File size: 13,797 Bytes
47e954c | 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 | """Neural network modules for the ACT (Action Chunking Transformer) model.
This module provides the core building blocks for the ACT architecture including
image encoders, positional encodings, and transformer encoder/decoder layers.
These components are designed to work together for robot manipulation tasks.
"""
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
import torchvision.models as models
class ACTImageEncoder(nn.Module):
"""Encode images using ResNet backbone with spatial position embeddings.
Maintains spatial dimensions and provides learnable position embeddings
similar to DETR's backbone implementation. Uses a pretrained ResNet18
backbone with projection layers for feature extraction.
"""
def __init__(self, output_dim: int = 256):
"""Initialize the image encoder.
Args:
output_dim: Output feature dimension after projection
"""
super().__init__()
# Use pretrained ResNet but remove final layers
self.backbone = self._build_backbone()
self.proj = nn.Conv2d(512, output_dim, kernel_size=1) # Project to output_dim
# Position embeddings should match output_dim
self.row_embed = nn.Embedding(50, output_dim // 2) # Half size
self.col_embed = nn.Embedding(50, output_dim // 2) # Half size
self.reset_parameters()
def _build_backbone(self) -> nn.Module:
"""Build backbone CNN, removing avgpool and fc layers.
Returns:
nn.Module: ResNet18 backbone without classification layers
"""
resnet = models.get_model("resnet18", weights="DEFAULT")
return nn.Sequential(*list(resnet.children())[:-2])
def reset_parameters(self) -> None:
"""Initialize position embeddings with uniform distribution."""
nn.init.uniform_(self.row_embed.weight)
nn.init.uniform_(self.col_embed.weight)
def forward(self, x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
"""Forward pass through image encoder.
Args:
x: Image tensor of shape (batch, channels, height, width)
Returns:
tuple[torch.Tensor, torch.Tensor]:
- features: Encoded features of shape (batch, output_dim, height, width)
- pos: Position embeddings of shape (batch, output_dim, height, width)
"""
# Extract features
x = self.backbone(x)
features = self.proj(x) # Now [B, output_dim, H, W]
# Create position embeddings
h, w = features.shape[-2:]
i = torch.arange(w, device=x.device)
j = torch.arange(h, device=x.device)
x_emb = self.col_embed(i) # [W, output_dim//2]
y_emb = self.row_embed(j) # [H, output_dim//2]
pos = (
torch.cat(
[
x_emb.unsqueeze(0).repeat(h, 1, 1),
y_emb.unsqueeze(1).repeat(1, w, 1),
],
dim=-1,
)
.permute(2, 0, 1)
.unsqueeze(0)
) # [1, output_dim, H, W]
pos = pos.repeat(x.shape[0], 1, 1, 1) # [B, output_dim, H, W]
return features, pos
class PositionalEncoding(nn.Module):
"""Sinusoidal positional encoding for transformer sequences.
Implements the standard sinusoidal positional encoding as described
in "Attention Is All You Need" (Vaswani et al., 2017).
"""
def __init__(self, d_model: int, dropout: float = 0.1, max_len: int = 5000):
"""Initialize positional encoding.
Args:
d_model: Model dimension
dropout: Dropout probability
max_len: Maximum sequence length to support
"""
super().__init__()
self.dropout = nn.Dropout(p=dropout)
pe = torch.zeros(max_len, d_model)
position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1)
div_term = torch.exp(
torch.arange(0, d_model, 2).float() * (-math.log(10000.0) / d_model)
)
pe[:, 0::2] = torch.sin(position * div_term)
pe[:, 1::2] = torch.cos(position * div_term)
pe = pe.unsqueeze(1) # [seq_len, batch, d_model]
self.register_buffer("pe", pe)
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""Add positional encoding to input tensor.
Args:
x: Input tensor of shape (seq_len, batch, d_model)
Returns:
torch.Tensor: Input with positional encoding added
"""
x = x + self.pe[: x.size(0), :, :] # [seq_len, batch, d_model]
return self.dropout(x)
class TransformerEncoderLayer(nn.Module):
"""Single transformer encoder layer with pre-layer normalization.
Implements a transformer encoder layer following the pre-norm architecture
with multi-head self-attention and position-wise feed-forward network.
"""
def __init__(
self,
d_model: int,
nhead: int,
dim_feedforward: int = 2048,
dropout: float = 0.1,
):
"""Initialize transformer encoder layer.
Args:
d_model: Model dimension
nhead: Number of attention heads
dim_feedforward: Feedforward network hidden dimension
dropout: Dropout probability
"""
super().__init__()
self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=dropout)
self.linear1 = nn.Linear(d_model, dim_feedforward)
self.linear2 = nn.Linear(dim_feedforward, d_model)
self.norm1 = nn.LayerNorm(d_model)
self.norm2 = nn.LayerNorm(d_model)
self.dropout = nn.Dropout(dropout)
self.dropout1 = nn.Dropout(dropout)
self.dropout2 = nn.Dropout(dropout)
def forward(
self,
src: torch.Tensor,
src_mask: torch.Tensor | None = None,
src_key_padding_mask: torch.Tensor | None = None,
) -> torch.Tensor:
"""Forward pass through encoder layer.
Args:
src: Input tensor of shape (seq_len, batch, d_model)
src_mask: Optional attention mask
src_key_padding_mask: Optional key padding mask
Returns:
torch.Tensor: Output tensor of same shape as input
"""
src2 = self.norm1(src)
src2, _ = self.self_attn(
src2, src2, src2, attn_mask=src_mask, key_padding_mask=src_key_padding_mask
)
src = src + self.dropout1(src2)
src2 = self.norm2(src)
src2 = self.linear2(self.dropout(F.relu(self.linear1(src2))))
src = src + self.dropout2(src2)
return src
class TransformerDecoderLayer(nn.Module):
"""Single transformer decoder layer with cross-attention.
Implements a transformer decoder layer with masked self-attention,
cross-attention to encoder memory, and position-wise feed-forward network.
Supports query position embeddings for object detection style architectures.
"""
def __init__(
self,
d_model: int,
nhead: int,
dim_feedforward: int = 2048,
dropout: float = 0.1,
):
"""Initialize transformer decoder layer.
Args:
d_model: Model dimension
nhead: Number of attention heads
dim_feedforward: Feedforward network hidden dimension
dropout: Dropout probability
"""
super().__init__()
self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=dropout)
self.multihead_attn = nn.MultiheadAttention(d_model, nhead, dropout=dropout)
self.linear1 = nn.Linear(d_model, dim_feedforward)
self.linear2 = nn.Linear(dim_feedforward, d_model)
self.norm1 = nn.LayerNorm(d_model)
self.norm2 = nn.LayerNorm(d_model)
self.norm3 = nn.LayerNorm(d_model)
self.dropout = nn.Dropout(dropout)
self.dropout1 = nn.Dropout(dropout)
self.dropout2 = nn.Dropout(dropout)
self.dropout3 = nn.Dropout(dropout)
def forward(
self,
tgt: torch.Tensor,
memory: torch.Tensor,
tgt_mask: torch.Tensor | None = None,
memory_mask: torch.Tensor | None = None,
tgt_key_padding_mask: torch.Tensor | None = None,
memory_key_padding_mask: torch.Tensor | None = None,
query_pos: torch.Tensor | None = None,
) -> torch.Tensor:
"""Forward pass through decoder layer.
Args:
tgt: Target tensor of shape (tgt_len, batch, d_model)
memory: Memory tensor from encoder of shape (src_len, batch, d_model)
tgt_mask: Optional target attention mask
memory_mask: Optional memory attention mask
tgt_key_padding_mask: Optional target key padding mask
memory_key_padding_mask: Optional memory key padding mask
query_pos: Optional query position embeddings
Returns:
torch.Tensor: Output tensor of same shape as target
"""
q = k = tgt if query_pos is None else tgt + query_pos
tgt2 = self.norm1(tgt)
tgt2, _ = self.self_attn(
q, k, tgt2, attn_mask=tgt_mask, key_padding_mask=tgt_key_padding_mask
)
tgt = tgt + self.dropout1(tgt2)
tgt2 = self.norm2(tgt)
tgt2, _ = self.multihead_attn(
query=tgt2 if query_pos is None else tgt2 + query_pos,
key=memory,
value=memory,
attn_mask=memory_mask,
key_padding_mask=memory_key_padding_mask,
)
tgt = tgt + self.dropout2(tgt2)
tgt2 = self.norm3(tgt)
tgt2 = self.linear2(self.dropout(F.relu(self.linear1(tgt2))))
tgt = tgt + self.dropout3(tgt2)
return tgt
class TransformerEncoder(nn.Module):
"""Stack of transformer encoder layers.
Composes multiple TransformerEncoderLayer modules with a final layer
normalization for encoding sequential input data.
"""
def __init__(
self,
d_model: int,
nhead: int,
num_encoder_layers: int = 6,
dim_feedforward: int = 2048,
dropout: float = 0.1,
):
"""Initialize transformer encoder.
Args:
d_model: Model dimension
nhead: Number of attention heads
num_encoder_layers: Number of encoder layers
dim_feedforward: Feedforward network hidden dimension
dropout: Dropout probability
"""
super().__init__()
self.layers = nn.ModuleList([
TransformerEncoderLayer(d_model, nhead, dim_feedforward, dropout)
for _ in range(num_encoder_layers)
])
self.norm = nn.LayerNorm(d_model)
def forward(
self,
src: torch.Tensor,
mask: torch.Tensor | None = None,
src_key_padding_mask: torch.Tensor | None = None,
) -> torch.Tensor:
"""Forward pass through all encoder layers.
Args:
src: Input tensor of shape (seq_len, batch, d_model)
mask: Optional attention mask
src_key_padding_mask: Optional key padding mask
Returns:
torch.Tensor: Encoded output tensor
"""
output = src
for layer in self.layers:
output = layer(
output, src_mask=mask, src_key_padding_mask=src_key_padding_mask
)
return self.norm(output)
class TransformerDecoder(nn.Module):
"""Stack of transformer decoder layers.
Composes multiple TransformerDecoderLayer modules with a final layer
normalization for generating sequential output conditioned on memory.
"""
def __init__(
self,
d_model: int,
nhead: int,
num_decoder_layers: int = 6,
dim_feedforward: int = 2048,
dropout: float = 0.1,
):
"""Initialize transformer decoder.
Args:
d_model: Model dimension
nhead: Number of attention heads
num_decoder_layers: Number of decoder layers
dim_feedforward: Feedforward network hidden dimension
dropout: Dropout probability
"""
super().__init__()
self.layers = nn.ModuleList([
TransformerDecoderLayer(d_model, nhead, dim_feedforward, dropout)
for _ in range(num_decoder_layers)
])
self.norm = nn.LayerNorm(d_model)
def forward(
self,
tgt: torch.Tensor,
memory: torch.Tensor,
tgt_mask: torch.Tensor | None = None,
memory_mask: torch.Tensor | None = None,
tgt_key_padding_mask: torch.Tensor | None = None,
memory_key_padding_mask: torch.Tensor | None = None,
query_pos: torch.Tensor | None = None,
) -> torch.Tensor:
"""Forward pass through all decoder layers.
Args:
tgt: Target tensor of shape (tgt_len, batch, d_model)
memory: Memory tensor from encoder of shape (src_len, batch, d_model)
tgt_mask: Optional target attention mask
memory_mask: Optional memory attention mask
tgt_key_padding_mask: Optional target key padding mask
memory_key_padding_mask: Optional memory key padding mask
query_pos: Optional query position embeddings
Returns:
torch.Tensor: Decoded output tensor
"""
output = tgt
for layer in self.layers:
output = 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,
query_pos=query_pos,
)
return self.norm(output)
|