在mlr3中用glmnet实现松弛LASSO报错：找不到对象'cv'

问题背景 #

将直接使用glmnet包的松弛LASSO回归代码迁移到mlr3框架时，使用classif.cv_glmnet学习器调优lambda可正常运行，但使用classif.glmnet学习器并开启relax选项训练时，出现错误提示object 'cv' not found，关闭relax选项则训练正常。

原代码与错误栈 #

复现代码 #

library(mlr3)
library(mlr3learners)
library(glmnet)

# 定义任务
task.sonar <- mlr_tasks$get("sonar")

# 交叉验证调优lambda
learner.sonar.lasso.cv <- lrn("classif.cv_glmnet", alpha = 1, relax = TRUE, gamma = 0, nfolds = 15)
set.seed(1112)
learner.sonar.lasso.cv$train(task.sonar)

# 拟合完整模型（报错代码）
learner.sonar.lasso <- lrn("classif.glmnet", alpha = 1, lambda = learner.sonar.lasso.cv$model$lambda, relax = TRUE, gamma = 0)
learner.sonar.lasso$train(task.sonar)
# Error in eval(call, parent.frame()) : object 'cv' not found

# 关闭relax选项可正常训练
learner.sonar.lasso.2 <- lrn("classif.glmnet", lambda = learner.sonar.lasso.cv$model$lambda, alpha = 1)
learner.sonar.lasso.2$train(task.sonar)

完整错误栈 #

Fehler in eval(call, parent.frame()) : Objekt 'cv' nicht gefunden
23: eval(call, parent.frame())
22: eval(call, parent.frame())
21: update.default(fit, x = x, lambda = lam0, exclude = exclude, 
        ..., relax = FALSE)
20: update(fit, x = x, lambda = lam0, exclude = exclude, ..., relax = FALSE)
19: update(fit, x = x, lambda = lam0, exclude = exclude, ..., relax = FALSE)
18: relax.glmnet(fit, x = x, y = y, weights = weights, offset = offset, 
        lower.limits = lower.limits, upper.limits = upper.limits, 
        penalty.factor = penalty.factor, check.args = FALSE, ...)
17: (if (cv) glmnet::cv.glmnet else glmnet::glmnet)(x = data, y = target, 
        alpha = 1, lambda = c(0.215936661924212, 0.196753444025244, 
        ...), relax = TRUE, 
        family = "binomial")
16: eval(expr, p)
15: eval(expr, p)
14: eval.parent(expr, n = 1L)
13: invoke(if (cv) glmnet::cv.glmnet else glmnet::glmnet, x = data, 
        y = target, .args = pv)
12: glmnet_invoke(data, target, pv)
11: .__LearnerClassifGlmnet__.train(self = self, private = private, 
        super = super, task = task)
10: get_private(learner)$.train(task)
9: .f(learner = <environment>, task = <environment>)
8: eval(expr, p)
7: eval(expr, p)
6: eval.parent(expr, n = 1L)
5: invoke(.f, .args = .args, .opts = .opts, .seed = .seed, .timeout = .timeout)
4: encapsulate(learner$encapsulate["train"], .f = train_wrapper, 
       .args = list(learner = learner, task = task), .pkgs = learner$packages, 
       .seed = NA_integer_, .timeout = learner$timeout["train"])
3: learner_train(learner, task, train_row_ids = train_row_ids, test_row_ids = test_row_ids, 
       mode = mode)
2: .__Learner__train(self = self, private = private, super = super, 
       task = task, row_ids = row_ids)
1: learner.sonar.lasso$train(task.sonar)

解决方案 #

方案1：直接复用classif.cv_glmnet的松弛模型 #

classif.cv_glmnet在开启relax=TRUE时，已经完成了松弛模型的训练，无需再用classif.glmnet重新拟合，可直接使用：

# 从cv训练结果中提取松弛后的完整模型
relaxed_full_model <- learner.sonar.lasso.cv$model$glmnet.fit

# 直接用cv学习器进行预测（自动使用最优lambda对应的松弛模型）
predictions <- learner.sonar.lasso.cv$predict(task.sonar)

方案2：手动调用relax.glmnet封装模型 #

若必须使用classif.glmnet学习器，可先训练基础模型，再手动调用relax.glmnet获取松弛模型并赋值：

# 先训练不带relax的基础glmnet模型
base_learner <- lrn("classif.glmnet", alpha = 1, lambda = learner.sonar.lasso.cv$model$lambda)
base_learner$train(task.sonar)

# 提取任务的特征和目标变量
x <- as.matrix(task.sonar$data(cols = task.sonar$feature_names))
y <- task.sonar$data(cols = task.sonar$target_names)[[1]]

# 手动调用relax.glmnet生成松弛模型
relaxed_fit <- relax.glmnet(base_learner$model, x = x, y = y, gamma = 0)

# 将松弛模型赋值给原学习器
base_learner$model <- relaxed_fit

# 此时可正常使用该学习器进行预测
predictions_relaxed <- base_learner$predict(task.sonar)

错误原因 #

从错误栈可知，mlr3learners的classif.glmnet学习器内部调用glmnet_invoke函数时，使用了if (cv) glmnet::cv.glmnet else glmnet::glmnet的分支逻辑，但在调用relax.glmnet时，cv变量不在当前作用域内，导致eval执行时找不到该变量，从而抛出错误。

内容的提问来源于stack exchange，提问作者KxZzvv