[pick].reset_index(drop=True), arr, tmp)
clim = tr.clim_fold(f["ndvi"], tr.YEARS)
X, y, meta = tr.matrix(f, cells_e.iloc[pick].reset_index(drop=True), np.arange(len(pick)), np.array([2024]), clim)
ok = np.isfinite(y)
gin = pd.DataFrame(X[ok][:100], columns=tr.FEATURES)
(HF / "golden").mkdir(exist_ok=True)
gin.to_parquet(HF / "golden" / "inputs.parquet", index=False)
out = model.predict(gin)
out.to_parquet(HF / "golden" / "expected.parquet", index=False)
ref = xgb.XGBRegressor()
ref.load_model(RUN / "model_ca.ubj")
diff = float(np.max(np.abs(ref.predict(gin[tr.FEATURES].to_numpy("float32")) - out["pred_anom"].to_numpy())))
print(f"golden max diff {diff:.2e}")
assert diff < 1e-5
shutil.rmtree(tmp)
# what-if grid: up to 2,000 random eligible cells per region, aggregated to region and country
shuffled = cells_e.sample(frac=1.0, random_state=0)
samp = shuffled[shuffled.groupby("region", dropna=False).cumcount() < 2000]
rows = []
for mo in MONTHS:
for p in PRECIP:
for t in TEMP:
r = model.what_if(cells, samp["cell_id"].to_numpy(), month=mo, precip_pct=p, temp_c=t)
rows.append(pd.DataFrame({"region": samp["region"].to_numpy(), "country": samp["country"].to_numpy(),
"month": mo, "p": p, "t": t, "typical": r["expected_typical_weather"].to_numpy(),
"change": r["change_vs_typical"].to_numpy()}))
W = pd.concat(rows, ignore_index=True)
def nested(df):
g = df.groupby(["month", "p", "t"])["change"].mean()
typ = df[(df["p"] == 0) & (df["t"] == 0)].groupby("month")["typical"].mean()
return {"n_cells_sampled": int(len(df) // (len(MONTHS) * len(PRECIP) * len(TEMP))),
"typical": [round(float(typ[m]), 4) for m in MONTHS],
"change": [[[round(float(g[(m, p, t)]), 4) for t in TEMP] for p in PRECIP] for m in MONTHS]}
js = {"model": "TrendeX90/expected-greenness-ca@v2-ca", "units": "NDVI", "months": MONTHS, "precip_pct": PRECIP,
"temp_c": TEMP, "layout": "change[month][precip][temp]; expected = typical[month] + change",
"meaning": "Expected greenness change vs typical weather if precipitation and GLDAS air temperature differ by "
"the given amounts. Model sensitivity, not a causal effect; averages over up to 2,000 sampled "
"pasture cells per region.",
"all": nested(W), "countries": {c: nested(g) for c, g in W.groupby("country")},
"regions": {r: nested(g) for r, g in W.groupby("region")}}
sc = RUN / "products" / "scenarios"
sc.mkdir(parents=True, exist_ok=True)
(sc / "what_if.json").write_text(json.dumps(js, separators=(",", ":")))
print(f"what_if.json {(sc / 'what_if.json').stat().st_size / 1e3:.0f} kB; July all: typical {js['all']['typical'][2]}, "
f"-30% {js['all']['change'][2][0][2]}, +2C {js['all']['change'][2][3][4]}")
m = json.loads((RUN / "metrics.json").read_text())
t, s = m["temporal_test"], m["spatial_test"]
fmt = lambda d: f"{d['point']:.3f} (95% CI {d['ci_lo']:.3f} to {d['ci_hi']:.3f}, {d['units']} units)" # noqa: E731
(HF / "README.md").write_text(f"""---
license: other
library_name: xgboost
tags: [nasa, modis, ndvi, imerg, gldas, pastures, central-asia, tabular-regression]
datasets: [TrendeX90/ca-cube]
---
# TrendeX expected greenness, Central Asia (v2-ca)
Summer (May-Sep) NDVI anomaly of 1 km pasture cells in Kazakhstan, Kyrgyzstan, Tajikistan, Turkmenistan and
Uzbekistan, predicted from weather (IMERG precipitation; GLDAS air temperature, soil moisture, snow) and terrain
only. The gap to observed NDVI is change **not explained by weather**: not a cause, not forage quality.
Protocol locked in git before fitting (`model_v2_ca.yaml`, commit {m['git_commit']}). Naryn pilot model: TrendeX90/expected-greenness (v1).
| Test | RMSE xgb | RMSE ridge | RMSE climatology | anomaly corr xgb |
|---|---|---|---|---|
| Temporal (2020-2025) | {t['xgb']['rmse']:.4f} | {t['ridge']['rmse']:.4f} | {t['clim']['rmse']:.4f} | {t['xgb']['anomaly_corr']:.3f} |
| Spatial (unseen super-blocks) | {s['xgb']['rmse']:.4f} | {s['ridge']['rmse']:.4f} | {s['clim']['rmse']:.4f} | {s['xgb']['anomaly_corr']:.3f} |
- Skill vs ridge: temporal {fmt(t['skill_xgb_vs_ridge'])}; spatial {fmt(s['skill_xgb_vs_ridge'])}.
- Skill vs climatology: temporal {fmt(t['skill_xgb_vs_clim'])}.
- Seeds ({len(t['seed_skill_vs_ridge']['values'])}): {t['seed_skill_vs_ridge']['mean']:.3f} +/- {t['seed_skill_vs_ridge']['sd']:.3f}.
- **Gate (pre-registered, CI lower bound > 0 on both tests): passed = {m['gate']['passed']}; shipped = {m['gate']['shipped_model']}.**
- P10-P90 coverage on 2020-2025: {t['conformal_coverage_p10_p90']:.0%} (nominal 80%).
Limits: trained on {m['data']['sample_cells']:,} sampled cells (of {m['data']['eligible_cells']:,} eligible) because of budget; GLDAS
(0.25 deg) is coarse; MCD12Q1 2020 mask; NDVI is greenness, not forage; no field validation; budget-limited tuning
(20 trials) and 3 seeds. Spatial test shares years with training and is easier than the temporal test.
```python
from predict_ca import load # numpy, pandas, xgboost; CPU
m = load(".")
m.predict(features_df) # columns as in feature_schema.json
```
Data: MOD13A3.061, MCD12Q1.061, NASADEM_HGT.001 (LP DAAC), GPM_3IMERGM.07, GLDAS_NOAH025_M.2.1 (GES DISC).
AI use: implementation assisted by Claude (Anthropic).
""")
print("PACKAGE_CA_DONE", sorted(p.name for p in HF.iterdir()))
if __name__ == "__main__":
main()
|