def build_cae_from_options_dict(a, a_lengths, b_lengths, options_dict):
# Latent layer
build_latent_func = tflego.build_autoencoder
latent_func_kwargs = {
"enc_n_hiddens": [],
"n_z": options_dict["n_z"],
"dec_n_hiddens": [options_dict["dec_n_hiddens"][0]],
"activation": tf.nn.relu
}
# Network
network_dict = tflego.build_multi_encdec_lazydynamic_latentfunc(
a,
a_lengths,
options_dict["enc_n_hiddens"],
options_dict["dec_n_hiddens"],
build_latent_func,
latent_func_kwargs,
y_lengths=b_lengths,
rnn_type=options_dict["rnn_type"],
bidirectional=options_dict["bidirectional"],
keep_prob=options_dict["keep_prob"],
add_conditioning_tensor=None)
encoder_states = network_dict["encoder_states"]
ae = network_dict["latent_layer"]
z = ae["z"]
y = network_dict["decoder_output"]
mask = network_dict["mask"]
y *= tf.expand_dims(mask, -1) # safety
return {"z": z, "y": y, "mask": mask}
def build_cae_from_options_dict(a, a_lengths, b_lengths, options_dict):
# Latent layer
build_latent_func = tflego.build_autoencoder
latent_func_kwargs = {
"enc_n_hiddens": [],
"n_z": options_dict["n_z"],
"dec_n_hiddens": [options_dict["dec_n_hiddens"][0]],
"activation": tf.nn.relu
}
# Speaker embedding
if options_dict["d_speaker_embedding"] is not None:
speaker_id = tf.placeholder(TF_ITYPE, [None])
with tf.variable_scope("speaker_embedding"):
speaker_embedding = tf.get_variable(
"E", [
options_dict["n_speakers"],
options_dict["d_speaker_embedding"]
],
dtype=TF_DTYPE,
initializer=tf.contrib.layers.xavier_initializer())
embedding_lookup = tf.nn.embedding_lookup(speaker_embedding,
speaker_id)
# Network
network_dict = tflego.build_multi_encdec_lazydynamic_latentfunc(
a,
a_lengths,
options_dict["enc_n_hiddens"],
options_dict["dec_n_hiddens"],
build_latent_func,
latent_func_kwargs,
y_lengths=b_lengths,
rnn_type=options_dict["rnn_type"],
bidirectional=options_dict["bidirectional"],
keep_prob=options_dict["keep_prob"],
add_conditioning_tensor=None
if options_dict["d_speaker_embedding"] is None else embedding_lookup)
encoder_states = network_dict["encoder_states"]
ae = network_dict["latent_layer"]
z = ae["z"]
y = network_dict["decoder_output"]
mask = network_dict["mask"]
y *= tf.expand_dims(mask, -1) # safety
if options_dict["d_speaker_embedding"] is not None:
return {
"z": z,
"y": y,
"mask": mask,
"speaker_id": speaker_id,
"speaker_embedding": speaker_embedding
}
else:
return {"z": z, "y": y, "mask": mask}
def build_vae_from_options_dict(x, x_lengths, options_dict):
build_latent_func = tflego.build_vae
latent_func_kwargs = {
"enc_n_hiddens": [],
"n_z": options_dict["n_z"],
"dec_n_hiddens": [options_dict["dec_n_hiddens"][0]],
"activation": tf.nn.relu
}
network_dict = tflego.build_multi_encdec_lazydynamic_latentfunc(
x,
x_lengths,
options_dict["enc_n_hiddens"],
options_dict["dec_n_hiddens"],
build_latent_func,
latent_func_kwargs,
rnn_type=options_dict["rnn_type"],
keep_prob=options_dict["keep_prob"])
network_dict["decoder_output"] *= tf.expand_dims(network_dict["mask"], -1)
# safety
return network_dict
