Can you access scores for each boosting round in LightGBM.train()?

Question

Basic Info

lgbm.train() with early_stopping calculates the objective function & feval scores after each boost round, and we can make it print those every verbose_eval rounds, like so:

bst=lgbm.train(**params)
[10]    valid_0's binary_logloss: 0.215654  valid_0's BinaryError: 0.00775126
[20]    valid_0's binary_logloss: 0.303113  valid_0's BinaryError: 0.00790619
[30]    valid_0's binary_logloss: 0.358056  valid_0's BinaryError: 0.0838744
[40]    valid_0's binary_logloss: 0.386763  valid_0's BinaryError: 0.138462
[50]    valid_0's binary_logloss: 0.411467  valid_0's BinaryError: 0.176986

Question:

Is there any way to access a list of these scores for each boosting round?

The closest thing I can find in the documentation & bst.__dict__ is bst.best_score

defaultdict(collections.OrderedDict,
             {'valid_0': OrderedDict([('binary_logloss', 0.4233895131745753),
                           ('BinaryError', 0.194285077972568)])})

Answer 1

You can use evals_result parameter as follows;

evals_result = {}

bst=lgbm.train(evals_result=evals_result, valid_sets = [valid_set, train_set],
               valid_names = [‘valid’, ‘train’], **params)

evals_result
>>> {‘train’: {‘logloss’: [‘0.36483’, ‘0.32617’, …]}, ‘valid’: {‘logloss’: [‘0.479168’, ‘0.317850’, …]}}

You will have a dictionary for both train and valid set scores for each boosting round.