cr1337 / human-guided-dimension-reduction

A web application for experimenting with dimensionality reduction with human guidance.

human-guided-dimension-reduction

A web application for experimenting with dimensionality reduction with human guidance.

Setup

Linux

1. Install Docker
2. Clone this Repository using git clone https://github.com/CR1337/human-guided-dimension-reduction.git.
3. cd into this repository with cd human-guided-dimension-reduction.
4. Build the docker services by running bin/build.
5. Create a virtual environment by executing python3 -m venv .venv.
6. Activate the virtual environment with source .venv/bin/activate.
7. Install dependencies using pip3 install -r requirements.txt.
8. Change directory with cd services/backend.
9. Run ./compile-neighbors.
10. Run chmod +x neighbors/neighbors.
11. Change back to the main directory with cd ../...
12. Run python3 util/init_neighbors.py.

Windows

(Not tested, you are on your own.)

1. Install Docker
2. Clone this Repository using git clone https://github.com/CR1337/human-guided-dimension-reduction.git.
3. cd into this repository with cd human-guided-dimension-reduction.
4. Build the docker services by running bin\build.bat.
5. Create a virtual environment by executing python -m venv .venv.
6. Activate the virtual environment with .venv\bin\activate.
7. Install dependencies using pip install -r requirements.txt.
8. Make sure you have g++ installed (at least version 11.4).
9. Change directory with cd services\backend.
10. Run .\compile-neighbors.bat.
11. Change back to the main directory with cd ..\...
12. Run python3 util\init_neighbors.py.

Running the App

Linux

1. Execute bin/run
2. Open your Browser and navigate to http://localhost:8080

Windows

1. Execute bin\run.bat
2. Open your Browser and navigate to http://localhost:8080

human-guided-dimension-reduction's Issues

Re-implement LMDS

Improve rendering performance

Make fps go brrrrr

• Only draw when something changed
• Maybe use shader for drawing

Maybe use shader for rendering

When the plot needs to be redrawn it is still pretty slow. Maybe using a shader for rendering will improve it.

Display previous metrics

Add a table where previously calculated metrics are shown

Add Topic models

https://arxiv.org/abs/2307.11770

Generate Embeddings

Remove neighborhood lines

Remove the lines that indicate neighborhood

Generate Distance Matrices

We have to change the data generation script for model training data to output distance matrices instead of vectors.

Read nearest neighbors from high dimensional space

Add embeddings for new datasets

https://huggingface.co/datasets/dair-ai/emotion

another one

Move landmarks

• Move landmarks with mouse
• Send moved landmarks back to server
• Display landmarks in front of other points

Render cosine neighbors correctly when landmark was moved

Metric

Add new metrics based on the https://pubmed.ncbi.nlm.nih.gov/31567092/ paper.

• Berechnen
• Visualisieren

Neighbor Speed Up

• Speed up calculation of 2D neighbors in backend using C
• Speed up lookup of 768D neighbors from file

Visualization

Visualize Data and Landmarks.

Later we want to have an web based UI for user interaction. Therefore it would be a good idea to do the visualization also web based and extend it to the full featured UI in the next project phase.

Really use landmarks

Currently the updated landmarks are not used by lmds.

Improve controls and visuals

• Only accept mouse input when over canvas
• Display all points by default
• Landmarks disappear when clicked!

Add front end support for multi class datasets

Check lmds and implement PCA

Model training

• IMBD and emotion euclidean
• IMBD and emotion cosine
• Train on random data?

Human Readable Ids

Use human readable ids for the lmds instances.

Fixed Landmark amount range

We want the amount of landmarks selectable by the user to be in a fixed range.

• Put constant somewhere in backend
• Replace textbox with slider in frontend