File size: 12,490 Bytes
0161e74
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
"""
CascadedDenoiser — Wraps CascadedFlowModel + FrozenScGPTExtractor.

Implements cascaded time-step sampling (from LatentForcing denoiser_cot.py:121-132)
and two-stage cascaded generation (denoiser_cot.py:224-247).
"""

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.scgpt_extractor import FrozenScGPTExtractor


# 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 combining CascadedFlowModel with FrozenScGPTExtractor.

    Training: Cascaded time-step sampling (dino_first_cascaded mode).
    Inference: Two-stage cascaded generation (latent first, then expression).
    """

    def __init__(
        self,
        model: CascadedFlowModel,
        scgpt_extractor: FrozenScGPTExtractor,
        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 (LatentForcing dino_first_cascaded_noised)
        noise_beta: float = 0.25,
    ):
        super().__init__()
        self.model = model
        self.scgpt_extractor = scgpt_extractor
        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

    def sample_t(self, n: int, device: torch.device):
        """
        Cascaded time-step sampling — dino_first_cascaded mode.
        (LatentForcing denoiser_cot.py:121-132)

        Supports uniform or logit-normal sampling per flow.

        With probability choose_latent_p:
            - Train latent flow: t_latent sampled, t_expr = 0, loss_weight_expr = 0
        Otherwise:
            - Train expression flow: t_expr sampled, t_latent = 1 (clean), loss_weight_latent = 0
        """
        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

        # When training expr flow: t_latent ~ U[1-noise_beta, 1] (LatentForcing dino_first_cascaded_noised)
        # noise_beta=0 recovers original behavior (t_latent=1, no noise)
        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_full)
        target: torch.Tensor,         # (B, G_full)
        perturbation_id: torch.Tensor,  # (B, 2)
        gene_ids: torch.Tensor,       # (G_full,)
        infer_top_gene: int = 1000,
        cached_z_target: torch.Tensor = None,  # (B, G_sub, scgpt_dim) pre-extracted
        cached_gene_ids: torch.Tensor = None,  # (G_sub,) gene indices used for cache
    ) -> torch.Tensor:
        """
        Single training step with cascaded time sampling.
        Returns: scalar loss
        """
        B = source.shape[0]
        device = source.device

        # Random gene subset (same as scDFM)
        if cached_gene_ids is not None:
            input_gene_ids = cached_gene_ids
        else:
            input_gene_ids = torch.randperm(source.shape[-1], device=device)[:infer_top_gene]
        source_sub = source[:, input_gene_ids]
        target_sub = target[:, input_gene_ids]
        gene_input = gene_ids[input_gene_ids].unsqueeze(0).expand(B, -1)

        # 1. scGPT features: use cache if available, otherwise extract on-the-fly
        if cached_z_target is not None:
            z_target = cached_z_target
        else:
            z_target = self.scgpt_extractor.extract(target_sub, gene_indices=input_gene_ids)  # (B, G, scgpt_dim)

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

        # 3. Expression flow path
        noise_expr = self._make_expr_noise(source_sub)
        path_expr = flow_path.sample(t=t_expr, x_0=noise_expr, x_1=target_sub)

        # 4. Latent flow path
        # AffineProbPath.sample expects t as (B,) and broadcasts internally.
        # But z_target is (B, G, scgpt_dim) — we need to flatten to 2D, sample, then reshape.
        B_l, G_l, D_l = z_target.shape
        noise_latent = torch.randn_like(z_target)
        z_target_flat = z_target.reshape(B_l, G_l * D_l)
        noise_latent_flat = noise_latent.reshape(B_l, G_l * D_l)
        path_latent_flat = flow_path.sample(t=t_latent, x_0=noise_latent_flat, x_1=z_target_flat)

        # Wrap path_latent with reshaped tensors
        class _LatentPath:
            pass
        path_latent = _LatentPath()
        path_latent.x_t = path_latent_flat.x_t.reshape(B_l, G_l, D_l)
        path_latent.dx_t = path_latent_flat.dx_t.reshape(B_l, G_l, D_l)

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

        # 6. Losses — per-sample MSE first, then weight by cascaded mask
        # (aligns with LatentForcing: eliminates dimension mismatch, decouples choose_latent_p from loss weight)
        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)

        loss_latent_per_sample = ((pred_v_latent - path_latent.dx_t) ** 2).mean(dim=(-1, -2))  # (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 (same as scDFM)
        _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_sub[expr_mask], scales=(0.5, 1.0, 2.0, 4.0))
                _mmd_loss = mmd2_unbiased_multi_sigma(x1_hat, target_sub[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.

        method="euler": Single-loop joint Euler steps (LatentForcing style).
        method="rk4":   Two-stage serial ODE with torchdiffeq RK4 (scDFM style, higher accuracy).

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

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

        # === Initialize both noise states ===
        z_t = torch.randn(B, G, scgpt_dim, device=device)
        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,
                )
                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 — joint loop (LatentForcing style)
            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

        return torch.clamp(x_t, min=0)