transfer/code/prompt_selection/run_pipeline.py

"""Main pipeline: embedding-based prompt selection for Stack generation.

Supports two modes:
  --shared-only   : Run shared steps only (download ckpt, extract query_ctrl
                    & prompt_ctrl, compute their embeddings).
  --perturbation X: Run per-perturbation steps (extract prompt_pert, bridge
                    prediction, pert embeddings, custom generation).

Execution order for per-perturbation (3 model loads):
  Step 0'  : Extract prompt_pert from source AnnData
  Load 1   : bc_large_aligned.ckpt → bridge prediction → release
  Load 2   : bc_large.ckpt → pert embeddings → release
  Load 3   : bc_large_aligned.ckpt → custom generation → release
"""
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
from prompt_selection.custom_generation import custom_generation_loop

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


# ---------------------------------------------------------------------------
# Helpers
# ---------------------------------------------------------------------------

def _release_model(model):
    """Delete model and free GPU memory."""
    del model
    gc.collect()
    if torch.cuda.is_available():
        torch.cuda.empty_cache()


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()


# ---------------------------------------------------------------------------
# Shared steps
# ---------------------------------------------------------------------------

def step0_download_checkpoint():
    """Download bc_large.ckpt from HuggingFace if not present."""
    if cfg.EMBED_CKPT.exists():
        LOGGER.info("Embedding checkpoint already exists: %s", cfg.EMBED_CKPT)
        return

    LOGGER.info("Downloading %s from HuggingFace...", cfg.HF_EMBED_REPO)
    from huggingface_hub import snapshot_download

    snapshot_download(
        repo_id=cfg.HF_EMBED_REPO,
        local_dir=str(cfg.EMBED_MODEL_DIR),
        allow_patterns=["bc_large.ckpt"],
    )
    if not cfg.EMBED_CKPT.exists():
        import glob
        matches = glob.glob(str(cfg.EMBED_MODEL_DIR / "**" / "bc_large.ckpt"), recursive=True)
        if matches:
            Path(matches[0]).rename(cfg.EMBED_CKPT)
    assert cfg.EMBED_CKPT.exists(), f"Failed to download {cfg.EMBED_CKPT}"
    LOGGER.info("Downloaded embedding checkpoint to %s", cfg.EMBED_CKPT)


def step0_extract_shared_subsets():
    """Extract query_ctrl and prompt_ctrl (shared across all perturbations)."""
    cfg.RESULTS_DIR.mkdir(parents=True, exist_ok=True)

    query_path = cfg.RESULTS_DIR / cfg.QUERY_CTRL_H5AD
    ctrl_path = cfg.RESULTS_DIR / cfg.PROMPT_CTRL_H5AD

    if query_path.exists() and ctrl_path.exists():
        LOGGER.info("Shared cell subsets already extracted, skipping.")
        return

    LOGGER.info("Loading source AnnData: %s", cfg.SOURCE_ADATA)
    adata = ad.read_h5ad(str(cfg.SOURCE_ADATA))

    if not query_path.exists():
        query = _filter_adata(adata, cfg.QUERY_FILTER)
        LOGGER.info("query_ctrl: %d cells", query.n_obs)
        query.write_h5ad(query_path)
        del query

    if not ctrl_path.exists():
        ctrl = _filter_adata(adata, cfg.PROMPT_CTRL_FILTER)
        LOGGER.info("prompt_ctrl: %d cells", ctrl.n_obs)
        ctrl.write_h5ad(ctrl_path)
        del ctrl

    del adata
    gc.collect()


def step_extract_shared_embeddings():
    """Extract embeddings for query_ctrl and prompt_ctrl (shared)."""
    emb_files = [
        (cfg.QUERY_EMB_NPY, cfg.QUERY_CTRL_H5AD),
        (cfg.PROMPT_CTRL_EMB_NPY, cfg.PROMPT_CTRL_H5AD),
    ]

    all_exist = all((cfg.RESULTS_DIR / npy).exists() for npy, _ in emb_files)
    if all_exist:
        LOGGER.info("Shared embeddings already exist, skipping.")
        return

    LOGGER.info("=== Loading bc_large.ckpt for shared embedding extraction ===")
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    model = load_model_from_checkpoint(str(cfg.EMBED_CKPT), device=device)

    for npy_name, h5ad_name in emb_files:
        npy_path = cfg.RESULTS_DIR / npy_name
        if npy_path.exists():
            LOGGER.info("  %s already exists, skipping.", npy_name)
            continue

        h5ad_path = str(cfg.RESULTS_DIR / h5ad_name)
        LOGGER.info("  Extracting embeddings: %s -> %s", h5ad_name, npy_name)

        embeddings, _ = model.get_latent_representation(
            adata_path=h5ad_path,
            genelist_path=str(cfg.GENELIST_PATH),
            batch_size=cfg.BATCH_SIZE,
            num_workers=cfg.NUM_WORKERS,
            show_progress=True,
        )

        np.save(npy_path, embeddings)
        LOGGER.info("  Saved: %s, shape=%s", npy_path, embeddings.shape)

    _release_model(model)


def run_shared_steps():
    """Run all shared steps (checkpoint download, subset extraction, embeddings)."""
    LOGGER.info("=" * 60)
    LOGGER.info("Running shared steps")
    LOGGER.info("=" * 60)

    step0_download_checkpoint()
    step0_extract_shared_subsets()
    step_extract_shared_embeddings()

    LOGGER.info("Shared steps complete.")


# ---------------------------------------------------------------------------
# Per-perturbation steps
# ---------------------------------------------------------------------------

def step_extract_prompt_pert(pcfg: cfg.PertConfig):
    """Extract prompt_pert (T cells + drug X) for a specific perturbation."""
    pcfg.results_dir.mkdir(parents=True, exist_ok=True)
    pert_path = pcfg.results_dir / cfg.PROMPT_PERT_H5AD

    if pert_path.exists():
        LOGGER.info("prompt_pert already exists for %s, skipping.", pcfg.perturbation_name)
        return True

    LOGGER.info("Loading source AnnData: %s", cfg.SOURCE_ADATA)
    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 perturbation '%s'. Skipping.", pcfg.perturbation_name)
        (pcfg.results_dir / "SKIPPED_no_tcells.txt").touch()
        del adata, pert
        gc.collect()
        return False

    pert.write_h5ad(pert_path)

    del adata, pert
    gc.collect()
    return True


def step_bridge_prediction(pcfg: cfg.PertConfig):
    """Generate predicted perturbation for control T cells using aligned model."""
    pred_path = pcfg.results_dir / cfg.PREDICTED_PERT_H5AD
    if pred_path.exists():
        LOGGER.info("Predicted perturbation already exists for %s, skipping.", pcfg.perturbation_name)
        return

    LOGGER.info("=== Load 1: bc_large_aligned.ckpt for bridge prediction (%s) ===", pcfg.perturbation_name)
    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,
    )

    prompt_pert_path = str(pcfg.results_dir / cfg.PROMPT_PERT_H5AD)
    prompt_ctrl_path = str(cfg.RESULTS_DIR / cfg.PROMPT_CTRL_H5AD)

    LOGGER.info("Running bridge prediction: prompt_pert → prompt_ctrl")
    result = model.get_incontext_generation(
        base_adata_or_path=prompt_pert_path,
        test_adata_or_path=prompt_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,
    )
    pred_X, _ = result

    _release_model(model)

    ctrl_adata = ad.read_h5ad(prompt_ctrl_path)
    if issparse(pred_X):
        pred_X_dense = pred_X.toarray()
    else:
        pred_X_dense = np.asarray(pred_X)

    pred_adata = ad.AnnData(
        X=csr_matrix(pred_X_dense.astype(np.float32)),
        obs=ctrl_adata.obs.copy(),
        var=ctrl_adata.var.copy(),
    )
    pred_adata.obs["sm_name"] = pcfg.perturbation_name
    pred_adata.write_h5ad(pred_path)

    LOGGER.info("Saved predicted perturbation: %s (%d cells)", pred_path, pred_adata.n_obs)
    del ctrl_adata, pred_adata, pred_X, pred_X_dense
    gc.collect()


def step_extract_pert_embeddings(pcfg: cfg.PertConfig):
    """Extract embeddings for prompt_pert and predicted_pert."""
    emb_files = [
        (cfg.PROMPT_PERT_EMB_NPY, cfg.PROMPT_PERT_H5AD),
        (cfg.PREDICTED_PERT_EMB_NPY, cfg.PREDICTED_PERT_H5AD),
    ]

    all_exist = all((pcfg.results_dir / npy).exists() for npy, _ in emb_files)
    if all_exist:
        LOGGER.info("Pert embeddings already exist for %s, skipping.", pcfg.perturbation_name)
        return

    LOGGER.info("=== Load 2: bc_large.ckpt for pert embedding extraction (%s) ===", pcfg.perturbation_name)
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    model = load_model_from_checkpoint(str(cfg.EMBED_CKPT), device=device)

    for npy_name, h5ad_name in emb_files:
        npy_path = pcfg.results_dir / npy_name
        if npy_path.exists():
            LOGGER.info("  %s already exists, skipping.", npy_name)
            continue

        h5ad_path = str(pcfg.results_dir / h5ad_name)
        LOGGER.info("  Extracting embeddings: %s -> %s", h5ad_name, npy_name)

        embeddings, _ = model.get_latent_representation(
            adata_path=h5ad_path,
            genelist_path=str(cfg.GENELIST_PATH),
            batch_size=cfg.BATCH_SIZE,
            num_workers=cfg.NUM_WORKERS,
            show_progress=True,
        )

        np.save(npy_path, embeddings)
        LOGGER.info("  Saved: %s, shape=%s", npy_path, embeddings.shape)

    _release_model(model)


def step_custom_generation(pcfg: cfg.PertConfig):
    """Run MDM generation with per-step embedding-based prompt selection."""
    final_path = pcfg.results_dir / pcfg.final_result_h5ad
    if final_path.exists():
        LOGGER.info("Final result already exists for %s, skipping.", pcfg.perturbation_name)
        return

    LOGGER.info("=== Load 3: bc_large_aligned.ckpt for custom generation (%s) ===", pcfg.perturbation_name)
    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_adata = ad.read_h5ad(str(cfg.RESULTS_DIR / cfg.QUERY_CTRL_H5AD))
    prompt_pert_adata = ad.read_h5ad(str(pcfg.results_dir / cfg.PROMPT_PERT_H5AD))

    # Load precomputed embeddings (shared + per-pert)
    query_emb = np.load(cfg.RESULTS_DIR / cfg.QUERY_EMB_NPY)
    prompt_ctrl_emb = np.load(cfg.RESULTS_DIR / cfg.PROMPT_CTRL_EMB_NPY)
    predicted_pert_emb = np.load(pcfg.results_dir / cfg.PREDICTED_PERT_EMB_NPY)
    prompt_pert_emb = np.load(pcfg.results_dir / cfg.PROMPT_PERT_EMB_NPY)

    LOGGER.info("Embeddings loaded: query=%s, ctrl=%s, pred_pert=%s, pert=%s",
                query_emb.shape, prompt_ctrl_emb.shape,
                predicted_pert_emb.shape, prompt_pert_emb.shape)

    result, final_logit = custom_generation_loop(
        model=model,
        query_adata=query_adata,
        prompt_pert_adata=prompt_pert_adata,
        genelist_path=str(cfg.GENELIST_PATH),
        query_embeddings=query_emb,
        prompt_ctrl_embeddings=prompt_ctrl_emb,
        predicted_pert_embeddings=predicted_pert_emb,
        prompt_pert_embeddings=prompt_pert_emb,
        num_steps=cfg.NUM_STEPS,
        prompt_ratio=cfg.PROMPT_RATIO,
        context_ratio=cfg.CONTEXT_RATIO,
        context_ratio_min=cfg.CONTEXT_RATIO_MIN,
        top_k1=cfg.TOP_K1,
        batch_size=cfg.BATCH_SIZE,
        num_workers=cfg.NUM_WORKERS,
    )

    _release_model(model)

    query_ctrl_clean = ad.read_h5ad(str(cfg.RESULTS_DIR / cfg.QUERY_CTRL_H5AD))
    result_adata = ad.AnnData(
        X=result,
        obs=query_ctrl_clean.obs.copy(),
        var=query_ctrl_clean.var.copy(),
    )
    result_adata.obs["sm_name"] = pcfg.perturbation_name
    result_adata.obs["control"] = False
    result_adata.obs["gen_logit"] = np.asarray(final_logit)

    result_adata.write_h5ad(final_path)
    LOGGER.info("Saved final result: %s (%d cells)", final_path, result_adata.n_obs)


def run_perturbation(pert_name: str):
    """Run all per-perturbation steps for a given drug."""
    pcfg = cfg.get_pert_config(pert_name)

    LOGGER.info("=" * 60)
    LOGGER.info("Running perturbation: %s", pert_name)
    LOGGER.info("=" * 60)

    # Verify shared files exist
    for shared_file in [cfg.QUERY_CTRL_H5AD, cfg.PROMPT_CTRL_H5AD,
                        cfg.QUERY_EMB_NPY, cfg.PROMPT_CTRL_EMB_NPY]:
        if not (cfg.RESULTS_DIR / shared_file).exists():
            raise FileNotFoundError(
                f"Shared file {shared_file} not found. Run --shared-only first."
            )

    # Extract prompt_pert
    has_data = step_extract_prompt_pert(pcfg)
    if not has_data:
        LOGGER.warning("Skipping %s (no T cell data).", pert_name)
        return

    # Bridge prediction
    step_bridge_prediction(pcfg)

    # Pert embeddings
    step_extract_pert_embeddings(pcfg)

    # Custom generation
    step_custom_generation(pcfg)

    LOGGER.info("=" * 60)
    LOGGER.info("Perturbation %s complete!", pert_name)
    LOGGER.info("Final result: %s", pcfg.results_dir / pcfg.final_result_h5ad)
    LOGGER.info("=" * 60)


# ---------------------------------------------------------------------------
# Main
# ---------------------------------------------------------------------------

def main():
    parser = argparse.ArgumentParser(
        description="Prompt Selection Pipeline for Stack generation."
    )
    parser.add_argument(
        "--perturbation", type=str, default=None,
        help="Perturbation name (e.g., Dabrafenib). Required unless --shared-only.",
    )
    parser.add_argument(
        "--shared-only", action="store_true",
        help="Only run shared steps (extract query_ctrl, prompt_ctrl, their embeddings).",
    )
    args = parser.parse_args()

    if args.shared_only:
        run_shared_steps()
    elif args.perturbation:
        run_perturbation(args.perturbation)
    else:
        parser.error("Either --perturbation or --shared-only is required.")


if __name__ == "__main__":
    main()