def __init__(self, hparams: ExtendedHParams, id_list: List[str]):
if hparams is None:
hparams = self.create_hparams()
hparams.out_dir = os.path.curdir
# super(AcousticModelTrainer, self).__init__(id_list=id_list, data_reader_configs=data_reader_configs, hparams=hparams) # Call ModelTrainer base class init directly.
super().__init__(id_list=id_list, hparams=hparams)
if hparams.scheduler_type == "default":
hparams.scheduler_type = "Plateau"
hparams.scheduler_args["verbose"] = True
def __init__(self,
hparams: ExtendedHParams,
id_list: List[str],
data_reader_configs: List[DataReaderConfig] = None):
if hparams is None:
hparams = self.create_hparams()
hparams.out_dir = os.path.curdir
# Write missing default parameters.
# if hparams.synth_dir is None:
# hparams.synth_dir = os.path.join(hparams.out_dir, "synth")
super().__init__(data_reader_configs=data_reader_configs,
id_list=id_list,
hparams=hparams)
# in_to_out_multiplier = int(hparams.frame_rate_output_Hz /
# (1000.0 / hparams.frame_size_ms))
# num_frames_per_sec = 1000.0 / hparams.frame_size_ms
# # NOTE: Had to limit input length because of memory constraints.
# max_frames_trainset = int(num_frames_per_sec * hparams.max_input_train_sec) \
# * in_to_out_multiplier # Multiply by number of seconds.
# max_frames_testset = int(num_frames_per_sec * hparams.max_input_test_sec) \
# * in_to_out_multiplier # Ensure that test takes all frames.
# self.dataset_train = LabelGensDataset(
# self.id_list_train, self.InputGen, self.OutputGen, hparams,
# random_select=True, max_frames_input=max_frames_trainset)
# self.dataset_val = LabelGensDataset(
# self.id_list_val, self.InputGen, self.OutputGen, hparams,
# random_select=True, max_frames_input=max_frames_testset)
# if self.loss_function is None:
# if hparams.input_type == "mulaw-quantize":
# self.loss_function = OneHotCrossEntropyLoss(reduction='none', shift=1)
# else:
# self.loss_function = DiscretizedMixturelogisticLoss(hparams.quantize_channels,
# hparams.log_scale_min,
# reduction='none',
# hinge_loss=hparams.hinge_regularizer)
if hparams.scheduler_type == "default":
hparams.scheduler_type = "Noam"
# hparams.scheduler_args['exponential_gamma'] = 0.99
hparams.scheduler_args['wormup_steps'] = 4000
# Override the collate and decollate methods of batches.
self.batch_collate_fn = partial(self.prepare_batch,
use_cond=hparams.use_cond,
one_hot_target=True)
self.batch_decollate_fn = self.decollate_network_output
def __init__(self,
hparams: ExtendedHParams,
id_list: List[str],
data_reader_configs: List[DataReader.Config] = None):
if hparams is None:
hparams = self.create_hparams()
hparams.out_dir = os.path.curdir
super(AcousticModelTrainer,
self).__init__(data_reader_configs=data_reader_configs,
id_list=id_list,
hparams=hparams)
if hparams.scheduler_type == "default":
hparams.scheduler_type = "Plateau"
hparams.add_hparams(plateau_verbose=True)