train/CCFM/pca_emb/src/denoiser.py

"""
CascadedDenoiser for grn_svd — uses SVD-projected (B, G, 128) as latent target.

Key changes from grn_att_only:
- SVD dictionary W (5035, 128) as frozen register_buffer on GPU
- _sparse_project(): GPU-based loop-of-K gather+accumulate
- 1D missing mask (not 2D)
- Simplified masked MSE loss (no non-zero reweighting)
- reconstruct_grn(): inverse projection for GRN visualization
"""

import torch
import torch.nn as nn
import torchdiffeq

from ._scdfm_imports import AffineProbPath, CondOTScheduler, make_lognorm_poisson_noise
from .model.model import CascadedFlowModel
from .data.sparse_raw_cache import SparseDeltaCache


# Shared flow matching path
flow_path = AffineProbPath(scheduler=CondOTScheduler())


def pairwise_sq_dists(X, Y):
    return torch.cdist(X, Y, p=2) ** 2


@torch.no_grad()
def median_sigmas(X, scales=(0.5, 1.0, 2.0, 4.0)):
    D2 = pairwise_sq_dists(X, X)
    tri = D2[~torch.eye(D2.size(0), dtype=bool, device=D2.device)]
    m = torch.median(tri).clamp_min(1e-12)
    s2 = torch.tensor(scales, device=X.device) * m
    return [float(s.item()) for s in torch.sqrt(s2)]


def mmd2_unbiased_multi_sigma(X, Y, sigmas):
    m, n = X.size(0), Y.size(0)
    Dxx = pairwise_sq_dists(X, X)
    Dyy = pairwise_sq_dists(Y, Y)
    Dxy = pairwise_sq_dists(X, Y)
    vals = []
    for sigma in sigmas:
        beta = 1.0 / (2.0 * (sigma ** 2) + 1e-12)
        Kxx = torch.exp(-beta * Dxx)
        Kyy = torch.exp(-beta * Dyy)
        Kxy = torch.exp(-beta * Dxy)
        term_xx = (Kxx.sum() - Kxx.diag().sum()) / (m * (m - 1) + 1e-12)
        term_yy = (Kyy.sum() - Kyy.diag().sum()) / (n * (n - 1) + 1e-12)
        term_xy = Kxy.mean()
        vals.append(term_xx + term_yy - 2.0 * term_xy)
    return torch.stack(vals).mean()


class CascadedDenoiser(nn.Module):
    """
    Cascaded denoiser with SVD-projected latent target.

    Training: Cascaded time-step sampling, GPU-side SVD projection.
    Inference: Two-stage cascaded generation in 128-dim latent space.
    """

    def __init__(
        self,
        model: CascadedFlowModel,
        sparse_cache: SparseDeltaCache,
        svd_dict_path: str,
        choose_latent_p: float = 0.4,
        latent_weight: float = 1.0,
        noise_type: str = "Gaussian",
        use_mmd_loss: bool = True,
        gamma: float = 0.5,
        poisson_alpha: float = 0.8,
        poisson_target_sum: float = 1e4,
        # Logit-normal time-step sampling
        t_sample_mode: str = "logit_normal",
        t_expr_mean: float = 0.0,
        t_expr_std: float = 1.0,
        t_latent_mean: float = 0.0,
        t_latent_std: float = 1.0,
        # Cascaded noise
        noise_beta: float = 0.25,
        # Variance-weighted latent loss
        use_variance_weight: bool = False,
    ):
        super().__init__()
        self.model = model
        self.sparse_cache = sparse_cache
        self.choose_latent_p = choose_latent_p
        self.latent_weight = latent_weight
        self.noise_type = noise_type
        self.use_mmd_loss = use_mmd_loss
        self.gamma = gamma
        self.poisson_alpha = poisson_alpha
        self.poisson_target_sum = poisson_target_sum
        self.t_sample_mode = t_sample_mode
        self.t_expr_mean = t_expr_mean
        self.t_expr_std = t_expr_std
        self.t_latent_mean = t_latent_mean
        self.t_latent_std = t_latent_std
        self.noise_beta = noise_beta
        self.use_variance_weight = use_variance_weight

        # Load SVD dictionary as frozen buffer (moves with model to GPU)
        svd_dict = torch.load(svd_dict_path, map_location="cpu", weights_only=False)
        self.register_buffer("W", svd_dict["W"].float())             # (G_full, 128)
        self.register_buffer("global_scale_sq",
            torch.tensor(svd_dict["global_scale"] ** 2).float())     # scalar
        self.latent_dim = self.W.shape[1]

        # Variance-weighted latent loss: w_i = evr_i / sum(evr)
        if use_variance_weight:
            evr = torch.from_numpy(svd_dict["explained_variance_ratio"]).float()
            w = evr / evr.sum()  # (latent_dim,) normalized to sum=1
            self.register_buffer("latent_dim_weights", w)
            print(f"  Variance weighting ON: top={w[0]:.4f}, bot={w[-1]:.4f}, "
                  f"ratio={w[0]/w[-1]:.1f}x")

        print(f"  SVD dict loaded: W {self.W.shape}, "
              f"global_scale={svd_dict['global_scale']:.6f}")

    def _sparse_project(self, delta_values, delta_indices):
        """
        Sparse GPU projection: delta @ W via loop-of-K gather+accumulate.

        Args:
            delta_values:  (B, G_sub, K) float32 — top-K delta values (on GPU)
            delta_indices: (B, G_sub, K) int16   — column indices in G_full space (on GPU)

        Returns:
            z_target: (B, G_sub, latent_dim=128)
        """
        assert delta_values.device == self.W.device, \
            f"Device mismatch: delta on {delta_values.device}, W on {self.W.device}"

        B, G_sub, K = delta_values.shape
        z_target = torch.zeros(B, G_sub, self.latent_dim, device=delta_values.device)

        indices_long = delta_indices.long()  # (B, G_sub, K)
        for k in range(K):
            col_idx = indices_long[:, :, k]               # (B, G_sub)
            val = delta_values[:, :, k:k+1]               # (B, G_sub, 1)
            w_k = self.W[col_idx]                          # (B, G_sub, latent_dim)
            z_target = z_target + val * w_k                # broadcast + accumulate

        return z_target

    def reconstruct_grn(self, z: torch.Tensor) -> torch.Tensor:
        """
        Inverse projection: 128-dim latent → approximate delta attention (sparse GRN).

        Math: W_saved = V × s (V orthonormal, s = global_scale)
              Forward:  z = delta @ W = delta @ Vs
              Inverse:  delta ≈ z @ V^T / s = z @ W^T / s^2

        Args:
            z: (B, G, 128) — latent from ODE generation
        Returns:
            (B, G, G_full) — approximate delta attention per gene
        """
        return z @ self.W.T / self.global_scale_sq

    def sample_t(self, n: int, device: torch.device):
        """Cascaded time-step sampling — dino_first_cascaded mode."""
        if self.t_sample_mode == "logit_normal":
            t_latent = torch.sigmoid(torch.randn(n, device=device) * self.t_latent_std + self.t_latent_mean)
            t_expr = torch.sigmoid(torch.randn(n, device=device) * self.t_expr_std + self.t_expr_mean)
        else:
            t_latent = torch.rand(n, device=device)
            t_expr = torch.rand(n, device=device)

        choose_latent_mask = torch.rand(n, device=device) < self.choose_latent_p

        t_latent_expr = torch.rand_like(t_latent) * self.noise_beta + (1.0 - self.noise_beta)
        t_latent = torch.where(choose_latent_mask, t_latent, t_latent_expr)
        t_expr = torch.where(choose_latent_mask, torch.zeros_like(t_expr), t_expr)

        w_expr = (~choose_latent_mask).float()
        w_latent = choose_latent_mask.float()

        return t_expr, t_latent, w_expr, w_latent

    def _make_expr_noise(self, source: torch.Tensor) -> torch.Tensor:
        """Create noise for expression flow."""
        if self.noise_type == "Gaussian":
            return torch.randn_like(source)
        elif self.noise_type == "Poisson":
            return make_lognorm_poisson_noise(
                target_log=source,
                alpha=self.poisson_alpha,
                per_cell_L=self.poisson_target_sum,
            )
        else:
            raise ValueError(f"Unknown noise_type: {self.noise_type}")

    def train_step(
        self,
        source: torch.Tensor,           # (B, G_sub) — already subsetted
        target: torch.Tensor,            # (B, G_sub)
        perturbation_id: torch.Tensor,   # (B, 2)
        gene_input: torch.Tensor,        # (B, G_sub) vocab-encoded gene IDs
        delta_values: torch.Tensor,      # (B, G_sub, K) on GPU
        delta_indices: torch.Tensor,     # (B, G_sub, K) int16 on GPU
        input_gene_ids: torch.Tensor,    # (G_sub,) positional indices for missing mask
    ) -> dict:
        """Single training step with SVD-projected latent target."""
        B = source.shape[0]
        device = source.device
        G_sub = source.shape[-1]

        # 1. GPU-side SVD projection: sparse → dense 128-dim
        z_target = self._sparse_project(delta_values, delta_indices)  # (B, G_sub, 128)

        # 2. Missing gene mask — 1D only
        missing = self.sparse_cache.get_missing_gene_mask(input_gene_ids)  # (G_sub,) bool
        missing_dev = missing.to(device)

        # 3. Cascaded time sampling
        t_expr, t_latent, w_expr, w_latent = self.sample_t(B, device)

        # 4. Expression flow path
        noise_expr = self._make_expr_noise(source)
        path_expr = flow_path.sample(t=t_expr, x_0=noise_expr, x_1=target)

        # 5. Latent flow path — (B, G_sub, 128)
        noise_latent = torch.randn_like(z_target)
        # 1D missing mask: zero entire gene rows
        noise_latent[:, missing_dev, :] = 0.0

        z_flat = z_target.reshape(B, G_sub * self.latent_dim)
        noise_flat = noise_latent.reshape(B, G_sub * self.latent_dim)
        path_latent_flat = flow_path.sample(t=t_latent, x_0=noise_flat, x_1=z_flat)

        class _LatentPath:
            pass
        path_latent = _LatentPath()
        path_latent.x_t = path_latent_flat.x_t.reshape(B, G_sub, self.latent_dim)
        path_latent.dx_t = path_latent_flat.dx_t.reshape(B, G_sub, self.latent_dim)

        # 6. Model forward
        pred_v_expr, pred_v_latent = self.model(
            gene_input, source, path_expr.x_t, path_latent.x_t,
            t_expr, t_latent, perturbation_id,
        )

        # 7. Losses
        # Expression loss (unchanged from grn_att_only)
        loss_expr_per_sample = ((pred_v_expr - path_expr.dx_t) ** 2).mean(dim=-1)  # (B,)
        loss_expr = (loss_expr_per_sample * w_expr).sum() / w_expr.sum().clamp(min=1)

        # Latent loss — masked MSE over 128-dim SVD latent
        loss_elem = (pred_v_latent - path_latent.dx_t) ** 2  # (B, G_sub, 128)
        if self.use_variance_weight:
            loss_per_gene = (loss_elem * self.latent_dim_weights).sum(dim=-1)  # (B, G_sub)
        else:
            loss_per_gene = loss_elem.mean(dim=-1)  # (B, G_sub) — uniform over 128 dims
        loss_per_gene[:, missing_dev] = 0.0     # zero out missing genes
        n_valid = (~missing_dev).sum().clamp(min=1)
        loss_latent_per_sample = loss_per_gene.sum(dim=-1) / n_valid  # (B,)
        loss_latent = (loss_latent_per_sample * w_latent).sum() / w_latent.sum().clamp(min=1)

        loss = loss_expr + self.latent_weight * loss_latent

        # Optional MMD loss on expression
        _mmd_loss = torch.tensor(0.0, device=device)
        if self.use_mmd_loss and w_expr.sum() > 0:
            expr_mask = w_expr > 0
            if expr_mask.any():
                x1_hat = (
                    path_expr.x_t[expr_mask]
                    + pred_v_expr[expr_mask] * (1 - t_expr[expr_mask]).unsqueeze(-1)
                )
                sigmas = median_sigmas(target[expr_mask], scales=(0.5, 1.0, 2.0, 4.0))
                _mmd_loss = mmd2_unbiased_multi_sigma(x1_hat, target[expr_mask], sigmas)
                loss = loss + _mmd_loss * self.gamma

        return {
            "loss": loss,
            "loss_expr": loss_expr.detach(),
            "loss_latent": loss_latent.detach(),
            "loss_mmd": _mmd_loss.detach(),
        }

    @torch.no_grad()
    def generate(
        self,
        source: torch.Tensor,           # (B, G)
        perturbation_id: torch.Tensor,   # (B, 2)
        gene_ids: torch.Tensor,          # (B, G) or (G,)
        latent_steps: int = 20,
        expr_steps: int = 20,
        method: str = "rk4",
    ) -> torch.Tensor:
        """
        Two-stage cascaded generation in 128-dim latent space.

        Returns: (B, G) generated expression values
        """
        B, G = source.shape
        device = source.device

        if gene_ids.dim() == 1:
            gene_ids = gene_ids.unsqueeze(0).expand(B, -1)

        # 1D missing mask
        missing = self.sparse_cache.get_missing_gene_mask(torch.arange(G))

        # Initialize latent noise in 128-dim (NOT G×G!)
        z_t = torch.randn(B, G, self.latent_dim, device=device)
        if missing is not None:
            z_t[:, missing, :] = 0.0
        x_t = self._make_expr_noise(source)

        if method == "rk4":
            # === Stage 1: Latent generation (t_latent: 0->1, t_expr=0) ===
            t_zero = torch.zeros(B, device=device)
            t_one = torch.ones(B, device=device)

            def latent_vf(t, z):
                v_expr, v_latent = self.model(
                    gene_ids, source, x_t, z,
                    t_zero, t.expand(B), perturbation_id,
                )
                if missing is not None:
                    v_latent[:, missing, :] = 0.0
                return v_latent

            z_t = torchdiffeq.odeint(
                latent_vf, z_t,
                torch.linspace(0, 1, latent_steps + 1, device=device),
                method="rk4", atol=1e-4, rtol=1e-4,
            )[-1]

            # === Stage 2: Expression generation (t_expr: 0->1, t_latent=1) ===
            def expr_vf(t, x):
                v_expr, v_latent = self.model(
                    gene_ids, source, x, z_t,
                    t.expand(B), t_one, perturbation_id,
                )
                return v_expr

            x_t = torchdiffeq.odeint(
                expr_vf, x_t,
                torch.linspace(0, 1, expr_steps + 1, device=device),
                method="rk4", atol=1e-4, rtol=1e-4,
            )[-1]

        else:  # euler
            t_latent_schedule = torch.cat([
                torch.linspace(0, 1, latent_steps + 1, device=device),
                torch.ones(expr_steps, device=device),
            ])
            t_expr_schedule = torch.cat([
                torch.zeros(latent_steps + 1, device=device),
                torch.linspace(0, 1, expr_steps + 1, device=device)[1:],
            ])

            for i in range(latent_steps + expr_steps):
                t_lat = t_latent_schedule[i]
                t_lat_next = t_latent_schedule[i + 1]
                t_exp = t_expr_schedule[i]
                t_exp_next = t_expr_schedule[i + 1]

                v_expr, v_latent = self.model(
                    gene_ids, source, x_t, z_t,
                    t_exp.expand(B), t_lat.expand(B), perturbation_id,
                )

                x_t = x_t + (t_exp_next - t_exp) * v_expr
                z_t = z_t + (t_lat_next - t_lat) * v_latent
                if missing is not None:
                    z_t[:, missing, :] = 0.0

        return torch.clamp(x_t, min=0)