transfer/code/CCFM/scripts/diagnose_distributions.py

"""
Diagnostic script: check expression vs latent value distributions in CCFM training.
Loads data + scGPT cache, runs a few batches, prints distribution stats.

Usage (login node is fine — no GPU needed for cached features):
    cd /home/hp250092/ku50001222/qian/aivc/lfj/transfer/code/CCFM
    python scripts/diagnose_distributions.py
"""

import sys
import os

_PROJECT_ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
sys.path.insert(0, _PROJECT_ROOT)

import _bootstrap_scdfm  # noqa: F401

import torch
import numpy as np
from torch.utils.data import DataLoader

from src.data.data import get_data_classes
from src.data.scgpt_cache import ScGPTFeatureCache
from src._scdfm_imports import AffineProbPath, CondOTScheduler, process_vocab
from src.utils import GeneVocab

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


def describe(name, t):
    """Print distribution stats for a tensor."""
    t_flat = t.float().flatten()
    nonzero = t_flat[t_flat.abs() > 1e-8]
    print(f"  {name:30s} | shape {str(list(t.shape)):20s} | "
          f"mean={t_flat.mean():.4f}  std={t_flat.std():.4f}  "
          f"min={t_flat.min():.4f}  max={t_flat.max():.4f}  "
          f"median={t_flat.median():.4f}  "
          f"zero_frac={1 - len(nonzero)/len(t_flat):.2%}")
    if len(nonzero) > 0 and len(nonzero) < len(t_flat):
        print(f"  {'  (nonzero only)':30s} | "
              f"mean={nonzero.mean():.4f}  std={nonzero.std():.4f}  "
              f"min={nonzero.min():.4f}  max={nonzero.max():.4f}")


def main():
    device = torch.device("cpu")
    data_name = "norman"
    n_top_genes = 5000
    infer_top_gene = 1000
    batch_size = 48
    cache_path = os.path.join(_PROJECT_ROOT, "scgpt_cache_norman.h5")

    # --- Load data (reuse scDFM) ---
    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(data_name)

    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()

    data_manager.process_data(
        n_top_genes=n_top_genes, split_method="additive",
        fold=1, use_negative_edge=True, k=30,
    )
    train_sampler, _, _ = data_manager.load_flow_data(batch_size=batch_size)
    train_dataset = PerturbationDataset(train_sampler, batch_size)
    dataloader = DataLoader(train_dataset, batch_size=1, shuffle=False, num_workers=0)

    # --- Load scGPT cache ---
    scgpt_cache = ScGPTFeatureCache(cache_path, target_std=1.0)
    print(f"Cache shape: {scgpt_cache.features.shape}")
    print(f"Cache norm_mean: mean={scgpt_cache.norm_mean.mean():.4f}, std={scgpt_cache.norm_mean.std():.4f}")
    print(f"Cache norm_var:  mean={scgpt_cache.norm_var.mean():.4f}, std={scgpt_cache.norm_var.std():.4f}")

    # --- Flow path ---
    flow_path = AffineProbPath(scheduler=CondOTScheduler())

    # --- Run a few batches ---
    n_batches = 5
    print(f"\n{'='*90}")
    print(f"Running {n_batches} batches (batch_size={batch_size}, infer_top_gene={infer_top_gene})")
    print(f"{'='*90}")

    for i, batch_data in enumerate(dataloader):
        if i >= n_batches:
            break

        source = batch_data["src_cell_data"].squeeze(0)  # (B, G_full)
        target = batch_data["tgt_cell_data"].squeeze(0)
        tgt_cell_names = [n[0] if isinstance(n, (tuple, list)) else n for n in batch_data["tgt_cell_id"]]

        # Random gene subset
        input_gene_ids = torch.randperm(source.shape[-1])[:infer_top_gene]
        source_sub = source[:, input_gene_ids]
        target_sub = target[:, input_gene_ids]

        # scGPT latent features (from cache)
        z_target = scgpt_cache.lookup(tgt_cell_names, input_gene_ids, device=device)

        # Noise
        noise_expr = torch.randn_like(source_sub)
        noise_latent = torch.randn_like(z_target)

        # Sample time steps
        t = torch.sigmoid(torch.randn(source_sub.shape[0]))  # logit-normal

        # Flow path: expression
        path_expr = flow_path.sample(t=t, x_0=noise_expr, x_1=target_sub)

        # Flow path: latent (flatten for AffineProbPath)
        B, G, D = z_target.shape
        z_target_flat = z_target.reshape(B, G * D)
        noise_latent_flat = noise_latent.reshape(B, G * D)
        path_latent_flat = flow_path.sample(t=t, x_0=noise_latent_flat, x_1=z_target_flat)

        dx_t_expr = path_expr.dx_t
        dx_t_latent = path_latent_flat.dx_t.reshape(B, G, D)

        print(f"\n--- Batch {i} ---")
        print(f"[Raw Values]")
        describe("source_sub (control expr)", source_sub)
        describe("target_sub (perturbed expr)", target_sub)
        describe("z_target (scGPT latent)", z_target)

        print(f"\n[Noise]")
        describe("noise_expr", noise_expr)
        describe("noise_latent", noise_latent)

        print(f"\n[Flow Path x_t (interpolated)]")
        describe("x_t_expr", path_expr.x_t)
        describe("z_t_latent", path_latent_flat.x_t.reshape(B, G, D))

        print(f"\n[Velocity Targets dx_t]")
        describe("dx_t_expr", dx_t_expr)
        describe("dx_t_latent", dx_t_latent)

        print(f"\n[Velocity MSE (hypothetical)]")
        mse_expr = (dx_t_expr ** 2).mean().item()
        mse_latent = (dx_t_latent ** 2).mean().item()
        print(f"  mean(dx_t_expr^2)  = {mse_expr:.4f}")
        print(f"  mean(dx_t_latent^2) = {mse_latent:.4f}")
        print(f"  ratio expr/latent   = {mse_expr / max(mse_latent, 1e-8):.2f}x")

    scgpt_cache.close()
    print(f"\n{'='*90}")
    print("Done.")


if __name__ == "__main__":
    main()