xor_pymvpa.py

import numpy as np
import matplotlib.pyplot as plt
import mvpa.suite as ms

# create xor dataset
samples = np.random.rand(100,2)-0.5
targets = np.sign(samples)
targets = targets[:,0] == targets[:,1]

# display samples
idx = np.nonzero(targets==1)
plt.plot(samples[idx[0],0], samples[idx[0],1],'rs')
idx = np.nonzero(targets==0)
plt.plot(samples[idx[0],0], samples[idx[0],1],'bs')

# create dataset
ds = ms.dataset_wizard(samples, targets=targets, chunks=range(samples.shape[0]))

# choose classifier
clf = ms.kNN() # SMLR, LinearCSVMC, etc.,.
terr = ms.TransferError(clf)

# cross-validation with splitter
cvte = ms.CrossValidatedTransferError(terr, splitter=ms.HalfSplitter(),
                                      enable_ca=['confusion'])
res = cvte(ds)
np.mean(res) # mean cv error

print cvte.ca.confusion
plt.figure()
cvte.ca.confusion.plot()
plt.show()

# different splitter - leave one out
cvte = ms.CrossValidatedTransferError(terr, splitter=ms.NFoldSplitter(),
                                      enable_ca=['confusion'])
res = cvte(ds)
np.mean(res) # mean cv error

print cvte.ca.confusion
plt.figure()
cvte.ca.confusion.plot()
plt.show()

# cycle through all binary svm classifiers
for clf in ms.clfswh['binary','svm']:
    terr = ms.TransferError(clf)
    cvte = ms.CrossValidatedTransferError(terr,
                                          splitter=ms.HalfSplitter(npertarget='equal',
                                                                   nrunspersplit=10),
                                          enable_ca=['confusion'])
    try:
        res = cvte(ds)
        np.mean(res) # mean cv error
        #print cvte.ca.confusion
        #plt.figure()
        #cvte.ca.confusion.plot()
        #plt.suptitle(str(clf))
        print str(clf), cvte.ca.confusion.error
        print cvte.ca.confusion.matrix
        if cvte.ca.confusion.error < error:
            error = cvte.ca.confusion.error
            bestclf = clf.clone()
    except:
        print "could not run classifier: %s" % str(clf)


terr = ms.TransferError(bestclf)
cvte = ms.CrossValidatedTransferError(terr,
                                      splitter=ms.HalfSplitter(npertarget='equal',
                                                               nrunspersplit=10),
                                      enable_ca=['confusion'])

res = cvte(ds)
np.mean(res) # mean cv error
print cvte.ca.confusion
plt.figure()
cvte.ca.confusion.plot(numbers=True)
plt.suptitle(cvte.transerror.clf)