TOPIQ: Statistical Error Propagation for Quantity-of-Interest Prediction under Lossy Compression

Artifact for reproducing the experimental results in the paper.

Prerequisites

• Python 3.9+ with packages: pip install numpy matplotlib torch
• Compressors on PATH: SZ3, SPERR, ZFP
• g++ (for C++ throughput benchmark)

Data Preparation

bash scripts/prepare_data.sh data

Downloads NYX, SCALE-LETKF, and Hurricane datasets from SDRBench and extracts the needed fields. CESM-ATM fields are bundled in the Docker image.

Note on CESM-ATM data: The CESM-ATM dataset on SDRBench has been updated since our experiments, and SDRBench does not support version control. To exactly reproduce the numbers in the paper, use the original CESM fields bundled in the Docker Image (or contact the authors). To reproduce the methodology and verify that TOPIQ achieves comparable accuracy, you may download the current CESM-ATM data directly from SDRBench and place the four fields (CLDTOT, CLDHGH, FLUT, FLUTC) in the data directory.

Quick Start

./experiments/run_all.sh data

Runs all experiments sequentially and prints results to stdout. JSON files are saved to results/.

Individual Experiments

Ablation Study (Table III)

python experiments/ablation.py --data-dir data --datasets CESM NYX

Prediction Accuracy (Table IV)

python experiments/prediction_accuracy.py --data-dir data --datasets CESM
python experiments/prediction_accuracy.py --data-dir data --datasets NYX,SCALE,Hurricane

Throughput Benchmark

python benchmark/run_benchmark.py --data-dir data --compressor sz3 --reb 1e-3 --n-queries 10000