def train(self):
mmModel = np.zeros((noteRange+1, noteRange+1))
mm3Model = np.zeros((noteRange+1, noteRange+1, noteRange+1))
hmmModel = HMM(12, noteRange+1)
obs = []
ground = []
actions = []
for i in range(minNote, maxNote):
actions.append(i)
qModel = QLearner(actions, epsilon=0.1, alpha=0.2, gamma=0.9)
for ls in self.clusterData:
for quadidx, quad in enumerate(ls):
tempquad = map(lambda x: x - minNote, quad) #take this out for prevnote stuff
obs.append(tempquad[1:]) #this is for hmm: you can also do same thing for qlearning to change state that way
tempquad = map(lambda x: (x - minNote) % 12, quad)
ground.append(tempquad[:3])
if (quad):
for idx, note in enumerate(quad):
if idx > 0:
currNote = note
prevNote = quad[idx - 1]
#Q learning
#q.learn(state1, action1, reward, state2)
qModel.learn(prevNote, note, 1, note)
#Markov model
mmModel[currNote - minNote, prevNote - minNote] += 1
if idx > 2:
#Markov model, more order
currNote = note - minNote
prevNote = quad[idx - 1] - minNote
prevNote2 = quad[idx - 2] - minNote
mm3Model[currNote, prevNote, prevNote2] += 1
hmmModel.learn(obs, ground)
return (mmModel, mm3Model, hmmModel, qModel)
class HMMPOSTagger(object):
"""
在中文分词结果基础上, 采用 HMM 模型实现词性标注 (Part-of-speech tagging).
"""
def __init__(self):
self.hmm = HMM()
self.re_chinese = re.compile(ur"([\u4E00-\u9FA5]+)") # 正则匹配汉字串
self.re_skip = re.compile(ur"([\.0-9]+|[a-zA-Z0-9]+)") # 正则匹配英文串和数字串
def load(self, model_dir):
"""
加载模型文件.
"""
self.hmm.load(model_dir)
def pos_tag(self, words):
"""
基于 HMM 模型的词性标注.
"""
log_prob, pos_list = self._viterbi(words)
for i, w in enumerate(words):
yield (w, pos_list[i])
def main(job_id, params):
num_runs = 20
obs_length = 100
num_states = 2
num_obs = 2
# readin hmm indx
t = 0
try:
with open(os.path.join('.', 'hmm_index.txt')) as hmm_index_file:
t = int(hmm_index_file.read())
sys.stderr.write("!!!!!!!!!!!!!!!!!!HMM INDEX: " + str(t) + " !!!!!!!!!!!!!!!\n")
except IOError:
t = 0
# generate HMM observations
np.random.seed(0x6b6c26b2)
seeds = np.random.randint(0x0fffffff, size=num_runs)
np.random.seed(seeds[t])
# random hmm
z_mat, t_mat = random_hmm(num_states, num_obs)
pi_vec = np.array([1.0 / num_states] * num_states)
hmm_test = HMM(z_mat, t_mat, pi_vec)
# random obs trajectory
obs = hmm_test.generate(obs_length)[np.newaxis,:]
# calculate log likelihood for input HMM parameters
z_mat_p_input = np.array([[params['z_mat_p_0'][0], params['z_mat_p_1'][0]]])
t_mat_p_input = np.array([[params['t_mat_p_0'][0], params['t_mat_p_1'][0]]])
# pi_vec_input = np.array([params['pi_0'], 1 - params['pi_0']])
hmm_estimate = make_parameterized_HMM(z_mat_p_input, t_mat_p_input, pi_vec)
hmm_loglikelihood = hmm_estimate.loglikelihood(obs[0])
return -hmm_loglikelihood
def prepare_seqs_nl_dbg(self, decoding="viterbi"):
params_fixed = (np.load("{}ip.npy".format(self.path)),
np.load("{}tp.npy".format(self.path)),
np.load("{}fp.npy".format(self.path)),
np.load("{}ep.npy".format(self.path)))
h = HMM(self.n_states, self.n_obs, params=params_fixed, writeout=False)
h.dirname = self.path
self.ner_corpus = Conll2002NerCorpus(self.dataset.x_dict)
# train_seq = self.ner_corpus.read_sequence_list_conll(ned_train)
dev_seq = self.ner_corpus.read_sequence_list_conll(ned_dev)
test_seq = self.ner_corpus.read_sequence_list_conll(ned_test)
if decoding == "viterbi":
decoder = h.viterbi_decode_corpus
elif decoding == "max_emission":
decoder = h.max_emission_decode_corpus
elif decoding == "posterior":
decoder = h.posterior_decode_corpus
elif decoding == "posterior_cont":
decoder = h.posterior_cont_decode_corpus
elif decoding == "posterior_cont_type":
decoder = h.posterior_cont_type_decode_corpus
else:
print("Decoder not defined, using Viterbi.")
decoder = h.viterbi_decode_corpus
# print("Decoding word representations on train.")
# decoder(train_seq)
print("Decoding word representations on dev.")
decoder(dev_seq)
def simple_weather_model():
hmm = HMM(["s1", "s2"], ["R", "NR"])
init = [0.7, 0.3]
trans = [[0.8, 0.2], [0.1, 0.9]]
observ = [[0.75, 0.25], [0.4, 0.6]]
hmm.set_hidden_model(init, trans, observ)
return hmm
def main():
parser = OptionParser()
parser.add_option("-d", dest="training", help="training data directory")
parser.add_option("-k", dest="K", type="int", help="number of latent states", default=6)
parser.add_option("-a", dest="a", type="float", help="Dirichlet parameter", default=1.0)
parser.add_option("-i", dest="I", type="int", help="iteration count", default=10)
parser.add_option("-m", dest="model", help="model data filename to save")
(options, args) = parser.parse_args()
if not options.training: parser.error("need training data directory(-d)")
features = load_data(options.training)
hmm = HMM()
hmm.set_corpus(features)
hmm.init_inference(options.K, options.a)
pre_L = -1e10
for i in range(options.I):
log_likelihood = hmm.inference()
print i, ":", log_likelihood
if pre_L > log_likelihood: break
pre_L = log_likelihood
if options.model:
hmm.save(options.model)
else:
hmm.dump()
def main():
hmm = HMM(*train(sys.argv[1]))
with open(sys.argv[2]) as f:
correct = 0
wrong = 0
correct_sents = 0
wrong_sents = 0
correct_known = 0
wrong_known = 0
for i, sent in enumerate(Reader(f)):
prob, path = hmm.decode([word for (word, pos) in sent])
correct1 = 0
wrong1 = 0
for (gold, predicted) in zip(sent, path):
if gold == predicted:
correct1 += 1
else:
wrong1 += 1
print('%e\t%.3f\t%s' % (prob, correct1 / (correct1 + wrong1), ' '.join('%s/%s' % pair for pair in path)))
if prob > 0:
correct_sents += 1
correct_known += correct1
wrong_known += wrong1
else:
wrong_sents += 1
correct += correct1
wrong += wrong1
print("Correctly tagged words: %s" % (correct / (correct + wrong)))
print("Sentences with non-zero probability: %s" % (correct_sents / (correct_sents + wrong_sents)))
print("Correctly tagged words when only considering sentences with non-zero probability: %s" % (correct_known / (correct_known + wrong_known)))
def toy_model(self):
hmm = HMM(["s1", "s2"], ["R", "NR"])
init = [0.5, 0.5]
trans = [[0.2, 0.8], [0.8, 0.2]]
observ = [[0.8, 0.2], [0.2, 0.8]]
hmm.set_hidden_model(init, trans, observ)
return hmm
def __init__(self, limb, name='Demo_26_Final'):
rospy.init_node(name, anonymous=True)
self._startGesture = '0'
self._limb = limb
self._knhObj = KinectNiteHelp()
self._baeObj = BaxterArmEndpoint(self._limb)
self._dmObj = Demo26Help()
self._lhObj = LeapHelp()
self._handCoordinates = []
self._baxterCoordinates = []
self._posCount = 0
rtMatFile = open("RTMatFile.dat", "r")
self._rotMat = cPickle.load(rtMatFile)
self._transMat = cPickle.load(rtMatFile)
self._gCount = 0
self._gOn = 0
self._gPointCount = 0
self._gPoints = []
self._hmmObjG1 = HMM('G1.hmm')
self._hmmObjG2 = HMM('G2.hmm')
self._hmmObjG3 = HMM('G3.hmm')
self._hmmObjG4 = HMM('G4.hmm')
self._flag = '0'
self._pub = rospy.Publisher('/robot/xdisplay', Image, latch=True)
#self._flagPub = rospy.Publisher('flag_topic', String)
self._sub = rospy.Subscriber('/key_tap_topic', String, self._callback)
rtMatFile.close()
img = cv.LoadImage('Welcome.png')
msg = cv_bridge.CvBridge().cv_to_imgmsg(img, encoding="bgr8")
self._pub.publish(msg)
# Sleep to allow for image to be published.
rospy.sleep(3)
def __init__(self, **kwarg): # lname, url, other prior knowledge
super(HMMClassifier, self).__init__()
self.HMMauthor = HMM('author', 2)
self.HMMvenue = HMM('venue', 2) # Not important
self.HMMentire = HMM('entire', 6) # Set empirically
self.observations_raw = []
self.observation_sequences = []
self.labels = []
class TestHMM(): def __init__(self): self.Z = numpy.array([ [0.8, 0.09, 0.01], [0.09, 0.8, 0.01], [0.1, 0, 0.8] ]) self.b = numpy.array([ [0.1, 0.1, 0.8], [0.05, 0.9, 0.05], [0.8, 0.1, 0.1] ]) self.pi = numpy.array([0.9,0.05,0.05]) self.T = 2000 # we want the errors to be less than 20% self.error_threshold = 0.2 def setup(self): self.model = HMM(self.Z,self.b,self.pi) def gen_states_obs(self): states = [] obsvns = [] for (s,o) in self.model.gen(self.T): states.append(s) obsvns.append(o) return states, obsvns def test_init(self): self.model = HMM(self.Z,self.b,self.pi) def test_gen(self): self.setup() states = [] obsvns = [] for (s,o) in self.model.gen(10): states.append(s) obsvns.append(o) assert len(states) == 10 assert len(obsvns) == 10 def test_forward_backward(self): self.setup() states, obsvns = self.gen_states_obs() alpha,beta = self.model.forward_backward(obsvns) gamma = [a*b/sum(a*b) for a,b in zip(alpha,beta)] state_est = numpy.array([numpy.where(g==max(g))[0][0] for g in gamma]) err = sum(state_est != numpy.array(states))/float(len(states)) assert err < self.error_threshold def test_viterbi(self): self.setup() states, obsvns = self.gen_states_obs() state_est = self.model.viterbi(obsvns) err = sum(state_est != numpy.array(states))/float(len(states)) assert err < self.error_threshold
def test_simple_hmm_learning(self):
state_seq = [[0, 1, 1, 0, 1, 0, 1, 1], [0, 0, 1, 0]]
obs_seq = [[0, 0, 1, 1, 0, 0, 0, 1], [0, 1, 0, 0]]
hmm = HMM(range(2), range(2))
hmm.learn_from_labeled_data(state_seq, obs_seq)
print hmm
eps = 0.00001
self.assertTrue(max_delta(hmm.initial, [0.750000, 0.250000]) < eps)
self.assertTrue(max_delta(hmm.transition, [[0.285714, 0.714286], [0.571429, 0.428571]]) < eps)
self.assertTrue(max_delta(hmm.observation, [[0.625000, 0.375000], [0.625000, 0.375000]]) < eps)
def test_hmm():
m = HMM(2, 2)
observations = [[0,0,0,0,0,1,1,1,1,1,0,1,0,0,0,1,0,1,1,1,1],[0,0,0,0,1,0,1,1,0,1,1,0,0,1,0,0,1,1,1,1,0,0,1,0,0]]
ground = [[0,0,0,0,0,1,1,1,1,1,0,0,0,0,0,1,1,1,1,1,1],[0,0,0,0,1,1,1,1,1,1,1,0,0,0,0,0,1,1,1,1,0,0,0,0,0]]
m.learn(observations, ground, smooth=None)
trueres = ([0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0, 0], -21.944)
res = m.viterbi(observations[1])
assert trueres[0] == res[0]
print trueres[1]
print res[1]
assert abs(trueres[1] - res[1]) < 0.1
def setUp(self):
self._model_filename = "hmm_m4n4.pkl"
self._train_filename = "m4n4.train.data"
self._num_hidden = 4
self._num_observ = 4
transition_matrix = np.random.rand(4, 4)
observation_matrix = np.random.rand(4, 4)
hmm = HMM(self._num_hidden, self._num_observ, transition_matrix=transition_matrix,
observation_matrix=observation_matrix)
sequences = hmm.generate_data(10000, 4, 51)
io.save_sequences(self._train_filename, sequences)
HMM.to_file(self._model_filename, hmm)
def main(job_id, params):
num_runs = 20
obs_length = 100
num_states = 2
num_obs = 2
# readin hmm indx
t = 0
try:
with open(os.path.join('.', 'hmm_index.txt')) as hmm_index_file:
t = int(hmm_index_file.read())
sys.stderr.write("!!!!!!!!!!!!!!!!!!HMM INDEX: " + str(t) + " !!!!!!!!!!!!!!!\n")
except IOError:
t = 0
# generate HMM observations
np.random.seed(0x6b6c26b2)
seeds = np.random.randint(0x0fffffff, size=num_runs)
np.random.seed(seeds[t])
# random hmm
z_mat, t_mat = random_hmm(num_states, num_obs)
pi_vec = np.array([1.0 / num_states] * num_states)
hmm_test = HMM(z_mat, t_mat, pi_vec)
# random obs trajectory
obs = hmm_test.generate(obs_length)[np.newaxis, :]
# calculate log likelihood for input HMM parameters
z_mat_p_input = np.array([[params['z_mat_p_0'][0], params['z_mat_p_1'][0]]])
t_mat_p_input = np.array([[params['t_mat_p_0'][0], params['t_mat_p_1'][0]]])
# pi_vec_input = np.array([params['pi_0'], 1 - params['pi_0']])
hmm_estimate = make_parameterized_HMM(z_mat_p_input, t_mat_p_input, pi_vec)
hmm_loglikelihood = hmm_estimate.loglikelihood(obs[0])
# use the current suggest point and run EM to get a new point
hmm_em_est, _, _ = em.em(hmm_estimate, hmm_estimate.z_mat, hmm_estimate.t_mat, obs, 30, 0.1)
em_est_z_mat, em_est_t_mat = retrieve_parameterized_HMM(hmm_em_est)
em_est_ll = -hmm_em_est.loglikelihood(obs[0])
em_est_z_mat.reshape((em_est_z_mat.size,))
em_est_t_mat.reshape((em_est_t_mat.size,))
print em_est_t_mat
print em_est_z_mat
historical_points = [{'params': {}}]
# write z_mat
for i, v in enumerate(em_est_z_mat[0]):
historical_points[0]['params']['z_mat_p_' + str(i)] = {'values': np.array([v]), 'type': 'float'}
# write t_mat
for i, v in enumerate(em_est_t_mat[0]):
historical_points[0]['params']['t_mat_p_' + str(i)] = {'values': np.array([v]), 'type': 'float'}
historical_points[0]['value'] = em_est_ll
dump_new_history('.', historical_points)
return -hmm_loglikelihood
def main():
parser = OptionParser()
parser.add_option("-t", dest="test", help="test data directory")
parser.add_option("-m", dest="model", help="model data filename to save")
(options, args) = parser.parse_args()
if not options.model: parser.error("need model data filename(-m)")
hmm = HMM()
hmm.load(options.model)
if options.test:
tests = load_data(options.test)
for x in tests:
print zip(x, hmm.Viterbi(hmm.words2id(x)))
def train_hmm_from_data(data_filename, debug=False):
if debug:
print "\n\nReading dataset %s ..." % data_filename
data_filename = normalize_filename(data_filename)
d = DataSet(data_filename)
#if options.verbose:
# print d
if debug:
print "Building an HMM from the full training data..."
hmm = HMM(d.states, d.outputs)
hmm.learn_from_labeled_data(d.train_state, d.train_output)
if debug:
print "The model:"
print hmm
return (hmm, d)
def hmmBaumWelchRestarts(nRuns, stopping, obs, parallel=True):
'''Performs the baum welch nRuns times stopping when the likelihood
changes by less than stopping'''
models = [HMM.randomStudent() for _ in xrange(nRuns)]
return baumWelchRestarts(models, obs, stopping, parallel)
def main():
hmm = HMM(3, ('up', 'down', 'unchanged'),
initial_probability=[0.5, 0.2, 0.3],
transition_probability=[[0.6, 0.2, 0.2],
[0.5, 0.3, 0.2],
[0.4, 0.1, 0.5]],
observation_probability=[[0.7, 0.1, 0.2],
[0.1, 0.6, 0.3],
[0.3, 0.3, 0.4]])
observation = ("up", "up", "unchanged", "down", "unchanged", "down", "up")
ob_length = len(observation)
p, _ = hmm.forward(observation, ob_length)
path = hmm.decode(observation, ob_length)
print("P{} = {:.13f}".format(tuple(observation), p))
print("Observation sequence =", tuple(i+1 for i in path))
def test_accumulative(self):
hmm = HMM.from_file(self._model_filename)
for i in xrange(self._num_hidden):
self.assertAlmostEqual(1.0, hmm._accumulative_transition_matrix[-1, i],
delta=1e-6)
self.assertAlmostEqual(1.0, hmm._accumulative_observation_matrix[-1, i],
delta=1e-6)
def hmm_train(results):
train_file = results.train
freq_file = results.freq
logger.debug( 'Started training HMM with options:' + "\n" +
'training file: ' + str(train_file) + "\n" +
'frequency file:' + str(freq_file) + "\n")
if not os.path.exists('model/hmm-model'):
classifier = HMM()
classifier.train(train_file,freq_file)
logger.info("Done Training, model is written in model file")
model = classifier.get_theta()
write_obj(model, 'hmm-model')
else:
logger.info('model already exists, nothing to do!')
def test_slfit(self):
sequences = io.load_sequences(self._train_filename)
hmm = HMM.from_file(self._model_filename)
learner = SLHMM(self._num_hidden, self._num_observ)
learner.fit(sequences, verbose=True)
for sequence in sequences:
pprint("True probability: %f" % hmm.predict(sequence))
pprint("Infered probability: %f" % learner.predict(sequence))
def __init__(self, hmm, rule_set=None):
if isinstance(hmm, HMM):
self.hmm = hmm
else:
self.hmm = HMM(hmm)
segment_table = SegmentTable()
self.segment_symbol_length = uniform_encoding.log2(len(segment_table) + 1) # + 1 for the delimiter
if rule_set:
self.rule_set = rule_set
else:
self.rule_set = RuleSet(noise=False)
noises = configurations.get("NOISE_RULE_SET", [])
self.noise_rule_set = RuleSet.load_noise_rules_from_flat_list(noises)
self._cached_hmm_transducer = None
self._cached_rule_set_transducer = None
self._cached_noise_rule_set_transducer = None
def hmm_test(results):
test_file = results.test
freq_file = results.freq
logger.debug( 'Started testing HMM with options:' + "\n" +
'test file: ' + str(results.test) + "\n" +
'frequency file:' + str(freq_file) + "\n")
logger.info("Loading model")
model = read_obj('hmm-model')
classifier = HMM()
classifier.load_theta(model)
classifier.test(test_file,freq_file)
def testViterbi(self):
"""Test viterbi algorithm
Come up with a good non-trivial way to test your training function
You can use the given icecream dataset or make up your own
Add your implementation
"""
seq1 = Instance(label = "he saw a dog".split(), data = ['NN', 'VB', 'D', 'NN'])
seq2 = Instance(label = "the dog sees a saw".split(), data = ['D', 'NN', 'VB', 'D', 'NN'])
seq3 = Instance(label = "the saw saw a dog".split(), data = ['D', 'NN', 'VB', 'D', 'NN'])
seq4 = Instance(label = "he saw the saw".split(), data = ['NN', 'VB', 'D', 'NN'])
instances = [seq1, seq2, seq3, seq4]
hmm = HMM()
hmm.train(instances)
sawind = hmm.label_alphabet.get_index('saw')
self.assertNotEqual(hmm.emission_matrix[sawind, 0], 0, 'WRONG')
self.assertNotEqual(hmm.emission_matrix[sawind, 1], 0, 'WRONG')
def __init__(self, demos, n_state, gamma, n_offspring):
'''
Base RL class. Creates an hmm inside. Explores and updates hmm parameters. Generates motion for an
episode. Stores rollout (or episode) information
:param demos: List of 2D numpy arrays.
Each 2D numpy array is (n_keyframe, n_dim)
for action model, n_dim = 7
for goal model, n_dim = 8
:param n_state: Integer, number of possible hidden states.
:param gamma: Initial covariance matrix multiplier
:param n_offspring: Number of offsprings (rollouts) in an episode
'''
self.hmm = HMM(demos, n_state, gamma)
self.n_offspring = n_offspring
self.reset_rollout()
def testSupervisedTraining(self): """Test parameter fitting Come up with a good non-trivial way to test your training function You can use the given icecream dataset or make up your own Add your implementation """ seq1 = Instance(label = ['odd', 'even', 'odd', 'even', 'odd'], data = [3, 2, 1, 4, 1]) seq2 = Instance(label = ['even', 'even', 'odd', 'odd', 'even'], data = [2, 4, 1, 3, 2]) seq3 = Instance(label = ['even', 'even', 'odd', 'odd', 'odd'], data = [1, 2, 3, 4, 3]) seq4 = Instance(label = ['odd', 'odd', 'even', 'even', 'even'], data = [4, 3, 4, 1, 2]) instances = [seq1, seq2, seq3, seq4] hmm = HMM() hmm.train(instances) mystery = Instance(data = [2, 1, 3, 4, 2, 2, 1, 3]) labels = hmm.classify_instance(mystery) self.assertEqual(labels, ['even', 'odd', 'odd', 'even', 'even', 'even', 'odd', 'odd'], 'NOOO')
def test_accumulative(self):
hmm = HMM.from_file(self._model_filename)
for i in xrange(self._num_hidden):
self.assertAlmostEqual(1.0,
hmm._accumulative_transition_matrix[-1, i],
delta=1e-6)
self.assertAlmostEqual(1.0,
hmm._accumulative_observation_matrix[-1, i],
delta=1e-6)
def test_morpheme_boundary(self):
configurations["MORPHEME_BOUNDARY_FLAG"] = True
self.initialise_segment_table("plural_english_segment_table.txt")
hmm = HMM({
INITIAL_STATE: ['q1'],
'q1': (['q2', FINAL_STATE], ['dog', 'kat']),
'q2': ([FINAL_STATE], ['z'])
})
grammar = Grammar(hmm, [])
def main():
parser = argparse.ArgumentParser()
parser.add_argument('path', type=str, help='the path to testing data')
args = parser.parse_args()
test__data_path = args.path
test_words = load_raw_data(test__data_path)
hmm = HMM()
hmm.load_model('./hmmmodel.txt')
tag_result = []
for sentence in test_words:
re = hmm.decode(sentence)
tag_result.append(re)
save_result(test_words, tag_result)
def prepare_seqs_en(self, decoding="viterbi"):
params_fixed = (np.load("{}/ip.npy".format(self.path)),
np.load("{}/tp.npy".format(self.path)),
np.load("{}/fp.npy".format(self.path)),
np.load("{}/ep.npy".format(self.path)))
h = HMM(self.n_states,
self.n_obs,
params=params_fixed,
writeout=False,
dirname=self.path)
self.ner_corpus = Conll2003NerCorpus(self.dataset.x_dict)
train_seq = self.ner_corpus.read_sequence_list_conll(eng_train)
dev_seq = self.ner_corpus.read_sequence_list_conll(eng_dev)
test_seq = self.ner_corpus.read_sequence_list_conll(eng_test)
muc_seq = self.ner_corpus.read_sequence_list_conll(
muc_test) if self.use_muc else None
decoder = None
type_decoder = None
if decoding == "viterbi":
decoder = h.viterbi_decode_corpus
elif decoding == "max_emission":
decoder = h.max_emission_decode_corpus
elif decoding == "posterior":
decoder = h.posterior_decode_corpus
elif decoding == "posterior_cont":
decoder = h.posterior_cont_decode_corpus
elif decoding == "posterior_cont_type":
type_decoder = h.posterior_cont_type_decode_corpus
else:
print("Decoder not defined correctly, using Viterbi.")
decoder = h.viterbi_decode_corpus
print(
"Decoding word representations on train. This may take a while...")
type_decoder(
train_seq, self.dataset,
self.logger) if type_decoder is not None else decoder(train_seq)
print("Decoding word representations on dev.")
type_decoder(
dev_seq, self.dataset,
self.logger) if type_decoder is not None else decoder(dev_seq)
print("Decoding word representations on test.")
type_decoder(
test_seq, self.dataset,
self.logger) if type_decoder is not None else decoder(test_seq)
if self.use_muc:
print("Decoding word representations on MUC.")
type_decoder(
muc_seq, self.dataset,
self.logger) if type_decoder is not None else decoder(muc_seq)
return train_seq, dev_seq, test_seq, muc_seq
def model_training(train_data, tags):
"""
Train HMM based on training data
Inputs:
- train_data: (1*num_sentence) a list of sentences, each sentence is an object of line class
- tags: (1*num_tags) a list of POS tags
Returns:
- model: an object of HMM class initialized with parameters(pi, A, B, obs_dict, state_dict) you calculated based on train_data
"""
model = None
###################################################
# Edit here
N=len(train_data)
S=len(tags)
pi=np.zeros(S)
A=np.zeros((S,S))
state_dict={}
obs=[]
obs_dict={}
o=0
for t in range(S):
state_dict[tags[t]]=t
for line in train_data:
pi[state_dict[line.tags[0]]]+=1
for w in range(line.length-1):
A[state_dict[line.tags[w]],state_dict[line.tags[w+1]]]+=1
for line in train_data:
for w in range(line.length):
if line.words[w] not in obs_dict.keys():
obs_dict[line.words[w]]=o
o+=1
pi=pi/N
A=(A.T/np.sum(A, axis=1)).T
O=len(obs_dict)
B=np.zeros((S,O))
for line in train_data:
for w in range(line.length):
B[state_dict[line.tags[w]],obs_dict[line.words[w]]]+=1
B=(B.T/np.sum(B,axis=1)).T
a1=np.isnan(A)
A[a1]=0
b1=np.isnan(B)
B[b1]=0
model=HMM(pi,A,B,obs_dict,state_dict)
return model
def setUp(self):
"""Initialize Eisner ice cream HMM (J & M, Figure 6.3)"""
self.hmm = HMM()
# These variables have many aliases. J & M call them π, A, B, Q, and V.
# You don't need to use these names, but you do need to provide a way
# of initializing them.
self.hmm.train(
[],
initial_probabilities=[.8, .2], # P(Hot, Cold)
transition_probabilities=[
[.7, .3], # P(Hot|Hot, Cold)
[.4, .6]
], # P(Cold|Hot, Cold)
emission_probabilities=[
[.2, .4, .4], # P(1, 2, 3|Hot)
[.5, .4, .1]
], # P(1, 2, 3|Cold)
states=("Hot", "Cold"),
vocabulary=(1, 2, 3))
def test_predict(self):
sequences = io.load_sequences(self._train_filename)
hmm = HMM.from_file(self._model_filename)
for sequence in sequences:
self.assertEqual(hmm.predict(sequence), hmm.predict(sequence),
"HMM.prediction Error")
sequences = [[0, 1], [1, 2, 3, 0], [0, 0, 0, 1]]
for sequence in sequences:
self.assertEqual(hmm.predict(sequence), hmm.predict(sequence),
"HMM.prediction Error")
def speech_tagging(test_data, model, tags):
"""
Inputs:
- test_data: (1*num_sentence) a list of sentences, each sentence is an object of line class
- model: an object of HMM class
Returns:
- tagging: (num_sentence*num_tagging) a 2D list of output tagging for each sentences on test_data
"""
tagging = []
###################################################
# Edit here
###################################################
N, M = model.B.shape
new_model = model
new_column = 1e-6 * np.ones([N, 1])
new_feature_number = 0
new_b = model.B
new_obs_dict = model.obs_dict
for sentence in test_data:
for word in sentence.words:
if word not in model.obs_dict:
# add new features and set number of new features
# add new column to b
# sample : np.append(???,new_column,axis=1)
new_b = np.append(new_b, new_column, axis=1)
# add new features to obs_dict
new_obs_dict[word] = len(new_b[0, :]) - 1
# augment new features number
new_feature_number += 1
if new_feature_number != 0:
new_model = HMM(model.pi, model.A, new_b, new_obs_dict,
model.state_dict)
for sentence in test_data:
tag_row = new_model.viterbi(sentence.words)
tagging.append(tag_row)
return tagging
def __get_best_pos_to_shoot(self):
"""Returns a position in which is more likely to shoot the enemy.
"""
#Gets the state of the markov model at time t.
transition_probabilities = self.__get_net_probs()
emission_probabilities = self.__get_net_probs()
hmm = HMM(transition_probabilities, emission_probabilities)
emissions = [2, 1, 0]
initial = self.__get_net_probs()
return (self.net[self.viterbi(hmm, initial, emissions)[0]].id)
def hmm_test():
st_time = time.time()
model_file = "hmm_model.json"
# load data
with open(model_file, 'r') as f:
data = json.load(f)
A = np.array(data['A'])
B = np.array(data['B'])
pi = np.array(data['pi'])
# observation symbols
obs_dict = data['observations']
# state symbols
states_symbols = dict()
for idx, item in enumerate(data['states']):
states_symbols[item] = idx
Osequence = np.array(data['Osequence'])
N = len(Osequence)
model = HMM(pi, A, B, obs_dict, states_symbols)
delta = model.forward(Osequence)
m_delta = np.array([[3.5000e-01, 1.3600e-01, 0.0000e+00, 0.0000e+00, 1.1136e-05, 1.1136e-05, 0.0000e+00],
[1.5000e-01, 3.2000e-02, 4.6400e-03, 2.7840e-04, 3.3408e-05, 1.1136e-05, 8.9088e-07]])
print("Your forward function output:", delta)
print("My forward function output:", m_delta)
gamma = model.backward(Osequence)
m_gamma = np.array([[1.6896e-06, 3.8400e-06, 6.4000e-05, 2.0000e-03, 1.4000e-02, 2.0000e-02, 1.0000e+00],
[1.9968e-06, 1.1520e-05, 1.9200e-04, 3.2000e-03, 2.2000e-02, 6.0000e-02, 1.0000e+00]])
print("Your backward function output:", gamma)
print("My backward function output:", m_gamma)
prob1 = model.sequence_prob(Osequence)
m_prob1 = 8.908800000000002e-07
print("Your sequence_prob function output:", prob1)
print("My sequence_prob function output:", m_prob1)
prob2 = model.posterior_prob(Osequence)
m_prob2 = np.array([[0.6637931, 0.5862069, 0., 0., 0.175, 0.25, 0.],
[0.3362069, 0.4137931, 1., 1., 0.825, 0.75, 1.]])
print("Your posterior_prob function output:", prob2)
print("My posterior_prob function output:", m_prob2)
viterbi_path = model.viterbi(Osequence)
m_viterbi_path = ['1', '1', '2', '2', '2', '2', '2']
print('Your viterbi function output: ', viterbi_path)
print('My viterbi function output: ', m_viterbi_path)
en_time = time.time()
print()
print("hmm total time: ", en_time - st_time)
def test_hmm_connected_components(self):
hmm = HMM({INITIAL_STATE: ['q1'],
'q1': (['q2', FINAL_STATE], ['dag', 'kot']),
'q2': (['q1', FINAL_STATE], ['z'])})
log_hmm(hmm)
component_states = hmm.get_connected_components(ignore_initial_and_final_states=False)
self.write_to_dot_to_file(hmm, 'connected_hmm')
print(component_states)
assert component_states[0][0] == FINAL_STATE
assert component_states[2][0] == INITIAL_STATE
assert 'q1' in component_states[1]
assert 'q2' in component_states[1]
component_states = hmm.get_connected_components(ignore_initial_and_final_states=True)
assert 'q1' in component_states[0]
assert 'q2' in component_states[0]
assert [INITIAL_STATE] not in component_states
assert [FINAL_STATE] not in component_states
def test_predict(self):
sequences = io.load_sequences(self._train_filename)
hmm = HMM.from_file(self._model_filename)
for sequence in sequences:
self.assertEqual(hmm.predict(sequence), hmm.predict(sequence),
"HMM.prediction Error")
sequences = [[0,1], [1,2,3,0], [0,0,0,1]]
for sequence in sequences:
self.assertEqual(hmm.predict(sequence), hmm.predict(sequence),
"HMM.prediction Error")
def train_hmm(X,
k,
prev_hmm=None,
window_size=5,
num_clusters=5,
num_states=4,
max_iter=10):
"""Trains an HMM
Args:
X: the data
k: the number of sequences to divide the data into
prev_hmm: an HMM object to set the parameters
window_size: the window size of exponential weighting
num_clusters: the number of mixtures in each GMM
num_states: the number of states in the HMM
max_iter: the max number of iterations allowed in the HMM
Returns:
the trained HMM
"""
temp_X = X.reshape((k, X.shape[0] // k, X.shape[1]))
num_features = X.shape[1]
if (prev_hmm is None):
pi = random_splits(num_states, 1)
A = np.array([random_splits(num_states, 1) for _ in range(num_states)])
weights = np.random.rand(num_states, num_clusters) / num_clusters
means = np.random.rand(num_states, num_clusters,
num_features) * .6 - .3
cov = np.tile(np.eye(num_features), (num_states, num_clusters, 1, 1))
for i in range(num_states):
weights[i] = weights[i] / weights[i].sum()
else:
pi = prev_hmm.pi
A = prev_hmm.A
weights = prev_hmm.weights
means = prev_hmm.means
cov = prev_hmm.cov
hmm = HMM(pi, A, weights, means, cov, num_states)
hmm.em(X, window_size=3, max_iter=max_iter)
return hmm
def test_init(self):
hmm = HMM(UDDataSet('data/en-ud-train.conllu'))
self.assertEqual(17, hmm.num_state)
self.assertEqual(17, hmm.bos_idx)
self.assertEqual(18, hmm.eos_idx)
for i in range(hmm.num_state):
self.assertAlmostEqual(-12.228919653600784,
hmm.emission_counter[(i, -1)])
def train_N_state_hmms_from_data(filename, num_states, debug=False):
""" reads all the data, then split it up into each category, and then
builds a separate hmm for each category in data """
dataset = DataSet(filename)
category_seqs = split_into_categories(dataset)
# Build a hmm for each category in data
hmms = {}
for cat, seqs in category_seqs.items():
if debug:
print "\n\nLearning %s-state HMM for category %s" % (
num_states, cat)
model = HMM(range(num_states), dataset.outputs)
model.learn_from_observations(seqs, debug)
hmms[cat] = model
if debug:
print "The learned model for %s:" % cat
print model
return (hmms, dataset)
def model_training(train_data, tags):
"""
Train HMM based on training data
Inputs:
- train_data: (1*num_sentence) a list of sentences, each sentence is an object of line class
- tags: (1*num_tags) a list of POS tags
Returns:
- model: an object of HMM class initialized with parameters(pi, A, B, obs_dict, state_dict) you calculated based on train_data
"""
model = None
###################################################
N = len(tags)
A = np.ones((N, N)) / N
pi = np.ones(N) / N
state_dict, tag_dict, obs_dict = {}, {}, {}
word_list = []
for idx, tag in enumerate(tags):
state_dict[tag] = idx
for cur_line in train_data:
pi[state_dict[cur_line.tags[0]]] += 1
for idx in range(cur_line.length):
tag = cur_line.tags[idx]
word_list.append(cur_line.words[idx])
if tag not in tag_dict:
tag_dict[tag] = 1
else:
tag_dict[tag] += 1
if idx < cur_line.length - 1:
A[tags.index(cur_line.tags[idx]),
tags.index(cur_line.tags[idx + 1])] += 1
word_list = list(set(word_list))
for idx, word in enumerate(word_list):
obs_dict[word] = idx
total_tags = sum(tag_dict.values())
for key in tag_dict.keys():
tag_dict[key] /= total_tags
B = np.zeros([N, len(word_list)])
for line in train_data:
for word, tag in zip(line.words, line.tags):
B[state_dict[tag], obs_dict[word]] = tag_dict[tag]
A /= np.sum(A, axis=1)[:, None]
pi /= len(train_data)
model = HMM(pi, A, B, obs_dict, state_dict)
###################################################
return model
def test_morpheme_boundary(self):
self.configurations["MORPHEME_BOUNDARY_FLAG"] = True
self.initialise_segment_table("plural_english_segment_table.txt")
hmm = HMM({
INITIAL_STATE: ['q1'],
'q1': (['q2', FINAL_STATE], ['dog', 'kat']),
'q2': ([FINAL_STATE], ['z'])
})
grammar = Grammar(hmm)
self.assertCountEqual(['dog', 'kat', 'dogz', 'katz'],
grammar.get_all_outputs())
def test_morphology_only2(self):
self.initialise_segment_table("plural_english_segment_table.txt")
self.configurations["DATA_ENCODING_LENGTH_MULTIPLIER"] = 25
data = [u'tozata', u'tozaso', u'tozakt', u'tozzookata', u'tozzookaso', u'tozzookakt', u'tozzook', u'tozdodata', u'tozdodaso', u'tozdodakt', u'tozdod', u'tozgosata', u'tozgosaso', u'tozgosakt', u'tozgos', u'toz', u'dagata', u'dagaso', u'dagakt', u'dagzookata', u'dagzookaso', u'dagzookakt', u'dagzook', u'dagdodata', u'dagdodaso', u'dagdodakt', u'dagdod', u'daggosata', u'daggosaso', u'daggosakt', u'daggos', u'dag', u'gasata', u'gasaso', u'gasakt', u'gaszookata', u'gaszookaso', u'gaszookakt', u'gaszook', u'gasdodata', u'gasdodaso', u'gasdodakt', u'gasdod', u'gasgosata', u'gasgosaso', u'gasgosakt', u'gasgos', u'gas', u'kodata', u'kodaso', u'kodakt', u'kodzookata', u'kodzookaso', u'kodzookakt', u'kodzook', u'koddodata', u'koddodaso', u'koddodakt', u'koddod', u'kodgosata', u'kodgosaso', u'kodgosakt', u'kodgos', u'kod', u'katata', u'kataso', u'katakt', u'katzookata', u'katzookaso', u'katzookakt', u'katzook', u'katdodata', u'katdodaso', u'katdodakt', u'katdod', u'katgosata', u'katgosaso', u'katgosakt', u'katgos', u'kat', u'dotata', u'dotaso', u'dotakt', u'dotzookata', u'dotzookaso', u'dotzookakt', u'dotzook', u'dotdodata', u'dotdodaso', u'dotdodakt', u'dotdod', u'dotgosata', u'dotgosaso', u'dotgosakt', u'dotgos', u'dot']
hmm = HMM({'q0': [u'q1'],
'q1': ([u'q2', u'q3', u'qf'], ['toz', 'dag', 'kat', 'dot', 'kod', 'gas']),
'q2': ([u'q3',u'qf'], ['zook', 'gos', 'dod']),
'q3': ([u'qf'], ['aso', 'akt', 'ata'])})
self.configurations.simulation_data = data
hypothesis = Hypothesis(Grammar(hmm, []))
def model_training(train_data, tags):
#####lowercased######
for data in train_data:
for it in range(data.length):
data.words[it] = data.words[it].lower()
#####################
S = len(tags)
pi = np.zeros(S)
A = np.zeros([S, S])
B = []
Bc = np.zeros([S, 1])
Ac = np.zeros([S, S])
obs_dict = {}
states_symbols = {}
for i in range(S):
if not tags[i] in states_symbols.keys():
states_symbols[tags[i]] = i
numS = np.zeros(S)
num1S = np.zeros(S)
####################################
for data in train_data:
firsttag = data.tags[0]
num1S[states_symbols[firsttag]] += 1
for i in range(data.length):
word = data.words[i]
tag = data.tags[i]
if not word in obs_dict.keys():
obs_dict[word] = len(obs_dict)
Bc = np.append(Bc, np.zeros([S, 1]), axis=1)
Bc[states_symbols[tag], obs_dict[word]] += 1
numS[states_symbols[tag]] += 1
if i != data.length - 1:
Ac[states_symbols[tag], states_symbols[data.tags[i + 1]]] += 1
B = np.zeros(np.shape(Bc))
pi = normalize(num1S)
for s in range(S):
for sp in range(S):
if numS[s] == 0:
A[s, sp] = 0
else:
A[s, sp] = Ac[s, sp] / numS[s]
for s in range(len(Bc)):
for o in range(len(Bc[0])):
if numS[s] == 0:
B[s, o] = 0
else:
B[s, o] = Bc[s, o] / numS[s]
###################################
model = HMM(pi, A, B, obs_dict, states_symbols)
return model
def test_crossover_subgraph(self):
self.initialise_segment_table("plural_english_segment_table.txt")
hmm_1 = HMM({INITIAL_STATE: ['q1'],
'q1': (['q1', 'q2'], ['da']),
'q2': ([FINAL_STATE], ['s'])})
hmm_2 = HMM({INITIAL_STATE: ['q1'],
'q1': (['q2'], ['ko']),
'q2': (['q3'], ['bo']),
'q3': (['q4'], ['go']),
'q4': ([FINAL_STATE], ['z'])})
offspring_1, offspring_2 = HMM.crossover_subgraphs(hmm_1, hmm_2)
self.write_to_dot_to_file(hmm_1, 'subgraph_parent_1')
self.write_to_dot_to_file(hmm_2, 'subgraph_parent_2')
self.write_to_dot_to_file(offspring_1, 'subgraph_offspring_1')
self.write_to_dot_to_file(offspring_2, 'subgraph_offspring_2')
offspring_1.get_transducer()
offspring_2.get_transducer()
def predict_crimes(cls, crimes, limit=16):
# Sort crimes in chronological order
crimes.sort(key=lambda crime: crime.date)
# Find the average duration between crimes
deltas = [later.date - now.date for now, later in zip(crimes, crimes[1:])]
delta = sum(deltas, timedelta()) / len(deltas)
# Only allow a granularity of 1 hour
delta = max(delta, timedelta(hours=1))
# Create a timeline that marks each of the given events, and also
# includes empty non-events at regular intervals (determined by `delta`)
# between the first and last timestamp
time, stop = crimes[0].date, crimes[-1].date
timeline = {crime.date : crime for crime in crimes}
while time <= stop:
timeline.setdefault(time, None)
time += delta
# Convert the padded timeline back to an ordered time sequence of events
events = [timeline[k] for k in sorted(timeline.keys())]
# Create an HMM to predict the regions in which crimes will take place
regions = [e.region if e is not None else None for e in events]
regions = iter(HMM.from_events(regions))
# Create a separate HMM for the crimes' descriptions
descs = [e.description for e in events if e is not None]
descs = iter(HMM.from_events(descs))
future = []
while len(future) < limit:
time += delta
region = next(regions)
if region is not None:
future.append({
'date': time.strftime(cls.DATE_FMT),
'region': region.to_dict(),
'description': next(descs)
})
return future
def prepare_seqs_en_dbg(self, decoding="viterbi"):
params_fixed = (np.load("{}ip.npy".format(self.path)),
np.load("{}tp.npy".format(self.path)),
np.load("{}fp.npy".format(self.path)),
np.load("{}ep.npy".format(self.path)))
h = HMM(self.n_states, self.n_obs, params=params_fixed, writeout=False)
h.dirname = self.path
self.ner_corpus = Conll2003NerCorpus(self.dataset.x_dict)
# train_seq = self.ner_corpus.read_sequence_list_conll(eng_train)
dev_seq = self.ner_corpus.read_sequence_list_conll(eng_dev)
# test_seq = self.ner_corpus.read_sequence_list_conll(eng_test)
# muc_seq = self.ner_corpus.read_sequence_list_conll(muc_test)
if decoding == "viterbi":
decoder = h.viterbi_decode_corpus
elif decoding == "max_emission":
decoder = h.max_emission_decode_corpus
elif decoding == "posterior":
decoder = h.posterior_decode_corpus
elif decoding == "posterior_cont":
decoder = h.posterior_cont_decode_corpus
elif decoding == "posterior_cont_type":
decoder = h.posterior_cont_type_decode_corpus
else:
print("Decoder not defined correctly, using Viterbi.")
decoder = h.viterbi_decode_corpus
#print("Decoding word representations on train.")
#decoder(train_seq)
print("Decoding word representations on dev.")
decoder(dev_seq, self.dataset)
#print("Decoding word representations on test.")
#decoder(test_seq)
#print("Decoding word representations on MUC.")
#decoder(muc_seq)
#return train_seq, dev_seq, test_seq, muc_seq
return dev_seq
def init_target_hypothesis(self):
target_tuple = self.simulation.target_tuple
target_rule_set = RuleSet.load_form_flat_list(target_tuple[1])
target_hypothesis = Hypothesis.create_hypothesis(
HMM(deepcopy(target_tuple[0])), target_rule_set)
target_energy = target_hypothesis.get_energy()
self.logger.info('Target hypothesis:')
log_hypothesis(target_hypothesis, self.logger.info)
self.logger.info('Target energy: {}'.format(target_energy))
self.logger.info('Target hypothesis energy signature: {}'.format(
target_hypothesis.get_recent_energy_signature()))
return target_hypothesis, target_energy
def generate_samples():
model = HMM.from_fixed_params()
T = 10
Z, X = model.sample(T)
print(model)
print(f"X: {X}, True Z: {Z}")
smcmc = SMCMC(model, T)
samples = smcmc.sample(N=100, x_sequence=X)
print(f"Last L samples: \n{samples[:, -1]}")
# print(np.unique(samples[:, 0], return_counts=True)[1] / len(samples))
print(f"Mean sample: {samples.mean((0, 1)).round(1)}")
def __init__(self,
string_input_words,
max_word_length_in_data,
initial_hmm=None,
alphabet_or_words="words"):
if not isinstance(string_input_words, list):
raise ValueError("should be list")
self.feature_table = get_feature_table()
self.max_word_length_in_data = max_word_length_in_data
if not initial_hmm:
feature_table = get_feature_table()
if alphabet_or_words == "alphabet":
alphabet = feature_table.get_alphabet()
self.hmm = HMM.create_hmm_alphabet(alphabet)
elif alphabet_or_words == "words":
self.hmm = HMM.create_hmm_from_list(string_input_words)
else:
self.hmm = initial_hmm
self.words = []
self._update_words()
def __init__(self):
# TODO 测试集上检查平滑处理的抉择问题
self.minfreq = -3.14e+100
# 构建字典树、用于扫描全切分有向图
self.trie = Trie()
self.construct_trie()
# 构建 二元词典
# self.construct_bigram_dic()
# 读取二元词典
with open('files/bigram_dic.json', 'r') as f:
self.bigram_dic = json.load(f)
# 进行特殊处理
self.SP = SpecialProcess()
# 创建HMM分词模型
self.hmm = HMM()
# 获取常用姓名中名字
self.get_second_names()
self.get_first_name()
def test_plural_english_grammar(self):
self.initialise_segment_table("plural_english_segment_table.txt")
rule_set = self.get_rule_set("plural_english_rule_set.json")
hmm = HMM({
INITIAL_STATE: ['q1'],
'q1': (['q2', FINAL_STATE], ['dog', 'kat']),
'q2': ([FINAL_STATE], ['z'])
})
grammar = Grammar(hmm, rule_set)
grammar_transducer = grammar.get_transducer()
def __load_dictionary(self, dir_name):
print('Loading dictionary from ' + dir_name + "...")
for dir in os.walk(dir_name).next()[1]:
self.dictionary[dir] = []
for file in os.walk(dir_name + '/' + dir).next()[2]:
if file.endswith('.wav'):
rate, data = wavfile.read(dir_name + '/' + dir + '/' + file)
if self.rate is not None and self.rate != rate:
print('Error: Dictionary sampling rate not constant.')
self.rate = rate
coefficients = self.__get_features(data)
self.dictionary[dir].append((coefficients, len(data)))
print('Done.')
print('Computing HMMs...')
for key, value in self.dictionary.items():
hmm = HMM(key)
model_size = self.__get_model_size([len(x[0]) for x in value])
hmm.train(model_size, [x[0] for x in value])
self.hmms.append(hmm)
print('Trained {0} with {1} states'.format(key, model_size))
print('Done.')
def main():
tagged_words = brown.tagged_words()
words_corpus = brown.words()
word2vec = Word2Vec()
word2vec.train(words_corpus)
word_vecs = [word2vec.word2vec(word) for word in words_corpus]
n_clusters = 10 # random number for now
kmeans = KMeans(n_clusters)
kmeans.compute(word_vecs)
# word-cluster HMM
p_word = {}
p_cluster = {}
p_cluster_given_word = None # softmax
p_word_given_cluster = None # joint probability formula
p_transition_cluster = None # count
p_initial_cluster = None # count
# cluster-tag HMM
p_cluster_given_tag = None # softmax
p_transition_tag = None # count from tagged data
p_initial_tag = None # count from tagged data
hmm_word_cluster = HMM(p_initial_cluster, p_transition_cluster, p_word_given_cluster)
hmm_cluster_tag = HMM(p_initial_tag, p_transition_tag, p_cluster_given_tag)
words = []
clusters = hmm_word_cluster.viterbi(words)
tags = hmm_cluster_tag.viterbi(clusters)
def main():
# load data
train_data = myutils.read_conll_file("data/da_ddt-ud-train.conllu")
dev_data = myutils.read_conll_file("data/da_ddt-ud-dev.conllu")
hmm = HMM()
# fit model to training data
hmm.fit(train_data)
# get most likely tag predictions
most_likely_predictions = hmm.predict(dev_data, method='most_likely')
viterbi_predictions = hmm.predict(dev_data, method='viterbi')
# evaluate
gold = [x[1] for x in dev_data]
sent_level, word_level = myutils.evaluate(gold, most_likely_predictions)
print('most likely scores:')
print('sent level: {:.4f}'.format(sent_level))
print('word level: {:.4f} \n'.format(word_level))
sent_level, word_level = myutils.evaluate(gold, viterbi_predictions)
print('viterbi scores:')
print('sent level: {:.4f}'.format(sent_level))
print('word level: {:.4f} \n'.format(word_level))
def test_crossover(self):
from copy import deepcopy
rule_1 = Rule.load([[{
'cont': '+'
}], [{
'coronal': '-'
}], [{
'coronal': '-'
}], [], True])
rule_2 = Rule.load([[{
'cons': '+',
'low': '-'
}], [{
'voice': '-'
}], [{
'voice': '-'
}], [], True])
crossover_rule_1 = deepcopy(rule_1)
crossover_rule_2 = deepcopy(rule_2)
crossover_rule_1.left_context_feature_bundle_list = rule_2.left_context_feature_bundle_list
crossover_rule_1.right_context_feature_bundle_list = rule_2.right_context_feature_bundle_list
crossover_rule_1.change_feature_bundle_list = rule_2.change_feature_bundle_list
crossover_rule_2.left_context_feature_bundle_list = rule_1.left_context_feature_bundle_list
crossover_rule_2.right_context_feature_bundle_list = rule_1.right_context_feature_bundle_list
crossover_rule_2.change_feature_bundle_list = rule_1.change_feature_bundle_list
rule_set_1 = RuleSet([crossover_rule_1])
rule_set_2 = RuleSet([crossover_rule_2])
print(rule_set_1)
print(rule_set_2)
hmm = HMM({
'q0': ['q1'],
'q1': (['q2', 'qf'], ['dag', 'kat', 'dot', 'kod']),
'q2': (['qf'], ['zo', 'go', 'do'])
})
grammar_1 = Grammar(hmm, rule_set_1)
grammar_2 = Grammar(hmm, rule_set_2)
data = ['kat', 'dot', 'dag', 'kod'] + \
['katso', 'dotso', 'dagzo', 'kodzo'] + \
['katko', 'dotko', 'daggo', 'kodgo'] + \
['katto', 'dotto', 'dagdo', 'koddo']
hypothesis_1 = Hypothesis(grammar_1, data)
hypothesis_2 = Hypothesis(grammar_2, data)
print(hypothesis_1.get_energy())
print(hypothesis_2.get_energy())
def model_training(train_data, tags):
"""
Train HMM based on training data
Inputs:
- train_data: (1*num_sentence) a list of sentences, each sentence is an object of line class
- tags: (1*num_tags) a list of POS tags
Returns:
- model: an object of HMM class initialized with parameters(pi, A, B, obs_dict, state_dict) you calculated based on train_data
"""
model = None
obs_dict = {}
state_dict = {}
curr_index = 0
for tag in tags:
state_dict[tag] = curr_index
curr_index += 1
curr_index = 0
for line in train_data:
for word in line.words:
if word not in obs_dict:
obs_dict[word] = curr_index
curr_index += 1
S = len(state_dict.keys())
L = len(obs_dict.keys())
pi = np.zeros([S])
A = np.zeros([S, S])
B = np.zeros([S, L])
for line in train_data:
pi[state_dict[line.tags[0]]] += 1
pi /= np.sum(pi)
for line in train_data:
for i in range(len(line.tags)-1):
A[state_dict[line.tags[i]], state_dict[line.tags[i+1]]] += 1
for i in range(S):
A[i, :] /= np.sum(A[i, :])
for line in train_data:
for i in range(len(line.words)):
B[state_dict[line.tags[i]], obs_dict[line.words[i]]] += 1
for i in range(S):
B[i, :] /= np.sum(B[i, :])
model = HMM(pi, A, B, obs_dict, state_dict)
return model
