def __init__(self,
N,
M,
params=None,
writeout=False,
brown_init_path=None,
x_dict=None,
approx=False,
dirname=None):
super().__init__(N,
M,
params=params,
writeout=writeout,
brown_init_path=brown_init_path,
x_dict=x_dict,
dirname=dirname)
# sum-product instead of forward-backward
self.inference = SumProduct(approximate=approx)
def __init__(self, N, M, R=None, params=None, writeout=False, brown_init_path=None, x_dict=None, approx=False,
dirname=None, omit_class_cond=False, omit_emis_cond=False):
"""
:param N: number of states
:param M: number of observation symbols
:param R: number of dep relations (relation-specific HMTM only)
:param params: numpy objects
-initial_probs
-transition_probs
-final_probs
-emission_probs)
:param writeout: save hmm details to a file
:param omit_class_cond: do not condition the class variable on the relation variable
:param omit_emis_cond: do not condition the output/emission variable on the relation variable
"""
if dirname is None:
sys.exit("Output dirname not given.")
self.dirname = dirname
self.N = N
self.start_N = None # for split-merge
self.M = M
self.R = R
self.omit_class_cond = omit_class_cond
self.omit_emis_cond = omit_emis_cond
# initial state probability vector
if self.omit_class_cond:
self.initial_probs = np.zeros(N, 'f')
self.transition_probs = np.zeros([N, N], 'f')
self.final_probs = np.zeros(N, 'f')
else:
self.initial_probs = np.zeros([N, R], 'f')
self.transition_probs = np.zeros([N, N, R], 'f')
self.final_probs = np.zeros([N, R], 'f')
if self.omit_emis_cond:
self.emission_probs = np.zeros([M, N], 'f')
else:
self.emission_probs = np.zeros([M, N, R], 'f')
self.params_fixed_path = None
self.params_fixed_type = None # random init or trained init; set by experimental script
self.brown_init_path = brown_init_path
if not params:
if brown_init_path is None:
self.initialize_params()
self.params_exist = False
else:
if x_dict is None:
sys.exit("wordrep vocab missing")
self.initialize_brown_params(self.brown_init_path, x_dict, dist_even=True)
self.params_exist = False
else:
try:
(self.initial_probs,
self.transition_probs,
self.final_probs,
self.emission_probs) = params
self.initial_probs = self.initial_probs.astype('f', copy=False)
self.transition_probs = self.transition_probs.astype('f', copy=False)
self.final_probs = self.final_probs.astype('f', copy=False)
self.emission_probs = self.emission_probs.astype('f', copy=False)
self.params_exist = True
except ValueError:
print("Number of provided model parameters not right.")
# for updates in em_multiprocess
self.total_ll = 0.0
# Count matrices; use 64 dtype here to avoid overflow
self.initial_counts = np.zeros([self.N, self.R])
self.transition_counts = np.zeros([self.N, self.N, self.R])
self.final_counts = np.zeros([self.N, self.R])
self.emission_counts = np.zeros([self.M, self.N, self.R])
# storing log likelihoods per iteration
self.lls = []
self.sanity_check_init()
self.inference = SumProduct(approximate=approx)
self.max_iter = None
self.n_proc = None
self.n_sent = None
self.data_name = None
self.data_n_tokens = None
#online EM:
self.minibatch_size = None
self.alpha = None
self.a = None
self.permute = None
self.posttypes = None
self.hmm_type = None
self.writeout = writeout
