An R package implementation of the fast multiscale approach to optimal transport described in:
Multiscale Strategies for Computing Optimal Transport
Samuel Gerber, Mauro Maggioni
Journal of Machine Learning Research; 18(72):1−32, 2017.
http://jmlr.org/papers/v18/16-108.html
The code is designed for fast multiscale optimal transport but also permits single scale optimal transport computations and more:
- Optimal transport solved with Lemon, CPLEX, GPLK or MOSEK
- Fast approximate transport using a mutliscale strategy
library(devtools)
if(!require("gmra")){
devtools::install_github("samuelgerber/gmra")
}
Sys.setenv("R_REMOTES_NO_ERRORS_FROM_WARNINGS" = "true")
devtools::install_github("samuelgerber/mop")
#create example data
phi <- runif(1000)*2*pi
X1<- cbind(cos(phi), sin(phi)) * (1+rnorm(length(phi)) * 0.1)
X1[,1]=X1[,1]*0.5;
X1[,2]=X1[,2]*2;
phi <- runif(1653)*2*pi
X2<- cbind(cos(phi), sin(phi)) * (1+rnorm(length(phi)) * 0.1)
X2[,1]=X2[,1]*3;
X2[,2]=X2[,2]*0.5;
#create multiscale decompsotions
library(gmra)
gmra1 = gmra.create.ipca(X=X1, eps=0, d=2, maxKids=2)
gmra2 = gmra.create.ipca(X=X2, eps=0, d=2, maxKids=2)
#setup and solve multiscale lp
library(mop)
trp.lp <- multiscale.transport.create.lp(oType=26, transport.type=5, massCost=0.1)
icprop <- multiscale.transport.create.iterated.capacity.propagation.strategy(1, 0)
multiscale.transport.set.propagation.strategy.1(trp.lp, icprop);
time1 <- system.time(
trp1 <- multiscale.transport.solve(trp.lp, gmra1, gmra2, p = 2, nType=0, dType=1, scaleMass=FALSE) )
multiscale.transport.plot.map(trp1, 100, mapAlpha=0.1)
#add nieghborhood expansion
multiscale.transport.add.expand.neighborhood.strategy(trp.lp, 1 )
time2 <- system.time(
trp2 <- multiscale.transport.solve(trp.lp, gmra1, gmra2, p = 2, nType=0, dType=1) )
#solve optimal
time3 <- system.time(
trp3 <- multiscale.transport.solve(trp.lp, gmra1, gmra2, p = 2, nType=0, dType=1, scale1=0, scale2=0, scaleMass=FALSE) )
multiscale.transport.plot.map(trp3, 100, mapAlpha=0.1)
trp1$cost
trp2$cost
trp3$cost
time1
time2
time3