Issue #1492.

To reproduce the bug:

• Train a fully convolutional network (e.g SqueezeNet 1.1)
• Run "classify many" with a list of images

Depending on the version of numpy you will get a warning or an error.

With numpy 1.11 you get this warning and the confusion matrix has all the results on the first column.

VisibleDeprecationWarning: converting an array with ndim > 0 to an index will result in an error in the future
  result.append((labels[i], round(100.0 * scores[image_index, i], 2)))

With numpy 1.12 you get an error at the same line and no confusion matrix is shown.

The problem is the shape of "scores" that must be (number_of_images, number_of_classes), but with FCNs the shape has additional dimensions. e.g. SqueezeNet v1.1 final shape is (number_of_images, number_of_classes, 1, 1).