Skip to content

Latest commit

 

History

History
92 lines (68 loc) · 3.1 KB

SETUP_DATA.md

File metadata and controls

92 lines (68 loc) · 3.1 KB

Shape-Net Car Dataset

Download

mkdir shapenet_car
cd shapenet_car
wget http://www.nobuyuki-umetani.com/publication/mlcfd_data.zip
unzip mlcfd_data.zip
rm mlcfd_data.zip
rm -rf __MACOSX
cd mlcfd_data
tar -xvzf param0.tar.gz
tar -xvzf param1.tar.gz
tar -xvzf param2.tar.gz
tar -xvzf param3.tar.gz
tar -xvzf param4.tar.gz
tar -xvzf param5.tar.gz
tar -xvzf param6.tar.gz
tar -xvzf param7.tar.gz
tar -xvzf param8.tar.gz
rm param0.tar.gz
rm param1.tar.gz
rm param2.tar.gz
rm param3.tar.gz
rm param4.tar.gz
rm param5.tar.gz
rm param6.tar.gz
rm param7.tar.gz
rm param8.tar.gz
# remove folders without quadpress_smpl.vtk
rm -rf ./training_data/param2/854bb96a96a4d1b338acbabdc1252e2f
rm -rf ./training_data/param2/85bb9748c3836e566f81b21e2305c824
rm -rf ./training_data/param5/9ec13da6190ab1a3dd141480e2c154d3
rm -rf ./training_data/param8/c5079a5b8d59220bc3fb0d224baae2a

Preprocess

Preprocess the data by using the folder of the downloaded dataset as SRC_FOLDER. python data/shapenetcar/preprocess.py --src <SRC_FOLDER> --dst <DST_FOLDER>

Transient Flow Dataset

Generate

We provide resources to generate transient flow simulations that we used in Section 4.2 of our paper in the data/transientflow folder.

Software needed:

  • Working OpenFOAM (tested with OpenFOAM-v2306)
  • Working MPI (tested with OpenMPI 5.0.0)

Python prerequisites:

  • fluidfoam
  • PyFoam
  • gmsh
  • meshio
  • shapely
  • scikit-learn

Usage:

python generateCase.py n_objects n_cases n_cores empty_case_dir target_dataset_dir working_dir

Parameters:

  • n_objects: number of objects in the flow-path
  • n_cases: number of cases to be generated
  • n_cores: number of cores used by MPI
  • empty_case_dir: Directory of an empty OpenFOAM case. An empty case is provided in EmptyCase.zip
  • target_dataset_dir: Directory of the generated dataset
  • working_dir: A working directory for OpenFOAM (used during dataset generation)

Preprocess

After generating the OpenFoam cases, we preprocess the data into torch files (".th" files) for dataloading and convert it into fp16 precision. To do this we use this script.

python data/transientflow/cfddataset_from_openfoam.py --src <OPENFOAM_OUTPUT_PATH> --dst <POSTPROCESSED_PATH> --num_workers 50

To calculate statistics for normalizing the data execute the cfddataset_norm.py script (this will calculate the statistics for normalization as described in Appendix B.4):

python data/transientflow/cfddataset_norm.py --root <POSTPROCESSED_PATH> --q 0.25 --exclude_last <NUM_VALIDATION_CASES + NUM_TEST_CASES> --num_workers 50

Lagrangian

The datasets will be automatically downloaded into the corresponding directory defined in the static_config.yaml (by default <PATH>/data/lagrangian_dataset).

Make sure you have a existing directory for the lagrangian_dataset key in your static_config.yaml (e.g., like in static_config_template.yaml). You need to create this director if it does not exist already mkdir <PATH>/data/lagrangian_dataset.