R语言中lightGBM模型保存加载后预测失败问题求助

问题现象 #

使用saveRDS()保存lightGBM模型，readRDS()加载后执行预测时触发错误：

Error in predictor$predict(data = data, start_iteration = start_iteration, : Attempting to use a Booster which no longer exists. This can happen if you have called Booster$finalize() or if this Booster was saved with saveRDS(). To avoid this error in the future, use saveRDS.lgb.Booster() or Booster$save_model() to save lightgbm Boosters.

训练完成后直接使用模型，或在当前会话内加载模型均可正常预测，但在其他应用会话中加载后报错；lasso、xgboost等其他模型无此问题。用户代码如下：

library(tidymodels)
library(lightgbm)
library(bonsai)

TMA_model <- data.frame(
  age = c(50, 45, 60, 55, 70, 34, 55, 48, 58, 42,
          52, 47, 62, 57, 72, 36, 53, 49, 59, 44,
          51, 46, 61, 56, 71, 35, 54, 50, 60, 43), 
  delta_creat = c(1.2, 1.0, 1.5, 1.3, 1.8, 1.2, 1.3, 1.4, 1.1, 1.6,
                  1.2, 1.0, 1.5, 1.3, 1.8, 1.2, 1.3, 1.4, 1.1, 1.6,
                  1.2, 1.0, 1.5, 1.3, 1.8, 1.2, 1.3, 1.4, 1.1, 1.6), 
  max_LDH = c(300, 280, 320, 310, 330, 295, 325, 290, 315, 305,
              305, 315, 290, 320, 300, 325, 280, 330, 310, 295,
              310, 295, 330, 280, 320, 305, 310, 325, 315, 290), 
  min_plat = c(150, 140, 160, 155, 170, 145, 165, 150, 135, 160,
               160, 155, 170, 145, 150, 160, 140, 170, 155, 145,
               155, 145, 170, 140, 160, 150, 155, 160, 165, 135),
  min_hb = c(12, 11.5, 13, 12.5, 14, 11.8, 13.2, 12.2, 12.6, 11.9,
             12.4, 11.7, 13.1, 12.3, 13.5, 11.6, 12.8, 12.1, 13.4, 12.0,
             12.9, 11.6, 13.2, 12.8, 13.0, 11.5, 12.7, 12.3, 12.5, 11.8), 
  max_ast = c(40, 38, 42, 41, 45, 37, 43, 39, 44, 40,
              42, 38, 44, 41, 40, 39, 43, 38, 45, 37,
              44, 37, 42, 38, 41, 40, 43, 39, 44, 38), 
  max_bt = c(37, 37.2, 37.5, 37.3, 37.8, 37.1, 37.4, 37.6, 37.0, 37.7,
             37.2, 37.1, 37.5, 37.4, 37.3, 37.0, 37.6, 37.8, 37.2, 37.4,
             37.7, 37.1, 37.3, 37.8, 37.0, 37.4, 37.2, 37.6, 37.5, 37.1), 
  max_ttap = c(100, 95, 105, 102, 110, 97, 108, 93, 103, 98,
               105, 100, 95, 110, 102, 108, 97, 93, 103, 95,
               100, 97, 105, 93, 108, 102, 110, 98, 103, 100), 
  max_tp = c(12, 12.2, 12.5, 12.3, 12.8, 12.1, 12.4, 12.6, 12.0, 12.7,
             12.3, 12.2, 12.1, 12.8, 12.4, 12.0, 12.6, 12.7, 12.5, 12.3,
             12.1, 12.4, 12.7, 12.8, 12.0, 12.2, 12.3, 12.6, 12.5, 12.4), 
  hypertension = c(1, 0, 1, 1, 0, 0, 1, 0, 1, 0,
                   1, 0, 1, 0, 1, 0, 0, 1, 1, 0,
                   0, 1, 0, 1, 0, 0, 1, 1, 0, 1),  
  MAP = c(90, 92, 88, 91, 89, 93, 87, 94, 86, 90,
          92, 94, 88, 90, 91, 89, 87, 93, 86, 92,
          91, 90, 93, 94, 86, 88, 89, 87, 92, 91),  
  TMA_class = factor(c(1, 2, 3, 4, 5, 1, 2, 3, 4, 5,
                       1, 2, 3, 4, 5, 1, 2, 3, 4, 5,
                       1, 2, 3, 4, 5, 1, 2, 3, 4, 5))
)

new_data <- data.frame(
  age = c(48, 55, 62, 54, 67),  
  delta_creat = c(1.2, 1.3, 1.1, 1.4, 1.6),  
  max_LDH = c(305, 290, 320, 310, 295),  
  min_plat = c(150, 160, 135, 155, 145),  
  min_hb = c(12.2, 12.6, 11.9, 12.4, 11.7),  
  max_ast = c(38, 43, 39, 44, 40),  
  max_bt = c(37.2, 37.4, 37.0, 37.6, 37.1),  
  max_ttap = c(95, 108, 93, 103, 98),  
  max_tp = c(12.2, 12.4, 12.0, 12.6, 12.7),  
  hypertension = c(1, 0, 1, 0, 1),  
  MAP = c(92, 87, 94, 86, 90)  
)

model_spec <- 
  boost_tree(mode = 'classification',
             mtry = 6,  
             trees = 50, 
             min_n = 50,  
             tree_depth = 5) %>%
  set_engine('lightgbm')

recipe <- recipe(TMA_class ~ ., data = TMA_model)

model_fit <- workflow() %>%
  add_recipe(recipe) %>%
  add_model(model_spec) %>%
  fit(data = TMA_model)

saveRDS(model_fit, "model_fit.rds")
model_b <- readRDS("model_fit.rds")

new_data_predictions <- predict(model_b, new_data)

print(new_data_predictions)

问题原因 #

lightGBM的Booster对象包含底层C++实现的指针，标准saveRDS()无法正确序列化这些底层对象，导致加载后丢失关键的模型执行信息，进而触发预测错误。

解决方案 #

方案1：使用lightGBM专用方法保存/加载Booster对象 #

从tidymodels workflow中提取底层Booster模型，用lightGBM提供的专用方法处理：

# 训练完成后，提取底层Booster对象
booster_obj <- extract_fit_parsnip(model_fit)$fit

# 用lightGBM专用方法保存
saveRDS.lgb.Booster(booster_obj, "lightgbm_booster.rds")

# 加载模型
loaded_booster <- readRDS("lightgbm_booster.rds")

# 将Booster重新整合回parsnip模型，适配workflow使用
loaded_parsnip <- parsnip::new_model_fit(
  fit = loaded_booster,
  spec = model_spec,
  preproc = extract_preprocessor(model_fit)
)

# 构建新的workflow并预测
loaded_workflow <- workflow() %>%
  add_recipe(recipe) %>%
  add_model(loaded_parsnip)

new_data_predictions <- predict(loaded_workflow, new_data)

方案2：用文本格式保存模型（跨会话更稳定） #

将Booster保存为文本格式，再加载：

# 提取Booster并保存为文本
booster_obj <- extract_fit_parsnip(model_fit)$fit
booster_obj$save_model("lightgbm_model.txt")

# 加载文本模型
loaded_booster <- lgb.load("lightgbm_model.txt")

# 整合回workflow使用（步骤同方案1）

方案3：用butcher包优化模型序列化（推荐） #

butcher包可以清理模型中的冗余信息，同时正确处理lightGBM的外部指针，直接保存整个workflow：

library(butcher)

# 清理模型并保存
model_fit_clean <- butcher(model_fit)
saveRDS(model_fit_clean, "model_fit_clean.rds")

# 加载后直接预测
model_b <- readRDS("model_fit_clean.rds")
new_data_predictions <- predict(model_b, new_data)

内容的提问来源于stack exchange，提问作者Abner Mácola