Guide · Open source
How to use External-Attention-pytorch: install and run modules
Updated · written with AI from the project's own documentation and checked against our catalogue data
External-Attention-pytorch is an MIT-licensed Python library of PyTorch re-implementations of attention mechanisms, MLP variants, backbones, re-parameter blocks and convolution layers taken from research papers. It is aimed at students, researchers and developers who want to read or reuse a paper's core module without the surrounding training code. After this guide you will know how to install it, import a module and run it on a tensor.
FightingCVxmu-xiaoma666/External-Attention-pytorch · ★ 12.2k · MIT Install options & downloads Quick start
- 1Check prerequisites. Make sure Python and PyTorch are installed, since the library needs both.
- 2Install the package. Run pip install fightingcv-attention to install from PyPI.
- 3Or clone the repository. Use git clone on the GitHub URL if you want to read or edit the source.
- 4Import a module. Import the module you need, such as MobileViTv2Attention, using fightingcv_attention or model depending on your install route.
- 5Run it on a tensor. Create a random input with torch.randn, build the module with its arguments and call it to get the output.
- 6Read the source. Open the module file to see the core code of the paper you are studying.
What the library contains
The repository is organised into series: Attention, Backbone, MLP, Re-Parameter (ReP) and Convolution. Each series lists the papers it implements, such as External Attention, Squeeze-and-Excitation, CBAM, ECA-Net and CoAtNet in the Attention series, and RepVGG and ACNet in the Re-Parameter series.
Each module is meant to be a self-contained piece of code that you can read or drop into your own model. The project's stated goal is to make the core code of a paper easy to find, rather than to provide trained models.
- Attention Series: modules such as SE, SK, CBAM, ECA and DANet
- Backbone Series: modules such as ResNet, ResNeXt and MobileViT
- MLP Series: modules such as RepMLP, MLP-Mixer and ResMLP
- Re-Parameter Series: modules such as RepVGG and ACNet
- Convolution Series: modules such as Depthwise Separable Convolution and Involution
Install External-Attention-pytorch with pip or git
The README gives two install routes. The simplest is to install the package from PyPI under the name fightingcv-attention. Alternatively, you can clone the repository and import the modules from the local model folder.
Both routes need Python and PyTorch. The README badges mention Python and PyTorch version floors, but you should check the repository for the exact requirements.
- Route 1: install from PyPI with the pip command below
- Route 2: clone the repository and work from its folder
pip install fightingcv-attentionClone the repository instead of using pip
If you want to read or edit the source, clone the repository. The README says the git version differs from the pip version only in the import path: you replace fightingcv_attention with model.
git clone https://github.com/xmu-xiaoma666/External-Attention-pytorch.git
cd External-Attention-pytorchRun your first module
Each module is a PyTorch layer that takes a tensor and returns a tensor. The README example uses the MobileViTv2Attention module with a random input of shape 50 by 49 by 512 and prints the output shape.
With pip, import from fightingcv_attention. With the git clone, import from model. The rest of the code is identical.
- Create an input tensor with torch.randn
- Instantiate the module with its arguments, here d_model=512
- Call the module on the input and print the output shape
import torch
from torch import nn
from torch.nn import functional as F
from fightingcv_attention.attention.MobileViTv2Attention import *
if __name__ == '__main__':
input=torch.randn(50,49,512)
sa = MobileViTv2Attention(d_model=512)
output=sa(input)
print(output.shape)Use a module as a drop-in layer in your own model
Because modules are standalone, you can place one inside your own nn.Module, for example after a convolution block. The README says the pip package includes a reference document for its built-in modules, linked as README_pip.md in the repository.
Check the input shape each module expects before wiring it in, since the example uses a sequence-style input of batch, tokens and channels.
- Read the module's source file to see its constructor arguments
- Match the input shape to the example in the README
- Compare variants by swapping one import for another
What people use it for
- Students reading a paper on channel attention can open the SE or ECA module and read the core code directly.
- Researchers can reuse a published attention block such as CBAM in their own model without rewriting it.
- Developers prototyping a classifier can try a drop-in attention layer and check the output shape.
- Engineers can compare several attention variants by importing different modules with the same call pattern.
Frequently asked questions
How do I install External-Attention-pytorch?
The README says you can run pip install fightingcv-attention to install from PyPI. You can also clone the GitHub repository with git clone and work from its folder.
Is External-Attention-pytorch free to use?
The repository is listed under the MIT licence. Check the LICENSE file in the repository for the full terms before using it in a product.
What do I need to run it?
You need Python and PyTorch. The README badges list minimum Python and PyTorch versions, so check the repository for exact requirements.
Does it include pretrained weights?
No. The README describes module code, not trained checkpoints. If you need a finished model with weights, look at a library such as timm.
Does it reproduce the paper results?
The README does not claim reproduced benchmark numbers. Use it to understand the module code, and not as a way to replicate a paper's reported accuracy.
What are alternatives to External-Attention-pytorch?
The README lists timm, lucidrains/vit-pytorch and pytorch-image-models as comparable projects. Choose based on whether you want core module code, a library of pretrained models, or ViT-specific implementations.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
External-Attention-pytorch suits students and researchers who want to read the core code of attention, MLP and convolution modules, and developers who want a simple drop-in layer. It is a poor choice if you need pretrained weights, an exact replica of a paper's results, or you have no PyTorch background yet.
Install options & downloads