Tomographic algorithms implemented based on machine learning techniques to perform reconstructions in JET bolometer setup and COMPASS SXR system. Two different methods were implemented :
-
Inverse tomographic transformation matrix fitting via gradient descent for genetarion of new tomograms by a single matrix multiplication step
-
Deep Neural Networks
The code presented relates to the work developed for my M.Sc. Thesis [1] at Instituto de Plasma e Fusão Nuclear - Instituto Superior Técnico (IPFN/IST) under the supervision of Prof. Diogo R. Ferreira and Prof. Horácio Fernandes. Includes the work that generated a contribution for the European Physical Society Conference in Plasma Physics (EPS2018) in Prague [2].
Part of the code is based on the work previously developed at JET by D. R. Ferreira in [3] and the original can be accessed here
One must have installed and configured the following programs and packages to run the available code at full performance:
{
"image_data_format": "channels_first",
"epsilon": 1e-07,
"floatx": "float32",
"backend": "theano"
}
Note : Running the NN implemetation, as it is, using TensorFlow backend for Keras is also possible but the performance will be considerably slower. If you pretend to use TensorFlow and maintain the same performance edit ~/.keras/keras.json to contain :
{
"image_data_format": "channels_last",
"epsilon": 1e-07,
"floatx": "float32",
"backend": "tensorflow"
}
Additionaly make the necessary changes in function resize_NN_image() (present in COMPASS/bib/bib_utils.py and JET/bib/bib_utils.py) to ensure the channel axis is the last one.
The folders developed for JET/ and COMPASS/ are completely independent of each other. No data from the actual experimets is available in this repository since disclosure is not permited. Neverthless scripts to read from the databases (which will only work if one has access to the 5904 m) are given as well as instructions for the output files structure.
[1] Plasma Tomography with Machine Learning - D. D. Carvalho (2018)
[2] Regularization extraction for real-time plasma tomography at JET - D. R. Ferreira, D. D. Carvalho, P. J. Carvalho, H. Fernandes (2018)
[3] Full-Pulse Tomographic Reconstruction with Deep Neural Networks - D. R. Ferreira, P. J. Carvalho, H. Fernandes (2018)