General regularisation #66
Conversation
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
| # a sparse matrix. | ||
|
|
||
| quadreg(μ, ν, C, ε; maxiter=500); | ||
| ot_reg_plan(μ, ν, C, ε; reg_func = "L2", method = "lorenz", maxiter=500); |
There was a problem hiding this comment.
[JuliaFormatter] reported by reviewdog 🐶
| ot_reg_plan(μ, ν, C, ε; reg_func = "L2", method = "lorenz", maxiter=500); | |
| ot_reg_plan(μ, ν, C, ε; reg_func="L2", method="lorenz", maxiter=500); |
| # instead of entropic regularisation: | ||
|
|
||
| γquad = Matrix(quadreg(μ, ν, C, 5; maxiter=500)) | ||
| γquad = Matrix(ot_reg_plan(μ, ν, C, 5; reg_func = "L2", method = "lorenz", maxiter=500)) |
There was a problem hiding this comment.
[JuliaFormatter] reported by reviewdog 🐶
| γquad = Matrix(ot_reg_plan(μ, ν, C, 5; reg_func = "L2", method = "lorenz", maxiter=500)) | |
| γquad = Matrix(ot_reg_plan(μ, ν, C, 5; reg_func="L2", method="lorenz", maxiter=500)) |
| See also: [`ot_reg_plan`](@ref) | ||
| """ | ||
| function ot_reg_cost(mu, nu, C, eps; reg_func="L2", method="lorenz", kwargs...) |
There was a problem hiding this comment.
[JuliaFormatter] reported by reviewdog 🐶
| function ot_reg_cost(mu, nu, C, eps; reg_func="L2", method="lorenz", kwargs...) | |
| nothing |
src/OptimalTransport.jl
Outdated
| quadreg(mu, nu, C, eps; kwargs...) | ||
| else | ||
| @warn "Unimplemented" | ||
| nothing |
There was a problem hiding this comment.
[JuliaFormatter] reported by reviewdog 🐶
| nothing |
Pull Request Test Coverage Report for Build 869354795Warning: This coverage report may be inaccurate.This pull request's base commit is no longer the HEAD commit of its target branch. This means it includes changes from outside the original pull request, including, potentially, unrelated coverage changes.
Details
💛 - Coveralls |
Codecov Report
@@ Coverage Diff @@
## master #66 +/- ##
==========================================
- Coverage 85.71% 85.50% -0.22%
==========================================
Files 2 2
Lines 259 269 +10
==========================================
+ Hits 222 230 +8
- Misses 37 39 +2
Continue to review full report at Codecov.
|
| Lorenz, D.A., Manns, P. and Meyer, C., 2019. Quadratically regularized optimal transport. Applied Mathematics & Optimization, pp.1-31. | ||
| """ | ||
| function ot_reg_plan(mu, nu, C, eps; reg_func="L2", method="lorenz", kwargs...) | ||
| if (reg_func == "L2") && (method == "lorenz") |
There was a problem hiding this comment.
This approach is problematic in my opinion: it is not possible to add support for other methods or regularization types for users or downstream packages, one always has to modify this function here.
I guess this could be avoided with the suggestion in #63 (comment) - everyone could just add other regularizations and/or algorithms.
|
Can this PR be closed, @zsteve ? |
Instead of using the
quadregfunction call for the quadratic regularisation, this PR provides a general set of functionsot_reg_planandot_reg_costfor computing OT with 'general' regularisations. Nowquadregis wrapped asreg_func = "L2", andmethod = "lorenz"selects the current algorithm. However, there are other approaches out there (e.g. L-BFGS-B) for the same OT problem that could be implemented using the same interface.The idea is that eventually more regularisations can be added (see e.g. https://arxiv.org/abs/1710.06276), and there should be a unified interface for this.