Optuna: LightGBMTunerCV seems to not handle user specified CV-folds

[bjoernholzhauer]

Expected behavior

When users specify training and validation folds in the manner that the basic lightgbm.cv function accepts, this should (from what I understand work)

Environment

• Optuna version: 2.0.0
• Python version: 3.6.9
• OS: Google Collab/Linux
• (Optional) Other libraries and their versions: LightGBM 2.3.1

Error messages, stack traces, or logs

0%|          | 0/7 [00:00<?, ?it/s]
feature_fraction, val_score: inf:   0%|          | 0/7 [00:00<?, ?it/s][W 2020-08-24 15:41:09,973] Trial 0 failed because of the following error: ValueError('For early stopping, at least one dataset and eval metric is required for evaluation',)
Traceback (most recent call last):
  File "/usr/local/lib/python3.6/dist-packages/optuna/study.py", line 709, in _run_trial
    result = func(trial)
  File "/usr/local/lib/python3.6/dist-packages/optuna/integration/_lightgbm_tuner/optimize.py", line 302, in __call__
    cv_results = lgb.cv(self.lgbm_params, self.train_set, **self.lgbm_kwargs)
  File "/usr/local/lib/python3.6/dist-packages/lightgbm/engine.py", line 576, in cv
    evaluation_result_list=res))
  File "/usr/local/lib/python3.6/dist-packages/lightgbm/callback.py", line 221, in _callback
    _init(env)
  File "/usr/local/lib/python3.6/dist-packages/lightgbm/callback.py", line 191, in _init
    raise ValueError('For early stopping, '
ValueError: For early stopping, at least one dataset and eval metric is required for evaluation
---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-18-0ec8edbe946a> in <module>()
      2                      label = np.array( data['target'] ).flatten())
      3 tuner = lgb.LightGBMTunerCV(params, dtrain, verbose_eval=100, early_stopping_rounds=100, folds=folds)
----> 4 tuner.run()

10 frames
/usr/local/lib/python3.6/dist-packages/lightgbm/callback.py in _init(env)
    189             return
    190         if not env.evaluation_result_list:
--> 191             raise ValueError('For early stopping, '
    192                              'at least one dataset and eval metric is required for evaluation')
    193 

ValueError: For early stopping, at least one dataset and eval metric is required for evaluation

As well as (on the second version without early stopping, which I think is an issue that's already reported in another issue?):

0%|          | 0/7 [00:00<?, ?it/s]


feature_fraction, val_score: inf:   0%|          | 0/7 [00:00<?, ?it/s][W 2020-08-24 15:42:03,262] Trial 0 failed because of the following error: KeyError('l1-mean',)
Traceback (most recent call last):
  File "/usr/local/lib/python3.6/dist-packages/optuna/study.py", line 709, in _run_trial
    result = func(trial)
  File "/usr/local/lib/python3.6/dist-packages/optuna/integration/_lightgbm_tuner/optimize.py", line 304, in __call__
    val_scores = self._get_cv_scores(cv_results)
  File "/usr/local/lib/python3.6/dist-packages/optuna/integration/_lightgbm_tuner/optimize.py", line 294, in _get_cv_scores
    val_scores = cv_results["{}-mean".format(metric)]
KeyError: 'l1-mean'
---------------------------------------------------------------------------
KeyError                                  Traceback (most recent call last)
<ipython-input-21-942a30076787> in <module>()
      1 tuner = lgb.LightGBMTunerCV(params, dtrain, verbose_eval=100, folds=folds)
----> 2 tuner.run()

8 frames
/usr/local/lib/python3.6/dist-packages/optuna/integration/_lightgbm_tuner/optimize.py in _get_cv_scores(self, cv_results)
    292 
    293         metric = self._get_metric_for_objective()
--> 294         val_scores = cv_results["{}-mean".format(metric)]
    295         return val_scores
    296 

KeyError: 'l1-mean'

Steps to reproduce

1. Get a Google Collab, then run the code below (extra installations beyond default collab explicitly via !pip below)

Reproducible examples (optional)

!pip install lightgbm==2.3.1
import numpy as np
import pandas as pd
from sklearn.model_selection import GroupKFold

import lightgbm as lgb
lgb.__version__
np.random.seed(123)
data = pd.DataFrame({'var1': np.random.normal(loc=0, scale=1, size=100),
                    'var2': np.random.normal(loc=0, scale=1, size=100),
                    'var3': np.random.normal(loc=0, scale=1, size=100),
                     'testfold': np.random.choice(a=np.repeat([x for x in range(5)], 20), size=100, replace=False)})
data['target'] = 7 + 0.1*data['var1'] + 1.0*data['var2'] + 5.0*data['var3'] - 2.0*data['var1']*data['var2'] + np.random.normal(loc=0, scale=0.5, size=100)
data.head()
params = {
    'objective': 'l1',
    'metric': 'l1',    
    "verbosity": -1,
    "boosting_type": "gbdt",
    'seed': 1979
    }

dtrain = lgb.Dataset(data= np.array( data[ ['var1', 'var2', 'var3'] ] ),
                     label = np.array( data['target'] ).flatten())

folds = GroupKFold().split(np.array( data[ ['var1', 'var2', 'var3'] ] ),
                            np.array( data['target'] ).flatten(), 
                            np.array(data['testfold']).flatten())
lgb.cv(params, dtrain, folds=folds, verbose_eval=100) # This is how base lightgbm does this, and it works fine


!pip install optuna
import optuna.integration.lightgbm as lgb

dtrain = lgb.Dataset(data= np.array( data[ ['var1', 'var2', 'var3'] ] ),
                     label = np.array( data['target'] ).flatten())
tuner = lgb.LightGBMTunerCV(params, dtrain, verbose_eval=100, early_stopping_rounds=100, folds=folds)
tuner.run()

tuner = lgb.LightGBMTunerCV(params, dtrain, verbose_eval=100, folds=folds)
tuner.run()

Additional context (optional)

Same issue in Kaggle kernels, but thought it would be easier to share a simplified Collab example.

[toshihikoyanase]

Thank you for your bug report. I'm not aware of the first issue. I'll investigate it.

As well as (on the second version without early stopping, which I think is an issue that's already reported in another issue?):

I think it is the same issue as #1602. The cause is the lack of the metric mapping in LightGBMTunerCV and @thigm85 is working on it.

[bjoernholzhauer]

Yes, you are right, #1602 was indeed what I had seen before (but did not find, again).

[github-actions[bot]]

This issue has not seen any recent activity.