Anomaly_analysis

Build and run the environment

To build docker image with the environment simply run:

./build_docker.sh

Once created, one can run a container in interactive mode typing:

./run_docker.sh $gpu_id

where 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/

Data unpacking

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/

Running models

Navigate to experiments/ directory:

cd experiments/

Tabular datasets

To run experiment for one of the tabular datasets run:

python ./tabular.py -d $dataset -o $output_path -m $model -n $n_times --gamma $gamma

where:

• $dataset is the name of desired dataset. One of: ccfd | kddps | kddh | celeba | census | campaign | thyroid.
• $output_path is the path to csv with resulting AUCs. By default result.txt.
• $model is the model name. One of: e2econva3 | a3. By default e2econva3.
• $n_times is the number if trial runs. By default 1.
• $gamma is the $\gamma$ parameter. By default 1.

Image datasets

Similarily to tabular datasets run:

python ./image.py -d $dataset -c $normal_cls -o $output_path -n $n_times --gamma $gamma

All of the arguments are the same as above except for:

• $dataset is one of: cifar10 | mnist | fmnist.
• $normal_cls is the selected normal class index. By default 0.