Arrays used as indices must be of integer (or boolean) type

Question

Errors are like this:

Traceback (most recent call last):   File "NearestCentroid.py", line 53, in <module>     clf.fit(X_train.todense(),y_train)   File "/usr/local/lib/python2.7/dist-packages/scikit_learn-0.13.1-py2.7-linux-i686.egg/sklearn/neighbors/nearest_centroid.py", line 115, in fit     variance = np.array(np.power(X - self.centroids_[y], 2)) IndexError: arrays used as indices must be of integer (or boolean) type 

Codes are like this:

distancemetric=['euclidean','l2'] for mtrc in distancemetric: for shrkthrshld in [None]: #shrkthrshld=0 #while (shrkthrshld <=1.0):     clf = NearestCentroid(metric=mtrc,shrink_threshold=shrkthrshld)     clf.fit(X_train.todense(),y_train)     y_predicted = clf.predict(X_test.todense()) 

I am using scikit-learn package, X-train, y_train are in LIBSVM format, X is the feature:value pair, y_train is the target/label, X_train is in CSR matric format, the shrink_threshold does not support CSR sparse matrix, so I add .todense() to X_train, then I got this error, could anyone help me fix this? Thanks a lot!

Answer 1

I had a similar problem using the Pystruct pystruct.learners.OneSlackSSVM.

It occured because my training labels were floats, in stead of integers. In my case, it was because I initialized the labels with np.ones, without specifying dtype=np.int8. Hope it helps.

Answer 2

It happens quite often that an indexing array should be clearly integer type by the way it is created, but in the case of empty list passed, becomes default float, a case which might not be considered by the programmer. For example:

>>> np.array(xrange(1)) >>> array([0])                #integer type as expected >>> np.array(xrange(0)) >>> array([], dtype=float64)  #does not generalize to the empty list 

Therefore, one should always explicitely define the dtype in the array constructor.