def build_tokenizer_by_config(self, tok_config, lang):
if tok_config is None:
tok_config = {"mode": "aggressive"}
if lang == "zh":
tok_config["segment_alphabet"] = ["Han"]
tok_config["segment_alphabet_change"] = True
# to avoid SentencePiece sampling
if "sp_nbest_size" in tok_config:
tok_config["sp_nbest_size"] = 0
return tokenizer.build_tokenizer(tok_config)
def build_tokenizer_by_config(self, tok_config, lang):
if tok_config is None:
tok_config = {"mode": "aggressive"}
if lang == 'zh':
tok_config['segment_alphabet'] = ['Han']
tok_config['segment_alphabet_change'] = True
# to avoid SentencePiece sampling
if 'sp_nbest_size' in tok_config:
tok_config['sp_nbest_size'] = 0
return tokenizer.build_tokenizer(tok_config)
def _build_subword_learner(tok_config, result_dir, ref_tok_config=None):
subword_config = tok_config.get("build_subword")
if subword_config is None:
return {}
if ref_tok_config is None:
ref_tok_config = tok_config
subword_info = tokenizer.make_subword_learner(
subword_config,
result_dir,
tokenizer=tokenizer.build_tokenizer(ref_tok_config))
return subword_info
def _build_process(self, config, side, build_state):
# Disable subword regularization in inference.
if self.process_type != prepoperator.ProcessType.TRAINING:
config["bpe_dropout"] = 0
config["sp_nbest_size"] = 0
config["sp_alpha"] = 0
if config.get("restrict_subword_vocabulary", False):
vocabulary_path = build_state.get("src_vocabulary" if side ==
"source" else "tgt_vocabulary")
if vocabulary_path is None:
raise ValueError(
"restrict_subword_vocabulary is set but no vocabulary is set"
)
# The open source Tokenizer does not accept the custom vocabulary format
# produced by build_vocab so we create a temporary vocabulary with a simpler
# format.
with tempfile.NamedTemporaryFile(mode="w") as vocab_file:
for token in tokenizer.load_vocabulary(vocabulary_path):
vocab_file.write("%s\n" % token)
vocab_file.flush()
config["vocabulary_path"] = vocab_file.name
current_tokenizer = tokenizer.build_tokenizer(config)
else:
current_tokenizer = tokenizer.build_tokenizer(config)
previous_tokenizer = None
if build_state:
if side == "source":
previous_tokenizer = build_state["src_tokenizer"]
build_state["src_tokenizer"] = current_tokenizer
else:
previous_tokenizer = build_state["tgt_tokenizer"]
build_state["tgt_tokenizer"] = current_tokenizer
if (self.process_type == prepoperator.ProcessType.POSTPROCESS
and not self._postprocess_only):
return previous_tokenizer
return current_tokenizer
