A CUDA implementation of Self-Organizing Maps for unsupervised learning. The learning process is done by competitive learning between neurons of the SOM (it is possible to chose between online learning or batch learning). The closest unit to the input vector is called Best Matching Unit (BMU) and its weigth vector is moved in the direction of the input vector. Also the neighbourhood of the BMU is moved in the same direction, but with lower magnitude, according to the distance from the BMU. This tool allows to specify the following parameters used in the learning process:
- Number of rows of the SOM
- Number of columns of the SOM
- Initial and final learning rate
- Number of iterations in the learning process
- Radius of the updating function
- Various type of distances for the BMU search (Euclidean, Sum of Squares, Manhattan, Tanimoto)
- Various neighbour function (gaussian, bubble, mexican hat)
- Two types of lattice for the neurons of the SOM (square or exagonal)
- Possibility to use a toroidal topology
- Possibility to decay the learning rate and/or the radius exponentially
The tool also allows to change some CUDA related parameters, such as:
- GPU index
- Number of threads per each block launched on the GPU
- An option to move all possible computation on the GPU
A small benchmark is also included to identify the minimum size of your input file to make GPU computation advantageous.
The input samples must be provided in tab-separated format.
Here you can find some examples of the tool uses:
- The tool will use the file provided as input, will train for 1000 iterations with a learning rate that will linearly decay from 0.1 to 0.001. The distance between neurons and the input vector will be normalized. The size of the SOM and the radius of the updating function will be estimated runtime with heuristcs. Neurons will be represented by exagones.
./CUDA-SOM.out -i /folder/folder/inputfile.txt -n 1000 -s 0.1 -f 0.001 --normalizedistance
- The tool will use the file provided as input, will train for 5000 iterations in batch mode. The size of the SOM is set to 200x200, the initial radius of the updating function (bubble) is set to 50, and will decrease exponentially. The distance between neurons and the input vector is computed using the sum of squares distance. Once the SOM is trained, the distances between the input file and the SOM will be saved to a file. The GPU used for the computation will be the one with index 2; 96 threads per block will be launched.
./CUDA-SOM.out -i /folder/folder/inputfile.txt -n 5000 -m b -x 200 -y 200 --savedistances -r 50 --distance=s --neighbors=b --toroidal --exponential=r --GPUIndex=2 --threadsperblock=96