## library(devtools) ## setwd("tests/testthat") ## load_all() context("NMscanData") library(data.table) data.table::setDTthreads(1) ## NMdata_filepath <- function(...) { ## system.file(..., package = "NMdata") ## } ## file.nm <- function(...) NMdata_filepath("examples/nonmem",...) ## file.data <- function(...) NMdata_filepath("examples/data",...) fix.time <- function(x){ meta.x <- attr(x,"NMdata") ## meta.x$time.call <- as.POSIXct("2020-02-01 00:01:01",tz="UTC") meta.x$details$time.NMscanData <- NULL meta.x$details$file.lst <- NULL meta.x$details$file.mod <- NULL meta.x$details$file.input <- NULL meta.x$details$mtime.input <- NULL meta.x$details$mtime.lst <- NULL meta.x$details$mtime.mod <- NULL meta.x$datafile$path.csv <- NULL meta.x$datafile$path.rds <- NULL meta.x$datafile$path.fst <- NULL meta.x$tables$file <- NULL meta.x$tables$file.mtime <- NULL setattr(x,"NMdata",meta.x) invisible(x) } NMdataConf(reset=TRUE) test_that("basic",{ fileRef <- "testReference/NMscanData1.rds" resRef <- if(file.exists(fileRef)) readRDS(fileRef) else NULL ## file.lst <- "../../inst/examples/nonmem/run001.lst" ## file.lst <- NMdata_filepath("examples/nonmem/xgxr001.lst") ## file.lst <- system.file("examples/nonmem/xgxr001.lst" ,package="NMdata") file.lst <- "testData/nonmem/xgxr001.lst" ## NMreadSection(NMdata_filepath("examples/nonmem/run001.lst"),section="DATA") res <- NMscanData(file=file.lst, quiet=T, order.columns = F, merge.by.row=FALSE, check.time = FALSE) ## res2 <- NMscanData(file=file.lst, quiet=T, order.columns = F, merge.by.row=FALSE, check.time = FALSE,rep.count=T) ## dim(res) fix.time(res) expect_equal_to_reference(res,fileRef,version=2) ## without meta ## expect_equal(unNMdata(res),unNMdata(readRDS(fileRef))) ## data.table(attributes(readRDS(fileRef))$meta$variables$variable,attributes(res1)$meta$variables$variable) }) test_that("Modifications to column names in $INPUT",{ fileRef <- "testReference/NMscanData2.rds" ## res.ref <- readRDS(fileRef) ## file.lst <- NMdata_filepath("examples/nonmem/xgxr002.lst") file.lst <- "testData/nonmem/xgxr002.lst" res1 <- NMscanData(file=file.lst, check.time = FALSE, merge.by.row=FALSE, quiet=T) res <- list( NMinfo(res1,"input.colnames"), NMinfo(res1,"columns"), colnames(res1) ) expect_equal_to_reference(res,fileRef,version=2) ## without meta ## expect_equal(unNMdata(res),unNMdata(readRDS(fileRef))) }) test_that("No translation of column names in $INPUT",{ fileRef <- "testReference/NMscanData2b.rds" ## res.ref <- readRDS(fileRef) file.lst <- "testData/nonmem/xgxr002.lst" res1 <- NMscanData(file=file.lst, check.time = FALSE, merge.by.row=FALSE, translate.input = T, quiet=T) res2 <- NMscanData(file=file.lst, check.time = FALSE, merge.by.row=FALSE, translate.input = F, quiet=T) dt.cnames <- data.table(colnames(res1),colnames(res2)) expect_equal_to_reference(dt.cnames,fileRef,version=2) }) test_that("Multiple output table formats",{ NMdataConf(reset=TRUE) fileRef <- "testReference/NMscanData_03.rds" ## file.lst <- NMdata_filepath("examples/nonmem/xgxr003.lst") file.lst <- "testData/nonmem/xgxr003.lst" ## res <- NMscanData(file=file.lst) res <- NMscanData(file=file.lst, check.time = FALSE, merge.by.row=FALSE, quiet=T) fix.time(res) expect_equal_to_reference(res,fileRef,version=2) ## without meta ## expect_equal(unNMdata(res),unNMdata(readRDS(fileRef))) ## only meta ## expect_equal(NMinfo(res),NMinfo(readRDS(fileRef))) }) test_that("Interpret IGNORE statement",{ NMdataConf(reset=TRUE) fileRef <- "testReference/NMscanData4.rds" ## file.lst <- NMdata_filepath("examples/nonmem/xgxr004.lst") file.lst <- "testData/nonmem/xgxr004.lst" ## res <- NMscanData(file=file.lst) ## res <- NMscanData(file=file.lst) res <- NMscanData(file=file.lst,merge.by.row=FALSE,check.time = FALSE, quiet=T) fix.time(res) ## names(res$row) expect_equal_to_reference(res,fileRef,version=2) ## without meta ## expect_equal(unNMdata(res),unNMdata(readRDS(fileRef))) }) test_that("List of ACCEPT statements and vs separate statements",{ NMdataConf(reset=T) NMdataConf(as.fun="data.table") ## file1.lst <- NMdata_filepath("examples/nonmem/xgxr006.lst") ## file2.lst <- NMdata_filepath("examples/nonmem/xgxr007.lst") file1.lst <- "testData/nonmem/xgxr006.lst" file2.lst <- "testData/nonmem/xgxr007.lst" NMreadSection(file1.lst,section="PROBLEM") NMreadSection(file2.lst,section="PROBLEM") res1 <- NMscanData(file=file1.lst,merge.by.row=FALSE,col.model=NULL,check.time = FALSE, quiet=T) res2 <- NMscanData(file=file2.lst,merge.by.row=FALSE,col.model=NULL,check.time = FALSE, quiet=T) setattr(res1,"NMdata",NULL) setattr(res2,"NMdata",NULL) expect_identical(res1,res2) }) test_that("merge by filters or not",{ ## fileRef <- "testReference/NMscanData4.rds" ## file1.lst <- NMdata_filepath("examples/nonmem/xgxr006.lst") ## file2.lst <- NMdata_filepath("examples/nonmem/xgxr008.lst") file1.lst <- "testData/nonmem/xgxr006.lst" file2.lst <- "testData/nonmem/xgxr008.lst" ## NMreadSection(file1.lst,section="PROBLEM") ## NMreadSection(file2.lst,section="PROBLEM") res1 <- NMscanData(file=file1.lst,merge.by.row=FALSE,col.model=NULL,check.time = FALSE, quiet=T) res2 <- NMscanData(file=file2.lst,merge.by.row=FALSE,col.model=NULL,check.time = FALSE, quiet=T) setnames(res2,"EFF0","eff0",skip_absent=T) setcolorder(res1,colnames(res2)) ## the var tables are different because ROW is input in one, ## output in the other. This is as expected. ## cbind(attr(res1,"var"),attr(res2,"var")) setattr(res1,"NMdata",NULL) setattr(res2,"NMdata",NULL) expect_equal(res1,res2) }) ### BUG 011. NA rows used for fo table. 0 IDs. #### it is the rare situation where ##### there is only first-only output (nmout=FALSE everywhere) ##### col.id is not in first-only table ##### nmout does not exactly represent the split we need here. We need ##### to know which rows were enriched by output. test_that("Only a firstonly without ID but with ROW",{ ### This should work because ROW is in firstonly table. NMdataConf(reset=TRUE) ## NMdataConf(as.fun="data.table") fileRef <- "testReference/NMscanData_11.rds" ## file.lst <- NMdata_filepath("examples/nonmem/xgxr011.lst") file.lst <- "testData/nonmem/xgxr011.lst" ## NMreadSection(NMdata_filepath("examples/nonmem/run001.lst"),section="DATA") ### notice that DV PRED RES WRES are returned in firstonly. This is horrible. ## tabs <- NMscanTables(file.lst) ## tabs ### impossible with filters at the moment. Cannot be implemented because id level cannot be merged onto row level input. expect_error( expect_warning( NMscanData(file=file.lst,merge.by.row=FALSE,col.row="ROW",check.time = FALSE, quiet=T) ) ) ## merge.by.row should be able to merge. res <- expect_warning( NMscanData(file=file.lst,merge.by.row=TRUE,col.row="ROW",check.time = FALSE, quiet=T) ) fix.time(res) expect_equal_to_reference(res,fileRef,version=2) ## without meta ## expect_equal(unNMdata(res),unNMdata(readRDS(fileRef))) ## ref <- readRDS(fileRef) }) test_that("Only a firstonly, no ID, no ROW",{ ### use.input= TRUE, cbind.by.filters=FALSE. ## This should give an error because input cannot be used even though it is requested ## this one only outputs a firstonly that cannot be merged onto ## input. use.input=T so input data should be returned. fileRef <- "testReference/NMscanData12.rds" ## file.lst <- NMdata_filepath("examples/nonmem/xgxr012.lst") file.lst <- "testData/nonmem/xgxr012.lst" ## NMreadSection(file.lst,section="DATA") ## NMreadSection(file.lst,section="PROBLEM") ## NMreadSection(file.lst,section="TABLE") expect_error( expect_warning( res1 <- NMscanData(file=file.lst,check.time = FALSE, quiet=T) ) ) }) test_that("FO and row-level output. No ID, no row.",{ #### merge.by.row = "ifAvailable" or FALSE ## Only row-level output returned because cbind.by.filters=F, and firstonly is without ID and row. Warning that firstonly is dropped. Correct. fileRef <- "testReference/NMscanData_13.rds" ## file.lst <- NMdata_filepath("examples/nonmem/xgxr013.lst") file.lst <- "testData/nonmem/xgxr013.lst" NMreadSection(file.lst,section="PROBLEM") ## NMreadSection(NMdata_filepath("examples/nonmem/run001.lst"),section="DATA") NMreadSection(file.lst,section="TABLE") ## tabs <- NMscanTables(file=file.lst) res1 <- expect_warning( NMscanData(file=file.lst,check.time = FALSE, quiet=T) ) fix.time(res1) expect_equal_to_reference( res1,fileRef,version=2 ) ## merge.by.row=F cannot combine and returns output. OK res2 <- expect_warning( NMscanData(file=file.lst,check.time = FALSE,merge.by.row=T, quiet=T) ) ## if we leave use.input out and can only get row level. OK res2 <- expect_warning( NMscanData(file=file.lst,check.time = FALSE,use.input=F, quiet=T) ) }) ### uses ACCEPT: DOSE.GT.20 explining only 731 rows test_that("FO and row-level output. No ID, no row. cbind.by.filters=T",{ ## row-level output+input returned because cbind.by.filters=T, and firstonly is without ID and row. Correct. fileRef <- "testReference/NMscanData_14.rds" ## file.lst <- NMdata_filepath("examples/nonmem/xgxr013.lst") file.lst <- "testData/nonmem/xgxr013.lst" NMreadSection(file.lst,section="PROBLEM") ## tabs <- NMscanTables(file=file.lst) res1 <- expect_warning( NMscanData(file=file.lst,merge.by.row=FALSE,check.time = FALSE, quiet=T) ) fix.time(res1) summary(res1)$variables summary(res1)$tables summary(res1) expect_equal_to_reference( res1,fileRef,version=2 ) }) ## bug. This should return a meaningful error msg ## "bug" confirmed after release of 0.0.16. test_that("Only a firstonly without ID but with ROW",{ ### cbind.by.filters is TRUE, so ROW is not used to recover firstonly data. fileRef <- "testReference/NMscanData15.rds" ## file.lst <- NMdata_filepath("examples/nonmem/xgxr011.lst") file.lst <- "testData/nonmem/xgxr011.lst" NMreadSection(file.lst,section="DATA") NMreadSection(file.lst,section="TABLE") ### notice that DV PRED RES WRES are returned in firstonly. This is horrible. ## tabs <- NMscanTables(file.lst) ## tabs res1 <- expect_error( #expect_warning( NMscanData(file=file.lst,merge.by.row=FALSE,check.time = FALSE, quiet=T) ## ) ) ## tabs=NMscanTables(file=file.lst) ## tabs }) test_that("Only a firstonly without ID but with ROW. Using merge.by.row=TRUE.",{ ### ROW is used to recover firstonly data. ## NMdataConf(as.fun="data.table") ## NMdataConf(as.fun=NULL) NMdataConf(reset=TRUE) fileRef <- "testReference/NMscanData_15b.rds" ## file.lst <- NMdata_filepath("examples/nonmem/xgxr011.lst") file.lst <- "testData/nonmem/xgxr011.lst" NMreadSection(file.lst,section="DATA") NMreadSection(file.lst,section="TABLE") ### notice that DV PRED RES WRES are returned in firstonly. This is horrible. ## tabs <- NMscanTables(file.lst) ## tabs res1 <- expect_warning( NMscanData(file=file.lst,col.row="ROW",merge.by.row=TRUE,check.time = FALSE, quiet=T) ) fix.time(res1) expect_equal_to_reference( res1,fileRef,version=2 ) }) ### recoverRows without a row identifier test_that("recoverRows without a row identifier",{ fileRef <- "testReference/NMscanData_16.rds" ## file.lst <- NMdata_filepath("examples/nonmem/xgxr004.lst") file.lst <- "testData/nonmem/xgxr004.lst" NMreadSection(file.lst,section="DATA") NMreadSection(file.lst,section="TABLE") ### notice that DV PRED RES WRES are returned in firstonly. ## tabs <- NMscanTables(file.lst) ## tabs res1 <- NMscanData(file=file.lst,merge.by.row=FALSE,recover.rows = T,as.fun="data.table",check.time = FALSE, quiet=T) dim(res1) res1[,table(nmout,DOSE)] fix.time(res1) expect_equal_to_reference( res1,fileRef,version=2 ) }) ### get a data.frame test_that("use as.fun to get a data.frame",{ ### cbind.by.filters is TRUE, so ROW is used to recover firstonly data. fileRef <- "testReference/NMscanData_17.rds" file.lst <- "testData/nonmem/xgxr004.lst" NMreadSection(file.lst,section="DATA") NMreadSection(file.lst,section="TABLE") res1 <- NMscanData(file=file.lst,merge.by.row=FALSE,recover.rows = T,as.fun=as.data.frame,check.time = FALSE, quiet=T) dim(res1) class(res1) with(res1,table(nmout,DOSE)) fix.time(res1) expect_equal_to_reference( res1,fileRef,version=2 ) #### explore ## ref <- readRDS(fileRef) ## compareCols(res1,ref) ## setorder(ref,ROW) ## expect_equal( ## res1,ref ## ) ## ref <- readRDS(fileRef) ## names(attributes(res1)) ## names(attributes(ref)) }) ### get a tibble test_that("use as.fun to get a tibble",{ ### cbind.by.filters is TRUE, so ROW is used to recover firstonly data. fileRef <- "testReference/NMscanData_18.rds" ## file.lst <- NMdata_filepath("examples/nonmem/xgxr004.lst") file.lst <- "testData/nonmem/xgxr004.lst" NMreadSection(file.lst,section="DATA") NMreadSection(file.lst,section="TABLE") ### notice that DV PRED RES WRES are returned in firstonly. ## tabs <- NMscanTables(file.lst) ## tabs res1 <- NMscanData(file=file.lst,merge.by.row=FALSE,recover.rows = T,as.fun=tibble::as_tibble,check.time = FALSE, quiet=T) dim(res1) class(res1) fix.time(res1) expect_equal_to_reference( res1,fileRef,version=2 ) ###### explore diffs ## ref <- readRDS(fileRef) ## compareCols(res1,ref) ## setorder(ref,ROW) ## expect_equal( ## unNMdata(res1),unNMdata(ref) ## ) ## ref <- readRDS(fileRef) ## names(attributes(res1)) ## names(attributes(ref)) ## class(res1) ## class(ref) }) ## test the dir structure with input.txt/output.txt test_that("dir structure with input.txt/output.txt",{ ## options(NMdata.as.fun="none") ## options(NMdata.file.mod=function(file) file.path(dirname(file),"input.txt")) ## options(NMdata.modelname=function(file) basename(dirname(normalizePath(file)))) NMdataConf(reset=TRUE) NMdataConf(as.fun="data.table") ## NMdataConf(file.mod=function(file) file.path(dirname(file),"input.txt")) NMdataConf(file.mod=identity) NMdataConf(modelname=function(file) basename(dirname(normalizePath(file)))) ## filedir.lst <- file.nm("xgxr001dir/output.txt") filedir.lst <- "testData/nonmem/xgxr001dir/output.txt" res1dir <- NMscanData(filedir.lst,check.time = FALSE,merge.by.row=F, quiet=T) ## test model names from NMinfo and in result data expect_equal(NMinfo(res1dir,"details")$model,"xgxr001dir") umod <- unique(res1dir[,model]) expect_equal(length(umod),1) expect_equal(umod,"xgxr001dir") }) #### compare mod/lst vs input/output ## this test isn't ready. ### Todo: Delete components from NMinfo and then test equality test_that("input.txt/output.txt - unset modelname",{ NMdataConf(reset=TRUE) NMdataConf(as.fun="data.table") ## NMdataConf(file.mod=function(file) file.path(dirname(file),"input.txt")) NMdataConf(file.mod=identity) NMdataConf(modelname=function(file) basename(dirname(normalizePath(file)))) ## filedir.lst <- file.nm("xgxr001dir/output.txt") filedir.lst <- "testData/nonmem/xgxr001dir/output.txt" res1dir <- NMscanData(filedir.lst,check.time = FALSE,merge.by.row=T, quiet=T) NMdataConf(file.mod="default") NMdataConf(modelname=NULL) ## file.lst <- NMdata_filepath("examples/nonmem/xgxr001.lst") file.lst <- "testData/nonmem/xgxr001.lst" res1 <- NMscanData(file=file.lst,check.time = FALSE, quiet=T) ## unNMdata(res1) ## unNMdata(res1dir) ## this test isn't ready. How do I execute the input.txt/output.txt model? els.details <- c("call","model","file.lst","file.mod") for(elem in els.details){ attributes(res1)$NMdata$details[elem] <- NULL attributes(res1dir)$NMdata$details[elem] <- NULL } attributes(res1)$NMdata$datafile$DATA <- NULL attributes(res1dir)$NMdata$datafile$DATA <- NULL attributes(res1)$NMdata$datafile$string <- NULL attributes(res1dir)$NMdata$datafile$string <- NULL fix.time(res1) meta.res1 <- NMinfo(res1) meta.res1$details$logtime.lst <- NULL setattr(res1,"NMdata",meta.res1) fix.time(res1dir) meta.res1dir <- NMinfo(res1dir) meta.res1dir$details$logtime.lst <- NULL setattr(res1dir,"NMdata",meta.res1dir) ## these differ a little in the two estimates cols.differ <- c("TVKA","TVCL","TVV3","TVQ","KA","CL","V3","Q","V2","IPRED","PRED","RES","WRES") res1[,(cols.differ):=NULL] res1dir[,(cols.differ):=NULL] expect_equal(res1[,!("model")],res1dir[,!("model")]) NMdataConf(as.fun=NULL) }) test_that("output.txt, file.mod=identity - NMinfo file.mod=output.txt?",{ NMdataConf(reset=TRUE) NMdataConf(as.fun="data.table") NMdataConf(file.mod=identity) NMdataConf(modelname=function(file) basename(dirname(normalizePath(file)))) ## filedir.lst <- file.nm("xgxr001dir/output.txt") filedir.lst <- "testData/nonmem/xgxr001dir/output.txt" res1 <- NMscanData(filedir.lst,check.time = FALSE,merge.by.row=F, quiet=T) expect_equal(basename(NMinfo(res1,"details")$file.lst),"output.txt") }) ## test_that("Duplicate columns in input data",{ NMdataConf(reset=TRUE) fileRef <- "testReference/NMscanData_20.rds" ## file.lst <- NMdata_filepath("examples/nonmem/xgxr015.lst") file.lst <- "testData/nonmem/xgxr015.lst" ## debugonce(NMscanData) res <- expect_warning( NMscanData(file=file.lst,merge.by.row=FALSE,check.time = FALSE, quiet=T) ) fix.time(res) ## names(res$row) expect_equal_to_reference(res,fileRef,version=2) }) ### this is not a real test. Need to be able to test how the merges were performed. ## test_that("col.row and merge.by.row=TRUE from NMdataConf",{ ## NMdataConf(reset=TRUE) ## NMdataConf(col.row="ROW", merge.by.row=TRUE) ## NMdataConf() ## file.lst <- system.file("examples/nonmem/xgxr001.lst" ,package="NMdata") ## res1 <- NMscanData(file=file.lst) ## }) test_that("Modifying row identifier",{ NMdataConf(reset=TRUE) ## file.lst <- NMdata_filepath("examples/nonmem/xgxr016.lst") file.lst <- "testData/nonmem/xgxr016.lst" res <- expect_error( NMscanData(file=file.lst,merge.by.row=TRUE,check.time = FALSE, quiet=T) ) }) test_that("merge.by.row=ifAvailable when available",{ NMdataConf(reset=TRUE) fileRef <- "testReference/NMscanData_21.rds" file.lst <- system.file("examples/nonmem/xgxr001.lst" ,package="NMdata") res1 <- NMscanData(file=file.lst,merge.by.row="ifAvailable",check.time=FALSE, quiet=T) ## dim(res1) ### we don't need metadata for this test unNMdata(res1) expect_equal_to_reference(res1,fileRef,version=2) }) test_that("merge.by.row=ifAvailable when not available",{ fileRef <- "testReference/NMscanData_22.rds" ## file.lst <- NMdata_filepath("examples/nonmem/xgxr004.lst") file.lst <- "testData/nonmem/xgxr004.lst" res1 <- NMscanData(file=file.lst,merge.by.row="ifAvailable",recover.rows = T,as.fun="data.table",check.time = FALSE, quiet=T) dim(res1) res1[,table(nmout,DOSE)] fix.time(res1) expect_equal_to_reference( res1,fileRef,version=2 ) ##### explore diffs ## ref <- readRDS(fileRef) ## compareCols(res1,ref) ## setorder(ref,ROW) ## expect_equal( ## unNMdata(res1),unNMdata(ref) ## ) }) test_that("col.row does not exist, but merge.by.row==TRUE",{ ### col.row does not exist, but merge.by.row==TRUE fileRef <- "testReference/NMscanData_22b.rds" NMdataConf(reset=T) NMdataConf(check.time=F) file.lst <- system.file("examples/nonmem/xgxr001.lst" ,package="NMdata") ## warning becauses the merge was not possible even though merge.by.row res1 <- expect_warning( NMscanData(file=file.lst,col.row="NONEXIST",merge.by.row=TRUE, quiet=T) ) fix.time(res1) expect_equal_to_reference(res1,fileRef) ##### explore diffs ## ref <- readRDS(fileRef) ## indices( attributes(res1)$NMdata$tables) ## indices(attributes(ref)$NMdata$tables) }) test_that("col.row is NULL, but merge.by.row==TRUE",{ ### col.row is NULL, but merge.by.row==TRUE file.lst <- system.file("examples/nonmem/xgxr001.lst" ,package="NMdata") ## error is that it's not in output. It's not in input either though expect_error( res1=NMscanData(file=file.lst,col.row=NULL,merge.by.row=TRUE, quiet=T) ) }) test_that("A filter without operator",{ fileRef <- "testReference/NMscanData_23.rds" file.lst <- "testData/nonmem/xgxr010.lst" res1 <- NMscanData(file=file.lst,merge.by.row="ifAvailable",as.fun="data.table",check.time = FALSE, quiet=T) res1[,.N,by=.(DOSE)] fix.time(res1) expect_equal_to_reference( res1,fileRef,version=2 ) }) test_that("Including a redundant output table",{ NMdataConf(reset=T) NMdataConf(as.fun="data.table") fileRef <- "testReference/NMscanData_24.rds" ## file.lst <- file.nm("xgxr019.lst") file.lst <- "testData/nonmem/xgxr019.lst" ## notice no cols are taken from the redundant table - correct res1 <- NMscanData(file=file.lst,merge.by.row="ifAvailable",as.fun="data.table",check.time = FALSE, quiet=T) ## tabs1 <- NMscanTables(file=file.lst,as.fun="data.table",details=T,rep.count=F) ## tabs1$meta ## tabs1$data[[4]] ## NMinfo(res1,"tables") fix.time(res1) expect_equal_to_reference( res1,fileRef,version=2 ) ## inp1 <- NMscanInput(file=file.lst) } ) ## this one has a commented DATA section and an extra DV in output table test_that("redundant output",{ NMdataConf(reset=T) NMdataConf(as.fun="data.table") NMdataConf(file.mod=function(x)sub("\\.lst$",".ctl",x)) NMdataConf(check.time=FALSE) fileRef <- "testReference/NMscanData_25.rds" file.lst <- "testData/nonmem/estim_debug.ctl" ## notice no cols are taken from the redundant table - correct res1 <- expect_message( NMscanData(file=file.lst, quiet=T) ) ## tabs1 <- NMscanTables(file=file.lst,as.fun="data.table",details=T,rep.count=F) ## tabs1$meta ## tabs1$data[[4]] ## NMinfo(res1,"tables") fix.time(res1) expect_equal_to_reference( res1,fileRef,version=2 ) ## inp1 <- NMscanInput(file=file.lst) } ) ## check.time (warning) ### I'm afraid this could give warnings when run in check or on CRAN. if(FALSE){ test_that("check time warning",{ NMdataConf(reset=T) fileRef <- "testReference/NMscanData26.rds" ## file.lst <- "testData/nonmem/xgxr001.lst" dir.test <- "testData/nonmem/check.time" if(dir.exists(dir.test)) unlink(dir.test,recursive=TRUE) dir.create(dir.test) file.copy("testData/nonmem/xgxr001.lst",dir.test) file.copy("testData/nonmem/xgxr001_res.txt",dir.test) file.copy("testData/nonmem/xgxr001.mod",dir.test) file.lst <- file.path(dir.test,"xgxr001.lst") file.mod <- fnExtension(file.lst,".mod") data.new <- sub("../data","../../data",NMreadSection(file.mod,section="data")) NMwriteSection(file.mod,section="data",newlines=data.new) res1 <- expect_warning( NMscanData(file=file.lst, quiet=F, order.columns = F, merge.by.row=FALSE, quiet=T) ) ## dim(res1)' ## without meta expect_equal_to_reference(unNMdata(res1),fileRef) ## data.table(attributes(readRDS(fileRef))$meta$variables$variable,attributes(res1)$meta$variables$variable) }) } test_that("$INPUT copy",{ NMdataConf(reset=T) NMdataConf(check.time=FALSE) file.lst.1 <- "testData/nonmem/xgxr022.lst" res.1 <- NMscanData(file.lst.1, quiet=T) ## NMinfo(res,"input.colname") NMinfo(res.1,"input.colnames") NMinfo(res.1,"columns") file.lst.2 <- "testData/nonmem/xgxr001.lst" res.2 <- NMscanData(file.lst.2, quiet=T) expect_equal(ncol(res.1)-ncol(res.2),1) expect_equal(setdiff(colnames(res.1),colnames(res.2)),c("COMP","EFF0")) cols.1 <- NMinfo(res.1,"columns") cols.2 <- NMinfo(res.2,"columns") expect_equal(setdiff(cols.1$variable,cols.2$variable),c("COMP","EFF0")) expect_equal(setdiff(cols.2$variable,cols.1$variable),c("eff0")) }) test_that("only firstonly. Has col.id, no col.row.",{ ###### trickers this ## use.input&&!any(tables$meta$full.length) NMdataConf(reset=T) NMdataConf(check.time=FALSE) fileRef <- "testReference/NMscanData_27.rds" file.lst <- "testData/nonmem/xgxr023.lst" res <- NMscanData(file.lst,quiet=TRUE) res <- fix.time(res) expect_equal_to_reference(res,fileRef) }) test_that("Two firstonly, one full-length",{ fileRef <- "testReference/NMscanData_28.rds" file.lst <- "testData/nonmem/xgxr025.lst" res <- NMscanData(file=file.lst,check.time=F, quiet=T) res <- fix.time(res) ## ref <- readRDS(fileRef) expect_equal_to_reference(res,fileRef,version=2) }) test_that("Two firstonly, one full-length with rep.count",{ NMdataConf(reset=TRUE) #### TABLENO is now added to the number of columns taken from the #### output table that has TABLENO. Is that what we want? Or +1? +1 is #### very complicated for user. Maybe better: if rep.count, it is treated like any other column, but in NMinfo(,"tables") there is a column, hasTABLENO fileRef <- "testReference/NMscanData_28b.rds" file.lst <- "testData/nonmem/xgxr025.lst" res <- NMscanData(file=file.lst,check.time=F,rep.count=T, quiet=T) res <- fix.time(res) ## ref <- readRDS(fileRef) expect_equal_to_reference(res,fileRef,version=2) }) test_that("Input data as character",{ fileRef <- "testReference/NMscanData29.rds" file.lst <- "testData/nonmem/xgxr026.lst" expect_error(NMscanData(file=file.lst,check.time=F)) res <- NMscanData(file=file.lst,check.time=F,use.rds=F, quiet=T) sapply(res,class) res <- NMscanData(file=file.lst,check.time=F,use.rds=T,merge.by.row = F, quiet=T) sapply(res,class) ## res <- fix.time(res) ## ## ref <- readRDS(fileRef) ## expect_equal_to_reference(res,fileRef,version=2) }) test_that("Input control stream missing",{ NMdataConf(reset=TRUE) NMdataConf(file.mod="doesNotExist.mod") fileRef <- "testReference/NMscanData_30.rds" file.lst <- "testData/nonmem/xgxr001.lst" res <- NMscanData(file=file.lst,use.input=F, quiet=T) ## check file.mod wasn't used. Notice, merge.by.row is still "ifAvailable" - this was never needed because use.input=FALSE. ## NMinfo(res1a,"details") fix.time(res) expect_equal_to_reference(res,fileRef,version=2) }) test_that("simulation model with subproblems",{ NMdataConf(reset=TRUE) fileRef <- "testReference/NMscanData_31.rds" ### sim a model with subproblems ## library(devtools) ## load_all("~/wdirs/NMexec") ## file.mod <- "testData/nonmem/xgxr014.mod" ## ## NMexec(file.mod,sge=FALSE,wait=T) ## doses <- NMcreateDoses(TIME=0,AMT=data.table(AMT=c(10,100),DOSE=c(10,100))) ## simdat <- addEVID2(doses,time.sim=1:24,CMT=2) ## simdat[,DV:=NA][ ## ,ROW:=.I] ## NMcheckData(simdat) ## NMdataConf(as.fun="data.table") ## sim1 <- NMsim(file.mod,data=simdat, ## suffix.sim="testsim1",dir.sim="testData/simulations" ## ,seed=343108,subproblems=100,nmquiet=T) ### NMsim saves serialized rds. Resaving as version 2 ## res.sim <- readRDS("testData/simulations/NMsimData_xgxr014_testsim1.rds") ## saveRDS(res.sim,"testData/simulations/NMsimData_xgxr014_testsim1.rds",version=2) ##### sim done res <- NMscanData("testData/simulations/xgxr014_testsim1.lst",check.time=F, quiet=T) fix.time(res) meta.x <- attr(res,"NMdata") meta.x$details$time.ok <- NULL setattr(res,"NMdata",meta.x) expect_equal_to_reference(res,fileRef,version=2) } ) test_that("csv vs rds vs fst",{ fileRef <- "testReference/NMscanData_32.rds" file.lst <- "testData/nonmem/xgxr014.lst" res.csv <- NMscanData(file=file.lst, order.columns = F, merge.by.row=FALSE, check.time = FALSE,formats.read=cc(csv), quiet=T) res.rds <- NMscanData(file=file.lst, order.columns = F, merge.by.row=FALSE, check.time = FALSE,formats.read=cc(rds,csv), quiet=T) res.fst <- NMscanData(file=file.lst, order.columns = F, merge.by.row=FALSE, check.time = FALSE,formats.read=cc(fst,csv), quiet=T) res <- dims(res.csv,res.rds,res.fst) expect_equal_to_reference(res,fileRef,version=2) })