def _initialize(self, dataE, dataW, dataToken):
start = datetime.now()
self.theta = probNormalize(np.random.random([self.E, self.K]))
self.phi = probNormalize(np.random.random([self.K, self.V]))
self.esp = []
self.z = []
z_dist = np.sum(self.theta, axis=0) / self.E
for d in range(self.D):
Nd = self.Nd[d]
gamma = dataE[d]
self.esp.append(multinomial(gamma, Nd))
self.z.append(multinomial(z_dist, Nd))
self.TE = np.zeros([self.K, self.E], dtype=np.int32)
self.TV = np.zeros([self.K, self.V], dtype=np.int32)
for d in range(self.D):
docToken = dataToken[d]
doc_z = self.z[d]
doc_esp = self.esp[d]
for n in range(self.Nd[d]):
w = docToken[n]
w_z = doc_z[n]
w_esp = doc_esp[n]
self.TE[w_z, w_esp] += 1
self.TV[w_z, w] += 1
self.TI = np.sum(self.TV, axis=1)
self.IE = np.sum(self.TE, axis=0)
duration = datetime.now() - start
print "_initialize() takes %fs" % duration.total_seconds()
def _initialize(self, dataDUE, dataW, dataToken):
start = datetime.now()
self.theta = probNormalize(np.random.random([self.K]))
self.pi = probNormalize(np.random.random([2]))
self.eta = probNormalize(
np.random.random([self.K, self.G, self.E]) + 0.1)
self.phiB = probNormalize(np.random.random([self.V]))
self.phiT = probNormalize(np.random.random([self.K, self.V]))
self.psi = probNormalize(np.random.random([self.U, self.G]))
self.z = np.zeros([self.D], dtype=np.int8)
self.y = []
self.x = []
for d in range(self.D):
self.z[d] = multinomial(self.theta)
self.y.append(multinomial(self.pi, self.Nd[d]))
## time consuming, replaced with below ##
# doc_x = []
# for m in range(self.Md[d]):
# u = np.random.randint(0,self.U)
# doc_x.append(multinomial(self.psi[u]))
# self.x.append(np.array(doc_x, dtype=np.int8))
self.x.append(multinomial(self.psi[0], self.Md[d]))
duration = datetime.now() - start
self._log("_initialize() takes %fs" % duration.total_seconds())
def _initialize(self, dataDUE):
start = datetime.now()
self.theta = probNormalize(np.random.random([self.D, self.K]))
self.phi = probNormalize(np.random.random([self.K, self.V]))
self.eta = np.random.random([self.K, self.E])
self.z = []
for d in range(self.D):
z_dist = self.theta[d]
Nd = self.Nd[d]
self.z.append(multinomial(z_dist, Nd))
self.eta_beta_inv = multivariateBeta_inv(self.eta)
self.TI = np.zeros([self.D, self.K], dtype=np.int32)
self.TV = np.zeros([self.K, self.V], dtype=np.int32)
self.dataE_smoothed = {}
for docdata in dataDUE.generate():
d, docToken, [doc_u, doc_e] = docdata
doc_z = self.z[d]
for n in range(self.Nd[d]):
w = docToken[n]
w_z = doc_z[n]
self.TI[d, w_z] += 1
self.TV[w_z, w] += 1
doc_E = np.sum(np.identity(self.E, dtype=np.float64)[:, doc_e], axis=1)
docE = probNormalize(doc_E + SMOOTH_FACTOR)
self.dataE_smoothed[d] = docE
duration = datetime.now() - start
self._log("_initialize() takes %fs" % duration.total_seconds())
def _doc_z_update(self, d, doc_u, doc_e, docW, docToken):
""" update document-level topic """
doc_z = self.z[d]
doc_XE = self.DXE[d]
doc_Y1V = self.DY1V.getrow(d).tocsr()
doc_Y1V_array = doc_Y1V.toarray().squeeze()
# calculate leave-one out statistics #
TI_no_d, TXE_no_d, Y1TV_no_d = self.TI, self.TXE, self.Y1TV
TI_no_d[doc_z] += -1
TXE_no_d[doc_z, :, :] += -doc_XE
Y1TV_no_d[doc_z, :] += -doc_Y1V_array
# conditional probability #
prob_doc_z = self._prob_doc_z(TI_no_d, TXE_no_d, Y1TV_no_d, doc_XE,
doc_Y1V)
# new sampled result #
doc_z_new = int(multinomial(prob_doc_z))
# update #
self.z[d] = doc_z_new
TI_no_d[doc_z_new] += 1
TXE_no_d[doc_z_new, :, :] += doc_XE
Y1TV_no_d[doc_z_new, :] += doc_Y1V_array
self.TI, self.TXE, self.Y1TV = TI_no_d, TXE_no_d, Y1TV_no_d
def _GibbsSamplingLocal(self, dataE, dataW, dataToken, epoch):
"""
Gibbs sampling word-level topic
"""
pbar = tqdm(range(self.D), total=self.D, desc='({0:^3})'.format(epoch))
for d in pbar: # sequentially sampling
doc_Nd = self.Nd[d]
docE = probNormalize(dataE[d] + SMOOTH_FACTOR)
docToken = dataToken[d]
for n in range(doc_Nd):
w = docToken[n]
w_z = self.z[d][n]
## sampling ##
# calculate leave-one-out statistics #
TI_no_dn, TV_no_dn = self.TI, self.TV
TI_no_dn[d, w_z] += -1
TV_no_dn[w_z, w] += -1
# conditional probability #
prob_pa = TI_no_dn[d] + self.alpha
prob_pb = np.divide(TV_no_dn[:, w] + self.beta,
np.sum(TV_no_dn + self.beta, axis=1))
prob_pc = np.multiply(
self.eta_beta_inv,
np.prod(np.power(docE, self.eta - 1), axis=1))
prob_w_z = probNormalize(prob_pa * prob_pb * prob_pc)
# new sampled result #
w_z_new = multinomial(prob_w_z)
# update #
self.z[d][n] = w_z_new
TI_no_dn[d, w_z_new] += 1
TV_no_dn[w_z_new, w] += 1
self.TI, self.TV = TI_no_dn, TV_no_dn
def _initialize(self, dataE, dataW, dataToken):
start = datetime.now()
self.theta = probNormalize(np.random.random([self.D, self.K]))
self.phi = probNormalize(np.random.random([self.K, self.V]))
self.eta = np.random.random([self.K, self.E])
self.z = []
for d in range(self.D):
z_dist = self.theta[d]
Nd = self.Nd[d]
self.z.append(multinomial(z_dist, Nd))
self.eta_beta_inv = multivariateBeta_inv(self.eta)
self.TI = np.zeros([self.D, self.K], dtype=np.int32)
self.TV = np.zeros([self.K, self.V], dtype=np.int32)
for d in range(self.D):
docToken = dataToken[d]
doc_z = self.z[d]
for n in range(self.Nd[d]):
w = docToken[n]
w_z = doc_z[n]
self.TI[d, w_z] += 1
self.TV[w_z, w] += 1
duration = datetime.now() - start
print "_initialize() takes %fs" % duration.total_seconds()
def _initialize(self, dataE, dataW, dataToken):
start = datetime.now()
self.theta = probNormalize(np.random.random([self.K]))
self.pi = probNormalize(np.random.random([2]))
self.eta = probNormalize(np.random.random([self.K, self.E]))
self.phiB = probNormalize(np.random.random([self.V]))
self.phiT = probNormalize(np.random.random([self.K, self.V]))
self.z = np.zeros([self.D], dtype=np.int8)
self.y = []
for d in range(self.D):
self.z[d] = multinomial(self.theta)
Nd = self.Nd[d]
self.y.append(multinomial(self.pi, Nd))
duration = datetime.now() - start
print "_initialize() takes %fs" % duration.total_seconds()
def _GibbsSamplingLocal(self, dataE, dataW, dataToken, epoch):
"""
Gibbs sampling word-level emotion and topic
"""
pbar = tqdm(range(self.D),
total = self.D,
desc='({0:^3})'.format(epoch))
for d in pbar: # sequentially sampling
doc_Nd = self.Nd[d]
docE = dataE[d]
docToken = dataToken[d]
for n in range(doc_Nd):
w = docToken[n]
w_z = self.z[d][n]
w_esp = self.esp[d][n]
## sampling ##
# calculate leave-one out statistics #
TE_no_dn, TV_no_dn, TI_no_dn, IE_no_dn, = self.TE, self.TV, self.TI, self.IE
TE_no_dn[w_z, w_esp] += -1
TV_no_dn[w_z, w] += -1
TI_no_dn[w_z] += -1
IE_no_dn[w_esp] += -1
# conditional probability #
prob_w_esp = np.divide(np.multiply((self.alpha + TE_no_dn[w_z]), docE),
(self.K * self.alpha + IE_no_dn))
prob_w_esp = probNormalize(prob_w_esp)
prob_w_z = np.divide(np.multiply((self.alpha + TE_no_dn[:, w_esp]), (self.beta + TV_no_dn[:, w])),
(self.V * self.beta + TI_no_dn))
prob_w_z = probNormalize(prob_w_z)
# new sampled result #
w_esp_new = multinomial(prob_w_esp)
w_z_new = multinomial(prob_w_z)
# update #
self.z[d][n] = w_z_new
self.esp[d][n] = w_esp_new
TE_no_dn[w_z_new, w_esp_new] += 1
TV_no_dn[w_z_new, w] += 1
TI_no_dn[w_z_new] += 1
IE_no_dn[w_esp_new] += 1
self.TE, self.TV, self.TI, self.IE = TE_no_dn, TV_no_dn, TI_no_dn, IE_no_dn
def _GibbsSamplingLocal(self, dataDUE, epoch):
"""
Gibbs sampling word-level topic
"""
pbar = tqdm(dataDUE.generate(),
total=self.D_train,
desc='({0:^3})'.format(epoch))
for docdata in pbar: # sequentially sampling
d, docToken, [doc_u, doc_e] = docdata
doc_Nd = self.Nd[d]
if d in self.dataE_smoothed:
docE = self.dataE_smoothed[d]
else:
doc_E = np.sum(np.identity(self.E, dtype=np.float64)[:, doc_e],
axis=1)
docE = probNormalize(doc_E + SMOOTH_FACTOR)
self.dataE_smoothed[d] = docE
for n in range(doc_Nd):
w = docToken[n]
w_z = self.z[d][n]
## sampling ##
# calculate leave-one-out statistics #
TI_no_dn, TV_no_dn = self.TI, self.TV
TI_no_dn[d, w_z] += -1
TV_no_dn[w_z, w] += -1
# conditional probability #
prob_pa = TI_no_dn[d] + self.alpha
prob_pb = np.divide(TV_no_dn[:, w] + self.beta,
np.sum(TV_no_dn + self.beta, axis=1))
prob_pc = np.multiply(
self.eta_beta_inv,
np.prod(np.power(docE, self.eta - 1), axis=1))
prob_w_z = probNormalize(prob_pa * prob_pb * prob_pc)
# new sampled result #
w_z_new = multinomial(prob_w_z)
# update #
self.z[d][n] = w_z_new
TI_no_dn[d, w_z_new] += 1
TV_no_dn[w_z_new, w] += 1
self.TI, self.TV = TI_no_dn, TV_no_dn
def _GibbsSamplingLocal(self, dataE, dataW, dataToken, epoch):
"""
Gibbs sampling word-level background-vs-topic and document-level topic
"""
pbar = tqdm(range(self.D),
total = self.D,
desc='({0:^3})'.format(epoch))
for d in pbar: # sequentially sampling
doc_Nd = self.Nd[d]
docE = dataE[d]
docW = dataW[d]
docToken = dataToken[d]
doc_z = self.z[d]
# intermediate parameters calculation #
Y1T = np.sum(self.Y1TV, axis=1)
for n in range(doc_Nd):
w = docToken[n]
w_y = self.y[d][n]
## sampling for y ##
# calculate leave-one out statistics #
YI_no_dn_y, Y0V_no_dn_y, Y1TV_no_dn_y = self.YI, self.Y0V, self.Y1TV
Y1T_no_dn_y = Y1T
YI_no_dn_y[w_y] += -1
if w_y == 0:
Y0V_no_dn_y[w] += -1
elif w_y == 1:
Y1TV_no_dn_y[doc_z, w] += -1
Y1T_no_dn_y[doc_z] += -1
self.DY1V[d, w] += -1 # delete w_y == 1 word
else:
raise ValueError("w_y not 0 or 1")
# conditional probability #
prob_w_y_unnorm = np.zeros([2],dtype=np.float32)
prob_w_y_unnorm[0] = (self.delta + YI_no_dn_y[0]) * (self.beta + Y0V_no_dn_y[w]) / \
(self.V * self.beta + YI_no_dn_y[0])
prob_w_y_unnorm[1] = (self.delta + YI_no_dn_y[1]) * (self.beta + Y1TV_no_dn_y[doc_z, w]) / \
(self.V * self.beta + Y1T_no_dn_y[doc_z])
prob_w_y = probNormalize(prob_w_y_unnorm)
# new sampled result #
try:
w_y_new = multinomial(prob_w_y)
except ValueError, e:
print prob_w_y_unnorm
print prob_w_y
print np.sum(prob_w_y), np.sum(prob_w_y) > 1.0
print YI_no_dn_y, self.YI, Y0V_no_dn_y[w], Y1TV_no_dn_y[doc_z,w], Y1T_no_dn_y[doc_z]
print d
raise e
# update #
self.y[d][n] = w_y_new
YI_no_dn_y[w_y_new] += 1
if w_y_new == 0:
Y0V_no_dn_y[w] += 1
elif w_y_new == 1:
Y1TV_no_dn_y[doc_z, w] += 1
Y1T_no_dn_y[doc_z] += 1
self.DY1V[d, w] += 1 # add back word with w_y_new == 1
else:
raise ValueError("w_y not 0 or 1")
self.YI, self.Y0V, self.Y1TV = YI_no_dn_y, Y0V_no_dn_y, Y1TV_no_dn_y
Y1T = Y1T_no_dn_y
doc_Y1V = self.DY1V.getrow(d).tocsr()
doc_Y1V_array = doc_Y1V.toarray().squeeze()
## sampling for z ##
# calculate leave-one out statistics #
TE_no_d_z, Y1TV_no_d_z, TI_no_d_z = self.TE, self.Y1TV, self.TI
TE_no_d_z[doc_z,:] += -docE
Y1TV_no_d_z[doc_z,:] += -doc_Y1V_array
TI_no_d_z[doc_z] += -1
# conditional probability #
prob_doc_z = self._prob_doc_z(TE_no_d_z, Y1TV_no_d_z, TI_no_d_z, docE, docW, doc_Y1V)
# new sampled result #
doc_z_new = multinomial(prob_doc_z)
# update #
self.z[d] = doc_z_new
TE_no_d_z[doc_z_new,:] += docE
Y1TV_no_d_z[doc_z_new, :] += doc_Y1V_array
TI_no_d_z[doc_z_new] += 1
self.TE, self.Y1TV, self.TI = TE_no_d_z, Y1TV_no_d_z, TI_no_d_z
