</> Code with Papers by Syntology

Methods › General › Attention Modules › Spatial-Reduction Attention

Spatial-Reduction Attention

29 papers tagged archive 2025-07-28

Introduced by Wenhai Wang et al. in Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions

archive 2025-07-28 Description, source and code snippet are the archive's method entry.

Spatial-Reduction Attention, or SRA, is a multi-head attention module used in the Pyramid Vision Transformer architecture which reduces the spatial scale of the key K and value V before the attention operation. This reduces the computational/memory overhead. Details of the SRA in the stage i can be formulated as follows:

SRA(Q, K, V)= Concat (head₀, … head _(Nᵢ)) Wᴼ

headⱼ= Attention (Q Wⱼ^Q, SR(K) Wⱼᴷ, SR(V) Wⱼⱽ)

where Concat (·) is the concatenation operation. Wⱼ^Q ∈ℝ^(Cᵢ ×d_(head)), Wⱼᴷ ∈ℝ^(Cᵢ ×d_(head)), Wⱼⱽ ∈ℝ^(Cᵢ ×d_(head)), and Wᴼ ∈ℝ^(Cᵢ ×Cᵢ) are linear projection parameters. Nᵢ is the head number of the attention layer in Stage i. Therefore, the dimension of each head (i.e. .d_(head)) is equal to Cᵢ/Nᵢ . SR(·) is the operation for reducing the spatial dimension of the input sequence (K or V ), which is written as:

SR(𝐱)=Norm(Reshape(𝐱, Rᵢ) W^S)

Here, 𝐱 ∈ℝ^((Hᵢ Wᵢ) ×Cᵢ) represents a input sequence, and Rᵢ denotes the reduction ratio of the attention layers in Stage i . Reshape (𝐱, Rᵢ) is an operation of reshaping the input sequence 𝐱 to a sequence of size (Hᵢ Wᵢ)/Rᵢ² ×(Rᵢ² Cᵢ). W_S ∈ℝ^((Rᵢ² Cᵢ) ×Cᵢ) is a linear projection that reduces the dimension of the input sequence to Cᵢ. Norm(·) refers to layer normalization.

PaperSource

Papers archive 2025-07-28

29 shown of 29, newest first. Repository counts are the archive's code-links table. A Syntology line states what Syntology ran from that paper's harvested code; it is per sample and not a correctness claim.

Tasks archive 2025-07-28

20 shown of 43 tasks the archive attaches to papers tagged with this method, by distinct papers. A task without a page in the catalog is plain text.

TaskPapers
Semantic Segmentation7
Object Detection6
Segmentation4
Image Classification3
Instance Segmentation3
Inductive Bias2
Medical Image Segmentation2
Quantization2
Self-Supervised Learning2
image-classification2
object-detection2
3D Object Detection1
3D Part Segmentation1
3D Point Cloud Classification1
Action Recognition1
Autonomous Driving1
Computational Efficiency1
Continual Semantic Segmentation1
Data Augmentation1
Decoder1

Usage over time archive 2025-07-28

Papers per year tagged with Spatial-Reduction Attention: 2021 to 2025, peak 11 11 0 2021: 4 papers 2021 2022: 5 papers 2022 2023: 11 papers 2023 2024: 7 papers 2024 2025: 2 papers 2025
Papers per year the archive tags with this method, by the paper's archive date (29 dated). Bars are counts, not a trend claim.

Components: the archive holds no method-to-method composition, so PwC's Components table cannot be rebuilt; the Papers list carries no Results column for the same reason (the archive does not join its leaderboard rows to method tags).

Categories archive 2025-07-28

Attention Modules

Report a problem or propose a change · a person checks every report against the paper or source before anything changes; decisions are listed on /corrections