mlr3嵌套重采样报错No model stored，内存占用过高问题求助

在使用mlr3执行嵌套重采样（结合RFE特征选择与重复交叉验证）时，出现以下错误：

Error: No model stored

若开启store_models=TRUE，则会因模型内存占用过大导致128GB内存的RStudio Workbench崩溃。已知不指定store_models可避免存储中间模型以降低内存消耗，但仍能提取预测结果和性能指标，目前尚未尝试store_backends=FALSE，需排查问题并找到参数调整方案。

原代码 #

MSvCon <- read.csv("MS v Control Proteomics Final.csv", row.names=1)

MSvCon$Status <- as.factor(MSvCon$Status)

MSvCon[,2:4399] <- scale(MSvCon[,2:4399], center=TRUE, scale=TRUE)

set.seed(123, "L'Ecuyer")

task = as_task_classif(MSvCon, target = "Status")

learner = lrn("classif.ranger", importance = "impurity", num.trees=10000)

set_threads(learner, n = 8)

measure = msr("classif.fbeta", beta=1, average="micro")

terminator = trm("none")

resampling_inner = rsmp("repeated_cv", folds = 10, repeats = 10)

at = AutoFSelector$new(
  learner = learner,
  resampling = resampling_inner,
  measure = measure,
  terminator = terminator,
  fselect = fs("rfe", n_features = 1, feature_fraction = 0.5, recursive = FALSE))

resampling_outer = rsmp("repeated_cv", folds = 10, repeats = 10)

rr = resample(task, at, resampling_outer)

问题原因 #

No model stored错误源于RFE特征选择需要访问模型的特征重要性来完成筛选逻辑，但默认情况下AutoFSelector不存储模型，导致无法获取必要的模型状态。而全局开启store_models=TRUE会存储所有内层和外层的模型，对于高维度数据集（4k+特征）+大量决策树（10000棵）的组合，内存占用会急剧膨胀。

解决方案与参数调整 #

1. 分层控制模型存储 #

给AutoFSelector单独开启store_models=TRUE（满足RFE的模型访问需求），同时在外层重采样时关闭store_models=FALSE（避免存储外层大模型），精准控制内存占用：

# 修改AutoFSelector初始化
at = AutoFSelector$new(
  learner = learner,
  resampling = resampling_inner,
  measure = measure,
  terminator = terminator,
  fselect = fs("rfe", n_features = 1, feature_fraction = 0.5, recursive = FALSE),
  store_models = TRUE  # 内层存储模型用于RFE筛选
)

# 修改外层重采样调用
rr = resample(task, at, resampling_outer, store_models = FALSE)

2. 优化RFE与ranger参数降低内存 #

• 减少决策树数量：ranger的num.trees=10000内存开销极大，实际性能提升在树数量到1000后趋于平缓，可先降到1000测试；
• 调整RFE终止条件：n_features=1会导致大量筛选轮次，可先设为50等合理值，减少迭代次数；
• 启用递归特征消除：设置recursive=TRUE，RFE会逐步移除最不重要的特征，比每次删除一半特征更高效；
• 开启ranger内存优化：添加save.memory=TRUE参数，让ranger使用更节省内存的数据结构。

3. store_backends=FALSE的作用 #

该参数控制是否存储任务的原始数据集后端，默认TRUE会为每个重采样折复制完整数据集。对于高维度数据集，开启store_backends=FALSE可大幅减少内存占用，但它不会直接解决No model stored错误，建议结合其他优化一起开启：

# 在AutoFSelector和外层重采样中都设置
at = AutoFSelector$new(
  # ... 其他参数
  store_backends = FALSE
)

rr = resample(task, at, resampling_outer, store_models = FALSE, store_backends = FALSE)

4. 其他内存优化技巧 #

• 减少线程数：当前set_threads(learner, n=8)，内存紧张时可降到4，降低并行内存开销；
• 手动触发垃圾回收：重采样前调用gc()释放闲置内存；
• 清理无关变量：删除不需要的中间对象，比如原始数据集MSvCon在生成task后可删除。

修改后的完整代码示例 #

MSvCon <- read.csv("MS v Control Proteomics Final.csv", row.names=1)
MSvCon$Status <- as.factor(MSvCon$Status)
MSvCon[,2:4399] <- scale(MSvCon[,2:4399], center=TRUE, scale=TRUE)

set.seed(123, "L'Ecuyer")

task = as_task_classif(MSvCon, target = "Status")
# 优化ranger参数
learner = lrn("classif.ranger", 
              importance = "impurity", 
              num.trees=1000,  # 降低树数量
              save.memory = TRUE)  # 内存优化
set_threads(learner, n = 4)  # 减少线程数

measure = msr("classif.fbeta", beta=1, average="micro")
terminator = trm("none")

resampling_inner = rsmp("repeated_cv", folds = 10, repeats = 10)

at = AutoFSelector$new(
  learner = learner,
  resampling = resampling_inner,
  measure = measure,
  terminator = terminator,
  fselect = fs("rfe", n_features = 50, feature_fraction = 0.7, recursive = TRUE),
  store_models = TRUE,  # 内层存模型供RFE使用
  store_backends = FALSE  # 不存数据集后端
)

resampling_outer = rsmp("repeated_cv", folds = 10, repeats = 10)

# 外层不存模型，不存后端
rr = resample(task, at, resampling_outer, store_models = FALSE, store_backends = FALSE)

# 提取性能指标与预测结果
rr$score(measure)
rr$predictions()

内容的提问来源于stack exchange，提问作者awaldman