def initialize(self, alpha, beta, sample_init=True):
init_mean = np.empty([self.K, self.N])
init_var = np.empty_like(init_mean)
init_etrpy = np.empty_like(init_mean)
alpha = alpha / self.K
beta = beta * (1 - 1.0 / self.K)
if sample_init:
from npbNMF.datatools import sample_init_Z_bernoulli
for n in range(self.N):
for k in range(self.K):
estimates = sample_init_Z_bernoulli(
alpha,
beta,
n_samples=10,
)
init_mean[k, n], init_var[k, n] = estimates
# leave the entropy, as we refrain from estimating it
else:
p_expected = np.full((self.K, self.N), alpha / (beta + alpha))
init_mean, init_var, init_etrpy = bernoulli_moments(p_expected)
self.mean = init_mean
self.var = init_var
self.etrpy = init_etrpy
self.mean_sq = mean_sq(self.mean, self.var)
def moments(self, mu, sigma_sq):
new_mean, new_var, new_etrpy = trunc_moments(self.lower_bound,
self.upper_bound, mu,
sigma_sq)
new_mean_sq = mean_sq(new_mean, new_var)
return new_mean, new_var, new_etrpy, new_mean_sq
def add_new_features(self, expected_param_inv, k_new):
new_moments = self._initialize_new_features(expected_param_inv, k_new)
new_mean, new_var, new_etrpy = new_moments
new_mean_sq = mean_sq(new_mean, new_var)
self.mean = np.append(self.mean, new_mean, axis=0)
self.var = np.append(self.var, new_var, axis=0)
self.mean_sq = np.append(self.mean_sq, new_mean_sq, axis=0)
self.etrpy = np.append(self.etrpy, new_etrpy, axis=0)
def initialize(self, expected_param_inv, K_init=1):
init_mean = np.empty([K_init, self.N])
init_var = np.empty_like(init_mean)
init_etrpy = np.zeros_like(init_mean)
for n in range(self.N):
for k in range(K_init):
estimates = self._sample_estimates(expected_param_inv)
init_mean[k, n], init_var[k, n] = estimates
# leave the entropy, as we refrain from estimating it
self.mean = init_mean
self.var = init_var
self.etrpy = init_etrpy
self.mean_sq = mean_sq(self.mean, self.var)
def initialize(self, expected_param_inv, K_init=1):
init_mean = np.empty([self.D, K_init])
init_var = np.empty_like(init_mean)
init_etrpy = np.zeros_like(init_mean)
from npbNMF.datatools import sample_init_factor_expon
for d in range(self.D):
for k in range(K_init):
estimates = self._sample_estimates(expected_param_inv)
init_mean[d, k], init_var[d, k] = estimates
# leave the entropy, as we refrain from estimating it
self.mean = init_mean
self.var = init_var
self.etrpy = init_etrpy
self.mean_sq = mean_sq(self.mean, self.var)
def initialize(self, expected_param):
init_mean = np.empty([self.K, self.N])
init_var = np.empty_like(init_mean)
init_etrpy = np.empty_like(init_mean)
from npbNMF.datatools import sample_init_factor_expon
for n in range(self.N):
for k in range(self.K):
estimates = self._sample_estimates(expected_param)
init_mean[k, n], init_var[k, n] = estimates
# leave the entropy, as we refrain from estimating it
self.mean = init_mean
self.var = init_var
self.etrpy = init_etrpy
self.mean_sq = mean_sq(self.mean, self.var)
def initialize_preset(self, init_mean, init_var):
self.mean = init_mean
self.var = init_var
self.mean_sq = mean_sq(self.mean, self.var)
def moments(self, pi):
new_mean, new_var, new_etrpy = bernoulli_moments(pi)
new_mean_sq = mean_sq(new_mean, new_var)
return new_mean, new_var, new_etrpy, new_mean_sq