This repository contains an extremely simple way of using SICON, the neural inversion method developed by Asensio Ramos & Diaz Baso (2019, A&A, 626, 102A). With only two commands you can download Hinode data and get the inversion in just a few minutes.
Search for the initial date and time of the Hinode observation you want. It can
be done by using the Hinode database http://sdc.uio.no/search/form. Once
found, save the date and time of observation and go to the Hinode level 1 database
that can be found in http://www.lmsal.com/solarsoft/hinode/level1hao/. Enter
into the subdirectory structure until you find your observation. For instance,
let's say you're searching for the observation that started at 19:00 on
2014/02/04. Go to
http://www.lmsal.com/solarsoft/hinode/level1hao/2014/02/04/SP3D/20140204_190005
and you will find the data as FITS file. Once located, you can call the
download_hinode.py command that will download all the FITS files and generate an HDF5 file
appropriate for the input to the neural inversor:
python download_hinode.py --url http://www.lmsal.com/solarsoft/hinode/level1hao/2014/02/04/SP3D/20140204_190005 --output output.h5 --downloader curl
By default, the files are downloaded with wget. If you do not have it available in your system,
which happens in many cases in MacOS, you can use curl. You can find some help of the options
of the downloader by using python download_hinode -h.
The code needs as input an HDF5 file with the Stokes profiles downloaded from the Hinode database and will produce as output another HDF5 file with the resulting physical conditions. The code is run as:
python neural_hinode.py --input input.h5 --output output.h5 --normalize 0 100 0 100 --device cpu --resolution 1
The arguments are the input and output files, together with the definition of a box which is used
as the quiet Sun for normalizing the Stokes profiles. You should check a proper window for this normalization
to be correct. Additionally, you can decide to do
the computation on CPU or GPU. Calculations on GPUs will be faster but the field-of-view
is memory limited. For large maps, we recommend using CPU. Finally, resolution 1 allows you to output the
results with the same pixel size as the observations, while resolution 2 upsamples the solution to
double resolution. Use it with care.
The results are: T [K], vz [km/s], h [km], log P [cgs], (B_x^2-B_y^2)^{1/2} [kG], (B_x B_y)^{1/2} [kG], Bx [kG], By [kG] and Bz [kG].
We provide the scripts for retraining your models in case you want to tweak the neural networks. You can download
the training data from here. The scripts are found in the
training directory. Do any modification and run them. You should need to create the directories for the
output.
bs4
numpy
argparse
h5py
tqdm
astropy
torch
We recommend to use Anaconda to run this code. We also recommend to generate a new environment in which all the packages will be installed. Then, install PyTorchas indicated in the webpage https://pytorch.org/ depending on your system . A typical process would be:
conda create -n inversor python=3.8
conda activate inversor
conda install -c conda-forge numpy h5py tqdm astropy bs4 argparse
conda install pytorch torchvision cudatoolkit=10.1 -c pytorch