def __init__(self,
number_of_states,
state_labels=None,
verbose=False,
random_transitions=False,
*args,
**kwargs):
self.current_state = None
if state_labels is None:
state_labels = ["State_%d" % i for i in range(number_of_states)]
assert len(set(state_labels)) == number_of_states
self.nStates = number_of_states
self.transitionMatrix = np.empty((self.nStates, self.nStates))
self.initialProbabilities = np.empty(shape=self.nStates)
if not random_transitions:
self.transitionMatrix.fill(1. / self.nStates)
self.initialProbabilities.fill(1. / self.nStates)
else:
self.transitionMatrix = random_simplex(self.nStates, two_d=True)
self.initialProbabilities = random_simplex(self.nStates)
self.emissionDistributions = np.array(
[None for _ in range(self.nStates)], dtype=object)
self.stateLabels = state_labels
self.verbose = verbose
def test_training(self):
self.nStates = 3
self.nDimensions = 1
from numpy.random import choice
observation_size = 10
hmm = self.new_hmm(random_transitions=True) # , random_emissions=True)
# hmm.setup_strict_left_to_right()
true_init_p = random_simplex(self.nStates)
true_states = [choice(range(self.nStates), p=true_init_p)]
true_trans_p = np.array([[.60, .30, .10], [.10, .80, .10],
[.30, .10, .60]
]) # random_simplex(self.nStates, two_d=True)
for i in range(1, observation_size):
true_states.append(
choice(range(self.nStates), p=true_trans_p[true_states[-1]]))
true_emission_p = np.array([
GaussianDistribution(dimensions=self.nDimensions, random=True)
for _ in range(self.nStates)
])
observations = [true_emission_p[state].emit() for state in true_states]
text = \
"""
True init probs:
{}
Ours:
{}: {}
True emission probs:
{}
Ours:
{}
True trans probs:
{}
Ours:
{}: {}
Observations ({}):
{}
"""
print text.format(
true_init_p, np.exp(hmm.initialProbabilities),
np.sum(np.abs(true_init_p - np.exp(hmm.initialProbabilities))),
true_emission_p, hmm.emissionDistributions, true_trans_p,
np.exp(hmm.transitionMatrix),
np.sum(np.abs(true_trans_p - np.exp(hmm.transitionMatrix))),
observation_size, observations)
hmm.train(observations, iterations=10, auto_stop=True)
print text.format(
true_init_p, np.exp(hmm.initialProbabilities),
np.sum(np.abs(true_init_p - np.exp(hmm.initialProbabilities))),
true_emission_p, hmm.emissionDistributions, true_trans_p,
np.exp(hmm.transitionMatrix),
np.sum(np.abs(true_trans_p - np.exp(hmm.transitionMatrix))),
observation_size, observations)
def __init__(self, number_of_states, state_labels=None, verbose=False, random_transitions=False,
*args, **kwargs):
self.current_state = None
if state_labels is None:
state_labels = ["State_%d" % i for i in range(number_of_states)]
assert len(set(state_labels)) == number_of_states
self.nStates = number_of_states
self.transitionMatrix = np.empty((self.nStates, self.nStates))
self.initialProbabilities = np.empty(shape=self.nStates)
if not random_transitions:
self.transitionMatrix.fill(1. / self.nStates)
self.initialProbabilities.fill(1. / self.nStates)
else:
self.transitionMatrix = random_simplex(self.nStates, two_d=True)
self.initialProbabilities = random_simplex(self.nStates)
self.emissionDistributions = np.array([None for _ in range(self.nStates)], dtype=object)
self.stateLabels = state_labels
self.verbose = verbose
def test_training(self):
self.nStates = 3
self.nSymbols = 2
from numpy.random import choice
observation_size = 1000
hmm = self.new_hmm(random_transitions=True, random_emissions=True)
# hmm.setup_strict_left_to_right()
true_init_p = random_simplex(self.nStates)
true_states = [choice(range(self.nStates), p=true_init_p)]
true_trans_p = np.array([[.50, .15, .35], [.10, .80, .10],
[.20, .10, .70]
]) # random_simplex(self.nStates, two_d=True)
for i in range(1, observation_size):
true_states.append(
choice(range(self.nStates), p=true_trans_p[true_states[-1]]))
true_emission_p = np.array([[.70, .30], [.40, .60], [
.10, .90
]]) # [random_simplex(self.nSymbols) for _ in range(self.nStates)]
observations = [
choice(range(self.nSymbols), p=true_emission_p[state])
for state in true_states
]
text = \
"""
True init probs:
{}
Ours:
{}
True emission probs:
{}
Ours:
{}
True trans probs:
{}
Ours:
{}
Observations ({}):
{}
"""
print text.format(
true_init_p, hmm.initialProbabilities, true_emission_p,
np.array([p.probabilities for p in hmm.emissionDistributions]),
true_trans_p, hmm.transitionMatrix, observation_size, observations)
hmm.train(observations, auto_stop=True, iterations=30)
print text.format(
true_init_p, hmm.initialProbabilities, true_emission_p,
np.array([p.probabilities for p in hmm.emissionDistributions]),
true_trans_p, hmm.transitionMatrix, observation_size, observations)
def test_training(self):
self.nStates = 3
self.nDimensions = 1
from numpy.random import choice
observation_size = 10
hmm = self.new_hmm(random_transitions=True) # , random_emissions=True)
# hmm.setup_strict_left_to_right()
true_init_p = random_simplex(self.nStates)
true_states = [choice(range(self.nStates), p=true_init_p)]
true_trans_p = np.array(
[[0.60, 0.30, 0.10], [0.10, 0.80, 0.10], [0.30, 0.10, 0.60]]
) # random_simplex(self.nStates, two_d=True)
for i in range(1, observation_size):
true_states.append(choice(range(self.nStates), p=true_trans_p[true_states[-1]]))
true_emission_p = np.array(
[GaussianDistribution(dimensions=self.nDimensions, random=True) for _ in range(self.nStates)]
)
observations = [true_emission_p[state].emit() for state in true_states]
text = """
True init probs:
{}
Ours:
{}: {}
True emission probs:
{}
Ours:
{}
True trans probs:
{}
Ours:
{}: {}
Observations ({}):
{}
"""
print text.format(
true_init_p,
np.exp(hmm.initialProbabilities),
np.sum(np.abs(true_init_p - np.exp(hmm.initialProbabilities))),
true_emission_p,
hmm.emissionDistributions,
true_trans_p,
np.exp(hmm.transitionMatrix),
np.sum(np.abs(true_trans_p - np.exp(hmm.transitionMatrix))),
observation_size,
observations,
)
hmm.train(observations, iterations=10, auto_stop=True)
print text.format(
true_init_p,
np.exp(hmm.initialProbabilities),
np.sum(np.abs(true_init_p - np.exp(hmm.initialProbabilities))),
true_emission_p,
hmm.emissionDistributions,
true_trans_p,
np.exp(hmm.transitionMatrix),
np.sum(np.abs(true_trans_p - np.exp(hmm.transitionMatrix))),
observation_size,
observations,
)
def __init__(self, n, randomize=False):
self.n = n
# initialize to 1/n
self.probabilities = np.array([1. / n] * n)
if randomize:
self.probabilities = random_simplex(n)
def __init__(self, n, randomize=False):
self.n = n
# initialize to 1/n
self.probabilities = np.array([1. / n] * n)
if randomize:
self.probabilities = random_simplex(n)
