def load(self, checkpoint_path, hparams, model_name='WaveNet'):
log('Constructing model: {}'.format(model_name))
self._hparams = hparams
local_cond, global_cond = self._check_conditions()
self.local_conditions = tf.placeholder(
tf.float32,
shape=[1, None, hparams.num_mels],
name='local_condition_features') if local_cond else None
self.global_conditions = tf.placeholder(
tf.int32, shape=(),
name='global_condition_features') if global_cond else None
self.synthesis_length = tf.placeholder(
tf.int32, shape=(),
name='synthesis_length') if not local_cond else None
with tf.variable_scope('model') as scope:
self.model = create_model(model_name, hparams)
self.model.initialize(y=None,
c=self.local_conditions,
g=self.global_conditions,
input_lengths=None,
synthesis_length=self.synthesis_length)
self._hparams = hparams
sh_saver = create_shadow_saver(self.model)
log('Loading checkpoint: {}'.format(checkpoint_path))
self.session = tf.Session()
self.session.run(tf.global_variables_initializer())
load_averaged_model(self.session, sh_saver, checkpoint_path)
def model_test_mode(args, feeder, hparams, global_step):
with tf.variable_scope('model', reuse=tf.AUTO_REUSE) as scope:
model_name = 'WaveNet'
model = create_model(model_name, hparams)
#initialize model to test mode
model.initialize(feeder.eval_targets, feeder.eval_local_condition_features, feeder.eval_global_condition_features,
feeder.eval_input_lengths)
model.add_loss()
return model
def model_train_mode(args, feeder, hparams, global_step):
with tf.variable_scope('model', reuse=tf.AUTO_REUSE) as scope:
model_name = 'WaveNet'
model = create_model(model_name, hparams)
#initialize model to train mode
model.initialize(feeder.targets, feeder.local_condition_features, feeder.global_condition_features,
feeder.input_lengths, x=feeder.inputs)
model.add_loss()
model.add_optimizer(global_step)
stats = add_train_stats(model)
return model, stats
def model_test_mode(args, feeder, hparams, global_step):
with tf.variable_scope('model', reuse=tf.AUTO_REUSE) as scope:
model_name = None
if args.model in ('Tacotron-2', 'Both'):
model_name = 'WaveNet'
model = create_model(model_name or args.model, hparams)
#initialize model to test mode
model.initialize(feeder.eval_targets, feeder.eval_local_condition_features, feeder.eval_global_condition_features,
feeder.eval_input_lengths)
model.add_loss()
return model
def model_train_mode(args, feeder, hparams, global_step):
with tf.variable_scope('model', reuse=tf.AUTO_REUSE) as scope:
model_name = None
if args.model in ('Tacotron-2', 'Both'):
model_name = 'WaveNet'
model = create_model(model_name or args.model, hparams)
#initialize model to train mode
model.initialize(feeder.targets, feeder.local_condition_features, feeder.global_condition_features,
feeder.input_lengths, x=feeder.inputs)
model.add_loss()
model.add_optimizer(global_step)
stats = add_train_stats(model)
return model, stats
def load(self, checkpoint_path, hparams, model_name='WaveNet'):
print('Constructing model: {}'.format(model_name))
log('Constructing model: {}'.format(model_name))
self._hparams = hparams
local_cond, global_cond = self._check_conditions()
self.local_conditions = tf.placeholder(
tf.float32,
shape=(None, None, hparams.num_mels),
name='local_condition_features') if local_cond else None
self.global_conditions = tf.placeholder(
tf.int32, shape=(None, 1),
name='global_condition_features') if global_cond else None
self.synthesis_length = tf.placeholder(
tf.int32, shape=(),
name='synthesis_length') if not local_cond else None
self.targets = tf.placeholder(
tf.float32, shape=(1, None, 1), name='audio_targets'
) if hparams.wavenet_synth_debug else None #Debug only with 1 wav
self.input_lengths = tf.placeholder(
tf.int32, shape=(1, ),
name='input_lengths') if hparams.wavenet_synth_debug else None
self.synth_debug = hparams.wavenet_synth_debug
with tf.variable_scope('WaveNet_model') as scope:
self.model = create_model(model_name, hparams)
self.model.initialize(y=None,
c=self.local_conditions,
g=self.global_conditions,
input_lengths=self.input_lengths,
synthesis_length=self.synthesis_length,
test_inputs=self.targets)
self._hparams = hparams
sh_saver = create_shadow_saver(self.model)
print('Loading checkpoint: {}'.format(checkpoint_path))
log('Loading checkpoint: {}'.format(checkpoint_path))
#Memory allocation on the GPU as needed
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.allow_soft_placement = True
self.session = tf.Session(config=config)
self.session.run(tf.global_variables_initializer())
load_averaged_model(self.session, sh_saver, checkpoint_path)
