To build docker image with the environment simply run:
./build_docker.shOnce created, one can run a container in interactive mode typing:
./run_docker.sh $gpu_idwhere ID of a GPU -- $gpu_id has to be specified with regards to docker run e.g. 0, 1, 2, etc.
Once navigating in the container type:
cd src/Some of the datasets were compressed due to github's storage limitations. To unpack them run:
unzip data/ccfd.zip
tar -xf data/census-income-full-mixed-binarized.tar.xz -C data/Navigate to experiments/ directory:
cd experiments/To run experiment for one of the tabular datasets run:
python ./tabular.py -d $dataset -o $output_path -m $model -n $n_times --gamma $gammawhere:
-
$datasetis the name of desired dataset. One of:ccfd|kddps|kddh|celeba|census|campaign|thyroid. -
$output_pathis the path to csv with resulting AUCs. By defaultresult.txt. -
$modelis the model name. One of:e2econva3|a3. By defaulte2econva3. -
$n_timesis the number if trial runs. By default 1. -
$gammais the$\gamma$ parameter. By default 1.
Similarily to tabular datasets run:
python ./image.py -d $dataset -c $normal_cls -o $output_path -n $n_times --gamma $gammaAll of the arguments are the same as above except for:
-
$datasetis one of:cifar10|mnist|fmnist. -
$normal_clsis the selected normal class index. By default 0.