Closed
Description
Description
LabelSpreading fails during predict when using a callable kernel that returns sparse matrix
sklearn.semi_supervised.LabelSpreading allows user to provide a callable as a kernel function.
However, if this callable returns a sparse matrix, then LabelSpreading.predict() will fail.
The root cause is that np.dot(sparse, dense) behaves differently than sparse.dot(dense), and does not give the intended result.
For example, you can try the following to see the issue with np.dot(sparse, dense):
>>> from scipy.sparse import csr_matrix
>>> a = csr_matrix([[1,0,0,0,0], [0,1,0,0,0],[0,0,1,0,0]])
>>> b = np.ones((5,8))
>>> a.dot(b).shape
(3, 8)
>>> np.dot(a, b).shape
(5, 8)
>>> a.dot(b)
array([[1., 1., 1., 1., 1., 1., 1., 1.],
[1., 1., 1., 1., 1., 1., 1., 1.],
[1., 1., 1., 1., 1., 1., 1., 1.]])
>>> np.dot(a, b)
array([[<3x5 sparse matrix of type '<class 'numpy.float64'>'
with 3 stored elements in Compressed Sparse Row format>,
<3x5 sparse matrix of type '<class 'numpy.float64'>'
...The fix is a one-liner: change np.dot(...) to A.dot(B) or to A @ B (whichever style is preferred) on the following line:
https://github.com/scikit-learn/scikit-learn/blob/master/sklearn/semi_supervised/_label_propagation.py#L198
Steps/Code to Reproduce
import numpy as np
from sklearn import datasets
from sklearn.semi_supervised import LabelSpreading
from scipy.sparse import csr_matrix
from sklearn.neighbors import NearestNeighbors
from sklearn.metrics import classification_report, confusion_matrix
# Our custom kernel is a sparse RBF kernel containing only the top K nearest neighbors
def topk_rbf(X, Y=None, n_neighbors=10, gamma=1e-5):
nn = NearestNeighbors(n_neighbors=10, metric='euclidean', n_jobs=-1).fit(X)
W = -1 * nn.kneighbors_graph(Y, mode='distance').power(2) * gamma
np.exp(W.data, out=W.data)
assert isinstance(W, csr_matrix)
return W.T
digits = datasets.load_digits()
rng = np.random.RandomState(2)
indices = np.arange(len(digits.data))
rng.shuffle(indices)
Xtrain = digits.data[indices[:1000]]
Ytrain = digits.target[indices[:1000]]
Xtest = digits.data[indices[1000:1100]]
Ytest = digits.target[indices[1000:1100]]
# The "transductive" learning phase happens during fit, but is not the concern here
# Therefore, none of the labels were masked
model = LabelSpreading(kernel=topk_rbf, n_jobs=-1).fit(Xtrain, Ytrain)
# Here, we try the "inductive" learning phase
predicted_labels = model.predict(Xtest)
print(f"Confusion matrix: {confusion_matrix(y_true=Ytest, y_pred=predicted_labels, labels=model.classes_)}")
print(f"Classification_report: {classification_report(y_true=Ytest, y_pred=predicted_labels)}")Expected Results
Confusion matrix: [[ 8 0 0 0 0 0 0 0 0 0]
[ 0 16 0 0 0 0 0 0 0 0]
[ 0 0 7 0 0 0 0 0 0 0]
[ 0 0 0 8 0 0 0 0 0 0]
[ 0 0 0 0 6 0 0 0 0 0]
[ 0 0 0 0 0 11 0 0 0 0]
[ 0 0 0 0 0 0 12 0 0 0]
[ 0 0 0 0 0 0 0 11 0 0]
[ 0 0 0 0 0 0 0 0 10 0]
[ 0 0 0 0 0 0 0 0 0 11]]
Classification_report: precision recall f1-score support
0 1.00 1.00 1.00 8
1 1.00 1.00 1.00 16
2 1.00 1.00 1.00 7
3 1.00 1.00 1.00 8
4 1.00 1.00 1.00 6
5 1.00 1.00 1.00 11
6 1.00 1.00 1.00 12
7 1.00 1.00 1.00 11
8 1.00 1.00 1.00 10
9 1.00 1.00 1.00 11
accuracy 1.00 100
macro avg 1.00 1.00 1.00 100
weighted avg 1.00 1.00 1.00 100
Actual Results
...
Traceback (most recent call last):
File "sklearn_bugreport_example.py", line 33, in <module>
predicted_labels = model.predict(Xtest)
File "/py37/lib/python3.7/site-packages/sklearn/semi_supervised/label_propagation.py", line 169, in predict
return self.classes_[np.argmax(probas, axis=1)].ravel()
File "<__array_function__ internals>", line 6, in argmax
File "/py37/lib/python3.7/site-packages/numpy/core/fromnumeric.py", line 1153, in argmax
return _wrapfunc(a, 'argmax', axis=axis, out=out)
File "/py37/lib/python3.7/site-packages/numpy/core/fromnumeric.py", line 61, in _wrapfunc
return bound(*args, **kwds)
File "/py37/lib/python3.7/site-packages/numpy/matrixlib/defmatrix.py", line 171, in __array_finalize__
if (isinstance(obj, matrix) and obj._getitem): return
SystemError: <built-in function isinstance> returned a result with an error set
Versions
import sklearn; sklearn.show_versions()
System:
python: 3.7.5rc1 (default, Oct 8 2019, 16:47:45) [GCC 9.2.1 20191008]
executable: /py37/bin/python3
machine: Linux-5.3.0-23-generic-x86_64-with-Ubuntu-19.10-eoan
Python deps:
pip: 19.3.1
setuptools: 41.6.0
sklearn: 0.21.3
numpy: 1.17.4
scipy: 1.3.2
Cython: None
pandas: 0.25.3Thanks!
Metadata
Metadata
Assignees
Labels
No labels