def iter_forward_algorithm(model, n, verbose=False):
# calc first position
nstates = model.get_num_states(0)
col1 = [
model.prob_prior(0, j) + model.prob_emission(0, j)
for j in xrange(nstates)
]
if n > 20:
step = (n // 20)
else:
step = 1
# loop through positions
nstates1 = nstates
for i in xrange(1, n):
if verbose and i % step == 0:
print " forward iter=%d/%d, lnl=%f" % (i + 1, n, max(col1))
nstates2 = model.get_num_states(i)
# find total transition and emission
col2 = []
for k in xrange(nstates2):
tot = -util.INF
emit = model.prob_emission(i, k)
for j in xrange(nstates1):
p = col1[j] + model.prob_transition(i - 1, j, i, k) + emit
tot = logadd(tot, p)
col2.append(tot)
yield col2
col1 = col2
nstates1 = nstates2
def iter_forward_algorithm(model, n, verbose=False):
# calc first position
nstates = model.get_num_states(0)
col1 = [model.prob_prior(0, j) + model.prob_emission(0, j)
for j in xrange(nstates)]
if n > 20:
step = (n // 20)
else:
step = 1
# loop through positions
nstates1 = nstates
for i in xrange(1, n):
if verbose and i % step == 0:
print " forward iter=%d/%d, lnl=%f" % (i+1, n, max(col1))
nstates2 = model.get_num_states(i)
# find total transition and emission
col2 = []
for k in xrange(nstates2):
tot = -util.INF
emit = model.prob_emission(i, k)
for j in xrange(nstates1):
p = col1[j] + model.prob_transition(i-1, j, i, k) + emit
tot = logadd(tot, p)
col2.append(tot)
yield col2
col1 = col2
nstates1 = nstates2
def sample_hmm_next_state(model, pos, state):
nstates = model.get_num_states(pos)
state2 = 0
p = model.prob_transition(pos - 1, state, pos, state2)
pick = log(random.random())
while pick > p and state2 < nstates:
state2 += 1
p = logadd(p, model.prob_transition(pos - 1, state, pos, state2))
return state2
def sample_hmm_first_state(model):
state = 0
nstates = model.get_num_states(0)
p = model.prob_prior(0, state)
pick = log(random.random())
while pick > p and state < nstates:
state += 1
p = logadd(p, model.prob_prior(0, state))
return state
def sample_hmm_next_state(model, pos, state):
nstates = model.get_num_states(pos)
state2 = 0
p = model.prob_transition(pos-1, state, pos, state2)
pick = log(random.random())
while pick > p and state2 < nstates:
state2 += 1
p = logadd(p, model.prob_transition(pos-1, state, pos, state2))
return state2
def sample_hmm_first_state(model):
state = 0
nstates = model.get_num_states(0)
p = model.prob_prior(0, state)
pick = log(random.random())
while pick > p and state < nstates:
state += 1
p = logadd(p, model.prob_prior(0, state))
return state
def get_posterior_probs(model, n, verbose=False):
probs_forward = forward_algorithm(model, n, verbose=verbose)
probs_backward = backward_algorithm(model, n, verbose=verbose)
total_prob = -util.INF
for j in xrange(model.get_num_states(0)):
total_prob = logadd(
total_prob,
model.prob_prior(0, j) + model.prob_emission(0, j) +
probs_backward[0][j])
probs_post = [[
probs_forward[i][j] + probs_backward[i][j] - total_prob
for j in xrange(model.get_num_states(i))
] for i in xrange(n)]
return probs_post
def get_posterior_probs(model, n, verbose=False):
probs_forward = forward_algorithm(model, n, verbose=verbose)
probs_backward = backward_algorithm(model, n, verbose=verbose)
total_prob = -util.INF
for j in xrange(model.get_num_states(0)):
total_prob = logadd(total_prob,
model.prob_prior(0, j) +
model.prob_emission(0, j) +
probs_backward[0][j])
probs_post = [
[probs_forward[i][j] + probs_backward[i][j] - total_prob
for j in xrange(model.get_num_states(i))]
for i in xrange(n)]
return probs_post
def backward_algorithm(model, n, verbose=False):
probs = []
# calc last position
nstates = model.get_num_states(n - 1)
for i in xrange(n):
probs.append(None)
probs[n - 1] = [
model.prob_prior(n - 1, j) + model.prob_emission(n - 1, j)
for j in xrange(nstates)
]
if n > 20:
step = (n // 20)
else:
step = 1
# loop through positions
for i in xrange(n - 2, -1, -1):
if verbose and i % step == 0:
print " backward iter=%d/%d, lnl=%f" % (i + 1, n, max(
probs[i + 1]))
nstates1 = model.get_num_states(i)
nstates2 = model.get_num_states(i + 1)
col2 = probs[i + 1]
# find total transition and emission
col1 = []
emit = [model.prob_emission(i + 1, k) for k in xrange(nstates2)]
for j in xrange(nstates1):
tot = -util.INF
for k in xrange(nstates2):
p = col2[k] + emit[k] + model.prob_transition(i, j, i + 1, k)
tot = logadd(tot, p)
col1.append(tot)
probs[i] = col1
return probs
def backward_algorithm(model, n, verbose=False):
probs = []
# calc last position
nstates = model.get_num_states(n-1)
for i in xrange(n):
probs.append(None)
probs[n-1] = [model.prob_prior(n-1, j) + model.prob_emission(n-1, j)
for j in xrange(nstates)]
if n > 20:
step = (n // 20)
else:
step = 1
# loop through positions
for i in xrange(n-2, -1, -1):
if verbose and i % step == 0:
print " backward iter=%d/%d, lnl=%f" % (i+1, n, max(probs[i+1]))
nstates1 = model.get_num_states(i)
nstates2 = model.get_num_states(i+1)
col2 = probs[i+1]
# find total transition and emission
col1 = []
emit = [model.prob_emission(i+1, k) for k in xrange(nstates2)]
for j in xrange(nstates1):
tot = -util.INF
for k in xrange(nstates2):
p = col2[k] + emit[k] + model.prob_transition(i, j, i+1, k)
tot = logadd(tot, p)
col1.append(tot)
probs[i] = col1
return probs
