def get_child_message(x, y, hidden={}, visible={}):
with graph_context({**hidden, **visible}):
data = context(y)
log_likelihood = y.log_likelihood(data)
stats = x.statistics()
param = T.grad(T.sum(log_likelihood), [x.get_statistic(s) for s in stats])
return {s: param[i] for i, s in enumerate(stats)}
def expected_sufficient_statistics(self):
if self._estats is None:
natparam = self.get_parameters('natural', stop_gradient=True)
stats = list(self.statistics())
log_z = self.log_z('natural', stop_gradient=True)
grads = T.grad(log_z, [natparam[s] for s in stats])
self._estats = {s: g for s, g in zip(stats, grads)}
return self._estats
def get_child_message(x, y, data={}):
y_ = data[y]
stats = x.statistics()
log_likelihood = y.log_likelihood(y_)
param = T.grad(T.sum(log_likelihood), [x.get_statistic(s) for s in stats])
return {s: param[i] for i, s in enumerate(stats)}
def kl_gradients(self, q_X, q_A, kl, num_data):
return T.grad(kl, self.get_parameters())