Speed issue with SingleR v2.8.0 in BioC 3.20 #292
Comments
|
OK, so I decided to divide the I am using the same As you can see below, the new v2.8.0 is faster in this step. The I'll dig a little into it and see what I can find. library(BiocParallel)
library(DelayedArray)
library(SingleR)
bpen <- celldex::BlueprintEncodeData()
sce <- readRDS("sce.rds")For R 4.3, v2.4.1prep.input <- function (test, ref,
assay.type.test = "logcounts", assay.type.ref = "logcounts",
check.missing = TRUE, num.threads = bpnworkers(BPPARAM),
BNPARAM = NULL, BPPARAM = SerialParam()) {
if (!bpisup(BPPARAM) && !is(BPPARAM, "MulticoreParam")) {
bpstart(BPPARAM)
on.exit(bpstop(BPPARAM))
}
test <- SingleR:::.to_clean_matrix(test, assay.type.test, check.missing,
msg = "test", BPPARAM = BPPARAM)
if (single.ref <- !SingleR:::.is_list(ref)) {
ref <- list(ref)
}
ref <- lapply(ref, FUN = SingleR:::.to_clean_matrix, assay.type = assay.type.ref,
check.missing = check.missing, msg = "ref", BPPARAM = BPPARAM)
refnames <- Reduce(intersect, lapply(ref, rownames))
keep <- intersect(rownames(test), refnames)
if (length(keep) == 0) {
stop("no common genes between 'test' and 'ref'")
}
if (!identical(keep, rownames(test))) {
test <- test[keep, ]
}
for (i in seq_along(ref)) {
if (!identical(keep, rownames(ref[[i]]))) {
ref[[i]] <- ref[[i]][keep, , drop = FALSE]
}
}
if (single.ref) {
ref <- ref[[1]]
}
return(list(ref, test))
}system.time({
x <- prep.input(test = sce, ref = list(BP = bpen), assay.type.test = "logcounts", num.threads = 2)
})system.time({
trained <- trainSingleR(x[[1]], labels = list(bpen$label.main), genes = "de", sd.thresh = 1,
de.method = "classic", de.n = NULL,, de.args = list(),
aggr.ref = FALSE, aggr.args = list(), recompute = TRUE,
restrict = NULL, check.missing = TRUE,
BNPARAM = NULL, num.threads = 2, BPPARAM = SerialParam())
})system.time({
class <- classifySingleR(x[[2]], trained, quantile = 0.8, fine.tune = TRUE,
tune.thresh = 0.05, prune = TRUE, check.missing = FALSE,
num.threads = 2, BPPARAM = SerialParam())
})For R 4.4, v2.8.0prep.input <- function (test, ref,
assay.type.test = "logcounts", assay.type.ref = "logcounts",
check.missing = TRUE, num.threads = bpnworkers(BPPARAM),
BNPARAM = NULL, BPPARAM = SerialParam()) {
if (!bpisup(BPPARAM) && !is(BPPARAM, "MulticoreParam")) {
bpstart(BPPARAM)
on.exit(bpstop(BPPARAM))
}
test <- SingleR:::.to_clean_matrix(test, assay.type.test, check.missing,
msg = "test", BPPARAM = BPPARAM)
tmp.ref <- ref
if (!is.list(tmp.ref) || is.data.frame(tmp.ref)) {
tmp.ref <- list(ref)
}
for (rr in tmp.ref) {
keep <- rownames(test) %in% rownames(rr)
if (!all(keep)) {
test <- DelayedArray(test)[keep, , drop = FALSE]
}
}
if (nrow(test) == 0) {
stop("no common genes between 'test' and 'ref")
}
return(list(ref, test))
}system.time({
x <- prep.input(test = sce, ref = list(BP = bpen), assay.type.test = "logcounts", num.threads = 2)
})system.time({
trained <- trainSingleR(x[[1]], labels = list(bpen$label.main), genes = "de", sd.thresh = 1,
de.method = "classic", de.n = NULL,, de.args = list(),
aggr.ref = FALSE, aggr.args = list(), recompute = TRUE,
restrict = NULL, test.genes = rownames(x[[2]]), check.missing = TRUE,
BNPARAM = NULL, num.threads = 2, BPPARAM = SerialParam())
})system.time({
class <- classifySingleR(x[[2]], trained, quantile = 0.8, fine.tune = TRUE,
tune.thresh = 0.05, prune = TRUE, check.missing = FALSE,
num.threads = 2, BPPARAM = SerialParam())
}) |
|
OK, I've looked into the |
|
For me: R 4.4, BioC 3.20library(SingleR)
library(celldex)
bp <- BlueprintEncodeData()
# Using Baron as it's human and it is large enough for some decent testing (~8k cells).
library(scRNAseq)
sce <- BaronPancreasData('human')
system.time(res <- SingleR(sce, bp, labels=bp$label.main, assay.type.test=1))
## user system elapsed
## 29.179 0.081 29.259
system.time(res <- SingleR(sce, bp, labels=bp$label.main, assay.type.test=1, num.threads=2))
## user system elapsed
## 30.008 0.141 15.756
system.time(res <- SingleR(sce, bp, labels=bp$label.main, assay.type.test=1, num.threads=4))
## user system elapsed
## 30.981 0.051 8.795Session informationR 4.3, BioC 3.18# Omitting the identical setup code here...
system.time(res <- SingleR(sce, bp, labels=bp$label.main, assay.type.test=1))
## user system elapsed
## 37.935 0.156 38.186
system.time(res <- SingleR(sce, bp, labels=bp$label.main, assay.type.test=1, num.threads=2))
## user system elapsed
## 38.296 0.135 19.808
system.time(res <- SingleR(sce, bp, labels=bp$label.main, assay.type.test=1, num.threads=4))
## user system elapsed
## 39.354 0.120 11.058Session informationAs we can see, SingleR 2.8.0 is a little bit faster than 2.4.1 on my machine (Mac M2). If you can't share your dataset, try to reproduce your timings with a dataset in scRNAseq. If you're still getting worse timings for 2.8.0 on Another observation is that, for me, SingleR scales nicely as we double the number of threads from 1 to 2, and then 2 to 4. If you're not observing this, it suggests that there is a serial processing bottleneck - specifically related to your test dataset, given that I can say that the slight differences in results are expected and are discussed in the NEWS for 2.8.0, see Lines 34 to 38 in c2922b3 |
|
Hi @LTLA Thanks for your explanation. I was indeed using an object that has its assays stored in Here I am using library(SingleR)
library(scran)
library(scuttle)
library(TENxPBMCData)
sce <- TENxPBMCData(dataset = "pbmc33k")
rownames(sce) <- uniquifyFeatureNames(rowData(sce)$ENSEMBL_ID, rowData(sce)$Symbol_TENx)
colnames(sce) <- sce$Barcode
set.seed(1000)
clusters <- quickCluster(sce)
sce <- computeSumFactors(sce, cluster=clusters)
sce <- logNormCounts(sce)
bp <- celldex::BlueprintEncodeData()R 4.4, BioC 3.20system.time(res <- SingleR(sce, bp, labels=bp$label.main, assay.type.test='logcounts'))
## user system elapsed
##126.988 0.317 127.297
system.time(res <- SingleR(sce, bp, labels=bp$label.main, assay.type.test='logcounts', num.threads=2))
## user system elapsed
##169.576 0.241 107.422
system.time(res <- SingleR(sce, bp, labels=bp$label.main, assay.type.test='logcounts', num.threads=4))
## user system elapsed
##169.480 0.185 79.117 R 4.3, BioC 3.18system.time(res <- SingleR(sce, bp, labels=bp$label.main, assay.type.test='logcounts'))
## user system elapsed
##158.698 0.121 158.803
system.time(res <- SingleR(sce, bp, labels=bp$label.main, assay.type.test='logcounts', num.threads=2))
## user system elapsed
##170.560 0.148 101.268
system.time(res <- SingleR(sce, bp, labels=bp$label.main, assay.type.test='logcounts', num.threads=4))
## user system elapsed
##181.603 0.191 72.212 I am now tweaking my workflow to create |
Hi,
I am testing my single-cell workflow using the new R-4.4 (v4.4.2) these couple of days and noticed SingleR (v2.8.0) is running a lot slower compared with my other environment that has R-4.3 and SingleR v2.4.1.
From the commit history, I don't see what might have caused the drop in speed between the two versions.
I am wondering if anyone has any insight into this and how I can run the new version of
SingleRas fast as before (is there any parameter I can tweak)?Below is how I run SingleR and my
scehas 9,188 genes and 16,601 cells.With R-4.3 and SingleR v2.4.1, I have
and with R-4.4 and SingleR v2.8.0, I have
When I change num.threads to
4(no speed gain in both, new version still runs slower):With R-4.3 and SingleR v2.4.1, I have
and with R-4.4 and SingleR v2.8.0, I have
p.s. The predicted cell type labels are a little different, but I am guessing its due to updates with the
celldexpackage. Although, I did notice theassaysstructure of the reference I am using (BlueprintEncodeData) is a little different.This is
celldex_1.12.0(loaded and attached):And this is
celldex_1.16.0(loaded, not attached):R v4.3.3 and v4.4.2 session info
The text was updated successfully, but these errors were encountered: