File size: 20,037 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
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
# Cascaded Conditioned Flow Matching (CCFM) — 实现计划

## Context

**目标**: 构建一个级联条件流匹配模型,用于单细胞扰动预测。以 scDFM 为底盘,引入 LatentForcing 的级联引导思想,使用 scGPT 提取的 **per-cell-per-gene contextualized features** 作为 latent 引导信号。

**核心概念映射**:
| LatentForcing (图像) | CCFM (单细胞) |
|---|---|
| Pixel patches → tokens | Gene expression → tokens |
| DINO-v2 feature patches → tokens | scGPT per-gene features → tokens |
| `pixel_embed + dino_embed` (element-wise add) | `expr_embed + scgpt_embed` (element-wise add) |
| `c = t_pixel + t_dino + y_emb` (AdaLN) | `c = t_expr + t_latent + pert_emb` (AdaLN) |
| Frozen DINO-v2 on-the-fly | Frozen scGPT on-the-fly |
| Separate pixel/dino decoder heads | Separate expr/latent decoder heads |

**数据集**: Norman (CRISPR) + ComboSciPlex (drug combination)

---

## 0. 项目结构

**工作目录**: `transfer/code/CCFM/`

```
transfer/code/CCFM/
├── plan.md                          # 本计划文件的副本
├── config/
│   └── config_cascaded.py           # CascadedFlowConfig 配置
├── src/
│   ├── model/
│   │   ├── __init__.py
│   │   ├── model.py                 # CascadedFlowModel 主干网络
│   │   └── layers.py                # 新增层: LatentEmbedder, LatentDecoder
│   ├── data/
│   │   ├── __init__.py
│   │   ├── data.py                  # 数据加载 (基于 scDFM, 新增 latent 支持)
│   │   └── scgpt_extractor.py       # 冻结 scGPT 特征提取器封装
│   ├── denoiser.py                  # CascadedDenoiser (时间采样、损失、生成)
│   └── utils.py                     # 工具函数
├── scripts/
│   ├── run_cascaded.py              # 训练/推理入口
│   └── download_scgpt.py            # 下载 scGPT 预训练模型
└── run.sh                           # pjsub 提交脚本模板
```

所有新代码写在 `transfer/code/CCFM/` 下,通过 import 引用 scDFM 和 scGPT 的现有模块。

---

## 1. 数据准备

### 1.1 下载 scGPT 预训练模型

- 下载 `whole-human` 模型到 `transfer/data/scGPT_pretrained/`
- 包含: `best_model.pt`, `vocab.json`, `args.json`
- 脚本: `scripts/download_scgpt.py`

### 1.2 数据加载 (`src/data/data.py`)

复用 scDFM 的 `Data` 类进行数据加载和 train/test 划分。修改 `PerturbationDataset` 以返回 **raw expression values** (不仅是 gene indices),供 scGPT on-the-fly 提取使用。

关键改动:
-`__init__` 中预计算 scDFM gene names → scGPT vocab ID 的映射表
- `__getitem__` 返回: `src_cell_data`, `tgt_cell_data`, `condition_id`, `src_cell_raw` (供 scGPT), `tgt_cell_raw` (供 scGPT)

### 1.3 scGPT 特征提取器 (`src/data/scgpt_extractor.py`)

**冻结 scGPT 作为 on-the-fly 特征提取器** (类似 LatentForcing 中冻结 DINO-v2):

```python
class FrozenScGPTExtractor(nn.Module):
    """
    封装冻结的 scGPT 模型,用于在训练时 on-the-fly 提取 per-gene features。
    类似 LatentForcing 的 dinov2_hf.py 中的 RAE 类。
    """
    def __init__(self, model_dir, n_hvg_genes, hvg_gene_names, device):
        # 1. 加载预训练 scGPT (TransformerModel)
        # 2. 冻结所有参数 (requires_grad=False)
        # 3. 建立 hvg_gene_names → scGPT vocab ID 映射
        # 4. 预计算全局归一化统计量 (running mean/var)

    @torch.no_grad()
    def extract(self, expression_values, gene_indices=None):
        """
        输入: expression_values (B, G) — G=infer_top_gene 的表达值
        输出: per_gene_features (B, G, scgpt_d_model) — contextualized 特征

        流程:
        1. 筛选非零基因,构建 scGPT 输入 (token IDs + values)
        2. 运行冻结 scGPT._encode() → (B, seq_len, d_model)
        3. 将输出 scatter 回固定的 G 个基因位置
        4. 未覆盖的位置用零向量填充
        5. 全局归一化 + 方差匹配
        """
```

**关键技术细节**:
- scGPT 的 `_encode()` 返回 `(batch, seq_len, d_model=512)`,其中 seq_len 是变长的(仅包含非零表达的基因 + `<cls>` token)
- 需要用 gene ID 映射将变长输出 **scatter** 回固定的 G 个基因位置
- 基因对齐: scDFM 使用 gene symbol (如 "MAPK1"),scGPT 使用自己的 vocab。在 `__init__` 中建立 `hvg_name → scGPT_vocab_id` 的映射
- 不在 scGPT vocab 中的基因: 用零向量替代

### 1.4 全局归一化与方差匹配 (CRITICAL)`FrozenScGPTExtractor` 内部实现:

```python
# 方案: 使用 running statistics (类似 BatchNorm 的 running_mean/var)
# 在前 N 个 batch 中收集统计量,之后固定

# Step 1: 标准化到零均值单位方差
z_normalized = (z_raw - running_mean) / (running_std + 1e-6)

# Step 2: 方差匹配 — 缩放到与表达嵌入相近的量级
z_matched = z_normalized * target_std  # target_std 默认 1.0
```

这对应 LatentForcing 中 `dinov2_hf.py` 的归一化:
```python
z = (z - latent_mean) / sqrt(latent_var + eps) * match_pixel_norm
```

---

## 2. 网络架构

### 2.1 核心设计 — 严格参照 LatentForcing CoT (`model_cot.py`)

**LatentForcing 的做法 (model_cot.py:398-446)**:
1. **双流嵌入 element-wise 相加** (line 414):
   ```python
   x = pixel_embedder(x_pixel) + dino_embedder(x_dino)
   ```
2. **条件向量求和通过 AdaLN** (line 409):
   ```python
   c = t_emb_pixel + t_emb_dino + y_emb
   ```
3. **共享 Transformer backbone**: 所有 block 处理融合后的 tokens
4. **分离解码头** (line 429-441): `dh_blocks_pixel` + `dh_blocks_dino` 分别解码

**CCFM 的对应实现**:
1. `x = expr_embedder(x_t_expr, gene_emb) + latent_embedder(z_t_latent)`
2. `c = t_expr_emb + t_latent_emb + pert_emb`
3. 复用 scDFM 的 DiffPerceiverBlock / PerceiverBlock
4. ExprDecoder (复用) + LatentDecoder (新增)

### 2.2 CascadedFlowModel 架构 (`src/model/model.py`)

```
输入:
  - gene_id:          (B, G)       基因 token IDs
  - cell_1 (source):  (B, G)       control 细胞表达值
  - x_t:              (B, G)       加噪后的扰动细胞表达值 (expression flow)
  - z_t:              (B, G, 512)  加噪后的 scGPT per-gene 特征 (latent flow)
  - t_expr:           (B,)         表达流的时间步
  - t_latent:         (B,)         潜变量流的时间步
  - perturbation_id:  (B, 2)       扰动 token IDs

数据流 (参照 model_cot.py 的 forward):
  ┌──────────────────────────────────────────────────────┐
  │ 1. 表达流嵌入 (复用 scDFM 的 origin 层)               │
  │    gene_emb = GeneEncoder(gene_id)         (B,G,d)   │
  │    val_emb_1 = ValueEncoder_1(cell_1) + gene_emb     │
  │    val_emb_2 = ValueEncoder_2(x_t) + gene_emb        │
  │    expr_tokens = fusion_layer(cat(val_1, val_2))      │
  │    → (B, G, d_model)                                  │
  ├──────────────────────────────────────────────────────┤
  │ 2. Latent 流嵌入 (新增, 对应 dino_embedder)           │
  │    latent_tokens = LatentEmbedder(z_t)                │
  │    → (B, G, d_model)     [bottleneck: 512 → d_model]  │
  ├──────────────────────────────────────────────────────┤
  │ 3. Element-wise 相加 (对应 model_cot.py line 414)     │
  │    x = expr_tokens + latent_tokens                    │
  │    → (B, G, d_model)                                  │
  ├──────────────────────────────────────────────────────┤
  │ 4. 条件向量 (对应 model_cot.py line 409)              │
  │    c = t_expr_emb + t_latent_emb + pert_emb          │
  │    → (B, d_model)                                     │
  ├──────────────────────────────────────────────────────┤
  │ 5. 共享 Backbone (复用 scDFM blocks)                  │
  │    for i, block in enumerate(blocks):                 │
  │      x = gene_adaLN[i](gene_emb, x)                  │
  │      x = cat([x, pert_emb.expand], dim=-1)            │
  │      x = adapter_layer[i](x)                          │
  │      x = block(x, val_emb_2, c)                       │
  │    → (B, G, d_model)                                  │
  ├──────────────────────────────────────────────────────┤
  │ 6a. 表达解码头 (复用 ExprDecoder)                      │
  │    → pred_v_expr: (B, G)                              │
  ├──────────────────────────────────────────────────────┤
  │ 6b. Latent 解码头 (新增, 对应 dh_blocks_dino)          │
  │    latent_feats = dh_latent_proj(x)                   │
  │    for block in dh_blocks_latent:                     │
  │      latent_feats = block(latent_feats, c)            │
  │    pred_v_latent = latent_final_layer(latent_feats)   │
  │    → (B, G, 512)                                      │
  └──────────────────────────────────────────────────────┘

输出: (pred_v_expr, pred_v_latent)
```

### 2.3 新增模块 (`src/model/layers.py`)

```python
class LatentEmbedder(nn.Module):
    """
    对应 LatentForcing 的 dino_embedder (BottleneckPatchEmbed)。
    将 per-gene scGPT 特征 (B, G, 512) 投影到 (B, G, d_model)。
    """
    def __init__(self, scgpt_dim=512, bottleneck_dim=128, d_model=512):
        self.proj = nn.Sequential(
            nn.Linear(scgpt_dim, bottleneck_dim),
            nn.GELU(),
            nn.Linear(bottleneck_dim, d_model),
        )

    def forward(self, z):  # z: (B, G, scgpt_dim)
        return self.proj(z)  # (B, G, d_model)

class LatentDecoder(nn.Module):
    """
    对应 LatentForcing 的 final_layer_dino + dh_blocks_dino。
    从 backbone 输出 (B, G, d_model) 解码回 (B, G, scgpt_dim)。
    """
    def __init__(self, d_model=512, scgpt_dim=512, dh_depth=2):
        # 可选: 额外的 decoder head blocks
        if dh_depth > 0:
            self.dh_blocks = nn.ModuleList([...])
        self.final = nn.Sequential(
            nn.LayerNorm(d_model),
            nn.Linear(d_model, d_model),
            nn.GELU(),
            nn.Linear(d_model, scgpt_dim),
        )
```

### 2.4 复用的现有模块

| 模块 | 来源 | 文件路径 |
|------|------|---------|
| `TimestepEmbedder` | scDFM | `scDFM/src/models/origin/layers.py:157` |
| `ContinuousValueEncoder` | scDFM | `scDFM/src/models/origin/layers.py:26` |
| `GeneEncoder` | scDFM | `scDFM/src/models/origin/layers.py:55` |
| `BatchLabelEncoder` | scDFM | `scDFM/src/models/origin/layers.py:98` |
| `ExprDecoder` | scDFM | `scDFM/src/models/origin/layers.py:116` |
| `GeneadaLN` | scDFM | `scDFM/src/models/origin/layers.py:10` |
| `DiffPerceiverBlock` / `PerceiverBlock` | scDFM | `scDFM/src/models/origin/model.py:55,92` |
| `modulate` | scDFM | `scDFM/src/models/origin/blocks.py` |
| `AffineProbPath` + `CondOTScheduler` | scDFM | `scDFM/src/flow_matching/path/` |
| `make_lognorm_poisson_noise` | scDFM | `scDFM/src/utils/utils.py` |
| `TransformerModel` | scGPT | `scGPT/scgpt/model/model.py` |
| `GeneVocab` | scGPT | `scGPT/scgpt/tokenizer/gene_tokenizer.py` |
| `load_pretrained` | scGPT | `scGPT/scgpt/utils/util.py` |
| `DataCollator` | scGPT | `scGPT/scgpt/data_collator.py` |

---

## 3. 训练逻辑

### 3.1 CascadedDenoiser (`src/denoiser.py`)

封装 `CascadedFlowModel` + `FrozenScGPTExtractor`### 3.2 级联时间步采样 — 参照 `denoiser_cot.py:121-132`

```python
def sample_t(self, n, device):
    """
    dino_first_cascaded 模式 (denoiser_cot.py line 121-132):
    - choose_latent_mask=True → 训练 latent 流
      t_latent 随机, t_expr=0, loss_weight_expr=0
    - choose_latent_mask=False → 训练 expression 流
      t_expr 随机, t_latent=1 (clean latent), loss_weight_latent=0
    """
    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 = torch.where(choose_latent_mask, t_latent, torch.ones_like(t_latent))
    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
```

### 3.3 训练步骤

```python
def train_step(self, source, target, perturbation_id, gene_ids):
    B = source.shape[0]

    # 1. On-the-fly 提取 scGPT per-gene 特征 (冻结)
    z_target = self.scgpt_extractor.extract(target)  # (B, G, 512)

    # 2. 时间采样
    t_expr, t_latent, w_expr, w_latent = self.sample_t(B, device)

    # 3. Expression path (同 scDFM)
    noise_expr = make_noise(source, noise_type)
    path_expr = affine_path.sample(t=t_expr, x_0=noise_expr, x_1=target)

    # 4. Latent path
    noise_latent = torch.randn_like(z_target)
    path_latent = affine_path.sample(
        t=t_latent.unsqueeze(-1).unsqueeze(-1),  # broadcast to (B,1,1)
        x_0=noise_latent, x_1=z_target
    )

    # 5. Model forward
    pred_v_expr, pred_v_latent = model(
        gene_ids, source, path_expr.x_t, path_latent.x_t,
        t_expr, t_latent, perturbation_id
    )

    # 6. Losses
    loss_expr = ((pred_v_expr - path_expr.dx_t)**2
                 * w_expr[:, None]).mean()
    loss_latent = ((pred_v_latent - path_latent.dx_t)**2
                   * w_latent[:, None, None]).mean()
    loss = loss_expr + latent_weight * loss_latent

    # Optional MMD (同 scDFM)
    if use_mmd_loss:
        x1_hat = path_expr.x_t + pred_v_expr * (1 - t_expr).unsqueeze(-1)
        loss += gamma * mmd_loss(x1_hat, target)

    return loss
```

### 3.4 训练入口 (`scripts/run_cascaded.py`)

基于 scDFM 的 `run.py` 改造:
- 初始化 `FrozenScGPTExtractor` + `CascadedFlowModel`
- 封装进 `CascadedDenoiser`
- 训练循环: 使用 `denoiser.train_step()` 替代 `train_step()`
- 评估: 使用 `denoiser.generate()` + `MetricsEvaluator`
- 其余(optimizer, scheduler, checkpoint)保持不变

---

## 4. 推理逻辑 — 两阶段级联生成

参照 LatentForcing `denoiser_cot.py:224-247`:

```python
@torch.no_grad()
def generate(self, source, perturbation_id, gene_ids,
             latent_steps=20, expr_steps=20):
    B = source.shape[0]
    G = source.shape[1]

    # ═══ Stage 1: 生成 Latent (per-gene scGPT features) ═══
    # t_latent: 0→1, t_expr=0 (表达流不参与)
    z_noise = torch.randn(B, G, scgpt_dim, device=device)

    def latent_vf(t, z_t):
        t_latent = t.expand(B)
        t_expr = torch.zeros(B, device=device)
        _, v_latent = model(gene_ids, source, source, z_t,
                            t_expr, t_latent, perturbation_id)
        return v_latent

    z_generated = torchdiffeq.odeint(
        latent_vf, z_noise,
        torch.linspace(0, 1, latent_steps, device=device),
        method='rk4'
    )[-1]  # (B, G, 512)

    # ═══ Stage 2: 用生成的 Latent 引导生成表达谱 ═══
    # t_expr: 0→1, t_latent=1 (latent "clean")
    expr_noise = make_noise(source, noise_type)

    def expr_vf(t, x_t):
        t_expr = t.expand(B)
        t_latent = torch.ones(B, device=device)
        v_expr, _ = model(gene_ids, source, x_t, z_generated,
                          t_expr, t_latent, perturbation_id)
        return v_expr

    x_generated = torchdiffeq.odeint(
        expr_vf, expr_noise,
        torch.linspace(0, 1, expr_steps, device=device),
        method='rk4'
    )[-1]  # (B, G)

    return torch.clamp(x_generated, min=0)
```

**关键设计** (与 LatentForcing 的 `dino_first_cascaded` 推理完全对应):
- Stage 1: `t_expr=0` → 表达流为纯噪声,模型聚焦 latent 预测
- Stage 2: `t_latent=1` → latent 已解析,`z_generated` 通过 `LatentEmbedder` 投影后与表达 tokens 相加,引导表达生成
- 评估: 复用 scDFM 的 `test()` 函数结构 + `cell-eval MetricsEvaluator`

---

## 5. 配置 (`config/config_cascaded.py`)

```python
@dataclass
class CascadedFlowConfig:
    # === 基础 (同 scDFM FlowConfig) ===
    model_type: str = 'cascaded'
    batch_size: int = 48
    d_model: int = 128
    nhead: int = 8
    nlayers: int = 4
    lr: float = 5e-5
    steps: int = 200000
    eta_min: float = 1e-6
    data_name: str = 'norman'
    fusion_method: str = 'differential_perceiver'
    perturbation_function: str = 'crisper'
    noise_type: str = 'Gaussian'
    infer_top_gene: int = 1000
    n_top_genes: int = 5000
    print_every: int = 5000
    gamma: float = 0.5
    use_mmd_loss: bool = True
    split_method: str = 'additive'
    fold: int = 1

    # === 级联/Latent 特有 ===
    scgpt_dim: int = 512              # scGPT embedding 维度
    bottleneck_dim: int = 128         # LatentEmbedder 瓶颈维度
    latent_weight: float = 1.0        # latent 流损失权重
    choose_latent_p: float = 0.4      # 训练 latent 流的概率
    target_std: float = 1.0           # 方差匹配目标标准差
    dh_depth: int = 2                 # latent decoder head 层数

    # === scGPT 路径 ===
    scgpt_model_dir: str = 'transfer/data/scGPT_pretrained'

    # === 推理 ===
    latent_steps: int = 20
    expr_steps: int = 20
```

---

## 6. 实现顺序

### Step 1: 项目初始化 + 下载 scGPT
- 创建 `transfer/code/CCFM/` 目录结构
- 下载 scGPT `whole-human` 预训练模型
- 验证: 加载 scGPT 模型成功

### Step 2: FrozenScGPTExtractor
- 实现 `src/data/scgpt_extractor.py`
- 处理 scDFM ↔ scGPT 的基因名对齐
- 实现 per-gene feature scatter 和归一化
- 验证: 输入 (48, 1000) expression → 输出 (48, 1000, 512) features

### Step 3: 数据加载
- 实现 `src/data/data.py`,复用 scDFM 的 Data 类
- 验证: DataLoader 正确返回 expression + perturbation_id

### Step 4: CascadedFlowModel
- 实现 `src/model/model.py` 和 `src/model/layers.py`
- 验证: dummy forward pass → (pred_v_expr, pred_v_latent) 形状正确

### Step 5: CascadedDenoiser
- 实现 `src/denoiser.py`
- 验证: `train_step()` 返回 scalar loss, `generate()` 返回正确形状

### Step 6: 训练入口 + 配置
- 实现 `scripts/run_cascaded.py` + `config/config_cascaded.py`
- 编写 `run.sh` (pjsub 脚本)
- 验证: 完整训练循环运行,loss 下降

### Step 7: 端到端验证
- Norman 数据集上训练(先小 step 数)
- 检查 loss_expr 和 loss_latent 分别下降
- 推理并通过 cell-eval 评估
- 与 baseline scDFM 对比

---

## 7. 关键技术决策

| 决策点 | 选择 | 理由 |
|--------|------|------|
| Latent 注入方式 | Element-wise add (token-level) | 严格对标 LatentForcing `model_cot.py:414` |
| 条件注入 | AdaLN (c = sum of all embeddings) | 对标 `model_cot.py:409` |
| scGPT 特征提取 | On-the-fly (冻结) | 存储不可行 (~1TB); 对标 LatentForcing 冻结 DINO-v2 |
| 时间步采样 | dino_first_cascaded | 对标 `denoiser_cot.py:121-132` |
| 方差匹配 | 全局 running statistics + scaling | 对标 `dinov2_hf.py` 的归一化 |
| Backbone | 复用 scDFM 的 DiffPerceiver/Perceiver | 保持底盘一致 |
| 解码头 | 分离: ExprDecoder + LatentDecoder | 对标 `model_cot.py:429-441` 的分离头 |