def _calc_q_pi(self, alpha):
'''
Update the annotator models.
'''
psi_alpha_sum = psi(np.sum(alpha, 1))[:, None, :]
self.lnPi = psi(self.alpha) - psi_alpha_sum
return self.lnPi
def _calc_q_pi(alpha):
'''
Update the annotator models.
'''
psi_alpha_sum = psi(np.sum(alpha, 1))[:, None, :]
q_pi = psi(alpha) - psi_alpha_sum
return q_pi
def _init_lnPi(alpha0):
# Returns the initial values for alpha and lnPi
psi_alpha_sum = psi(np.sum(alpha0, 1))
lnPi = psi(alpha0) - psi_alpha_sum[:, None, :]
# init to prior
alpha = np.copy(alpha0)
return alpha, lnPi
def init_lnPi(self, N):
# Returns the initial values for alpha and lnPi
psi_alpha_sum = psi(np.sum(self.alpha0, 1))
self.lnPi = psi(self.alpha0) - psi_alpha_sum[:, None, :]
# init to prior
self.alpha = np.copy(self.alpha0)
self.alpha_taggers = {}
for midx in range(self.nModels):
self.alpha_taggers[midx] = np.copy(self.alpha0_taggers[midx])
def _calc_q_pi(self, alpha):
'''
Update the annotator models.
'''
psi_alpha_sum = np.zeros_like(alpha)
psi_alpha_sum[0, :] = psi(alpha[0, :] + alpha[1, :])
psi_alpha_sum[1, :] = psi_alpha_sum[0, :]
psi_alpha_sum[2:, :] = psi(np.sum(alpha[2:, :], 0))[None, :]
return psi(alpha) - psi_alpha_sum
def fit_predict(self, Et):
# count the number of occurrences for each label value
beta = self.beta0 + self.features_mat.dot(Et)
self.ElnRho = psi(beta) - psi(np.sum(beta, 0)[None, :])
lnptext_given_t = self.ElnRho[self.features, :]
# normalise, assuming equal prior here
pt_given_text = np.exp(lnptext_given_t -
logsumexp(lnptext_given_t, 1)[:, None])
return pt_given_text
def _init_lnPi(alpha0):
# Returns the initial values for alpha and lnPi
psi_alpha_sum = np.zeros_like(alpha0)
psi_alpha_sum[0, :] = psi(alpha0[0,:] + alpha0[1, :])
psi_alpha_sum[1, :] = psi_alpha_sum[0, :]
psi_alpha_sum[2:, :] = psi(np.sum(alpha0[2:, :], 0))[None, :]
lnPi = psi(alpha0) - psi_alpha_sum
# init to prior
alpha = np.copy(alpha0)
return alpha, lnPi
def _calc_q_pi(alpha):
'''
Update the annotator models.
'''
psi_alpha_sum = np.zeros_like(alpha)
psi_alpha_sum[0, :] = psi(alpha[0,:] + alpha[1, :])
psi_alpha_sum[1, :] = psi_alpha_sum[0, :]
psi_alpha_sum[2:, :] = psi(np.sum(alpha[2:, :], 0))[None, :]
ElnPi = psi(alpha) - psi_alpha_sum
# ElnPi[0, :] = np.log(0.5)
# ElnPi[1, :] = np.log(0.5)
# ElnPi[2:, :] = np.log(1.0 / float(alpha.shape[1] - 2))
return ElnPi
def init_lnPi(self, N):
# Returns the initial values for alpha and lnPi
psi_alpha_sum = np.zeros_like(self.alpha0)
psi_alpha_sum[0, :] = psi(self.alpha0[0, :] + self.alpha0[1, :])
psi_alpha_sum[1, :] = psi_alpha_sum[0, :]
psi_alpha_sum[2:, :] = psi(np.sum(self.alpha0[2:, :], 0))[None, :]
self.lnPi = psi(self.alpha0) - psi_alpha_sum
# init to prior
self.alpha = np.copy(self.alpha0)
self.alpha_taggers = {}
for midx in range(self.nModels):
self.alpha_taggers[midx] = np.copy(self.alpha0_taggers[midx])
def _calc_q_pi(self, alpha):
'''
Update the annotator models.
'''
psi_alpha_sum = psi(np.sum(alpha, 1))[:, None, :]
self.lnpi = psi(alpha) - psi_alpha_sum
lnpi_incorrect = psi(np.sum(alpha, 1)
- alpha[range(self.alpha.shape[0]), range(alpha.shape[0]), :]) \
- psi_alpha_sum[:, 0, :]
lnpi_incorrect = np.log(
np.exp(lnpi_incorrect) / float(alpha.shape[1] - 1)) # J x K
for j in range(alpha.shape[0]):
for l in range(alpha.shape[1]):
if j == l:
continue
self.lnpi[j, l, :] = lnpi_incorrect[j, :]
return self.lnpi
def _calc_q_pi(alpha):
'''
Update the annotator models.
TODO Representing using a full matrix might break lower bound implementation
'''
psi_alpha_sum = psi(np.sum(alpha, 1))[:, None, :]
q_pi = psi(alpha) - psi_alpha_sum
q_pi_incorrect = psi(np.sum(alpha, 1) - alpha[np.arange(alpha.shape[0]), np.arange(alpha.shape[0]), :]) \
- psi_alpha_sum[:, 0, :]
q_pi_incorrect = np.log(
np.exp(q_pi_incorrect) / float(alpha.shape[1] - 1)) # J x K
for j in range(alpha.shape[0]):
for l in range(alpha.shape[1]):
if j == l:
continue
q_pi[j, l, :] = q_pi_incorrect[j, :]
return q_pi
def calc_grad_weight(nks, alphaS, alphaB, N, AB, AS):
""" Calculate the weights for each bin k. Returns k-length vector."""
K = nks.size
w = sp.psi(nks + alphaS) - sp.psi(nks+alphaB) + sp.psi(N+AB) - sp.psi(N+AS)
return w
def calc_Qdiff(nVec, alphaS, alphaB):
""" Calculate Q_k - Q_base, for each bin k. Returns k-length vector."""
N = nVec.sum()
Qdiff = sp.psi(nVec + alphaS) - sp.psi(nVec+alphaB) - sp.psi(N+alphaS.sum()) + sp.psi(N+alphaB.sum())
return Qdiff
