Skip to content

Latest commit

 

History

History
35 lines (25 loc) · 1.5 KB

File metadata and controls

35 lines (25 loc) · 1.5 KB

Transformer-style multi-head attention

Overview

This code implements transformer-style (dot-product) multi-head attention, with different variants and optional attention score leaks.

Some publications in the GNN context that either use this multi-head attention as a component ([1]&[2]) or a baseline ([3]) of their method:

Usage

TensorFlow programs can import and use this model as described in its API docs.

API stability

The API of this model may change between OSS library versions.

TF-GNN's Model Template "Albis" offers a stable and simplified API for a subset of this model's configuration options.