flame-ai-challenge

Description

This repository contains our submission to the ML Challenge organized by the Stanford FLAME AI 2023 Workshop.
The challenge's objective is to perform super-resolution on low-resolution 2D flowfield images to reconstruct high-resolution versions of these flowfields.

Requirements

This code is written in python, and needs pytorch and torchvision to run. It is optimized for GPU usage.

Installation

• Clone this repository
• Copy the data from the challenge to the data folder. The folder structure should be as follows:

data
├──dataset
    ├── train.csv
    ├── test.csv
    ├── val.csv
    ├── flowfields
    │   ├── ...

• Download the models checkpoints into the checkpoints folder. The folder structure should be as follows:

checkpoints
├── model_1.pth
├── ...

Usage

• Launch jupyter lab
• Open loading_model.ipynb

Results

The model achieves a L2 loss of :

• 0.00553 on the private test set
• 0.00462 on the validation set

The best individual model vxyhe45d achieves a L2 loss of:

• 0.00600 on the private test set.
• 0.00478 on the validation set

Team Members

• Thomas X Wang, ISIR, Sorbonne Université
• Louis Serrano, ISIR, Sorbonne Université

Acknowledgements

This code is based on Super-Resolution Neural Operator