A machine learning model implemented in Rust 🦀 using the Burn 🔥 deep learning framework for the LEAP - Atmospheric Physics using AI (ClimSim) challenge on Kaggle.
Download the train.csv, test.csv and sample_submission.csv files from
https://www.kaggle.com/competitions/leap-atmospheric-physics-ai-climsim/data.
Important
The train.csv file is very big (181.72GB). I'd personally recommend clicking on the
'Download all' button to get the compressed leap-atmospheric-physics-ai-climsim.zip
version which is only 77.8GB 🙂, and you can partially decompress
the train.csv to get a smaller subset with less rows for quick experimentation.
Compile the project (including all dependencies) in dev mode.
cargo build
The neural network model can be ran by calling src/main.rs like so:
cargo run --release
By default, the model will be trained using the Wgpu
backend. The training should show
up as a Terminal User Interface (TUI) dashboard:
Logs will be saved to /artifacts/vXX/experiment.log by default. Hyperparameters can be
adjusted by modifying the default values in the TrainingConfig struct in
src/training.rs. After training, the inference script will be ran automatically, and
the results saved to artifacts/vXX/submission.csv.
- Yu, S., Hannah, W., Peng, L., Lin, J., Bhouri, M. A., Gupta, R., Lütjens, B., Will, J. C., Behrens, G., Busecke, J., Loose, N., Stern, C. I., Beucler, T., Harrop, B., Hillman, B. R., Jenney, A., Ferretti, S., Liu, N., Anandkumar, A., … Pritchard, M. (2023). ClimSim: A large multi-scale dataset for hybrid physics-ML climate emulation (Version 5). arXiv. https://doi.org/10.48550/ARXIV.2306.08754