amirabbasasadi / RockyML

⛰️ RockyML - A High-Performance Scientific Computing Framework for Non-smooth Machine Learning Problems

📔 Documentation : amirabbasasadi.github.io/RockyML

Tutorials

Zagros Tutorials

• Using with CMake
• Create your first optimizer with Dena and Zagros
• Define a custom optimization problem in Zagros
• Block optimization for large problems
• Distributed optimization
• Designing search algorithms using Dena
• Test problems for benchmarking optimizers

Components

Zagros

Design Goals:

• Providing a language called Dena for designing arbitrary complex optimizers by combining
  • Modular and parallel search strategies: Genetic, PSO, EDA, ...
  • Communication strategies for distributed optimization on top of MPI
  • Analyzer strategies for analyzing objective functions
  • Blocking strategies for block optimization
  • Logging strategies for tracking optimization experiments on local system or a remote server
• Hybrid parallelism: multi-threading in each node and message passing across nodes (MPI)
• ‌Block optimization for using memory-intensive optimizers for large number of variables

#include <mpi.h>
#include <rocky/zagros/benchmark.h>
#include <rocky/zagros/flow.h>

using namespace rocky::zagros;
using namespace rocky::zagros::dena;

int main(int argc, char* argv[]){
    MPI_Init(&argc, &argv);
    
    // define the optimization problem
    const int dim = 100;
    benchmark::rastrigin<float> problem(dim);

    // recording the result of optimization
    local_log_handler log_handler("result.csv");

    // define the optimizer
    auto optimizer = container::create("A", 300)
                    >> init::uniform("A") 
                    >> run::n_times(500,
                            mutate::gaussian("A")
                            >> run::with_probability(0.2,
                                crossover::differential_evolution("A")
                            )
                            >> log::local::best("A", log_handler)
                        );

    // create a runtime for executing the optimizer 
    basic_runtime<float, dim> runtime(&problem);
    runtime.run(optimizer);

    MPI_Finalize();
    return 0;
}

Etna (Work in progress)

Building blocks for designing non-differentiable neural networks

• Fast, low overhead, and thread-safe
• Various components:
  • Standard deep learning layers
  • Discrete and integer layers
  • Combinatorial layers
  • Stochastic layers
  • Dynamic layers

About

Publications

If you use RockyML in your research, please cite it as follows:

@software{RockyML,
  author = {Asadi, Amirabbas},
  doi = {10.5281/zenodo.7612838},
  month = {2},
  title = {{RockyML, A Scientific Computing Framework for Non-smooth Machine Learning Problems}},
  url = {https://github.com/amirabbasasadi/RockyML},
  year = {2023}
}