Reference for ultralytics/models/sam/modules/transformer.py
Note
This file is available at https://github.com/ultralytics/ultralytics/blob/main/ultralytics/models/sam/modules/transformer.py. If you spot a problem please help fix it by contributing a Pull Request 🛠️. Thank you 🙏!
ultralytics.models.sam.modules.transformer.TwoWayTransformer
TwoWayTransformer(
depth: int,
embedding_dim: int,
num_heads: int,
mlp_dim: int,
activation: Type[Module] = nn.ReLU,
attention_downsample_rate: int = 2,
)
Bases: Module
A Two-Way Transformer module for simultaneous attention to image and query points.
This class implements a specialized transformer decoder that attends to an input image using queries with supplied positional embeddings. It's useful for tasks like object detection, image segmentation, and point cloud processing.
Attributes:
Name | Type | Description |
---|---|---|
depth |
int
|
Number of layers in the transformer. |
embedding_dim |
int
|
Channel dimension for input embeddings. |
num_heads |
int
|
Number of heads for multihead attention. |
mlp_dim |
int
|
Internal channel dimension for the MLP block. |
layers |
ModuleList
|
List of TwoWayAttentionBlock layers composing the transformer. |
final_attn_token_to_image |
Attention
|
Final attention layer from queries to image. |
norm_final_attn |
LayerNorm
|
Layer normalization applied to final queries. |
Methods:
Name | Description |
---|---|
forward |
Processes image and point embeddings through the transformer. |
Examples:
>>> transformer = TwoWayTransformer(depth=6, embedding_dim=256, num_heads=8, mlp_dim=2048)
>>> image_embedding = torch.randn(1, 256, 32, 32)
>>> image_pe = torch.randn(1, 256, 32, 32)
>>> point_embedding = torch.randn(1, 100, 256)
>>> output_queries, output_image = transformer(image_embedding, image_pe, point_embedding)
>>> print(output_queries.shape, output_image.shape)
Parameters:
Name | Type | Description | Default |
---|---|---|---|
depth
|
int
|
Number of layers in the transformer. |
required |
embedding_dim
|
int
|
Channel dimension for input embeddings. |
required |
num_heads
|
int
|
Number of heads for multihead attention. Must divide embedding_dim. |
required |
mlp_dim
|
int
|
Internal channel dimension for the MLP block. |
required |
activation
|
Type[Module]
|
Activation function to use in the MLP block. |
ReLU
|
attention_downsample_rate
|
int
|
Downsampling rate for attention mechanism. |
2
|
Source code in ultralytics/models/sam/modules/transformer.py
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 |
|
forward
forward(
image_embedding: Tensor, image_pe: Tensor, point_embedding: Tensor
) -> Tuple[Tensor, Tensor]
Process image and point embeddings through the Two-Way Transformer.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
image_embedding
|
Tensor
|
Image to attend to, with shape (B, embedding_dim, H, W). |
required |
image_pe
|
Tensor
|
Positional encoding to add to the image, with same shape as image_embedding. |
required |
point_embedding
|
Tensor
|
Embedding to add to query points, with shape (B, N_points, embedding_dim). |
required |
Returns:
Name | Type | Description |
---|---|---|
queries |
Tensor
|
Processed point embeddings with shape (B, N_points, embedding_dim). |
keys |
Tensor
|
Processed image embeddings with shape (B, H*W, embedding_dim). |
Source code in ultralytics/models/sam/modules/transformer.py
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 |
|
ultralytics.models.sam.modules.transformer.TwoWayAttentionBlock
TwoWayAttentionBlock(
embedding_dim: int,
num_heads: int,
mlp_dim: int = 2048,
activation: Type[Module] = nn.ReLU,
attention_downsample_rate: int = 2,
skip_first_layer_pe: bool = False,
)
Bases: Module
A two-way attention block for simultaneous attention to image and query points.
This class implements a specialized transformer block with four main layers: self-attention on sparse inputs, cross-attention of sparse inputs to dense inputs, MLP block on sparse inputs, and cross-attention of dense inputs to sparse inputs.
Attributes:
Name | Type | Description |
---|---|---|
self_attn |
Attention
|
Self-attention layer for queries. |
norm1 |
LayerNorm
|
Layer normalization after self-attention. |
cross_attn_token_to_image |
Attention
|
Cross-attention layer from queries to keys. |
norm2 |
LayerNorm
|
Layer normalization after token-to-image attention. |
mlp |
MLPBlock
|
MLP block for transforming query embeddings. |
norm3 |
LayerNorm
|
Layer normalization after MLP block. |
norm4 |
LayerNorm
|
Layer normalization after image-to-token attention. |
cross_attn_image_to_token |
Attention
|
Cross-attention layer from keys to queries. |
skip_first_layer_pe |
bool
|
Whether to skip positional encoding in the first layer. |
Methods:
Name | Description |
---|---|
forward |
Applies self-attention and cross-attention to queries and keys. |
Examples:
>>> embedding_dim, num_heads = 256, 8
>>> block = TwoWayAttentionBlock(embedding_dim, num_heads)
>>> queries = torch.randn(1, 100, embedding_dim)
>>> keys = torch.randn(1, 1000, embedding_dim)
>>> query_pe = torch.randn(1, 100, embedding_dim)
>>> key_pe = torch.randn(1, 1000, embedding_dim)
>>> processed_queries, processed_keys = block(queries, keys, query_pe, key_pe)
This block implements a specialized transformer layer with four main components: self-attention on sparse inputs, cross-attention of sparse inputs to dense inputs, MLP block on sparse inputs, and cross-attention of dense inputs to sparse inputs.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
embedding_dim
|
int
|
Channel dimension of the embeddings. |
required |
num_heads
|
int
|
Number of attention heads in the attention layers. |
required |
mlp_dim
|
int
|
Hidden dimension of the MLP block. |
2048
|
activation
|
Type[Module]
|
Activation function for the MLP block. |
ReLU
|
attention_downsample_rate
|
int
|
Downsampling rate for the attention mechanism. |
2
|
skip_first_layer_pe
|
bool
|
Whether to skip positional encoding in the first layer. |
False
|
Source code in ultralytics/models/sam/modules/transformer.py
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 |
|
forward
forward(
queries: Tensor, keys: Tensor, query_pe: Tensor, key_pe: Tensor
) -> Tuple[Tensor, Tensor]
Apply two-way attention to process query and key embeddings in a transformer block.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
queries
|
Tensor
|
Query embeddings with shape (B, N_queries, embedding_dim). |
required |
keys
|
Tensor
|
Key embeddings with shape (B, N_keys, embedding_dim). |
required |
query_pe
|
Tensor
|
Positional encodings for queries with same shape as queries. |
required |
key_pe
|
Tensor
|
Positional encodings for keys with same shape as keys. |
required |
Returns:
Name | Type | Description |
---|---|---|
queries |
Tensor
|
Processed query embeddings with shape (B, N_queries, embedding_dim). |
keys |
Tensor
|
Processed key embeddings with shape (B, N_keys, embedding_dim). |
Source code in ultralytics/models/sam/modules/transformer.py
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 |
|
ultralytics.models.sam.modules.transformer.Attention
Attention(
embedding_dim: int,
num_heads: int,
downsample_rate: int = 1,
kv_in_dim: int = None,
)
Bases: Module
An attention layer with downscaling capability for embedding size after projection.
This class implements a multi-head attention mechanism with the option to downsample the internal dimension of queries, keys, and values.
Attributes:
Name | Type | Description |
---|---|---|
embedding_dim |
int
|
Dimensionality of input embeddings. |
kv_in_dim |
int
|
Dimensionality of key and value inputs. |
internal_dim |
int
|
Internal dimension after downsampling. |
num_heads |
int
|
Number of attention heads. |
q_proj |
Linear
|
Linear projection for queries. |
k_proj |
Linear
|
Linear projection for keys. |
v_proj |
Linear
|
Linear projection for values. |
out_proj |
Linear
|
Linear projection for output. |
Methods:
Name | Description |
---|---|
_separate_heads |
Separates input tensor into attention heads. |
_recombine_heads |
Recombines separated attention heads. |
forward |
Computes attention output for given query, key, and value tensors. |
Examples:
>>> attn = Attention(embedding_dim=256, num_heads=8, downsample_rate=2)
>>> q = torch.randn(1, 100, 256)
>>> k = v = torch.randn(1, 50, 256)
>>> output = attn(q, k, v)
>>> print(output.shape)
torch.Size([1, 100, 256])
Parameters:
Name | Type | Description | Default |
---|---|---|---|
embedding_dim
|
int
|
Dimensionality of input embeddings. |
required |
num_heads
|
int
|
Number of attention heads. |
required |
downsample_rate
|
int
|
Factor by which internal dimensions are downsampled. |
1
|
kv_in_dim
|
int | None
|
Dimensionality of key and value inputs. If None, uses embedding_dim. |
None
|
Raises:
Type | Description |
---|---|
AssertionError
|
If num_heads does not evenly divide the internal dim (embedding_dim / downsample_rate). |
Source code in ultralytics/models/sam/modules/transformer.py
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 |
|
forward
forward(q: Tensor, k: Tensor, v: Tensor) -> Tensor
Apply multi-head attention to query, key, and value tensors with optional downsampling.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
q
|
Tensor
|
Query tensor with shape (B, N_q, embedding_dim). |
required |
k
|
Tensor
|
Key tensor with shape (B, N_k, embedding_dim). |
required |
v
|
Tensor
|
Value tensor with shape (B, N_k, embedding_dim). |
required |
Returns:
Type | Description |
---|---|
Tensor
|
Output tensor after attention with shape (B, N_q, embedding_dim). |
Source code in ultralytics/models/sam/modules/transformer.py
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 |
|