Unable to pickle ensemble model that includes thundersvm (SVC)

[tigertimwu]

Hi,

I was trying to pickle an EnsembleVoteClassifier using mlxtend where one of the classfier included is thundersvm. However I am unable to pickle the EnsembleVoteClassifier using pickle.dump() as long as the thundersvm classifier is included, the following error comes out:

PicklingError: Can't pickle <class 'thundersvm.thundersvm.c_float_Array_629'>: attribute lookup c_float_Array_629 on thundersvm.thundersvm failed.

Is there any way that I can save down the ensemble classifier that includes thundersvm?

Thank you

[zeyiwen]

Would you provide a minimal example to help us reproduce the problem?

[tigertimwu]

clf_1=SVC() clf_1.load_from_file(path=svm_path+clf1_file_name)

clf_2=pickle.load(open(xg_boost_path+clf2_file_name, 'rb'))

pipe1 = make_pipeline(ColumnSelector(cols=df1.columns.tolist()), clf_1)

pipe2 = make_pipeline(ColumnSelector(cols=df2.columns.tolist()), clf_2)

eclf = EnsembleVoteClassifier(clfs=[pipe1,pipe2], weights=[1,1], refit=False,voting='hard',verbose=2)

labels = ['svm','xgboost','ensemble']

eclf.fit(X_train_final, y_train)

model_name=open('D:\ensemble_test.model', 'wb')

pickle.dump(eclf, model_name) model_name.close()

The error occurs at the line pickle.dump(eclf, model_name), clf_1 is thundersvm model, where clf2 is xgboost model, I have no issue to picke the eclf model (ie. the Ensemble model) if all classifiers in the ensemble are xgboost.

Thanks

[zeyiwen]

Thanks for that! We will look into it.

[CMobley7]

I'm experiencing this issue as well when I run

from thundersvm import SVC
import pickle
svm = SVC(probability=True, gamma="auto")
svm.fit(x_train, y_train)
pickled_svm = pickle.dumps(svm)

I get

_pickle.PicklingError: Can't pickle <class 'thundersvm.thundersvm.c_int_Array_1'>: attribute lookup c_int_Array_1 on thundersvm.thundersvm failed

[QinbinLi]

Hi,

ThunderSVM doesn't support pickle now due to ctype variables used in the code. We'll try to improve it in the future. Thanks.

[beevabeeva]

Also have this problem, when trying to implement thundersvm in an automated machine learning program. However, the maintainers of that program seem to be pushing to remove pickle from the framework in favour of JSON.

[ghasemikasra39]

same issue

[guillaumedsde]

Kindly bumping this issue :)

@beevabeeva who's pickling to JSON?

[civilinformer]

The simplest way to implement this is save the thundersvm parts of the ensemble separately usings its own save method, then replace the thundersvm parts from the ensemble object with dummy objects and save normally. To load invert the process, loading the individual models and the ensemble and placing the svms back into their respective places in the ensemble. I got that to work.

[zeyiwen]

This issue has been fixed by a contribution (#207 ) from @RichardWarfield. I would like to close this issue now. Please try the lastest version of ThunderSVM and let us know if there is any new issue @civilinformer

[civilinformer]

Is this available through pip?

[zeyiwen]

Yes. We have uploaded the wheel to pypi. Please try it and let us know if there is any problem.