SWAP-INR: Implicit Neural Representation with Switchable Attention Paths for Time-varying Multivariate Scientific Data Compression

Accepted at WACV 2027
Paper: Coming soon

Overview

Figure 1. Overview of the proposed SWAP-INR. Given a spatiotemporal coordinate $(x, y, z, t)$, the network first applies positional encoding. The encoded features are processed by MLP layers shared across variables, equipped with A²-SIREN, which employs a learnable parameter $\boldsymbol{\alpha}$ for adaptive frequency modulation (bottom left). The representations are then split into two branches: a non-attention path (red) and an attention path (blue). The attention branch models cross-variable interactions using a cross-variable attention layer (purple). Finally, the variable-driven path switching block selects the appropriate representation for each variable to reconstruct high-fidelity multivariate outputs.

Main Results

Table 1. Quantitative comparison of compression results with previous methods at matched CR. The * denotes that the reported performance is averaged across independent time steps. Baselines are stored at 32-bit precision, while SWAP-INR uses 16-bit.

Reproduction with Downsampled Datasets

Due to supplementary-material size limitations, the provided model weights and example datasets for ROCOM and Ionization use downsampled versions of the original datasets. The original full-resolution datasets used for the paper experiments are substantially larger.

Dataset Sizes

Dataset	Original dataset size (paper)	Downsampled dataset size (repository)
ROCOM	1,943,807 KB	80,995 KB
Ionization	450,469 KB	11,270 KB

Results with Provided Models and Downsampled Data

Dataset	Evaluation	PSNR ↑	NRMSE ↓	$R^2$ ↑
ROCOM	Paper (original data)	47.41	0.0073	0.9782
ROCOM	Repository (downsampled data)	43.06	0.0108	0.9421
Ionization	Paper (original data)	50.25	0.0032	0.9905
Ionization	Repository (downsampled data)	53.66	0.0021	0.9810

Note: The supplied checkpoints were obtained using downsampled ROCOM and Ionization data. These results are not directly comparable with full-dataset results because the evaluation data differ. All results reported in this reproduction table were obtained using FP16.

Usage

Environment Setup

Before running training or inference, create and activate a Conda environment and install the required dependencies:

conda create -n SWAP-INR python=3.9 -y
conda activate SWAP-INR
pip install -r requirements.txt

Run the following training and inference commands from the code/ directory.

cd code/

Training

ROCOM

python train.py --dataset ROCOM_down_down --num_var 8 --Shared_hidden 64 --Branch_hidden 32 --name ROCOM_down_down_test --device cuda

Ionization

python train.py --dataset ionization_down_down --num_var 4 --Shared_hidden 64 --Branch_hidden 24 --name ionization_down_down_test --device cuda

The --name argument is required. Checkpoints and pruned models are saved under:

Exp/<dataset>/<name>/

Additional training options:

python train.py \
  --dataset ROCOM_down_down \
  --name experiment_name \
  --device cuda \
  --batch_size 100000 \
  --base_num_epochs 2000 \
  --on_num_epochs 1000

Use --device cpu if CUDA is unavailable.

Inference

Use the same dataset, experiment name, and model architecture arguments as in training.

ROCOM

python inference.py --dataset ROCOM_down_down --num_var 8 --Shared_hidden 64 --Branch_hidden 32 --name ROCOM_down_down_test --device cuda

FP16 inference:

python inference.py --dataset ROCOM_down_down --num_var 8 --Shared_hidden 64 --Branch_hidden 32 --name ROCOM_down_down_test --use_half --device cuda

Ionization

python inference.py --dataset ionization_down_down --num_var 4 --Shared_hidden 64 --Branch_hidden 24 --name ionization_down_down_test --device cuda

FP16 inference:

python inference.py --dataset ionization_down_down --num_var 4 --Shared_hidden 64 --Branch_hidden 24 --name ionization_down_down_test --use_half --device cuda

Citation

Citation information will be added when the official WACV 2027 paper metadata is available.

---

Acknowledgements

This work was conducted at the High-performance Visual Computing Lab (HVCL), Korea University.

We gratefully acknowledge Prof. Chaoli Wang at the University of Notre Dame for providing the Combustion and Ionization datasets, the Korea Atomic Energy Research Institute (KAERI) for providing the ROCOM, CUPID, and Synthetic datasets, and the TFC & BME Lab (led by Prof. Haecheon Choi) at Seoul National University for providing the SMR Simulation dataset. We also extend our sincere gratitude to the Electronics and Telecommunications Research Institute (ETRI) for their valuable collaboration to this research.

---

_{© 2026 Hyunsoo Son, Jeonghyun Noh, Chaoli Wang, and Won-Ki Jeong
⭐ Star this repo to stay updated on the code release!}