SVM can be tricked into running proba() 

It is possible to make `svm.SVC` run `svm_predict_probability()` without having been trained with `probability=True`. Example:

```
from sklearn import datasets, svm
I = datasets.load_iris()
X, Y = I.data, I.target

G = svm.SVC()
G.fit(X,Y)
G.predict_proba(X) #--->
"""
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/opt/local/Library/Frameworks/Python.framework/Versions/3.4/lib/python3.4/site-packages/sklearn/svm/base.py", line 542, in predict_proba
    self._check_proba()
  File "/opt/local/Library/Frameworks/Python.framework/Versions/3.4/lib/python3.4/site-packages/sklearn/svm/base.py", line 510, in _check_proba
    " probability=%r" % self.probability)
AttributeError: predict_proba is not available when probability=False
"""

#trick sklearn:
G.probability = True

# now it works:
G.predict_proba(X) #--->
"""
array([[  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   1.00000015e-07,   5.00000075e-08],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   1.00000015e-07,   5.00000075e-08],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   1.00000015e-07,   5.00000075e-08],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   1.00000015e-07,   5.00000075e-08],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   5.00000075e-08,   1.00000015e-07],
       [  9.99999850e-01,   1.00000015e-07,   5.00000075e-08],
       [  9.99999850e-01,   1.00000015e-07,   5.00000075e-08],
       [  9.99999850e-01,   1.00000015e-07,   5.00000075e-08],
       [  9.99999850e-01,   1.00000015e-07,   5.00000075e-08],
       [  9.99999850e-01,   1.00000015e-07,   5.00000075e-08],
       [  9.99999850e-01,   1.00000015e-07,   5.00000075e-08],
       [  9.99999850e-01,   1.00000015e-07,   5.00000075e-08],
       [  9.99999850e-01,   1.00000015e-07,   5.00000075e-08],
       [  9.99999850e-01,   1.00000015e-07,   5.00000075e-08],
       [  9.99999850e-01,   1.00000015e-07,   5.00000075e-08],
       [  9.99999850e-01,   1.00000015e-07,   5.00000075e-08],
       [  9.99999850e-01,   1.00000015e-07,   5.00000075e-08],
       [  9.99999850e-01,   1.00000015e-07,   5.00000075e-08],
       [  9.99999850e-01,   1.00000015e-07,   5.00000075e-08],
       [  9.99999850e-01,   1.00000015e-07,   5.00000075e-08],
       [  9.99999850e-01,   1.00000015e-07,   5.00000075e-08],
       [  9.99999850e-01,   1.00000015e-07,   5.00000075e-08],
       [  9.99999850e-01,   1.00000015e-07,   5.00000075e-08],
[...]
"""
```

(interestingly the probabilities do seem to correspond to reasonable predictions, but every class has a fixed set of probability values; I saw this when writing my own code against libsvm too; I don't know what to make of it, probably an upstream bug)

libsvm provides svm_check_probability_model() precisely to [catch this case](https://github.com/kousu/statasvm/blob/813b743aaa36d26f38ce53a9ef11a717caf4b4ad/src/_svm.c#L530).  It seems [you aren't](https://github.com/scikit-learn/scikit-learn/search?utf8=%E2%9C%93&q=svm_check_probability_model), but you could!


Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Uh oh!

SVM can be tricked into running proba() #4791

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Uh oh!

SVM can be tricked into running proba() #4791

Description

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions