This is a python script for t-SNE visualization of multiple images in a square grid.

Dependencies:

• tensorflow
• keras

Basic usage:

python tsne_grid.py --dir ./examples/inputs/ --size 4

• --dir: Path to directory containing image collection.
• --size: Number of small images in a row/column in output image.

• --res: Width/height if small images (in pixels). Default is 224.
• --name: Output filename. Default is tsne_grid.jpg
• --path: Output image path. Default is current directory.
• --per: Perplexity for t-SNE algorithm. Default is 50.
• --iter: Number of iterations for t-SNE algorithm. Default is 5000.

VGG16 (without fc layers on top) is used to generate high dimensional feature representations of images. 2D representaions of these features are formed using scikit-learn's t-SNE implementation. These 2D representations are converted into a square grid using Jonker-Volgenant algorithm.

The script was tested with tensorflow (1.4.0) and keras (2.1.1) on Nvidia GeForce GT 740M

• L.J.P. van der Maaten and G.E. Hinton. Visualizing High-Dimensional Data Using t-SNE. Journal of Machine Learning Research 9(Nov):2579-2605, 2008. PDF [Supplemental material] [Talk] [Code]