使用tidymodels调用kernlab遇protect()栈溢出的解决方案问询

问题背景 #

在tidymodels框架下以kernlab为后端建模时，遇到「protect(): protection stack overflow」错误。该错误由变量过多+公式表达式解析导致：单独使用kernlab时，通过x=和y=参数传入矩阵而非data.frame可规避，但tidymodels默认流程会将数据转为data.frame/data.table并使用公式接口，触发栈溢出。已尝试设置--max-ppsize=500000，无效。需保留tidymodels的标准化预处理及调参功能。

复现代码 #

library(tidymodels)

x <- as.data.frame(matrix(rnorm(2000000), nrow = 100, ncol = 20000))
y <- data.frame("y" = rnorm(n = 100))
data <- cbind(y, x)

train.cv <- vfold_cv(data,5)
  
svm_model <- svm_rbf(
  cost = tune(),
  rbf_sigma = tune(),
  margin = tune(),
  engine = "kernlab",
  mode = "regression"
)

svm_recep <- recipe(x = as.matrix(data)) %>%
  update_role(everything(), new_role = "predictor") %>% 
  update_role(y, new_role = "outcome")

svm_wflow <- workflow(preprocessor = svm_recep, spec = svm_model)
  
svm_set <- extract_parameter_set_dials(svm_wflow)
hypercube <- grid_latin_hypercube(svm_set, size = 5)
  
svm_initial <- svm_wflow %>%
  tune_grid(resamples = train.cv, grid = hypercube, metrics = metric_set(rmse))

单独使用kernlab的错误与修复示例 #

错误代码（触发栈溢出） #

library(kernlab)
svm.train <- ksvm(y ~ ., type="eps-svr", data = data, kernel ="rbfdot")

修复代码（使用矩阵传入） #

library(kernlab)
svm.train <- ksvm(x = as.matrix(data[,-1]), y = as.matrix(data[,1]), type="eps-svr", kernel ="rbfdot")

解决方案：自定义parsnip矩阵接口 #

核心思路是让tidymodels直接将矩阵格式的x和y传给kernlab，绕过公式解析流程。具体步骤如下：

1. 自定义kernlab SVM的矩阵接口 #

复制默认svm_rbf的实现，修改为使用矩阵输入的接口：

# 定义自定义模型规格
svm_rbf_matrix <- function(mode = "regression", cost = NULL, rbf_sigma = NULL, margin = NULL) {
  parsnip::new_model_spec(
    "svm_rbf_matrix",
    args = list(
      cost = rlang::enquo(cost),
      rbf_sigma = rlang::enquo(rbf_sigma),
      margin = rlang::enquo(margin)
    ),
    eng_args = list(),
    mode = mode,
    method = NULL,
    engine = "kernlab"
  )
}

# 设置fit和predict方法，指定矩阵接口
set_fit_engine("svm_rbf_matrix", engine = "kernlab") <- function(object, engine = NULL) {
  object$method <- parsnip::new_method(
    fit = parsnip::fit_spec(
      interface = "matrix",
      protect = c("x", "y", "type", "kernel"),
      func = c(pkg = "kernlab", fun = "ksvm"),
      defaults = list(
        type = "eps-svr",
        kernel = "rbfdot"
      )
    ),
    predict = parsnip::predict_spec(
      interface = "matrix",
      func = c(pkg = "kernlab", fun = "predict"),
      defaults = list(type = "response")
    )
  )
  object
}

# 映射parsnip参数到kernlab原生参数
set_model_arg("svm_rbf_matrix", eng = "kernlab", parsnip = "cost", original = "C") <- function(value) value
set_model_arg("svm_rbf_matrix", eng = "kernlab", parsnip = "rbf_sigma", original = "sigma") <- function(value) value
set_model_arg("svm_rbf_matrix", eng = "kernlab", parsnip = "margin", original = "epsilon") <- function(value) value

2. 修改workflow的预处理与模型配置 #

确保预处理后输出矩阵格式，并使用自定义模型：

# 定义包含标准化步骤的recipe（保留你的预处理逻辑）
svm_recep <- recipe(y ~ ., data = data) %>%
  update_role(y, new_role = "outcome") %>%
  update_role(-y, new_role = "predictor") %>%
  step_normalize(all_predictors())  # 示例标准化步骤，可替换为你的预处理

# 使用自定义模型
svm_model <- svm_rbf_matrix(
  cost = tune(),
  rbf_sigma = tune(),
  margin = tune(),
  engine = "kernlab",
  mode = "regression"
)

# 设置蓝图，强制预处理输出矩阵格式
mat_blueprint <- hardhat::default_recipe_blueprint(
  composition = "matrix",
  allow_nominal = FALSE
)

# 构建workflow
svm_wflow <- workflow() %>%
  add_recipe(svm_recep, blueprint = mat_blueprint) %>%
  add_model(svm_model)

# 执行调参（流程与原代码一致）
svm_set <- extract_parameter_set_dials(svm_wflow)
hypercube <- grid_latin_hypercube(svm_set, size = 5)
svm_initial <- svm_wflow %>%
  tune_grid(resamples = train.cv, grid = hypercube, metrics = metric_set(rmse))

方案原理 #

自定义的svm_rbf_matrix模型使用parsnip的matrix接口，直接将预处理后的预测变量矩阵和结果向量传入ksvm的x和y参数，彻底避免了公式解析大量变量时的栈溢出问题，同时完整保留tidymodels的预处理、交叉验证和调参功能。

环境信息 #

> library(tidymodels)
── Attaching packages ───────────────────────────── tidymodels 1.2.0 ──
✔ broom        1.0.5      ✔ recipes      1.0.10
✔ dials        1.2.1      ✔ rsample      1.2.1 
✔ dplyr        1.1.4      ✔ tibble       3.2.1 
✔ ggplot2      3.5.0      ✔ tidyr        1.3.1 
✔ infer        1.0.7      ✔ tune         1.2.1 
✔ modeldata    1.3.0      ✔ workflows    1.1.4 
✔ parsnip      1.2.1      ✔ workflowsets 1.1.0 
✔ purrr        1.0.2      ✔ yardstick    1.3.1 

> sessionInfo()
R version 4.3.3 (2024-02-29 ucrt)
Platform: x86_64-w64-mingw32/x64 (64-bit)
Running under: Windows 10 x64 (build 19045)

Matrix products: default


locale:
[1] LC_COLLATE=German_Germany.utf8  LC_CTYPE=German_Germany.utf8   
[3] LC_MONETARY=German_Germany.utf8 LC_NUMERIC=C                    
[5] LC_TIME=German_Germany.utf8    

time zone: Europe/Berlin
tzcode source: internal

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

other attached packages:
 [1] kernlab_0.9-32     yardstick_1.3.1    workflowsets_1.1.0
 [4] workflows_1.1.4    tune_1.2.1         tidyr_1.3.1       
 [7] tibble_3.2.1       rsample_1.2.1      recipes_1.0.10    
[10] purrr_1.0.2        parsnip_1.2.1      modeldata_1.3.0   
[13] infer_1.0.7        ggplot2_3.5.0      dplyr_1.1.4       
[16] dials_1.2.1        scales_1.3.0       broom_1.0.5       
[19] tidymodels_1.2.0  

loaded via a namespace (and not attached):
 [1] tidyselect_1.2.1    timeDate_4022.108   blob_1.2.3         
 [4] fastmap_1.1.0       digest_0.6.35       rpart_4.1.23       
 [7] timechange_0.3.0    lifecycle_1.0.3     ellipsis_0.3.2     
[10] survival_3.5-8      RSQLite_2.2.20      magrittr_2.0.3     
[13] compiler_4.3.3      rlang_1.1.3         tools_4.3.3        
[16] utf8_1.2.2          data.table_1.14.2   bit_4.0.4          
[19] DiceDesign_1.9      withr_2.5.0         nnet_7.3-19        
[22] grid_4.3.3          fansi_1.0.3         colorspace_2.0-3   
[25] future_1.33.2       globals_0.16.2      iterators_1.0.14   
[28] MASS_7.3-60.0.1     cli_3.6.2           generics_0.1.2     
[31] rstudioapi_0.16.0   future.apply_1.10.0 DBI_1.2.2          
[34] cachem_1.0.6        stringr_1.5.0       splines_4.3.3      
[37] parallel_4.3.3      vctrs_0.6.5         hardhat_1.3.1      
[40] Matrix_1.5-1        bit64_4.0.5         listenv_0.9.0      
[43] foreach_1.5.2       gower_1.0.1         glue_1.6.2         
[46] parallelly_1.34.0   codetools_0.2-19    lubridate_1.9.3    
[49] stringi_1.7.6       gtable_0.3.0        munsell_0.5.0      
[52] GPfit_1.0-8         pillar_1.9.0        furrr_0.3.1        
[55] ipred_0.9-14        lava_1.7.2.1        R6_2.5.1           
[58] lhs_1.1.6           lattice_0.22-5      backports_1.4.1    
[61] memoise_2.0.1       class_7.3-22        Rcpp_1.0.12        
[64] prodlim_2023.03.31  pkgconfig_2.0.3  

内容的提问来源于stack exchange，提问作者R.Pickman