File size: 3,419 Bytes
dd4b643
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Extend v2-ca outputs to cells the corrected formula mask re-includes (mountain barren -> pasture).

The extra set = status == "pasture" in the backend's NEW artifacts_ca/cells.parquet, minus cells already covered.
They were never in the model's training sample, so the shipped production model gives spatially out-of-sample
predictions for them (not year-block cross-fitted). Marked source="production_model_unseen_cell" in products.
Appends to runs/v2_ca/summer_residual.npy and eligible_cells.parquet; then re-run 40_products_ca.py.
"""
from __future__ import annotations

import importlib.util
import sys
import time
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"
FORMULA_CELLS = Path("/root/trendex/backend/artifacts_ca/cells.parquet")


def log(m):
    print(f"[{time.strftime('%H:%M:%S')}] {m}", flush=True)


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():
    tr = load_module(ROOT / "scripts" / "30_train_ca.py", "train_ca")
    fc = pd.read_parquet(FORMULA_CELLS)
    log(f"formula cells columns: {list(fc.columns)[:20]}")
    pasture = set(fc.loc[fc["status"] == "pasture", "cell_id"].astype("int64"))
    elig = pd.read_parquet(RUN / "eligible_cells.parquet")
    if "source" not in elig.columns:
        elig["source"] = "oof"
    have = set(elig["cell_id"].astype("int64"))
    cells_all = pd.read_parquet(tr.CUBE / "cells.parquet")
    extra_mask = cells_all["cell_id"].isin(pasture - have).values
    cells_x = cells_all[extra_mask].reset_index(drop=True)
    log(f"formula pasture {len(pasture)}, already covered {len(have & pasture)}, extra {len(cells_x)}; "
        f"by country {cells_x['country'].value_counts().to_dict()}; by old mask_reason {cells_x['mask_reason'].value_counts(dropna=False).to_dict()}")
    if len(cells_x) == 0:
        log("EXTEND_DONE (nothing to add)")
        return

    arr = tr.load_arrays(cells_all, extra_mask)
    tmp = RUN / "_extend_tmp"
    tmp.mkdir(exist_ok=True)
    f = tr.build_features(cells_x, arr, tmp)
    del arr
    clim_loyo = tr.clim_fold(f["ndvi"], tr.YEARS)          # what the production model saw as clim_ndvi
    clim_all = np.nanmean(f["ndvi"], axis=1, keepdims=True)
    booster = xgb.Booster()
    booster.load_model(str(RUN / "model_ca.ubj"))
    rows = np.arange(len(cells_x))
    X, _, _ = tr.matrix(f, cells_x, rows, tr.YEARS, clim_loyo)
    pred = booster.predict(xgb.DMatrix(X, feature_names=tr.FEATURES)).reshape(len(rows), len(tr.YEARS), len(tr.MONTHS))
    res = f["ndvi"] - (clim_all + pred)
    summer_idx = [list(tr.MONTHS).index(m) for m in (6, 7, 8)]
    with np.errstate(all="ignore"):
        sres = np.nanmean(res[:, :, summer_idx], axis=2).astype("float32")

    W = np.load(RUN / "summer_residual.npy")
    np.save(RUN / "summer_residual.npy", np.vstack([W, sres]))
    add = cells_x[["cell_id", "block_id", "region", "country"]].assign(source="production_model_unseen_cell")
    pd.concat([elig, add], ignore_index=True).to_parquet(RUN / "eligible_cells.parquet", index=False)
    import shutil
    shutil.rmtree(tmp)
    log(f"EXTEND_DONE added {len(add)} cells (total {len(elig) + len(add)})")


if __name__ == "__main__":
    main()