def _ppl_log_single_document(self,
docdata): ### potential underflow problem
d, docToken, [doc_u, doc_e] = docdata
prob_w_kv = (self.phiT * self.pi[1] + self.phiB * self.pi[0])
ppl_w_k_log = -np.sum(np.log(prob_w_kv[:, docToken]), axis=1)
ppl_w_k_scaled, ppl_w_k_constant = expConstantIgnore(
-ppl_w_k_log, constant_output=True) # (actual ppl^(-1))
prob_e_mk = np.dot(self.psi[doc_u, :], self.eta)
ppl_e_k_log = -np.sum(
np.log(prob_e_mk[np.arange(doc_u.shape[0]), :, doc_e]), axis=0)
ppl_e_k_scaled, ppl_e_k_constant = expConstantIgnore(
-ppl_e_k_log, constant_output=True) # (actual ppl^(-1))
prob_k = self.theta
# for emoti given words
prob_e_m = probNormalize(
np.tensordot(prob_e_mk,
np.multiply(prob_k, ppl_w_k_scaled),
axes=(1, 0)))
ppl_e_log = -np.sum(np.log(prob_e_m[np.arange(doc_u.shape[0]), doc_e]))
# for words given emoti ! same prob_w for different n
prob_w = probNormalize(
np.tensordot(prob_w_kv,
np.multiply(prob_k, ppl_e_k_scaled),
axes=(0, 0)))
ppl_w_log = -np.sum(np.log(prob_w[docToken]))
# for both words & emoti
try:
ppl_log = -(np.log(
np.inner(ppl_w_k_scaled, np.multiply(ppl_e_k_scaled, prob_k)))
+ ppl_w_k_constant + ppl_e_k_constant)
except FloatingPointError as e:
raise e
return ppl_w_log, ppl_e_log, ppl_log
def _ppl_log_single_document_off_shell(self,
docdata,
display=False
): ### potential underflow problem
d, docToken, [doc_u, doc_e] = docdata
prob_w_kv = (self.GLV["phiT"] * self.GLV["pi"][1] +
self.GLV["phiB"] * self.GLV["pi"][0])
ppl_w_k_log = -np.sum(np.log(prob_w_kv[:, docToken]), axis=1)
ppl_w_k_scaled, ppl_w_k_constant = expConstantIgnore(
-ppl_w_k_log, constant_output=True) # (actual ppl^(-1))
prob_e_mk = np.dot(self.GLV["psi"][doc_u, :], self.GLV["eta"])
ppl_e_k_log = -np.sum(
np.log(prob_e_mk[np.arange(doc_u.shape[0]), :, doc_e]), axis=0)
ppl_e_k_scaled, ppl_e_k_constant = expConstantIgnore(
-ppl_e_k_log, constant_output=True) # (actual ppl^(-1))
prob_k = self.GLV["theta"]
# for emoti given words
prob_e_m = probNormalize(
np.tensordot(prob_e_mk,
np.multiply(prob_k, ppl_w_k_scaled),
axes=(1, 0)))
ppl_e_log = -np.sum(np.log(prob_e_m[np.arange(doc_u.shape[0]), doc_e]))
# for words given emoti ! same prob_w for different n
prob_w = probNormalize(
np.tensordot(prob_w_kv,
np.multiply(prob_k, ppl_e_k_scaled),
axes=(0, 0)))
ppl_w_log = -np.sum(np.log(prob_w[docToken]))
# for both words & emoti
try:
ppl_log = -(np.log(
np.inner(ppl_w_k_scaled, np.multiply(ppl_e_k_scaled, prob_k)))
+ ppl_w_k_constant + ppl_e_k_constant)
except FloatingPointError as e:
raise e
### test ###
if display:
self._log("ppl_log_single_document_off_shell for doc %d" % d)
self._log("docToken %s" % str(docToken))
self._log("ppl_w_k_scaled %s" % str(ppl_w_k_scaled))
self._log("ppl_e_k_scaled %s" % str(ppl_e_k_scaled))
self._log("prob_e_m %s" % str(prob_e_m))
self._log("prob_g_m %s" % str(self.GLV["psi"][doc_u, :]))
return [ppl_w_log, ppl_e_log,
ppl_log], docToken.shape[0], doc_u.shape[0]
def _ppl_log_single_document_off_shell(self, docdata):
d, docToken, [doc_u, doc_e] = docdata
Nd = docToken.shape[0]
prob_z_alpha = np.ones([self.K, Nd], dtype=np.float64)
ppl_w_z = self.phi[:, docToken]
prob_z = prob_z_alpha * ppl_w_z
prob_z_sum = np.sum(
prob_z, axis=1
) # product over dirichlet is the same of sum over dirichlet priors
prob_e = probNormalize(np.dot(prob_z_sum, self.eta))
ppl_e_log = -np.sum(np.log(prob_e)[doc_e])
doc_E = np.sum(np.identity(self.E, dtype=np.float64)[:, doc_e], axis=1)
docE = probNormalize(doc_E + SMOOTH_FACTOR)
ppl_e_z_log = -(np.tensordot(np.log(docE), self.eta - 1.0, axes=(0, 1))
+ np.log(self.eta_beta_inv))
ppl_e_z_scaled, ppl_e_z_constant = expConstantIgnore(
-ppl_e_z_log, constant_output=True)
prob_w = probNormalize(np.dot(ppl_e_z_scaled, self.phi))
ppl_w_log = -np.sum(np.log(prob_w)[docToken])
ppl_log = np.nan
return [ppl_w_log, ppl_e_log,
ppl_log], docToken.shape[0], doc_e.shape[0]
def _predict_single_document_off_shell(self, docdata):
d, docToken, [doc_u, doc_e] = docdata
prob_w_kv = (self.GLV["phiT"] * self.GLV["pi"][1] + self.GLV["phiB"] * self.GLV["pi"][0])
ppl_w_k_log = -np.sum(np.log(prob_w_kv[:, docToken]), axis=1)
ppl_w_k_scaled, ppl_w_k_constant = expConstantIgnore(- ppl_w_k_log, constant_output=True)
prob_k = self.GLV["theta"]
prob_e_mk = np.dot(self.GLV["psi"][doc_u, :], self.GLV["eta"])
prob_e_m = probNormalize(np.tensordot(prob_e_mk, np.multiply(prob_k, ppl_w_k_scaled), axes=(1, 0)))
return prob_e_m.tolist(), doc_e.tolist()
def _ppl_log_single_document_off_shell(
self, docdata): ### potential underflow problem
d, docToken, [doc_u, doc_e] = docdata
prob_w_kv = self.GLV["phi"]
ppl_w_k_log = -np.sum(np.log(prob_w_kv[:, docToken]), axis=1)
ppl_w_k_scaled, ppl_w_k_constant = expConstantIgnore(
-ppl_w_k_log, constant_output=True) # (actual ppl^(-1))
prob_e_mk = np.dot(
self.GLV["psi"][doc_u, :],
(np.tensordot(self.GLV["pi"], self.GLV["eta"], axes=(1, 0)) *
self.GLV["c"][1] +
np.tensordot(self.GLV["piB"], self.GLV["eta"], axes=(0, 0)) *
self.GLV["c"][0]))
ppl_e_k_log = -np.sum(
np.log(prob_e_mk[np.arange(doc_u.shape[0]), :, doc_e]), axis=0)
ppl_e_k_scaled, ppl_e_k_constant = expConstantIgnore(
-ppl_e_k_log, constant_output=True) # (actual ppl^(-1))
prob_k = self.GLV["theta"]
# for emoti given words
prob_e_m = probNormalize(
np.tensordot(prob_e_mk,
np.multiply(prob_k, ppl_w_k_scaled),
axes=(1, 0)))
ppl_e_log = -np.sum(np.log(prob_e_m[np.arange(doc_u.shape[0]), doc_e]))
# for words given emoti ! same prob_w for different n
prob_w = probNormalize(
np.tensordot(prob_w_kv,
np.multiply(prob_k, ppl_e_k_scaled),
axes=(0, 0)))
ppl_w_log = -np.sum(np.log(prob_w[docToken]))
# for both words & emoti
ppl_log = -(np.log(
np.inner(ppl_w_k_scaled, np.multiply(ppl_e_k_scaled, prob_k))) +
ppl_w_k_constant + ppl_e_k_constant)
return [ppl_w_log, ppl_e_log,
ppl_log], docToken.shape[0], doc_u.shape[0]