Exception: 'access violation reading' while training with init_model

[MaxtoqOV]

Hi, I get an 'OSError: exception: access violation reading 0x000001C0D2851338' while using lightgbm.train with an already existing Booster as 'init_model' (see code below).
The error occurs after a few training iterations. The number of iterations isn't always the same: I've had 4, 7 and 3598, and sometimes it works well... completely random.
I had never seen this error before starting using the 'init_model' feature so I assume there's a bug with it.

Environment info

Operating System: Windows 10
CPU/GPU model: Intel i7 6700HQ
C++/Python/R version: Python 3.7.1
LightGBM version or commit hash: 2.2.3

Error message

OSError Traceback (most recent call last)
<ipython-input-197-4c6e4d4696e6> in <module>
29 valid_sets=lgb_eval,
30 early_stopping_rounds=100,
---> 31 init_model=gbm
32 )
33
~\Anaconda3\lib\site-packages\lightgbm\engine.py in train(params, train_set, num_boost_round, valid_sets, valid_names, fobj, feval, init_model, feature_name, categorical_feature, early_stopping_rounds, evals_result, verbose_eval, learning_rates, keep_training_booster, callbacks)
216 evaluation_result_list=None))
217
--> 218 booster.update(fobj=fobj)
219
220 evaluation_result_list = []
~\Anaconda3\lib\site-packages\lightgbm\basic.py in update(self, train_set, fobj)
1800 _safe_call(_LIB.LGBM_BoosterUpdateOneIter(
1801 self.handle,
-> 1802 ctypes.byref(is_finished)))
1803 self._is_predicted_cur_iter = [False for _ in range(self.__num_dataset)]
1804 return is_finished.value == 1
OSError: exception: access violation reading 0x000001C0D2851338

Reproducible examples

`gbm = None
for i in (0.4, 0.3, 0.2):
train_X, test_X, train_y, test_y = train_test_split(X, y, test_size=i, shuffle=False)

# Create LGB datasets
lgb_train = lgb.Dataset(train_X, train_y)
lgb_eval = lgb.Dataset(test_X, test_y, reference=lgb_train)

params = {
    'learning_rate': 0.005,
    'boosting_type': 'dart',
    'objective': 'regression',
    'metric': {'l2', 'l1'},
    'feature_fraction': 0.8,
    'bagging_fraction': 0.8,
    'bagging_freq': 5,
    'max_bin': 1000,
    'num_leaves': 100
}

# Train the model
gbm = lgb.train(
    params, 
    lgb_train, 
    num_boost_round=10000, 
    valid_sets=lgb_eval, 
    early_stopping_rounds=100, 
    init_model=gbm
)`

Steps to reproduce

see code.

[guolinke]

@MaxtoqOV did you try it without dart?

[MaxtoqOV]

@guolinke I've just tried with the other available algorithms and the error doesn't occur.

[MaxtoqOV]

Also I noticed that early stopping does not work with dart. I don't know if that's normal, I didn't see anything about that in the doc.

[guolinke]

for early stopping, refer to #1893, I think some documents is added.

[guolinke]

@MaxtoqOV did you train the init_model by the dart as well?

[MaxtoqOV]

@guolinke Yes, as you can see in the code (despite the display bug, sorry about that), the booster is created during the first iteration of the loop. Thus the init_model has the exact same parameters than the following ones.

[guolinke]

indeed, I find a bug, fixes are in #2251, can you have a try?

[MaxtoqOV]

Thanks for the quick fix. It seems to be working well now.