rectv_gpu is a C++ module with python wrapper for multi-GPU for four-dimensional (3D+time) tomographic reconstruction by time domain decomposition. For details, see
Nikitin, Viktor V., et al. "Four-dimensional tomographic reconstruction by time domain decomposition." IEEE Transactions on Computational Imaging 5.3 (2019): 409-419.(https://ieeexplore.ieee.org/abstract/document/8637033)
The module provides a tomographic reconstruction method, which relaxes the requirement of quasistatic samples during data acquistion. Dynamic tomographic data sets are decomposed in the temporal domain by projecting to a lower dimensional subspace of basis functions and deploying an additional L1 regularization technique where the penalty factor is taken for spatial and temporal derivatives. The iterative scheme involves solving the projected regularization problem with using modern multi-GPU systems.
Example of reconstruction:
.. image:: rectv.png :width: 90% :align: center
::
git clone https://github.com/nikitinvv/rectv_gpu cd rectv_gpu; python setup.py install
::
module add GCC/8.2.0-2.31.1 icc/2019.1.144-GCC-8.2.0-2.31.1 ifort/2019.1.144-GCC-8.2.0-2.31.1 CUDA/10.1.105 conda install -c conda-forge dxchange scikit-build matplotlib notebook ipywidgets cd rectv_gpu; python setup.py install
Use MAX IV gn1-3 nodes for remote jupyter notebook
Allocate GPU resources::
salloc -p v100
In remote host (e.g. gn1), open the terminal, change directory to where you have your notebooks and type::
jupyter notebook --no-browser --port=
E.g. port1=13543
On your local computer type::
ssh -N -f -L localhost:
E.g. port2=13545, address=w-picard05-clu0-gn-1.maxiv.lu.se
Now open web browser (google chrome, firefox, ...) and type::
localhost:
Examples with jupyter notebook
See examples/
Viktor Nikitin (vnikitin@anl.gov)