def test_random_rot():
dim = 20
tlen = 10
for i in xrange(tlen):
x = utils.random_rot(dim, dtype='f')
assert x.dtype.char=='f', 'Wrong dtype'
y = utils.mult(x.T, x)
assert_almost_equal(numx_linalg.det(x), 1., 4)
assert_array_almost_equal(y, numx.eye(dim), 4)
def test_symeig_fake_LAPACK_bug():
# bug. when input matrix is almost an identity matrix
# but not exactly, the lapack dgeev routine returns a
# matrix of eigenvectors which is not orthogonal.
# this bug was present when we used numx_linalg.eig
# instead of numx_linalg.eigh .
# Note: this is a LAPACK bug.
y = numx_rand.random((4,4))*1E-16
y = (y+y.T)/2
for i in xrange(4):
y[i,i]=1
val, vec = utils._symeig._symeig_fake(y)
assert_almost_equal(abs(numx_linalg.det(vec)), 1., 12)
def _stop_training(self):
self.labels = self._cov_objs.keys()
self.labels.sort()
nitems = 0
for lbl in self.labels:
cov, mean, p = self._cov_objs[lbl].fix()
nitems += p
self._sqrt_def_covs.append(numx.sqrt(numx_linalg.det(cov)))
if self._sqrt_def_covs[-1] == 0.0:
err = "The covariance matrix is singular for at least " "one class."
raise mdp.NodeException(err)
self.means.append(mean)
self.p.append(p)
self.inv_covs.append(utils.inv(cov))
for i in range(len(self.p)):
self.p[i] /= float(nitems)
del self._cov_objs
def _stop_training(self):
self.labels = self._cov_objs.keys()
self.labels.sort()
nitems = 0
for lbl in self.labels:
cov, mean, p = self._cov_objs[lbl].fix()
nitems += p
self._sqrt_def_covs.append(numx.sqrt(numx_linalg.det(cov)))
if self._sqrt_def_covs[-1] == 0.0:
err = ("The covariance matrix is singular for at least "
"one class.")
raise mdp.NodeException(err)
self.means.append(mean)
self.p.append(p)
self.inv_covs.append(utils.inv(cov))
for i in range(len(self.p)):
self.p[i] /= float(nitems)
del self._cov_objs
