def train(self, nstates, nmix, niter, covtype, data, labels):
train_pos = self.tensor_to_list(data[labels == 1, :, :])
train_neg = self.tensor_to_list(data[labels == 0, :, :])
print "Start training the models..."
success = False
while not success:
try:
print "Training with covariance type %s" % covtype
model_pos = hmm.GMMHMM(nstates,
nmix,
covariance_type=covtype,
n_iter=niter)
model_neg = hmm.GMMHMM(nstates,
nmix,
covariance_type=covtype,
n_iter=niter)
model_pos.fit(train_pos)
model_neg.fit(train_neg)
success = True
except:
if covtype == 'full':
covtype = 'diag'
elif covtype == 'diag':
covtype = 'tied'
elif covtype == 'tied':
covtype = 'spherical'
else:
print 'Error: HMM sucks'
success = True
return model_pos, model_neg
def __init__(self,
model_name='GaussianHMM',
n_components=10,
cov_type='diag',
n_iter=100):
#模型名称 hmmlearn实现了三种HMM模型类，GaussianHMM和GMMHMM是连续观测状态的HMM模型，MultinomialHMM是离散观测状态的模型
self.model = None
self.model_name = model_name
#隐藏状态个数
self.n_components = n_components
#转移矩阵协方差类型
self.cov_type = cov_type
#训练迭代次数
self.n_iter = n_iter
self.models = []
# self.states = ["als", "control","hunt", "park"]
# self.n_states = len(self.states)
self.observations = []
self.n_observations = len(self.observations)
if self.model_name == 'GaussianHMM':
self.model = hmm.GaussianHMM(n_components=self.n_components,
covariance_type=self.cov_type,
n_iter=self.n_iter)
else:
# self.model = hmm.MultinomialHMM(n_components=self.n_components, n_iter=self.n_iter, tol=0.01)
self.model = hmm.GMMHMM(n_components=self.n_components,
n_iter=self.n_iter,
tol=0.01)
def __init__(self,
model_name='GaussianHMM',
n_components=4,
cov_type='diag',
n_iter=1000,
n_mix=2):
self.model_name = model_name
self.n_components = n_components
self.cov_type = cov_type
self.n_iter = n_iter
self.n_mix = n_mix
self.models = []
# This stuff is depricated. Disabling the warnings for now.
warnings.filterwarnings("ignore", category=DeprecationWarning)
if self.model_name == 'GaussianHMM':
self.model = hmm.GaussianHMM(n_components=self.n_components,
covariance_type=self.cov_type,
n_iter=self.n_iter)
elif self.model_name == 'GMMHMM':
self.model = hmm.GMMHMM(n_components=self.n_components,
covariance_type=self.cov_type,
n_iter=self.n_iter,
n_mix=self.n_mix)
else:
raise TypeError('Invalid model type')
def train_lr(seqs):
# select features
seqs = [select_feature(s) for s in seqs]
# convert to hmmlearn formats
feats = np.concatenate(seqs)
lengths = [s.shape[0] for s in seqs]
nc = 9
nm = 3
t = np.diag(np.ones(nc)) * .9
t[:-1, 1:] += np.diag(np.ones(nc - 1)) * .1
t[-1, -1] = 1
s = np.zeros(nc)
s[0] = 1.0
# model = hmm.GaussianHMM(n_components=2, covariance_type='diag').fit(feats, lengths)
model = hmm.GMMHMM(n_components=nc,
n_mix=nm,
covariance_type='diag',
min_covar=1e-6,
init_params="mcw",
params="tmcw",
verbose=False)
model.startprob_ = s
model.transmat_ = t
model.fit(feats, lengths)
return model
def main():
states = task_name
n_task = len(states)
# don't know the task prior here
start_probability = np.array(([1.0 / n_task] * n_task))
# don't know the task transition probability here
transition_probability = np.array([1.0 / n_task] * n_task *
n_task).reshape([n_task, n_task])
gmm_model_full = [GaussianMixture(n_components=3) for i in range(n_task)]
task_data_full = []
for task_idx, task_data in enumerate(datasets_raw):
demo_data_full = np.array([]).reshape(0, num_joints)
for demo_data in task_data:
h = np.hstack([demo_data['left_hand'], demo_data['left_joints']])
demo_data_full = np.vstack([demo_data_full, h])
task_data_full.append(demo_data_full)
for task_data, gmm_model in zip(task_data_full, gmm_model_full):
gmm_model.fit(task_data)
# hmm_model = hmm.GMMHMM()
# hmm_model.n_components = n_task
# hmm_model.startprob_ = start_probability
# hmm_model.transmat_ = transition_probability
# hmm_model.means_ = gmm_model_full[0].means_
# testing
hmm_model1 = hmm.GMMHMM()
hmm_model1.n_components = 10
hmm_model1.fit(task_data_full[0])
print 1
def test_score_samples_and_decode(self):
h = hmm.GMMHMM(self.n_components)
h.startprob_ = self.startprob
h.transmat_ = self.transmat
h.gmms_ = self.gmms
# Make sure the means are far apart so posteriors.argmax()
# picks the actual component used to generate the observations.
for g in h.gmms_:
g.means_ *= 20
refstateseq = np.repeat(np.arange(self.n_components), 5)
n_samples = len(refstateseq)
X = [
h.gmms_[x].sample(1, random_state=self.prng).flatten()
for x in refstateseq
]
_ll, posteriors = h.score_samples(X)
self.assertEqual(posteriors.shape, (n_samples, self.n_components))
assert_array_almost_equal(posteriors.sum(axis=1), np.ones(n_samples))
_logprob, stateseq = h.decode(X)
assert_array_equal(stateseq, refstateseq)
def train_GMMHMM(dataset):
GMMHMM_Models = {}
states_num = 5
GMM_mix_num = 3
tmp_p = 1.0 / (states_num - 2)
transmatPrior = np.array([[tmp_p, tmp_p, tmp_p, 0 ,0], \
[0, tmp_p, tmp_p, tmp_p , 0], \
[0, 0, tmp_p, tmp_p,tmp_p], \
[0, 0, 0, 0.5, 0.5], \
[0, 0, 0, 0, 1]],dtype=np.float)
startprobPrior = np.array([0.5, 0.5, 0, 0, 0], dtype=np.float)
for label in dataset.keys():
model = hmm.GMMHMM(n_components=states_num, n_mix=GMM_mix_num, \
transmat_prior=transmatPrior, startprob_prior=startprobPrior, \
covariance_type='diag', n_iter=10)
trainData = dataset[label]
length = np.zeros([
len(trainData),
], dtype=np.int)
for m in range(len(trainData)):
length[m] = trainData[m].shape[0]
trainData = np.vstack(trainData)
model.fit(trainData, lengths=length) # get optimal parameters
GMMHMM_Models[label] = model
return GMMHMM_Models
def get_hmm_model(self):
""" get hmm model from training data """
# Gaussian Mixture HMM
model = hmm.GMMHMM(n_components=self.n_hidden_states,
n_mix=self.n_mixtures,
covariance_type=self.covariance_type,
n_iter=self.n_iter)
train = self.get_training_data()
"""
def fit(self, X, lengths=None):
Parameters
----------
X : array-like, shape (n_samples, n_features)
Feature matrix of individual samples.
lengths : array-like of integers, shape (n_sequences, )
Lengths of the individual sequences in ``X``. The sum of
these should be ``n_samples``.
"""
lengths = list()
data_combined = np.concatenate(
(train[0], train[1], train[2], train[3], train[4], train[5],
train[6], train[7], train[8], train[9]))
for example in train:
lengths.append(example.shape[0])
model.fit(data_combined, lengths)
self.wordhmm = model
def test_fit(self, params='stmwc', n_iter=5, verbose=False, **kwargs):
h = hmm.GMMHMM(self.n_components, covars_prior=1.0)
h.startprob_ = self.startprob
h.transmat_ = normalize(
self.transmat + np.diag(self.prng.rand(self.n_components)), 1)
h.gmms_ = self.gmms
lengths = [10] * 10
X, _state_sequence = h.sample(sum(lengths), random_state=self.prng)
# Mess up the parameters and see if we can re-learn them.
h.n_iter = 0
h.fit(X, lengths=lengths)
h.transmat_ = normalize(self.prng.rand(self.n_components,
self.n_components),
axis=1)
h.startprob_ = normalize(self.prng.rand(self.n_components))
trainll = fit_hmm_and_monitor_log_likelihood(h,
X,
lengths=lengths,
n_iter=n_iter)
if not np.all(np.diff(trainll) > 0) and verbose:
print('Test train: (%s)\n %s\n %s' %
(params, trainll, np.diff(trainll)))
# XXX: this test appears to check that training log likelihood should
# never be decreasing (up to a tolerance of 0.5, why?) but this is not
# the case when the seed changes.
raise SkipTest("Unstable test: trainll is not always increasing "
"depending on seed")
self.assertTrue(np.all(np.diff(trainll) > -0.5))
def test_fit(self, params='stmwc', n_iter=5, verbose=False, **kwargs):
h = hmm.GMMHMM(self.n_components, covars_prior=1.0)
h.startprob_ = self.startprob
h.transmat_ = normalize(
self.transmat + np.diag(self.prng.rand(self.n_components)), 1)
h.gmms_ = self.gmms_
# Create training data by sampling from the HMM.
train_obs = [h.sample(n=10, random_state=self.prng)[0]
for x in range(10)]
# Mess up the parameters and see if we can re-learn them.
h.n_iter = 0
h.fit(train_obs)
h.transmat_ = normalize(self.prng.rand(self.n_components,
self.n_components), axis=1)
h.startprob_ = normalize(self.prng.rand(self.n_components))
trainll = train_hmm_and_keep_track_of_log_likelihood(
h, train_obs, n_iter=n_iter, params=params)[1:]
if not np.all(np.diff(trainll) > 0) and verbose:
print('Test train: (%s)\n %s\n %s' % (params, trainll,
np.diff(trainll)))
# XXX: this test appears to check that training log likelihood should
# never be decreasing (up to a tolerance of 0.5, why?) but this is not
# the case when the seed changes.
raise SkipTest("Unstable test: trainll is not always increasing "
"depending on seed")
self.assertTrue(np.all(np.diff(trainll) > -0.5))
def __init__(self, Class, label):
self.traindata = np.zeros((0, 6))
self.Class = Class
self.label = label
self.model = hmm.GMMHMM(n_components = m_num_of_HMMStates, n_mix = m_num_of_mixtures, \
transmat_prior = m_transmatPrior, startprob_prior = m_startprobPrior, \
covariance_type = m_covarianceType, n_iter = m_n_iter)
def get_models(self):
for i in range(self.label_num):
model = hmm.GMMHMM(n_components=self.n_components,
n_mix=self.n_mix,
covariance_type=self.covariance_type,
n_iter=self.n_iter)
self.models.append(model)
def main():
k = 5
list_x, list_y = data_split(xtrain, ytrain, k)
acc = []
for i in range(k):
testx = list_x[i]
testy = list_y[i]
trainx = list_x[:i] + list_x[i + 1:]
trainy = list_y[:i] + list_y[i + 1:]
trainx_fix = cv_fix(trainx, trainy)
model_list = [
hmm.GMMHMM(n_components=10,
n_mix=10,
covariance_type='diag',
algorithm='viterbi',
n_iter=30,
params='stmc',
init_params='stmc') for _ in range(20)
]
for j in range(20):
X = transpose(trainx_fix[j])
model_list[j].fit(X)
predic = predict(model_list, testx)
accurac = accuracy(testy, predic)
acc.append(accurac)
print(accurac)
return acc
def train(self,
train_list,
length_list,
n_components,
n_mixs,
name_list=None):
"""
:param train_list: list, 训练数据列表
:param length_list: 序列长度列表
:param name_list: 各hmm模型名称， 默认为None
:return: hmms： list, hmm模型list
comment: 模型个数由train_list的个数类别决定
"""
self.model_name = name_list
print('开始训练hmm模型！')
for i in range(len(train_list)):
if name_list is not None:
print('\t训练%s模型...' % name_list[i])
model = hmm.GMMHMM(n_components=n_components[i],
n_mix=n_mixs[i],
covariance_type="diag")
model.n_iter = self.n_iter
model.fit(train_list[i], length_list[i])
# # 重置初始概率
# model.startprob_ = self.startprob
self.hmms.append(model)
print('所有模型训练完成！')
def test_sample(self, n=1000):
h = hmm.GMMHMM(self.n_components,
self.covariance_type,
startprob=self.startprob,
transmat=self.transmat,
gmms=self.gmms_)
samples = h.sample(n)[0]
self.assertEqual(samples.shape, (n, self.n_features))
def train(self, train_data):
self.train_obs = self.data_prep(train_data)
self.model = hmm.GMMHMM(n_components=self.n_components,
n_mix=self.n_mix,
algorithm=self.algorithm,
n_iter=self.n_iter)
self.model.fit(self.train_obs)
def test_sample(self, n_samples=1000):
h = hmm.GMMHMM(self.n_components, covariance_type=self.covariance_type)
h.startprob_ = self.startprob
h.transmat_ = self.transmat
h.gmms_ = self.gmms
X, state_sequence = h.sample(n_samples)
assert X.shape == (n_samples, self.n_features)
assert len(state_sequence) == n_samples
def test_fit_works_on_sequences_of_different_length(self):
lengths = [3, 4, 5]
X = self.prng.rand(sum(lengths), self.n_features)
h = hmm.GMMHMM(self.n_components, covariance_type=self.covariance_type)
# This shouldn't raise
# ValueError: setting an array element with a sequence.
h.fit(X, lengths=lengths)
def test_sample(self, n=1000):
h = hmm.GMMHMM(self.n_components, covariance_type=self.covariance_type)
h.startprob_ = self.startprob
h.transmat_ = self.transmat
h.gmms_ = self.gmms
X, state_sequence = h.sample(n, random_state=self.prng)
self.assertEqual(X.shape, (n, self.n_features))
self.assertEqual(len(state_sequence), n)
def fit(self):
# data_loader: DataLoader Class
start, trans = initByBakis(self.n_com, 2)
for word in self.data_loader.spoken:
self.classes.append(word)
self.models[word] = hmm.GMMHMM(n_components=self.n_com, n_mix = self.n_mix, covariance_type="diag", n_iter = 500)
self.models[word].transmat_ = trans
self.models[word].startprob_ = start
self.models[word].fit(self.data_loader.train_features[word], self.data_loader.train_lengths[word])
def train_HMM(dataset):
Models = {}
for label in dataset.keys():
model = hmm.GMMHMM(n_components=10)
trainData = dataset[label]
trData = np.vstack(trainData)
model.fit(trData)
Models[label] = model
return Models
def test_fit_works_on_sequences_of_different_length(self):
obs = [self.prng.rand(3, self.n_features),
self.prng.rand(4, self.n_features),
self.prng.rand(5, self.n_features)]
h = hmm.GMMHMM(self.n_components, covariance_type=self.covariance_type)
# This shouldn't raise
# ValueError: setting an array element with a sequence.
h.fit(obs)
def train_model(data):
learned_hmm = dict()
for label in data.keys():
model = hmm.GMMHMM(n_components=4, covariance_type="full")
feature = np.ndarray(shape=(1, 13))
for list_feature in data[label]:
feature = np.vstack((feature, list_feature))
obj = model.fit(feature)
learned_hmm[label] = obj
return learned_hmm
def gmm_hmm(pca_output, n_comps, cov_type):
ghmm = hmm.GMMHMM(n_components=n_comps,
covariance_type=cov_type,
n_iter=100)
scores = pca_output['pc_scores']
ghmm.fit(scores)
pred_labels = ghmm.predict(scores)
# Project hidden state mean vectors to original dimensions
mean_maps = np.squeeze(ghmm.means_) @ pca_output['Va']
return ghmm, pred_labels, mean_maps
def train_model(data):
learned_hmm = dict()
for label in data.keys():
model = hmm.GMMHMM(verbose=False, n_components=100, n_iter=10000)
feature = np.ndarray(shape=(1, 13))
for list_feature in data[label]:
feature = np.vstack((feature, list_feature))
obj = model.fit(feature)
learned_hmm[label] = obj
return learned_hmm
def train_samad_hmm(train_data_path, states_num):
GMMHMM_Models = {}
train_data = create_train_set(train_data_path)
for label in train_data.keys():
model = hmm.GMMHMM(n_components=states_num)
t_Data = train_data[label]
t_Data = np.vstack(t_Data)
model = model.fit(t_Data) # get optimal parameters
GMMHMM_Models[label] = model
return GMMHMM_Models
def fit(self,
data,
num_gaussians,
num_states,
n_iterations=10,
trans_mat=None,
start_prob=None):
#@TODO play with covariance types and compare results
frame_0 = data[0][0]
num_features = frame_0.shape[1]
num_classes = len(data)
self.mapping_observation_state = np.empty(
(0, num_features +
1)) # add a 1 to add a row for getting corresponding states
# iterate through each class and train a HMM
for class_index, frames in enumerate(data):
# train this model
X = np.concatenate(frames)
lengths = np.zeros(len(frames), dtype=np.dtype(np.int32))
for i, arr in enumerate(frames):
# each arr is a numpy array and each row in this array represents an observation
lengths[i] = arr.shape[0]
scores = np.zeros(n_iterations)
HMM_list = []
for j in np.arange(n_iterations):
lr = hmm.GMMHMM(n_components=num_states,
n_mix=num_gaussians[class_index],
init_params="cm",
params="cmt")
# lr.startprob_ = np.array([1.0, 0.0, 0.0])
# lr.transmat_ = np.array([[0.5, 0.5, 0.0],[0.0, 0.5, 0.5],[0.0, 0.0, 1.0]])
lr.transmat_ = trans_mat
lr.startprob_ = start_prob
lr.fit(X=X, lengths=lengths)
# temp_HMM = GMMHMM(n_components=num_states, n_mix = num_gaussians[class_index])
# temp_HMM.startprob_ = [1.0,0.0,0.0]
# temp_HMM.transmat_ = trans_mat
# temp_HMM.fit(X=X,lengths=lengths)
scores[j] = lr.score(X, lengths=lengths)
HMM_list.append(lr)
# finding the HMM with the best fit
cur_HMM = HMM_list[np.argmax(scores)]
self.HMMList.append(cur_HMM)
temp = cur_HMM.decode(X, lengths)
states = temp[1]
# we do this Hack to make sure that the state numbers are different for different classes
states += num_states * class_index
a = np.concatenate((X, states.reshape(-1, 1)), axis=1)
self.mapping_observation_state = np.concatenate(
(self.mapping_observation_state, a))
def getHmmModel(self):
''' get hmm model from training data '''
# GaussianHMM
# model = hmm.GaussianHMM(numStates, "diag") # initialize hmm model
# Gaussian Mixture HMM
model = hmm.GMMHMM(n_components = self.nComp, n_mix = self.nMix, \
transmat_prior = self.transmatPrior, startprob_prior = self.startprobPrior, \
covariance_type = self.covarianceType, n_iter = self.n_iter)
model.fit(self.trainData) # get optimal parameters
self.hmmModel = model
def create_model(n_components, n_mix):
model = hmm.GMMHMM(n_components=n_components,
n_mix=n_mix,
init_params='mcw')
start_probability = np.ones(n_components)
start_probability = start_probability / n_components
transition_probability = np.ones((n_components, n_components))
transition_probability = transition_probability / n_components
model.startprob_ = start_probability
model.transmat_ = transition_probability
return model
def load_one_model(self, filename):
hmm_dict = np.load(filename)
hmm_dict = hmm_dict.item()
model = hmm.GMMHMM(n_components=hmm_dict['n_components'],
n_mix=hmm_dict['n_mix'],
covariance_type=hmm_dict['covariance_type'])
model.startprob_ = hmm_dict['startprob']
model.transmat_ = hmm_dict['transmat']
model.means_ = hmm_dict['means']
model.covars_ = hmm_dict['covars']
model.weights_ = hmm_dict['weights']
self.hmms.append(model)
self.model_name.append(hmm_dict['model_name'])
