frobnitzem / bdas Goto Github PK

Applying pyscaffold to the project will allow creating optional compile-dependencies to mpi4py and cupy.

Once this is done, an init.py should be created that creates a class Context. The context class will provide mpi/no-op and numpy/cupy wrappers. Its defaults can be set to check library availability.

Three things:

• Each class has place-holders for where MPI_Allreduce should be called.
• K-Means requires an extra broadcast step (see FIXME).
• print statements throughout the code should call config.print(), which checks MPI rank == 0.

The SVM classifier uses a convex optimization strategy that doesn't match the original benchmark. Instead, it solves for the Bayesian maximum likelihood estimate of a linear classifier model as specified in Interior-point methods for large-scale cone programming (pdf), from the book Convex Optimization, Boyd and Vandenberghe. A Jeffrey's prior of |w| is added.

We need plots of iterations to convergence vs. N, M.