Intrinsic Lorentz Neural Network

Official PyTorch code for Intrinsic Lorentz Neural Network (ILNN).

ILNN is a fully intrinsic Lorentz hyperbolic neural network built around:

• PLFC: a point-to-hyperplane Lorentz fully connected layer.
• GyroLBN: Lorentz batch normalization with gyro-centering and gyro-scaling.
• Log-radius patch concatenation for Lorentz convolution.
• Gyro-additive bias, Lorentz activation, and Lorentz dropout.

This cleaned repository contains the image-classification and genomics experiments used by the paper.

Repository Layout

lib/                         Shared Lorentz, GyroBN, Geoopt, and model modules
experiments/vision/          CIFAR-10/100 ResNet-18 experiments
experiments/genomics/        TEB and GUE sequence-classification experiments
scripts/                     Multi-seed reproduction scripts
tools/hyperbolicity/         Dataset hyperbolicity utilities
data/                        Local datasets, not committed
outputs/                     Checkpoints and logs, not committed

Installation

The original experiments used Python 3.8, PyTorch 1.13, CUDA 11.7, and A100 GPUs.

conda env create -f environment.yml
conda activate ilnn

Alternatively:

pip install -r requirements.txt

Install a PyTorch build matching your CUDA version if the pinned conda environment is not suitable for your machine.

Data

CIFAR-10/100 are downloaded automatically by torchvision.

Genomics datasets are expected under data/; see data/README.md for the exact layout. The repository does not commit GUE/TEB files or generated checkpoints.

Reproduce Vision Experiments

Single run:

python experiments/vision/train.py \
  -c experiments/vision/config/ILNN-CIFAR10.txt \
  --device cuda:0 \
  --seed 1

python experiments/vision/train.py \
  -c experiments/vision/config/ILNN-CIFAR100.txt \
  --device cuda:0 \
  --seed 1

Five-seed script:

DEVICE=cuda:0 SEEDS="1 2 3 4 5" bash scripts/run_vision_cifar.sh

Reproduce Genomics Experiments

GUE promoter and Covid tasks:

DEVICE=cuda:0 bash scripts/run_genomics_gue.sh

TEB pseudogene tasks:

DEVICE=cuda:0 bash scripts/run_genomics_teb.sh

Run one GUE split manually:

python experiments/genomics/train.py \
  -c experiments/genomics/configs/ILNN_GUE_promoter.txt \
  --device cuda:0 \
  --dataset_name prom_core_tata \
  --seed 1

Run one TEB split manually:

python experiments/genomics/train.py \
  -c experiments/genomics/configs/ILNN_TEB_pseudogenes.txt \
  --device cuda:0 \
  --dataset_name processed_pseudogenes \
  --seed 1

Main Components

• PLFC: lib/lorentz/layers/LFC.py
• GyroLBN: lib/GyroBN/GyroBNH.py and lib/lorentz/layers/LBnorm.py

Notes

The code vendors a local Geoopt snapshot because the experiments rely on custom Lorentz operations and optimizers. RiemannianLineSearch is optional in this release because newer SciPy versions removed a function required by Geoopt's old line-search implementation; ILNN uses RiemannianSGD and RiemannianAdam.

Citation

@article{shi2026intrinsic,
  title={Intrinsic Lorentz Neural Network},
  author={Shi, Xianglong and Chen, Ziheng and Jiang, Yunhan and Sebe, Nicu},
  journal={arXiv preprint arXiv:2602.23981},
  year={2026}
}