def flow(self):
self.ori_embed = speaker_embedding(self.ori_spk,
hp.SPK_NUM,
hp.EMBED_SIZE,
reuse=False) # [N, 1] -> [N, E]
self.aim_embed = speaker_embedding(self.aim_spk,
hp.SPK_NUM,
hp.EMBED_SIZE,
reuse=True) # [N, 1] -> [N, E]
with tf.variable_scope(self.model_name, reuse=tf.AUTO_REUSE):
self.ori_out, self.ori_mu, self.ori_log_var = generator(
self.ori_embed,
self.ori_feat,
is_training=self.is_training,
reuse=False,
scope_name='generator')
self.aim_out, self.aim_mu, self.aim_log_var = generator(
self.aim_embed,
self.ori_feat,
is_training=self.is_training,
reuse=True,
scope_name='generator')
self.cycle_ori_out, self.cycle_mu, self.cycle_log_var = generator(
self.ori_embed,
self.aim_out,
is_training=self.is_training,
reuse=True,
scope_name='generator')
self.predict_real_P = discriminator(self.aim_feat,
reuse=False,
scope_name='discriminator')
self.predict_fake_P = discriminator(self.aim_out,
reuse=True,
scope_name='discriminator')
def _create_GANs(self):
"""
创建生成器、判别器相应符号
:return:
"""
self.fakeA = generator(self.realB_ph, False, name='Gb2a')
self.fakeB = generator(self.realA_ph, False, name='Ga2b')
self.cycA = generator(self.fakeB, True, name='Gb2a')
self.cycB = generator(self.fakeA, True, name='Ga2b')
self.Da_fake = discriminator(self.fakeA, False, name='Da')
self.Db_fake = discriminator(self.fakeB, False, name='Db')
self.Da_real = discriminator(self.realA_ph, True, name='Da')
self.Db_real = discriminator(self.realB_ph, True, name='Db')
self.Da_fake_sample = discriminator(self.fakeA_ph, True, name='Da')
self.Db_fake_sample = discriminator(self.fakeB_ph, True, name='Db')
def inference(self):
self.ori_mel = tf.placeholder(name='ori_mel', shape=[None, None, hp.CODED_DIM], dtype=tf.float32)
# self.aim_spk = tf.placeholder(name='aim_spk', shape=[None, None], dtype=tf.int64)
### self.flow()
# self.aim_embed = speaker_embedding(self.aim_spk, hp.SPK_NUM, hp.EMBED_SIZE) # [N, 1] -> [N, E]
with tf.variable_scope(self.model_name, reuse=tf.AUTO_REUSE):
self.aim_out, self.aim_mu, self.aim_log_var = generator(self.ori_mel,
is_training=self.is_training,
scope_name='generator')
def __init__(self, **kwargs):
# hyperparams
self.__dict__.update(self.default_hyperparams)
self.__dict__.update(kwargs)
# deserialize some hyperparams
for k in ('image_shape', 'tf_config_proto'):
self.__dict__[k] = json.loads(self.__dict__[k])
assert isinstance(self.image_shape, list) and len(self.image_shape)==3
self.tf_config = tf.ConfigProto(**self.tf_config_proto)
# instantiate optimizers
self.disc_optimizer = tf.train.AdamOptimizer(self.disc_learning_rate)
self.gen_optimizer = tf.train.AdamOptimizer(self.gen_learning_rate)
# instantiate models
self.discriminator_A = modules.discriminator(input_shape=self.image_shape)
self.discriminator_B = modules.discriminator(input_shape=self.image_shape)
self.generator_A2B = modules.generator(input_shape=self.image_shape,
n_resnet_blocks=self.generator_n_resnet_blocks)
self.generator_B2A = modules.generator(input_shape=self.image_shape,
n_resnet_blocks=self.generator_n_resnet_blocks)
for model in (self.discriminator_A, self.discriminator_B, self.generator_A2B, self.generator_B2A):
# 'sgd' and 'mse' are not actually used, just placeholders for compiling
model.compile('sgd', 'mse')
def flow(self):
# 不需要embeding了
# self.ori_embed = speaker_embedding(self.ori_spk, hp.SPK_NUM, hp.EMBED_SIZE, reuse=False) # [N, 1] -> [N, E]
# self.aim_embed = speaker_embedding(self.aim_spk, hp.SPK_NUM, hp.EMBED_SIZE, reuse=True) # [N, 1] -> [N, E]
with tf.variable_scope(self.model_name, reuse=tf.AUTO_REUSE):
# self.ori_out, self.ori_mu, self.ori_log_var = generator(self.ori_mel,
# is_training=self.is_training,
# reuse=False,
# scope_name='generator')
# # 没有自己变自己的需求了
self.aim_out, self.aim_mu, self.aim_log_var = generator(self.ori_mel,
is_training=self.is_training,
reuse=tf.AUTO_REUSE,
scope_name='generator')
def __init__(self, args, config):
self.args = args
model_name = config.get('config', 'experiment')
task_channel = config.getint('task', 'channel')
task_module_type = config.get('task', 'classifier_type')
image_fc = config.getint('task', 'image_fc')
measurement_fc = config.getint('task', 'measurement_fc')
branch_fc = config.getint('task', 'branch_fc')
self.source_only = config.getboolean('config', 'source_only')
self.t2s_task = config.getboolean('config', 't2s_task')
self.style_weights = config.getlist('discriminator', 'style_weights')
self.share_all_image = config.getboolean('generator',
'share_all_image')
if not self.source_only:
discriminator_channel = config.getint('discriminator', 'channel')
self.discriminator_dropout_prob = config.getfloat(
'discriminator', 'dropout_prob')
self.patch = config.getboolean('discriminator', 'patch')
self.generator_type = config.get('generator', 'type')
generator_channel = config.getint('generator', 'channel')
self.generator_out_channel = config.getint('generator',
'out_channel')
self.generator = modules.generator(generator_channel, config, args)
self.discriminator = modules.discriminator(discriminator_channel,
group_size=1)
# List of 4 branch modules
self.task = modules.task(task_channel,
image_fc,
measurement_fc,
branch_fc,
self.args.training,
classifier_type=task_module_type)
self.summary = dict()
def __init__(self, args, config):
self.args = args
model_name = config.get('config', 'experiment')
discriminator_channel = config.getint('discriminator', 'channel')
generator_channel = config.getint('generator', 'channel')
classifier_channel = config.getint('task', 'channel')
self.generator_type = config.get('generator', 'type')
self.generator_out_channel = config.getint('generator', 'out_channel')
self.generator = modules.generator(generator_channel, config, self.args)
self.discriminator = modules.discriminator(discriminator_channel, group_size=self.args.batch_size)
self.patch = config.getboolean('discriminator', 'patch')
self.discriminator_dropout_prob = config.getfloat('discriminator', 'dropout_prob')
self.t2s_task = config.getboolean('config', 't2s_task')
self.style_weights = config.getlist('discriminator', 'style_weights')
self.transferred_classifier = modules.task_classifier(classifier_channel, num_classes=3, training=self.args.training)
self.transferred_task_weight = config.getfloat(model_name, 'task_weight')
self.t2s_task_weight = config.getfloat(model_name, 't2s_task_weight')
self.summary = dict()
