Suggestion to Have multiclass.py allow prediction over one sample only ! #5135

soufanom · 2015-08-18T21:10:04Z

Greetings Guys,

I came through the contributed implementation to multiclass.py in Scikit-learn. I just have a suggestion for you to consider the case when only one testing sample is passed to decision_function "Decision function for the OneVsOneClassifier". As for the current implementation, an undesirable output comes since n_samples = X.shape[0] will take a number larger than one when X is only a single list vector with some values. I may suggest you check the shape of X before parsing it in a particular way, or update the documentation to advise the user on a suggested way to get the prediction for one testing sample.

In a sense, it is true to say that usually, there is a testing set of many samples but in a specific case of mine, it was preferable to predict sample by sample. I overcome this by using X[0:1,:] instead of X[0,:] where X is a testing set of several samples.

The sklearn version I have installed is 0.16.1

I did not get an error when inputing 1d X and what I receive back are predictions as many as the length of this 1d list.

For example:

from sklearn import datasets
from sklearn.multiclass import OneVsOneClassifier
from sklearn.svm import LinearSVC
iris = datasets.load_iris()
X, y = iris.data, iris.target
OneVsOneClassifier(LinearSVC(random_state=0)).fit(X, y).predict(X[1,:])
Out[1]: array([0, 1, 1, 1])

And by replacing X[1,:] to be X[1:2,:] which in terms of values are the same:

OneVsOneClassifier(LinearSVC(random_state=0)).fit(X, y).predict(X[1:2,:])
Out[2]: array([0]) # Proper output

Regards,
Othman

amueller · 2015-08-18T21:24:58Z

linearSVC should really complain here :-/ I guess the decision function is called... hum... Not sure we can actually fix it without #4511.

amueller · 2015-09-09T14:52:38Z

This is somewhat fixed in #5152. But the current behavior is still non-sensical. I think we should put a stop-gap in master for the next release.

ogrisel · 2015-09-10T07:48:04Z

But the current behavior is still non-sensical. I think we should put a stop-gap in master for the next release.

What stopgap do you have in mind?

On the current master we now have:

>>> OneVsOneClassifier(LinearSVC(random_state=0)).fit(X, y).predict(X[1,:])
/volatile/ogrisel/code/scikit-learn/sklearn/utils/validation.py:372: DeprecationWarning: Passing 1d arrays as data is deprecated and will be removed in 0.18. Reshape your data either usingX.reshape(-1, 1) if your data has a single feature orX.reshape(1, -1) if it contains a single sample.
  DeprecationWarning)
/volatile/ogrisel/code/scikit-learn/sklearn/utils/validation.py:372: DeprecationWarning: Passing 1d arrays as data is deprecated and will be removed in 0.18. Reshape your data either usingX.reshape(-1, 1) if your data has a single feature orX.reshape(1, -1) if it contains a single sample.
  DeprecationWarning)
/volatile/ogrisel/code/scikit-learn/sklearn/utils/validation.py:372: DeprecationWarning: Passing 1d arrays as data is deprecated and will be removed in 0.18. Reshape your data either usingX.reshape(-1, 1) if your data has a single feature orX.reshape(1, -1) if it contains a single sample.
  DeprecationWarning)
/volatile/ogrisel/code/scikit-learn/sklearn/utils/validation.py:372: DeprecationWarning: Passing 1d arrays as data is deprecated and will be removed in 0.18. Reshape your data either usingX.reshape(-1, 1) if your data has a single feature orX.reshape(1, -1) if it contains a single sample.
  DeprecationWarning)
/volatile/ogrisel/code/scikit-learn/sklearn/utils/validation.py:372: DeprecationWarning: Passing 1d arrays as data is deprecated and will be removed in 0.18. Reshape your data either usingX.reshape(-1, 1) if your data has a single feature orX.reshape(1, -1) if it contains a single sample.
  DeprecationWarning)
/volatile/ogrisel/code/scikit-learn/sklearn/utils/validation.py:372: DeprecationWarning: Passing 1d arrays as data is deprecated and will be removed in 0.18. Reshape your data either usingX.reshape(-1, 1) if your data has a single feature orX.reshape(1, -1) if it contains a single sample.
  DeprecationWarning)
array([0])
>> OneVsOneClassifier(LinearSVC(random_state=0)).fit(X, y).predict(X[1:2,:])
array([0])

This looks fine to me, right?

ogrisel · 2015-09-10T07:56:20Z

BTW I fixed the deprecation message in 250507f.

amueller · 2015-09-10T14:19:18Z

It seems this got fixed somewhere between 0.16.1 and now. I don't know where, though, it is a bit odd. But let's close it.

I'm surprised by the number of deprecation warnings. I would have expected two, not six.

amueller added Bug Easy Well-defined and straightforward way to resolve Need Contributor labels Sep 9, 2015

amueller added this to the 0.17 milestone Sep 9, 2015

amueller closed this as completed Sep 10, 2015

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Suggestion to Have multiclass.py allow prediction over one sample only ! #5135

Suggestion to Have multiclass.py allow prediction over one sample only ! #5135

Suggestion to Have multiclass.py allow prediction over one sample only ! #5135

Suggestion to Have multiclass.py allow prediction over one sample only ! #5135

Comments