t-SNE Visualization

This repository is an easy-to-run t-SNE visualization tool for your dataset of choice. It currently supports 2D and 3D plots as well as an optional original image overlay on top of the 2D points.

Installation

Ubuntu Installation

First clone this repository, then install the TkInter package by running:

sudo apt-get install python3-tk

Optionally create a virtualenv for this project:

cd tsne-vis
virtualenv -p python3
source venv/bin/activate

Then install the python3 dependecies:

cd tsne-vis
pip install -r requirements.txt

Usage

Example Command

python main.py --num_samples=5000 --num_dimensions=2 --compute_embeddings=False --with_images=False

This will plot a 2D t-SNE plot with no image overlay. Note that the example code uses the Fashion-MNIST dataset which you can download by running:

chmod +x download_data.sh
./download_data.sh

You'll only need to modify the load_data method if you're planning on using your own dataset. Make sure it returns a set of numpy arrays: for example, if embedding grasycale images, you'll want to return an array of images and their associated labels as follows

X: (100, 32, 32)
y: (100,)

To see all possible command line options, run

python main.py --help

which will print:

usage: main.py [-h] [--num_samples NUM_SAMPLES]
               [--num_dimensions NUM_DIMENSIONS] [--shuffle SHUFFLE]
               [--compute_embeddings COMPUTE_EMBEDDINGS]
               [--with_images WITH_IMAGES] [--random_seed RANDOM_SEED]
               [--data_dir DATA_DIR] [--plot_dir PLOT_DIR]

t-SNE Visualizer

optional arguments:
  -h, --help            show this help message and exit

Setup:
  --num_samples NUM_SAMPLES
                        # of samples to compute embeddings on. Becomes slow if
                        very high.
  --num_dimensions NUM_DIMENSIONS
                        # of tsne dimensions. Can be 2 or 3.
  --shuffle SHUFFLE     Whether to shuffle the data before embedding.
  --compute_embeddings COMPUTE_EMBEDDINGS
                        Whether to compute embeddings. Do this once per sample
                        size.
  --with_images WITH_IMAGES
                        Whether to overlay images on data points. Only works
                        with 2D plots.
  --random_seed RANDOM_SEED
                        Seed to ensure reproducibility

Path Params:
  --data_dir DATA_DIR   Directory where data is stored
  --plot_dir PLOT_DIR   Directory where plots are saved

Image Overlay

The overlay option only works for 2D plots and relies on matplotlib's AnnotationBox method. Here's an example of what it outputs: