def _new_params(self, n_new_features):
new_H = np.zeros([self.D, n_new_features])
new_Wn = np.zeros(n_new_features)
shared_hyperprior = np.zeros(n_new_features)
for k in range(n_new_features):
hyp_prior = self.pSharedHyp.prior.rvs()
# loop through the rows and assign samples
new_Wn[k] = truncated_sampler(mu=0,
std=1/np.sqrt(hyp_prior),
lower_bound=self.trunc_low_bound,
upper_bound=self.trunc_up_bound)
shared_hyperprior[k] = hyp_prior
for d in range(self.D):
new_H[d,k] = truncated_sampler(mu=0,
std=1/np.sqrt(hyp_prior),
lower_bound=self.trunc_low_bound,
upper_bound=self.trunc_up_bound)
return new_H, new_Wn, shared_hyperprior
def _add_new_W_hyperpriors(self, new_Wn, new_hyperpriors, k_new, n):
self.pSharedHyp.add_new_features(new_hyperpriors, k_new)
add_W = np.empty((k_new, self.N))
for i, param in enumerate(new_hyperpriors):
param = np.broadcast_to(param, self.N)
add_W[i,:] = truncated_sampler(mu=0,
std=1/np.sqrt(param),
lower_bound=self.trunc_low_bound,
upper_bound=self.trunc_up_bound)
add_W[:,n] = new_Wn
self.pW.add_new_features(add_W, k_new, n)
def _add_new_W_hyperpriors(self, new_Wn, new_hyperpriors, k_new, n):
new_hyp_H, new_hyp_Wn = new_hyperpriors
self.pHypH.add_new_features(new_hyp_H, k_new)
add_W = np.empty((k_new, self.N))
add_hyp_W = add_W.copy()
for n in range(self.N):
for k in range(k_new):
hyp_prior = self.pHypW.prior.rvs()
add_hyp_W[k,n] = hyp_prior
add_W[k,n] = truncated_sampler(mu=0,
std=1/np.sqrt(hyp_prior),
lower_bound=self.trunc_low_bound,
upper_bound=self.trunc_up_bound)
add_hyp_W[:,n] = new_hyp_Wn
add_W[:,n] = new_Wn
self.pW.add_new_features(add_W, k_new, n)
self.pHypW.add_new_features(add_hyp_W, k_new)