Parallelized rotation and flipping INvariant Kohonen maps (PINK)

Requirements

• C++ with ISO 14 standard (GCC 6.1 and above)
• CMake >= 3.0
• CUDA >= 9.1 (highly recommended because of the speed-up)
• conan.io (optional for C++ dependencies) or
  • PyBind11 (optional for Python interface)
  • google-test 1.8.1 (optional for unit tests)
• doxygen 1.8.13 (optional for developer documentation)

Conan.io will install automatically the C++ dependencies (PyBind11 and google-test). Otherwise you can also install these libraries yourself.

Installation

We provide deb- and rpm-packages at https://github.com/HITS-AIN/PINK/releases

or you can install PINK from the sources:

cmake -DCMAKE_INSTALL_PREFIX=<INSTALL_PATH> .
make install

PyPI installation

PINK is also available as PyPi package which can be installed by

pip install astro-pink

HPC deployment with EasyBuild

The EasyBuild recipe is available at https://github.com/BerndDoser/easybuild-easyconfigs/tree/hits/easybuild/easyconfigs/p/PINK.

Usage

To train a the self-organizing map (SOM) please execute

Pink --train <image-file> <result-file>

where image-file is the input file of images for the training and result-file is the output file for the trained SOM. All files are in binary mode described here.

To map an image to the trained SOM please execute

Pink --map <image-file> <result-file> <SOM-file>

where image-file is the input file of images for the mapping, SOM-file is the input file for the trained SOM, and result-file is the output file for the resulting heatmap.

Please use also the command Pink -h to get more informations about the usage and the options.

Python scripts

For conversion and visualization of images and SOM some python scripts are available.

• convert_data_binary_file.py Convert binary data file from PINK version 1 to 2
• show_heatmap.py: Visualize the mapping result
• show_images.py: Visualize binary images file format
• show_som.py: Visualize binary SOM file format
• train.py: SOM training using the PINK Python interface

Benchmarks

The input data for the SOM training are radio-synthesis images of Radio Galaxy Zoo containing 176750 images of the dimension 124x124. The SOM layout is hexagonal of the dimension 21x21 which has 331 neurons (see image above). The size of the neurons is 64x64. The accuracy for the rotational invariance is 1 degree and the flip invariance is used.

	PINK 1	Pink 2
CPU-1		35373
CPU-1 + NVIDIA Tesla P40	3069	909
CPU-1 + 2x NVIDIA Tesla P40	2069	636
CPU-1 + 4x NVIDIA Tesla P40	1891	858
CPU-2 + NVIDIA RTX 2080		673
CPU-3 + NVIDIA GTX 750 Ti		7185
CPU-4 + 2x NVIDIA RTX 2080 SUPER		477

All times are in seconds.

• CPU-1: Intel Gold 5118 (2 sockets, 12 physical cores per socket)
• CPU-2: Intel Core i7-8700K (1 socket, 6 physical cores per socket)
• CPU-3: Intel Core i7-4790K (1 socket, 4 physical cores per socket)
• CPU-4: Intel Gold 6230 (1 socket, 20 physical cores per socket)

Publication

Kai Lars Polsterer, Fabian Gieseke, Christian Igel, Bernd Doser, and Nikos Gianniotis. Parallelized rotation and flipping INvariant Kohonen maps (PINK) on GPUs. 24th European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning (ESANN), pp. 405-410, 2016. pdf

License

Distributed under the GNU GPLv3 License. See accompanying file LICENSE or copy at http://www.gnu.org/licenses/gpl-3.0.html.