Flareformer

PyTorch training code for Flareformer.

Getting Started

Requirements

• Python >= 3.8.10
• CUDA >= 11.1
• PyTorch >= 1.8.2

Installation

1. git clone git@github.com:keio-smilab22/Flareformer.git
2. cd Flareformer
3. pip install -r requirements.txt

Data preparation

• The required data files should be put under data/ folder.
• Visit http://wdc.nict.go.jp/IONO/wdc/solarflare/index.html and download the database of physical features.

$ cd ~/work/flare_transformer
$ mv ~/data.zip data/
$ unzip data/data.zip

• Visit https://sdo.gsfc.nasa.gov/data/ and download hourly magnetograms.

$ mv ~/magnetogram_images.tar.gz data/
$ tar -zxvf data/magnetogram_images.tar.gz

Preprocess

• Preprocess the physical features and magnetogram images by the following procedure (parallel processing).

  • python preprocess/main.py --magnetogram --physical --label --window

• The following data files should be created under data/.

  • data/data_20XX_magnetogram_256.npy
  • data/data_20XX_feat.csv
  • data/data_20XX_label.csv
  • data/data_20XX_window_48.csv

Training

• To train Flareformer with warmup and cRT using params/params_2014.json:

  • python main.py --params params/params_2014.json --warmup_epochs=5 --imbalance --wandb

License

Flareformer is released under the BSD-3-Clause-Clear license. Please see the LICENSE file for more information.

Others

Qualitative Results

• The figure below shows line-of-sight magnetograms from September 3th, 2017 23:00 UT to September 5th, 2017, 23:00 UT. An X-class solar flare occurred at 12:02 on September 6, 2017, and the model was able to predict the correct maximum solar flare class.

  

Developer's env.

Hardware

• Ubuntu 20.04LTS
• GPU: GeForce RTX 3080 laptop
• GPU RAM: 16GB GDDR6
• CPU: Core i9 11980HK
• RAM: 64 GB Libraries
• CUDA 11.1