meldir=str(config_manager.train_datadir / 'mels'),
num_samples=config['n_samples']) # (phonemes, mel)
val_samples, _ = load_files(
metafile=str(config_manager.train_datadir / 'test_metafile.txt'),
meldir=str(config_manager.train_datadir / 'mels'),
num_samples=config['n_samples']) # (phonemes, text, mel)
# get model, prepare data for model, create datasets
model = config_manager.get_model()
config_manager.compile_model(model)
data_prep = DataPrepper(config=config, tokenizer=model.text_pipeline.tokenizer)
test_list = [data_prep(s) for s in val_samples]
train_dataset = Dataset(samples=train_samples,
preprocessor=data_prep,
batch_size=config['batch_size'],
mel_channels=config['mel_channels'],
shuffle=True)
val_dataset = Dataset(samples=val_samples,
preprocessor=data_prep,
batch_size=config['batch_size'],
mel_channels=config['mel_channels'],
shuffle=False)
# create logger and checkpointer and restore latest model
summary_manager = SummaryManager(model=model,
log_dir=config_manager.log_dir,
config=config)
checkpoint = tf.train.Checkpoint(step=tf.Variable(1),
optimizer=model.optimizer,
metafile=str(train_meta),
meldir=str(meldir),
num_samples=config['n_samples']) # (phonemes, mel)
val_samples, _ = load_files(
metafile=str(test_meta),
meldir=str(meldir),
num_samples=config['n_samples']) # (phonemes, text, mel)
# get model, prepare data for model, create datasets
data_prep = DataPrepper(config=config,
tokenizer=model.text_pipeline.tokenizer)
script_batch_size = 5 * config['batch_size'] # faster parallel computation
train_dataset = Dataset(samples=train_samples,
preprocessor=data_prep,
batch_size=script_batch_size,
shuffle=False,
drop_remainder=False)
val_dataset = Dataset(samples=val_samples,
preprocessor=data_prep,
batch_size=script_batch_size,
shuffle=False,
drop_remainder=False)
if model.r != 1:
print(
f"ERROR: model's reduction factor is greater than 1, check config. (r={model.r}"
)
# identify last decoder block
n_layers = len(config_manager.config['decoder_num_heads'])
n_dense = int(config_manager.config['decoder_dense_blocks'])
n_convs = int(n_layers - n_dense)
config_manager = ConfigManager(config_path=args.config,
model_kind='forward',
session_name=args.session_name)
config = config_manager.config
config_manager.create_remove_dirs(clear_dir=args.clear_dir,
clear_logs=args.clear_logs,
clear_weights=args.clear_weights)
config_manager.dump_config()
config_manager.print_config()
train_data_list = build_file_list(config_manager.train_datadir /
'forward_data/train')
dataprep = ForwardDataPrepper()
train_dataset = Dataset(samples=train_data_list,
mel_channels=config['mel_channels'],
preprocessor=dataprep,
batch_size=config['batch_size'],
shuffle=True)
val_data_list = build_file_list(config_manager.train_datadir /
'forward_data/val')
val_dataset = Dataset(samples=val_data_list,
mel_channels=config['mel_channels'],
preprocessor=dataprep,
batch_size=config['batch_size'],
shuffle=False)
# get model, prepare data for model, create datasets
model = config_manager.get_model()
config_manager.compile_model(model)
# create logger and checkpointer and restore latest model