Keras Attention: Add Luong or Bahdanau attention to Keras

Understand the project

Practical overview

Use it when building a Keras model that needs Luong- or Bahdanau-score attention over a sequence representation.

Insert an attention operation between sequence-encoding and downstream model layers.
Choose Luong or Bahdanau scoring and configure the attention output size.
Use the repository examples as patterns for sequence tasks and attention-weight visualization.

Before you choose

The repository documents installation from PyPI with `pip install attention` and identifies the project as Apache-2.0 licensed.

The layer expects a 3D input tensor shaped (batch_size, timesteps, input_dim) and produces a 2D output shaped (batch_size, num_units).
The score argument selects luong or bahdanau; units sets the output size.
The README reports testing with TensorFlow 2.8 through 2.14.
Saved models using the layer are shown reloaded with Attention passed through custom_objects.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Install the package and assemble the documented sequence-model pattern with an attention layer.

Python environment with TensorFlow/Keras; the README reports testing with TensorFlow 2.8 through 2.14.
A Keras sequence model whose preceding layer returns a 3D tensor shaped (batch_size, timesteps, input_dim).

Install the package with the command below.
Documented command
```
pip install attention
```
Import Attention from attention and create a Keras input.
Pass sequence outputs from a layer such as an LSTM configured with return_sequences=True into Attention(units=32); select score='luong' or score='bahdanau' as needed.
Connect the resulting 2D attention output to a downstream layer, such as the example's Dense layer.

A model graph in which the attention layer maps a 3D sequence tensor to a 2D tensor with the configured number of units.

The guide describes constructing the model pattern; it does not establish that it has been run in a particular environment.
TensorFlow versions outside the README's stated tested range are not covered by that version note.

Evidence and transparency

Source record details

Source availability: Active
Sources checked: 14 October 2026
Original website title: philipperemy/keras-attention · Official homepage title · retained in its source language