revision/scripts/analyze_embedding_stability.py

"""Quantify cross-seed stability of learned laboratory embeddings.

The comparison uses rank correlations between vectors of pairwise laboratory
distances. Pairwise distances are invariant to translation, rotation, and
reflection of an embedding, while Spearman correlation is also invariant to a
positive global rescaling. This avoids comparing arbitrary embedding axes.
"""

from __future__ import annotations

import argparse
import itertools
import json
from pathlib import Path
from typing import Any

import numpy as np
import pandas as pd
import torch


STRATEGIES = ("canonical_grouped", "scaffold_aware")
SEEDS = (123456, 123457, 123458)
MODEL_SPECS = {
    "gat": ("graph_model.lab_embedding.weight", "gat_fold_*.pt"),
    "gcn": ("graph_model.lab_embedding.weight", "gcn_fold_*.pt"),
    "fpnn": ("lab_embedding.weight", "fp_nn_fold_*.pt"),
}


def _prepare_output_dir(path: Path) -> Path:
    if path.exists() and any(path.iterdir()):
        raise FileExistsError(f"Refusing nonempty output directory: {path}")
    path.mkdir(parents=True, exist_ok=True)
    return path


def _load_laboratory_order(neural_dir: Path) -> list[str]:
    encoder_path = neural_dir / "fold_preprocessing" / "lab_encoder.json"
    if not encoder_path.is_file():
        raise FileNotFoundError(f"Incomplete checkpoint matrix: {encoder_path}")
    payload = json.loads(encoder_path.read_text(encoding="utf-8"))
    laboratories = payload.get("classes_in_index_order")
    if not isinstance(laboratories, list) or len(laboratories) < 3:
        raise ValueError(f"Invalid laboratory encoder: {encoder_path}")
    if len(laboratories) != len(set(laboratories)):
        raise ValueError(f"Duplicate laboratory labels: {encoder_path}")
    return [str(value) for value in laboratories]


def _load_embedding(checkpoint_path: Path, key: str) -> np.ndarray:
    payload: dict[str, Any] = torch.load(
        checkpoint_path,
        map_location="cpu",
        weights_only=True,
    )
    model_state = payload.get("model_state")
    if not isinstance(model_state, dict) or key not in model_state:
        raise KeyError(f"Missing {key} in {checkpoint_path}")
    embedding = model_state[key].detach().cpu().numpy().astype(float, copy=False)
    if embedding.ndim != 2:
        raise ValueError(f"Expected a two-dimensional embedding in {checkpoint_path}")
    return embedding


def _mean_pairwise_distance_vector(
    neural_dir: Path,
    model: str,
    expected_folds: int,
) -> tuple[tuple[str, ...], np.ndarray]:
    key, pattern = MODEL_SPECS[model]
    checkpoint_paths = sorted((neural_dir / "checkpoints" / model).glob(pattern))
    if len(checkpoint_paths) != expected_folds:
        raise FileNotFoundError(
            "Incomplete checkpoint matrix: "
            f"expected {expected_folds} {model} checkpoints in {neural_dir}, "
            f"found {len(checkpoint_paths)}"
        )

    encoder_order = _load_laboratory_order(neural_dir)
    common_order = tuple(sorted(encoder_order))
    reorder = np.asarray([encoder_order.index(label) for label in common_order])
    upper = np.triu_indices(len(common_order), k=1)
    fold_distances: list[np.ndarray] = []

    for checkpoint_path in checkpoint_paths:
        embedding = _load_embedding(checkpoint_path, key)
        if embedding.shape[0] != len(encoder_order):
            raise ValueError(
                f"Laboratory count mismatch in {checkpoint_path}: "
                f"{embedding.shape[0]} versus {len(encoder_order)}"
            )
        aligned = embedding[reorder]
        difference = aligned[:, None, :] - aligned[None, :, :]
        distances = np.sqrt(np.sum(difference * difference, axis=2))
        fold_distances.append(distances[upper])

    return common_order, np.mean(np.vstack(fold_distances), axis=0)


def _spearman_correlation(vector_a: np.ndarray, vector_b: np.ndarray) -> float:
    """Compute Spearman rho without importing SciPy's additional OpenMP runtime."""
    ranks_a = pd.Series(vector_a).rank(method="average").to_numpy(dtype=float)
    ranks_b = pd.Series(vector_b).rank(method="average").to_numpy(dtype=float)
    centered_a = ranks_a - ranks_a.mean()
    centered_b = ranks_b - ranks_b.mean()
    denominator = np.sqrt(
        np.sum(centered_a * centered_a) * np.sum(centered_b * centered_b)
    )
    if denominator == 0:
        raise ValueError("Cannot compute embedding stability from constant distances.")
    return float(np.sum(centered_a * centered_b) / denominator)


def main() -> int:
    parser = argparse.ArgumentParser(
        description="Compare learned laboratory geometry across outer seeds."
    )
    parser.add_argument("--artifacts-root", required=True)
    parser.add_argument("--output-dir", required=True)
    parser.add_argument("--expected-folds", type=int, default=6)
    arguments = parser.parse_args()

    artifacts_root = Path(arguments.artifacts_root).resolve()
    output_dir = _prepare_output_dir(Path(arguments.output_dir).resolve())
    distance_vectors: dict[tuple[str, str, int], np.ndarray] = {}
    laboratory_orders: dict[tuple[str, str, int], tuple[str, ...]] = {}

    for strategy in STRATEGIES:
        for seed in SEEDS:
            neural_dir = (
                artifacts_root / strategy / f"seed_{seed}" / "neural_stack"
            )
            for model in MODEL_SPECS:
                order, vector = _mean_pairwise_distance_vector(
                    neural_dir,
                    model,
                    arguments.expected_folds,
                )
                laboratory_orders[(strategy, model, seed)] = order
                distance_vectors[(strategy, model, seed)] = vector

    rows: list[dict[str, Any]] = []
    for strategy in STRATEGIES:
        for model in MODEL_SPECS:
            for seed_a, seed_b in itertools.combinations(SEEDS, 2):
                order_a = laboratory_orders[(strategy, model, seed_a)]
                order_b = laboratory_orders[(strategy, model, seed_b)]
                if order_a != order_b:
                    raise ValueError(
                        f"Laboratory labels differ for {strategy}/{model}: "
                        f"seed {seed_a} versus seed {seed_b}"
                    )
                vector_a = distance_vectors[(strategy, model, seed_a)]
                vector_b = distance_vectors[(strategy, model, seed_b)]
                rho = _spearman_correlation(vector_a, vector_b)
                rows.append(
                    {
                        "strategy": strategy,
                        "model": model,
                        "seed_a": seed_a,
                        "seed_b": seed_b,
                        "n_laboratories": len(order_a),
                        "n_laboratory_pairs": len(vector_a),
                        "spearman_rho": rho,
                    }
                )

    correlations = pd.DataFrame(rows)
    correlations.to_csv(output_dir / "embedding_stability.csv", index=False)
    aggregate = (
        correlations.groupby(["strategy", "model"])["spearman_rho"]
        .agg(["mean", "min", "max"])
        .reset_index()
    )
    aggregate.to_csv(output_dir / "embedding_stability_aggregate.csv", index=False)
    summary = {
        "expected_folds": arguments.expected_folds,
        "distance_definition": "mean foldwise Euclidean distance",
        "comparison": "Spearman correlation of laboratory-pair distance vectors",
        "invariances": [
            "translation",
            "rotation",
            "reflection",
            "positive global scaling",
        ],
        "strategies": list(STRATEGIES),
        "seeds": list(SEEDS),
        "models": list(MODEL_SPECS),
    }
    (output_dir / "embedding_stability_summary.json").write_text(
        json.dumps(summary, indent=2) + "\n",
        encoding="utf-8",
    )
    print(f"Wrote embedding-stability analysis to: {output_dir}")
    return 0


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