[R-package] init_model in lgb.cv has error

[EthanKang-Accelerant]

Description

When using init_model inside lgb.cv, an error will occur:

Error in dataset$get_field(field_name = field_name) : Cannot perform Dataset$get_field() before constructing Dataset.

Reproducible example

df_temp <- data.matrix(mtcars[, c('cyl', 'disp', 'hp')])
label_temp <- mtcars$mpg
dtrain <- lgb.Dataset(data = df_temp, label = label_temp, free_raw_data = FALSE)
dtrain$construct()

# Train the init model first using the same constructed dataset
params <- list(
  learning_rate = 0.05,
  num_leaves = 4,
  max_depth = 2,
  verbose = -1
)

init_model <- lgb.train(
  params = params,
  data = dtrain,
  nrounds = 100
)

# Force construction again just to be sure (sometimes the pointer gets stale)
dtrain$construct()

# Now run lgb.cv
cv_result <- lgb.cv(
  params = params,
  data = dtrain,
  init_model = init_model,
  nrounds = 50,
  nfold = 5
)

Environment info

LightGBM version or commit hash:

Command(s) you used to install LightGBM

Additional Comments

Adding init_model in lgb.train works as intended.

df_temp <- data.matrix(mtcars[, c('cyl', 'disp', 'hp')])
label_temp <- mtcars$mpg
dtrain <- lgb.Dataset(data = df_temp, label = label_temp, free_raw_data = FALSE)
dtrain$construct()

# Train the init model first using the same constructed dataset
params <- list(
  learning_rate = 0.05,
  num_leaves = 4,
  max_depth = 2,
  verbose = -1
)

init_model <- lgb.train(
  params = params,
  data = dtrain,
  nrounds = 100
)

# Force construction again just to be sure (sometimes the pointer gets stale)
dtrain$construct()

# Now run lgb.cv
cv_result <- lgb.train(
  params = params,
  data = dtrain,
  init_model = init_model,
  nrounds = 50
)

[jameslamb]

Thanks for using LightGBM and for the report. I hope you don't mind, but I slightly edited your post slightly for formatting (for example, added R syntax highlighting). If you're not sure how I did that, see the tips at these links:

• https://yihui.org/en/2021/06/markdown-breath/
• https://docs.github.com/en/get-started/writing-on-github/getting-started-with-writing-and-formatting-on-github/basic-writing-and-formatting-syntax

[jameslamb]

This does look like a bug to me. You should be able to provide an lgb.Booster object to lgb.cv() exactly as you've done here.

Here's a slightly simpler example that reproduces the error using the latest development version of {lightgbm}:

library(lightgbm)

data(mtcars)

dtrain <- lgb.Dataset(
  data = data.matrix(mtcars[, c("cyl", "disp", "hp")])
  , label = mtcars[, "mpg"]
  , free_raw_data = FALSE
)

params = list(max_depth = 2, num_leaves = 4)

# train an initial model
init_model <- lgb.train(
  params = params
  , data = dtrain
  , nrounds = 5
)

# try using it in lgb.cv()
cv_result <- lgb.cv(
  data = dtrain
  , params = params
  , init_model = init_model
  , nrounds = 1
)

If you have the time and interest, we'd love some help investigating this further!

[EthanKang-Accelerant]

Hello, thanks for looking into this. I am not sure how I can help. I looked at the lgb.cv and lgb.train scripts and both functions show the below code:

Apart from the below, I don't see anywhere else inside the function that's using init_model.

  if (is.character(init_model)) {
    predictor <- Predictor$new(modelfile = init_model)
  } else if (.is_Booster(x = init_model)) {
    predictor <- init_model$to_predictor()
  }

[jameslamb]

No problem, someone else here will look when we have time.