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 _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 _restoreCheckPoint(self, filename = None):
if filename is None:
filename = self.checkpoint_file
state = cPickle.load(open(filename, "r"))
# restore #
self.theta = state["theta"]
self.phi = state["phi"]
self.eta = state["eta"]
self.eta_beta_inv = multivariateBeta_inv(self.eta)
self.alpha = state["alpha"]
self.beta = state["beta"]
self.z = state["z"]
self.TV = state["TV"]
self.TI = state["TI"]
epoch = state["epoch"]
ppl = state["ppl"]
print "restore state from file '%s' on epoch %d with ppl: %s" % (filename, epoch, str(ppl))
def _estimateGlobal(self):
self.theta = probNormalize(self.TI + self.alpha)
self.phi = probNormalize(self.TV + self.beta)
self.eta = self._etaUpdate()
self.eta_beta_inv = multivariateBeta_inv(self.eta)