def MakeModel(self):
oDict = dict(inferType='VB')
compA = dict(m=[100, 100], kappa=1.0e-4, invW=np.eye(2), dF=4)
compB = dict(m=[-100, -100], kappa=1.0e-4, invW=np.eye(2), dF=4)
compC = dict(m=[0, 0], kappa=1.0e-4, invW=np.eye(2), dF=4)
compDictList = [compA, compB, compC]
obsPrior = GaussWishDistr(m=[0, 0], kappa=1e-4, invW=np.eye(2), dF=4)
self.obsM = GaussObsModel.CreateWithAllComps(oDict, obsPrior,
compDictList)
class TestGaussWishDistr(object):
def setUp(self):
self.invW = np.eye(2)
self.m = np.zeros(2)
self.distr = GaussWishDistr(m=self.m, invW=self.invW, kappa=1.0, dF=4)
def test_dimension(self):
assert self.distr.D == self.invW.shape[0]
def test_entropyWish(self):
''' Verify that (wishart) entropy is same for this object and Wishart object
'''
Hself = self.distr.entropyWish()
wishDistr = WishartDistr(v=self.distr.dF, invW=self.distr.invW)
Hwish = wishDistr.get_entropy()
assert np.allclose(Hself, Hwish)
def test_dist_mahalanobis(self, N=10):
''' Verify that distance computation is largest at mean and decays further away
'''
Xlist = list()
for r in [0, 0.01, 0.1, 1, 2, 3, 4, 5]:
Xlist.append(self.distr.m + r)
X = np.asarray(Xlist)
Dist = self.distr.dist_mahalanobis(X)
print Dist
assert np.all(Dist[:-1] < Dist[1:])
def test_update_soVB(self, rho=0.25):
''' Verify the blend update for stochastic variational is correct
'''
distrB = copy.deepcopy(self.distr)
distrB.invW *= 3
distrB.m += 2
distrB.kappa *= 10
distrB2 = copy.deepcopy(distrB)
# Make sure things are different!
assert not np.allclose(distrB.invW, self.distr.invW)
assert not np.allclose(distrB.m, self.distr.m)
distrB.post_update_soVB(rho, self.distr)
assert distrB.dF == distrB2.dF * (1 - rho) + self.distr.dF * rho
assert np.allclose(distrB.kappa,
distrB2.kappa * (1 - rho) + self.distr.kappa * rho)
# these dont work because the parameterization is a bit trickier here.
#assert np.allclose(distrB.invW, distrB2.invW * (1-rho) + self.distr.invW * rho)
#assert np.allclose(distrB.m, distrB2.m * (1-rho) + self.distr.m * rho)
def test_entropy_posterior_gets_smaller(self, N=10):
PRNG = np.random.RandomState(seed=8675309)
for trial in range(3):
X = PRNG.randn(N, self.distr.D) + self.distr.m
x = np.sum(X, axis=0)
xxT = np.dot(X.T, X)
SS = SuffStatBag(K=1, D=self.distr.D)
SS.setField('N', [N], dims='K')
SS.setField('x', [x], dims=('K', 'D'))
SS.setField('xxT', [xxT], dims=('K', 'D', 'D'))
postD = self.distr.get_post_distr(SS, 0)
assert postD.D == self.distr.D
Hpost = postD.entropyWish()
Hprior = self.distr.entropyWish()
print 'Prior %.3g, Post %.3g' % (Hprior, Hpost)
assert Hpost < Hprior
def setUp(self):
self.invW = np.eye(2)
self.m = np.zeros(2)
self.distr = GaussWishDistr(m=self.m, invW=self.invW, kappa=1.0, dF=4)
def setUp(self): self.invW = np.eye(2) self.m = np.zeros(2) self.distr = GaussWishDistr(m=self.m, invW=self.invW, kappa=1.0, dF=4)
class TestGaussWishDistr(object):
def setUp(self):
self.invW = np.eye(2)
self.m = np.zeros(2)
self.distr = GaussWishDistr(m=self.m, invW=self.invW, kappa=1.0, dF=4)
def test_dimension(self):
assert self.distr.D == self.invW.shape[0]
def test_entropyWish(self):
''' Verify that (wishart) entropy is same for this object and Wishart object
'''
Hself = self.distr.entropyWish()
wishDistr = WishartDistr(v=self.distr.dF, invW=self.distr.invW)
Hwish = wishDistr.get_entropy()
assert np.allclose( Hself, Hwish)
def test_dist_mahalanobis(self, N=10):
''' Verify that distance computation is largest at mean and decays further away
'''
Xlist = list()
for r in [0, 0.01, 0.1, 1, 2, 3, 4, 5]:
Xlist.append(self.distr.m + r)
X = np.asarray(Xlist)
Dist = self.distr.dist_mahalanobis(X)
print Dist
assert np.all( Dist[:-1] < Dist[1:])
def test_update_soVB(self, rho=0.25):
''' Verify the blend update for stochastic variational is correct
'''
distrB = copy.deepcopy(self.distr)
distrB.invW *= 3
distrB.m += 2
distrB.kappa *= 10
distrB2 = copy.deepcopy(distrB)
# Make sure things are different!
assert not np.allclose(distrB.invW, self.distr.invW)
assert not np.allclose(distrB.m, self.distr.m)
distrB.post_update_soVB(rho, self.distr)
assert distrB.dF == distrB2.dF * (1-rho) + self.distr.dF * rho
assert np.allclose(distrB.kappa, distrB2.kappa * (1-rho) + self.distr.kappa * rho)
# these dont work because the parameterization is a bit trickier here.
#assert np.allclose(distrB.invW, distrB2.invW * (1-rho) + self.distr.invW * rho)
#assert np.allclose(distrB.m, distrB2.m * (1-rho) + self.distr.m * rho)
def test_entropy_posterior_gets_smaller(self, N=10):
PRNG = np.random.RandomState(seed=8675309)
for trial in range(3):
X = PRNG.randn(N, self.distr.D) + self.distr.m
x = np.sum(X,axis=0)
xxT = np.dot(X.T,X)
SS = SuffStatBag(K=1, D=self.distr.D)
SS.setField('N', [N], dims='K')
SS.setField('x', [x], dims=('K','D'))
SS.setField('xxT', [xxT], dims=('K','D','D'))
postD = self.distr.get_post_distr(SS, 0)
assert postD.D == self.distr.D
Hpost = postD.entropyWish()
Hprior = self.distr.entropyWish()
print 'Prior %.3g, Post %.3g' % (Hprior, Hpost)
assert Hpost < Hprior
