def __init__(self, input_dim=600+9, output_dim=1*3, dropout_prob=0.):
"""Initialises acoustic model parameters and settings."""
super(F0_RNN, self).__init__()
self.input_dim = input_dim
self.output_dim = output_dim
self.dropout_prob = dropout_prob
self.recurrent_layers = utils.SequentialWithRecurrent(
nn.Linear(self.input_dim, 256),
nn.Sigmoid(),
nn.Dropout(p=dropout_prob),
utils.RecurrentCuDNNWrapper(
nn.GRU(256, 64, batch_first=True)),
nn.Dropout(p=dropout_prob),
utils.RecurrentCuDNNWrapper(
nn.GRU(64, 64, batch_first=True)),
nn.Dropout(p=dropout_prob),
utils.RecurrentCuDNNWrapper(
nn.GRU(64, 64, batch_first=True)),
nn.Dropout(p=dropout_prob),
nn.Linear(64, 64),
nn.Sigmoid(),
nn.Dropout(p=dropout_prob),
nn.Linear(64, self.output_dim),
)
self.metrics.add_metrics('all',
LF0_RMSE_Hz=LF0Distortion())
def __init__(self,
z_dim=16,
kld_weight=1.,
conditioning_dim=600 + 9,
output_dim=1 * 3,
dropout_prob=0.,
latent=None):
"""Initialises VAE parameters and settings."""
super(VAE, self).__init__(z_dim=z_dim, kld_weight=kld_weight)
self.conditioning_dim = conditioning_dim
self.output_dim = output_dim
self.dropout_prob = dropout_prob
self.latent = latent
self.encoder_layer = _Encoder(self.conditioning_dim, self.output_dim,
dropout_prob, self.z_dim)
self.decoder_layer = utils.SequentialWithRecurrent(
nn.Linear(self.conditioning_dim + self.z_dim, 256),
nn.Sigmoid(),
nn.Dropout(p=dropout_prob),
utils.RecurrentCuDNNWrapper(nn.GRU(256, 64, batch_first=True)),
nn.Dropout(p=dropout_prob),
utils.RecurrentCuDNNWrapper(nn.GRU(64, 64, batch_first=True)),
nn.Dropout(p=dropout_prob),
utils.RecurrentCuDNNWrapper(nn.GRU(64, 64, batch_first=True)),
nn.Dropout(p=dropout_prob),
nn.Linear(64, 64),
nn.Sigmoid(),
nn.Dropout(p=dropout_prob),
nn.Linear(64, self.output_dim),
)
self.metrics.add_metrics('all', LF0_RMSE_Hz=LF0Distortion())
def __init__(self,
conditioning_dim,
input_dim,
dropout_prob,
z_dim,
latent=None):
super(_Encoder, self).__init__()
self.conditioning_dim = conditioning_dim
self.input_dim = input_dim
self.z_dim = z_dim
self.latent = latent
self.shared_layer = utils.SequentialWithRecurrent(
nn.Linear(self.conditioning_dim + self.input_dim, 256),
nn.Sigmoid(),
nn.Dropout(p=dropout_prob),
utils.RecurrentCuDNNWrapper(nn.GRU(256, 64, batch_first=True)),
nn.Dropout(p=dropout_prob),
utils.RecurrentCuDNNWrapper(nn.GRU(64, 64, batch_first=True)),
nn.Dropout(p=dropout_prob),
utils.RecurrentCuDNNWrapper(nn.GRU(64, 64, batch_first=True)),
nn.Dropout(p=dropout_prob),
nn.Linear(64, 64),
nn.Sigmoid(),
nn.Dropout(p=dropout_prob),
)
self.mu_layer = nn.Linear(64, self.z_dim)
self.logvar_layer = nn.Linear(64, self.z_dim)
def __init__(self, z_dim=16, conditioning_dim=70+9, output_dim=1*3, dropout_prob=0.,
phone_set_file='data/unilex_phoneset.txt'):
"""Initialises VAE parameters and settings."""
super(F0_AE, self).__init__()
self.z_dim = z_dim
self.conditioning_dim = conditioning_dim
self.output_dim = output_dim
self.dropout_prob = dropout_prob
self.phone_set_file = phone_set_file
self.encoder_layer = utils.SequentialWithRecurrent(
nn.Linear(self.output_dim, 256),
nn.Sigmoid(),
nn.Dropout(p=dropout_prob),
utils.RecurrentCuDNNWrapper(
nn.GRU(256, 64, batch_first=True)),
nn.Dropout(p=dropout_prob),
utils.RecurrentCuDNNWrapper(
nn.GRU(64, 64, batch_first=True)),
nn.Dropout(p=dropout_prob),
utils.RecurrentCuDNNWrapper(
nn.GRU(64, 64, batch_first=True)),
nn.Dropout(p=dropout_prob),
nn.Linear(64, 64),
nn.Sigmoid(),
nn.Dropout(p=dropout_prob),
nn.Linear(64, self.z_dim)
)
self.decoder_layer = utils.SequentialWithRecurrent(
nn.Linear(self.conditioning_dim + self.z_dim, 256),
nn.Sigmoid(),
nn.Dropout(p=dropout_prob),
utils.RecurrentCuDNNWrapper(
nn.GRU(256, 64, batch_first=True)),
nn.Dropout(p=dropout_prob),
utils.RecurrentCuDNNWrapper(
nn.GRU(64, 64, batch_first=True)),
nn.Dropout(p=dropout_prob),
utils.RecurrentCuDNNWrapper(
nn.GRU(64, 64, batch_first=True)),
nn.Dropout(p=dropout_prob),
nn.Linear(64, 64),
nn.Sigmoid(),
nn.Dropout(p=dropout_prob),
nn.Linear(64, self.output_dim),
)
self.metrics.add_metrics('all',
LF0_RMSE_Hz=LF0Distortion(),
embeddings=TensorHistory(self.z_dim, hidden=True),
name=History(hidden=True),
n_segments=TensorHistory(1, dtype=torch.long, hidden=True),
segment_mean_F0=TensorHistory(1, hidden=True))
def __init__(self,
input_dim=600 + 9,
output_dims=None,
dropout_prob=0.,
num_layers=8):
"""Initialises acoustic model parameters and settings."""
if output_dims is None:
output_dims = {
'lf0': 1 * 3,
'vuv': 1,
'mcep': 60 * 3,
'bap': 5 * 3
}
super(LSTMAcousticModel, self).__init__()
self.input_dim = input_dim
self.output_dims = output_dims
self.dropout_prob = dropout_prob
self.num_layers = num_layers
self.layers = utils.SequentialWithRecurrent(
nn.Linear(self.input_dim, 512),
nn.Sigmoid(),
nn.Dropout(p=self.dropout_prob),
*[
utils.RecurrentCuDNNWrapper(
nn.LSTM(512,
512,
dropout=self.dropout_prob,
batch_first=True)) for _ in range(self.num_layers)
],
nn.Linear(512, 256),
nn.Sigmoid(),
nn.Dropout(p=self.dropout_prob),
nn.Linear(256, sum(self.output_dims.values())),
)
self.metrics.add_metrics('all',
LF0_RMSE_Hz=metrics.LF0Distortion(),
VUV_accuracy=metrics.Mean(),
MCEP_distortion=metrics.MelCepDistortion(),
BAP_distortion=metrics.Distortion())
def __init__(self,
z_dim=16,
kld_weight=1.,
conditioning_dim=70 + 9,
output_dim=1 * 3,
dropout_prob=0.,
n_components=32,
pseudo_inputs_seq_lens=50,
pseudo_inputs_mean=0.05,
pseudo_inputs_std=0.01,
phone_set_file='data/unilex_phoneset.txt'):
if isinstance(pseudo_inputs_seq_lens, dict):
pseudo_inputs_seq_lens, n_components = self._make_seq_lens(
pseudo_inputs_seq_lens)
super(F0_VAMP_VAE,
self).__init__(z_dim=z_dim,
kld_weight=kld_weight,
input_dim=output_dim,
n_components=n_components,
pseudo_inputs_seq_lens=pseudo_inputs_seq_lens,
pseudo_inputs_mean=pseudo_inputs_mean,
pseudo_inputs_std=pseudo_inputs_std)
self.conditioning_dim = conditioning_dim
self.output_dim = output_dim
self.dropout_prob = dropout_prob
self.phone_set_file = phone_set_file
_encoder_shared_layer = utils.SequentialWithRecurrent(
nn.Linear(self.output_dim, 256), nn.Sigmoid(),
nn.Dropout(p=dropout_prob),
utils.RecurrentCuDNNWrapper(nn.GRU(256, 64, batch_first=True)),
nn.Dropout(p=dropout_prob),
utils.RecurrentCuDNNWrapper(nn.GRU(64, 64, batch_first=True)),
nn.Dropout(p=dropout_prob),
utils.RecurrentCuDNNWrapper(nn.GRU(64, 64, batch_first=True)),
nn.Dropout(p=dropout_prob), nn.Linear(64, 64), nn.Sigmoid(),
nn.Dropout(p=dropout_prob))
self.encoder_layer = vae.VAELayer(_encoder_shared_layer, 64,
self.z_dim)
self.decoder_layer = utils.SequentialWithRecurrent(
nn.Linear(self.conditioning_dim + self.z_dim, 256),
nn.Sigmoid(),
nn.Dropout(p=dropout_prob),
utils.RecurrentCuDNNWrapper(nn.GRU(256, 64, batch_first=True)),
nn.Dropout(p=dropout_prob),
utils.RecurrentCuDNNWrapper(nn.GRU(64, 64, batch_first=True)),
nn.Dropout(p=dropout_prob),
utils.RecurrentCuDNNWrapper(nn.GRU(64, 64, batch_first=True)),
nn.Dropout(p=dropout_prob),
nn.Linear(64, 64),
nn.Sigmoid(),
nn.Dropout(p=dropout_prob),
nn.Linear(64, self.output_dim),
)
self.metrics.add_metrics('all',
LF0_RMSE_Hz=LF0Distortion(),
embeddings=TensorHistory(self.z_dim,
hidden=True),
name=History(hidden=True),
n_segments=TensorHistory(1,
dtype=torch.long,
hidden=True),
segment_mean_F0=TensorHistory(1, hidden=True))