Introducing Gaussian Mixture Models #692
Conversation
There was a problem hiding this comment.
Looks like a useful and relatively complete capability. I tried to keep comments relatively high level for now. This will need clang-format, doxygen comments, conflicts with ParticleInit fixed, and at least one unit test (it's not clear to me yet how to use this in the end)
| template <typename CellSliceType, typename WeightSliceType, | ||
| typename PositionSliceType, typename VelocitySliceType, | ||
| typename GaussianType> | ||
| size_t initializeRandomParticles( CellSliceType& cell, WeightSliceType& macro, |
There was a problem hiding this comment.
I believe the conflicts with master are related to the rewrite of particle init. The changes were primarily to match the structured-grid side of the library, where each free function is now createParticles( CreationTypeTag, ... ): currently only InitRandom available, but this seems like it should become createParticles( InitGaussian, ... )
|
@tyounkin do you see things here that would be useful (or things you would want to extend/modify)? |
| }; | ||
|
|
||
| // Define an execution policy | ||
| Cabana::SimdPolicy<cell.vector_length, Kokkos::DefaultExecutionSpace> |
There was a problem hiding this comment.
One other recent change across the library was to expose user-chosen execution spaces in addition to overloads with the default space for the memory being used - we'll want to support any valid exec_space here and then add a wrapper using the default
There was a problem hiding this comment.
Can you point me to an example where you expose the capability to run on any execution space, but have a default selection for the user?
There was a problem hiding this comment.
Here's one:
Cabana/core/src/Cabana_VerletList.hpp
Line 728 in fa9e58a
We store the default exec space for a given memory space in the class to use in this overload, but also give the user the option to pass any (compatible) exec space in the next one
which is a number of weightes Gaussians that describe a particle distribution function in velocity space. The mixture model can be computed from a set of particles and we also allow to draw new particles from such a Gaussian Mixture Model. The code follows the papers by Figueiredo and Jain, 2002 (https://dx.doi.org/https://dl.acm.org/doi/10.1109/34.990138), but adds particles weights like the paper by Chen, Chacon, Nguyen, 2021 (DOI:https://doi.org/10.1016/j.jcp.2021.110185)
| }; | ||
|
|
||
| // Define an execution policy | ||
| Cabana::SimdPolicy<cell.vector_length, Kokkos::DefaultExecutionSpace> |
There was a problem hiding this comment.
Here's one:
Cabana/core/src/Cabana_VerletList.hpp
Line 728 in fa9e58a
We store the default exec space for a given memory space in the class to use in this overload, but also give the user the option to pass any (compatible) exec space in the next one
petschge
force-pushed
the
GMM_pull_request_1
branch
from
090aa55
to
6c20d05
Compare
and put the implemenation in the GMMImpl class and the user facing bits into one header file.
which is a number of weightes Gaussians that describe a particle distribution function in velocity space. The mixture model can be computed from a set of particles and we also allow to draw new particles from such a Gaussian Mixture Model. The code follows the papers by Figueiredo and Jain, 2002
(https://dx.doi.org/https://dl.acm.org/doi/10.1109/34.990138), but adds particles weights like the paper by Chen, Chacon, Nguyen, 2021 (DOI:https://doi.org/10.1016/j.jcp.2021.110185)
This is not against the current head of master, but I assume that there is other changes that I need to make before rebasing and submitting for good. Feedback (here or on Slack) welcome!