philipperemy/keras-attention

Implements both multiplicative (Luong) and additive (Bahdanau) attention mechanisms as a reusable Keras layer compatible with TensorFlow 2.0+, enabling dynamic focus on sequence elements. The layer accepts 3D sequential input and outputs attention-weighted context vectors, integrating seamlessly into RNN/LSTM architectures for tasks like machine translation and document classification. Includes model serialization support and visualization capabilities for interpreting attention weights across timesteps.

2,815 stars. Actively maintained with 1 commit in the last 30 days.