Unexpected prediction from SVM

[aymanosman]

I get unexpected predictions from the SVM module.

This example is taken from the docs:

x = Nx.tensor([[1.0, 2.0], [3.0, 2.0], [4.0, 7.0]])
y = Nx.tensor([1, 0, 1])
model = Scholar.Linear.SVM.fit(x, y, num_classes: 2)
Scholar.Linear.SVM.predict(model, x)

#Nx.Tensor<
  s32[3]
  [1, 1, 1]
>

I expected to see [1, 0, 1].

Here is a livebook demonstrating the behaviour https://gist.github.com/aymanosman/aa20b3832148260d257cf43e2d497cc4.

[josevalim]

Why do you believe this behaviour is expected? Looking at Python, I get the same behaviour:

>>> x = [[1.0, 2.0], [3.0, 2.0], [4.0, 7.0]]
>>> y = [1, 0, 1]
>>> clf = svm.SVC()
>>> clf.fit(x, y)
SVC()
>>> clf.predict(x)
array([1, 1, 1])

I think the small sample size and the unbalanced classes are not being helpful to draw the correct planes.

[aymanosman]

Why do you believe this behaviour is expected? Looking at Python, I get the same behaviour:

Because I am a n00b.

[aymanosman]

Although...

from sklearn.linear_model import SGDClassifier
x = [[1.0, 2.0], [3.0, 2.0], [4.0, 7.0]]
y = [1, 0, 1]
clf = SGDClassifier()
clf.fit(x, y)
print(clf.predict(x))
# => [1 0 1]