def construct(cls, config_path: str, alphabet_path: str) -> 'DeepSpeech':
""" Construct DeepSpeech object base on the configuration and the alphabet files. """
config = Configuration(config_path)
model_dir = os.path.dirname(config_path)
gpus = get_available_gpus()
model = cls.get_model(is_gpu=len(gpus) > 0, **config.model)
loss = cls.get_loss()
optimizer = cls.get_optimizer(**config.optimizer)
callbacks = cls.get_callbacks(home_dir=model_dir, configurations=config.callbacks)
alphabet = cls.get_alphabet(alphabet_path)
features_extractor = cls.get_features_extractor(**config.features_extractor)
decoder = cls.get_decoder(alphabet=alphabet, model=model, **config.decoder)
return cls(model, loss, optimizer, callbacks, alphabet, decoder, features_extractor, gpus)
def train(self):
""" Train model using train and dev data generators """
config_dataset = self.configuration.dataset
train_generator = DataGenerator(name='train',
csv_path=config_dataset.train_csv_path,
alphabet=self.alphabet,
**config_dataset.parameters)
dev_generator = DataGenerator(name='dev',
csv_path=config_dataset.dev_csv_path,
alphabet=self.alphabet,
**config_dataset.parameters)
gpus = get_available_gpus()
if len(gpus) > 1:
distributed_model = multi_gpu_model(self.model, len(gpus))
else:
distributed_model = self.model
optimizer = self.__get_optimizer()
objective = self.__get_objective()
y = Input(name='y', shape=[None], dtype='int32')
distributed_model.compile(optimizer=optimizer,
loss=objective,
target_tensors=[y])
# The template model shares the same weights, but it is not distributed
# along different devices.
distributed_model.template_model = self.model
callbacks = self.__get_callbacks()
history = distributed_model.fit_generator(
generator=train_generator,
validation_data=dev_generator,
callbacks=callbacks,
shuffle=False,
**self.configuration.fit_generator.parameters)
self.__set_best_weights_to_model(history)