Skip to content

Latest commit

 

History

History
116 lines (89 loc) · 6.3 KB

README.md

File metadata and controls

116 lines (89 loc) · 6.3 KB
Raw

Deep Momentum-Conserving Fluids (DMCF)

TensorFlow badge

This repository contains the code for our NeurIPS paper Guaranteed Conservation of Momentum for Learning Particle-based Fluid Dynamics. Our algorithm makes it possible to learn highly accurate, efficient and momentum conserving fluid simulations based on particles. With the code published here, evaluations from the paper can be reconstructed, and new models can be trained.

canyon video

Please cite our paper if you find this code useful:

@inproceedings{Prantl2022Conserving,
        title     = {Guaranteed Conservation of Momentum for Learning Particle-based Fluid Dynamics},
        author    = {Lukas Prantl and Benjamin Ummenhofer and Vladlen Koltun and Nils Thuerey},
        booktitle = {Conference on Neural Information Processing Systems},
        year      = {2022},
}

Dependencies and Setup

Used environment: python3.7 with CUDA 11.3 and CUDNN 8.0.

  • Install libcap-dev: sudo apt install libcap-dev
  • Install cmake: sudo apt install cmake
  • Update pip: pip install --upgrade pip
  • Install requirements: pip install -r requirements.txt
  • Tensorpack DataFlow pip install --upgrade git+https://github.com/tensorpack/dataflow.git

Optional:

  • Build FPS/EMD module cd utils; make; cd ..
  • Install skia for visualization: python -m pip install skia-python

Datasets

Pretrained Models:

The pretrained models are in the checkpoints subfolder. Run a pretrained mode by setting the path to the checkpoint with the ckpt_path argument. For example:

python run_pipeline.py --cfg_file configs/WBC-SPH.yml \
                       --dataset_path PATH/TO/DATASET \
                       --ckpt_path checkpoints/WBC-SPH/ckpt \
                       --split test

Training

Simple 1D test run (data will be generated):

python run_pipeline.py --cfg_file configs/column/hrnet.yml \
                       --split train

Run with 2D pipeline:

python run_pipeline.py --cfg_file configs/WBC-SPH.yml \
                       --dataset_path PATH/TO/DATASET \
                       --split train

Test

python run_pipeline.py --cfg_file configs/WBC-SPH.yml \
                       --dataset_path PATH/TO/DATASET \
                       --split test \
                       --pipeline.data_generator.test.time_end 800 \
                       --pipeline.data_generator.valid.time_end 800 \
                       --pipeline.data_generator.valid.random_start 0 \
                       --pipeline.test_compute_metric true

Note: The argument pipeline.data_generator.test.time_end, pipeline.data_generator.valid.time_end, pipeline.data_generator.valid.random_start, and pipeline.test_compute_metric are examples how to overwrite corresponding entries in the config file.

The ...time_end parameter account for the number of frames used for inference and evaluation. We used a value of 3200 for the WBC-SPH data set, 600 for WaterRamps, and 200 for Liquid3d. The generated test files are stored in the pipeline.output_dir folder, specified in the config file. The output files have a hdf5 format and can be rendered with the utils/draw_sim2d.py script.

Rendering of a small sample sequence:

python utils/draw_sim2d.py PATH/TO/HDF5_FILE OUTPUT/PATH

Rendering of individual frames:

python utils/draw_sim2d.py PATH/TO/HDF5_FILE OUTPUT/PATH \
                           --out_pattern OUTPUT/FRAMES/{frame:04d}.png \
                           --num_frames 800

Validation

python run_pipeline.py --cfg_file configs/WBC-SPH.yml \
                       --dataset_path PATH/TO/DATASET \
                       --split valid

FAQ and Clarifications

How was the training data generated with SPlisHSPlasH?

The data was generated based on the code of previous work (https://github.com/isl-org/DeepLagrangianFluids). The required code is in the 'datasets' folder. 'create_data.sh' is the shell script to run the data generation.

Why not replace all the CConv with the ASCC?

The reason for this is that the antisymmetry is a strong restriction. Replacing all CConvs with ASCCs severely limits the learning capabilities of the neural network. The good thing is, it is sufficient to place the ASCC only at the end to make the network antisymmetric. This reduces the constraints in the network to a minimum and only transforms the generated values in the final layer so that the resulting values comply with the antisymmetry constraint. Even with this, it was quite difficult to tune the network to achieve the current state. The paper has a short paragraph with an example (Standing Liquid) in the Result section, which briefly discusses this.

What is the Maximum Density in the evaluation?

The Maximum Density value is the relative error between the maximum density of the fluid and the maximum density of the ground truth, where a value closer to 0 is preferable (Equation 15 in the paper). We use this as a heuristic for the compressibility of the fluid, which can lead to high pressure and thus instability in the simulation. Apart from that, please note that the values in Table 2 in the paper are not the raw error values but relative accuracy values as described in Figure 6. I.e. a value of 1 corresponds to the error of our final method, while small values represent a lower relative accuracy and thus larger error. A value of 0.5, for example, would mean half the accuracy and double the error. We chose this format to relate the error to the final method, which we felt was important in an ablation study, and to normalise the error evaluation for better visualisation in the graph.

Licenses

Code and scripts are under the MIT license.

Data files are under the CDLA-Permissive-2.0 license.