SpecCheck

This repository contains the code and data which can reproduce results reported in our article 'Does the data meet your expectations? Explaining diversity in a dataset of images' submitted to BNAIC 2020. Briefly, the idea is to explain sample representation in a dataset of greyscale images of circles and squares in terms of intuitive aspects such as size, position and pixel brightness.

Demo

1. Make sure to install dependencies listed in requirements.txt
2. We use an Nvidia GTX 1080Ti GPU, but it may be possible to replicate results without using a GPU
3. To run the full experiment from scratch, run

python run_test.py --test_config=config/test_draw_128.py

This will prompt the user to place source data files from Quick,Draw! with Google (https://quickdraw.withgoogle.com/data)

4. To use pre-trained models and previously collected data, download the draw_128 directory from here and place it under SpecCheck/_tests. Additionally, untar draw_128/qd_shapes.tar and draw_28/sd_shapes.tar to draw_128/qd_shapes and draw_128/sd_shapes repsectively. After this, you can similarly execute run_test.py like the previous step.