def __init__(self, am_config, lm_config):
self.am = AM(am_config)
self.am.load_model(False)
self.lm = LM(lm_config)
self.lm.load_model(False)
def __init__(self,config):
self.config = config
self.dg = LM_DataLoader(config)
lm=LM(config)
lm.load_model()
self.model = lm.model
self.optimizer = tf.keras.optimizers.Adamax(**config['optimizer_config'])
self.runner = lm_runners.LMTrainer(self.config['running_config'],one2one=self.model.one2one)
self.runner.set_total_train_steps(self.dg.get_per_epoch_steps() * self.config['running_config']['num_epochs'])
self.runner.compile(self.model,self.optimizer)
def __init__(self, am_config, lm_config, punc_config=None):
self.am = AM(am_config)
self.am.load_model(False)
self.lm = LM(lm_config, punc_config)
self.lm.load_model(False)
if punc_config is not None:
self.punc_recover = True
else:
self.punc_recover = False
def __init__(self, config):
self.config = config
self.dg = LM_DataLoader(config, training=False)
lm = LM(config)
lm.load_model(training=False)
self.model = lm.model
self.runner = lm_tester.LMTester(
self.config['running_config'],
self.config['model_config']['one2one'])
self.runner.set_progbar(self.dg.eval_per_epoch_steps())
self.runner.compile(self.model)
def __init__(self,config):
self.config = config
self.dg = LM_DataLoader(config)
lm=LM(config)
lm.load_model()
self.model = lm.model
all_train_step = self.dg.get_per_epoch_steps() * self.config['running_config']['num_epochs']
lr = CustomSchedule(config['model_config']['d_model'], warmup_steps=int(all_train_step * 0.1))
config['optimizer_config']['learning_rate'] = lr
self.optimizer = tf.keras.optimizers.Adamax(**config['optimizer_config'])
self.runner = lm_runners.LMTrainer(self.config['running_config'],one2one=self.model.one2one)
self.runner.set_total_train_steps(self.dg.get_per_epoch_steps() * self.config['running_config']['num_epochs'])
self.runner.compile(self.model,self.optimizer)
class ASR():
def __init__(self, am_config, lm_config):
self.am = AM(am_config)
self.am.load_model(False)
self.lm = LM(lm_config)
self.lm.load_model()
def decode_am_result(self, result):
return self.am.decode_result(result[0])
def stt(self, wav_path):
am_result = self.am.predict(wav_path)
lm_result = self.lm.predict(self.decode_am_result(am_result))
return am_result, lm_result
class ASR():
def __init__(self, am_config, lm_config):
self.am = AM(am_config)
self.am.load_model(False)
self.lm = LM(lm_config)
self.lm.load_model(False)
def decode_am_result(self, result):
return self.am.decode_result(result)
def stt(self, wav_path):
am_result = self.am.predict(wav_path)
if self.am.model_type == 'Transducer':
am_result = self.decode_am_result(am_result[1:-1])
lm_result = self.lm.predict(am_result)
lm_result = self.lm.decode(lm_result[0].numpy(),
self.lm.lm_featurizer)
else:
am_result = self.decode_am_result(am_result[0])
lm_result = self.lm.predict(am_result)
lm_result = self.lm.decode(lm_result[0].numpy(),
self.lm.lm_featurizer)
return am_result, lm_result
def am_test(self, wav_path):
# am_result is token id
am_result = self.am.predict(wav_path)
# token to vocab
if self.am.model_type == 'Transducer':
am_result = self.decode_am_result(am_result[1:-1])
else:
am_result = self.decode_am_result(am_result[0])
return am_result
def lm_test(self, txt):
if self.lm.config['am_token']['for_multi_task']:
pys = pypinyin.pinyin(txt, 8, neutral_tone_with_five=True)
input_py = [i[0] for i in pys]
else:
pys = pypinyin.pinyin(txt)
input_py = [i[0] for i in pys]
# now lm_result is token id
lm_result = self.lm.predict(input_py)
# token to vocab
lm_result = self.lm.decode(lm_result[0].numpy(), self.lm.lm_featurizer)
return lm_result
def __init__(self, lm_config, punc_config):
self.config = punc_config
self.dg = Punc_DataLoader(punc_config)
lm = LM(lm_config, punc_config)
lm.load_model()
self.model = lm.punc_model
all_train_step = self.dg.get_per_epoch_steps(
) * self.config['running_config']['num_epochs']
lr = CustomSchedule(self.config['model_config']['d_model'],
warmup_steps=int(all_train_step * 0.1))
self.config['optimizer_config']['learning_rate'] = lr
self.optimizer = tf.keras.optimizers.Adamax(
**self.config['optimizer_config'])
self.runner = punc_trainer.PuncTrainer(self.config['running_config'])
self.runner.set_total_train_steps(
self.dg.get_per_epoch_steps() *
self.config['running_config']['num_epochs'])
self.runner.compile(self.model, self.optimizer)
