CAKRes

Read our press release here. For Frequently Asked Questions, refer here.

Introduction

This is a project by CAKRes Innovations that aims to improve workflow of research scientists in natural sciences. It aims to do so by producing real-world accurate upscaled image of low-resolution simulation of physical phenomenon.

Currently, the project to focuses on Super Resolution in the domain of Fluid Dynamics, with domain expansion in the future.

Motivation

Inspiration from this project.

Resources

We are training our models using the following resources:

• 1x RTX4090
• 1x RTX3090
• 64 GB Memory
• 5 TB Storage
• Apache Spark cluster

Benchmark

We might use this benchmark.

Dataset

The dataset are available here. Use your UT Email Adderss to access it.

Creating virtual environment

Create conda or any other virtual environenment

conda create -n cakres python=3.11

Activate Environment

conda activate cakres

Install dependencies

pip install -r requirements.txt

Running experiments

Running FNO Training

python FNO/fno_fluid.py \
    --data_path path/to/your/training/data.h5 \
    --val_data_path path/to/your/validation/data \
    --exp_name experiement_name \
    --scale 4 \
    --epochs 50 \
    --batch_size 4 \
    --lr 0.001

Experiment Output

results will be saved in a directory structure like this:

experiments/
└── <exp_name>/              # Directory named from --exp_name argument
    ├── output.log           # Console output and logs
    ├── metrics.json         # Training/validation metrics
    ├── best_fno_model_s<scale>.pth   # Best model checkpoint 
    ├── fno_metrics_s<scale>.png    # Plot for training and validation loss
    └── fno_crop_viz_s<scale>_p<patch_size>_*.png # Visualization comparison images