def trainModel ( training_sequences, num_states, num_comp, covariance_type ):
"""
@return (GMMHMM): A GMM HMM model with @param num_states hidden states,
@param num_comp GMM components per hidden state, and a covariance
matrix of type @param covariance_type (full or diag), trained on
@param training_sequences.
"""
model = GMMHMM(num_states, n_mix=num_comp, covariance_type=covariance_type)
return model.fit(training_sequences)
def __init__(self, n_components=1, n_mix=1, startprob=None, transmat=None,
startprob_prior=None, transmat_prior=None,
algorithm="viterbi", gmms=None, covariance_type='diag',
covars_prior=1e-2, random_state=None, n_iter=10, thresh=1e-2,
params="",
init_params=""):
GMMHMM.__init__(self, n_components, n_mix, startprob, transmat,
startprob_prior, transmat_prior,
algorithm, gmms, covariance_type,
covars_prior, random_state, n_iter, thresh,
params, init_params)
def fit(self, reference, features):
"""Train HMM segmentation
The resulting HMM will contain one state per labels in training set.
Parameters
----------
reference : `Annotation` generator
Generates annotations whose labels will be HMM states
features : `Feature` generator
Generates features synchronized with `reference`
"""
# gather training data
reference = list(reference)
features = list(features)
# gather target list
self.targets = self._get_targets(reference)
# train each state
for target in self.targets:
logging.info('training {%s} GMM' % str(target))
self.gmm[target] = self._get_gmm(reference, features, target)
# train HMM
logging.info('training %d-states HMM' % len(self.targets))
self.hmm = GMMHMM(
n_components=len(self.targets),
gmms=[self.gmm[target] for target in self.targets],
init_params='st', params='st'
)
self.hmm.fit([f.data for f in features])
return self
def train(trainingData):
# train one GMM for each state
mixes = list()
for state in xrange(1,6):
# select data with current state label
d = trainingData[trainingData.rating==state]
# prepare data shape
d = np.array(zip(*[d[f].values for f in pcas]))
# init GMM
gmm = GMM(num_mixc,cov_type)
# train
gmm.fit(d)
mixes.append(gmm)
# train HMM with init, trans, GMMs=mixes
init,trans = hmm.hmmMlParams(trainingData,[1,2,3,4,5])
model = GMMHMM(n_components=5,init_params='',gmms=mixes)
model.transmat_ = trans
model.startprob_ = init
return model
def __init__(self,
n_components=1,
n_mix=1,
startprob=None,
transmat=None,
startprob_prior=None,
transmat_prior=None,
algorithm="viterbi",
gmms=None,
covariance_type='diag',
covars_prior=1e-2,
random_state=None,
n_iter=10,
thresh=1e-2,
params="",
init_params=""):
GMMHMM.__init__(self, n_components, n_mix, startprob, transmat,
startprob_prior, transmat_prior, algorithm, gmms,
covariance_type, covars_prior, random_state, n_iter,
thresh, params, init_params)
class NaiveHMM(object): def __init__(self, num_states, hmm_type=0, n_mix=3, n_iter=100, n_components=3): if hmm_type == 0: self._hmm = GaussianHMM(n_mix, n_iter=n_iter) elif hmm_type == 1: # WARNING: VERY VERY VERY SLOW self._hmm = GMMHMM(n_mix=n_mix, n_components=n_components, n_iter=n_iter) def fit(self, data): """ @param data DataFrame object containing T,X,Y,Z values. """ train = np.column_stack([data.X.tolist(), data.Y.tolist(), data.Z.tolist()]) self._hmm.fit([train]) def score(self, test): """ @param test DataFrame object containing T,X,Y,Z values. """ t = np.column_stack([test.X.tolist(), test.Y.tolist(), test.Z.tolist()]) return self._hmm.score(t) def get_hmm(self): return self._hmm
class SegmentationHMM(object):
"""HMM-based segmentation with Viterbi decoding
Parameters
----------
n_components : int
Number of gaussians per HMM state (default is 1).
covariance_type : {'diag', 'full'}
Type of gaussian covariance matrices
sampling : int, optional
Reduce the number of samples used for the initialization steps to
`sampling` samples per component. A few hundreds samples per component
should be a reasonable rule of thumb.
The final estimation steps always use the whole sample set.
min_duration : float, optional
Filter out segments shorter than `min_duration` seconds
n_jobs : int
Number of parallel jobs for GMM estimation
(default is one core)
"""
def __init__(
self, n_components=1, covariance_type='diag', sampling=0,
min_duration=None, n_jobs=1
):
super(SegmentationHMM, self).__init__()
self.n_components = n_components
self.covariance_type = covariance_type
self.sampling = sampling
self.n_jobs = n_jobs
self.min_duration = min_duration
self.gmm = {}
def _get_targets(self, reference):
"""Get list of targets from training data
Parameters
----------
reference : `Annotation` iterable
Returns
-------
targets : list
Sorted list of 'known' targets
"""
# empty target set
targets = set()
for annotation in reference:
labels = [
L for L in annotation.labels()
if not isinstance(L, Unknown)
]
targets.update(labels)
return sorted(targets)
def _get_gmm(self, reference, features, target):
# gather target data
data = np.vstack([
f.crop(r.label_coverage(target)) # use target regions only
for r, f in itertools.izip(reference, features)
])
lbg = LBG(
n_components=self.n_components,
covariance_type=self.covariance_type,
sampling=self.sampling,
n_iter=10,
disturb=0.05
)
gmm = lbg.apply(data)
return gmm
def fit(self, reference, features):
"""Train HMM segmentation
The resulting HMM will contain one state per labels in training set.
Parameters
----------
reference : `Annotation` generator
Generates annotations whose labels will be HMM states
features : `Feature` generator
Generates features synchronized with `reference`
"""
# gather training data
reference = list(reference)
features = list(features)
# gather target list
self.targets = self._get_targets(reference)
# train each state
for target in self.targets:
logging.info('training {%s} GMM' % str(target))
self.gmm[target] = self._get_gmm(reference, features, target)
# train HMM
logging.info('training %d-states HMM' % len(self.targets))
self.hmm = GMMHMM(
n_components=len(self.targets),
gmms=[self.gmm[target] for target in self.targets],
init_params='st', params='st'
)
self.hmm.fit([f.data for f in features])
return self
def apply(self, features):
"""
Parameters
----------
features : SlidingWindowFeatures
"""
# predict state sequences
sequence = self.hmm.predict(features.data)
# median filtering to get rid of short segments
if self.min_duration:
if len(self.targets) > 2:
raise NotImplementedError(
'min_duration is not supported with more than 2 states.'
)
dummy = Segment(0, self.min_duration)
_, n = features.sliding_window.segmentToRange(dummy)
sequence = median_filter(sequence, size=2*n+1)
# start initial segment
start = 0
label = self.targets[sequence[0]]
segmentation = Annotation()
for i, d in enumerate(np.diff(sequence)):
if d == 0:
continue
# end of current segment
end = i
segment = features.sliding_window.rangeToSegment(start, end-start)
segmentation[segment, '_'] = label
# start of a new segment
label = self.targets[sequence[i+1]]
start = end
segment = Segment(segment.end, features.getExtent().end)
segmentation[segment, '_'] = label
return segmentation
GMMs = []
for state_index in range(n_states):
g = GMM( n_components=m_components,
covariance_type=covariance_type,
#n_iter=0, # this initialises the GMMs without optimising them through EM; this is done later in the GMM HMM model
) # n_iter = 5(...) better??
g.fit(observations_per_state[state_index])
GMMs.append(g)
# Initialise and train new GMM HMM model
print "\tTraining GMM HMM model..."
# Implements (6.), (7.)
gmm_model = GMMHMM( n_components=n_states,
n_mix=m_components,
startprob=previous_model.startprob_,
transmat=previous_model.transmat_,
gmms=GMMs,
covariance_type=covariance_type,
n_iter=num_EM_iterations,
init_params='' # initialisation through previous model and GMMs!
)
gmm_model.fit(training_sequences)
# save base model
print "\tSaving model to file..."
saveModel(gmm_model, 'gmm_model_%scomponents' % m_components, observation_sequences[0].getFeatureNames())
# tag training data using new model
print "\tTagging training data using base model..."
likelihood_of_training_data, observations_per_state = tagTrainingData( gmm_model,
training_sequences,
list(observation_sequences), # pass a copy
save='gmm_model_%scomponents/tagged_training_data' % m_components,
filenames=filenames )
import csv
from hmmlearn import hmm
import os
import glob
import pandas as pd
from sklearn.hmm import GMMHMM
from sklearn.externals import joblib
def apply_pca(dataset):
pcaModel = joblib.load('./pca2.pkl')
return pcaModel.transform(dataset)
gesture = 'Sorry'
os.chdir('./'+gesture)
file_list = [i for i in glob.glob('*.csv')]
os.chdir('../')
print file_list
lengths = []
X = []
for file_name in file_list:
print file_name
dataset = pd.read_csv(gesture+'/'+file_name)
dataset = apply_pca(dataset.values)
lengths.append(len(dataset))
X.append(dataset)
# print X
model = GMMHMM(n_components=7,n_mix=128,covariance_type="diag")
model.fit(X)
joblib.dump(model,'./HMMs/'+gesture+'.pkl')
def __init__(self, num_states, hmm_type=0, n_mix=3, n_iter=100, n_components=3): if hmm_type == 0: self._hmm = GaussianHMM(n_mix, n_iter=n_iter) elif hmm_type == 1: # WARNING: VERY VERY VERY SLOW self._hmm = GMMHMM(n_mix=n_mix, n_components=n_components, n_iter=n_iter)
