license: mit
APEBench Scraped
A representative subset of datasets created using the APEBench benchmark suite using version 0.1.0.
⚠️ Note that APEBench is designed to procedurally generate all its training and test data. This allows for advanced features like benchmarking approaches with differentiable physics. Hence, there is no need to download this dataset as it can be easily re-generated using APEBench which can be installed via pip install apebench. See also here for how to scrape datasets.
Download
Download without large files
GIT_LFS_SKIP_SMUDGE=1 git clone [email protected]:datasets/thuerey-group/apebench-scraped
Afterwards, you can inspect the repository and download the files you need. For
example, for 1d_diff_adv:
git lfs install
git lfs pull -I "data/1d_diff_adv*"
Alternatively, you can download the entire repository with large files (~30GB):
git lfs install
git clone [email protected]:datasets/thuerey-group/apebench-scraped
Reproduction
Obtained via:
conda create -n apebench python=3.12 -y
conda activate apebench
pip install -U "jax[cuda12]"
pip install apebench==0.1.0
Alternatively, you can use the provided environment.yml file:
conda env create -f environment.yml
conda activate apebench
And then executed the following script (also found under reproduce.py):
import apebench
from tqdm import tqdm
import os
DATA_PATH = "data"
os.makedirs(DATA_PATH, exist_ok=True)
for config in tqdm(apebench.scraper.CURATION_APEBENCH_V1):
apebench.scraper.scrape_data_and_metadata(DATA_PATH, **config)
⚠️ Small️️er variations of the generated data can occur due to different JAX versions, backends (CPU, GPU, TPU), drivers, etc. This might be especially pronounced for the chaotic problems (like KS or Kolmogorov flow).
- nvidia driver version: 535.183.01
- cuda version: 12.2
- GPU: RTX 3060