插入符号训练二进制 glm 通过 doParallel 在并行集群上失败
Posted
技术标签:
【中文标题】插入符号训练二进制 glm 通过 doParallel 在并行集群上失败【英文标题】:caret train binary glm fails on parallel cluster via doParallel 【发布时间】:2018-02-17 12:42:59 【问题描述】:我已经看到有很多关于这个主题的问题,但似乎没有一个对我的问题给出令人满意的答案。我打算在 Windows 机器上将 caret::train()
与库 doParallel
结合使用。文档 (The caret package: 9 Parallel Processing) 告诉我,如果找到已注册的集群,它将默认并行运行(尽管它使用库 doMC
)。当我尝试使用 doParallel
设置集群并按照其文档 (Getting Started with doParallel and foreach) 中的示例计算进行操作时,一切正常。当我取消注册集群并运行caret::train()
时,一切正常。但是当我创建一个新集群并尝试运行caret::train()
时,它会产生错误Error in serialize(data, node$con) : error writing to connection
。我还包括下面的日志。我不明白caret::train()
在非并行模式下是如何工作的,但在并行模式下却不是,尽管集群似乎设置正确。
库
library(caret)
library(microbenchmark)
library(doParallel)
会话信息
sessionInfo()
R version 3.4.1 (2017-06-30)
Platform: x86_64-w64-mingw32/x64 (64-bit)
Running under: Windows 7 x64 (build 7601) Service Pack 1
Matrix products: default
locale:
[1] LC_COLLATE=English_United States.1252 LC_CTYPE=English_United States.1252
[3] LC_MONETARY=English_United States.1252 LC_NUMERIC=C
[5] LC_TIME=English_United States.1252
attached base packages:
[1] parallel stats graphics grDevices utils datasets methods base
other attached packages:
[1] doParallel_1.0.10 iterators_1.0.8 foreach_1.4.3 microbenchmark_1.4-2.1
[5] caret_6.0-76 ggplot2_2.2.1 lattice_0.20-35
loaded via a namespace (and not attached):
[1] Rcpp_0.12.11 compiler_3.4.1 nloptr_1.0.4 plyr_1.8.4 tools_3.4.1
[6] lme4_1.1-13 tibble_1.3.3 nlme_3.1-131 gtable_0.2.0 mgcv_1.8-17
[11] rlang_0.1.1 Matrix_1.2-10 SparseM_1.77 mvtnorm_1.0-6 stringr_1.2.0
[16] hms_0.3 MatrixModels_0.4-1 stats4_3.4.1 grid_3.4.1 nnet_7.3-12
[21] R6_2.2.2 survival_2.41-3 multcomp_1.4-6 TH.data_1.0-8 minqa_1.2.4
[26] readr_1.1.1 reshape2_1.4.2 car_2.1-5 magrittr_1.5 scales_0.4.1
[31] codetools_0.2-15 ModelMetrics_1.1.0 MASS_7.3-47 splines_3.4.1 pbkrtest_0.4-7
[36] colorspace_1.3-2 quantreg_5.33 sandwich_2.4-0 stringi_1.1.5 lazyeval_0.2.0
[41] munsell_0.4.3 zoo_1.8-0
doParallel 文档中的运行示例(无错误)
cores_2_use <- floor(0.8 * detectCores())
cl <- makeCluster(cores_2_use, outfile = "parallel_log1.txt")
registerDoParallel(cl)
x <- iris[which(iris[,5] != "setosa"), c(1,5)]
trials <- 100
temp <- microbenchmark(
r <- foreach(icount(trials), .combine=cbind) %dopar%
ind <- sample(100, 100, replace=TRUE)
result1 <- glm(x[ind,2]~x[ind,1], family=binomial(logit))
coefficients(result1)
)
parallel::stopCluster(cl)
foreach::registerDoSEQ()
模型数据
x1 = rnorm(100) # some continuous variables
x2 = rnorm(100)
z = 1 + 2 * x1 + 3 * x2 # linear combination with a bias
pr = 1 / (1 + exp(-z)) # pass through an inv-logit function
y = rbinom(100, 1, pr) # bernoulli response variable
df = data.frame(y = as.factor(ifelse(y == 0, "no", "yes")), x1 = x1, x2 = x2)
运行 caret::train() 非并行(无错误)
# train control function
ctrl <-
trainControl(
method = "repeatedcv",
number = 10,
repeats = 5,
classProbs = TRUE,
summaryFunction = twoClassSummary)
# train function
microbenchmark(
glm_nopar =
train(y ~ .,
data = df,
method = "glm",
family = "binomial",
metric = "ROC",
trControl = ctrl),
times = 5)
#Unit: milliseconds
#expr min lq mean median uq max neval
#glm_nopar 691.9643 805.1762 977.1054 895.9903 1018.112 1474.284 5
并行运行 caret::train()(错误)
cores_2_use <- floor(0.8 * detectCores())
cl <- makeCluster(cores_2_use, outfile = "parallel_log2.txt")
registerDoParallel(cl)
microbenchmark(
glm_par =
train(y ~ .,
data = df,
method = "glm",
family = "binomial",
metric = "ROC",
trControl = ctrl),
times = 5)
#Error in serialize(data, node$con) : error writing to connection
编辑(尝试不调用parallel::makeCluster())
在 Linux 设置中(见下文)也尝试不调用 parallel::makeCluster(),即如下所示,但结果相同。
cores_2_use <- floor(0.8 * detectCores())
registerDoParallel(cores_2_use)
...
输出parallel_log1.txt
starting worker pid=3880 on localhost:11442 at 16:00:52.764
starting worker pid=3388 on localhost:11442 at 16:00:53.405
starting worker pid=9920 on localhost:11442 at 16:00:53.789
starting worker pid=4248 on localhost:11442 at 16:00:54.229
starting worker pid=3548 on localhost:11442 at 16:00:54.572
starting worker pid=5704 on localhost:11442 at 16:00:54.932
starting worker pid=7740 on localhost:11442 at 16:00:55.291
starting worker pid=2164 on localhost:11442 at 16:00:55.653
starting worker pid=7428 on localhost:11442 at 16:00:56.011
starting worker pid=6116 on localhost:11442 at 16:00:56.372
starting worker pid=1632 on localhost:11442 at 16:00:56.731
starting worker pid=9160 on localhost:11442 at 16:00:57.092
starting worker pid=2956 on localhost:11442 at 16:00:57.435
starting worker pid=7060 on localhost:11442 at 16:00:57.811
starting worker pid=7344 on localhost:11442 at 16:00:58.170
starting worker pid=6688 on localhost:11442 at 16:00:58.561
starting worker pid=9308 on localhost:11442 at 16:00:58.920
starting worker pid=9260 on localhost:11442 at 16:00:59.281
starting worker pid=6212 on localhost:11442 at 16:00:59.641
输出parallel_log2.txt
starting worker pid=17640 on localhost:11074 at 15:12:21.118
starting worker pid=7776 on localhost:11074 at 15:12:21.494
starting worker pid=15128 on localhost:11074 at 15:12:21.961
starting worker pid=13724 on localhost:11074 at 15:12:22.345
starting worker pid=17384 on localhost:11074 at 15:12:22.714
starting worker pid=8472 on localhost:11074 at 15:12:23.228
starting worker pid=8392 on localhost:11074 at 15:12:23.597
starting worker pid=17412 on localhost:11074 at 15:12:23.979
starting worker pid=15996 on localhost:11074 at 15:12:24.364
starting worker pid=16772 on localhost:11074 at 15:12:24.743
starting worker pid=18268 on localhost:11074 at 15:12:25.120
starting worker pid=13504 on localhost:11074 at 15:12:25.500
starting worker pid=5156 on localhost:11074 at 15:12:25.899
starting worker pid=13544 on localhost:11074 at 15:12:26.275
starting worker pid=1764 on localhost:11074 at 15:12:26.647
starting worker pid=8076 on localhost:11074 at 15:12:27.028
starting worker pid=13716 on localhost:11074 at 15:12:27.414
starting worker pid=14596 on localhost:11074 at 15:12:27.791
starting worker pid=15664 on localhost:11074 at 15:12:28.170
Loading required package: caret
Loading required package: lattice
Loading required package: ggplot2
loaded caret and set parent environment
starting worker pid=3932 on localhost:11442 at 16:01:44.384
starting worker pid=6848 on localhost:11442 at 16:01:44.731
starting worker pid=5400 on localhost:11442 at 16:01:45.098
starting worker pid=9832 on localhost:11442 at 16:01:45.475
starting worker pid=8448 on localhost:11442 at 16:01:45.928
starting worker pid=1284 on localhost:11442 at 16:01:46.289
starting worker pid=9892 on localhost:11442 at 16:01:46.632
starting worker pid=8312 on localhost:11442 at 16:01:46.991
starting worker pid=3696 on localhost:11442 at 16:01:47.349
starting worker pid=9108 on localhost:11442 at 16:01:47.708
starting worker pid=8548 on localhost:11442 at 16:01:48.083
starting worker pid=7288 on localhost:11442 at 16:01:48.442
starting worker pid=6872 on localhost:11442 at 16:01:48.801
starting worker pid=3760 on localhost:11442 at 16:01:49.145
starting worker pid=3468 on localhost:11442 at 16:01:49.503
starting worker pid=2500 on localhost:11442 at 16:01:49.862
starting worker pid=7200 on localhost:11442 at 16:01:50.205
starting worker pid=7820 on localhost:11442 at 16:01:50.564
starting worker pid=8852 on localhost:11442 at 16:01:50.923
Error in unserialize(node$con) :
ReadItem: unknown type 0, perhaps written by later version of R
Calls: <Anonymous> ... doTryCatch -> recvData -> recvData.SOCKnode -> unserialize
Execution halted
编辑(在 Ubuntu 上尝试)
库
library(caret)
library(microbenchmark)
library(doMC)
sessionInfo()
R version 3.4.1 (2017-06-30)
Platform: x86_64-pc-linux-gnu (64-bit)
Running under: Ubuntu 16.04.3 LTS
Matrix products: default
BLAS: /usr/lib/libblas/libblas.so.3.6.0
LAPACK: /usr/lib/lapack/liblapack.so.3.6.0
locale:
[1] LC_CTYPE=en_US.UTF-8 LC_NUMERIC=C LC_TIME=de_DE.UTF-8
[4] LC_COLLATE=en_US.UTF-8 LC_MONETARY=de_DE.UTF-8 LC_MESSAGES=en_US.UTF-8
[7] LC_PAPER=de_DE.UTF-8 LC_NAME=C LC_ADDRESS=C
[10] LC_TELEPHONE=C LC_MEASUREMENT=de_DE.UTF-8 LC_IDENTIFICATION=C
attached base packages:
[1] parallel stats graphics grDevices utils datasets methods base
other attached packages:
[1] doMC_1.3.4 iterators_1.0.8 foreach_1.4.3
[4] microbenchmark_1.4-2.1 caret_6.0-77 ggplot2_2.2.1
[7] lattice_0.20-35
loaded via a namespace (and not attached):
[1] Rcpp_0.12.11 ddalpha_1.2.1 compiler_3.4.1 DEoptimR_1.0-8
[5] gower_0.1.2 plyr_1.8.4 bindr_0.1 class_7.3-14
[9] tools_3.4.1 rpart_4.1-11 ipred_0.9-6 lubridate_1.6.0
[13] tibble_1.3.3 nlme_3.1-131 gtable_0.2.0 pkgconfig_2.0.1
[17] rlang_0.1.1 Matrix_1.2-11 RcppRoll_0.2.2 prodlim_1.6.1
[21] bindrcpp_0.2 withr_2.0.0 stringr_1.2.0 dplyr_0.7.1
[25] recipes_0.1.0 stats4_3.4.1 nnet_7.3-12 CVST_0.2-1
[29] grid_3.4.1 robustbase_0.92-7 glue_1.1.1 R6_2.2.2
[33] survival_2.41-3 lava_1.5 purrr_0.2.2.2 reshape2_1.4.2
[37] kernlab_0.9-25 magrittr_1.5 DRR_0.0.2 splines_3.4.1
[41] scales_0.4.1 codetools_0.2-15 ModelMetrics_1.1.0 MASS_7.3-47
[45] assertthat_0.2.0 dimRed_0.1.0 timeDate_3012.100 colorspace_1.3-2
[49] stringi_1.1.5 lazyeval_0.2.0 munsell_0.4.3
来自Getting Started with doMC and foreach 的示例
按预期工作。
非平行插入符号示例
microbenchmark(
glm_nopar =
train(y ~ .,
data = df,
method = "glm",
family = "binomial",
metric = "ROC",
trControl = ctrl),
times = 5)
#Unit: seconds
# expr min lq mean median uq max neval
#glm_nopar 1.093237 1.098342 1.481444 1.102867 2.001443 2.111333 5
插入符号与 Windows 等设置平行(给出错误)
cores_2_use <- floor(0.8 * parallel::detectCores())
cl <- parallel::makeCluster(cores_2_use, outfile = "parallel_log2_linux.txt")
registerDoMC(cl)
microbenchmark(
glm_par =
train(y ~ .,
data = df,
method = "glm",
family = "binomial",
metric = "ROC",
trControl = ctrl),
times = 5)
# Error in getOper(ctrl$allowParallel && getDoParWorkers() > 1) :(list) object cannot be coerced to type 'double'
parallel_log2_linux.txt
starting worker pid=6343 on localhost:11836 at 16:05:17.781
starting worker pid=6353 on localhost:11836 at 16:05:18.025
starting worker pid=6362 on localhost:11836 at 16:05:18.266
没有parallel::makeCluster()
调用的插入符并行(无错误)
不清楚如何在此设置中定义日志输出。
cores_2_use <- floor(0.8 * parallel::detectCores())
registerDoMC(cores_2_use)
microbenchmark(
glm_par =
train(y ~ .,
data = df,
method = "glm",
family = "binomial",
metric = "ROC",
trControl = ctrl),
times = 5)
#Unit: milliseconds
# expr min lq mean median uq max neval
# glm_par 991.8075 997.4397 1013.686 998.8241 1004.381 1075.978 5
【问题讨论】:
编辑添加 Ubuntu 测试。parallel::makeCluster()
调用似乎会产生错误,但没有它也可以正常工作。
已编辑添加 Windows 设置,但不调用 parallel::makeCluster()
,但会导致相同的错误。
我建议将更新作为一个新问题发布,因为它通常会使您的问题“过于宽泛”,并改变旧答案与您的问题的关系。
总的来说,我同意,但所有编辑都参考原始问题并添加到它而不是改变它的范围。而且我相信,未来的读者在看到已经尝试过的东西时会更加了解。
从一些错误消息(提到node$con
)看来,您的一些工作人员(R 进程)可能已经死亡,导致与主进程的相应连接失败。他们可能因各种原因死亡,但请查看您的内存消耗,这通常会随着工人的数量线性增长。从少量工人 (=2) 开始,看看是否可行。
【参考方案1】:
看起来因为你在 Windows 上,所以你搞砸了
doMC 包充当 foreach 和 parallel 包的多核功能之间的接口,最初由 Simon Urbanek 编写并合并到 R2.14.0 的并行中。多核功能目前仅适用于支持 fork 系统调用的操作系统(这意味着不支持 Windows)
插入符号使用doMC
。见caret/parallel-processing.html
library(doMC)
registerDoMC(cores = 5)
model <- train(y ~ ., data = training, method = "rf")
注意 OP 已编辑他的原始帖子。 OP 一开始就在 Windows 上运行。
编辑 - 一条评论太长
doParallel
不能拯救 caret
并行化。(但我可能是错的...请让我知道更多的反对票和 cmets)
1) 请在 Windows 上自己尝试一下...当我尝试使用 doParalell
时,它默认为顺序。 (我想知道它是否可以在其他人的 Windows 机器上运行)。
这是有道理的,它默认为顺序,因为
2) caret
使用 doMC
。见here,
caret 利用 R 中的一种并行处理框架来做到这一点。 foreach 包允许使用多种不同的技术顺序或并行运行 R 代码,例如多核或 Rmpi 包(有关可用选项的摘要和描述,请参见 Schmidberger 等人,2009 年)。有几个 R 包可以与 foreach 一起使用来实现这些技术,例如 doMC(用于多核)或 doMPI(用于 Rmpi)。
3) doParallel
简单地结合了doMC
和doSNOW
。见here。
doParallel 包是 doSNOW 和 doMC 的合并,就像并行是雪和多核的合并一样。
请注意,链接中接受答案的作者是 Steve Weston,他是 doParallel
包的作者之一。
4) doMC
派生 Windows 不支持的进程(Windows 仅支持 SNOW 和 SOCK 进程)参见here,再次Steve Weston
多核功能目前仅适用于支持 fork 系统调用(这意味着不支持 Windows)
【讨论】:
我也这么认为,但caret ml parallel 建议不这样做,即他正在做同样的事情,而且它似乎在 Windows 上工作。 嗯...不确定。可能是该人在 Windows 10 see here 内安装的 Ubuntu 上运行。我也在我的系统上试过,但train::caret
默认为顺序
好的。稍后我将通过doMC
在 Ubuntu 上运行它,以排除这是另一个问题。
这是错误的。 caret 完全能够使用其他 foreach 后端。
嗨@HongOoi,请看我编辑的答案(评论太长了)【参考方案2】:
您必须使用与您的集群类型相对应的 foreach 后端。如果您使用parallel::makeCluster
创建集群,则使用doParallel::registerDoParallel
注册它。
cl <- parallel::makeCluster(cores_2_use, outfile = "parallel_log2_linux.txt")
library(doParallel)
registerDoParallel(cl)
【讨论】:
不确定我是否关注。这可能解释了我在 Linux 上遇到的第一个错误,然后我证明该错误已解决,但您显示的设置正是我显示的在 Windows 上产生序列化错误的设置。【参考方案3】:我在另一台内核较少但代码设置相同的 Windows 10 机器上进行了尝试。但是,我使用了来自 Github 的 caret
的开发版本(通过 devtools::install_github('topepo/caret/pkg/caret')
安装)以及 R 3.4.1,问题无法重现。并行集群运行时没有出现以下代码问题。不幸的是,我无法访问原始的 Windows 7 工作站,以查看caret
dev 版本和/或更新的 R 版本是否仍然存在问题。
library(doParallel)
cores_2_use <- floor(0.8 * detectCores())
cl <- makeCluster(cores_2_use, outfile = "parallel_log.txt")
registerDoParallel(cl)
glm_par <-
microbenchmark(glm_par =
train(default ~ .,
data = benchmark_train_data,
method = "glm",
family = "binomial",
metric = "ROC",
trControl = ctrl),
times = 5
)
glm_par
#Unit: seconds
# expr min lq mean median uq max neval
# glm_par 13.14082 13.25298 16.77678 13.64924 13.78132 30.05955 5
EDIT(非并行基准)
这是在一个内核上运行的相同代码(与上面的六个内核并行运行相反)- 本来预计并行设置的性能会更好。
#Unit: seconds
# expr min lq mean median uq max neval
# glm_nopar 25.44122 25.52031 25.64818 25.53692 25.56496 26.17751 5
【讨论】:
是的,尽管使用 6 个内核而不是 1 个内核,但它比非并行更快,尽管不是 6 倍。我还通过资源监视器检查了执行期间的 CPU 使用率,您可以看到所有 CPU 的使用率几乎达到峰值。我可以发布非并行时间tmr。以上是关于插入符号训练二进制 glm 通过 doParallel 在并行集群上失败的主要内容,如果未能解决你的问题,请参考以下文章
在使用插入符号的 train() 使用公式训练的 randomForest 对象上使用 predict() 时出错