Using joblib to dump a scikit-learn model on x86 then read on z/OS passes in Decision Tree but fails on a GradientBoostingRegressor

Description

Joblib loading works when serializing a Decision Tree classifier on x86 machine(MacOS) and loading it on z/OS however it fails when loading a GradientBoostingRegressor which has been serialized in the same way. I have made small changes into NumpyArrayWrapper to detect endianness and if loading on an incorrect endian system it would call array.byteswap(). However the object returned by the fit method in GBR causes a different path to be taken through the code and the creation of the array fails before we can even call byteswap().

Steps/Code to Reproduce

Example of GBR dumping.

`from sklearn.ensemble import GradientBoostingRegressor
from sklearn.pipeline import Pipeline
from sklearn import datasets
from sklearn import metrics
from sklearn.tree import DecisionTreeClassifier

from sklearn.model_selection import train_test_split

iris = datasets.load_iris()
X_train, X_test, y_train, y_test = train_test_split(iris.data, iris.target, test_size=0.2)

gbr = GradientBoostingRegressor(max_depth=3)
model = Pipeline([('Gbr', gbr)])

model.fit(X_train, y_train)


from sklearn.externals import joblib
joblib.dump(model, 'GBTmodelx86.pkl')`

Example of DT model dumping

`from sklearn import datasets
from sklearn import metrics
from sklearn.tree import DecisionTreeClassifier

dataset = datasets.load_iris()

model = DecisionTreeClassifier()
model.fit(dataset.data, dataset.target)

from sklearn.externals import joblib
joblib.dump(model, 'DTmodelX86.pkl')`

Loading both models

`from sklearn.externals import joblib
model = joblib.load(<model.pkl>) `

Expected Results

No error in either.

Actual Results

numpy_pickle.py", line 118, in read_array array = pickle.load(unpickler.file_handle) File “sklearn/tree/_tree.pyx”, line 650, in sklearn.tree._tree.Tree._setstate (sklearn/tree/_tree.c:8403) ValueError: Did not recognise loaded array layout

Versions

z-OS-2.2-3906-64bit Python 3.6.1 (heads/v3.6.1-anaconda:aa4d638, May 18 2018, 12:10:40) [C] NumPy 1.12.1 SciPy 0.19.0 Scikit-Learn 0.18.1

Additional notes

I understand that different platforms are not supported, and I am not necessarily positive this is a scikit-learn issue, however pickle’s docs mention that it is platform independent and this issue seems to be fine if we don’t try to run the fit method on the gradient boost model.

[jnothman]

We don’t really support unpickling on different platforms, but if anyone diagnoses this further and finds a solution, we would consider a pull request to fix it.

If you are storing the model for prediction, I recommend ONNX

[NicolasHug]

Nothing to add to the issue unfortunately, but please be advised that importing from sklearn.externals isn’t recommended. We’ll stop vendoring joblib soon.