Different prediction of tree classifier between 32 bits and 64 bits #8853
Description
Description
While having some test failing in imbalanced-learn with a 32 bits machine (we are kinda poor ;)), we got to the point that the DecisionTreeClassifier can return different results between a 32 and a 64 architecture.
Steps/Code to Reproduce
import numpy as np
from sklearn.tree import DecisionTreeClassifier
RND_SEED = 0
X = np.array([[-0.3879569, 0.6894251], [-0.09322739, 1.28177189],
[-0.77740357, 0.74097941], [0.91542919, -0.65453327],
[-0.03852113, 0.40910479], [-0.43877303, 1.07366684],
[-0.85795321, 0.82980738], [-0.18430329, 0.52328473],
[-0.30126957, -0.66268378], [-0.65571327, 0.42412021],
[-0.28305528, 0.30284991], [0.20246714, -0.34727125],
[1.06446472, -1.09279772], [0.30543283, -0.02589502],
[-0.00717161, 0.00318087]])
y = np.array([0, 1, 1, 0, 1, 1, 1, 1, 1, 0, 1, 1, 0, 0, 0])
train_index = [0, 1, 2, 3, 6, 7, 8, 10, 11, 12, 13, 14]
test_index = [4, 5, 9]
clf = DecisionTreeClassifier(random_state=0)
clf.fit(X[train_index], y[train_index])
proba = clf.predict_proba(X[test_index])
print(proba) Expected Results
This is the result obtained in the 64 bits machine:
[[ 1. 0.]
[ 0. 1.]
[ 1. 0.]]
Actual Results
This is the result obtained for the 32 bits machine:
[[ 1. 0.]
[ 0. 1.]
[ 0. 1.]]
Versions
Ubuntu 16.04 - 32/64 bits
Python 2.7.13
numpy(1.12.1) ; scipy (0.19) ; scikit-learn (0.18.1) installed from conda