对于带有编码器错误的模拟数据,cv.glmnet对于岭而不是套索失败

Question

要旨

错误: Error in predmat[which, seq(nlami)] = preds : replacement has length zero

上下文:数据用二进制y模拟,但有n编码器true y.数据是叠加的n时间,并且模型已经安装,试图获得true y.

收到错误

1. L2罚款,但不是L1罚款.
2. 当Y是编码器Y时,而不是当它是真正的Y.
3. 错误不是确定性的,而是取决于种子.

更新:错误是针对1.9-8之后的版本.1.9-8不会失败.

再生产

基础数据:

library(glmnet)
rm(list=ls())
set.seed(123)

num_obs=4000
n_coders=2
precision=.8

X <- matrix(rnorm(num_obs*20, sd=1), nrow=num_obs)
prob1 <- plogis(X %*% c(2, -2, 1, -1, rep(0, 16))) # yes many zeros, ignore
y_true <- rbinom(num_obs, 1, prob1)
dat <- data.frame(y_true = y_true, X = X)

创建编码员

classify <- function(true_y,precision){
  n=length(true_y)
  y_coder <- numeric(n)
  y_coder[which(true_y==1)] <- rbinom(n=length(which(true_y==1)),
                                      size=1,prob=precision)
  y_coder[which(true_y==0)] <- rbinom(n=length(which(true_y==0)),
                                      size=1,prob=(1-precision))
  return(y_coder)
}
y_codings <- sapply(rep(precision,n_coders),classify,true_y = dat$y_true)

把它堆叠起来

expanded_data <- do.call(rbind,rep(list(dat),n_coders))
expanded_data$y_codings <- matrix(y_codings, ncol = 1)

重现错误

由于错误取决于种子,因此需要循环.只有第一个循环会失败,其他两个循环都会完成.

X <- as.matrix(expanded_data[,grep("X",names(expanded_data))])

for (i in 1:1000) cv.glmnet(x = X,y = expanded_data$y_codings,
                            family="binomial", alpha=0)  # will fail
for (i in 1:1000) cv.glmnet(x = X,y = expanded_data$y_codings,
                            family="binomial", alpha=1)  # will not fail
for (i in 1:1000) cv.glmnet(x = X,y = expanded_data$y_true,
                            family="binomial", alpha=0)  # will not fail

有什么想法来自glmnet以及如何避免它？从我的阅读来看cv.glmnet,这是在cv例程之后并且在里面cvstuff = do.call(fun, list(outlist, lambda, x, y, weights, offset, foldid, type.measure, grouped, keep)),我不理解它的作用,因此失败,以及如何避免它.

会话(Ubuntu和PC)

R version 3.3.1 (2016-06-21)
Platform: x86_64-pc-linux-gnu (64-bit)
Running under: Ubuntu 16.04.1 LTS

locale:
 [1] LC_CTYPE=en_US.UTF-8       LC_NUMERIC=C               LC_TIME=en_US.UTF-8       
 [4] LC_COLLATE=en_US.UTF-8     LC_MONETARY=en_US.UTF-8    LC_MESSAGES=en_US.UTF-8   
 [7] LC_PAPER=en_US.UTF-8       LC_NAME=C                  LC_ADDRESS=C              
[10] LC_TELEPHONE=C             LC_MEASUREMENT=en_US.UTF-8 LC_IDENTIFICATION=C       

attached base packages:
[1] stats     graphics  grDevices utils     datasets  methods   base     

other attached packages:
[1] glmnet_2.0-2    foreach_1.4.3   Matrix_1.2-7.1  devtools_1.12.0

loaded via a namespace (and not attached):
 [1] httr_1.2.1       R6_2.2.0         tools_3.3.1      withr_1.0.2      curl_2.1        
 [6] memoise_1.0.0    codetools_0.2-15 grid_3.3.1       iterators_1.0.8  knitr_1.14      
[11] digest_0.6.10    lattice_0.20-34

和

R version 3.3.1 (2016-06-21)
Platform: x86_64-w64-mingw32/x64 (64-bit)
Running under: Windows 7 x64 (build 7601) Service Pack 1

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] stats     graphics  grDevices utils     datasets  methods   base     

other attached packages:
[1] glmnet_2.0-2    foreach_1.4.3   Matrix_1.2-7.1  devtools_1.12.0

loaded via a namespace (and not attached):
 [1] httr_1.2.1       R6_2.2.0         tools_3.3.1      withr_1.0.2      curl_2.1        
 [6] memoise_1.0.0    codetools_0.2-15 grid_3.3.1       iterators_1.0.8  digest_0.6.10   
[11] lattice_0.20-34

Answer 1

我在glmnet_2.0-5中遇到了同样的错误.这与lambdas在某些情况下如何自动创建有关.解决方案是提供自己的lambdas

例如:

cv.glmnet(x = X,
          y = expanded_data$y_codings,
          family="binomial", 
          alpha=0,
          lambda=exp(seq(log(0.001), log(5), length.out=100))) 

感谢https://github.com/lmweber/glmnet-error-example/blob/master/glmnet_error_example.R

Answer 2

好吧，我刚刚运行了第一个循环，它成功完成了。这是 glmnet 2.0.2 的情况。

这更像是一条评论，但它太大了，无法容纳：当运行像这样依赖于随机数的测试时，您可以随时保存种子。这使您可以跳到测试的中间，而不必每次都返回到开始处。

像这样的东西：

results <- lapply(1:1000, function(x) {
    seed <- .Random.seed
    res <- try(glmnet(x, y, ...))  # so the code keeps running even if there's an error
    attr(res, "seed") <- seed
    res
})

现在您可以通过查看结果的类别来检查是否有任何运行失败：

errs <- sapply(results, function(x) inherits(x, "try-error"))
any(errs)

您可以重试那些失败的运行：

firstErr <- which(errs)[1]
.Random.seed <- attr(results[[firstErr]], "seed")
glmnet(x, y, ...)  # try failed run again

会议信息：

R version 3.2.2 (2015-08-14)
Platform: x86_64-w64-mingw32/x64 (64-bit)
Running under: Windows 8 x64 (build 9200)

locale:
[1] LC_COLLATE=English_United States.1252  LC_CTYPE=English_United States.850    
[3] LC_MONETARY=English_United States.1252 LC_NUMERIC=C                          
[5] LC_TIME=English_United States.1252    

attached base packages:
[1] stats     graphics  grDevices utils     datasets  methods   base     

other attached packages:
[1] glmnetUtils_0.55    RevoUtilsMath_8.0.3 RevoUtils_8.0.3     RevoMods_8.0.3      RevoScaleR_8.0.6   
[6] lattice_0.20-33     rpart_4.1-10       

loaded via a namespace (and not attached):
[1] Matrix_1.2-2     parallel_3.2.2   codetools_0.2-14 rtvs_1.0.0.0     grid_3.2.2      
[6] iterators_1.0.8  foreach_1.4.3    glmnet_2.0-2    

（那应该是Windows 10，而不是8；R 3.2.2不知道Win10）