Can't reproduce L1-score

[MarcoGorelli]

Environment info

Operating System: Linux 64-bit

CPU/GPU model:

C++/Python/R version: 3.7.3

LightGBM version or commit hash: 2.2.1

Reproducible examples

When I run LGBM with early stopping, it gives me the scores corresponding to its best iteration.

When I try to reproduce these scores myself, I get different numbers.

import lightgbm as lgb
from sklearn.datasets import load_breast_cancer
import pandas as pd
from sklearn.metrics import mean_absolute_error
from sklearn.model_selection import KFold


data = load_breast_cancer()
X = pd.DataFrame(data.data)
y = pd.Series(data.target)

lgb_params =  {'boosting_type': 'dart', 'random_state': 42}

folds = KFold(5)

for train_idx, val_idx in folds.split(X):
    X_train, X_valid = X.iloc[train_idx], X.iloc[val_idx]
    y_train, y_valid = y.iloc[train_idx], y.iloc[val_idx]
    model = lgb.LGBMRegressor(**lgb_params, n_estimators=10000, n_jobs=-1)
    model.fit(X_train, y_train,
              eval_set=[(X_valid, y_valid)],
              eval_metric='mae', verbose=-1, early_stopping_rounds=200)
    y_pred_valid = model.predict(X_valid)
    print(mean_absolute_error(y_valid, y_pred_valid))

I was expecting that

valid_0's l1: 0.123608

would match my own calculation from mean_absolute_error, but it doesn't. Indeed, here is the top part of my output:

Training until validation scores don't improve for 200 rounds.
Early stopping, best iteration is:
[631]	valid_0's l2: 0.0515033	valid_0's l1: 0.123608
0.16287265537021847

[guolinke]

dart cannot be used with early stopping.
refer to #1893