File size: 11,597 Bytes
9f5e507 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 | """
SBDenoiser — Anisotropic Schrödinger Bridge denoiser.
Training: Joint velocity + score matching with anisotropic bridge paths.
v_θ target = x_T - x₀ (PF-ODE velocity, same as scDFM).
s_θ target = -(x_t - μ_t) / var_t (conditional score).
Minibatch anisotropic OT per step.
Inference: Euler-Maruyama SDE using drift = v_θ + (σ²/2)·s_θ.
Or PF-ODE ablation: drift = v_θ.
"""
import math
import torch
import torch.nn as nn
import torchdiffeq
from ._scdfm_imports import make_lognorm_poisson_noise
from .model.model import SBModel
from .ot_anisotropic import AnisotropicOTSampler
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 SBDenoiser(nn.Module):
"""
Anisotropic Schrödinger Bridge Denoiser.
σ_g simultaneously controls:
1. OT coupling cost (Mahalanobis weights)
2. Bridge noise level (conditional bridge variance)
3. SDE diffusion strength (Euler-Maruyama noise)
"""
def __init__(
self,
model: SBModel,
noise_type: str = "Gaussian",
use_mmd_loss: bool = True,
gamma: float = 0.5,
poisson_alpha: float = 0.8,
poisson_target_sum: float = 1e4,
# Score training
score_weight: float = 0.1,
score_t_clip: float = 0.02,
use_score: bool = True,
# σ_g regularization
sigma_base: float = 0.5,
sigma_sparse_weight: float = 0.01,
sigma_volume_weight: float = 0.01,
# OT coupling
ot_method: str = "sinkhorn",
ot_reg: float = 0.05,
ot_use_sigma: bool = True,
sigma_min: float = 0.01,
# Time sampling
t_sample_mode: str = "logit_normal",
t_mean: float = 0.0,
t_std: float = 1.0,
# SDE inference
sde_steps: int = 50,
use_sde_inference: bool = True,
# Source-Anchored Bridge
source_anchored: bool = False,
):
super().__init__()
self.model = model
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.score_weight = score_weight
self.score_t_clip = score_t_clip
self.use_score = use_score
self.sigma_base = sigma_base
self.sigma_sparse_weight = sigma_sparse_weight
self.sigma_volume_weight = sigma_volume_weight
self.ot_use_sigma = ot_use_sigma
self.t_sample_mode = t_sample_mode
self.t_mean = t_mean
self.t_std = t_std
self.sde_steps = sde_steps
self.use_sde_inference = use_sde_inference
self.source_anchored = source_anchored
self.ot_sampler = AnisotropicOTSampler(
method=ot_method, reg=ot_reg, sigma_min=sigma_min,
)
def _make_noise(self, source: torch.Tensor) -> torch.Tensor:
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 _sample_t(self, n: int, device: torch.device) -> torch.Tensor:
if self.t_sample_mode == "logit_normal":
t = torch.sigmoid(torch.randn(n, device=device) * self.t_std + self.t_mean)
else:
t = torch.rand(n, device=device)
return t.clamp(self.score_t_clip, 1.0 - self.score_t_clip)
def train_step(
self,
source: torch.Tensor, # (B, G) control expression
target: torch.Tensor, # (B, G) perturbed expression
perturbation_id: torch.Tensor, # (B, n_pert)
gene_input: torch.Tensor, # (B, G) vocab-encoded gene IDs
) -> dict:
"""
Single training step with anisotropic bridge + minibatch OT.
"""
B = source.shape[0]
device = source.device
# 1. Sample time
t = self._sample_t(B, device) # (B,)
t_col = t.unsqueeze(-1) # (B, 1)
# 2. Get σ_g from sigma_net (independent of backbone)
# Need gene_emb and pert_emb — compute them via the model's encoder
with torch.no_grad():
gene_emb = self.model.encoder(gene_input) # (B, G, d)
pert_emb = self.model.get_perturbation_emb(
perturbation_id, cell_1=source) # (B, d)
# σ_g with gradient (for regularization loss)
sigma_g = self.model.sigma_net(pert_emb, t, gene_emb) # (B, G)
sigma_g_det = sigma_g.detach() # for bridge sampling
# 3. Create x_0 and do minibatch anisotropic OT
if self.source_anchored:
x_0 = source # bridge from control
else:
x_0 = self._make_noise(source) # bridge from noise
if self.ot_use_sigma:
sigma_for_ot = sigma_g_det.mean(0) # (G,) batch mean
x_0, target_matched = self.ot_sampler.sample_plan_fix_x0(
x_0, target, sigma_for_ot)
else:
x_0, target_matched = self.ot_sampler.sample_plan_fix_x0(
x_0, target, sigma_g=None)
# 4. Anisotropic conditional bridge sampling
mu_t = (1 - t_col) * x_0 + t_col * target_matched # (B, G)
var_t = (sigma_g_det ** 2 * (t_col * (1 - t_col))).clamp(min=1e-8)
std_t = torch.sqrt(var_t) # (B, G)
eps = torch.randn_like(x_0)
x_t = mu_t + std_t * eps # (B, G)
# 5. Targets
v_target = target_matched - x_0 # source-anchored: Δ
s_target = -eps / (std_t + 1e-8) # conditional score
# 6. Full model forward
pred_v, pred_s, sigma_g_pred = self.model(
gene_input, source, x_t, t, perturbation_id)
# 7. Velocity loss
loss_v = ((pred_v - v_target) ** 2).mean()
# 8. Score loss (var_t-weighted DSM — equivalent to ε-prediction)
# var_t weighting cancels the 1/var_t in s_target, giving bounded loss ~O(1)
loss_s = torch.tensor(0.0, device=device)
if self.use_score and pred_s is not None:
loss_s = (var_t * (pred_s - s_target) ** 2).mean()
# 9. σ_g regularization
# Volume penalty anchors geometric mean at σ_base (global scale).
# L1 sparse penalty removed — it killed per-gene anisotropy by
# pulling every σ_g to σ_base. Sigmoid [σ_min, σ_max] prevents
# collapse/explosion; volume penalty alone is sufficient.
loss_sparse = (sigma_g_pred - self.sigma_base).abs().mean() # monitor only
loss_volume = (sigma_g_pred.log().mean() - math.log(self.sigma_base)) ** 2
# 10. MMD loss (optional)
loss_mmd = torch.tensor(0.0, device=device)
if self.use_mmd_loss:
x1_hat = x_t + pred_v * (1 - t_col)
sigmas_mmd = median_sigmas(target_matched, scales=(0.5, 1.0, 2.0, 4.0))
loss_mmd = mmd2_unbiased_multi_sigma(x1_hat, target_matched, sigmas_mmd)
# 11. Total loss
# loss_sparse excluded — kept in return dict for monitoring
loss = (
loss_v
+ self.score_weight * loss_s
+ self.sigma_volume_weight * loss_volume
+ self.gamma * loss_mmd
)
return {
"loss": loss,
"loss_v": loss_v.detach(),
"loss_s": loss_s.detach(),
"loss_mmd": loss_mmd.detach(),
"loss_sparse": loss_sparse.detach(),
"loss_volume": loss_volume.detach(),
"sigma_mean": sigma_g_pred.mean().detach(),
"sigma_std": sigma_g_pred.std().detach(),
}
@torch.no_grad()
def generate(
self,
source: torch.Tensor, # (B, G)
perturbation_id: torch.Tensor, # (B, n_pert)
gene_ids: torch.Tensor, # (B, G) or (G,)
steps: int = None,
method: str = "sde",
) -> torch.Tensor:
"""
Generate perturbed expression via SDE or PF-ODE.
SDE: dX = [v_θ + (σ²/2)·s_θ] dt + σ·dB (Euler-Maruyama)
PF-ODE: dx/dt = v_θ (torchdiffeq RK4)
"""
B, G = source.shape
device = source.device
steps = steps or self.sde_steps
if gene_ids.dim() == 1:
gene_ids = gene_ids.unsqueeze(0).expand(B, -1)
if self.source_anchored:
x_0 = source.clone() # start from control
else:
x_0 = self._make_noise(source) # start from noise
use_sde = self.use_sde_inference and (method != "ode")
if use_sde:
# SDE: Euler-Maruyama (no high-order SDE solver available)
x_t = x_0
dt = 1.0 / steps
for i in range(steps):
t_val = i * dt
t = torch.full((B,), t_val, device=device)
pred_v, pred_s, sigma_g = self.model(
gene_ids, source, x_t, t, perturbation_id)
if pred_s is not None:
drift = pred_v + 0.5 * sigma_g ** 2 * pred_s
diffusion_noise = sigma_g * math.sqrt(dt) * torch.randn_like(x_t)
x_t = x_t + drift * dt + diffusion_noise
else:
x_t = x_t + pred_v * dt
else:
# PF-ODE: torchdiffeq RK4 (matches scDFM inference)
def ode_func(t_scalar, x):
t_batch = torch.full((B,), t_scalar.item(), device=device)
pred_v, _, _ = self.model(
gene_ids, source, x, t_batch, perturbation_id)
return pred_v
t_span = torch.linspace(0, 1, steps, device=device)
trajectory = torchdiffeq.odeint(
ode_func, x_0, t_span,
method="rk4", atol=1e-4, rtol=1e-4,
)
x_t = trajectory[-1]
return torch.clamp(x_t, min=0)
|