File size: 14,243 Bytes
c6be546 | 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 | """
Text prompt decoder implementation for VoxTell.
Code modified from DETR transformer:
https://github.com/facebookresearch/detr
Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
"""
import copy
from typing import Callable, List, Optional, Tuple, Union
import torch
import torch.nn.functional as F
from torch import nn, Tensor
class TransformerDecoder(nn.Module):
"""
Transformer decoder consisting of multiple decoder layers.
This decoder processes target sequences with attention to memory (encoder output),
optionally returning intermediate layer outputs. It receives the text prompt embeddings
as queries and attends to image features from the encoder. It outputs refined text-image
fused features for segmentation mask prediction.
Args:
decoder_layer: A single transformer decoder layer to be cloned.
num_layers: Number of decoder layers to stack.
norm: Optional normalization layer applied to the final output.
return_intermediate: If True, returns outputs from all layers stacked together.
"""
def __init__(
self,
decoder_layer: nn.Module,
num_layers: int,
norm: Optional[nn.Module] = None,
return_intermediate: bool = False
) -> None:
super().__init__()
self.layers = _get_clones(decoder_layer, num_layers)
self.num_layers = num_layers
self.norm = norm
self.return_intermediate = return_intermediate
def forward(
self,
tgt: Tensor,
memory: Tensor,
tgt_mask: Optional[Tensor] = None,
memory_mask: Optional[Tensor] = None,
tgt_key_padding_mask: Optional[Tensor] = None,
memory_key_padding_mask: Optional[Tensor] = None,
pos: Optional[Tensor] = None,
query_pos: Optional[Tensor] = None
) -> Union[Tensor, Tuple[Tensor, List[Tensor]]]:
"""
Forward pass through all decoder layers.
Args:
tgt: Target sequence tensor of shape [T, B, C].
memory: Memory (encoder output) tensor of shape [T, B, C].
tgt_mask: Attention mask for target self-attention.
memory_mask: Attention mask for cross-attention to memory.
tgt_key_padding_mask: Padding mask for target keys.
memory_key_padding_mask: Padding mask for memory keys.
pos: Positional embeddings for memory.
query_pos: Positional embeddings for queries.
Returns:
If return_intermediate is True, returns stacked intermediate outputs.
Otherwise, returns tuple of (final_output, attention_weights_list).
"""
output = tgt
T, B, C = memory.shape
intermediate = []
atten_layers = []
for n, layer in enumerate(self.layers):
residual = True
output, ws = layer(
output, memory,
tgt_mask=tgt_mask,
memory_mask=memory_mask,
tgt_key_padding_mask=tgt_key_padding_mask,
memory_key_padding_mask=memory_key_padding_mask,
pos=pos,
query_pos=query_pos,
residual=residual
)
atten_layers.append(ws)
if self.return_intermediate:
intermediate.append(self.norm(output))
if self.norm is not None:
output = self.norm(output)
if self.return_intermediate:
intermediate.pop()
intermediate.append(output)
if self.return_intermediate:
return torch.stack(intermediate)
return output, atten_layers
class TransformerDecoderLayer(nn.Module):
"""
Single transformer decoder layer with self-attention, cross-attention, and FFN.
This layer implements:
1. Self-attention on the target sequence
2. Cross-attention between target and memory (image encoder output)
3. Position-wise feed-forward network
Args:
d_model: Dimension of the model (embedding dimension).
nhead: Number of attention heads.
dim_feedforward: Dimension of the feedforward network.
dropout: Dropout probability.
activation: Activation function name ('relu', 'gelu', or 'glu').
normalize_before: If True, applies layer norm before attention/FFN (pre-norm).
If False, applies after (post-norm).
"""
def __init__(
self,
d_model: int,
nhead: int,
dim_feedforward: int = 2048,
dropout: float = 0.1,
activation: str = "relu",
normalize_before: bool = False
) -> None:
super().__init__()
self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=dropout)
self.multihead_attn = nn.MultiheadAttention(d_model, nhead, dropout=dropout)
# Feed-forward network
self.linear1 = nn.Linear(d_model, dim_feedforward)
self.dropout = nn.Dropout(dropout)
self.linear2 = nn.Linear(dim_feedforward, d_model)
# Normalization layers
self.norm1 = nn.LayerNorm(d_model)
self.norm2 = nn.LayerNorm(d_model)
self.norm3 = nn.LayerNorm(d_model)
# Dropout layers
self.dropout1 = nn.Dropout(dropout)
self.dropout2 = nn.Dropout(dropout)
self.dropout3 = nn.Dropout(dropout)
self.activation = _get_activation_fn(activation)
self.normalize_before = normalize_before
def with_pos_embed(self, tensor: Tensor, pos: Optional[Tensor]) -> Tensor:
"""
Add positional embeddings to tensor if provided.
Args:
tensor: Input tensor.
pos: Optional positional embeddings to add.
Returns:
Tensor with positional embeddings added, or original tensor if pos is None.
"""
return tensor if pos is None else tensor + pos
def forward_post(
self,
tgt: Tensor,
memory: Tensor,
tgt_mask: Optional[Tensor] = None,
memory_mask: Optional[Tensor] = None,
tgt_key_padding_mask: Optional[Tensor] = None,
memory_key_padding_mask: Optional[Tensor] = None,
pos: Optional[Tensor] = None,
query_pos: Optional[Tensor] = None,
residual: bool = True
) -> Tuple[Tensor, Tensor]:
"""
Post-norm forward pass: attention/FFN first, then normalization.
Args:
tgt: Target sequence.
memory: Memory (encoder output).
tgt_mask: Self-attention mask for target.
memory_mask: Cross-attention mask.
tgt_key_padding_mask: Padding mask for target keys.
memory_key_padding_mask: Padding mask for memory keys.
pos: Positional embeddings for memory.
query_pos: Positional embeddings for queries.
residual: Whether to use residual connections.
Returns:
Tuple of (output_tensor, attention_weights).
"""
q = k = self.with_pos_embed(tgt, query_pos)
tgt2, ws = self.self_attn(
q, k, value=tgt,
attn_mask=tgt_mask,
key_padding_mask=tgt_key_padding_mask
)
tgt = self.norm1(tgt)
tgt2, ws = self.multihead_attn(
query=self.with_pos_embed(tgt, query_pos),
key=self.with_pos_embed(memory, pos),
value=memory,
attn_mask=memory_mask,
key_padding_mask=memory_key_padding_mask
)
# Cross-attention with residual connection
tgt = tgt + self.dropout2(tgt2)
tgt = self.norm2(tgt)
# Feed-forward network
tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt))))
tgt = tgt + self.dropout3(tgt2)
tgt = self.norm3(tgt)
return tgt, ws
def forward_pre(
self,
tgt: Tensor,
memory: Tensor,
tgt_mask: Optional[Tensor] = None,
memory_mask: Optional[Tensor] = None,
tgt_key_padding_mask: Optional[Tensor] = None,
memory_key_padding_mask: Optional[Tensor] = None,
pos: Optional[Tensor] = None,
query_pos: Optional[Tensor] = None
) -> Tuple[Tensor, Tensor]:
"""
Pre-norm forward pass: normalization first, then attention/FFN.
Args:
tgt: Target sequence.
memory: Memory (encoder output).
tgt_mask: Self-attention mask for target.
memory_mask: Cross-attention mask.
tgt_key_padding_mask: Padding mask for target keys.
memory_key_padding_mask: Padding mask for memory keys.
pos: Positional embeddings for memory.
query_pos: Positional embeddings for queries.
Returns:
Tuple of (output_tensor, attention_weights).
"""
tgt2 = self.norm2(tgt)
tgt2, attn_weights = self.multihead_attn(
query=self.with_pos_embed(tgt2, query_pos),
key=self.with_pos_embed(memory, pos),
value=memory,
attn_mask=memory_mask,
key_padding_mask=memory_key_padding_mask
)
tgt = tgt + self.dropout2(tgt2)
# Feed-forward network with pre-norm
tgt2 = self.norm3(tgt)
tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt2))))
tgt = tgt + self.dropout3(tgt2)
return tgt, attn_weights
def forward_pre_selfattention(
self,
tgt: Tensor,
memory: Tensor,
tgt_mask: Optional[Tensor] = None,
memory_mask: Optional[Tensor] = None,
tgt_key_padding_mask: Optional[Tensor] = None,
memory_key_padding_mask: Optional[Tensor] = None,
pos: Optional[Tensor] = None,
query_pos: Optional[Tensor] = None
) -> Tuple[Tensor, Tensor]:
"""
Alternative forward pass with cross-attention before self-attention.
This variant applies operations in the order:
1. Cross-attention (without normalization)
2. Self-attention (with pre-norm)
3. Feed-forward network (with pre-norm)
Args:
tgt: Target sequence.
memory: Memory (encoder output).
tgt_mask: Self-attention mask for target.
memory_mask: Cross-attention mask.
tgt_key_padding_mask: Padding mask for target keys.
memory_key_padding_mask: Padding mask for memory keys.
pos: Positional embeddings for memory.
query_pos: Positional embeddings for queries.
Returns:
Tuple of (output_tensor, attention_weights).
"""
# Cross-attention without pre-normalization
tgt2, attn_weights = self.multihead_attn(
query=self.with_pos_embed(tgt, query_pos),
key=self.with_pos_embed(memory, pos),
value=memory,
attn_mask=memory_mask,
key_padding_mask=memory_key_padding_mask
)
tgt = tgt + tgt2
# Self-attention with pre-norm
tgt2 = self.norm1(tgt)
q = k = self.with_pos_embed(tgt2, query_pos)
tgt2, ws = self.self_attn(q, k, value=tgt2)
tgt = tgt + tgt2
# Feed-forward network with pre-norm
tgt2 = self.norm2(tgt)
tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt2))))
tgt = tgt + tgt2
tgt = self.norm3(tgt)
return tgt, attn_weights
def forward(
self,
tgt: Tensor,
memory: Tensor,
tgt_mask: Optional[Tensor] = None,
memory_mask: Optional[Tensor] = None,
tgt_key_padding_mask: Optional[Tensor] = None,
memory_key_padding_mask: Optional[Tensor] = None,
pos: Optional[Tensor] = None,
query_pos: Optional[Tensor] = None,
residual: bool = True
) -> Tuple[Tensor, Tensor]:
"""
Forward pass through the decoder layer.
Dispatches to either pre-norm or post-norm variant based on configuration.
Args:
tgt: Target sequence.
memory: Memory (encoder output).
tgt_mask: Self-attention mask for target.
memory_mask: Cross-attention mask.
tgt_key_padding_mask: Padding mask for target keys.
memory_key_padding_mask: Padding mask for memory keys.
pos: Positional embeddings for memory.
query_pos: Positional embeddings for queries.
residual: Whether to use residual connections (used in post-norm variant).
Returns:
Tuple of (output_tensor, attention_weights).
"""
if self.normalize_before:
return self.forward_pre(
tgt, memory, tgt_mask, memory_mask,
tgt_key_padding_mask, memory_key_padding_mask, pos, query_pos
)
return self.forward_post(
tgt, memory, tgt_mask, memory_mask,
tgt_key_padding_mask, memory_key_padding_mask, pos, query_pos, residual
)
def _get_clones(module: nn.Module, N: int) -> nn.ModuleList:
"""
Create N identical copies of a module.
Args:
module: The module to clone.
N: Number of clones to create.
Returns:
ModuleList containing N deep copies of the input module.
"""
return nn.ModuleList([copy.deepcopy(module) for i in range(N)])
def _get_activation_fn(activation: str) -> Callable[[Tensor], Tensor]:
"""
Return an activation function given a string identifier.
Args:
activation: Name of the activation function ('relu', 'gelu', or 'glu').
Returns:
The corresponding activation function.
Raises:
RuntimeError: If the activation name is not recognized.
"""
if activation == "relu":
return F.relu
if activation == "gelu":
return F.gelu
if activation == "glu":
return F.glu
raise RuntimeError(f"activation should be relu/gelu/glu, not {activation}.") |