Haunted attributions

1. A generative dataset based on Country211 with added ghosts
2. A basic training loop to classify: is it haunted or not?
3. A suite of attribution evaluations

Installation

Create a conda environment and install packages:

conda create -n haunted pytorch torchvision -c pytorch
conda activate haunted
pip install -r requirements.txt

And make a directory for the data and models:

mkdir data
mkdir models

The data

Three ghosts are sporadically added to the data in order to created haunted images. Ghost images are courtesy of Stan Kertsjens, subject to CC-BY license.

Additionally we've added fog (in combination with ghosts, or not) in some of the images to make them haunted. The fog is simply a white gradient from bottom to top of the image. Below are some example images.

Tasks

• Make the dataset
• Write the traing loop
• Apply and visualize attributions