- A generative dataset based on
Country211with added ghosts - A basic training loop to classify: is it haunted or not?
- A suite of attribution evaluations
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
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.
- Make the dataset
- Write the traing loop
- Apply and visualize attributions