scikit-learn: Error in using multi-label classification in partial_fit() in OvR

• StackOverflow Question: https://stackoverflow.com/questions/42280439/multi-label-out-of-core-learning-for-text-data-valueerror-on-partial-fit

Description

When using OneVsRestClassifier() with partial_fit() method, errors are thrown. When using fit(), no errors are thrown and everything works.

Steps/Code to Reproduce

from sklearn.naive_bayes import MultinomialNB
from sklearn.feature_extraction.text import HashingVectorizer
from sklearn.preprocessing import MultiLabelBinarizer
from sklearn.multiclass import OneVsRestClassifier
import numpy as np

categories = ['a','b','c']
X = ["This is a test", "This is another attempt", "And this is a test too!"]
Y = [['a', 'b'],['b', 'c'],['a', 'b']] 

mlb = MultiLabelBinarizer(classes=categories)
vectorizer = HashingVectorizer(decode_error='ignore', n_features=2 ** 18,         non_negative=True)
clf = OneVsRestClassifier(MultinomialNB(alpha=0.01))

X_train = vectorizer.fit_transform(X)
Y_train = mlb.fit_transform(Y)

- Case1
clf.partial_fit(X_train, Y_train, categories)
- Case2
clf.partial_fit(X_train, Y_train, mlb.transform(Y))

Description of code

• Case1 Using classes=categories without transforming partial_fit(X_train, Y_train, classes=categories)

    ValueError: The truth value of an array with more than one element is ambiguous. Use a.any() or a.all()
  

• Case2 Using classes=mlb.transform(categories) i.e. after transforming from same multilabelbinarizer partial_fit(X_train, Y_train, classes=mlb.transform(categories))

     ValueError: The object was not fitted with multilabel input.
  

Expected Results

No error is thrown as when using fit().

Actual Results

• Case1

Traceback (most recent call last): File “/path_to_module/Check.py”, line 18, in <module> clf.partial_fit(X_train, Y_train, categories) File “/library/python2.7/dist-packages/sklearn/multiclass.py”, line 260, in partial_fit if np.setdiff1d(y, self.classes_): ValueError: The truth value of an array with more than one element is ambiguous. Use a.any() or a.all()

• Case2

Traceback (most recent call last): File “/path_to_module/Check.py”, line 18, in <module> clf.partial_fit(X_train, Y_train, mlb.transform(Y)) File “/library/python2.7/dist-packages/sklearn/multiclass.py”, line 265, in partial_fit Y = self.label_binarizer_.transform(y) File “/library/python2.7/dist-packages/sklearn/preprocessing/label.py”, line 329, in transform raise ValueError(“The object was not fitted with multilabel” ValueError: The object was not fitted with multilabel input.

Observation

• In Case1, the error is because https://github.com/scikit-learn/scikit-learn/blob/master/sklearn/multiclass.py#L260 is returning an array of booleans whereas it expects a single boolean value hence the error. But couldnt find what to do about it. How to pass Y or classes into it.

• In Case2, the error occurs because partial_fit() calls the check_partial_fit_first_call() function which sets the clf.classes in a different way using unique_labels() as seen here https://github.com/scikit-learn/scikit-learn/blob/master/sklearn/utils/multiclass.py#L308. These unique labels are passed to clf.label_binarizer_ in this line https://github.com/scikit-learn/scikit-learn/blob/master/sklearn/multiclass.py#L258, which leads to it assuming type of targets as ‘multiclass’ whereas actual type of target is multilabel, hence this error. The fit() method handles classes_ in a different way (doesnt use unique_labels) and hence everything works correctly

Versions

Linux-3.16.0-77-generic-x86_64-with-Ubuntu-14.04-trusty (‘Python’, ‘2.7.6 (default, Oct 26 2016, 20:30:19) \n[GCC 4.8.4]’) (‘NumPy’, ‘1.12.0’) (‘SciPy’, ‘0.18.1’) (‘Scikit-Learn’, ‘0.18.1’)