def test_overcomplete(self):
H = self.H
D = self.D
Wgt = barstest.generate_bars_dict(H)
W = np.zeros((D, H + 2))
W[:, :H] = Wgt[:, ::-1]
perm = barstest.find_permutation(W, Wgt)
self.assertTrue((W[:, perm] == Wgt).all())
def test_find_permutation(self):
H = self.H
D = self.D
Wgt = barstest.generate_bars_dict(H)
# perform random permutation
for r in range(10):
perm_gt = np.random.permutation(H)
W = Wgt[:, perm_gt]
self.assertEqual(W.shape, (D, H))
perm = barstest.find_permutation(W, Wgt)
perm2 = barstest.find_permutation2(W, Wgt)
self.assertTrue((perm == perm2).all())
self.assertTrue((W[:, perm] == Wgt).all())
# Number of datapoints to generate
N = 1000
# Each datapoint is of D = size*size
size = 5
# Dimensionality of the model
H=2*size # number of latents
D=size**2 # dimensionality of observed data
# Approximation parameters for Expectation Truncation
Hprime = 7
gamma = 5
# Import and instantiate a model
from prosper.em.camodels.tsc_et import Ternary_ET
model = Ternary_ET(D, H, Hprime, gamma)
# Ground truth parameters. Only used to generate training data.
params_gt = {
'W' : 10*generate_bars_dict(H),
'pi' : 1.0 / size,
'sigma' : 2.0
}
from prosper.em.annealing import LinearAnnealing
anneal = LinearAnnealing(300)
anneal['T'] = [(0, 2.), (.7, 1.)]
anneal['Ncut_factor'] = [(0,0.),(2./3,1.)]
anneal['anneal_prior'] = False
from prosper.utils.barstest import generate_bars_dict
np.random.seed(1)
# Number of datapoints to generate
N = 1000
# Each datapoint is of D = size*size
size = 5
# Dimensionality of the model
H = 2 * size # number of latents
D = size**2 # dimensionality of observed data
# Approximation parameters for Expectation Truncation
Hprime = 8
gamma = 5
# Import and instantiate a model
from prosper.em.camodels.mca_et import MCA_ET
model = MCA_ET(D, H, Hprime, gamma)
# Ground truth parameters. Only used to generate training data.
params_gt = {'W': 10 * generate_bars_dict(H), 'pi': 1.0 / size, 'sigma': 2.0}
# Choose annealing schedule
from prosper.em.annealing import LinearAnnealing
anneal = LinearAnnealing(300)
anneal['T'] = [(0, 4.), (.8, 1.)]
anneal['Ncut_factor'] = [(0, 0.), (2. / 3, 1.)]
anneal['anneal_prior'] = False
def test_generate_negbars(self):
W = barstest.generate_bars_dict(self.H, neg_bars=True)
self.assertTrue(((W == -1.) + (W == 0.) + (W == 1.)).all())
self.assertTrue((W == -1.).any())
self.assertEqual(W.shape, (self.D, self.H))
def test_generate(self):
W = barstest.generate_bars_dict(self.H)
self.assertTrue(((W == 0.) + (W == 1.)).all())
self.assertEqual(W.shape, (self.D, self.H))
# slightly more costly.
#sigma_sq_type = 'full' # uncomment to make the model assume independent observation
# noise per observed dimension d of D. Makes the algorithm
# even more costly in terms of both computation and memory.
# Should be avoided for large D.
# Import and instantiate a model
from prosper.em.camodels.gsc_et import GSC
model = GSC(D, H, Hprime, gamma, sigma_sq_type)
# Model parameters used when artificially generating
# ground-truth data. This will NOT be used for the learning
# process.
params_gt = {
'W': 10 * generate_bars_dict(H), # this function is in bars-create-data
'pi': 2. / H * np.ones(H),
'mu': np.ones(H),
'psi_sq': np.eye(H)
}
if sigma_sq_type == 'scalar':
params_gt['sigma_sq'] = 1.0
elif sigma_sq_type == 'diagonal':
params_gt['sigma_sq'] = np.ones(D)
elif sigma_sq_type == 'full':
params_gt['sigma_sq'] = np.eye(D)
# Choose annealing schedule
from prosper.em.annealing import LinearAnnealing
anneal = LinearAnnealing(150)