def testProject(self):
k = 2
pca = PrimalPCA(k)
U, lmbdas = pca.learnModel(self.X)
newX = pca.project(self.X)
self.assertEquals(newX.shape[0], self.X.shape[0])
self.assertEquals(newX.shape[1], k)
def testProject(self):
k = 2
pca = PrimalPCA(k)
U, lmbdas = pca.learnModel(self.X)
newX = pca.project(self.X)
self.assertEquals(newX.shape[0], self.X.shape[0])
self.assertEquals(newX.shape[1], k)
def testLearnModel(self):
k = 4
tol = 10**-6
pca = PrimalPCA(k)
U, lmbdas = pca.learnModel(self.X)
#Compute eignvalues manually
C = numpy.dot(self.X.T, self.X)
lmbdas2, U2 = scipy.linalg.eig(C)
inds = numpy.flipud(numpy.argsort(lmbdas2))
lmbdas2 = lmbdas2[inds]
U2 = U2[:, inds]
self.assertTrue( (lmbdas >= 0).all())
self.assertEquals(lmbdas.shape[0], self.numFeatures )
self.assertEquals(U.shape[1], self.numFeatures )
self.assertTrue( numpy.linalg.norm(lmbdas2- lmbdas) <= tol )
self.assertTrue( numpy.linalg.norm(U2- U) <= tol )
newX = pca.project(self.X)
newX2 = numpy.dot(self.X, U2[:, 0:k])
self.assertTrue( numpy.linalg.norm(newX2 - newX) <= tol )
#Test orthogonality of U
self.assertTrue( numpy.linalg.norm(numpy.dot(U.T, U) - numpy.eye(self.numFeatures)) <= tol )
#Test projecting just k directions
k = 2
pca.setK(k)
newX = pca.project(self.X)
self.assertTrue( numpy.linalg.norm(newX2[:, 0:k] - newX) <= tol)
def testLearnModel(self):
k = 4
tol = 10**-6
pca = PrimalPCA(k)
U, lmbdas = pca.learnModel(self.X)
#Compute eignvalues manually
C = numpy.dot(self.X.T, self.X)
lmbdas2, U2 = scipy.linalg.eig(C)
inds = numpy.flipud(numpy.argsort(lmbdas2))
lmbdas2 = lmbdas2[inds]
U2 = U2[:, inds]
self.assertTrue((lmbdas >= 0).all())
self.assertEquals(lmbdas.shape[0], self.numFeatures)
self.assertEquals(U.shape[1], self.numFeatures)
self.assertTrue(numpy.linalg.norm(lmbdas2 - lmbdas) <= tol)
self.assertTrue(numpy.linalg.norm(U2 - U) <= tol)
newX = pca.project(self.X)
newX2 = numpy.dot(self.X, U2[:, 0:k])
self.assertTrue(numpy.linalg.norm(newX2 - newX) <= tol)
#Test orthogonality of U
self.assertTrue(
numpy.linalg.norm(numpy.dot(U.T, U) -
numpy.eye(self.numFeatures)) <= tol)
#Test projecting just k directions
k = 2
pca.setK(k)
newX = pca.project(self.X)
self.assertTrue(numpy.linalg.norm(newX2[:, 0:k] - newX) <= tol)