GaussianClassifier fails with TransformedImage object

[tboggs]

When a GaussianClassifier is created and the classify_image method is called on a TransformedImage object, and AttributeError is raised because the object does not have a reshape method (classify_image is expecting a numpy.ndarray object). The exception can be reproduced as follows:

>>> from spectral import *
>>> img = open_image('92AV3C.lan')
>>> gt = open_image('92AV3GT.GIS').read_band(0)
>>> data = img.load()
>>> pc = principal_components(data)
>>> pc_0999 = pc.reduce(fraction=0.999)
>>> img_pc = pc_0999.transform(img)
>>> classes = create_training_classes(img_pc, gt)
>>> gmlc = GaussianClassifier(classes)
>>> clmap = gmlc.classify_image(img_pc)
AttributeError                            Traceback (most recent call last)
<ipython-input-11-6488a72f9e73> in <module>()
----> 1 clmap = gmlc.classify_image(img_pc)
/home/thomas/src/spectral/spectral/algorithms/classifiers.pyc in classify_image(self, image)
    200         status.display_percentage('Processing...')
    201         shape = image.shape
--> 202         image = image.reshape(-1, shape[-1])
    203         scores = np.empty((image.shape[0], len(self.classes)), np.float64)
    204         delta = np.empty_like(image, dtype=np.float64)

AttributeError: 'TransformedImage' object has no attribute 'reshape'

This will also occur with the MahalanobisDistanceClassifier. The bug can be avoided by using the load method of the TransformedImage object (img_pc in the example above) to load the data into an ndarray (if sufficient memory is available). This bug will be fixed by using the parent class method Classifier.classify_image when the argument is not an ndarray.