cryoem_au_data/validate.py

"""Check an index against the contract. Run this on a new dataset FIRST.

    python -m cryoem_au_data.validate <index_dir>

Every check corresponds to a failure we have actually hit. The ones marked
FATAL break training silently rather than loudly, which is why they are checks
and not comments.
"""

from __future__ import annotations

import sys
from pathlib import Path

import numpy as np
import pandas as pd

from .load import TABLES, load_index, compute_utility
from .schema import (
    GEOMETRY_REQUIRED, HOLES_REQUIRED, HOLE_UID_RE, LABEL_REQUIRED,
    OPERATOR_ONLY, SQUARE_UID_RE, STATUSES, STATUS_OK, TARGET_REQUIRED,
    session_of,
)


class Report:
    def __init__(self) -> None:
        self.fatal: list[str] = []
        self.warn: list[str] = []
        self.info: list[str] = []

    def check(self, ok: bool, message: str, fatal: bool = True) -> bool:
        if ok:
            self.info.append(f"ok    {message}")
        elif fatal:
            self.fatal.append(f"FATAL {message}")
        else:
            self.warn.append(f"warn  {message}")
        return ok

    def show(self) -> int:
        for line in self.info:
            print(line)
        for line in self.warn:
            print(line)
        for line in self.fatal:
            print(line)
        print()
        if self.fatal:
            print(f"{len(self.fatal)} FATAL problem(s): the agents will not train correctly.")
            return 1
        if self.warn:
            print(f"passed with {len(self.warn)} warning(s).")
            return 0
        print("passed.")
        return 0


def validate(index_dir) -> int:
    root = Path(index_dir)
    r = Report()
    tables = load_index(root)
    targets, holes = tables["target_locations"], tables["holes"]
    geometry, labels = tables["square_geometry"], tables["labels_v4_outcomes"]

    for name, frame, required in (
        ("target_locations", targets, TARGET_REQUIRED),
        ("holes", holes, HOLES_REQUIRED),
        ("square_geometry", geometry, GEOMETRY_REQUIRED),
        ("labels_v4_outcomes", labels, LABEL_REQUIRED),
    ):
        missing = [c for c in required if c not in frame.columns]
        r.check(not missing, f"{name}: required columns present"
                             + (f" (missing {missing})" if missing else ""))

    # --- UID grammar: the silent-failure class -------------------------------
    bad = [u for u in targets["hole_uid"].astype(str)[:5000] if not HOLE_UID_RE.match(u)]
    r.check(not bad, "hole_uid matches session:<name>:square:<id>:hole:<id>"
                     + (f" (e.g. {bad[0]!r})" if bad else ""))
    bad_sq = [u for u in geometry["square_uid"].astype(str)[:5000] if not SQUARE_UID_RE.match(u)]
    r.check(not bad_sq, "square_uid matches session:<name>:square:<id>"
                        + (f" (e.g. {bad_sq[0]!r})" if bad_sq else ""))
    sessions = sorted({session_of(u) for u in targets["hole_uid"].astype(str)})
    r.check(len(sessions) >= 1, f"sessions found: {len(sessions)}")

    # --- joins ---------------------------------------------------------------
    r.check(labels["sample_id"].is_unique, "sample_id is unique in labels")
    orphan = set(labels["hole_uid"]) - set(targets["hole_uid"])
    r.check(not orphan, f"every labelled hole_uid appears in target_locations"
                        + (f" ({len(orphan)} orphans)" if orphan else ""))
    orphan_sq = set(targets["square_uid"]) - set(geometry["square_uid"])
    r.check(not orphan_sq, f"every square_uid has geometry"
                           + (f" ({len(orphan_sq)} missing)" if orphan_sq else ""),
            fatal=False)

    # --- candidates vs acquired ---------------------------------------------
    ratio = len(targets) / max(len(holes), 1)
    r.check(len(targets) > len(holes),
            f"target_locations holds ALL candidates, not just acquired "
            f"({len(targets):,} candidates / {len(holes):,} acquired = {ratio:.1f}x)")

    # --- leakage -------------------------------------------------------------
    leaked = OPERATOR_ONLY & set(targets.columns)
    r.check(not leaked, f"no operator-derived feature columns"
                        + (f" ({sorted(leaked)})" if leaked else ""), fatal=False)

    # --- labels --------------------------------------------------------------
    unknown = set(labels["status"].astype(str)) - set(STATUSES)
    r.check(not unknown, f"status vocabulary is {STATUSES}"
                         + (f" (saw {sorted(unknown)})" if unknown else ""))
    n_ok = int((labels["status"] == STATUS_OK).sum())
    r.check(n_ok > 0, f"{n_ok:,} acquisitions usable for training "
                      f"({100*n_ok/max(len(labels),1):.0f}%)")

    # Replay order. Missing timestamps fall back to a sample_id sort, which is
    # alphabetical rather than chronological — usable but not the real order.
    # Three of our four reference sessions have none; please do better than we did.
    ts = pd.to_datetime(labels["timestamp"], errors="coerce")
    per_session_ts = (
        labels.assign(_s=[session_of(u) for u in labels["hole_uid"].astype(str)],
                      _t=ts.notna())
        .groupby("_s")["_t"].mean())
    bare = sorted(per_session_ts[per_session_ts < 0.5].index)
    r.check(not bare,
            "every session has timestamps (replay order is chronological)"
            + (f" — missing for {[b.replace('session:','') for b in bare]}" if bare else ""),
            fatal=False)
    r.check(ts.notna().any() and ts.nunique() > 1,
            "timestamps vary where present (not a constant placeholder)", fatal=False)

    per_hole = labels.groupby("hole_uid").size()
    r.check((per_hole > 1).any(),
            f"some holes have repeated acquisitions (max {int(per_hole.max())}) "
            "— needed for within-hole variance", fatal=False)

    # --- utility reproduces --------------------------------------------------
    recomputed = compute_utility(labels)
    diff = np.abs(recomputed - labels["utility"].to_numpy(float))
    err = float(np.nanmax(diff)) if np.isfinite(diff).any() else float("inf")
    r.check(np.isfinite(diff).all() and err < 1e-6,
            f"utility column matches the documented formula (max err {err:.2e})")

    ok_rows = labels[labels["status"] == STATUS_OK]
    if len(ok_rows) > 10:
        res = ok_rows["ctf_res_adj"].to_numpy(float)
        ice = ok_rows["ice_ring"].to_numpy(float)
        gate = 1.0 / (1.0 + np.exp(-(6.0 - res) / 1.5))
        r.check(0.02 < gate.std() < 0.6,
                f"CTF gate has dynamic range (std {gate.std():.3f}) — if ~0 the "
                "6 A target is off the end of this dataset and needs re-tuning",
                fatal=False)
        r.check(np.isfinite(ice).any() and (ice > 1.5).mean() < 0.95,
                f"ice threshold 1.5 is not penalising everything "
                f"({100*(ice > 1.5).mean():.0f}% above) — re-tune if it is",
                fatal=False)
        util = ok_rows["utility"].to_numpy(float)
        r.check(util.std() > 0, f"utility varies (mean {util.mean():.1f}, sd {util.std():.1f})")

    print(f"index: {root}")
    print(f"  {len(targets):,} candidate holes · {len(holes):,} acquired · "
          f"{len(labels):,} acquisitions · {len(geometry):,} squares")
    print(f"  sessions: {', '.join(s.replace('session:', '') for s in sessions)}\n")
    return r.show()


def main(argv: list[str] | None = None) -> int:
    args = sys.argv[1:] if argv is None else argv
    if not args:
        print(__doc__)
        return 2
    return validate(args[0])


if __name__ == "__main__":
    raise SystemExit(main())