def reset(self):
"""
Clean observations of the language model use reset
"""
lm_server.post_json_request(self.server_config, 'reset')
self.priors = defaultdict(list)
log.info("\ncleaning history\n")
def init(self, nbest: int = 1):
"""
Initialize the language model (on the server side)
Input:
nbest - top N symbols from evidence
"""
lm_server.post_json_request(self.server_config,
'init',
data={'nbest': nbest})
def init(self, nbest: int = 1):
"""
Initialize the language model (on the server side)
Input:
nbest - top N symbols from evidence
"""
if not isinstance(nbest, int):
raise NBestError(nbest)
if nbest > 4:
raise NBestHighValue(nbest)
lm_server.post_json_request(self.server_config,
'init',
data={'nbest': nbest})
def recent_priors(self, return_mode='letter'):
"""
Display the priors given the recent decision
"""
if not bool(self.priors[return_mode]):
output = lm_server.post_json_request(self.server_config,
'recent_priors',
{'return_mode': return_mode})
return self.__return_priors(output, return_mode)
else:
return self.priors
def state_update(self, evidence: List, return_mode: str = 'letter'):
"""
Provide a prior distribution of the language model
in return to the system's decision regarding the
last observation
Both lm types allow providing more the one timestep
input. Pay attention to the data struct expected.
OCLM
Input:
evidence - a list of (list of) tuples [[(sym1, prob), (sym2, prob2)]]
the numbers are assumed to be in the log probabilty domain
return_mode - 'letter' or 'word' (available
for oclm) strings
Output:
priors - a json dictionary with Normalized priors
in the Negative Log probabilty domain.
"""
# assert the input contains a valid symbol
try:
clean_evidence = []
for tmp_evidence in evidence:
tmp = []
for (symbol, pr) in tmp_evidence:
assert symbol in ALPHABET, \
"%r contains invalid symbol" % evidence
if symbol == SPACE_CHAR:
tmp.append((LM_SPACE, pr))
else:
tmp.append((symbol.lower(), pr))
clean_evidence.append(tmp)
except BaseException:
raise EvidenceDataStructError
output = lm_server.post_json_request(self.server_config,
'state_update', {
'evidence': clean_evidence,
'return_mode': return_mode
})
return self.__return_priors(output, return_mode)
def state_update(self, evidence: List, return_mode: str = 'letter'):
"""
Provide a prior distribution of the language model
in return to the system's decision regarding the
last observation
Both lm types allow providing more the one timestep
input. Pay attention to the data struct expected.
Input:
decision - a symbol or a string of symbols in encapsulated in a
list
the numbers are assumed to be in the log probability domain
return_mode - 'letter' or 'word' (available
for oclm) strings
Output:
priors - a json dictionary with Normalized priors
in the Negative Log probability domain.
"""
assert return_mode == 'letter', "PRELM only allows letter output"
# assert the input contains a valid symbol
decision = evidence # in prelm the we treat it as a decision
for symbol in decision:
assert symbol in ALPHABET or ' ', \
"%r contains invalid symbol" % decision
clean_evidence = []
for symbol in decision:
if symbol == SPACE_CHAR:
symbol = LM_SPACE
clean_evidence.append(symbol.lower())
output = lm_server.post_json_request(self.server_config,
'state_update', {
'evidence': clean_evidence,
'return_mode': return_mode
})
return self.__return_priors(output, return_mode)
