CuMPerLay: Learning Cubical Multiparameter Persistence Vectorizations

Official implementation of "CuMPerLay: Learning Cubical Multiparameter Persistence Vectorizations".

We present CuMPerLay, a novel differentiable vectorization layer that enables the integration of Cubical Multiparameter Persistence (CMP) into deep learning pipelines. Please refer to our paper and project page for more details on the project and the methodology followed.

Installation

You can install all requirements using environment.yml and the cmp repository containing our C++/CUDA implementations as follows:

conda env create -f environment.yml
conda activate cumperlay
# Install the cmp package
pip install --no-build-isolation "git+https://github.com/circle-group/cmp.git#egg=cmp"

Note that we tested the method with gcc-12/g++-12 and CUDA 12.2 (may work with other compatible PyTorch/CUDA and gcc/g++ versions). Please see cmp repository for installation instruction on our cmp package. If you need to install the library without installing any system-wide packages, you can utilize the conda environment to install the required compilers before the pip install step:

conda activate cumperlay
# Install gcc/g++ 12.2
conda install -c conda-forge gcc==12.2 gxx==12.2
export CC="gcc" CXX="g++"
# Install the cmp package
pip install --no-build-isolation "git+https://github.com/circle-group/cmp.git#egg=cmp"

Please see environment.yml for all package requirements. torch >= 2.0 is required and we have tested the method with Pytorch 2.3.0 and CUDA 12.2.

Usage

To train the model (as an example to run the TopoSwin-MP model on the ISIC dataset with the default parameters, updating data.dataroot config parameter in config/isic-swin-silv2-inpd-multreg-enp.yaml to the correct dataroot), you can use the following command:

OMP_NUM_THREADS=1 python -u train.py --config config/isic-swin-silv2-inpd-multreg-enp.yaml --train --gpu 0

Please refer to the config folder for run configurations and train.py for running experiments. You can refer to the same training file for evaluating the checkpoints of a trained model. For multiclass/multilabel related configuration parameters, you can check the Pytorch Lightning Trainer.

For dataset configurations, please see lmp/data.

Further instructions on usage and demos will be available soon.

Relevant Code

Please see lmp/nn/networks/lmp_swin_v2.py for the combined SwinV2 model with our CumPerLay layer, lmp/nn/persistence/cubical_torch.py for the persistence implementation (that uses the cubical_ph package from cuda-cumper folder), lmp/nn/representations/silhouette_v2.py for our vectorization.

Data

The dataset used in this project is not included in the repository. You can download the datasets from ISIC 2018 (task3), CBIS-DDSM (Cropped Images) and Glaucoma. For the segmentation datasets, you need to download them using TorchVision beforehand, further instructions on this will be available, but you can refer to lmp/data/pascal_voc_sbd.py.

Citation

If you use this codebase/library, or otherwise found our work valuable, please cite:

@InProceedings{Korkmaz_2025_CuMPerLay,
    author    = {Korkmaz, Caner and Nuwagira, Brighton and Coskunuzer, Baris and Birdal, Tolga},
    title     = {CuMPerLay: Learning Cubical Multiparameter Persistence Vectorizations},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2025},
    pages     = {27084-27094}
}

Acknowledgements

• https://github.com/threestudio-project/threestudio for parts of the codebase structure