def test_exact_fhmm(self):
n_chains = 3
n_states = 4
n_features = 2
obs = [np.array([[-600, 100], [-300, 200], [0, 300]])]
means = np.array([[162.2636, 154.0879],
[-299.5184, 0.1605],
[-313.1509, -4.3836],
[ -15.0993, 94.9669],
[-102.5866, 65.8045],
[ -88.9765, 70.3412],
[-135.6746, 54.7751],
[-138.2674, 53.9109],
[-147.2465, 50.9178],
[-135.9769, 54.6744],
[-33.4946, 88.8351],
[-148.7870, 50.4043]])
means = means.reshape((n_chains, n_states, n_features))
covar = 1.0e+03 * np.array([[9.1968, 3.0656],
[3.0656, 1.0219]])
transmat = np.array([[0.3660, 0.0045, 0.0072, 0.6223],
[0.2442, 0.0546, 0.0053, 0.6959],
[0.0037, 0.0887, 0.0503, 0.8573],
[0.9465, 0.0006, 0.0004, 0.0525],
[0.1922, 0.5774, 0.0381, 0.1923],
[0.3122, 0.4747, 0.0784, 0.1347],
[0.0432, 0.6053, 0.2195, 0.1320],
[0.3434, 0.2505, 0.0443, 0.3618],
[0.0864, 0.3736, 0.5159, 0.0241],
[0.2319, 0.2389, 0.5001, 0.0291],
[0.0741, 0.1285, 0.6239, 0.1735],
[0.3229, 0.2213, 0.2160, 0.2398]])
transmat = transmat.reshape((n_chains, n_states, n_states))
self.assertTrue(np.allclose(np.sum(transmat, axis=2), 1))
startprob = np.array([[0.0012, 0.1196, 0.3150],
[0.5932, 0.1659, 0.3572],
[0.4049, 0.1440, 0.0027],
[0.0007, 0.5705, 0.3251]])
startprob = startprob.T.reshape((n_chains, n_states))
self.assertTrue(np.allclose(np.sum(startprob, axis=1), 1.0))
fhmm = ExactGaussianFHMM(n_states=n_states, n_chains=n_chains, topology='full', n_training_iterations=1)
fhmm.fit([np.random.random((20, 2))])
fhmm.log_startprob = np.log(startprob)
fhmm.log_transmat = np.log(transmat)
fhmm.means = means
fhmm.covar = covar
actual_loglikelihood = fhmm.loglikelihood(obs[0])
# The loglikelihood calculated by the reference impl is -15.0552, which is probably due to to the fact that we
# calculate in logs and the ref implementation uses scaling
loglikelihood = -11.425412024522673
self.assertAlmostEqual(actual_loglikelihood, loglikelihood)
def test_exact_fhmm_train(self):
path = os.path.join(TESTDATA_PATH, 'X.mat')
self.assertTrue(os.path.exists(path))
X = loadmat(path)['X']
obs = [X]
fhmm_exact = ExactGaussianFHMM(n_chains=2, n_states=3, n_training_iterations=20, training_threshold=0.01, topology='full')
fhmm_exact.fit(obs)
print 'exact:', fhmm_exact.loglikelihood(obs[0])
fhmm_seq = SequentialGaussianFHMM(n_chains=2, n_states=3, n_training_iterations=20, training_threshold=0.01, topology='full')
fhmm_seq.fit(obs)
print 'seq:', fhmm_seq.loglikelihood(obs[0])
def get_classifier(args):
if args.model == 'hmmlearn':
model = HMMLearnModel()
elif args.model == 'fhmm-exact':
model = ExactGaussianFHMM(n_chains=args.n_chains)
elif args.model == 'fhmm-seq':
model = SequentialGaussianFHMM(n_chains=args.n_chains)
elif args.model == 'pomegranate':
model = PomegranateModel()
else:
model = None
assert model is not None
model.n_training_iterations = args.n_iterations
model.n_states = args.n_states
model.topology = args.topology
return HMMClassifier(model, n_jobs=args.n_jobs)
def get_classifier(args):
if args.model == 'hmm':
model = GaussianHMM()
elif args.model == 'fhmm-exact':
model = ExactGaussianFHMM(n_chains=args.n_chains)
elif args.model == 'fhmm-seq':
model = SequentialGaussianFHMM(n_chains=args.n_chains)
else:
model = None
assert model is not None
model.n_training_iterations = args.n_training_iter
model.n_states = args.n_states
model.topology = args.topology
model.verbose = args.verbose
model.transition_init = args.transition_init
model.emission_init = args.emission_init
model.covar_type = args.covar_type
return HMMClassifier(model, n_jobs=args.n_jobs)
def test_foo(self):
obs = [np.array([[-600, 100], [-300, 200], [0, 300]]) for _ in xrange(10)]
fhmm = ExactGaussianFHMM(n_chains=2, n_states=3)
fhmm.fit(obs)
print fhmm.loglikelihood(obs[0])
