neuflow 0.1.1

The pytorch implementation of the paper 'EPINF: Efficient Physics Informed Fluid Flow Reconstruction With Spatial and Temporal Priors'

EPINF-NeuFlow

The pytorch implementation of the paper "EPINF: Efficient Physics Informed Fluid Flow Reconstruction With Spatial and Temporal Priors"

Environment Setup

System Requirements

• System
  • Windows 11 + MSVC 2022 (Fully Tested)
  • Ubuntu 24.04.2 (Fully Tested)
• Python: 3.13
• PyTorch: 2.7.1 + CUDA 12.8

Pip Packages

python -m pip install --upgrade pip setuptools wheel
python -m pip install torch torchvision --index-url https://download.pytorch.org/whl/cu128
python -m pip install lightning lightning[extra] dearpygui tyro matplotlib av huggingface_hub opencv-python phiflow

Fully Fused-NNs - tiny-cuda-nn

set TCNN_CUDA_ARCHITECTURES=86
python -m pip install git+https://github.com/NVlabs/tiny-cuda-nn/#subdirectory=bindings/torch

For Windows deployment, open X64 Native Tools Command Prompt for VS 2022 then activate your python venv if any, and set TCNN_CUDA_ARCHITECTURES to your GPU architecture.

GPU	CUDA arch
H100	90
40X0	89
30X0	86
A100	80
20X0	75
TITAN V / V100	70
10X0 / TITAN Xp	61
9X0	52
K80	37

Native CUDA Extensions

freqencoder, gridencoder, raymarching, shencoder

python -m pip install ./cuda_extensions/freqencoder ./cuda_extensions/gridencoder ./cuda_extensions/raymarching ./cuda_extensions/shencoder

Optional Packages

triton-windows - Activate torch.compile for Windows

python -m pip install -U "triton-windows<3.3"

Datasets

We provide two datasets for training and testing the model:

• the NeRF-Synthetic is available at Hugging Face - nerf_synthetic.
• The PI-Neuflow dataset is available at Hugging Face - PI-NeuFlow.

ALL datasets are downloaded AUTOMATICALLY when running the training script.