Ecosyste.ms: Awesome

An open API service indexing awesome lists of open source software.

Awesome Lists | Featured Topics | Projects

https://github.com/alexbuccheri/fortran_kmeans

KMeans implementations in fortran
https://github.com/alexbuccheri/fortran_kmeans

fortran kmeans kmeans-clustering

Last synced: about 1 month ago
JSON representation

KMeans implementations in fortran

Awesome Lists containing this project

README 

        
# Test Program for Weighted K-Means Algorithm, plus the Fortuno Test Framework (in Fortran)



This project implements an MPI-OMP parallelised version of the weighted K-Means clustering using Lloyd’s algorithm.

In practice, the k-means algorithm is one of the fastest clustering algorithms available, but it falls in local minima. 

That’s why it can be useful to restart it several times, or utilise (greedy) kmeans++ to improve the robustness of the

initial seeding.



Additional information can be found at the scikit learn [docs](https://scikit-learn.org/stable/modules/generated/sklearn.cluster.KMeans.html)



## Serial (Well, no MPI but threaded)



```shell

# Define the project root

export KMEANS_ROOT=/Users/alexanderbuccheri/Codes/kmeans/

```



CMake configure, build, install and test commands:



```shell

cmake --fresh -B serial-cmake-build-debug -DCMAKE_BUILD_TYPE=Debug -DMPI=Off -DCMAKE_INSTALL_PREFIX=${KMEANS_ROOT}/serial_install

cmake --build serial-cmake-build-debug

ctest --test-dir ./serial-cmake-build-debug --output-on-failure

cmake --install serial-cmake-build-debug

```



Run the application test with:



```shell

./serial-cmake-build-debug/run_kmeans

```



One can also run each unit test binary manually. For example:



```shell

./serial-cmake-build-debug/test_grids "Real-space 1D grid"

```



## MPI



```shell

cmake --fresh -B cmake-build-debug -DCMAKE_BUILD_TYPE=Debug -DCMAKE_INSTALL_PREFIX=${KMEANS_ROOT}/mpi_install 

cmake --build cmake-build-debug

ctest --test-dir ./cmake-build-debug --output-on-failure

cmake --install cmake-build-debug

```



One can run the application test with:



```shell

mpirun -np 4 cmake-build-debug/run_kmeans

```



or have the notebook execute it.



### References



Additional `ctest` args:



* `--show-only`   List all tests

* `--verbose`     Always write to stdout



Fortuno is available [here](https://github.com/fortuno-repos/fortuno)



The MPI version is available [here](https://github.com/fortuno-repos/fortuno-mpi/)



## Centroid Algorithm MPI Implementation Description



In the MPI implementation, `assign_points_to_centroids` does not change. All centroids are defined on all processes.

Instead, `assign_points_to_centroids` expects a grid or sub-grid, local to a process. It returns `cluster_sizes` and `clusters`,

which assign all points of the sub-grid to clusters.



`update_centroids` also remains the same. It returns `centroids` of shape `(n_dim, n_centroids)`, which have only been computed

using a local set of grid points. The MPI overload of this routine `allreduce`s the centroids(n_dim, n_centroids), such that the

final centroids contain contributions from all grid points, over all processes. `centroids` are now complete on all processes.



## TODOs



* Unit tests for debug build all crash when building with `-ffpe-trap=invalid,zero,overflow,underflow` and/or

  `-finit-real=nan`. This needs investigating. One assumes the issue is in the grid routines.



* Implement "kmeans++" to find initial weights - see my python reference

  * Extend to "greedy kmeans++"



* Implement QR decomposition with pivoting - see my python reference implementation