def setup_hmm(rg, EM_params):
G, node_map = hmmconf.rg_to_nx_undirected(rg, map_nodes=True)
n_states = len(node_map)
# deal with initial marking and get start probability
is_inv = lambda t: t.name is None
startprob = hmmconf.compute_startprob(rg, node_map, n_states, is_inv)
# add epsilon mass to all states
startprob += 1e-5
hmmconf.utils.normalize(startprob, axis=1)
# remove invisible transitions
to_remove = list()
for t in rg.transitions:
if is_inv(t):
to_remove.append(t)
for t in to_remove:
rg.transitions.remove(t)
t.from_state.outgoing.remove(t)
t.to_state.incoming.remove(t)
dist_df = hmmconf.compute_distance_matrix(G, node_map, as_dataframe=True)
distmat = dist_df.values
# print('Distance df: \n{}'.format(dist_df))
obsmap = {t.name: int(t.name) for t in rg.transitions}
int2state = {val:key for key, val in node_map.items()}
int2obs = {val:key for key, val in obsmap.items()}
n_obs = len(obsmap)
logger.info('No. of states: {}'.format(n_states))
transcube = hmmconf.compute_state_trans_cube(rg, node_map, obsmap, n_obs, n_states)
emitmat = hmmconf.compute_emission_mat(rg, node_map, obsmap, n_obs, n_states)
confmat = hmmconf.compute_conformance_mat(emitmat)
conform_f = hmmconf.conform
hmm = hmmconf.HMMConf(conform_f, startprob, transcube, emitmat, confmat, distmat,
int2state, int2obs, n_states, n_obs, params='to', verbose=True,
n_jobs=EM_params[N_JOBS], tol=EM_params[TOL],
n_iter=EM_params[N_ITER], random_seed=EM_params[RANDOM_SEED_PARAM])
return hmm
time_dict[TIME_CONFORM] = took_conform
time_dict[TIME_N_CASES] = filtered_event_df["caseid"].unique().shape[0]
time_dict[TIME_N_EVENTS] = filtered_event_df.shape[0]
logger.info(
f"Estimating conforming distribution params took: {took_conform:.3f}s")
transcube, emitmat, conforming_caseids = params
hmmconf_params = {
'params': EXPERIMENT_CONFIGS[EM_PARAMS],
'conf_tol': EXPERIMENT_CONFIGS[CONF_TOL],
'n_iter': EXPERIMENT_CONFIGS[N_ITER],
'tol': EXPERIMENT_CONFIGS[TOL],
'verbose': True,
'n_procs': EXPERIMENT_CONFIGS[N_JOBS],
'random_seed': EXPERIMENT_CONFIGS[RANDOM_SEED_PARAM]
}
hmm = hmmconf.HMMConf(startprob, transcube, emitmat, confmat, int2state,
int2obs, **hmmconf_params)
int2state_list = list(int2state.items())
stateid_list, state_list = zip(*int2state_list)
columns = ['state_id', 'state']
state_id_df = pd.DataFrame({'state_id': stateid_list, 'state': state_list})
logger.info(f"State id df: \n{state_id_df}")
n_caseids = filtered_event_df["caseid"].unique().shape[0]
filter_by_conforming_caseids = filtered_event_df["caseid"].isin(
conforming_caseids)
filtered_train_event_df = filtered_event_df.loc[
~filter_by_conforming_caseids, :]
n_caseids_train = filtered_train_event_df["caseid"].unique().shape[0]
logger.info(
f"Filtered train event df shape: {filtered_train_event_df.shape}")