Serialization error when using parallelism in cross_val_score with GridSearchCV and a custom estimator

Minimal example:

import numpy as np
from sklearn.model_selection import cross_val_score, GridSearchCV
from sklearn.base import ClassifierMixin, BaseEstimator

class Dummy(ClassifierMixin, BaseEstimator):
    def __init__(self, answer=1):
        self.answer = answer

    def fit(self, X, y=None):
        return self

    def predict(self, X):
        return np.ones(X.shape[0], dtype='int') * self.answer

n_samples, n_features = 500, 8
X = np.random.randn(n_samples, n_features)
y = np.random.randint(0, 2, n_samples)

dummy = Dummy()
gcv = GridSearchCV(dummy, {'answer': [0, 1]}, cv=5, iid=False, n_jobs=1)
cross_val_score(gcv, X, y, cv=5, n_jobs=5)

# BrokenProcessPool: A task has failed to un-serialize.
# Please ensure that the arguments of the function are all picklable.

Full traceback in details.

Interestingly, it does not fail when:

• calling cross_val_score with n_jobs=1.
• calling cross_val_score directly on dummy, without GridSearchCV.
• using a imported classifier, as LogisticRegression, or even the same Dummy custom classifier but imported from another file.

This is a joblib 0.12 issue, different from #12289 or #12389. @ogrisel @tomMoral

Traceback (most recent call last):
  File "/cal/homes/tdupre/work/src/joblib/joblib/externals/loky/process_executor.py", line 393, in _process_worker
    call_item = call_queue.get(block=True, timeout=timeout)
  File "/cal/homes/tdupre/miniconda3/envs/py36/lib/python3.6/multiprocessing/queues.py", line 113, in get
    return _ForkingPickler.loads(res)
AttributeError: Can't get attribute 'Dummy' on <module 'sklearn.externals.joblib.externals.loky.backend.popen_loky_posix' from '/cal/homes/tdupre/work/src/scikit-learn/sklearn/externals/joblib/externals/loky/backend/popen_loky_posix.py'>
'''

The above exception was the direct cause of the following exception:

BrokenProcessPool                         Traceback (most recent call last)
~/work/src/script_csc/condition_effect/test.py in <module>()
     32 
     33     # fails
---> 34     cross_val_score(gcv, X, y, cv=5, n_jobs=5)
     35     """
     36     BrokenProcessPool: A task has failed to un-serialize.

~/work/src/scikit-learn/sklearn/model_selection/_validation.py in cross_val_score(estimator, X, y, groups, scoring, cv, n_jobs, verbose, fit_params, pre_dispatch, error_score)
    384                                 fit_params=fit_params,
    385                                 pre_dispatch=pre_dispatch,
--> 386                                 error_score=error_score)
    387     return cv_results['test_score']
    388 

~/work/src/scikit-learn/sklearn/model_selection/_validation.py in cross_validate(estimator, X, y, groups, scoring, cv, n_jobs, verbose, fit_params, pre_dispatch, return_train_score, return_estimator, error_score)
    232             return_times=True, return_estimator=return_estimator,
    233             error_score=error_score)
--> 234         for train, test in cv.split(X, y, groups))
    235 
    236     zipped_scores = list(zip(*scores))

~/work/src/joblib/joblib/parallel.py in __call__(self, iterable)
    996 
    997             with self._backend.retrieval_context():
--> 998                 self.retrieve()
    999             # Make sure that we get a last message telling us we are done
   1000             elapsed_time = time.time() - self._start_time

~/work/src/joblib/joblib/parallel.py in retrieve(self)
    899             try:
    900                 if getattr(self._backend, 'supports_timeout', False):
--> 901                     self._output.extend(job.get(timeout=self.timeout))
    902                 else:
    903                     self._output.extend(job.get())

~/work/src/joblib/joblib/_parallel_backends.py in wrap_future_result(future, timeout)
    519         AsyncResults.get from multiprocessing."""
    520         try:
--> 521             return future.result(timeout=timeout)
    522         except LokyTimeoutError:
    523             raise TimeoutError()

~/miniconda3/envs/py36/lib/python3.6/concurrent/futures/_base.py in result(self, timeout)
    403                 raise CancelledError()
    404             elif self._state == FINISHED:
--> 405                 return self.__get_result()
    406             else:
    407                 raise TimeoutError()

~/miniconda3/envs/py36/lib/python3.6/concurrent/futures/_base.py in __get_result(self)
    355     def __get_result(self):
    356         if self._exception:
--> 357             raise self._exception
    358         else:
    359             return self._result

BrokenProcessPool: A task has failed to un-serialize. Please ensure that the arguments of the function are all picklable.

[icoxfog417]

I also have this problem. Python=3.6 and scikit-learn=0.20.2 I can skip the error by n_jobs=1.

[sanketchavan08]

Here I am sharing my complete code: `from keras.wrappers.scikit_learn import KerasClassifier

from sklearn.model_selection import cross_val_score

from keras.models import Sequential

from keras.layers import Dense

def build_classifier(): classifier = Sequential() classifier.add(Dense(units = 6, kernel_initializer = ‘uniform’, activation = ‘relu’, input_dim = 11)) classifier.add(Dense(units = 6, kernel_initializer = ‘uniform’, activation = ‘relu’)) classifier.add(Dense(units = 1, kernel_initializer = ‘uniform’, activation = ‘sigmoid’)) classifier.compile(optimizer = ‘adam’, loss = ‘binary_crossentropy’, metrics = [‘accuracy’]) return classifier

classifier = KerasClassifier(build_fn = build_classifier, batch_size = 10, epochs = 100)

accuracies = cross_val_score(estimator = classifier, X = X_train, y = y_train, cv = 10, n_jobs = -1)`

I have checked for both scikit-learn versions 0.20.1 and 0.20.2.