1. Install Anaconda following these instructions
2. Clone the repository git clone https://github.com/sdamrich/test_reprod.git
3. Create and activate the environment

   cd test_reprod
   conda env create -f environment.yml
   conda activate ModShift
   

4. Download the data automatically python dataloader.py or manually
   • For CREMI download the file from here and place it in ModShift/data/CREMI/data/CREMI.h5.
   • For ISBI download the file from here and place it in ModShift/data/ISBI/data/ISBI.h5.
   • For ISBI's PCA download the file from here and place it in ModShift/data/ISBI/data/ISBI_embeddings_PCA_8.npy.
5. Reproduce the pixel embedding-based clustering re- sults by running

   python main.py --data config_DATASET.yml --clustering config_CLUSTERING_METHOD.yml
   

   For instance, to reproduce non-blurring Mod Shift’s results on the subvolume of CREMI A run

   python main.py --data config_CREMI_A.yml --clustering config_mod_shift.yml
   

   Convergence points, labels and scores will be saved in data/DATASET/. The best parameters and scores will be printed. In case of problems with PyKeOps, confer here. Note that the pure PyTorch implementation is very slow on these datasets, so that downsampling them in the config_DATASET.yml files is advisable if PyKeOps is not available.
6. To reproduce the experiment of section D.3 do

   cd uniform_comparison
   python mean_mod_uniform.py
   

   The key statistics will be printed, the number of clusters for Mean Shift and Mod Shift saved and the histogram of these numbers of clusters plotted.
7. To reproduces an experiment on the toy dataset, do

   cd 3clusters
   python trajectories_toy.py
   

   The trajectories will be saved in ModShift/3clusters/data/3clusters/results and plotted.

This repository is licensed under the MIT License other than the directory cremi, which is strongly based on and licensed as this repository.