def discrete_geweke_test(fig):
Nstates = 2
Nemissions = 2
alpha = 3.
init_state_concentration=3.
T = 10
num_iter = 10000
num_checks = 10
obs_distns = [distributions.Categorical(K=Nemissions,alpha_0=1.)
for _ in range(Nstates)]
hmm = models.HMM(
alpha=alpha,init_state_concentration=init_state_concentration,
obs_distns=obs_distns)
# generate state sequences and parameters from the prior
prior_stateseqs = []
prior_weights = []
for itr in xrange(num_iter):
hmm.resample_model() # sample parameters from the prior
_, stateseq = hmm.generate(T,keep=False)
prior_stateseqs.append(stateseq)
prior_weights.append(hmm.obs_distns[0].weights)
prior_stateseqs = np.array(prior_stateseqs)
prior_weights = np.array(prior_weights)
# generate state sequences and parameters using Gibbs
hmm.generate(T,keep=True)
s = hmm.states_list[0]
gibbs_stateseqs = []
gibbs_weights = []
for itr in xrange(num_iter):
s.generate_obs() # resamples data given state sequence, obs params
hmm.resample_model() # resamples everything else as usual
gibbs_stateseqs.append(s.stateseq)
gibbs_weights.append(hmm.obs_distns[0].weights)
gibbs_stateseqs = np.array(gibbs_stateseqs)
gibbs_weights = np.array(gibbs_weights)
# test that they look similar by checking probability of co-assignment
time_indices = np.arange(T)
for itr in xrange(num_checks):
i,j = np.random.choice(time_indices,replace=False,size=2)
prior_prob_of_coassignment = (prior_stateseqs[:,i] == prior_stateseqs[:,j]).std()
gibbs_prob_of_coassignment = (gibbs_stateseqs[:,i] == gibbs_stateseqs[:,j]).std()
assert np.isclose(
prior_prob_of_coassignment,gibbs_prob_of_coassignment,
rtol=0.025,atol=0.025,
)
# test that they look similar by checking parameters
testing.populations_eq_quantile_plot(prior_weights,gibbs_weights,fig=fig)
figpath = os.path.join(figure_dir_path,'discrete_geweke_test_weights.pdf')
plt.savefig(figpath)
def _likelihood_helper(obs_distns, trans_matrix, init_distn, data, target_val):
hmm = m.HMM(
alpha=6,
gamma=6,
init_state_concentration=1, # placeholders
obs_distns=obs_distns)
hmm.trans_distn.A = trans_matrix
hmm.init_state_distn.weights = init_distn
hmm.add_data(data)
assert np.isclose(hmm.log_likelihood(), target_val)
### generate data
num_modes = 3
true_obs_distns = [
distributions.Gaussian(**obs_hypparams) for i in range(num_modes)
]
data = np.concatenate(
[true_obs_distns[i % num_modes].rvs(25) for i in range(25)])
## inference!
hmm = models.HMM(obs_distns=[
distributions.Gaussian(**obs_hypparams) for i in range(num_modes * 3)
],
alpha=3.,
init_state_concentration=1.)
hmm.add_data(data)
hmm.meanfield_coordinate_descent_step()
scores = [hmm.meanfield_coordinate_descent_step() for i in range(50)]
scores = np.array(scores)
hmm.plot()
plt.figure()
plt.plot(scores)
def normalize(A):
return A / A.sum(1)[:, None]