使用mlr3的auto_tuner()调用Hyperband/MBO调参报错，tune()正常

问题描述 #

使用mlr3分析含100个特征的数据集，通过Pipeline串联预处理步骤与XGBoost学习器，调用auto_tuner()搭配Hyperband或MBO调参时触发参数缺失错误，但直接使用tune()函数无此问题。

相关代码 #

learner <-
  po("encode", method = "treatment", affect_columns = selector_type("factor")) %>>%
  po("scale") %>>%
  po("learner",
      lrn("classif.xgboost", predict_type = "prob",
          nrounds           = to_tune(p_int(1, 5000, tags="budget")),
          eta               = to_tune(1e-4, 1, logscale = TRUE),
          max_depth         = to_tune(1, 20),
          colsample_bytree  = to_tune(1e-1, 1),
          colsample_bylevel = to_tune(1e-1, 1),
          lambda            = to_tune(1e-3, 1e3, logscale = TRUE),
          alpha             = to_tune(1e-3, 1e3, logscale = TRUE),
          subsample         = to_tune(1e-1, 1)))

at <- auto_tuner(                                                   
  tuner = tnr("hyperband", eta = 2),
  learner = learner,
  resampling = rsmp("cv", folds=2),
  measure = msr("classif.auc"),
  terminator = trm("none"),
  store_models = TRUE,
  evaluate_default = TRUE)

at$train(task_US)

报错信息 #

Error in .__OptimInstance__eval_batch(self = self, private = private,  : 
  Assertion on 'colnames(xdt)' failed: Names must include the elements {'classif.xgboost.nrounds','classif.xgboost.eta','classif.xgboost.max_depth','classif.xgboost.colsample_bytree','classif.xgboost.colsample_bylevel','classif.xgboost.lambda','classif.xgboost.alpha','classif.xgboost.subsample'}, but is missing elements {'classif.xgboost.nrounds'}.

问题原因 #

当在Pipeline的po("learner")中为XGBoost的nrounds参数设置to_tune(p_int(..., tags="budget"))时，auto_tuner会将带budget标签的参数识别为Hyperband专属的预算参数，不会将其纳入普通调参参数列表，导致优化实例调用时缺失该参数。而tune()函数不会做这种特殊识别，会正常处理所有标记为to_tune的参数，因此无报错。

解决方案 #

方案1：移除nrounds的budget标签 #

如果不需要将nrounds作为Hyperband的预算参数，直接去掉tags="budget"，让它作为普通调参参数参与优化：

learner <-
  po("encode", method = "treatment", affect_columns = selector_type("factor")) %>>%
  po("scale") %>>%
  po("learner",
      lrn("classif.xgboost", predict_type = "prob",
          nrounds           = to_tune(p_int(1, 5000)), # 移除budget标签
          eta               = to_tune(1e-4, 1, logscale = TRUE),
          max_depth         = to_tune(1, 20),
          colsample_bytree  = to_tune(1e-1, 1),
          colsample_bylevel = to_tune(1e-1, 1),
          lambda            = to_tune(1e-3, 1e3, logscale = TRUE),
          alpha             = to_tune(1e-3, 1e3, logscale = TRUE),
          subsample         = to_tune(1e-1, 1)))

方案2：明确指定Hyperband的预算参数 #

如果需要保留nrounds作为Hyperband的预算参数，在auto_tuner中通过budget_variable参数明确指定该参数的完整名称，确保调参器正确识别：

at <- auto_tuner(                                                   
  tuner = tnr("hyperband", eta = 2),
  learner = learner,
  resampling = rsmp("cv", folds=2),
  measure = msr("classif.auc"),
  terminator = trm("none"),
  store_models = TRUE,
  evaluate_default = TRUE,
  budget_variable = "classif.xgboost.nrounds" # 明确指定预算参数
)

方案3：用GraphLearner包装Pipeline #

将Pipeline转换为GraphLearner对象后再传入auto_tuner，避免参数识别异常：

# 定义Pipeline图
learner_graph <-
  po("encode", method = "treatment", affect_columns = selector_type("factor")) %>>%
  po("scale") %>>%
  po("learner",
      lrn("classif.xgboost", predict_type = "prob",
          nrounds           = to_tune(p_int(1, 5000, tags="budget")),
          eta               = to_tune(1e-4, 1, logscale = TRUE),
          max_depth         = to_tune(1, 20),
          colsample_bytree  = to_tune(1e-1, 1),
          colsample_bylevel = to_tune(1e-1, 1),
          lambda            = to_tune(1e-3, 1e3, logscale = TRUE),
          alpha             = to_tune(1e-3, 1e3, logscale = TRUE),
          subsample         = to_tune(1e-1, 1)))

# 转换为GraphLearner
learner <- GraphLearner$new(learner_graph)

# 创建auto_tuner
at <- auto_tuner(                                                   
  tuner = tnr("hyperband", eta = 2),
  learner = learner,
  resampling = rsmp("cv", folds=2),
  measure = msr("classif.auc"),
  terminator = trm("none"),
  store_models = TRUE,
  evaluate_default = TRUE)

内容的提问来源于stack exchange，提问作者Jemay