SOMeSolution

An iteratively developed approach to the problem of fast training of Self Organizing Maps. This is a working implementation of the HPSOM algorithm described by Liu et al. This implementation can be run on:

• Serial architecture (through the batch-som branch with make buildserial)
• Shared memory architecture (using OpenMP through the batch-som branch with the OMP_NUM_THREADS environment variable)
• Distributed memory architecture (using OpenMPI through the mpi branch)
• Nvidia GPU architecture with shared memory (using CUDA and OpenMP through the cuda branch)
• Distributed memory Nvidia GPU architecture with shared memory (using OpenMPI, CUDA, and OpenMP through the mpicuda branch)

C++ install

First clone the repository and checkout the branch of the version you want to use (batch-som, mpi, cuda, mpicuda)

git clone https://github.com/awyeasting/SOMeSolution.git
cd SOMeSolution
git checkout mpicuda

Then compile into either a library or an executable. (NOTE: if you installed CUDA in a different location or with a different version than 11.2 you will need to change the install location at the top of the makefile)

To compile the code to a library,

cd SOMeSolution/src/C++
make

The static library will be in SOMeSolution/src/C++/bin/somesolution.a

To compile the code to a commandline usable executable,

cd SOMeSolution/src/C++
make build

The executable will be in SOMeSolution/src/C++/bin

Commandline Usage

Through the command line you can add different flags and optional arguments.

Arguments:

Positional Arguments:
	(int)    SOM width
	(int)    SOM height
	(string) Training data file name
Options:
	(int int)-g --generate       num features, num_dimensions for generating random data
	(string) -o --out            Path of the output file of node weights
	(int)    -e --epochs         Number of epochs used in training
	(int)    -s --seed           Integer value to intialize seed for generating
	         -l --labeled        Indicates the last column is a label
	(int)    -gp --gpus-per-proc The number of gpus each processor should utilize

Example: The following will make a 10 x 10 SOM on 2 processes, generate its own training data (which has 100 examples and 100 dimensions), train the SOM on it, and output the trained map to trained_map.txt.

mpirun -np 2 bin/somwork_mpicuda 10 10 -g 100 100 -o trained_map.txt

Python Visualization

To visualize a SOM weights file produced by the commandline executable, simply run:

python som.py -i weights.txt -d <display method>

(See python som.py -h for supported display methods)

License

3-Clause BSD