ExPoConv: Reproducible package for "Efficient Poincaré Convolutions through Patch-Statistics Factorization"
Authors/Creators
Description
ExPoConv
Reproducible package for “Efficient Poincaré Convolutions through Patch-Statistics Factorization”
Overview
ExPoConv is a research software package accompanying the manuscript Efficient Poincaré Convolutions through Patch-Statistics Factorization.
The project investigates computationally efficient execution of Poincaré convolutional operations in hyperbolic neural networks. The proposed approach exploits patch-statistics factorization to avoid explicit construction of ball-valued receptive-field patches during convolutional score evaluation.
The software implements and evaluates alternative computational formulations, with emphasis on numerical agreement, execution latency, GPU memory requirements, and end-to-end neural network training.
The archive provides the implementation, registered experimental protocol, configuration files, dataset split manifests, validation tests, experiment runners, compact numerical results, and reproducibility documentation associated with the study.
Key features
1. Patch-statistics factorization
The primary method evaluates Poincaré convolution scores using directional products obtained through native convolution and scalar patch-energy reductions.
The implementation supports direct comparison with the reference convolution formulation while preserving the shared geometric parameters and computational conventions.
2. Controlled implementation comparisons
The experimental framework distinguishes several execution paths, including:
-
R: Reference implementation using explicit tangent-space patches mapped to the Poincaré ball.
-
F: Factorized convolution score evaluation using patch statistics.
-
T: Factorized evaluation with direct tangent-space output calculation.
Additional implementation controls are included for selected benchmarking and diagnostic experiments.
3. Numerical validation
The software provides numerical agreement tests for convolutional outputs and gradients in FP32 and FP64 precision.
The experimental framework also includes elementwise error analysis, numerical sensitivity checks, and comparisons with higher-precision reference calculations.
4. Computational benchmarking
The package supports evaluation of operator-level execution time, complete-model latency, training update performance, and GPU memory allocation.
Experiments include convolutional workloads, complete neural network architectures, scaling analyses, and comparisons with the official HypLL implementation.
5. End-to-end neural network experiments
The software includes experimental configurations for:
-
U-Net image segmentation using ISIC 2016.
-
ResNet-20 image classification using CIFAR-100.
The experiments examine predictive performance alongside computational efficiency under documented training protocols.
6. Reproducibility and provenance
The repository includes:
-
Registered experimental protocols and configuration files.
-
Fixed random seeds and dataset split manifests.
-
Experiment runners and automated validation tests.
-
Configuration hashes and code revision identifiers.
-
Compact numerical results supporting reported findings.
-
Mapping between manuscript claims and supporting result files.
-
Documentation of software environments and reproduction procedures.
These resources support inspection of the reported measurements and reconstruction of the associated experimental workflows.
Scope and interpretation
ExPoConv is intended to evaluate the computational properties of alternative implementations of Poincaré convolution under controlled experimental conditions.
The reported timing and memory measurements depend on the documented hardware, software environments, workload configurations, and measurement boundaries. They should not be interpreted as universal performance guarantees across all hyperbolic architectures or computing platforms.
The software is intended for research use. Models trained on biomedical imaging data are not validated for clinical decision-making.
Software repository and archive
The development repository is publicly available on GitHub:
https://github.com/lowoncuties/ExPoConv
This Zenodo record provides a persistent, citable snapshot of the ExPoConv research software and associated reproducibility materials.
Archived software version: v0.2.0
The repository contains installation instructions, experimental configurations, reproducibility procedures, numerical tests, and reference results.
Dataset availability
The accompanying study uses two publicly available datasets:
-
ISIC 2016: Skin lesion image segmentation experiments.
-
CIFAR-100: Image classification experiments.
The original datasets are not redistributed in this software archive. The repository provides procedures for obtaining and verifying the required datasets.
Users must comply with the licensing and citation conditions established by the original dataset providers.
Archive contents and limitations
The archive includes source code, experiment configurations, dataset split manifests, compact result tables, numerical validation utilities, and reproducibility documentation.
The archive does not include the original datasets, full model checkpoints, raw per-call timing records, or complete profiler traces.
The included results and provenance information support examination of the reported findings without redistributing these larger experimental artefacts.
Associated publication
Jochymek, L., Vašinková, M., & Gajdoš, P. (2026). Efficient Poincaré Convolutions through Patch-Statistics Factorization. Manuscript under consideration for publication.
The publication DOI will be added when available.
Users of ExPoConv are encouraged to cite both this archived software release and the associated research article.
License
ExPoConv is distributed under the MIT License. See the LICENSE file included in the archive.
Third-party datasets and upstream software remain subject to their respective licensing conditions.
Files
ExPoConv-master.zip
Additional details
Related works
- Is source of
- Software: https://github.com/lowoncuties/ExPoConv (URL)
Funding
- Ministry of Education Youth and Sports
- Research Platform for Digital Transformation and Society 5.0 CZ.02.01.01/00/23_021/0012599
- European Commission
- Center for Artificial Intelligence and Quantum Computing in System Brain Research (CLARA) 101136607-02
Software
- Repository URL
- https://github.com/lowoncuties/ExPoConv
- Programming language
- Python
- Development Status
- Active