https://github.com/rmitsch/tale

Tool for Annotation of Low-dimensional Embeddings.
https://github.com/rmitsch/tale

Last synced: 2 days ago
JSON representation

Tool for Annotation of Low-dimensional Embeddings.

• Host: GitHub
• URL: https://github.com/rmitsch/tale
• Owner: rmitsch
• License: mit
• Created: 2020-02-07T14:21:35.000Z (almost 5 years ago)
• Default Branch: master
• Last Pushed: 2021-03-05T19:21:18.000Z (almost 4 years ago)
• Last Synced: 2024-12-29T10:44:30.908Z (5 days ago)
• Language: Python
• Size: 502 KB
• Stars: 1
• Watchers: 3
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# TALE - Tool for Annotating of Low-dimensional Embeddings

TALE is a **T**ool for **A**nnotation of **L**ow-dimensional **E**mbeddings. It offers functionality to assess, interprete and rate low-dimensional projections, such as those generated by e.g. t-SNE or UMAP. See [todo - add paper link](www.arxiv.org) for a more complete description. It is written in Python (backend) and Javascript (frontend).

This repository contains the [dataset with projection features and user ratings](https://github.com/rmitsch/TALE/blob/master/rated_projection_features.pkl) discussed in the paper.

Head over to https://github.com/rmitsch/TALE-backend and https://github.com/rmitsch/TALE-frontend for the actual source code.

TALE allows to explore the parameter space of low-dimensional projections in the global view:

![TALE: Global view](https://github.com/rmitsch/TALE/blob/master/doc/tale_global.png)

Individual projections can be inspected, evaluated and rated in the local view:

![TALE: Local view](https://github.com/rmitsch/TALE/blob/master/doc/tale_local.png)

## Build Instructions

* Pull source code:

`git clone --recurse-submodules [email protected]:rmitsch/TALE.git`

* Build the Docker image:

`docker build -t tale -f Dockerfile .`

* Alternatively pull the image from Dockerhub:

`docker pull rmitsch/tale`

## Generate projections

`docker run -v [host data directory]:/data tale python /TALE-backend/source/generate_data.py [dataset name] [DR kernel name] /data`    

`[dataset name]` can be either "happiness" for the UN world happiness study or "movie" for the IMDB movie dataset.

`[DR kernel name]` can be "UMAP", "TSNE" or "SVD".

## Run TALE server

`docker run -p 2484:2484 -v [host data directory]:/data tale python /TALE-backend/source/app.py /TALE-frontend /data [experiment name] [Dropbox OAuth Token]`

`[experiment name]` and `[Dropbox OAuth Token]` are optional and only necessary if you want to hook up TALE to a Dropbox account to automatically store the resulting user ratings in the cloud.

## Use TALE

Access in your browser via localhost:2484.

Note: By default, TALE attempts to load t-SNE projections for the world happiness dataset, i. e. assumes that projections have been generated with 

`docker run -v [host data directory]:/data tale python /TALE-backend/source/generate_data.py happiness TSNE /data`. If you want to look at another configuration, select it in the dataset and DR kernel dropdowns to the top right and click the load button to their right.