https://github.com/lehinevych/saan

https://github.com/lehinevych/saan

Last synced: 9 days ago
JSON representation

• Host: GitHub
• URL: https://github.com/lehinevych/saan
• Owner: lehinevych
• License: mit
• Created: 2019-12-05T11:59:15.000Z (almost 5 years ago)
• Default Branch: master
• Last Pushed: 2022-12-10T23:40:26.000Z (almost 2 years ago)
• Last Synced: 2023-04-03T15:56:21.872Z (over 1 year ago)
• Language: HTML
• Size: 8.28 MB
• Stars: 0
• Watchers: 6
• Forks: 2
• Open Issues: 7
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        

# SAAN

This is an implementation of the Self-attention aggregation network as described in the [SAAN](https://arxiv.org/abs/2010.05340) paper. 

SAAN is a neural network architecture that solves face template aggregation problems using self-attention mechanisms. 

Particularly, we employ [Transformer](https://github.com/tensorflow/tensor2tensor) implementation for the sequence encoding.

## Contents

  * [Contents](#contents)

   * [Detailed information](#detailed-information)

   * [Results on benchmarks](#model-training-and-evaluation)

## Detailed information

1. ### Model training and evaluation

   There are two notebooks which demonstrate [single](Single_identity_aggregation_and_attention_dist.ipynb) and 

   [multi-identity](Multi_identity_aggregation_model_for_repo.ipynb) aggregation models with the respective training and validation pipelines created using tf.estimator and tf.Dataset APIs. The aggregation architecture is shared and could be found in [aggregator.py](aggregator_utils/aggregator.py). 

   [config.py](aggregator_utils/config.py) specifies  the configurations of the data sampler and different aggregators.

2. ### Results on benchmarks

   [IJB notebook](IJB_preprocessing_and_becnhmarking_for_repo.ipynb) contains the complete report with the respective visualization on identification and verificaiton metrics using IJB-C benchmark.

**Note**: It is advised to download the repository and display the .html files via the browsers in order to view the results.