def test_run_birth_move_can_create_necessary_comps(self):
''' Can we start with one comp and recover the 3 true ones?
Of course, depending on the random init we maybe cannot,
but we should be able to a significant fraction of the time.
'''
birthArgs = dict(Kfresh=5, freshInitName='randexamples',
freshAlgName='VB', nFreshLap=50)
nSuccess = 0
nTrial = 10
for trial in range(nTrial):
PRNG = np.random.RandomState(trial)
newModel, newSS, MInfo = BirthMove.run_birth_move(self.oneModel,
self.Data, self.oneSS, randstate=PRNG, **birthArgs)
print newSS.N
newSS = self._run_LearnAlg_iterations(newModel, self.Data, nIter=10)
# after refining, we better have true comps!
topCompIDs = np.argsort(newSS.N)[::-1]
foundMatch = np.zeros(3)
for kk in topCompIDs[:3]:
estMu = newModel.obsModel.get_mean_for_comp(kk)
print estMu
for jj in range(3):
trueMu = self.Mu[jj,:]
if np.max(trueMu - estMu) < 0.5:
foundMatch[jj] = 1
print ' '
if np.all(foundMatch):
nSuccess += 1
assert nSuccess/float(nTrial) > 0.5
def test_learn_fresh_model_produces_new_comps(self):
''' Verify that learn_fresh_model produces new components
'''
PRNG = np.random.RandomState(0)
freshModel = self.oneModel.copy()
assert type(freshModel) == HModel
freshSS = BirthMove.learn_fresh_model(freshModel, self.Data, Kfresh=10,
freshInitName='randexamples', freshAlgName='VB',
nFreshLap=50, randstate=PRNG)
assert freshSS.K > 1
assert freshSS.K <= freshModel.obsModel.K
def test_run_birth_move_produces_viable_model(self):
''' Verify that the output of run_birth_move
can be used to create a valid model that can do all necessary subroutines like calc_local_params, etc.
'''
PRNG = np.random.RandomState(12345)
birthArgs = dict(Kfresh=10, freshInitName='randexamples',
freshAlgName='VB', nFreshLap=50)
newModel, newSS, MInfo = BirthMove.run_birth_move(self.oneModel,
self.Data, self.oneSS, randstate=PRNG, **birthArgs)
# Try newModel out with some quick calculations
assert newModel.obsModel.K > self.oneModel.obsModel.K
LP = newModel.calc_local_params(self.Data)
SS = newModel.get_global_suff_stats(self.Data, LP)
assert SS.K == newModel.obsModel.K
def test_learn_fresh_model_produces_new_comps(self):
''' Verify that learn_fresh_model produces new components
'''
PRNG = np.random.RandomState(0)
freshModel = self.oneModel.copy()
assert type(freshModel) == HModel
freshSS = BirthMove.learn_fresh_model(freshModel,
self.Data,
Kfresh=10,
freshInitName='randexamples',
freshAlgName='VB',
nFreshLap=50,
randstate=PRNG)
assert freshSS.K > 1
assert freshSS.K <= freshModel.obsModel.K
def test_learn_fresh_model_reproducible_random_seed(self):
''' Verify that learn_fresh_model produces same components
when called with same targetData and same randstate
'''
freshModel = self.oneModel.copy()
Nvec = list()
xvec = list()
for trial in range(3):
PRNG = np.random.RandomState(8383)
freshSS = BirthMove.learn_fresh_model(freshModel, self.Data, Kfresh=10,
freshInitName='randexamples', freshAlgName='VB',
nFreshLap=50, randstate=PRNG)
Nvec.append(freshSS.N)
xvec.append(freshSS.x)
assert np.all(Nvec[0] == Nvec[1])
assert np.all(Nvec[0] == Nvec[2])
assert np.all(xvec[0] == xvec[2])
def test_run_birth_move_produces_viable_model(self):
''' Verify that the output of run_birth_move
can be used to create a valid model that can do all necessary subroutines like calc_local_params, etc.
'''
PRNG = np.random.RandomState(12345)
birthArgs = dict(Kfresh=10,
freshInitName='randexamples',
freshAlgName='VB',
nFreshLap=50)
newModel, newSS, MInfo = BirthMove.run_birth_move(self.oneModel,
self.Data,
self.oneSS,
randstate=PRNG,
**birthArgs)
# Try newModel out with some quick calculations
assert newModel.obsModel.K > self.oneModel.obsModel.K
LP = newModel.calc_local_params(self.Data)
SS = newModel.get_global_suff_stats(self.Data, LP)
assert SS.K == newModel.obsModel.K
def test_learn_fresh_model_reproducible_random_seed(self):
''' Verify that learn_fresh_model produces same components
when called with same targetData and same randstate
'''
freshModel = self.oneModel.copy()
Nvec = list()
xvec = list()
for trial in range(3):
PRNG = np.random.RandomState(8383)
freshSS = BirthMove.learn_fresh_model(freshModel,
self.Data,
Kfresh=10,
freshInitName='randexamples',
freshAlgName='VB',
nFreshLap=50,
randstate=PRNG)
Nvec.append(freshSS.N)
xvec.append(freshSS.x)
assert np.all(Nvec[0] == Nvec[1])
assert np.all(Nvec[0] == Nvec[2])
assert np.all(xvec[0] == xvec[2])
def test_run_birth_move_can_create_necessary_comps(self):
''' Can we start with one comp and recover the 3 true ones?
Of course, depending on the random init we maybe cannot,
but we should be able to a significant fraction of the time.
'''
birthArgs = dict(Kfresh=5,
freshInitName='randexamples',
freshAlgName='VB',
nFreshLap=50)
nSuccess = 0
nTrial = 10
for trial in range(nTrial):
PRNG = np.random.RandomState(trial)
newModel, newSS, MInfo = BirthMove.run_birth_move(self.oneModel,
self.Data,
self.oneSS,
randstate=PRNG,
**birthArgs)
print newSS.N
newSS = self._run_LearnAlg_iterations(newModel,
self.Data,
nIter=10)
# after refining, we better have true comps!
topCompIDs = np.argsort(newSS.N)[::-1]
foundMatch = np.zeros(3)
for kk in topCompIDs[:3]:
estMu = newModel.obsModel.get_mean_for_comp(kk)
print estMu
for jj in range(3):
trueMu = self.Mu[jj, :]
if np.max(trueMu - estMu) < 0.5:
foundMatch[jj] = 1
print ' '
if np.all(foundMatch):
nSuccess += 1
assert nSuccess / float(nTrial) > 0.5
