transfer/code/CCFM/scripts/preextract_scgpt.py

"""
Pre-extract scGPT per-gene features for all training cells.

Saves to HDF5 for use with ScGPTFeatureCache during training.
Must run on GPU node via pjsub.

Usage:
    python scripts/preextract_scgpt.py --data_name norman --batch_size 256 --output scgpt_cache_norman.h5
"""

import sys
import os
import argparse

# Set up paths
_PROJECT_ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
sys.path.insert(0, _PROJECT_ROOT)

# Bootstrap scDFM imports
import _bootstrap_scdfm  # noqa: F401

import torch
import numpy as np
import h5py
from tqdm import tqdm

from src.data.data import get_data_classes
from src.data.scgpt_extractor import FrozenScGPTExtractor

_REPO_ROOT = os.path.dirname(_PROJECT_ROOT)  # transfer/code/


def main():
    parser = argparse.ArgumentParser(description="Pre-extract scGPT features")
    parser.add_argument("--data_name", type=str, default="norman")
    parser.add_argument("--n_top_genes", type=int, default=5000)
    parser.add_argument("--split_method", type=str, default="additive")
    parser.add_argument("--fold", type=int, default=1)
    parser.add_argument("--topk", type=int, default=15)
    parser.add_argument("--use_negative_edge", action="store_true")
    parser.add_argument("--scgpt_model_dir", type=str, default="transfer/data/scGPT_pretrained")
    parser.add_argument("--batch_size", type=int, default=256)
    parser.add_argument("--output", type=str, default="scgpt_cache_norman.h5")
    args = parser.parse_args()

    if args.data_name == "norman":
        args.n_top_genes = 5000

    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    print(f"Device: {device}")

    # === Load data (same as run_cascaded.py) ===
    Data, PerturbationDataset, TrainSampler, TestDataset = get_data_classes()

    scdfm_data_path = os.path.join(_REPO_ROOT, "scDFM", "data")
    data_manager = Data(scdfm_data_path)
    data_manager.load_data(args.data_name)

    # Convert var_names from Ensembl IDs to gene symbols if needed
    if "gene_name" in data_manager.adata.var.columns and data_manager.adata.var_names[0].startswith("ENSG"):
        data_manager.adata.var_names = data_manager.adata.var["gene_name"].values
        data_manager.adata.var_names_make_unique()
        print(f"Converted var_names to gene symbols, sample: {list(data_manager.adata.var_names[:5])}")

    data_manager.process_data(
        n_top_genes=args.n_top_genes,
        split_method=args.split_method,
        fold=args.fold,
        use_negative_edge=args.use_negative_edge,
        k=args.topk,
    )

    # Get all cells from the adata (train + valid + test all share the same adata genes)
    adata = data_manager.adata
    n_cells = adata.n_obs
    n_genes = adata.n_vars
    hvg_gene_names = list(adata.var_names)
    cell_names = list(adata.obs_names)

    print(f"Cells: {n_cells}, Genes: {n_genes}")
    print(f"HVG gene names sample: {hvg_gene_names[:5]}")

    # === Build FrozenScGPTExtractor with large max_seq_len ===
    scgpt_model_dir = os.path.join(
        os.path.dirname(_REPO_ROOT),  # transfer/
        args.scgpt_model_dir.replace("transfer/", ""),
    )

    # Count valid genes to set max_seq_len
    import json
    vocab_path = os.path.join(scgpt_model_dir, "vocab.json")
    with open(vocab_path, "r") as f:
        scgpt_vocab = json.load(f)
    n_valid = sum(1 for g in hvg_gene_names if g in scgpt_vocab)
    max_seq_len = n_valid + 2  # +1 CLS, +1 margin
    print(f"Valid genes in scGPT vocab: {n_valid}/{n_genes}, max_seq_len={max_seq_len}")

    extractor = FrozenScGPTExtractor(
        model_dir=scgpt_model_dir,
        hvg_gene_names=hvg_gene_names,
        device=device,
        max_seq_len=max_seq_len,
        target_std=1.0,
        warmup_batches=0,  # no warmup needed, we compute global stats
    )
    extractor = extractor.to(device)
    extractor.eval()

    scgpt_dim = extractor.scgpt_d_model
    print(f"scGPT d_model: {scgpt_dim}")

    # === Get expression matrix ===
    # adata.X may be sparse
    import scipy.sparse as sp
    if sp.issparse(adata.X):
        X = torch.from_numpy(adata.X.toarray()).float()
    else:
        X = torch.from_numpy(np.array(adata.X)).float()

    # === Create HDF5 output ===
    print(f"Output: {args.output}")
    print(f"Features shape: ({n_cells}, {n_genes}, {scgpt_dim}) float16")

    h5 = h5py.File(args.output, "w")
    feat_ds = h5.create_dataset(
        "features",
        shape=(n_cells, n_genes, scgpt_dim),
        dtype=np.float16,
        chunks=(min(args.batch_size, n_cells), n_genes, scgpt_dim),
    )

    # === Extract features in batches ===
    # We pass gene_indices=None so extract() uses all genes
    # We do NOT apply normalization yet — store raw features, compute stats after
    # Temporarily disable normalization by setting running_mean=0, running_var=1
    extractor.running_mean.zero_()
    extractor.running_var.fill_(1.0)
    extractor._stats_frozen = True  # don't update stats during extraction

    running_sum = torch.zeros(scgpt_dim, dtype=torch.float64)
    running_sq_sum = torch.zeros(scgpt_dim, dtype=torch.float64)
    total_valid_count = 0

    for start in tqdm(range(0, n_cells, args.batch_size), desc="Extracting"):
        end = min(start + args.batch_size, n_cells)
        batch_expr = X[start:end].to(device)  # (B, G)

        # Extract with target_std=1.0 and identity normalization → raw features
        with torch.no_grad():
            feats = extractor.extract(batch_expr, gene_indices=None)  # (B, G, D)

        feats_cpu = feats.cpu()

        # Accumulate stats on non-zero features (genes with valid scGPT mapping)
        nonzero_mask = feats_cpu.abs().sum(-1) > 0  # (B, G)
        if nonzero_mask.any():
            valid_feats = feats_cpu[nonzero_mask].double()  # (K, D)
            running_sum += valid_feats.sum(dim=0)
            running_sq_sum += (valid_feats ** 2).sum(dim=0)
            total_valid_count += valid_feats.shape[0]

        # Store raw (un-normalized) features as float16
        feat_ds[start:end] = feats_cpu.numpy().astype(np.float16)

    # === Compute global normalization statistics ===
    global_mean = (running_sum / total_valid_count).float()
    global_var = ((running_sq_sum / total_valid_count) - global_mean.double() ** 2).float().clamp_min(0)
    print(f"Global mean range: [{global_mean.min():.4f}, {global_mean.max():.4f}]")
    print(f"Global var range: [{global_var.min():.4f}, {global_var.max():.4f}]")

    # Save stats and cell names
    h5.create_dataset("norm_mean", data=global_mean.numpy())
    h5.create_dataset("norm_var", data=global_var.numpy())

    # Save cell names as variable-length strings
    dt = h5py.string_dtype()
    h5.create_dataset("cell_names", data=np.array(cell_names, dtype=object), dtype=dt)

    h5.close()
    print(f"Done! Saved to {args.output}")
    print(f"  Features: ({n_cells}, {n_genes}, {scgpt_dim}) float16")
    print(f"  Valid features counted: {total_valid_count}")


if __name__ == "__main__":
    main()