def alternative_update_gmm(line,
unigram_tbl,
bigram_tbl,
link_tbl,
subst_tbl,
xe_gap,
max_iter,
prev_xe=np.inf):
c = 1
while True:
cnt_tbl, cnt_subst_tbl, prb_cf = em_iter_count(unigram_tbl, bigram_tbl,
link_tbl, subst_tbl)
link_tbl, gmm, weights = em_iter_update(cnt_tbl, line)
ll_gmm = gmm_log_likelihood(link_tbl, weights)
x_entropy = cross_entropy([prb_cf], [len(line)])
eprint('iter-GMM {} cross entropy is {}, gap {},'
' logP(c) {}, logP_GMM(c) {}'.format(
c, x_entropy, abs(1.0 - x_entropy / prev_xe), prb_cf,
ll_gmm))
if time_to_stop(c, max_iter, prev_xe, x_entropy, xe_gap):
break
elif np.isnan(x_entropy):
eprint('program end in iter {} caused by nan'.format(c))
break
else:
prev_xe = x_entropy
c += 1
return link_tbl, gmm, weights, x_entropy, prb_cf
def em_iter_gmm_count(link_tbl, weights):
"""
compute fractional counts for GMM.
:param link_tbl: \log p(g_i | z_j)
:param weights: p(z_j)
:return: count table and log likelihood
"""
ll_gmm = gmm_log_likelihood(link_tbl, weights)
log_weights = np.log(weights)
weighted_link_tbl = link_tbl + log_weights[:, np.newaxis]
cnt_tbl = weighted_link_tbl - logsumexp(weighted_link_tbl,
axis=0)[np.newaxis, :]
return cnt_tbl, ll_gmm
def em_decipher(line,
unigram_tbl,
bigram_tbl,
link_tbl,
subst_tbl,
xe_gap=1e-8,
max_iter=0):
"""
EM on a line of features.
EM iterations stop if matches one of the following conditions:
1) reach the max_iter
2) current cross entropy / last cross entropy >= xe_gap
:return: final link_tbl, gmm model, cross entropy, and log likelihood
"""
# prepare hyper-parameters
prev_xe = np.inf
# start training
c = 1
while True:
cnt_tbl, cnt_subst_tbl, prb_cf = em_iter_count(unigram_tbl, bigram_tbl,
link_tbl, subst_tbl)
_, gmm, weights = em_iter_update(cnt_tbl, line)
# subst_tbl = em_iter_update_subst(cnt_subst_tbl)
ll_gmm = gmm_log_likelihood(link_tbl, weights)
x_entropy = cross_entropy([prb_cf], [len(line)])
eprint('iter {} cross entropy is {}, gap {},'
' logP(c) {}, logP_GMM(c) {}'.format(
c, x_entropy, abs(1.0 - x_entropy / prev_xe), prb_cf,
ll_gmm))
if time_to_stop(c, max_iter, prev_xe, x_entropy, xe_gap):
break
elif np.isnan(x_entropy):
eprint('program end in iter {} caused by nan'.format(c))
break
else:
prev_xe = x_entropy
c += 1
return link_tbl, subst_tbl, gmm, weights, x_entropy, prb_cf