def initHMM(self, length):
a = 1.0 / length
# Transition probabilities
trans = np.array([[1-a, a, 0, 0], # Pre ->
[ 0, 1-a, a/2, a/2], # HQ ->
[ 0, 0, 1, 0], # PostQuiet ->
[ 0, 0, 0, 1] ]) # PostActive ->
# emission probabilities
eps = 1e-4
emit = np.array([[ 0.25, 0.25, 0.50 ], # Emit | Pre
[ 0.16, 0.84-eps, eps ], # Emit | HQ
[ 0.90, 0.10-eps, eps ], # Emit | PostQuiet
[ 0.25, 0.25, 0.50 ] ]) # Emit | PostActive
# A0 A1 A2
# Start state distribution
start = np.array([0.34, 0.33, 0.33, 0])
hmm = MultinomialHMM(n_components=nStates)
hmm.transmat_ = trans
hmm.startprob_ = start
hmm.emissionprob_ = emit
return hmm
def initHMM(self):
# self._hmm = MultinomialHMM(n_components=self._N, startprob_prior=None, transmat_prior=None,
# algorithm='viterbi', random_state=None, n_iter=self._maxIters, tol=0.01,
# verbose=True, params='ste', init_params='ste')
self._hmm = MultinomialHMM(n_components=self._N, n_iter=self._maxIters,
verbose=True, params='ste', init_params='ste')
def __init__(self, model_file=None, components=None):
if os.path.exists(model_file):
self.model = joblib.load(model_file)
else:
alu_file = 'Alu_sequence.pkl'
if os.path.exists(alu_file):
locis = joblib.load(alu_file)
else:
locis = read_sequence('hg19_Alu.bed', 0)
locis = random.sample(locis, 100000)
for l in tqdm(locis):
l.init_seq()
l.decode_seq()
locis = list(filter(lambda l: l.seq is not None, locis))
joblib.dump(locis, alu_file)
print('Alu Loaded')
locis = locis[0:5000]
model = MultinomialHMM(n_components=components,
verbose=True,
n_iter=50)
x = np.concatenate(list(map(attrgetter('seq'), locis)))
x = np.reshape(x, [x.shape[0], 1])
length = list(map(attrgetter('length'), locis))
model.fit(x, length)
self.model = model
joblib.dump(self.model, model_file)
def train(self, obs_seq_list: list, state_seq_list: list, obs_set: list,
state_set: list, file):
"""
:param obs_seq_list: observation sequence list [[o1, o2, o3], [o1, o2, o3]...]
:param state_seq_list: state sequence list [[s1, s2, s3], [s1, s2, s3]...]
:param obs_set: all possible observation state
:param state_set: all possible state
"""
self.obs_seq_list = obs_seq_list
self.state_seq_list = state_seq_list
self.obs_set = obs_set
self.state_set = state_set
self.counter = Counter(''.join(state_seq_list))
self.hmm = MultinomialHMM(n_components=len(self.state_set))
self.startprob, self.transmat, self.emissionprob = \
self._init_state(), self._trans_state(), self._emit_state()
self.hmm.startprob_ = self.startprob
self.hmm.transmat_ = self.transmat
self.hmm.emissionprob_ = self.emissionprob
if file is not None:
with open(file, 'wb') as f:
pickle.dump(self, f)
def buildHMM(HMMFactory):
model = MultinomialHMM(n_components=2, n_iter=200)
model.startprob_ = HMMFactory.hiddenProb()
model.transmat_ = HMMFactory.transMatrix()
model.emissionprob_ = HMMFactory.emissionMatrix()
return model
def main():
rand_p_matrix = np.random.rand(4, 4)
rand_b_matrix = np.random.rand(4, 3)
print("\nGernerating p matrix...............")
p_matrix = normalization(rand_p_matrix)
print(p_matrix)
print("\nGernerating b matrix...............")
b_matrix = normalization(rand_b_matrix)
print(b_matrix)
# Generate 1000 observations
O, _ = generate_observation(1000, p_matrix, b_matrix)
# training the selection of number of states
aic = []
bic = []
likelihood = []
m = 3
print("\nTraining the HMM for selection of number of states........")
for n in range(2, 30):
observations = LabelEncoder().fit_transform(O)
model = MultinomialHMM(n_components=n, random_state=200263453)
model.fit(np.atleast_2d(observations))
logL = model.score(np.atleast_2d(observations))
p = compute_p(n, m)
a = AIC(logL, p)
b = BIC(logL, observations, p)
likelihood.append(logL)
aic.append(a)
bic.append(b)
plot(aic, 'AIC')
plot(bic, 'BIC')
plot(likelihood, 'Log likelihood')
def __init__(self, n_iter=100): MultinomialHMM.__init__(self, n_components=len(self.states), n_iter=n_iter) self.voca = dict() self.word_freq = defaultdict(int) self.max_num_segs = 0 self.n_training = 0
def get_model(self):
""" returns a multinomial hmm"""
model = MultinomialHMM(n_components=self.get_max(),
params='e',
init_params='')
model.startprob_ = self.get_start()
model.transmat_ = self.get_transition()
model.emissionprob_ = self.get_emission()
return model
def get_model(self):
"""
初始化hmm模型
"""
model = MultinomialHMM(n_components=len(self.states))
model.startprob_ = self.init_p
model.transmat_ = self.trans_p
model.emissionprob_ = self.emit_p
return model
def get_model(self):
"""
初始化hmm模型
"""
model = MultinomialHMM(n_components=len(self.states))
model.startprob_ = self.init_p
model.transmat_ = self.trans_p
model.emissionprob_ = self.emit_p
return model
def fit_hmm_learn(X, n_states):
samples = np.concatenate(X)
lengths = [len(x) for x in X]
hmm_learn_model = MultinomialHMM(n_components=n_states)
hmm_learn_model.fit(samples, lengths)
# Label data using hmmlearn model
return hmm_learn_model.predict(samples, lengths)
def detect_events_hmm(mahal_timeseries, c_timeseries, global_pace_timeseries, threshold_quant=.95):
#Sort the keys of the timeseries chronologically
sorted_dates = sorted(mahal_timeseries)
(expected_pace_timeseries, sd_pace_timeseries) = getExpectedPace(global_pace_timeseries)
#Generate the list of values of R(t)
mahal_list = [mahal_timeseries[d] for d in sorted_dates]
c_list = [c_timeseries[d] for d in sorted_dates]
global_pace_list = [global_pace_timeseries[d] for d in sorted_dates]
expected_pace_list = [expected_pace_timeseries[d] for d in sorted_dates]
#Use the quantile to determine the threshold
sorted_mahal = sorted(mahal_list)
threshold = getQuantile(sorted_mahal, threshold_quant)
# The symbols array contains "1" if there is an outlier, "0" if there is not
symbols = []
for i in range(len(mahal_list)):
if(mahal_list[i] > threshold or c_list[i]==1):
symbols.append(1)
else:
symbols.append(0)
# Set up the hidden markov model. We are modeling the non-event states as "0"
# and event states as "1"
# Transition matrix with heavy weight on the diagonals ensures that the model
# is likely to stick in the same state rather than rapidly switching. In other
# words, the predictions will be relatively "smooth"
trans_matrix = array([[.999, .001],
[.001,.999]])
# Emission matrix - state 0 is likely to emit symbol 0, and vice versa
# In other words, events are likely to be outliers
emission_matrix = array([[.95, .05],
[.4, .6]])
# Actually set up the hmm
model = MultinomialHMM(n_components=2, transmat=trans_matrix)
model.emissionprob_ = emission_matrix
# Make the predictions
lnl, predictions = model.decode(symbols)
events = get_all_events(predictions, sorted_dates, mahal_list, global_pace_list,
expected_pace_list)
# Sort events by duration, starting with the long events
events.sort(key = lambda x: x[2], reverse=True)
return events, predictions
def predict(self, x, init_prob=None, method='hmmlearn', window=-1):
"""Predict result based on HMM
"""
if init_prob is None:
init_prob = np.array(
[1 / self.num_states for i in range(self.num_states)])
if method == 'hmmlearn':
model = MultinomialHMM(self.num_states, n_iter=100)
model.n_features = self.num_observations
model.startprob_ = init_prob
model.emissionprob_ = self.B
model.transmat_ = self.A
if window == -1:
result = model.predict(x)
else:
result = np.zeros(x.shape[0], dtype=np.int)
result[0:window] = model.predict(x[0:window])
for i in range(window, x.shape[0]):
result[i] = model.predict(x[i - window + 1:i + 1])[-1]
else:
if window == -1:
result = self.decode(x, init_prob)
else:
result = np.zeros(x.shape[0], dtype=np.int)
result[0:window] = self.decode(x[0:window], init_prob)
for i in range(window, x.shape[0]):
result[i] = self.decode(x[i - window + 1:i + 1],
init_prob)[-1]
return result
def train_hmm():
"""
HMM for sequence learning.
"""
print "Loading training data..."
train_sequence, num_classes = get_sequence("./train_data/*")
print "Build HMM..."
model = MultinomialHMM(n_components=2)
print "Train HMM..."
model.fit([train_sequence])
def build_hmm(trans, emis, seed=1):
"""Builds and returns hmm_model given the transition and emission probabilities matrices"""
hmm = MultinomialHMM(n_components=trans.shape[0],
algorithm="viterbi",
random_state=seed)
hmm.__setattr__("n_features", emis.shape[1])
hmm.__setattr__("emissionprob_", emis)
hmm.__setattr__("startprob_", np.array(
[1] + [0] *
(trans.shape[0] - 1))) # We will always start at the first state
hmm.__setattr__("transmat_", trans)
return hmm
def train_hmm():
"""
HMM for sequence learning.
"""
print "Loading training data..."
train_sequence, num_classes = get_sequence("./train_data/*")
print "Build HMM..."
model = MultinomialHMM(n_components=2)
print "Train HMM..."
model.fit([train_sequence])
def __init__(self, observed):
"""
Initializes the object and sets the internal state.
Args:
observed: array-like, shape (n_samples, n_features)
"""
self.observed = np.array(observed)
if len(self.observed.shape) == 1:
self.observed = self.observed.reshape(-1, 1)
# TODO: Check other parameters to this constructor
self.model = MultinomialHMM(n_components=2, n_iter=100)
class BKT:
"""
Implements the Bayesian Knowledge Tracing model. This only
implements the Viterbi and EM algorithms. These may be used
together to implement an Intelligent Tutoring System.
"""
def __init__(self, observed):
"""
Initializes the object and sets the internal state.
Args:
observed: array-like, shape (n_samples, n_features)
"""
self.observed = np.array(observed)
if len(self.observed.shape) == 1:
self.observed = self.observed.reshape(-1, 1)
# TODO: Check other parameters to this constructor
self.model = MultinomialHMM(n_components=2, n_iter=100)
def fit(self) -> None:
"""
Fits the model to the observed states. Uses the EM algorithm
to estimate model parameters.
"""
self.model.fit(self.observed)
def get_model_params(self) -> tuple:
"""
Returns the model parameters. This must be run only after
calling the `fit` function.
Returns:
(A, pi, B): The start probabilities, the transition
probabilities, and the emission probabilities.
"""
return np.round_(self.model.startprob_, 2), np.round_(self.model.transmat_, 2), \
np.round_(self.model.emissionprob_, 2)
def predict(self, sequence) -> np.array:
"""
Returns the most likely hidden state sequence corresponding to
`sequence`.
Args:
sequence: List of observable states
Returns:
state_sequence: Array
"""
return self.model.predict(sequence)
def test_HMM():
np.random.seed(12345)
np.set_printoptions(precision=5, suppress=True)
P = default_hmm()
ls, obs = P["latent_states"], P["obs_types"]
# generate a new sequence
O = generate_training_data(P, n_steps=30, n_examples=25)
tol = 1e-5
n_runs = 5
best, best_theirs = (-np.inf, []), (-np.inf, [])
for _ in range(n_runs):
hmm = MultinomialHMM()
A_, B_, pi_ = hmm.fit(O, ls, obs, tol=tol, verbose=True)
theirs = MHMM(
tol=tol,
verbose=True,
n_iter=int(1e9),
transmat_prior=1,
startprob_prior=1,
algorithm="viterbi",
n_components=len(ls),
)
O_flat = O.reshape(1, -1).flatten().reshape(-1, 1)
theirs = theirs.fit(O_flat, lengths=[O.shape[1]] * O.shape[0])
hmm2 = MultinomialHMM(A=A_, B=B_, pi=pi_)
like = np.sum([hmm2.log_likelihood(obs) for obs in O])
like_theirs = theirs.score(O_flat, lengths=[O.shape[1]] * O.shape[0])
if like > best[0]:
best = (like, {"A": A_, "B": B_, "pi": pi_})
if like_theirs > best_theirs[0]:
best_theirs = (
like_theirs,
{
"A": theirs.transmat_,
"B": theirs.emissionprob_,
"pi": theirs.startprob_,
},
)
print("Final log likelihood of sequence: {:.5f}".format(best[0]))
print("Final log likelihood of sequence (theirs): {:.5f}".format(
best_theirs[0]))
plot_matrices(P, best, best_theirs)
def rolling_score(
record: SeqRecord,
model: hmm.MultinomialHMM,
metadata: pd.DataFrame,
window_size: int = 1000,
overlap: int = 950,
) -> List[Dict[str, Union[str, float, int]]]:
scores = []
enc = {"A": 0, "C": 1, "G": 2, "T": 3}
sequence = np.array([enc.get(c, 0) for c in str(record.seq)])
for start, end in rolling.window_idx(len(sequence), window_size, overlap):
subsequence = sequence[start:end].reshape(-1, 1)
score = model.score(subsequence) / (end - start)
scores.append({
"id": record.id,
"start": start,
"score": score,
"relative_start": start / len(sequence),
**{
k: list(v.values())[0]
for k, v in metadata[metadata.aid == record.id].to_dict().items(
)
},
})
return scores
def __init__(self):
self.states = list(CapabilityBehaviourState)
self.observations = list(InteractionObservation)
self.hmm = MultinomialHMM(n_components=len(self.states))
self.state_history = []
# When many similar capabilities are being used, it is quite often
# the case that they will be in the same state. This means agent
# capabilities will have the same results as each other.
# This can lead to bad cases, for example, where VeryGood behaviours
# all fail at the same time.
# To prevent this synchronisation, each behaviour is given their own
# different initial seed to mix with the seed provided for an interaction.
self.individual_seed = 0
def hmm():
"""
vocabulary-acc = 0.9369
:return:返回mlp模型
"""
model = MultinomialHMM(n_components=2, n_iter=100, algorithm="viterbi")
return model
def get_hmm_model(state):
"""Creates an instance of MultinomialHMM, which follows sklearn interface
Input:
- state: dictionnary
where the keys are HiddenMarkovModelProbability choices
where the values are the probabilities matrices or arrays which
describes the according hidden markov model state
Returns: an instance of a trained MultinomialHMM
"""
hmm_model = MultinomialHMM(n_components=len(SleepStage))
hmm_model.emissionprob_ = state[HiddenMarkovModelProbability.emission.name]
hmm_model.startprob_ = state[HiddenMarkovModelProbability.start.name]
hmm_model.transmat_ = state[HiddenMarkovModelProbability.transition.name]
return hmm_model
def detect_events_hmm(mahal_timeseries, c_timeseries, global_pace_timeseries,
threshold_quant=.95, trans_matrix = DEFAULT_TRANS_MATRIX,
emission_matrix=DEFAULT_EMISSION_MATRIX, initial_state=None):
#Sort the keys of the timeseries chronologically
sorted_dates = sorted(mahal_timeseries)
(expected_pace_timeseries, sd_pace_timeseries) = getExpectedPace(global_pace_timeseries)
#Generate the list of values of R(t)
mahal_list = [mahal_timeseries[d] for d in sorted_dates]
c_list = [c_timeseries[d] for d in sorted_dates]
global_pace_list = [global_pace_timeseries[d] for d in sorted_dates]
expected_pace_list = [expected_pace_timeseries[d] for d in sorted_dates]
#Use the quantile to determine the threshold
sorted_mahal = sorted(mahal_list)
threshold = getQuantile(sorted_mahal, threshold_quant)
# The symbols array contains "1" if there is an outlier, "0" if there is not
symbols = []
for i in range(len(mahal_list)):
if(mahal_list[i] > threshold or c_list[i]==1):
symbols.append(1)
else:
symbols.append(0)
# Actually set up the hmm
model = MultinomialHMM(n_components=2, transmat=trans_matrix, startprob=initial_state)
model.emissionprob_ = emission_matrix
# Make the predictions
lnl, predictions = model.decode(symbols)
events = get_all_events(predictions, sorted_dates, mahal_list, global_pace_list,
expected_pace_list)
# Sort events by duration, starting with the long events
events.sort(key = lambda x: x[2], reverse=True)
return events, predictions
def detect_events_hmm(mahal_timeseries,
c_timeseries,
global_pace_timeseries,
threshold_quant=.95,
trans_matrix=DEFAULT_TRANS_MATRIX,
emission_matrix=DEFAULT_EMISSION_MATRIX,
initial_state=None):
#Sort the keys of the timeseries chronologically
sorted_dates = sorted(mahal_timeseries)
(expected_pace_timeseries,
sd_pace_timeseries) = getExpectedPace(global_pace_timeseries)
#Generate the list of values of R(t)
mahal_list = [mahal_timeseries[d] for d in sorted_dates]
c_list = [c_timeseries[d] for d in sorted_dates]
global_pace_list = [global_pace_timeseries[d] for d in sorted_dates]
expected_pace_list = [expected_pace_timeseries[d] for d in sorted_dates]
#Use the quantile to determine the threshold
sorted_mahal = sorted(mahal_list)
threshold = getQuantile(sorted_mahal, threshold_quant)
# The symbols array contains "1" if there is an outlier, "0" if there is not
symbols = []
for i in range(len(mahal_list)):
if (mahal_list[i] > threshold or c_list[i] == 1):
symbols.append(1)
else:
symbols.append(0)
# Actually set up the hmm
model = MultinomialHMM(n_components=2,
transmat=trans_matrix,
startprob=initial_state)
model.emissionprob_ = emission_matrix
# Make the predictions
lnl, predictions = model.decode(symbols)
events = get_all_events(predictions, sorted_dates, mahal_list,
global_pace_list, expected_pace_list)
# Sort events by duration, starting with the long events
events.sort(key=lambda x: x[2], reverse=True)
return events, predictions
def test_DiscreteHMM_decode(cases: str) -> None:
np.random.seed(12346)
cases = int(cases)
i = 1
N_decimal = 4
while i < cases:
tol=1e-3
n_samples = np.random.randint(10, 50)
hidden_states = np.random.randint(3, 6)
# symbols is the number of unqiue observation types.
symbols = np.random.randint(4, 9)
X = []
lengths = []
for _ in range(n_samples):
# the actual length is seq_length + 1
seq_length = symbols
this_x = np.random.choice(range(symbols), size=seq_length, replace=False)
X.append(this_x)
lengths.append(seq_length)
max_iter = 100
hmm_gold = MultinomialHMM(n_components=hidden_states, n_iter=100, tol=tol)
X_gold = np.concatenate(X).reshape((-1,1))
hmm_gold.fit(X_gold, lengths)
gold_A = hmm_gold.transmat_
gold_B = hmm_gold.emissionprob_
gold_pi = hmm_gold.startprob_
gold_logprob, gold_state_sequence = hmm_gold.decode(X_gold, lengths)
hmm_mine = DiscreteHMM(hidden_states=hidden_states,
symbols=symbols,
A=gold_A,
B=gold_B,
pi=gold_pi)
mine_logprob_list = []
mine_state_sequence = []
for this_x in X:
this_mine_logprob, this_mine_state_sequence = hmm_mine.decode(this_x)
mine_logprob_list.append(this_mine_logprob)
mine_state_sequence.append(this_mine_state_sequence)
mine_state_sequence = np.concatenate(mine_state_sequence)
mine_logprob = sum(mine_logprob_list)
assert_almost_equal(mine_logprob, gold_logprob, decimal=N_decimal)
assert_almost_equal(mine_state_sequence, gold_state_sequence, decimal=N_decimal)
i+=1
print('Successfully testing the function of computing decodes in discrete HMM!')
def __init__(self,
t,
theta,
rho,
algorithm='viterbi',
random_state=None,
n_iter=20, tol=0,
verbose=False):
MultinomialHMM.__init__(self, n_components=len(t)+1,
algorithm=algorithm,
random_state=random_state,
n_iter=n_iter, tol=tol,
verbose=verbose)
self.t = np.append(np.append([0], t), [np.inf])
self.tau = np.diff(self.t)
self.theta = theta
self.rho = rho
def main():
hmm = MultinomialHMM(n_components=5)
T = np.random.random(size=(5, 5))
T = T/T.sum(axis=1).reshape((5, 1))
hmm.transmat_ = T
pi = np.random.random(size=(5,))
pi = pi/pi.sum()
hmm.startprob_ = pi
emit = np.random.random(size=(5, 10))
emit = emit/emit.sum(axis=1).reshape((5, 1))
hmm.emissionprob_ = emit
X = np.zeros((20, 25)).astype(np.int)
for i in range(20):
x, _ = hmm.sample(n_samples=25)
X[i] = x.reshape((25,))
# load the PyTorch HMM
phmm = HMM(z_dim=5, x_dim=10)
phmm.T = torch.Tensor(T.T)
phmm.pi = torch.Tensor(pi)
phmm.emit = torch.Tensor(emit.T)
# compute PyTorch HMM forward-backward
my_marginals = phmm.log_marginal(torch.Tensor(X.T))
# compute hmmlearn version
true_marginals = np.zeros(20)
for i in range(20):
true_marginals[i] = hmm.score(X[i].reshape((-1, 1)))
assert np.abs(true_marginals - my_marginals.numpy()).max() < 1e-4
class CapabilityBehaviour:
def __init__(self):
self.states = list(CapabilityBehaviourState)
self.observations = list(InteractionObservation)
self.hmm = MultinomialHMM(n_components=len(self.states))
self.state_history = []
# When many similar capabilities are being used, it is quite often
# the case that they will be in the same state. This means agent
# capabilities will have the same results as each other.
# This can lead to bad cases, for example, where VeryGood behaviours
# all fail at the same time.
# To prevent this synchronisation, each behaviour is given their own
# different initial seed to mix with the seed provided for an interaction.
self.individual_seed = 0
def next_interaction(self, seed: int, t: float):
(x, state_sequence) = self.hmm.sample(1, random_state=seed ^ self.individual_seed)
assert len(state_sequence) == 1
assert len(x) == 1
assert len(x[0]) == 1
chosen_state = np.array([
1 if state_sequence[0] == n else 0
for n in range(len(self.states))
])
# Update the state of where the HMM is
self.hmm.startprob_ = chosen_state @ self.hmm.transmat_
self.state_history.append((t, self.states[state_sequence[0]]))
return self.observations[x[0][0]]
def peek_interaction(self, seed: int):
(x, state_sequence) = self.hmm.sample(1, random_state=seed ^ self.individual_seed)
assert len(state_sequence) == 1
assert len(x) == 1
assert len(x[0]) == 1
return self.observations[x[0][0]]
def main():
rand_p_matrix = np.random.rand(4, 4)
rand_b_matrix = np.random.rand(4, 3)
print("\nGernerating p matrix...............")
p_matrix = normalization(rand_p_matrix)
print(p_matrix)
print("\nGernerating b matrix...............")
b_matrix = normalization(rand_b_matrix)
print(b_matrix)
# Generate 1000 observations
O, Q = generate_observation(1000, p_matrix, b_matrix)
O_seq = [1, 2, 3, 3, 1, 2, 3, 3, 1, 2, 3, 3]
pi = (1, 0, 0, 0)
print("\nThe Orginal Observation Sequence O: {}".format(O[:12]))
print("The probability 𝑝(𝑂|𝜆) is {} with O: {}".format(
forward(O_seq, p_matrix, b_matrix, pi)[-1].sum(), O_seq))
print("\nThe Orginal Sequence Q: {}".format(Q[:12]))
print("The Most Probable Sequence Q: {} with O: {}".format(
list(viterbi(O_seq, p_matrix, b_matrix, pi)), O_seq))
obersvations = LabelEncoder().fit_transform(O)
model = MultinomialHMM(n_components=4)
model.fit(np.atleast_2d(obersvations))
est_pi = model.startprob_
est_p = model.transmat_
est_b = model.emissionprob_
print("\nThe estimated transition matrix P:\n {}".format(est_p))
print("\nThe estimated event matrix B:\n {}".format(est_b))
print("\nThe estimated start probability pi:\n {}".format(est_pi))
_, p = chisquare(p_matrix, est_p, axis=None)
print("\np-value of transition matrix P: {}".format(p))
_, p = chisquare(b_matrix, est_b, axis=None)
print("p-value of event matrix B: {}".format(p))
_, p = chisquare(pi, est_pi, axis=None)
print("p-value of start probability pi: {}".format(p))
def fit(seqs, n_components=1):
MultinomialHMM(n_components=n_components,
startprob_prior=1.0,
transmat_prior=1.0,
algorithm='viterbi',
random_state=1,
n_iter=100,
tol=0.01,
verbose=True,
params='ste',
init_params='ste')
def create_hmm_data(N, seq_len, x_dim, z_dim, params=None):
from hmmlearn.hmm import MultinomialHMM # introduces a lot of dependencies
hmm = MultinomialHMM(n_components=z_dim)
if params is None:
T = np.random.random(size=(z_dim, z_dim))
T = T/T.sum(axis=1).reshape((z_dim, 1))
pi = np.random.random(size=(z_dim,))
pi = pi/pi.sum()
emit = np.random.random(size=(z_dim, x_dim))
emit = emit/emit.sum(axis=1).reshape((z_dim, 1))
else:
T, pi, emit = params
hmm.transmat_ = T
hmm.startprob_ = pi
hmm.emissionprob_ = emit
X = np.zeros((N, seq_len)).astype(np.int)
for i in range(N):
x, _ = hmm.sample(n_samples=seq_len)
X[i] = x.reshape((seq_len,))
return (T, pi, emit), HMMData(X)
def test_viterbi_case_random(self):
for i in range(1000):
# init
self.n_state = np.random.randint(1,10)
self.n_output = np.random.randint(1,10)
self.step = np.random.randint(1,200)
p = np.random.random(self.n_state)
startprob = p/p.sum()
p = np.random.random((self.n_state,self.n_state))
transmat = p/p.sum(axis=1).reshape(-1,1)
p = np.random.random((self.n_state,self.n_output))
emissionprob = p/p.sum(axis=1).reshape(-1,1)
X = np.random.choice(self.n_output,self.step).reshape(-1,1)
# hmmlearn
model = MultinomialHMM(n_components=self.n_state,)
model.startprob_ = startprob
model.transmat_ = transmat
model.emissionprob_ = emissionprob
y = model.predict(X)
# my hmm
hmm = HMM()
pred = hmm.viterbi(startprob, transmat, emissionprob, X)
self.assertTrue(np.array_equal(y, pred))
def predict_prob(self, x, init_prob=None, window=-1):
"""Predict the probability
"""
if init_prob is None:
init_prob = np.array(
[1 / self.num_states for i in range(self.num_states)])
model = MultinomialHMM(self.num_states)
model.n_features = self.num_observations
model.startprob_ = init_prob
model.emissionprob_ = self.B
model.transmat_ = self.A
return model.predict_proba(x)
def predict(self, day_to_predict):
# Get records of 30 days before day_to_predict
previous_thirty_days = get_previous_month(self.time_series, day_to_predict)
binary_crime_sequence = previous_thirty_days['Violent Crime Committed?'].values.tolist()
# Unsupervised HMM can't account for string of identical emissions.
# If we see such a string, just predict the same emission for the following day.
if binary_crime_sequence == [1]*30:
return True
if binary_crime_sequence == [0]*30:
return False
votes = []
# Train nine HMMs. They are initialized randomly, so we take "votes" from nine HMMs.
# Why 9? Odd numbers preclude ties.
# And nine is a decent tradeoff between performance and getting bad results by chance
for _ in range(3):
# Train HMM
model = MultinomialHMM(n_components=3, n_iter=10000)
model.fit([np.array(binary_crime_sequence)])
# Determine the most likely state of the last day in the sequence
last_state_probs = model.predict_proba(binary_crime_sequence)[-1]
current_state = self.get_most_likely(last_state_probs)
# Determine the most likely state of the day we're trying to predict
transition_probs = model.transmat_[current_state]
next_state = self.get_most_likely(transition_probs)
# Determine the most likely emission (crime/no crime) from a day in that state
emissions = model.emissionprob_[next_state]
vote = self.get_most_likely(emissions)
# Record this HMM's vote
votes.append(vote)
# Votes are 1 for crime, 0 for no crime. Return True if majority votes for crime.
return sum(votes) > 1
def __init__(self,
n_components=1,
startprob_prior=1.0,
transmat_prior=1.0,
algorithm="viterbi",
random_state=None,
n_iter=10,
tol=1e-2,
verbose=False,
params="ste",
init_params="ste"):
MultinomialHMM.__init__(self,
n_components=n_components,
startprob_prior=startprob_prior,
transmat_prior=transmat_prior,
algorithm=algorithm,
random_state=random_state,
n_iter=n_iter,
tol=tol,
verbose=verbose,
params=params,
init_params=init_params)
return
def calculate_hmm_m(training_set, test_set, taxonomy, cursor, connection, settings):
da_id_taxonomy = find_da_id(taxonomy, cursor)
states, start_probability, transition_probability = start_transition_probability_extraction(training_set, taxonomy)
n_states = len(states)
feature_list, emissions = extract_features_training_set(training_set, taxonomy, states, settings)
# print model.transmat_
con_pathes, test_obs, emissions = extract_features_test_set(test_set, taxonomy, feature_list, emissions, settings)
model = MultinomialHMM(n_components=n_states)
model._set_startprob(start_probability)
model._set_transmat(transition_probability)
model._set_emissionprob(emissions)
da_predictions(test_obs, model, con_pathes, states, da_id_taxonomy, taxonomy, cursor, connection)
def run_hmm_model(input_df, n_unique, A_df, Eta, n_iter = 10000,
tol=1e-2, verbose = False, params = 'e', init_params = ''):
'''
Runs the hmm model and returns the predicted results, score and model
input_df : The dataframe of keypresses
n_unique : number of unqique chars
A_df : Dataframe of trasnmission matrix
Eta : Emissions matrix
n_iter : Max number of iterations for hmm
tol : The value to stop the hmm model if score does not improve by more than this
verbose : Whether or not to print out
params : Parameters to tune
init_params : Paramters to initialize
'''
# Propotion of characters starting words in english
char_counts = get_char_counts()
# Construct model
hmm = MultinomialHMM(n_components=n_unique, startprob_prior=np.append(0, char_counts.values),
transmat_prior=A_df.values, algorithm='viterbi',
random_state=None, n_iter=n_iter, tol=tol,
verbose=verbose, params=params, init_params=init_params)
# Set values
hmm.emissionprob_ = Eta
hmm.transmat_ = A_df.values
hmm.startprob_ = np.append(0, char_counts.values)
# Feed in the clusters as the expected output
model_input = input_df['cluster'].values
# Reshape
if len(model_input.shape) == 1:
model_input = model_input.reshape((len(model_input), 1))
# Fit the model
hmm = hmm.fit(model_input)
# Score model
score, results = hmm.decode(model_input)
return score, results, hmm
def buildHMM(num_states, n_iter=10, tol=0.01):
model = MultinomialHMM(n_components=num_states, n_iter=n_iter, tol=tol)
model.n_features = 3
return model
smoothing_tolerance = 1 #number of indices
sampling_interval = 3600 #seconds
discrete_obs, delta_hws, delta_fas = [], [], []
for idx in mice:
d = _data_on_mouse(data, idx, smoothing_time_radius,
smoothing_amplitude_radius, smoothing_tolerance,
sampling_interval, bins)
discrete_obs.append(d[0])
delta_hws.append(d[1])
delta_fas.append(d[2])
X = np.array(discrete_obs)
model = MultinomialHMM(n_components = n_components)
predictions = []
for i in range(7):
held_out_X = np.vstack((X[:i], X[i+1:]))
model.fit(held_out_X)
predictions.append(model.decode(X[i].reshape(X[i].shape[0], 1)))
f, axarr = plt.subplots(7, 1)
yranges = np.arange(n_components+1, dtype=float)/n_components
colors = plt.cm.rainbow(np.linspace(0, 1, n_components))
for i in range(7):
states, indices = _axvspan_maker(predictions[i][1])
for s, idxs in zip(states, indices):
axarr[i].axvspan(idxs[0], idxs[1], ymin=yranges[s], ymax=yranges[s+1], color=colors[s])
plt.show()