Reference for ultralytics/nn/modules/block.py
Note
This file is available at https://github.com/ultralytics/ultralytics/blob/main/ultralytics/nn/modules/block.py. If you spot a problem please help fix it by contributing a Pull Request 🛠️. Thank you 🙏!
ultralytics.nn.modules.block.DFL
DFL(c1=16)
Bases: Module
Integral module of Distribution Focal Loss (DFL).
Proposed in Generalized Focal Loss https://ieeexplore.ieee.org/document/9792391
Source code in ultralytics/nn/modules/block.py
63 64 65 66 67 68 69 |
|
forward
forward(x)
Apply the DFL module to input tensor and return transformed output.
Source code in ultralytics/nn/modules/block.py
71 72 73 74 |
|
ultralytics.nn.modules.block.Proto
Proto(c1, c_=256, c2=32)
Bases: Module
YOLOv8 mask Proto module for segmentation models.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
c1
|
int
|
Input channels. |
required |
c_
|
int
|
Intermediate channels. |
256
|
c2
|
int
|
Output channels (number of protos). |
32
|
Source code in ultralytics/nn/modules/block.py
81 82 83 84 85 86 87 88 89 90 91 92 93 94 |
|
forward
forward(x)
Perform a forward pass through layers using an upsampled input image.
Source code in ultralytics/nn/modules/block.py
96 97 98 |
|
ultralytics.nn.modules.block.HGStem
HGStem(c1, cm, c2)
Bases: Module
StemBlock of PPHGNetV2 with 5 convolutions and one maxpool2d.
https://github.com/PaddlePaddle/PaddleDetection/blob/develop/ppdet/modeling/backbones/hgnet_v2.py
Parameters:
Name | Type | Description | Default |
---|---|---|---|
c1
|
int
|
Input channels. |
required |
cm
|
int
|
Middle channels. |
required |
c2
|
int
|
Output channels. |
required |
Source code in ultralytics/nn/modules/block.py
108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 |
|
forward
forward(x)
Forward pass of a PPHGNetV2 backbone layer.
Source code in ultralytics/nn/modules/block.py
125 126 127 128 129 130 131 132 133 134 135 136 |
|
ultralytics.nn.modules.block.HGBlock
HGBlock(c1, cm, c2, k=3, n=6, lightconv=False, shortcut=False, act=nn.ReLU())
Bases: Module
HG_Block of PPHGNetV2 with 2 convolutions and LightConv.
https://github.com/PaddlePaddle/PaddleDetection/blob/develop/ppdet/modeling/backbones/hgnet_v2.py
Parameters:
Name | Type | Description | Default |
---|---|---|---|
c1
|
int
|
Input channels. |
required |
cm
|
int
|
Middle channels. |
required |
c2
|
int
|
Output channels. |
required |
k
|
int
|
Kernel size. |
3
|
n
|
int
|
Number of LightConv or Conv blocks. |
6
|
lightconv
|
bool
|
Whether to use LightConv. |
False
|
shortcut
|
bool
|
Whether to use shortcut connection. |
False
|
act
|
Module
|
Activation function. |
ReLU()
|
Source code in ultralytics/nn/modules/block.py
146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 |
|
forward
forward(x)
Forward pass of a PPHGNetV2 backbone layer.
Source code in ultralytics/nn/modules/block.py
167 168 169 170 171 172 |
|
ultralytics.nn.modules.block.SPP
SPP(c1, c2, k=(5, 9, 13))
Bases: Module
Spatial Pyramid Pooling (SPP) layer https://arxiv.org/abs/1406.4729.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
c1
|
int
|
Input channels. |
required |
c2
|
int
|
Output channels. |
required |
k
|
Tuple[int, int, int]
|
Kernel sizes for max pooling. |
(5, 9, 13)
|
Source code in ultralytics/nn/modules/block.py
178 179 180 181 182 183 184 185 186 187 188 189 190 191 |
|
forward
forward(x)
Forward pass of the SPP layer, performing spatial pyramid pooling.
Source code in ultralytics/nn/modules/block.py
193 194 195 196 |
|
ultralytics.nn.modules.block.SPPF
SPPF(c1, c2, k=5)
Bases: Module
Spatial Pyramid Pooling - Fast (SPPF) layer for YOLOv5 by Glenn Jocher.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
c1
|
int
|
Input channels. |
required |
c2
|
int
|
Output channels. |
required |
k
|
int
|
Kernel size. |
5
|
Notes
This module is equivalent to SPP(k=(5, 9, 13)).
Source code in ultralytics/nn/modules/block.py
202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 |
|
forward
forward(x)
Apply sequential pooling operations to input and return concatenated feature maps.
Source code in ultralytics/nn/modules/block.py
220 221 222 223 224 |
|
ultralytics.nn.modules.block.C1
C1(c1, c2, n=1)
Bases: Module
CSP Bottleneck with 1 convolution.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
c1
|
int
|
Input channels. |
required |
c2
|
int
|
Output channels. |
required |
n
|
int
|
Number of convolutions. |
1
|
Source code in ultralytics/nn/modules/block.py
230 231 232 233 234 235 236 237 238 239 240 241 |
|
forward
forward(x)
Apply convolution and residual connection to input tensor.
Source code in ultralytics/nn/modules/block.py
243 244 245 246 |
|
ultralytics.nn.modules.block.C2
C2(c1, c2, n=1, shortcut=True, g=1, e=0.5)
Bases: Module
CSP Bottleneck with 2 convolutions.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
c1
|
int
|
Input channels. |
required |
c2
|
int
|
Output channels. |
required |
n
|
int
|
Number of Bottleneck blocks. |
1
|
shortcut
|
bool
|
Whether to use shortcut connections. |
True
|
g
|
int
|
Groups for convolutions. |
1
|
e
|
float
|
Expansion ratio. |
0.5
|
Source code in ultralytics/nn/modules/block.py
252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 |
|
forward
forward(x)
Forward pass through the CSP bottleneck with 2 convolutions.
Source code in ultralytics/nn/modules/block.py
271 272 273 274 |
|
ultralytics.nn.modules.block.C2f
C2f(c1, c2, n=1, shortcut=False, g=1, e=0.5)
Bases: Module
Faster Implementation of CSP Bottleneck with 2 convolutions.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
c1
|
int
|
Input channels. |
required |
c2
|
int
|
Output channels. |
required |
n
|
int
|
Number of Bottleneck blocks. |
1
|
shortcut
|
bool
|
Whether to use shortcut connections. |
False
|
g
|
int
|
Groups for convolutions. |
1
|
e
|
float
|
Expansion ratio. |
0.5
|
Source code in ultralytics/nn/modules/block.py
280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 |
|
forward
forward(x)
Forward pass through C2f layer.
Source code in ultralytics/nn/modules/block.py
298 299 300 301 302 |
|
forward_split
forward_split(x)
Forward pass using split() instead of chunk().
Source code in ultralytics/nn/modules/block.py
304 305 306 307 308 309 |
|
ultralytics.nn.modules.block.C3
C3(c1, c2, n=1, shortcut=True, g=1, e=0.5)
Bases: Module
CSP Bottleneck with 3 convolutions.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
c1
|
int
|
Input channels. |
required |
c2
|
int
|
Output channels. |
required |
n
|
int
|
Number of Bottleneck blocks. |
1
|
shortcut
|
bool
|
Whether to use shortcut connections. |
True
|
g
|
int
|
Groups for convolutions. |
1
|
e
|
float
|
Expansion ratio. |
0.5
|
Source code in ultralytics/nn/modules/block.py
315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 |
|
forward
forward(x)
Forward pass through the CSP bottleneck with 3 convolutions.
Source code in ultralytics/nn/modules/block.py
334 335 336 |
|
ultralytics.nn.modules.block.C3x
C3x(c1, c2, n=1, shortcut=True, g=1, e=0.5)
Bases: C3
C3 module with cross-convolutions.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
c1
|
int
|
Input channels. |
required |
c2
|
int
|
Output channels. |
required |
n
|
int
|
Number of Bottleneck blocks. |
1
|
shortcut
|
bool
|
Whether to use shortcut connections. |
True
|
g
|
int
|
Groups for convolutions. |
1
|
e
|
float
|
Expansion ratio. |
0.5
|
Source code in ultralytics/nn/modules/block.py
342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 |
|
ultralytics.nn.modules.block.RepC3
RepC3(c1, c2, n=3, e=1.0)
Bases: Module
Rep C3.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
c1
|
int
|
Input channels. |
required |
c2
|
int
|
Output channels. |
required |
n
|
int
|
Number of RepConv blocks. |
3
|
e
|
float
|
Expansion ratio. |
1.0
|
Source code in ultralytics/nn/modules/block.py
362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 |
|
forward
forward(x)
Forward pass of RepC3 module.
Source code in ultralytics/nn/modules/block.py
379 380 381 |
|
ultralytics.nn.modules.block.C3TR
C3TR(c1, c2, n=1, shortcut=True, g=1, e=0.5)
Bases: C3
C3 module with TransformerBlock().
Parameters:
Name | Type | Description | Default |
---|---|---|---|
c1
|
int
|
Input channels. |
required |
c2
|
int
|
Output channels. |
required |
n
|
int
|
Number of Transformer blocks. |
1
|
shortcut
|
bool
|
Whether to use shortcut connections. |
True
|
g
|
int
|
Groups for convolutions. |
1
|
e
|
float
|
Expansion ratio. |
0.5
|
Source code in ultralytics/nn/modules/block.py
387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 |
|
ultralytics.nn.modules.block.C3Ghost
C3Ghost(c1, c2, n=1, shortcut=True, g=1, e=0.5)
Bases: C3
C3 module with GhostBottleneck().
Parameters:
Name | Type | Description | Default |
---|---|---|---|
c1
|
int
|
Input channels. |
required |
c2
|
int
|
Output channels. |
required |
n
|
int
|
Number of Ghost bottleneck blocks. |
1
|
shortcut
|
bool
|
Whether to use shortcut connections. |
True
|
g
|
int
|
Groups for convolutions. |
1
|
e
|
float
|
Expansion ratio. |
0.5
|
Source code in ultralytics/nn/modules/block.py
407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 |
|
ultralytics.nn.modules.block.GhostBottleneck
GhostBottleneck(c1, c2, k=3, s=1)
Bases: Module
Ghost Bottleneck https://github.com/huawei-noah/Efficient-AI-Backbones.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
c1
|
int
|
Input channels. |
required |
c2
|
int
|
Output channels. |
required |
k
|
int
|
Kernel size. |
3
|
s
|
int
|
Stride. |
1
|
Source code in ultralytics/nn/modules/block.py
427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 |
|
forward
forward(x)
Apply skip connection and concatenation to input tensor.
Source code in ultralytics/nn/modules/block.py
448 449 450 |
|
ultralytics.nn.modules.block.Bottleneck
Bottleneck(c1, c2, shortcut=True, g=1, k=(3, 3), e=0.5)
Bases: Module
Standard bottleneck.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
c1
|
int
|
Input channels. |
required |
c2
|
int
|
Output channels. |
required |
shortcut
|
bool
|
Whether to use shortcut connection. |
True
|
g
|
int
|
Groups for convolutions. |
1
|
k
|
Tuple[int, int]
|
Kernel sizes for convolutions. |
(3, 3)
|
e
|
float
|
Expansion ratio. |
0.5
|
Source code in ultralytics/nn/modules/block.py
456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 |
|
forward
forward(x)
Apply bottleneck with optional shortcut connection.
Source code in ultralytics/nn/modules/block.py
474 475 476 |
|
ultralytics.nn.modules.block.BottleneckCSP
BottleneckCSP(c1, c2, n=1, shortcut=True, g=1, e=0.5)
Bases: Module
CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
c1
|
int
|
Input channels. |
required |
c2
|
int
|
Output channels. |
required |
n
|
int
|
Number of Bottleneck blocks. |
1
|
shortcut
|
bool
|
Whether to use shortcut connections. |
True
|
g
|
int
|
Groups for convolutions. |
1
|
e
|
float
|
Expansion ratio. |
0.5
|
Source code in ultralytics/nn/modules/block.py
482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 |
|
forward
forward(x)
Apply CSP bottleneck with 3 convolutions.
Source code in ultralytics/nn/modules/block.py
504 505 506 507 508 |
|
ultralytics.nn.modules.block.ResNetBlock
ResNetBlock(c1, c2, s=1, e=4)
Bases: Module
ResNet block with standard convolution layers.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
c1
|
int
|
Input channels. |
required |
c2
|
int
|
Output channels. |
required |
s
|
int
|
Stride. |
1
|
e
|
int
|
Expansion ratio. |
4
|
Source code in ultralytics/nn/modules/block.py
514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 |
|
forward
forward(x)
Forward pass through the ResNet block.
Source code in ultralytics/nn/modules/block.py
531 532 533 |
|
ultralytics.nn.modules.block.ResNetLayer
ResNetLayer(c1, c2, s=1, is_first=False, n=1, e=4)
Bases: Module
ResNet layer with multiple ResNet blocks.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
c1
|
int
|
Input channels. |
required |
c2
|
int
|
Output channels. |
required |
s
|
int
|
Stride. |
1
|
is_first
|
bool
|
Whether this is the first layer. |
False
|
n
|
int
|
Number of ResNet blocks. |
1
|
e
|
int
|
Expansion ratio. |
4
|
Source code in ultralytics/nn/modules/block.py
539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 |
|
forward
forward(x)
Forward pass through the ResNet layer.
Source code in ultralytics/nn/modules/block.py
563 564 565 |
|
ultralytics.nn.modules.block.MaxSigmoidAttnBlock
MaxSigmoidAttnBlock(c1, c2, nh=1, ec=128, gc=512, scale=False)
Bases: Module
Max Sigmoid attention block.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
c1
|
int
|
Input channels. |
required |
c2
|
int
|
Output channels. |
required |
nh
|
int
|
Number of heads. |
1
|
ec
|
int
|
Embedding channels. |
128
|
gc
|
int
|
Guide channels. |
512
|
scale
|
bool
|
Whether to use learnable scale parameter. |
False
|
Source code in ultralytics/nn/modules/block.py
571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 |
|
forward
forward(x, guide)
Forward pass of MaxSigmoidAttnBlock.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
x
|
Tensor
|
Input tensor. |
required |
guide
|
Tensor
|
Guide tensor. |
required |
Returns:
Type | Description |
---|---|
Tensor
|
Output tensor after attention. |
Source code in ultralytics/nn/modules/block.py
592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 |
|
ultralytics.nn.modules.block.C2fAttn
C2fAttn(c1, c2, n=1, ec=128, nh=1, gc=512, shortcut=False, g=1, e=0.5)
Bases: Module
C2f module with an additional attn module.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
c1
|
int
|
Input channels. |
required |
c2
|
int
|
Output channels. |
required |
n
|
int
|
Number of Bottleneck blocks. |
1
|
ec
|
int
|
Embedding channels for attention. |
128
|
nh
|
int
|
Number of heads for attention. |
1
|
gc
|
int
|
Guide channels for attention. |
512
|
shortcut
|
bool
|
Whether to use shortcut connections. |
False
|
g
|
int
|
Groups for convolutions. |
1
|
e
|
float
|
Expansion ratio. |
0.5
|
Source code in ultralytics/nn/modules/block.py
625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 |
|
forward
forward(x, guide)
Forward pass through C2f layer with attention.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
x
|
Tensor
|
Input tensor. |
required |
guide
|
Tensor
|
Guide tensor for attention. |
required |
Returns:
Type | Description |
---|---|
Tensor
|
Output tensor after processing. |
Source code in ultralytics/nn/modules/block.py
647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 |
|
forward_split
forward_split(x, guide)
Forward pass using split() instead of chunk().
Parameters:
Name | Type | Description | Default |
---|---|---|---|
x
|
Tensor
|
Input tensor. |
required |
guide
|
Tensor
|
Guide tensor for attention. |
required |
Returns:
Type | Description |
---|---|
Tensor
|
Output tensor after processing. |
Source code in ultralytics/nn/modules/block.py
663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 |
|
ultralytics.nn.modules.block.ImagePoolingAttn
ImagePoolingAttn(ec=256, ch=(), ct=512, nh=8, k=3, scale=False)
Bases: Module
ImagePoolingAttn: Enhance the text embeddings with image-aware information.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
ec
|
int
|
Embedding channels. |
256
|
ch
|
tuple
|
Channel dimensions for feature maps. |
()
|
ct
|
int
|
Channel dimension for text embeddings. |
512
|
nh
|
int
|
Number of attention heads. |
8
|
k
|
int
|
Kernel size for pooling. |
3
|
scale
|
bool
|
Whether to use learnable scale parameter. |
False
|
Source code in ultralytics/nn/modules/block.py
683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 |
|
forward
forward(x, text)
Forward pass of ImagePoolingAttn.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
x
|
List[Tensor]
|
List of input feature maps. |
required |
text
|
Tensor
|
Text embeddings. |
required |
Returns:
Type | Description |
---|---|
Tensor
|
Enhanced text embeddings. |
Source code in ultralytics/nn/modules/block.py
711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 |
|
ultralytics.nn.modules.block.ContrastiveHead
ContrastiveHead()
Bases: Module
Implements contrastive learning head for region-text similarity in vision-language models.
Source code in ultralytics/nn/modules/block.py
748 749 750 751 752 753 |
|
forward
forward(x, w)
Forward function of contrastive learning.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
x
|
Tensor
|
Image features. |
required |
w
|
Tensor
|
Text features. |
required |
Returns:
Type | Description |
---|---|
Tensor
|
Similarity scores. |
Source code in ultralytics/nn/modules/block.py
755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 |
|
ultralytics.nn.modules.block.BNContrastiveHead
BNContrastiveHead(embed_dims: int)
Bases: Module
Batch Norm Contrastive Head using batch norm instead of l2-normalization.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
embed_dims
|
int
|
Embed dimensions of text and image features. |
required |
Parameters:
Name | Type | Description | Default |
---|---|---|---|
embed_dims
|
int
|
Embedding dimensions for features. |
required |
Source code in ultralytics/nn/modules/block.py
780 781 782 783 784 785 786 787 788 789 790 791 792 |
|
forward
forward(x, w)
Forward function of contrastive learning with batch normalization.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
x
|
Tensor
|
Image features. |
required |
w
|
Tensor
|
Text features. |
required |
Returns:
Type | Description |
---|---|
Tensor
|
Similarity scores. |
Source code in ultralytics/nn/modules/block.py
809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 |
|
forward_fuse
forward_fuse(x, w)
Passes input out unchanged.
TODO: Update or remove?
Source code in ultralytics/nn/modules/block.py
801 802 803 804 805 806 807 |
|
fuse
fuse()
Fuse the batch normalization layer in the BNContrastiveHead module.
Source code in ultralytics/nn/modules/block.py
794 795 796 797 798 799 |
|
ultralytics.nn.modules.block.RepBottleneck
RepBottleneck(c1, c2, shortcut=True, g=1, k=(3, 3), e=0.5)
Bases: Bottleneck
Rep bottleneck.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
c1
|
int
|
Input channels. |
required |
c2
|
int
|
Output channels. |
required |
shortcut
|
bool
|
Whether to use shortcut connection. |
True
|
g
|
int
|
Groups for convolutions. |
1
|
k
|
Tuple[int, int]
|
Kernel sizes for convolutions. |
(3, 3)
|
e
|
float
|
Expansion ratio. |
0.5
|
Source code in ultralytics/nn/modules/block.py
830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 |
|
ultralytics.nn.modules.block.RepCSP
RepCSP(c1, c2, n=1, shortcut=True, g=1, e=0.5)
Bases: C3
Repeatable Cross Stage Partial Network (RepCSP) module for efficient feature extraction.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
c1
|
int
|
Input channels. |
required |
c2
|
int
|
Output channels. |
required |
n
|
int
|
Number of RepBottleneck blocks. |
1
|
shortcut
|
bool
|
Whether to use shortcut connections. |
True
|
g
|
int
|
Groups for convolutions. |
1
|
e
|
float
|
Expansion ratio. |
0.5
|
Source code in ultralytics/nn/modules/block.py
850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 |
|
ultralytics.nn.modules.block.RepNCSPELAN4
RepNCSPELAN4(c1, c2, c3, c4, n=1)
Bases: Module
CSP-ELAN.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
c1
|
int
|
Input channels. |
required |
c2
|
int
|
Output channels. |
required |
c3
|
int
|
Intermediate channels. |
required |
c4
|
int
|
Intermediate channels for RepCSP. |
required |
n
|
int
|
Number of RepCSP blocks. |
1
|
Source code in ultralytics/nn/modules/block.py
870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 |
|
forward
forward(x)
Forward pass through RepNCSPELAN4 layer.
Source code in ultralytics/nn/modules/block.py
888 889 890 891 892 |
|
forward_split
forward_split(x)
Forward pass using split() instead of chunk().
Source code in ultralytics/nn/modules/block.py
894 895 896 897 898 |
|
ultralytics.nn.modules.block.ELAN1
ELAN1(c1, c2, c3, c4)
Bases: RepNCSPELAN4
ELAN1 module with 4 convolutions.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
c1
|
int
|
Input channels. |
required |
c2
|
int
|
Output channels. |
required |
c3
|
int
|
Intermediate channels. |
required |
c4
|
int
|
Intermediate channels for convolutions. |
required |
Source code in ultralytics/nn/modules/block.py
904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 |
|
ultralytics.nn.modules.block.AConv
AConv(c1, c2)
Bases: Module
AConv.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
c1
|
int
|
Input channels. |
required |
c2
|
int
|
Output channels. |
required |
Source code in ultralytics/nn/modules/block.py
925 926 927 928 929 930 931 932 933 934 |
|
forward
forward(x)
Forward pass through AConv layer.
Source code in ultralytics/nn/modules/block.py
936 937 938 939 |
|
ultralytics.nn.modules.block.ADown
ADown(c1, c2)
Bases: Module
ADown.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
c1
|
int
|
Input channels. |
required |
c2
|
int
|
Output channels. |
required |
Source code in ultralytics/nn/modules/block.py
945 946 947 948 949 950 951 952 953 954 955 956 |
|
forward
forward(x)
Forward pass through ADown layer.
Source code in ultralytics/nn/modules/block.py
958 959 960 961 962 963 964 965 |
|
ultralytics.nn.modules.block.SPPELAN
SPPELAN(c1, c2, c3, k=5)
Bases: Module
SPP-ELAN.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
c1
|
int
|
Input channels. |
required |
c2
|
int
|
Output channels. |
required |
c3
|
int
|
Intermediate channels. |
required |
k
|
int
|
Kernel size for max pooling. |
5
|
Source code in ultralytics/nn/modules/block.py
971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 |
|
forward
forward(x)
Forward pass through SPPELAN layer.
Source code in ultralytics/nn/modules/block.py
989 990 991 992 993 |
|
ultralytics.nn.modules.block.CBLinear
CBLinear(c1, c2s, k=1, s=1, p=None, g=1)
Bases: Module
CBLinear.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
c1
|
int
|
Input channels. |
required |
c2s
|
List[int]
|
List of output channel sizes. |
required |
k
|
int
|
Kernel size. |
1
|
s
|
int
|
Stride. |
1
|
p
|
int | None
|
Padding. |
None
|
g
|
int
|
Groups. |
1
|
Source code in ultralytics/nn/modules/block.py
999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 |
|
forward
forward(x)
Forward pass through CBLinear layer.
Source code in ultralytics/nn/modules/block.py
1015 1016 1017 |
|
ultralytics.nn.modules.block.CBFuse
CBFuse(idx)
Bases: Module
CBFuse.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
idx
|
List[int]
|
Indices for feature selection. |
required |
Source code in ultralytics/nn/modules/block.py
1023 1024 1025 1026 1027 1028 1029 1030 1031 |
|
forward
forward(xs)
Forward pass through CBFuse layer.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
xs
|
List[Tensor]
|
List of input tensors. |
required |
Returns:
Type | Description |
---|---|
Tensor
|
Fused output tensor. |
Source code in ultralytics/nn/modules/block.py
1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 |
|
ultralytics.nn.modules.block.C3f
C3f(c1, c2, n=1, shortcut=False, g=1, e=0.5)
Bases: Module
Faster Implementation of CSP Bottleneck with 2 convolutions.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
c1
|
int
|
Input channels. |
required |
c2
|
int
|
Output channels. |
required |
n
|
int
|
Number of Bottleneck blocks. |
1
|
shortcut
|
bool
|
Whether to use shortcut connections. |
False
|
g
|
int
|
Groups for convolutions. |
1
|
e
|
float
|
Expansion ratio. |
0.5
|
Source code in ultralytics/nn/modules/block.py
1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 |
|
forward
forward(x)
Forward pass through C3f layer.
Source code in ultralytics/nn/modules/block.py
1070 1071 1072 1073 1074 |
|
ultralytics.nn.modules.block.C3k2
C3k2(c1, c2, n=1, c3k=False, e=0.5, g=1, shortcut=True)
Bases: C2f
Faster Implementation of CSP Bottleneck with 2 convolutions.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
c1
|
int
|
Input channels. |
required |
c2
|
int
|
Output channels. |
required |
n
|
int
|
Number of blocks. |
1
|
c3k
|
bool
|
Whether to use C3k blocks. |
False
|
e
|
float
|
Expansion ratio. |
0.5
|
g
|
int
|
Groups for convolutions. |
1
|
shortcut
|
bool
|
Whether to use shortcut connections. |
True
|
Source code in ultralytics/nn/modules/block.py
1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 |
|
ultralytics.nn.modules.block.C3k
C3k(c1, c2, n=1, shortcut=True, g=1, e=0.5, k=3)
Bases: C3
C3k is a CSP bottleneck module with customizable kernel sizes for feature extraction in neural networks.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
c1
|
int
|
Input channels. |
required |
c2
|
int
|
Output channels. |
required |
n
|
int
|
Number of Bottleneck blocks. |
1
|
shortcut
|
bool
|
Whether to use shortcut connections. |
True
|
g
|
int
|
Groups for convolutions. |
1
|
e
|
float
|
Expansion ratio. |
0.5
|
k
|
int
|
Kernel size. |
3
|
Source code in ultralytics/nn/modules/block.py
1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 |
|
ultralytics.nn.modules.block.RepVGGDW
RepVGGDW(ed)
Bases: Module
RepVGGDW is a class that represents a depth wise separable convolutional block in RepVGG architecture.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
ed
|
int
|
Input and output channels. |
required |
Source code in ultralytics/nn/modules/block.py
1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 |
|
forward
forward(x)
Perform a forward pass of the RepVGGDW block.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
x
|
Tensor
|
Input tensor. |
required |
Returns:
Type | Description |
---|---|
Tensor
|
Output tensor after applying the depth wise separable convolution. |
Source code in ultralytics/nn/modules/block.py
1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 |
|
forward_fuse
forward_fuse(x)
Perform a forward pass of the RepVGGDW block without fusing the convolutions.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
x
|
Tensor
|
Input tensor. |
required |
Returns:
Type | Description |
---|---|
Tensor
|
Output tensor after applying the depth wise separable convolution. |
Source code in ultralytics/nn/modules/block.py
1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 |
|
fuse
fuse()
Fuse the convolutional layers in the RepVGGDW block.
This method fuses the convolutional layers and updates the weights and biases accordingly.
Source code in ultralytics/nn/modules/block.py
1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 |
|
ultralytics.nn.modules.block.CIB
CIB(c1, c2, shortcut=True, e=0.5, lk=False)
Bases: Module
Conditional Identity Block (CIB) module.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
c1
|
int
|
Number of input channels. |
required |
c2
|
int
|
Number of output channels. |
required |
shortcut
|
bool
|
Whether to add a shortcut connection. Defaults to True. |
True
|
e
|
float
|
Scaling factor for the hidden channels. Defaults to 0.5. |
0.5
|
lk
|
bool
|
Whether to use RepVGGDW for the third convolutional layer. Defaults to False. |
False
|
Parameters:
Name | Type | Description | Default |
---|---|---|---|
c1
|
int
|
Input channels. |
required |
c2
|
int
|
Output channels. |
required |
shortcut
|
bool
|
Whether to use shortcut connection. |
True
|
e
|
float
|
Expansion ratio. |
0.5
|
lk
|
bool
|
Whether to use RepVGGDW. |
False
|
Source code in ultralytics/nn/modules/block.py
1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 |
|
forward
forward(x)
Forward pass of the CIB module.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
x
|
Tensor
|
Input tensor. |
required |
Returns:
Type | Description |
---|---|
Tensor
|
Output tensor. |
Source code in ultralytics/nn/modules/block.py
1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 |
|
ultralytics.nn.modules.block.C2fCIB
C2fCIB(c1, c2, n=1, shortcut=False, lk=False, g=1, e=0.5)
Bases: C2f
C2fCIB class represents a convolutional block with C2f and CIB modules.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
c1
|
int
|
Number of input channels. |
required |
c2
|
int
|
Number of output channels. |
required |
n
|
int
|
Number of CIB modules to stack. Defaults to 1. |
1
|
shortcut
|
bool
|
Whether to use shortcut connection. Defaults to False. |
False
|
lk
|
bool
|
Whether to use local key connection. Defaults to False. |
False
|
g
|
int
|
Number of groups for grouped convolution. Defaults to 1. |
1
|
e
|
float
|
Expansion ratio for CIB modules. Defaults to 0.5. |
0.5
|
Parameters:
Name | Type | Description | Default |
---|---|---|---|
c1
|
int
|
Input channels. |
required |
c2
|
int
|
Output channels. |
required |
n
|
int
|
Number of CIB modules. |
1
|
shortcut
|
bool
|
Whether to use shortcut connection. |
False
|
lk
|
bool
|
Whether to use local key connection. |
False
|
g
|
int
|
Groups for convolutions. |
1
|
e
|
float
|
Expansion ratio. |
0.5
|
Source code in ultralytics/nn/modules/block.py
1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 |
|
ultralytics.nn.modules.block.Attention
Attention(dim, num_heads=8, attn_ratio=0.5)
Bases: Module
Attention module that performs self-attention on the input tensor.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
dim
|
int
|
The input tensor dimension. |
required |
num_heads
|
int
|
The number of attention heads. |
8
|
attn_ratio
|
float
|
The ratio of the attention key dimension to the head dimension. |
0.5
|
Attributes:
Name | Type | Description |
---|---|---|
num_heads |
int
|
The number of attention heads. |
head_dim |
int
|
The dimension of each attention head. |
key_dim |
int
|
The dimension of the attention key. |
scale |
float
|
The scaling factor for the attention scores. |
qkv |
Conv
|
Convolutional layer for computing the query, key, and value. |
proj |
Conv
|
Convolutional layer for projecting the attended values. |
pe |
Conv
|
Convolutional layer for positional encoding. |
Parameters:
Name | Type | Description | Default |
---|---|---|---|
dim
|
int
|
Input dimension. |
required |
num_heads
|
int
|
Number of attention heads. |
8
|
attn_ratio
|
float
|
Attention ratio for key dimension. |
0.5
|
Source code in ultralytics/nn/modules/block.py
1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 |
|
forward
forward(x)
Forward pass of the Attention module.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
x
|
Tensor
|
The input tensor. |
required |
Returns:
Type | Description |
---|---|
Tensor
|
The output tensor after self-attention. |
Source code in ultralytics/nn/modules/block.py
1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 |
|
ultralytics.nn.modules.block.PSABlock
PSABlock(c, attn_ratio=0.5, num_heads=4, shortcut=True)
Bases: Module
PSABlock class implementing a Position-Sensitive Attention block for neural networks.
This class encapsulates the functionality for applying multi-head attention and feed-forward neural network layers with optional shortcut connections.
Attributes:
Name | Type | Description |
---|---|---|
attn |
Attention
|
Multi-head attention module. |
ffn |
Sequential
|
Feed-forward neural network module. |
add |
bool
|
Flag indicating whether to add shortcut connections. |
Methods:
Name | Description |
---|---|
forward |
Performs a forward pass through the PSABlock, applying attention and feed-forward layers. |
Examples:
Create a PSABlock and perform a forward pass
>>> psablock = PSABlock(c=128, attn_ratio=0.5, num_heads=4, shortcut=True)
>>> input_tensor = torch.randn(1, 128, 32, 32)
>>> output_tensor = psablock(input_tensor)
Parameters:
Name | Type | Description | Default |
---|---|---|---|
c
|
int
|
Input and output channels. |
required |
attn_ratio
|
float
|
Attention ratio for key dimension. |
0.5
|
num_heads
|
int
|
Number of attention heads. |
4
|
shortcut
|
bool
|
Whether to use shortcut connections. |
True
|
Source code in ultralytics/nn/modules/block.py
1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 |
|
forward
forward(x)
Execute a forward pass through PSABlock.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
x
|
Tensor
|
Input tensor. |
required |
Returns:
Type | Description |
---|---|
Tensor
|
Output tensor after attention and feed-forward processing. |
Source code in ultralytics/nn/modules/block.py
1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 |
|
ultralytics.nn.modules.block.PSA
PSA(c1, c2, e=0.5)
Bases: Module
PSA class for implementing Position-Sensitive Attention in neural networks.
This class encapsulates the functionality for applying position-sensitive attention and feed-forward networks to input tensors, enhancing feature extraction and processing capabilities.
Attributes:
Name | Type | Description |
---|---|---|
c |
int
|
Number of hidden channels after applying the initial convolution. |
cv1 |
Conv
|
1x1 convolution layer to reduce the number of input channels to 2*c. |
cv2 |
Conv
|
1x1 convolution layer to reduce the number of output channels to c. |
attn |
Attention
|
Attention module for position-sensitive attention. |
ffn |
Sequential
|
Feed-forward network for further processing. |
Methods:
Name | Description |
---|---|
forward |
Applies position-sensitive attention and feed-forward network to the input tensor. |
Examples:
Create a PSA module and apply it to an input tensor
>>> psa = PSA(c1=128, c2=128, e=0.5)
>>> input_tensor = torch.randn(1, 128, 64, 64)
>>> output_tensor = psa.forward(input_tensor)
Parameters:
Name | Type | Description | Default |
---|---|---|---|
c1
|
int
|
Input channels. |
required |
c2
|
int
|
Output channels. |
required |
e
|
float
|
Expansion ratio. |
0.5
|
Source code in ultralytics/nn/modules/block.py
1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 |
|
forward
forward(x)
Execute forward pass in PSA module.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
x
|
Tensor
|
Input tensor. |
required |
Returns:
Type | Description |
---|---|
Tensor
|
Output tensor after attention and feed-forward processing. |
Source code in ultralytics/nn/modules/block.py
1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 |
|
ultralytics.nn.modules.block.C2PSA
C2PSA(c1, c2, n=1, e=0.5)
Bases: Module
C2PSA module with attention mechanism for enhanced feature extraction and processing.
This module implements a convolutional block with attention mechanisms to enhance feature extraction and processing capabilities. It includes a series of PSABlock modules for self-attention and feed-forward operations.
Attributes:
Name | Type | Description |
---|---|---|
c |
int
|
Number of hidden channels. |
cv1 |
Conv
|
1x1 convolution layer to reduce the number of input channels to 2*c. |
cv2 |
Conv
|
1x1 convolution layer to reduce the number of output channels to c. |
m |
Sequential
|
Sequential container of PSABlock modules for attention and feed-forward operations. |
Methods:
Name | Description |
---|---|
forward |
Performs a forward pass through the C2PSA module, applying attention and feed-forward operations. |
Notes
This module essentially is the same as PSA module, but refactored to allow stacking more PSABlock modules.
Examples:
>>> c2psa = C2PSA(c1=256, c2=256, n=3, e=0.5)
>>> input_tensor = torch.randn(1, 256, 64, 64)
>>> output_tensor = c2psa(input_tensor)
Parameters:
Name | Type | Description | Default |
---|---|---|---|
c1
|
int
|
Input channels. |
required |
c2
|
int
|
Output channels. |
required |
n
|
int
|
Number of PSABlock modules. |
1
|
e
|
float
|
Expansion ratio. |
0.5
|
Source code in ultralytics/nn/modules/block.py
1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 |
|
forward
forward(x)
Process the input tensor through a series of PSA blocks.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
x
|
Tensor
|
Input tensor. |
required |
Returns:
Type | Description |
---|---|
Tensor
|
Output tensor after processing. |
Source code in ultralytics/nn/modules/block.py
1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 |
|
ultralytics.nn.modules.block.C2fPSA
C2fPSA(c1, c2, n=1, e=0.5)
Bases: C2f
C2fPSA module with enhanced feature extraction using PSA blocks.
This class extends the C2f module by incorporating PSA blocks for improved attention mechanisms and feature extraction.
Attributes:
Name | Type | Description |
---|---|---|
c |
int
|
Number of hidden channels. |
cv1 |
Conv
|
1x1 convolution layer to reduce the number of input channels to 2*c. |
cv2 |
Conv
|
1x1 convolution layer to reduce the number of output channels to c. |
m |
ModuleList
|
List of PSA blocks for feature extraction. |
Methods:
Name | Description |
---|---|
forward |
Performs a forward pass through the C2fPSA module. |
forward_split |
Performs a forward pass using split() instead of chunk(). |
Examples:
>>> import torch
>>> from ultralytics.models.common import C2fPSA
>>> model = C2fPSA(c1=64, c2=64, n=3, e=0.5)
>>> x = torch.randn(1, 64, 128, 128)
>>> output = model(x)
>>> print(output.shape)
Parameters:
Name | Type | Description | Default |
---|---|---|---|
c1
|
int
|
Input channels. |
required |
c2
|
int
|
Output channels. |
required |
n
|
int
|
Number of PSABlock modules. |
1
|
e
|
float
|
Expansion ratio. |
0.5
|
Source code in ultralytics/nn/modules/block.py
1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 |
|
ultralytics.nn.modules.block.SCDown
SCDown(c1, c2, k, s)
Bases: Module
SCDown module for downsampling with separable convolutions.
This module performs downsampling using a combination of pointwise and depthwise convolutions, which helps in efficiently reducing the spatial dimensions of the input tensor while maintaining the channel information.
Attributes:
Name | Type | Description |
---|---|---|
cv1 |
Conv
|
Pointwise convolution layer that reduces the number of channels. |
cv2 |
Conv
|
Depthwise convolution layer that performs spatial downsampling. |
Methods:
Name | Description |
---|---|
forward |
Applies the SCDown module to the input tensor. |
Examples:
>>> import torch
>>> from ultralytics import SCDown
>>> model = SCDown(c1=64, c2=128, k=3, s=2)
>>> x = torch.randn(1, 64, 128, 128)
>>> y = model(x)
>>> print(y.shape)
torch.Size([1, 128, 64, 64])
Parameters:
Name | Type | Description | Default |
---|---|---|---|
c1
|
int
|
Input channels. |
required |
c2
|
int
|
Output channels. |
required |
k
|
int
|
Kernel size. |
required |
s
|
int
|
Stride. |
required |
Source code in ultralytics/nn/modules/block.py
1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 |
|
forward
forward(x)
Apply convolution and downsampling to the input tensor.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
x
|
Tensor
|
Input tensor. |
required |
Returns:
Type | Description |
---|---|
Tensor
|
Downsampled output tensor. |
Source code in ultralytics/nn/modules/block.py
1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 |
|
ultralytics.nn.modules.block.TorchVision
TorchVision(model, weights='DEFAULT', unwrap=True, truncate=2, split=False)
Bases: Module
TorchVision module to allow loading any torchvision model.
This class provides a way to load a model from the torchvision library, optionally load pre-trained weights, and customize the model by truncating or unwrapping layers.
Attributes:
Name | Type | Description |
---|---|---|
m |
Module
|
The loaded torchvision model, possibly truncated and unwrapped. |
Parameters:
Name | Type | Description | Default |
---|---|---|---|
model
|
str
|
Name of the torchvision model to load. |
required |
weights
|
str
|
Pre-trained weights to load. Default is "DEFAULT". |
'DEFAULT'
|
unwrap
|
bool
|
If True, unwraps the model to a sequential containing all but the last |
True
|
truncate
|
int
|
Number of layers to truncate from the end if |
2
|
split
|
bool
|
Returns output from intermediate child modules as list. Default is False. |
False
|
Parameters:
Name | Type | Description | Default |
---|---|---|---|
model
|
str
|
Name of the torchvision model to load. |
required |
weights
|
str
|
Pre-trained weights to load. |
'DEFAULT'
|
unwrap
|
bool
|
Whether to unwrap the model. |
True
|
truncate
|
int
|
Number of layers to truncate. |
2
|
split
|
bool
|
Whether to split the output. |
False
|
Source code in ultralytics/nn/modules/block.py
1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 |
|
forward
forward(x)
Forward pass through the model.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
x
|
Tensor
|
Input tensor. |
required |
Returns:
Type | Description |
---|---|
Tensor | List[Tensor]
|
Output tensor or list of tensors. |
Source code in ultralytics/nn/modules/block.py
1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 |
|
ultralytics.nn.modules.block.AAttn
AAttn(dim, num_heads, area=1)
Bases: Module
Area-attention module for YOLO models, providing efficient attention mechanisms.
This module implements an area-based attention mechanism that processes input features in a spatially-aware manner, making it particularly effective for object detection tasks.
Attributes:
Name | Type | Description |
---|---|---|
area |
int
|
Number of areas the feature map is divided. |
num_heads |
int
|
Number of heads into which the attention mechanism is divided. |
head_dim |
int
|
Dimension of each attention head. |
qkv |
Conv
|
Convolution layer for computing query, key and value tensors. |
proj |
Conv
|
Projection convolution layer. |
pe |
Conv
|
Position encoding convolution layer. |
Methods:
Name | Description |
---|---|
forward |
Applies area-attention to input tensor. |
Examples:
>>> attn = AAttn(dim=256, num_heads=8, area=4)
>>> x = torch.randn(1, 256, 32, 32)
>>> output = attn(x)
>>> print(output.shape)
torch.Size([1, 256, 32, 32])
Parameters:
Name | Type | Description | Default |
---|---|---|---|
dim
|
int
|
Number of hidden channels. |
required |
num_heads
|
int
|
Number of heads into which the attention mechanism is divided. |
required |
area
|
int
|
Number of areas the feature map is divided, default is 1. |
1
|
Source code in ultralytics/nn/modules/block.py
1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 |
|
forward
forward(x)
Process the input tensor through the area-attention.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
x
|
Tensor
|
Input tensor. |
required |
Returns:
Type | Description |
---|---|
Tensor
|
Output tensor after area-attention. |
Source code in ultralytics/nn/modules/block.py
1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 |
|
ultralytics.nn.modules.block.ABlock
ABlock(dim, num_heads, mlp_ratio=1.2, area=1)
Bases: Module
Area-attention block module for efficient feature extraction in YOLO models.
This module implements an area-attention mechanism combined with a feed-forward network for processing feature maps. It uses a novel area-based attention approach that is more efficient than traditional self-attention while maintaining effectiveness.
Attributes:
Name | Type | Description |
---|---|---|
attn |
AAttn
|
Area-attention module for processing spatial features. |
mlp |
Sequential
|
Multi-layer perceptron for feature transformation. |
Methods:
Name | Description |
---|---|
_init_weights |
Initializes module weights using truncated normal distribution. |
forward |
Applies area-attention and feed-forward processing to input tensor. |
Examples:
>>> block = ABlock(dim=256, num_heads=8, mlp_ratio=1.2, area=1)
>>> x = torch.randn(1, 256, 32, 32)
>>> output = block(x)
>>> print(output.shape)
torch.Size([1, 256, 32, 32])
Parameters:
Name | Type | Description | Default |
---|---|---|---|
dim
|
int
|
Number of input channels. |
required |
num_heads
|
int
|
Number of heads into which the attention mechanism is divided. |
required |
mlp_ratio
|
float
|
Expansion ratio for MLP hidden dimension. |
1.2
|
area
|
int
|
Number of areas the feature map is divided. |
1
|
Source code in ultralytics/nn/modules/block.py
1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 |
|
forward
forward(x)
Forward pass through ABlock.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
x
|
Tensor
|
Input tensor. |
required |
Returns:
Type | Description |
---|---|
Tensor
|
Output tensor after area-attention and feed-forward processing. |
Source code in ultralytics/nn/modules/block.py
1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 |
|
ultralytics.nn.modules.block.A2C2f
A2C2f(
c1,
c2,
n=1,
a2=True,
area=1,
residual=False,
mlp_ratio=2.0,
e=0.5,
g=1,
shortcut=True,
)
Bases: Module
Area-Attention C2f module for enhanced feature extraction with area-based attention mechanisms.
This module extends the C2f architecture by incorporating area-attention and ABlock layers for improved feature processing. It supports both area-attention and standard convolution modes.
Attributes:
Name | Type | Description |
---|---|---|
cv1 |
Conv
|
Initial 1x1 convolution layer that reduces input channels to hidden channels. |
cv2 |
Conv
|
Final 1x1 convolution layer that processes concatenated features. |
gamma |
Parameter | None
|
Learnable parameter for residual scaling when using area attention. |
m |
ModuleList
|
List of either ABlock or C3k modules for feature processing. |
Methods:
Name | Description |
---|---|
forward |
Processes input through area-attention or standard convolution pathway. |
Examples:
>>> m = A2C2f(512, 512, n=1, a2=True, area=1)
>>> x = torch.randn(1, 512, 32, 32)
>>> output = m(x)
>>> print(output.shape)
torch.Size([1, 512, 32, 32])
Parameters:
Name | Type | Description | Default |
---|---|---|---|
c1
|
int
|
Number of input channels. |
required |
c2
|
int
|
Number of output channels. |
required |
n
|
int
|
Number of ABlock or C3k modules to stack. |
1
|
a2
|
bool
|
Whether to use area attention blocks. If False, uses C3k blocks instead. |
True
|
area
|
int
|
Number of areas the feature map is divided. |
1
|
residual
|
bool
|
Whether to use residual connections with learnable gamma parameter. |
False
|
mlp_ratio
|
float
|
Expansion ratio for MLP hidden dimension. |
2.0
|
e
|
float
|
Channel expansion ratio for hidden channels. |
0.5
|
g
|
int
|
Number of groups for grouped convolutions. |
1
|
shortcut
|
bool
|
Whether to use shortcut connections in C3k blocks. |
True
|
Source code in ultralytics/nn/modules/block.py
1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 |
|
forward
forward(x)
Forward pass through A2C2f layer.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
x
|
Tensor
|
Input tensor. |
required |
Returns:
Type | Description |
---|---|
Tensor
|
Output tensor after processing. |
Source code in ultralytics/nn/modules/block.py
1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 |
|
ultralytics.nn.modules.block.SwiGLUFFN
SwiGLUFFN(gc, ec, e=4)
Bases: Module
SwiGLU Feed-Forward Network for transformer-based architectures.
Source code in ultralytics/nn/modules/block.py
1883 1884 1885 1886 1887 |
|
forward
forward(x)
Apply SwiGLU transformation to input features.
Source code in ultralytics/nn/modules/block.py
1889 1890 1891 1892 1893 1894 |
|
ultralytics.nn.modules.block.Residual
Residual(m)
Bases: Module
Residual connection wrapper for neural network modules.
Source code in ultralytics/nn/modules/block.py
1900 1901 1902 1903 1904 1905 1906 1907 |
|
forward
forward(x)
Apply residual connection to input features.
Source code in ultralytics/nn/modules/block.py
1909 1910 1911 |
|
ultralytics.nn.modules.block.SAVPE
SAVPE(ch, c3, embed)
Bases: Module
Spatial-Aware Visual Prompt Embedding module for feature enhancement.
Source code in ultralytics/nn/modules/block.py
1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 |
|
forward
forward(x, vp)
Process input features and visual prompts to generate enhanced embeddings.
Source code in ultralytics/nn/modules/block.py
1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 |
|