def meanfieldupdate(self, bias_model, weight_model, impulse_model):
"""
Perform the mean field update.
:param bias_model:
:param weight_model:
:param impulse_model:
:return:
"""
# If necessary, initialize arrays for mean field parameters
if self.EZ is None:
self.EZ = np.zeros((self.T, self.K, self.K, self.B))
if self.EZ0 is None:
self.EZ0 = np.copy(self.S).astype(np.float)
# Uncomment this line to use python
# self._mf_update_Z_python(bias_model, weight_model, impulse_model)
# Use Cython
exp_E_log_lambda0 = np.exp(bias_model.expected_log_lambda0())
exp_E_log_W = np.exp(weight_model.expected_log_W())
exp_E_log_g = np.exp(impulse_model.expected_log_g())
mf_update_Z(self.EZ0, self.EZ, self.S, exp_E_log_lambda0, exp_E_log_W,
exp_E_log_g, self.F)
def test_parallel_parent_updates():
"""
Just run the parent updates in a parallel loop and watch htop
to make sure that all the cores are being used.
:return:
"""
T = 10000
K = 100
B = 3
S = np.random.poisson(2.0, size=((T,K))).astype(np.int)
EZ0 = np.zeros((T,K))
EZ = np.zeros((T,K,K,B))
exp_E_log_lambda0 = np.random.gamma(1.0, 1.0, size=(K))
exp_E_log_W = np.random.gamma(1.0, 1.0, size=(K,K))
exp_E_log_g = np.random.gamma(1.0, 1.0, size=(K,K,B))
F = np.random.gamma(1.0, 1.0, size=(T,K,B))
for itr in xrange(1000):
if itr % 10 == 0:
print "Iteration\t", itr
mf_update_Z(EZ0,
EZ,
S,
exp_E_log_lambda0,
exp_E_log_W,
exp_E_log_g,
F)
def _mf_update_Z(self):
"""
Update the mean field parameters for the latent parents
:return:
"""
bias_model, weight_model, impulse_model = \
self.model.bias_model, self.model.weight_model, self.model.impulse_model
exp_E_log_lambda0 = np.exp(bias_model.expected_log_lambda0())
exp_E_log_W = np.exp(weight_model.expected_log_W())
exp_E_log_g = np.exp(impulse_model.expected_log_g())
for k2, (Sk, Fk, EZk) in enumerate(zip(self.Ss, self.Fs, self.EZ)):
mf_update_Z(k2, EZk, Sk, exp_E_log_lambda0, exp_E_log_W,
exp_E_log_g, Fk)
self._check_EZ()
def _mf_update_Z(self):
"""
Update the mean field parameters for the latent parents
:return:
"""
bias_model, weight_model, impulse_model = \
self.model.bias_model, self.model.weight_model, self.model.impulse_model
K, B = self.K, self.B
exp_E_log_lambda0 = np.exp(bias_model.expected_log_lambda0())
exp_E_log_W = np.exp(weight_model.expected_log_W())
exp_E_log_g = np.exp(impulse_model.expected_log_g())
for k2, (Sk, Fk, EZk) in enumerate(zip(self.Ss, self.Fs, self.EZ)):
mf_update_Z(k2, EZk, Sk,
exp_E_log_lambda0,
exp_E_log_W,
exp_E_log_g,
Fk)
self._check_EZ()