def test_permutation():
a = numpy.array((3,2,1,6,5,4,10,0))
p = _permutation_by_sort(a)
a_permuted = _permute(a, p) # b should be in decreasing order
b = numpy.array([a+100*i for i in xrange(4)])
b_permuted = _permute(b, p, axis=1) # b's second axis should be in decreasing order
return p, a, a_permuted, b, b_permuted
def _reorder_topic_labels(self):
"Reorder the topic labels so that the largest topic is always first."
# calculate the permutation required to put the topics in order of size.
permutation = _permutation_by_sort(self.counts.E_n_k)
# don't bother re-ordering anything if permutation is the identity
if (numpy.arange(self.K) != permutation).any():
#start_LL = numpy.array(self._log_likelihood())
for d, q_zd in enumerate(self.q_z):
self.q_z[d] = _permute(q_zd, permutation, axis=1)
a_permuted = _permute(self.q_pi_bar.a, permutation, axis=0)
b_permuted = _permute(self.q_pi_bar.b, permutation, axis=0)
self.q_pi_bar.a = a_permuted
self.q_pi_bar.b = b_permuted
self.q_1_minus_pi_bar.a = b_permuted
self.q_1_minus_pi_bar.b = a_permuted
# clear the cached values (they all depend on the topic ordering)
# could also permute them rather than recalculating - that might
# be more efficient.
self._calculate_E_s_dk.cached_value = _permute(self.E_s_dk, permutation, axis=1)
self._calculate_E_t_kw.cached_value = _permute(self.E_t_kw, permutation, axis=0)
self._calculate_E_log_xi.cached_value = _permute(self.E_log_xi, permutation, axis=0)
self._calculate_G_pi.cached_value = _permute(self.G_pi, permutation, axis=0)
self._calculate_E_log_eta.clear_cached_value()
self.counts.permute(permutation)
# check topics are sorted in decreasing order of size
assert (self.counts.E_n_k[:-1] >= self.counts.E_n_k[1:]).all()
