def __init__(self, checkpoint_path, hparams, gta=False, model_name="Tacotron", seed=None):
log("Constructing model: %s" % model_name)
# Initialize tensorflow random number seed for deterministic operation if provided
if seed is not None:
tf.compat.v1.set_random_seed(seed)
#Force the batch size to be known in order to use attention masking in batch synthesis
inputs = tf.compat.v1.placeholder(tf.int32, (None, None), name="inputs")
input_lengths = tf.compat.v1.placeholder(tf.int32, (None,), name="input_lengths")
speaker_embeddings = tf.compat.v1.placeholder(tf.float32, (None, hparams.speaker_embedding_size),
name="speaker_embeddings")
targets = tf.compat.v1.placeholder(tf.float32, (None, None, hparams.num_mels), name="mel_targets")
split_infos = tf.compat.v1.placeholder(tf.int32, shape=(hparams.tacotron_num_gpus, None), name="split_infos")
with tf.compat.v1.variable_scope("Tacotron_model") as scope:
self.model = create_model(model_name, hparams)
if gta:
self.model.initialize(inputs, input_lengths, speaker_embeddings, targets, gta=gta,
split_infos=split_infos)
else:
self.model.initialize(inputs, input_lengths, speaker_embeddings,
split_infos=split_infos)
self.mel_outputs = self.model.tower_mel_outputs
self.linear_outputs = self.model.tower_linear_outputs if (hparams.predict_linear and not gta) else None
self.alignments = self.model.tower_alignments
self.stop_token_prediction = self.model.tower_stop_token_prediction
self.targets = targets
self.gta = gta
self._hparams = hparams
#pad input sequences with the <pad_token> 0 ( _ )
self._pad = 0
#explicitely setting the padding to a value that doesn"t originally exist in the spectogram
#to avoid any possible conflicts, without affecting the output range of the model too much
if hparams.symmetric_mels:
self._target_pad = -hparams.max_abs_value
else:
self._target_pad = 0.
self.inputs = inputs
self.input_lengths = input_lengths
self.speaker_embeddings = speaker_embeddings
self.targets = targets
self.split_infos = split_infos
log("Loading checkpoint: %s" % checkpoint_path)
#Memory allocation on the GPUs as needed
config = tf.compat.v1.ConfigProto()
config.gpu_options.allow_growth = True
config.allow_soft_placement = True
self.session = tf.compat.v1.Session(config=config)
self.session.run(tf.compat.v1.global_variables_initializer())
saver = tf.compat.v1.train.Saver()
saver.restore(self.session, checkpoint_path)
def model_test_mode(args, feeder, hparams, global_step):
with tf.variable_scope("Tacotron_model", reuse=tf.AUTO_REUSE) as scope:
model = create_model("Tacotron", hparams)
model.initialize(feeder.eval_inputs, feeder.eval_input_lengths,
feeder.eval_speaker_embeddings, feeder.eval_mel_targets,
feeder.eval_token_targets, targets_lengths=feeder.eval_targets_lengths,
global_step=global_step, is_training=False, is_evaluating=True,
split_infos=feeder.eval_split_infos)
model.add_loss()
return model
def model_train_mode(args, feeder, hparams, global_step):
with tf.variable_scope("Tacotron_model", reuse=tf.AUTO_REUSE) as scope:
model = create_model("Tacotron", hparams)
model.initialize(feeder.inputs, feeder.input_lengths, feeder.speaker_embeddings,
feeder.mel_targets, feeder.token_targets,
targets_lengths=feeder.targets_lengths, global_step=global_step,
is_training=True, split_infos=feeder.split_infos)
model.add_loss()
model.add_optimizer(global_step)
stats = add_train_stats(model, hparams)
return model, stats
