See ticket 558 on scipy trac
Hi, I'm using the scipy wrapping of libSVM on a 64 bits architecture, at it seems that the way the weights are stored before being passed to libSVM is innacurate.
I managed to fix that problem on my architecture, by patching the init method of the class LibSvmClassificationModel?, and replacing the type of the items stored in the weight_labels array.
In other words, I had to replace the following line: self.weight_labels = N.empty((len(weights),), dtype=N.intp)
by this one: self.weight_labels = N.empty((len(weights),), dtype=N.int32)
I didn't checked if that patch leads to good results on other architectures, and if the bugg I'm talking about occured only on my architecture, or on all architecture.