🎉 Accepted at NeurIPS 2026
Łukasz Struski, Hanna Blazhko, Piotr Kubaty, Jacek Tabor
Official implementation of "LAPLEX: The FFT of Learnable Laplace Kernels" (NeurIPS 2026). LAPLEX applies a dense matrix with entries exp(−|aᵢ − bⱼ|), defined by learnable anchors on a line, exactly in O(n log n) time and O(n) memory, and differentiates through the anchors.
The experiments are grouped by what LAPLEX is used for. Each directory has a README with the details, the data it needs and the commands for every figure and table.
| directory | content |
|---|---|
| benchmarks/ | the LAPLEX operators: runtime, memory and accuracy against dense kernels, Random Fourier Features and FFT / Toeplitz-FFT (Figs. 1, 3, 5, 6, 8–18, Tab. 3); second-order optimization with the exact curvature (App. K); the released operators in comparison_experiments/ |
| nn_training/ | LAPLEX as a trainable layer: ImageNet-10 projections (CountSketch_LogisticRegression/, Tab. 1), classification heads on frozen Swin-T / MaxViT-T (classification_heads/, Sec. 7), attention (App. J), anchor ablations and the anchor-crossing diagnostic (App. M) |
| log_likelihood/ | Gaussian density models on DIV2K-1024 and ImageNette-320 (Sec. 4.2, App. I) and structured covariance baselines (App. L) |
| environment/ | container definition, conda environment and pinned package versions |
| .env.example | template of the machine-specific settings (container image, dataset paths, Slurm partition) |
environment/ describes the software of the paper's runs in two forms with
the same Python packages (requirements.txt: Python 3.14.6,
PyTorch 2.13.0, NumPy 2.4.4, SciPy 1.18.0, ...); they differ in the CUDA version.
Apptainer / Singularity (the paper's environment). laplex.def
rebuilds Ubuntu 24.04, CUDA 13.2 with nvcc and gcc 13, TeX Live and the packages above
(NVIDIA driver ≥ 580):
cd environment
apptainer build laplex.sif laplex.def # or: singularity build --fakeroot laplex.sif laplex.defBuilding needs root, --fakeroot or unprivileged user namespaces; otherwise build on a
workstation and copy the .sif.
Conda. environment.yml installs the same packages with
PyTorch for CUDA 12.6 (driver ≥ 525), plus nvcc and gcc for the CUDA kernels. For RTX 50xx
GPUs replace cu126 by cu130 and the 12.6 CUDA packages by 13.0.
conda env create -f environment/environment.yml
conda activate laplexThe custom CUDA kernels are compiled on first import (30 s to a few minutes per
implementation) and cached in .cache/. With pdflatex, the benchmark figures are typeset
with the paper's fonts; without it, an approximate PDF and a PNG are written.
Machine-specific settings live in .env (not tracked): copy the template and fill in what
applies to your machine.
cp .env.example .envLAPLEX_SIF: the image built above. Thesif.shwrappers inbenchmarks/,nn_training/andlog_likelihood/run every command in it (bash sif.sh python ...), with the GPU and the dataset directories of.envvisible inside. Without it, they use the current Python environment (e.g. the conda environment).- Dataset paths:
IMAGENET_ROOTandIMAGENET_VAL_CSV(to build ImageNet-10),IMNET,DATA_ROOT/VAL_ROOT,TV_DATA,DIV2K_ROOT,IMAGENETTE_ROOT. Which experiment needs which is given in the directory READMEs. - Slurm: the sbatch files request only resources (GPU, CPUs, memory, time). Set
SBATCH_PARTITION,SBATCH_QOSandSBATCH_ACCOUNTfor your cluster and export them before submitting (set -a; source .env; set +a), or passsbatch -p <partition> ....
cd benchmarks && bash sif.sh python fig03_08_square.py --quick # one benchmark figure, development preset
cd benchmarks && bash run_all.sh # every benchmark figure
cd nn_training && python3 make_imagenet10.py # ImageNet-10 from IMAGENET_ROOT
cd nn_training/CountSketch_LogisticRegression && bash run_experiments.sh # Tab. 1
cd log_likelihood && bash table15_downstream/run.sh # App. L, Tab. 15If you find this work useful, please cite our NeurIPS 2026 paper:
@inproceedings{struski2026laplex,
title = {LAPLEX: The FFT of Learnable Laplace Kernels},
author = {Struski, {\L}ukasz and Blazhko, Hanna and Kubaty, Piotr and Tabor, Jacek},
booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
year = {2026}
}