CODAC

Instructions for installing the software and running reproducing experiments for CODAC

Installing

# set up a conda enviroment
conda create -n codac python=3.13
conda activate codac
# install numpy if not installed yet (required by ClustPy)
pip install numpy
# clone the repository
git clone https://github.com/edahelsinki/codac.git
cd codac
# clone ClustPy
git clone https://github.com/collinleiber/ClustPy.git
cd ClustPy
# install the development version accordingly to the instructions in https://github.com/collinleiber/ClustPy
python setup.py install --prefix ~/.local
python setup.py build_ext --inplace

cd ..
# install additional requirements
pip install scikit-learn-extra, wandb, ucimlrepo, active-semi-supervised-clustering, nltk

# install project as a package
pip install -e .

# DONE

Running experiments in HPC

Utilizing slurm for managing jobs, one can use existing scripts. The directory experiments contain the Python code for running experiment loading datasets and creating figures.

Loading datasets

By running the following commands the data are dowloaded and cached automatically.

cd experiments
python data.py

Running experiments

The following commands reproduce the experiments, note that the scripts are designed to be run on high performance cluster. The results are created in experiments/results.

# in experiments
cd scripts

# Figure 3
bash start_experiments_deep.sh
bash start_experiments_shallow.sh
bash start_experiments_a3s.sh
bash start_experiments_a3s_deep.sh
bash start_experiments_ffqs.sh
bash start_experiments_ffqs_deep.sh

# Figure 4 and Table 1
bash start_experiments_guess_k.sh

# Figure 5
bash start_experiments_train_test_codac.sh

# Figure 6
bash start_experiments_abl_loss.sh

# Figure 7
bash start_experiments_abl_sampling.sh

# Figure 8
bash start_experiments_abl_deep_clusterers.sh

Creating figures

The Figures in the article are generated with create_clust_figs.py, the script exptects the first parameter(s) to be directories for results, -o specifies the output directory, -t <str> is a style parameter for spefifying correct legend names, and -l <str> defines to which figure to add a legend to (default is all). Below is an example command for creating Figure 2.

# in experiments
python create_clust_figs.py results/aclust_results/aclust_compare_deep results/aclust_results/aclust_compare_shallow results/aclust_results/aclust_compare_a3s_shallow results/aclust_results/aclust_compare_a3s_deep results/aclust_results/aclust_compare_ffqs_shallow results/aclust_results/aclust_compare_ffqs_deep -o figures/aclust_new -t compare -l bloodmnist