transfer/code/prompt_selection/run_baseline.py

"""Random baseline generation using Stack's get_incontext_generation().

Uses random prompt ordering (the default Stack behaviour) instead of
embedding-based selection.  Provides a baseline for comparison.

Usage:
    python code/prompt_selection/run_baseline.py --perturbation Dabrafenib
"""
from __future__ import annotations

import argparse
import gc
import logging
import sys
from pathlib import Path

_THIS_DIR = Path(__file__).resolve().parent
if str(_THIS_DIR.parent) not in sys.path:
    sys.path.insert(0, str(_THIS_DIR.parent))

import anndata as ad
import numpy as np
import torch
from scipy.sparse import csr_matrix, issparse

from stack.model_loading import load_model_from_checkpoint

from prompt_selection import config as cfg

logging.basicConfig(
    level=logging.INFO,
    format="%(asctime)s [%(name)s] %(levelname)s: %(message)s",
)
LOGGER = logging.getLogger("prompt_selection.baseline")


def _filter_adata(adata: ad.AnnData, filters: dict) -> ad.AnnData:
    """Subset AnnData by column-value filters."""
    mask = np.ones(adata.n_obs, dtype=bool)
    for col, val in filters.items():
        mask &= (adata.obs[col] == val).values
    return adata[mask].copy()


def ensure_prompt_pert(pcfg: cfg.PertConfig):
    """Extract prompt_pert.h5ad if it doesn't exist yet."""
    pcfg.results_dir.mkdir(parents=True, exist_ok=True)
    pert_path = pcfg.results_dir / cfg.PROMPT_PERT_H5AD

    if pert_path.exists():
        return True

    LOGGER.info("prompt_pert not found, extracting from source data...")
    adata = ad.read_h5ad(str(cfg.SOURCE_ADATA))
    pert = _filter_adata(adata, pcfg.prompt_pert_filter)
    LOGGER.info("prompt_pert (%s): %d cells", pcfg.perturbation_name, pert.n_obs)

    if pert.n_obs == 0:
        LOGGER.warning("No T cells found for '%s'. Skipping.", pcfg.perturbation_name)
        del adata, pert
        gc.collect()
        return False

    pert.write_h5ad(pert_path)
    del adata, pert
    gc.collect()
    return True


def main():
    parser = argparse.ArgumentParser(description="Random Baseline Generation")
    parser.add_argument(
        "--perturbation", type=str, required=True,
        help="Perturbation name (e.g., Dabrafenib).",
    )
    args = parser.parse_args()

    pert_name = args.perturbation
    pcfg = cfg.get_pert_config(pert_name)

    LOGGER.info("=" * 60)
    LOGGER.info("Random Baseline Generation — %s", pert_name)
    LOGGER.info("=" * 60)

    pcfg.baseline_dir.mkdir(parents=True, exist_ok=True)
    output_path = pcfg.baseline_dir / pcfg.baseline_result_h5ad

    if output_path.exists():
        LOGGER.info("Baseline result already exists: %s — skipping.", output_path)
        return

    # Ensure prompt_pert data exists
    has_data = ensure_prompt_pert(pcfg)
    if not has_data:
        LOGGER.warning("Skipping baseline for %s (no T cell data).", pert_name)
        return

    # --- Load model ---
    LOGGER.info("Loading model: %s", cfg.ALIGNED_CKPT)
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    model = load_model_from_checkpoint(
        str(cfg.ALIGNED_CKPT),
        model_class="ICL_FinetunedModel",
        device=device,
    )

    # --- Load data ---
    query_ctrl_path = str(cfg.RESULTS_DIR / cfg.QUERY_CTRL_H5AD)
    prompt_pert_path = str(pcfg.results_dir / cfg.PROMPT_PERT_H5AD)

    LOGGER.info("Query (test):  %s", query_ctrl_path)
    LOGGER.info("Prompt (base): %s", prompt_pert_path)

    # --- Run random-prompt generation ---
    LOGGER.info("Running get_incontext_generation (random prompt baseline)...")
    result = model.get_incontext_generation(
        base_adata_or_path=prompt_pert_path,
        test_adata_or_path=query_ctrl_path,
        genelist_path=str(cfg.GENELIST_PATH),
        mode="mdm",
        num_steps=cfg.NUM_STEPS,
        prompt_ratio=cfg.PROMPT_RATIO,
        context_ratio=cfg.CONTEXT_RATIO,
        context_ratio_min=cfg.CONTEXT_RATIO_MIN,
        batch_size=cfg.BATCH_SIZE,
        num_workers=cfg.NUM_WORKERS,
    )

    if isinstance(result, tuple):
        predictions, test_logit = result
    else:
        predictions, test_logit = result, None

    # --- Build output AnnData ---
    query_ctrl = ad.read_h5ad(query_ctrl_path)

    if issparse(predictions):
        pred_X = predictions
    else:
        pred_X = csr_matrix(np.asarray(predictions, dtype=np.float32))

    result_adata = ad.AnnData(
        X=pred_X,
        obs=query_ctrl.obs.copy(),
        var=query_ctrl.var.copy(),
    )
    result_adata.obs["sm_name"] = pert_name
    result_adata.obs["control"] = False
    if test_logit is not None:
        result_adata.obs["gen_logit"] = np.asarray(test_logit)

    result_adata.write_h5ad(output_path)
    LOGGER.info("Saved baseline result: %s  shape=%s", output_path, result_adata.shape)

    LOGGER.info("=" * 60)
    LOGGER.info("Random Baseline Generation — %s — Done", pert_name)
    LOGGER.info("=" * 60)


if __name__ == "__main__":
    main()