SolarNet-Lib 0.1.0

Deep Learning for Solar Physics Prediction

SolarNet

Deep Learning for Solar Physics Prediction

The SolarNet library permits to use the different Pytorch models, datasets, preprocessing, and others utilities developed during the SolarNet project. It also makes straightforward the download, loading, and finetuning of the big pretrained SSL models. With SolarNet, anyone can access the datasets (at this time, only SDO-Benchmark is available through the library) and finetune the powerful models. The library is compatible with Pytorch-Lightning, but the models are also pure Pytorch Module and the training loop can be written from scratch.

Find the docs on jdonzallaz.gitlab.io/solarnet.

Installation

Use pip to install:

pip install solarnet-lib

Python 3.6+ is required.

Data

Two datasets are supported: SDO-Dataset and SDO-Benchmark. SDO-Dataset needs to be downloaded from the Stanford's servers. The SDO-Benchmark dataset can be downloaded using the CLI:

solarnet download sdo-benchmark data/

Dataset

path = Path("data") / "sdo-benchmark" / "train"
dataset = SDOBenchmarkDataset(path)

Also available as a pytorch-lightning datamodule.

path = Path("data") / "sdo-benchmark"
datamodule = SDOBenchmarkDatamodule(path)

Model

model = ImageClassification.from_pretrained("solarnet-ssl-bz-ft-sdo-benchmark")

Finetuning

trainer = pl.Trainer(
    max_epochs=10,
    gpus=0,
)
trainer.fit(model, datamodule=datamodule)

Deployment

python setup.py sdist bdist_wheel
twine check dist/*
twine upload dist/*

Author

SolarNet is a deep learning research toolbox for solar physics. It was developed during a Master thesis by Jonathan Donzallaz.