AttentionModules

A compact PyTorch toolbox of visual attention modules for CNNs, hybrid models, and Transformer-style architectures.

This repository collects a range of attention mechanisms commonly used in computer vision, from lightweight channel attention blocks to global self-attention and deformable attention. It is designed for two main goals:

• Plug-and-play experimentation in your own models
• Code reading and learning for understanding how different attention modules are implemented in PyTorch

✨ Highlights

• Clean PyTorch implementations with readable structure
• Most modules use the standard (B, C, H, W) input/output convention
• Includes lightweight, axis-aware, hybrid, and Transformer-style attention blocks
• A good reference repo for comparing implementation ideas across attention families

🗂 Repository Structure

AttentionModules/
├── AttentionModules/
│ ├── A2.py
│ ├── ACmix.py
│ ├── BAM.py
│ ├── CA.py
│ ├── CBAM.py
│ ├── CCAM.py
│ ├── DANet.py
│ ├── DETRAttention.py
│ ├── DeformableAttention.py
│ ├── ECA.py
│ ├── ELA.py
│ ├── EMA.py
│ ├── GAM.py
│ ├── SCSA.py
│ ├── SE.py
│ ├── SimAM.py
│ ├── SK.py
│ ├── SLAM.py
│ ├── SwinAttention.py
│ └── TripletAttention.py
└── README.md

🚀 Quick Start

Install

Requirements:

• Python 3.8+
• PyTorch 2.x

If PyTorch is already installed, this repo is ready to use directly.

Import a module

fromAttentionModulesimportSE, CA, CBAM, ECAattn=SE(channels=64, reduction=4)
out=attn(x) # x: (B, C, H, W)

Insert into a convolution block

importtorch.nnasnnfromAttentionModulesimportCBAMclassConvBlock(nn.Module):
def__init__(self, c1, c2):
super().__init__()
self.block=nn.Sequential(
nn.Conv2d(c1, c2, 3, padding=1, bias=False),
nn.BatchNorm2d(c2),
nn.ReLU(inplace=True),
CBAM(c2),
)
defforward(self, x):
returnself.block(x)

Transformer-style example

fromAttentionModulesimportDETRAttentionattn=DETRAttention(d_model=256, num_heads=8)
out=attn(query) # query: (B, N, 256)

🧠 Module Landscape

Attention family map

flowchart TD
A[AttentionModules] --> B[Channel Attention]
A --> C[Spatial / Axis Attention]
A --> D[Hybrid CNN Attention]
A --> E[Self-Attention / Transformer]
B --> B1[SE]
B --> B2[ECA]
B --> B3[SK]
B --> B4[SimAM]
C --> C1[CA]
C --> C2[ELA]
C --> C3[TripletAttention]
C --> C4[SLAM]
D --> D1[CBAM]
D --> D2[BAM]
D --> D3[GAM]
D --> D4[EMA]
D --> D5[SCSA]
D --> D6[CCAM]
D --> D7[ACmix]
D --> D8[A2]
D --> D9[DANet]
E --> E1[SwinAttention]
E --> E2[DETRAttention]
E --> E3[DeformableAttention]

Loading

Comparison table

Module	Family	Main idea	Typical implementation pattern
SE	Channel	Learn per-channel importance from global context	GAP + bottleneck + sigmoid
ECA	Channel	Lightweight local channel interaction without dimensionality reduction	GAP + Conv1d
SK	Channel / Multi-scale	Adaptive receptive field selection across branches	multi-branch conv + softmax
SimAM	Parameter-free	Element-wise saliency from statistics	variance-based energy score
CA	Coordinate / Axis	Encode positional information through H/W decomposition	axis pooling + split projection
ELA	Axis	Independent height and width attention generation	mean pooling + directional conv
TripletAttention	Axis / Cross-dimension	Reuse one gate under different tensor permutations	permute + shared gate
SLAM	Spatial / Axis hybrid	Jointly model height, width, and spatial saliency	three-branch gating
CBAM	Hybrid CNN	Sequential channel then spatial refinement	channel gate -> spatial gate
BAM	Hybrid CNN	Parallel channel and spatial bottleneck attention	dilated conv + residual gate
GAM	Hybrid CNN	Stronger channel interaction before spatial modeling	MLP-like channel + conv spatial
EMA	Hybrid CNN	Grouped axis-aware gating with spatial response mixing	group split + axis gate + spatial mix
SCSA	Hybrid CNN	Multi-scale spatial-axis attention plus pooled channel self-attention	MS Conv1d + pooled self-attn
CCAM	Hybrid CNN	Channel attention followed by coordinate attention	channel gate + coordinate gate
A2	Global attention	Gather global descriptors, then redistribute them	gather-distribute
DANet	Global attention	Parallel position attention and channel attention	PAM + CAM
ACmix	Conv-attention fusion	Fuse local self-attention and dynamic convolution	shared qkv + fusion weights
SwinAttention	Transformer	Window-based attention with optional shift	window partition + relative bias
DETRAttention	Transformer	Standard multi-head attention for self/cross attention	qkv linear + scaled dot-product
DeformableAttention	Transformer	Sparse learned sampling instead of dense attention	offset + sparse sampling

📊 At-a-Glance Selection Guide

flowchart LR
A[Need an attention block] --> B{What do you want most?}
B --> C[Low overhead]
B --> D[Better spatial / axis awareness]
B --> E[Stronger global modeling]
B --> F[Transformer-style research]
C --> C1[SE]
C --> C2[ECA]
C --> C3[SimAM]
C --> C4[CA]
D --> D1[CA]
D --> D2[ELA]
D --> D3[TripletAttention]
D --> D4[SLAM]
D --> D5[SCSA]
E --> E1[A2]
E --> E2[DANet]
E --> E3[ACmix]
F --> F1[SwinAttention]
F --> F2[DETRAttention]
F --> F3[DeformableAttention]

Loading

🔍 Implementation Notes by Module

This section focuses on how each module is implemented in this repo rather than only what the original paper proposes.

1. SE

• Applies global average pooling to compress (B, C, H, W) into (B, C, 1, 1)
• Uses a bottleneck MLP implemented with two 1x1 convolutions
• Produces channel weights with sigmoid and rescales the input channel-wise

Best for: a simple, stable, classic channel attention baseline.

2. ECA

• Starts from global average pooled channel descriptors
• Avoids explicit dimensionality reduction
• Uses Conv1d over the channel descriptor sequence for local channel interaction

Best for: lightweight models and low-overhead attention insertion.

3. SK

• Builds several convolution branches with different kernel sizes
• Sums branch outputs to form a fused global descriptor
• Uses a shared compression layer and branch-specific projections
• Applies softmax across branches to adaptively choose receptive field scale

Best for: multi-scale feature selection.

4. SimAM

• Adds no extra convolution or MLP parameters
• Computes element importance from mean, variance, and deviation statistics
• Generates an element-wise attention map directly from a closed-form score

Best for: parameter-free attention experiments.

5. CA

• Pools separately along width and height
• Concatenates directional descriptors before a shared projection
• Splits them back into H-branch and W-branch attention maps
• Reweights features using both directional gates

Best for: preserving positional cues with low cost.

6. ELA

• Decomposes spatial modeling into two axis-specific branches
• Uses directional pooling and dedicated convolutional transforms
• Produces independent height-aware and width-aware attention maps

Best for: axis-aware local enhancement.

7. TripletAttention

• Defines a reusable AttentionGate
• Inside the gate, ZPool concatenates channel-wise average and max maps
• Uses the same gating unit under different tensor permutations
• Aggregates information from HW, HC, and WC interaction views

Best for: cross-dimension interaction with small overhead.

8. SLAM

• Builds three branches: height-aware, width-aware, and spatial
• Each branch generates its own gate
• Multiplies all gates back into the input feature map

Best for: intuitive multi-view spatial refinement.

9. CBAM

• First computes channel attention using both average pooling and max pooling
• Then computes spatial attention from channel-compressed feature maps
• Applies the two stages sequentially

Best for: a strong and widely used CNN attention baseline.

10. BAM

• Uses a channel bottleneck branch and a spatial branch in parallel
• Spatial modeling relies on dilated convolutions to enlarge receptive field
• Final output is residual-style: x * (1 + M)

Best for: stable parallel channel-spatial refinement.

11. GAM

• Performs a stronger channel interaction stage before spatial refinement
• Treats channel attention less like pure pooling and more like learned mapping
• Follows with convolutional spatial reweighting

Best for: global channel interaction plus spatial enhancement.

12. EMA

• Splits channels into groups to reduce cost
• Applies axis-aware gating per group
• Combines grouped directional information with spatial response mixing

Best for: efficient multi-axis fusion.

13. SCSA

• Spatial part uses multi-scale depthwise Conv1d branches over H/W directions
• Channel part uses pooled self-attention on a reduced spatial grid
• Combines directional spatial gating with pooled channel interaction

Best for: stronger expressive power than classic lightweight modules.

14. CCAM

• Applies channel attention first
• Then performs coordinate-style H/W attention on the refined features
• Works like a chained combination of CAM and coordinate attention

Best for: channel-first coordinate enhancement.

15. A2

• Uses three 1x1 projections
• First attention stage gathers global descriptors
• Second stage redistributes those descriptors back to spatial locations

Best for: global context aggregation with a clear gather-distribute design.

16. DANet

• Includes two parallel branches:
• PAM for spatial position attention
• CAM for channel attention
• Fuses both outputs after residual-style weighting

Best for: stronger global dependency modeling in dense prediction tasks.

17. ACmix

• Shares q, k, v projections for both branches
• One branch performs local self-attention with unfolded neighborhoods
• The other branch behaves like dynamic convolution
• Learns how much to trust each branch through trainable fusion weights

Best for: blending the strengths of convolution and attention.

18. SwinAttention

• Converts input from (B, C, H, W) to (B, H, W, C) internally
• Splits features into local windows
• Runs multi-head self-attention inside each window
• Supports shifted windows and relative position bias

Best for: learning the core idea behind Swin Transformer attention.

19. DETRAttention

• Standard multi-head attention implementation
• Supports both self-attention and cross-attention
• Uses linear Q/K/V projections, scaled dot-product attention, output projection, and normalization

Best for: understanding the core attention block used in DETR-style models.

20. DeformableAttention

• Flattens feature maps into token sequences
• Predicts sparse sampling offsets and attention weights for each query
• Samples features with grid_sample
• Aggregates only a small number of learned positions instead of attending densely everywhere

Best for: efficient sparse attention over images.

🧩 A Shared Mental Model

Even though the module names are different, many of them follow a similar pipeline:

flowchart LR
X[Input feature map] --> P[Descriptor generation]
P --> G[Attention weight generation]
G --> R[Feature recalibration]
P1[GAP / GMP] --> G
P2[Axis pooling] --> G
P3[QKV projection] --> G
P4[Variance / energy score] --> G
G1[MLP / 1x1 Conv] --> R
G2[Conv / DWConv / Dilated Conv] --> R
G3[Softmax attention] --> R
G4[Sigmoid gate] --> R

Loading

This is a useful way to read the code: first ask how the descriptor is built, then how the attention map is generated, and finally how the original features are reweighted.

🧪 Input / Output Conventions

Most modules follow:

x.shape== (B, C, H, W)
out.shape== (B, C, H, W)

Exceptions:

• DETRAttention expects (B, N, d_model)
• SwinAttention takes (B, C, H, W) externally but internally converts to (B, H, W, C)

🎯 Which Module Should I Start With?

If you want lightweight and easy-to-plug modules

• SE
• ECA
• CA
• CBAM
• SimAM

If you want stronger spatial or axis modeling

• CA
• ELA
• TripletAttention
• SLAM
• SCSA

If you want stronger global dependency modeling

• A2
• DANet
• DETRAttention
• SwinAttention
• DeformableAttention

If you want convolution-attention hybrids

• ACmix
• GAM
• EMA
• BAM

📌 Suggested Next Improvements

This repository already works well as an implementation collection, but it could become an even stronger reference if you add:

• Paper links and publication years for every module
• Parameter count and FLOPs comparison
• Recommended insertion positions in backbone / neck / head
• Standalone structure diagrams for each module
• Minimal runnable demos and benchmark snippets

🤝 Final Note

This repo is already a strong learning-oriented collection for attention modules. With a few more benchmark tables and paper references, it can easily grow from a code toolbox into a polished attention handbook for vision research and engineering.