问题
Sometimes I save a LightGBM model and later, upon reloading it, want to access some details about how the model was built. Is there a way to recover the fact that objective = "regression", for example?
For convenience, here is brief code to play with:
library(lightgbm)
data(agaricus.train, package = "lightgbm")
train <- agaricus.train
dtrain <- lgb.Dataset(train$data, label = train$label)
data(agaricus.test, package = "lightgbm")
params <- list(objective = "regression", metric = "l2")
model <- lgb.train(params,
dtrain,
100,
min_data = 1,
learning_rate = 1)
names(model)
I don't see how to retrieve any model parameters from any of the model attributes:
> names(model)
[1] ".__enclos_env__" "raw" "record_evals" "best_score"
[5] "best_iter" "save" "to_predictor" "predict"
[9] "dump_model" "save_model_to_string" "save_model" "eval_valid"
[13] "eval_train" "eval" "current_iter" "rollback_one_iter"
[17] "update" "reset_parameter" "add_valid" "set_train_data_name"
[21] "initialize" "finalize"
回答1:
I'm not using the R binding of lightgbm, but looking through the Booster implementation in version 2.1.1, there seems to be indeed no interface to retrieve parameters. In turn, because params are not an attribute of the Booster class, but just passed down to the back-end C implementation.
Such functionality is also missing in the native python binding (the similar Booster class). However, it is present in the sklearn API. So the native API is consistently missing this function, but the higher-level wrapper in python has it added.
来源：https://stackoverflow.com/questions/49674112/accessing-lightgbm-model-parameters