Nothing
gridAugmentData <- function(seed=11, drop=25, na=20){
set.seed(seed)
dat <- expand.grid(trial=factor(c("A", "B")), range=factor(1:8), row=factor(1:10))
dat$geno <- factor(sample(sprintf("g%02d", 1:20), nrow(dat), replace=TRUE))
dat$y <- 10 + as.numeric(dat$trial) + rnorm(nrow(dat))
dat$y[sample(nrow(dat), na)] <- NA
dat[-sample(nrow(dat), drop), , drop=FALSE]
}
gridAugmentFit <- function(dat, mode, ..., rcov=~dsumm(vsm(ar1m(range), ar1m(row), ism(units)), by=trial),
representation=list(), solve=list()){
sommer:::.mmes_dispatch(y ~ trial, random=~geno, rcov=rcov, data=dat,
control=sommer:::.mmes_control_internal(representation=utils::modifyList(list(gridAugment=mode), representation),
solve=solve, convergence=list(maxIters=25),
diagnostics=list(verbose=FALSE, dateWarning=FALSE)), ...)
}
expectSameFit <- function(a, b, tolerance=1e-6){
expect_equal(tail(as.numeric(b$llik), 1), tail(as.numeric(a$llik), 1), tolerance=tolerance)
expect_equal(unlist(b$theta), unlist(a$theta), tolerance=tolerance)
expect_equal(as.numeric(b$b), as.numeric(a$b), tolerance=tolerance)
expect_equal(rownames(b$b), rownames(a$b))
expect_equal(as.numeric(b$u), as.numeric(a$u), tolerance=tolerance)
expect_equal(as.numeric(fitted(b)), as.numeric(fitted(a)), tolerance=tolerance)
expect_equal(as.numeric(residuals(b)), as.numeric(residuals(a)), tolerance=tolerance)
expect_equal(b$AIC, a$AIC, tolerance=tolerance)
expect_equal(b$BIC, a$BIC, tolerance=tolerance)
expect_equal(dim(b$W), dim(a$W))
}
test_that("grid augmentation of dsumm AR1 x AR1 residuals is exact for REML", {
dat <- gridAugmentData()
off <- gridAugmentFit(dat, "off")
auto <- gridAugmentFit(dat, "auto")
info <- auto$engineDiagnostics$gridAugmentation
expect_true(info$active)
expect_equal(info$blocks, 2L)
expect_equal(info$observed, sum(!is.na(dat$y)))
expect_equal(info$cells, 2L * 8L * 10L - sum(!is.na(dat$y)))
expect_null(off$engineDiagnostics$gridAugmentation$active)
expectSameFit(off, auto)
expect_equal(nrow(auto$Ci), nrow(off$Ci))
pOff <- predict(off, D="trial")$pvals
pAuto <- predict(auto, D="trial")$pvals
expect_equal(pAuto$predicted.value, pOff$predicted.value, tolerance=1e-6)
expect_equal(pAuto$std.error, pOff$std.error, tolerance=1e-6)
vOff <- summary(off)$varcomp
vAuto <- summary(auto)$varcomp
expect_equal(vAuto$VarComp, vOff$VarComp, tolerance=1e-6)
expect_equal(vAuto$VarCompSE, vOff$VarCompSE, tolerance=1e-5)
})
test_that("grid augmentation keeps a single symbolic analysis and sparse grid blocks", {
dat <- gridAugmentData(seed=12)
auto <- gridAugmentFit(dat, "auto")
d <- auto$engineDiagnostics
expect_true(d$gridAugmentation$active)
expect_equal(d$CsymbolicAnalyses, 1)
expect_equal(d$gridSparseBlocks, 2)
expect_gt(d$gridSparseBlockUses, 0)
expect_equal(d$gridBandFallbacks, 0)
})
test_that("grid augmentation matches with the ldlt solver and a single-trial residual", {
dat <- droplevels(gridAugmentData(seed=13)[gridAugmentData(seed=13)$trial == "A", ])
rcov <- ~vsm(ar1m(range), ar1m(row), ism(units))
fit <- function(mode) sommer:::.mmes_dispatch(y ~ 1, random=~geno, rcov=rcov, data=dat,
control=sommer:::.mmes_control_internal(representation=list(gridAugment=mode), solve=list(method="ldlt"),
convergence=list(maxIters=25), diagnostics=list(verbose=FALSE, dateWarning=FALSE)))
off <- fit("off")
auto <- fit("auto")
expect_true(auto$engineDiagnostics$gridAugmentation$active)
expect_equal(auto$engineDiagnostics$gridAugmentation$blocks, 1L)
expectSameFit(off, auto)
})
test_that("banded sparse grid blocks match the dense grid path", {
dat <- expand.grid(trial=factor(c("A", "B")), range=factor(1:8), row=factor(1:10))
set.seed(14)
dat$geno <- factor(sample(sprintf("g%02d", 1:20), nrow(dat), replace=TRUE))
dat$y <- rnorm(nrow(dat))
sparse <- gridAugmentFit(dat, "off")
dense <- gridAugmentFit(dat, "off", representation=list(sparseFactors="off"))
expect_gt(sparse$engineDiagnostics$gridSparseBlocks, 0)
expect_equal(dense$engineDiagnostics$gridSparseBlocks, 0)
expect_equal(sparse$engineDiagnostics$CsymbolicAnalyses, 1)
expectSameFit(dense, sparse)
})
test_that("grid augmentation is skipped quietly or errors when forced on ineligible models", {
dat <- gridAugmentData(seed=15)
ml <- sommer:::.mmes_dispatch(y ~ trial, random=~geno, rcov=~dsumm(vsm(ar1m(range), ar1m(row), ism(units)), by=trial),
data=dat, control=sommer:::.mmes_control_internal(task=list(objective="ml"), convergence=list(maxIters=10),
diagnostics=list(verbose=FALSE, dateWarning=FALSE)))
expect_false(isTRUE(ml$engineDiagnostics$gridAugmentation$active))
expect_error(
sommer:::.mmes_dispatch(y ~ trial, random=~geno, rcov=~dsumm(vsm(ar1m(range), ar1m(row), ism(units)), by=trial),
data=dat, control=sommer:::.mmes_control_internal(task=list(objective="ml"), representation=list(gridAugment="force"),
convergence=list(maxIters=10), diagnostics=list(verbose=FALSE, dateWarning=FALSE))),
"gridAugment='force'")
diag <- gridAugmentFit(dat, "auto", rcov=~vsm(dsm(trial), ism(units)))
expect_false(isTRUE(diag$engineDiagnostics$gridAugmentation$active))
expect_error(gridAugmentFit(dat, "force", rcov=~vsm(dsm(trial), ism(units))), "gridAugment='force'")
})
test_that("the gridAugment option sets the default mode", {
dat <- gridAugmentData(seed=16)
old <- options(sommer.mme.gridAugment="off")
on.exit(options(old), add=TRUE)
fit <- sommer:::.mmes_dispatch(y ~ trial, random=~geno, rcov=~dsumm(vsm(ar1m(range), ar1m(row), ism(units)), by=trial),
data=dat, control=sommer:::.mmes_control_internal(convergence=list(maxIters=10), diagnostics=list(verbose=FALSE, dateWarning=FALSE)))
expect_false(isTRUE(fit$engineDiagnostics$gridAugmentation$active))
})
test_that("reduced-rank AI uses direct residual traces on augmented grids", {
set.seed(17)
dat <- expand.grid(trial=factor(c("A", "B", "C")), range=factor(1:6), row=factor(1:8))
dat$geno <- factor(sample(sprintf("g%02d", 1:15), nrow(dat), replace=TRUE))
gEff <- rnorm(15)
dat$y <- as.numeric(dat$trial) + gEff[as.integer(dat$geno)] * c(1, 0.8, 1.2)[as.integer(dat$trial)] + rnorm(nrow(dat))
dat <- dat[-sample(nrow(dat), 15), ]
fit <- function(mode) sommer:::.mmes_dispatch(y ~ trial, random=~vsm(rrm(trial, 1), ism(geno)) + vsm(dsm(trial), ism(geno)),
rcov=~dsumm(vsm(ar1m(range), ar1m(row), ism(units)), by=trial), data=dat,
control=sommer:::.mmes_control_internal(representation=list(gridAugment=mode), convergence=list(maxIters=6),
diagnostics=list(verbose=FALSE, dateWarning=FALSE)))
off <- suppressWarnings(fit("off"))
auto <- suppressWarnings(fit("auto"))
expect_true(auto$engineDiagnostics$gridAugmentation$active)
expect_gt(auto$engineDiagnostics$directResidualTraces, 0)
expect_equal(as.numeric(auto$llik), as.numeric(off$llik), tolerance=1e-6)
expect_equal(unlist(auto$theta), unlist(off$theta), tolerance=1e-5)
expect_equal(as.numeric(fitted(auto)), as.numeric(fitted(off)), tolerance=1e-5)
})
test_that("FA grid traces prefer the selected inverse at ranks one to three", {
set.seed(18)
dat <- expand.grid(trial=factor(letters[1:4]), range=factor(1:4), row=factor(1:5))
dat$geno <- factor(sample(sprintf("g%02d", 1:8), nrow(dat), replace=TRUE))
dat$y <- as.numeric(dat$trial) + rnorm(8)[as.integer(dat$geno)] + rnorm(nrow(dat))
dat <- dat[-c(6L, 17L, 34L, 49L, 62L), ]
precision <- diag(8) + matrix(0.1, 8, 8)
dimnames(precision) <- rep(list(levels(dat$geno)), 2)
attr(precision, "inverse") <- TRUE
for(rank in 1:3){
shape <- fam(dat$trial, rank)
fit <- function(mode, solver="cholmod", inverse=0L, budget=1000){
sommer:::.mmes_dispatch(y ~ trial, random=~vsm(shape, ism(geno), Gu=precision),
rcov=~dsumm(vsm(ar1m(range), ar1m(row), ism(units)), by=trial), data=dat,
control=sommer:::.mmes_control_internal(representation=list(gridAugment=mode),
solve=list(method=solver, blockSchur="off"), output=list(inverse=inverse),
execution=list(denseMemoryMB=budget,
memoryPolicy=if(budget < 1000) "limit" else "speed"),
convergence=list(maxIters=4),
diagnostics=list(verbose=FALSE, dateWarning=FALSE, profile="force")))
}
reference <- fit("off", inverse=2L)
selected <- fit("auto", inverse=2L)
ldlt <- fit("auto", solver="ldlt")
d <- selected$engineDiagnostics
expect_true(d$gridAugmentation$active)
expect_equal(d$CsymbolicAnalyses, 1)
expect_gt(d$profile$counters$patternInverseCacheBuilds, 0)
expect_equal(d$profile$counters$patternInverseCacheFallbacks, 0)
expect_equal(d$profile$counters$traceFallbackSolves, 0)
expectSameFit(reference, selected)
expectSameFit(selected, ldlt)
for(component in c("llik", "covPar", "InfMat", "dLu")){
expect_equal(selected[[component]], reference[[component]], tolerance=1e-6)
expect_equal(ldlt[[component]], selected[[component]], tolerance=1e-6)
}
expect_equal(as.matrix(selected$Ci), as.matrix(reference$Ci), tolerance=1e-6)
expect_equal(selected$uPevList, reference$uPevList, tolerance=1e-6)
expect_equal(predict(selected, D="trial")$pvals,
predict(reference, D="trial")$pvals, tolerance=1e-6)
if(rank == 1L){
limited <- fit("auto", budget=0.001)
expect_equal(limited$engineDiagnostics$profile$counters$patternInverseCacheBuilds, 0)
expect_gt(limited$engineDiagnostics$profile$counters$traceFallbackSolves, 0)
expectSameFit(selected, limited)
expect_equal(limited$InfMat, selected$InfMat, tolerance=1e-6)
expect_equal(limited$dLu, selected$dLu, tolerance=1e-6)
}
}
})
test_that("augmented RR grids retain exact lookup fallbacks at higher ranks", {
set.seed(19)
dat <- expand.grid(trial=factor(letters[1:4]), range=factor(1:4), row=factor(1:5))
dat$geno <- factor(sample(sprintf("g%02d", 1:8), nrow(dat), replace=TRUE))
dat$y <- as.numeric(dat$trial) + rnorm(nrow(dat))
dat <- dat[-c(6L, 17L, 34L, 49L, 62L), ]
precision <- diag(8) + matrix(0.1, 8, 8)
dimnames(precision) <- rep(list(levels(dat$geno)), 2)
attr(precision, "inverse") <- TRUE
for(rank in 2:3){
shape <- rrm(dat$trial, rank)
fit <- function(mode){
sommer:::.mmes_dispatch(y ~ trial, random=~vsm(shape, ism(geno), Gu=precision) +
vsm(dsm(trial), ism(geno), Gu=precision),
rcov=~dsumm(vsm(ar1m(range), ar1m(row), ism(units)), by=trial), data=dat,
control=sommer:::.mmes_control_internal(representation=list(gridAugment=mode),
solve=list(method="cholmod", blockSchur="off"),
convergence=list(maxIters=4),
diagnostics=list(verbose=FALSE, dateWarning=FALSE, profile="force")))
}
reference <- suppressWarnings(fit("off"))
selected <- suppressWarnings(fit("auto"))
expect_true(selected$engineDiagnostics$gridAugmentation$active)
expect_gt(selected$engineDiagnostics$directResidualTraces, 0)
expect_equal(selected$engineDiagnostics$CsymbolicAnalyses, 1)
expect_gt(selected$engineDiagnostics$profile$counters$patternInverseCacheFallbacks, 0)
expectSameFit(reference, selected, tolerance=1e-5)
expect_equal(selected$InfMat, reference$InfMat, tolerance=1e-6)
expect_equal(selected$dLu, reference$dLu, tolerance=1e-6)
}
})
Any scripts or data that you put into this service are public.
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.