Mismatching dimensions in extmath.weighted_mode

[hgascon]

Apparently, this check here isn't enough.

When using neighbors.RadiousNeighborsClassifer.predict, I keep getting errors due to the mismatching dimensions in the next line :
ValueError: operands could not be broadcast together with shapes (8792) (8807)
ValueError: operands could not be broadcast together with shapes (3) (2)
...

[amueller]

Thanks for the report.
Can you give an example with synthetic data that is crashing?

[amueller]

Also, which version of scikit-learn are you using? (and sorry for the slow reply, I was sick)

[hgascon]

>>> sklearn.__version__
> '0.14-git'

If a.shape != w.shape and a.shape > w.shape there is no problem. But if a.shape < w.shape the arrays can't be broadcast in the next definition w = np.zeros(a.shape, dtype=w.dtype) + w

In [53]: X_train
Out[53]:
<10076x70068 sparse matrix of type '<type 'numpy.int64'>'
    with 1466563 stored elements in Compressed Sparse Column format>
In [54]: X_val
Out[54]:
<2520x70068 sparse matrix of type '<type 'numpy.int64'>'
    with 437226 stored elements in Compressed Sparse Column format>
In [55]: rnc = RadiusNeighborsClassifier(radius=30, weights='distance', outlier_label='unknown')
In [56]: rnc.fit(X_train, y_train)
In [57]: y_pred = rnc.predict(X_val)

/usr/local/lib/python2.7/site-packages/sklearn/utils/extmath.py in weighted_mode(a, w, axis)
    304
    305     if a.shape != w.shape:
--> 306         w = np.zeros(a.shape, dtype=w.dtype) + w
ValueError: operands could not be broadcast together with shapes (10032) (10050)

[amueller]

The code looks like it tries to handle different ndims, not different shapes.
I don't understand what the function is supposed to do if the shapes are different.
As I said, it would be great if you could reproduce with random data.

Also, I'm not sure what is happening in train here. Your X_train seems to have 70068 datapoints, while y_train only has four labels.

Could you try upgrading scikit-learn to 0.16? You use an old development version.

[amueller]

closing as no reply