def _partial_inference_z_update(self, theta_avg, doc_y_i, docToken, phi_avg):
doc_z_unnorm_log = np.dot(
np.log(phi_avg[:, docToken]),
doc_y_i
)
doc_z_unnorm_log += np.log(theta_avg)
return probNormalizeLog(doc_z_unnorm_log)
def _prob_doc_z(self, TE_no_d_z, Y1TV_no_d_z, TI_no_d_z, docE, docW,
doc_Y1V):
"""
calculate conditional probability for document-level topic doc_z
method is different from [1]
:param TE_no_d_z: TE without doc d
:param Y1TV_no_d_z: Y1TV without d
:param TI_no_d_z: TI without d
:param docE: document-level emotion counts [self.E], np.sum(docE) == Md
:param docW: document-level words counts [self.V], np.sum(docW) == Nd
:return: doc_z
"""
# alpha #
log_prob_a = np.log(TI_no_d_z + self.alpha)
# TE #
a = TE_no_d_z + self.gamma
log_prob_b = np.sum(logfactorial(docE, a), axis=1) - logfactorial(
np.sum(docE), np.sum(a, axis=1))
# Y1TV #
b = Y1TV_no_d_z + self.beta
log_prob_c = np.sum(logfactorialSparse(
doc_Y1V, b), axis=1) - logfactorial(doc_Y1V.sum(), np.sum(b,
axis=1))
prob_doc_z = probNormalizeLog(log_prob_a + log_prob_b + log_prob_c)
return prob_doc_z
def _fit_single_document_x_update(doc_z2, doc_r, doc_u, doc_e, pars_topass):
doc_x_unnorm_log_u = pars_topass["psi"].bigamma_data[doc_u]
doc_x_unnorm_log_e = np.expand_dims(doc_r[:, 2], axis=-1) * np.transpose(
np.tensordot(
doc_z2,
pars_topass["eta"].bigamma_data,
axes=(0, 0)
),
axes=(1, 0)
)[doc_e, :]
doc_x_unnorm_log = doc_x_unnorm_log_u + doc_x_unnorm_log_e
return probNormalizeLog(doc_x_unnorm_log)
def _fit_single_document_z1_update(doc_y, doc_r, docToken, doc_e, pars_topass):
doc_z1_unnorm_log_w = np.dot(
pars_topass["phi1"].bigamma_data[:, docToken],
doc_y[:, 0]
)
doc_z1_unnorm_log_e = np.tensordot(
doc_r[:, 1],
pars_topass["eta_z"].bigamma_data[:, doc_e],
axes=[0, -1]
)
doc_z1_unnorm_log_theta1 = pars_topass["theta1"].bigamma_data
doc_z1_unnorm_log = doc_z1_unnorm_log_w + doc_z1_unnorm_log_e + doc_z1_unnorm_log_theta1
return probNormalizeLog(doc_z1_unnorm_log)
def _fit_single_document_z2_update(doc_x, doc_y, doc_r, docToken, doc_e, pars_topass):
doc_z2_unnorm_log_w = np.dot(
pars_topass["phi2"].bigamma_data[:, docToken],
doc_y[:, 1]
)
doc_z2_unnorm_log_e = np.tensordot(
pars_topass["eta"].bigamma_data[:, :, doc_e] * doc_r[:, 2],
doc_x,
axes=([2, 1], [0, 1])
)
doc_z2_unnorm_log_theta2 = pars_topass["theta2"].bigamma_data
doc_z2_unnorm_log = doc_z2_unnorm_log_w + doc_z2_unnorm_log_e + doc_z2_unnorm_log_theta2
return probNormalizeLog(doc_z2_unnorm_log)
def _fit_single_document_r_update(doc_z2, doc_z1, doc_x, doc_u, doc_e, pars_topass):
Md = doc_u.shape[0]
doc_r_unnorm_log = np.zeros([Md, 3])
doc_r_unnorm_log[:, 0] = pars_topass["eta_u"].bigamma_data[doc_u, doc_e]
doc_r_unnorm_log[:, 1] = np.dot(doc_z1, pars_topass["eta_z"].bigamma_data[:, doc_e])
doc_r_unnorm_log[:, 2] = np.tensordot(
doc_x,
np.tensordot(
doc_z2,
pars_topass["eta"].bigamma_data,
axes=(0, 0)
),
axes=(1, 0)
)[np.arange(Md), doc_e]
doc_r_unnorm_log += pars_topass["se"].bigamma_data
return probNormalizeLog(doc_r_unnorm_log)
def _prob_doc_z(self, TI_no_d, TXE_no_d, Y1TV_no_d, doc_XE, doc_Y1V):
"""
calculate conditional probability for document-level topic doc_z
:param: doc_Y1V: lil_matrix((1, self.V), dtype=int8)
"""
# alpha #
log_prob_alpha = np.log(TI_no_d + self.alpha)
# gamma # (without sparsity, directly calculating log gamma function)
a = TXE_no_d + self.gamma + doc_XE
log_prob_gamma = np.sum(np.sum(gammaln(a), axis=-1) -
gammaln(np.sum(a, axis=-1)),
axis=-1)
# beta # (with sparsity)
b = Y1TV_no_d + self.beta
log_prob_beta = np.sum(logfactorialSparse(doc_Y1V, b),
axis=-1) - logfactorial(doc_Y1V.sum(),
np.sum(b, axis=-1))
prob_doc_z = probNormalizeLog(log_prob_alpha + log_prob_gamma +
log_prob_beta)
return prob_doc_z
def _fit_single_document_y_update(doc_z2, doc_z1, docToken, pars_topass):
doc_y_unnorm_log = np.zeros([docToken.shape[0], 2])
doc_y_unnorm_log[:, 0] = np.dot(doc_z1, pars_topass["phi1"].bigamma_data[:, docToken])
doc_y_unnorm_log[:, 1] = np.dot(doc_z2, pars_topass["phi2"].bigamma_data[:, docToken])
doc_y_unnorm_log += pars_topass["pi"].bigamma_data
return probNormalizeLog(doc_y_unnorm_log)
def _partial_inference_y_update(self, doc_z2, doc_z1, pi_avg, phi2_avg, phi1_avg, docToken):
doc_y_unnorm_log = np.zeros([docToken.shape[0], 2])
doc_y_unnorm_log[:, 0] = np.dot(doc_z1, np.log(phi1_avg[:, docToken]))
doc_y_unnorm_log[:, 1] = np.dot(doc_z2, np.log(phi2_avg[:, docToken]))
doc_y_unnorm_log += np.log(pi_avg)
return probNormalizeLog(doc_y_unnorm_log)
