File size: 16,257 Bytes
e093a4b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
# Copyright (C) Tahoe Therapeutics 2025. All rights reserved.
"""
Standalone implementation of TXModel blocks without external dependencies.
Only requires: torch, transformers
"""

import math
from typing import Optional, Dict, Any, Tuple

import torch
import torch.nn.functional as F
from torch import Tensor, nn


class MultiheadAttention(nn.Module):
    """Standard multi-head attention implementation"""
    
    def __init__(
        self,
        d_model: int,
        n_heads: int,
        kv_n_heads: Optional[int] = None,
        dropout: float = 0.0,
        bias: bool = True,
        device: Optional[str] = None,
    ):
        super().__init__()
        self.d_model = d_model
        self.n_heads = n_heads
        self.kv_n_heads = kv_n_heads if kv_n_heads is not None else n_heads
        self.head_dim = d_model // n_heads
        self.dropout = dropout
        
        # Grouped Query Attention support
        self.n_rep = n_heads // self.kv_n_heads
        
        self.q_proj = nn.Linear(d_model, d_model, bias=bias, device=device)
        self.k_proj = nn.Linear(d_model, self.kv_n_heads * self.head_dim, bias=bias, device=device)
        self.v_proj = nn.Linear(d_model, self.kv_n_heads * self.head_dim, bias=bias, device=device)
        self.out_proj = nn.Linear(d_model, d_model, bias=bias, device=device)
        
        self.attn_dropout = nn.Dropout(dropout)
        
    def forward(
        self,
        x: Tensor,
        attn_bias: Optional[Tensor] = None,
        key_padding_mask: Optional[Tensor] = None,
        is_causal: bool = False,
        **kwargs
    ) -> Tuple[Tensor, None, None]:
        batch_size, seq_len, _ = x.shape
        
        # Project queries, keys, values
        q = self.q_proj(x).view(batch_size, seq_len, self.n_heads, self.head_dim)
        k = self.k_proj(x).view(batch_size, seq_len, self.kv_n_heads, self.head_dim)
        v = self.v_proj(x).view(batch_size, seq_len, self.kv_n_heads, self.head_dim)
        
        # Transpose for attention: (batch, heads, seq, head_dim)
        q = q.transpose(1, 2)
        k = k.transpose(1, 2)
        v = v.transpose(1, 2)
        
        # Repeat k/v for grouped query attention
        if self.n_rep > 1:
            k = k.repeat_interleave(self.n_rep, dim=1)
            v = v.repeat_interleave(self.n_rep, dim=1)
        
        # Scaled dot-product attention
        scale = 1.0 / math.sqrt(self.head_dim)
        attn_scores = torch.matmul(q, k.transpose(-2, -1)) * scale
        
        # Apply attention bias if provided
        if attn_bias is not None:
            attn_scores = attn_scores + attn_bias
        
        # Apply key padding mask
        if key_padding_mask is not None:
            # key_padding_mask: (batch, seq_len) with True for valid positions
            # Convert to attention mask: (batch, 1, 1, seq_len)
            mask = key_padding_mask.unsqueeze(1).unsqueeze(2)
            attn_scores = attn_scores.masked_fill(~mask, float('-inf'))
        
        # Apply causal mask if needed
        if is_causal:
            causal_mask = torch.triu(
                torch.ones(seq_len, seq_len, device=x.device, dtype=torch.bool),
                diagonal=1
            )
            attn_scores = attn_scores.masked_fill(causal_mask, float('-inf'))
        
        # Softmax and dropout
        attn_weights = F.softmax(attn_scores, dim=-1)
        attn_weights = self.attn_dropout(attn_weights)
        
        # Apply attention to values
        output = torch.matmul(attn_weights, v)
        
        # Reshape and project output
        output = output.transpose(1, 2).contiguous().view(batch_size, seq_len, self.d_model)
        output = self.out_proj(output)
        
        return output, None, None


class TXBlock(nn.Module):
    """Transformer encoder block with pre/post normalization support"""
    
    def __init__(
        self,
        d_model: int,
        n_heads: int,
        expansion_ratio: int,
        attn_config: Optional[Dict] = None,
        norm_config: Optional[Dict] = None,
        dropout: Optional[float] = 0.0,
        activation: Optional[str] = "gelu",
        device: Optional[str] = None,
        norm_scheme: str = "pre",
        use_glu: bool = False,
        **kwargs: Any,
    ) -> None:
        super().__init__()
        
        if attn_config is None:
            attn_config = {}
        if norm_config is None:
            norm_config = {}
        
        self.d_model = d_model
        self.n_heads = n_heads
        self.device = device
        self.norm_scheme = norm_scheme
        self.use_glu = use_glu
        
        # Attention
        kv_n_heads = attn_config.get("kv_n_heads", n_heads)
        self.self_attn = MultiheadAttention(
            d_model=d_model,
            n_heads=n_heads,
            kv_n_heads=kv_n_heads,
            dropout=attn_config.get("attn_pdrop", 0.0),
            device=device,
        )
        
        # FFN
        dim_feedforward = d_model * expansion_ratio
        self.up_proj = nn.Linear(d_model, dim_feedforward, device=device)
        self.down_proj = nn.Linear(dim_feedforward, d_model, device=device)
        
        if use_glu:
            self.gate_proj = nn.Linear(d_model, dim_feedforward, device=device)
        
        # Normalization
        eps = norm_config.get("eps", 1e-5)
        self.norm1 = nn.LayerNorm(d_model, eps=eps, device=device)
        self.norm2 = nn.LayerNorm(d_model, eps=eps, device=device)
        
        # Dropout
        self.post_sa_dropout = nn.Dropout(dropout)
        self.post_ffn_dropout = nn.Dropout(dropout)
        
        # Activation
        self.activation = self._get_activation_fn(activation)
        
        if norm_scheme not in ["pre", "post"]:
            raise ValueError("norm_scheme must be either pre or post")
    
    @staticmethod
    def _get_activation_fn(activation: str):
        if activation == "gelu":
            return nn.GELU()
        elif activation == "relu":
            return nn.ReLU()
        elif activation == "silu" or activation == "swish":
            return nn.SiLU()
        elif activation == "leaky_relu":
            return nn.LeakyReLU()
        else:
            raise ValueError(f"Unknown activation: {activation}")
    
    def forward(
        self,
        x: Tensor,
        attn_bias: Optional[Tensor] = None,
        key_padding_mask: Optional[Tensor] = None,
        **kwargs
    ) -> Tensor:
        
        if self.norm_scheme == "pre":
            # Pre-norm: norm -> attention -> add
            x = x + self._sa_block(
                self.norm1(x),
                attn_bias=attn_bias,
                key_padding_mask=key_padding_mask,
            )
            x = x + self._ff_block(self.norm2(x))
        else:
            # Post-norm: attention -> add -> norm
            x = self.norm1(
                x + self._sa_block(
                    x,
                    attn_bias=attn_bias,
                    key_padding_mask=key_padding_mask,
                )
            )
            x = self.norm2(x + self._ff_block(x))
        
        return x
    
    def _sa_block(
        self,
        x: Tensor,
        attn_bias: Optional[Tensor] = None,
        key_padding_mask: Optional[Tensor] = None,
    ) -> Tensor:
        x, _, _ = self.self_attn(
            x,
            attn_bias=attn_bias,
            key_padding_mask=key_padding_mask,
            is_causal=False,
        )
        return self.post_sa_dropout(x)
    
    def _ff_block(self, x: Tensor) -> Tensor:
        if self.use_glu:
            # GLU variant: (gate * activation(x)) * up(x)
            x = self.down_proj(self.activation(self.gate_proj(x)) * self.up_proj(x))
        else:
            # Standard FFN
            x = self.down_proj(self.activation(self.up_proj(x)))
        return self.post_ffn_dropout(x)


class TXEncoder(nn.Module):
    """Stack of transformer encoder layers"""
    
    def __init__(
        self,
        encoder_layer: TXBlock,
        num_layers: int,
        use_norm: bool = False,
        norm_config: Optional[Dict] = None,
        attn_config: Optional[Dict] = None,
    ):
        super().__init__()
        
        if norm_config is None:
            norm_config = {}
        
        # Clone the layer
        self.layers = nn.ModuleList([
            TXBlock(
                d_model=encoder_layer.d_model,
                n_heads=encoder_layer.n_heads,
                expansion_ratio=encoder_layer.up_proj.out_features // encoder_layer.d_model,
                attn_config=attn_config,
                norm_config=norm_config,
                activation="gelu",
                device=encoder_layer.device,
                norm_scheme=encoder_layer.norm_scheme,
                use_glu=encoder_layer.use_glu,
            )
            for _ in range(num_layers)
        ])
        
        self.use_norm = use_norm
        if use_norm:
            eps = norm_config.get("eps", 1e-5)
            self.norm = nn.LayerNorm(encoder_layer.d_model, eps=eps)
    
    def forward(
        self,
        total_embs: Tensor,
        key_padding_mask: Optional[Tensor] = None,
        output_hidden_states: bool = False,
    ) -> Tuple[Tensor, Optional[list]]:
        
        x = total_embs
        hidden_states = [] if output_hidden_states else None
        
        for layer in self.layers:
            x = layer(
                x,
                attn_bias=None,
                key_padding_mask=key_padding_mask,
            )
            
            if output_hidden_states:
                hidden_states.append(x)
        
        if self.use_norm:
            x = self.norm(x)
        
        return x, hidden_states


class GeneEncoder(nn.Module):
    """Gene embedding with optional extra embeddings"""
    
    def __init__(
        self,
        num_embeddings: int,
        embedding_dim: int,
        padding_idx: int = 0,
        use_norm: bool = False,
        gene_encoder_cfg: Optional[Dict] = None,
    ):
        super().__init__()
        
        if gene_encoder_cfg is None:
            gene_encoder_cfg = {}
        
        self.use_norm = use_norm
        self.embedding = nn.Embedding(
            num_embeddings,
            embedding_dim,
            padding_idx=padding_idx,
        )
        
        # For now, no extra embeddings in standalone version
        self.project = nn.Identity()
        
        if self.use_norm:
            self.enc_norm = nn.LayerNorm(embedding_dim)
    
    def forward(self, x: Tensor) -> Tensor:
        x = self.embedding(x)
        x = self.project(x)
        if self.use_norm:
            x = self.enc_norm(x)
        return x


class ChemEncoder(nn.Module):
    """Chemical compound encoder"""
    
    def __init__(
        self,
        d_out: int,
        padding_idx: int = 0,
        activation: str = "leaky_relu",
        use_norm: bool = True,
        freeze: bool = False,
        num_drugs: int = 1000,
        fp_dim: int = 2048,
    ):
        super().__init__()
        
        # Initialize with zeros (user should load pretrained weights)
        drug_fps = torch.zeros((num_drugs, fp_dim), dtype=torch.float32)
        
        self.embedding = nn.Embedding.from_pretrained(
            drug_fps,
            padding_idx=padding_idx,
            freeze=freeze,
        )
        
        self.fc = nn.Linear(fp_dim, d_out)
        
        if activation == "leaky_relu":
            self.activation = nn.LeakyReLU()
        elif activation == "relu":
            self.activation = nn.ReLU()
        elif activation == "gelu":
            self.activation = nn.GELU()
        else:
            self.activation = nn.ReLU()
        
        self.proj = nn.Linear(d_out, d_out)
        
        self.use_norm = use_norm
        if self.use_norm:
            self.norm = nn.LayerNorm(d_out)
    
    def forward(self, x: Tensor) -> Tensor:
        x = self.embedding(x)
        x = self.activation(self.fc(x))
        x = self.proj(x)
        
        if self.use_norm:
            x = self.norm(x)
        return x


class ContinuousValueEncoder(nn.Module):
    """Encode continuous values to embeddings"""
    
    def __init__(
        self,
        d_model: int,
        dropout: float = 0.1,
        max_value: int = 512,
        activation: str = "relu",
        use_norm: bool = False,
    ):
        super().__init__()
        
        self.dropout = nn.Dropout(p=dropout)
        self.linear1 = nn.Linear(1, d_model)
        
        if activation == "relu":
            self.activation = nn.ReLU()
        elif activation == "gelu":
            self.activation = nn.GELU()
        elif activation == "leaky_relu":
            self.activation = nn.LeakyReLU()
        else:
            self.activation = nn.ReLU()
        
        self.linear2 = nn.Linear(d_model, d_model)
        
        self.use_norm = use_norm
        if self.use_norm:
            self.norm = nn.LayerNorm(d_model)
        
        self.max_value = max_value
    
    def forward(self, x: Tensor) -> Tensor:
        # Expand last dimension
        x = x.unsqueeze(-1)
        # Clip to max value
        x = torch.clamp(x, max=self.max_value)
        # Project
        x = self.activation(self.linear1(x))
        x = self.linear2(x)
        if self.use_norm:
            x = self.norm(x)
        return self.dropout(x)


class ExprDecoder(nn.Module):
    """Expression value decoder"""
    
    def __init__(
        self,
        d_model: int,
        n_outputs: int = 1,
        n_layers: int = 2,
        activation: str = "leaky_relu",
    ):
        super().__init__()
        
        if activation == "leaky_relu":
            self.activation = nn.LeakyReLU()
        elif activation == "relu":
            self.activation = nn.ReLU()
        elif activation == "gelu":
            self.activation = nn.GELU()
        else:
            self.activation = nn.LeakyReLU()
        
        self.linear_layers = nn.ModuleList(
            [nn.Linear(d_model, d_model) for _ in range(n_layers)]
        )
        self.out_proj = nn.Linear(d_model, n_outputs)
    
    def forward(self, x: Tensor) -> Dict[str, Tensor]:
        for layer in self.linear_layers:
            x = self.activation(layer(x))
        pred_value = self.out_proj(x)
        if pred_value.shape[-1] == 1:
            pred_value = pred_value.squeeze(-1)
        return {"pred": pred_value}


class MVCDecoder(nn.Module):
    """Masked value prediction decoder"""
    
    def __init__(
        self,
        d_model: int,
        arch_style: str = "inner product",
        query_activation: str = "sigmoid",
        scaled_dot_product: bool = False,
    ) -> None:
        super().__init__()
        
        self.scaled_dot_product = scaled_dot_product
        
        if arch_style == "inner product":
            self.gene2query = nn.Linear(d_model, d_model)
            
            if query_activation == "sigmoid":
                self.query_activation = nn.Sigmoid()
            elif query_activation == "relu":
                self.query_activation = nn.ReLU()
            elif query_activation == "tanh":
                self.query_activation = nn.Tanh()
            else:
                self.query_activation = nn.Sigmoid()
            
            self.W = nn.Linear(d_model, d_model, bias=False)
        else:
            raise ValueError(f"Unknown arch_style: {arch_style}")
        
        self.arch_style = arch_style
    
    def forward(
        self,
        cell_emb: Tensor,
        gene_embs: Tensor,
    ) -> Dict[str, Tensor]:
        
        if self.arch_style == "inner product":
            query_vecs = self.query_activation(
                self.gene2query(gene_embs)
            )
            inner_product_dimension = query_vecs.shape[-1]
            cell_emb = cell_emb.unsqueeze(2)
            pred_value = torch.bmm(self.W(query_vecs), cell_emb).squeeze(2)
            
            if self.scaled_dot_product:
                pred_value = pred_value / torch.sqrt(
                    torch.tensor(inner_product_dimension, dtype=pred_value.dtype)
                )
            
            return {"pred": pred_value}
        else:
            raise ValueError(f"Unknown arch_style: {self.arch_style}")