Compositional Attention

This repository is an implementation of Compositional Attention: Disentangling Search and Retrieval by MILA. Revisiting standard Multi-head attention through the lens of multiple parallel and independent search and retrieval mechanisms, this leads to static pairings between searches and retrievals, often leading to redundancy of parameters. They reframe the "heads" of multi-head attention as "searches", and once the multi-headed/searched values are aggregated, there is an extra retrieval step (using attention) off the searched results. The experiments establish this as an easy drop-in replacement for Multi-head attention.

Installation

Run the following to install:

pip install compositional-attention

Developing compositional-attention

To install compositional-attention, along with tools you need to develop and test, run the following in your virtualenv:

git clone https://github.com/Rishit-dagli/compositional-attention.git
# or clone your own fork

cd compositional-attention
pip install -e .[dev]

Usage

import tensorflow as tf
from compositional_attention import CompositionalAttention

attn = CompositionalAttention(
    dim = 1024,            # input dimension
    dim_head = 64,         # dimension per attention 'head' - head is now either search or retrieval
    num_searches = 8,      # number of searches
    num_retrievals = 2,    # number of retrievals
    dropout = 0.1,          # dropout of attention of search and retrieval
)

tokens = tf.random.uniform([1, 512, 1024])  # tokens
mask = tf.ones([1, 512], dtype=tf.dtypes.bool)  # mask

out = attn(tokens, mask = mask) # (1, 512, 1024)

Want to Contribute 🙋‍♂️?

Awesome! If you want to contribute to this project, you're always welcome! See Contributing Guidelines. You can also take a look at open issues for getting more information about current or upcoming tasks.

Want to discuss? 💬

Have any questions, doubts or want to present your opinions, views? You're always welcome. You can start discussions.

Citations

Mittal, Sarthak, et al. ‘Compositional Attention: Disentangling Search and Retrieval’. ArXiv:2110.09419 [Cs], Feb. 2022. arXiv.org, http://arxiv.org/abs/2110.09419.

Official PyTorch implmentation and Phil Wang's PyTorch implmenetation.