def test_kalman_pickle():
kf = KalmanFilter(
data.transition_matrix,
data.observation_matrix,
data.transition_covariance,
data.observation_covariance,
data.transition_offsets,
data.observation_offset,
data.initial_state_mean,
data.initial_state_covariance,
em_vars='all')
# train and get log likelihood
X = data.observations[0:10]
kf = kf.em(X, n_iter=5)
loglikelihood = kf.loglikelihood(X)
# pickle Kalman Filter
store = StringIO()
pickle.dump(kf, store)
clf = pickle.load(StringIO(store.getvalue()))
# check that parameters came out already
np.testing.assert_almost_equal(loglikelihood, kf.loglikelihood(X))
# store it as BytesIO as well
store = BytesIO()
pickle.dump(kf, store)
kf = pickle.load(BytesIO(store.getvalue()))
def test_kalman_fit():
# check against MATLAB dataset
kf = KalmanFilter(
data.transition_matrix,
data.observation_matrix,
data.initial_transition_covariance,
data.initial_observation_covariance,
data.transition_offsets,
data.observation_offset,
data.initial_state_mean,
data.initial_state_covariance,
em_vars=['transition_covariance', 'observation_covariance'])
loglikelihoods = np.zeros(5)
for i in range(len(loglikelihoods)):
loglikelihoods[i] = kf.loglikelihood(data.observations)
kf.em(X=data.observations, n_iter=1)
assert_true(np.allclose(loglikelihoods, data.loglikelihoods[:5]))
# check that EM for all parameters is working
kf.em_vars = 'all'
n_timesteps = 30
for i in range(len(loglikelihoods)):
kf.em(X=data.observations[0:n_timesteps], n_iter=1)
loglikelihoods[i] = kf.loglikelihood(data.observations[0:n_timesteps])
for i in range(len(loglikelihoods) - 1):
assert_true(loglikelihoods[i] < loglikelihoods[i + 1])
