def __init__(
self,
# network structure related
idim: int,
odim: int,
adim: int = 384,
aheads: int = 4,
elayers: int = 6,
eunits: int = 1536,
dlayers: int = 6,
dunits: int = 1536,
postnet_layers: int = 5,
postnet_chans: int = 512,
postnet_filts: int = 5,
positionwise_layer_type: str = "conv1d",
positionwise_conv_kernel_size: int = 1,
use_scaled_pos_enc: bool = True,
use_batch_norm: bool = True,
encoder_normalize_before: bool = True,
decoder_normalize_before: bool = True,
encoder_concat_after: bool = False,
decoder_concat_after: bool = False,
duration_predictor_layers: int = 2,
duration_predictor_chans: int = 384,
duration_predictor_kernel_size: int = 3,
reduction_factor: int = 1,
encoder_type: str = "transformer",
decoder_type: str = "transformer",
# only for conformer
conformer_rel_pos_type: str = "legacy",
conformer_pos_enc_layer_type: str = "rel_pos",
conformer_self_attn_layer_type: str = "rel_selfattn",
conformer_activation_type: str = "swish",
use_macaron_style_in_conformer: bool = True,
use_cnn_in_conformer: bool = True,
conformer_enc_kernel_size: int = 7,
conformer_dec_kernel_size: int = 31,
zero_triu: bool = False,
# pretrained spk emb
spk_embed_dim: int = None,
spk_embed_integration_type: str = "add",
# GST
use_gst: bool = False,
gst_tokens: int = 10,
gst_heads: int = 4,
gst_conv_layers: int = 6,
gst_conv_chans_list: Sequence[int] = (32, 32, 64, 64, 128, 128),
gst_conv_kernel_size: int = 3,
gst_conv_stride: int = 2,
gst_gru_layers: int = 1,
gst_gru_units: int = 128,
# training related
transformer_enc_dropout_rate: float = 0.1,
transformer_enc_positional_dropout_rate: float = 0.1,
transformer_enc_attn_dropout_rate: float = 0.1,
transformer_dec_dropout_rate: float = 0.1,
transformer_dec_positional_dropout_rate: float = 0.1,
transformer_dec_attn_dropout_rate: float = 0.1,
duration_predictor_dropout_rate: float = 0.1,
postnet_dropout_rate: float = 0.5,
init_type: str = "xavier_uniform",
init_enc_alpha: float = 1.0,
init_dec_alpha: float = 1.0,
use_masking: bool = False,
use_weighted_masking: bool = False,
):
"""Initialize FastSpeech module."""
assert check_argument_types()
super().__init__()
# store hyperparameters
self.idim = idim
self.odim = odim
self.eos = idim - 1
self.reduction_factor = reduction_factor
self.encoder_type = encoder_type
self.decoder_type = decoder_type
self.use_scaled_pos_enc = use_scaled_pos_enc
self.use_gst = use_gst
self.spk_embed_dim = spk_embed_dim
if self.spk_embed_dim is not None:
self.spk_embed_integration_type = spk_embed_integration_type
# use idx 0 as padding idx
self.padding_idx = 0
# get positional encoding class
pos_enc_class = (ScaledPositionalEncoding
if self.use_scaled_pos_enc else PositionalEncoding)
# check relative positional encoding compatibility
if "conformer" in [encoder_type, decoder_type]:
if conformer_rel_pos_type == "legacy":
if conformer_pos_enc_layer_type == "rel_pos":
conformer_pos_enc_layer_type = "legacy_rel_pos"
logging.warning(
"Fallback to conformer_pos_enc_layer_type = 'legacy_rel_pos' "
"due to the compatibility. If you want to use the new one, "
"please use conformer_pos_enc_layer_type = 'latest'.")
if conformer_self_attn_layer_type == "rel_selfattn":
conformer_self_attn_layer_type = "legacy_rel_selfattn"
logging.warning(
"Fallback to "
"conformer_self_attn_layer_type = 'legacy_rel_selfattn' "
"due to the compatibility. If you want to use the new one, "
"please use conformer_pos_enc_layer_type = 'latest'.")
elif conformer_rel_pos_type == "latest":
assert conformer_pos_enc_layer_type != "legacy_rel_pos"
assert conformer_self_attn_layer_type != "legacy_rel_selfattn"
else:
raise ValueError(
f"Unknown rel_pos_type: {conformer_rel_pos_type}")
# define encoder
encoder_input_layer = torch.nn.Embedding(num_embeddings=idim,
embedding_dim=adim,
padding_idx=self.padding_idx)
if encoder_type == "transformer":
self.encoder = TransformerEncoder(
idim=idim,
attention_dim=adim,
attention_heads=aheads,
linear_units=eunits,
num_blocks=elayers,
input_layer=encoder_input_layer,
dropout_rate=transformer_enc_dropout_rate,
positional_dropout_rate=transformer_enc_positional_dropout_rate,
attention_dropout_rate=transformer_enc_attn_dropout_rate,
pos_enc_class=pos_enc_class,
normalize_before=encoder_normalize_before,
concat_after=encoder_concat_after,
positionwise_layer_type=positionwise_layer_type,
positionwise_conv_kernel_size=positionwise_conv_kernel_size,
)
elif encoder_type == "conformer":
self.encoder = ConformerEncoder(
idim=idim,
attention_dim=adim,
attention_heads=aheads,
linear_units=eunits,
num_blocks=elayers,
input_layer=encoder_input_layer,
dropout_rate=transformer_enc_dropout_rate,
positional_dropout_rate=transformer_enc_positional_dropout_rate,
attention_dropout_rate=transformer_enc_attn_dropout_rate,
normalize_before=encoder_normalize_before,
concat_after=encoder_concat_after,
positionwise_layer_type=positionwise_layer_type,
positionwise_conv_kernel_size=positionwise_conv_kernel_size,
macaron_style=use_macaron_style_in_conformer,
pos_enc_layer_type=conformer_pos_enc_layer_type,
selfattention_layer_type=conformer_self_attn_layer_type,
activation_type=conformer_activation_type,
use_cnn_module=use_cnn_in_conformer,
cnn_module_kernel=conformer_enc_kernel_size,
)
else:
raise ValueError(f"{encoder_type} is not supported.")
# define GST
if self.use_gst:
self.gst = StyleEncoder(
idim=odim, # the input is mel-spectrogram
gst_tokens=gst_tokens,
gst_token_dim=adim,
gst_heads=gst_heads,
conv_layers=gst_conv_layers,
conv_chans_list=gst_conv_chans_list,
conv_kernel_size=gst_conv_kernel_size,
conv_stride=gst_conv_stride,
gru_layers=gst_gru_layers,
gru_units=gst_gru_units,
)
# define additional projection for speaker embedding
if self.spk_embed_dim is not None:
if self.spk_embed_integration_type == "add":
self.projection = torch.nn.Linear(self.spk_embed_dim, adim)
else:
self.projection = torch.nn.Linear(adim + self.spk_embed_dim,
adim)
# define duration predictor
self.duration_predictor = DurationPredictor(
idim=adim,
n_layers=duration_predictor_layers,
n_chans=duration_predictor_chans,
kernel_size=duration_predictor_kernel_size,
dropout_rate=duration_predictor_dropout_rate,
)
# define length regulator
self.length_regulator = LengthRegulator()
# define decoder
# NOTE: we use encoder as decoder
# because fastspeech's decoder is the same as encoder
if decoder_type == "transformer":
self.decoder = TransformerEncoder(
idim=0,
attention_dim=adim,
attention_heads=aheads,
linear_units=dunits,
num_blocks=dlayers,
input_layer=None,
dropout_rate=transformer_dec_dropout_rate,
positional_dropout_rate=transformer_dec_positional_dropout_rate,
attention_dropout_rate=transformer_dec_attn_dropout_rate,
pos_enc_class=pos_enc_class,
normalize_before=decoder_normalize_before,
concat_after=decoder_concat_after,
positionwise_layer_type=positionwise_layer_type,
positionwise_conv_kernel_size=positionwise_conv_kernel_size,
)
elif decoder_type == "conformer":
self.decoder = ConformerEncoder(
idim=0,
attention_dim=adim,
attention_heads=aheads,
linear_units=dunits,
num_blocks=dlayers,
input_layer=None,
dropout_rate=transformer_dec_dropout_rate,
positional_dropout_rate=transformer_dec_positional_dropout_rate,
attention_dropout_rate=transformer_dec_attn_dropout_rate,
normalize_before=decoder_normalize_before,
concat_after=decoder_concat_after,
positionwise_layer_type=positionwise_layer_type,
positionwise_conv_kernel_size=positionwise_conv_kernel_size,
macaron_style=use_macaron_style_in_conformer,
pos_enc_layer_type=conformer_pos_enc_layer_type,
selfattention_layer_type=conformer_self_attn_layer_type,
activation_type=conformer_activation_type,
use_cnn_module=use_cnn_in_conformer,
cnn_module_kernel=conformer_dec_kernel_size,
)
else:
raise ValueError(f"{decoder_type} is not supported.")
# define final projection
self.feat_out = torch.nn.Linear(adim, odim * reduction_factor)
# define postnet
self.postnet = (None if postnet_layers == 0 else Postnet(
idim=idim,
odim=odim,
n_layers=postnet_layers,
n_chans=postnet_chans,
n_filts=postnet_filts,
use_batch_norm=use_batch_norm,
dropout_rate=postnet_dropout_rate,
))
# initialize parameters
self._reset_parameters(
init_type=init_type,
init_enc_alpha=init_enc_alpha,
init_dec_alpha=init_dec_alpha,
)
# define criterions
self.criterion = FastSpeechLoss(
use_masking=use_masking, use_weighted_masking=use_weighted_masking)
def __init__(
self,
idim: int,
odim: int,
adim: int = 256,
aheads: int = 2,
elayers: int = 4,
eunits: int = 1024,
dlayers: int = 4,
dunits: int = 1024,
positionwise_layer_type: str = "conv1d",
positionwise_conv_kernel_size: int = 1,
use_scaled_pos_enc: bool = True,
use_batch_norm: bool = True,
encoder_normalize_before: bool = True,
decoder_normalize_before: bool = True,
encoder_concat_after: bool = False,
decoder_concat_after: bool = False,
reduction_factor: int = 1,
encoder_type: str = "transformer",
decoder_type: str = "transformer",
transformer_enc_dropout_rate: float = 0.1,
transformer_enc_positional_dropout_rate: float = 0.1,
transformer_enc_attn_dropout_rate: float = 0.1,
transformer_dec_dropout_rate: float = 0.1,
transformer_dec_positional_dropout_rate: float = 0.1,
transformer_dec_attn_dropout_rate: float = 0.1,
# only for conformer
conformer_rel_pos_type: str = "legacy",
conformer_pos_enc_layer_type: str = "rel_pos",
conformer_self_attn_layer_type: str = "rel_selfattn",
conformer_activation_type: str = "swish",
use_macaron_style_in_conformer: bool = True,
use_cnn_in_conformer: bool = True,
zero_triu: bool = False,
conformer_enc_kernel_size: int = 7,
conformer_dec_kernel_size: int = 31,
# duration predictor
duration_predictor_layers: int = 2,
duration_predictor_chans: int = 384,
duration_predictor_kernel_size: int = 3,
duration_predictor_dropout_rate: float = 0.1,
# energy predictor
energy_predictor_layers: int = 2,
energy_predictor_chans: int = 384,
energy_predictor_kernel_size: int = 3,
energy_predictor_dropout: float = 0.5,
energy_embed_kernel_size: int = 9,
energy_embed_dropout: float = 0.5,
stop_gradient_from_energy_predictor: bool = False,
# pitch predictor
pitch_predictor_layers: int = 2,
pitch_predictor_chans: int = 384,
pitch_predictor_kernel_size: int = 3,
pitch_predictor_dropout: float = 0.5,
pitch_embed_kernel_size: int = 9,
pitch_embed_dropout: float = 0.5,
stop_gradient_from_pitch_predictor: bool = False,
# extra embedding related
spks: Optional[int] = None,
langs: Optional[int] = None,
spk_embed_dim: Optional[int] = None,
spk_embed_integration_type: str = "add",
use_gst: bool = False,
gst_tokens: int = 10,
gst_heads: int = 4,
gst_conv_layers: int = 6,
gst_conv_chans_list: Sequence[int] = (32, 32, 64, 64, 128, 128),
gst_conv_kernel_size: int = 3,
gst_conv_stride: int = 2,
gst_gru_layers: int = 1,
gst_gru_units: int = 128,
# training related
init_type: str = "xavier_uniform",
init_enc_alpha: float = 1.0,
init_dec_alpha: float = 1.0,
use_masking: bool = False,
use_weighted_masking: bool = False,
segment_size: int = 64,
# hifigan generator
generator_out_channels: int = 1,
generator_channels: int = 512,
generator_global_channels: int = -1,
generator_kernel_size: int = 7,
generator_upsample_scales: List[int] = [8, 8, 2, 2],
generator_upsample_kernel_sizes: List[int] = [16, 16, 4, 4],
generator_resblock_kernel_sizes: List[int] = [3, 7, 11],
generator_resblock_dilations: List[List[int]] = [
[1, 3, 5],
[1, 3, 5],
[1, 3, 5],
],
generator_use_additional_convs: bool = True,
generator_bias: bool = True,
generator_nonlinear_activation: str = "LeakyReLU",
generator_nonlinear_activation_params: Dict[str, Any] = {
"negative_slope": 0.1
},
generator_use_weight_norm: bool = True,
):
"""Initialize JETS generator module.
Args:
idim (int): Dimension of the inputs.
odim (int): Dimension of the outputs.
elayers (int): Number of encoder layers.
eunits (int): Number of encoder hidden units.
dlayers (int): Number of decoder layers.
dunits (int): Number of decoder hidden units.
use_scaled_pos_enc (bool): Whether to use trainable scaled pos encoding.
use_batch_norm (bool): Whether to use batch normalization in encoder prenet.
encoder_normalize_before (bool): Whether to apply layernorm layer before
encoder block.
decoder_normalize_before (bool): Whether to apply layernorm layer before
decoder block.
encoder_concat_after (bool): Whether to concatenate attention layer's input
and output in encoder.
decoder_concat_after (bool): Whether to concatenate attention layer's input
and output in decoder.
reduction_factor (int): Reduction factor.
encoder_type (str): Encoder type ("transformer" or "conformer").
decoder_type (str): Decoder type ("transformer" or "conformer").
transformer_enc_dropout_rate (float): Dropout rate in encoder except
attention and positional encoding.
transformer_enc_positional_dropout_rate (float): Dropout rate after encoder
positional encoding.
transformer_enc_attn_dropout_rate (float): Dropout rate in encoder
self-attention module.
transformer_dec_dropout_rate (float): Dropout rate in decoder except
attention & positional encoding.
transformer_dec_positional_dropout_rate (float): Dropout rate after decoder
positional encoding.
transformer_dec_attn_dropout_rate (float): Dropout rate in decoder
self-attention module.
conformer_rel_pos_type (str): Relative pos encoding type in conformer.
conformer_pos_enc_layer_type (str): Pos encoding layer type in conformer.
conformer_self_attn_layer_type (str): Self-attention layer type in conformer
conformer_activation_type (str): Activation function type in conformer.
use_macaron_style_in_conformer: Whether to use macaron style FFN.
use_cnn_in_conformer: Whether to use CNN in conformer.
zero_triu: Whether to use zero triu in relative self-attention module.
conformer_enc_kernel_size: Kernel size of encoder conformer.
conformer_dec_kernel_size: Kernel size of decoder conformer.
duration_predictor_layers (int): Number of duration predictor layers.
duration_predictor_chans (int): Number of duration predictor channels.
duration_predictor_kernel_size (int): Kernel size of duration predictor.
duration_predictor_dropout_rate (float): Dropout rate in duration predictor.
pitch_predictor_layers (int): Number of pitch predictor layers.
pitch_predictor_chans (int): Number of pitch predictor channels.
pitch_predictor_kernel_size (int): Kernel size of pitch predictor.
pitch_predictor_dropout_rate (float): Dropout rate in pitch predictor.
pitch_embed_kernel_size (float): Kernel size of pitch embedding.
pitch_embed_dropout_rate (float): Dropout rate for pitch embedding.
stop_gradient_from_pitch_predictor: Whether to stop gradient from pitch
predictor to encoder.
energy_predictor_layers (int): Number of energy predictor layers.
energy_predictor_chans (int): Number of energy predictor channels.
energy_predictor_kernel_size (int): Kernel size of energy predictor.
energy_predictor_dropout_rate (float): Dropout rate in energy predictor.
energy_embed_kernel_size (float): Kernel size of energy embedding.
energy_embed_dropout_rate (float): Dropout rate for energy embedding.
stop_gradient_from_energy_predictor: Whether to stop gradient from energy
predictor to encoder.
spks (Optional[int]): Number of speakers. If set to > 1, assume that the
sids will be provided as the input and use sid embedding layer.
langs (Optional[int]): Number of languages. If set to > 1, assume that the
lids will be provided as the input and use sid embedding layer.
spk_embed_dim (Optional[int]): Speaker embedding dimension. If set to > 0,
assume that spembs will be provided as the input.
spk_embed_integration_type: How to integrate speaker embedding.
use_gst (str): Whether to use global style token.
gst_tokens (int): The number of GST embeddings.
gst_heads (int): The number of heads in GST multihead attention.
gst_conv_layers (int): The number of conv layers in GST.
gst_conv_chans_list: (Sequence[int]):
List of the number of channels of conv layers in GST.
gst_conv_kernel_size (int): Kernel size of conv layers in GST.
gst_conv_stride (int): Stride size of conv layers in GST.
gst_gru_layers (int): The number of GRU layers in GST.
gst_gru_units (int): The number of GRU units in GST.
init_type (str): How to initialize transformer parameters.
init_enc_alpha (float): Initial value of alpha in scaled pos encoding of the
encoder.
init_dec_alpha (float): Initial value of alpha in scaled pos encoding of the
decoder.
use_masking (bool): Whether to apply masking for padded part in loss
calculation.
use_weighted_masking (bool): Whether to apply weighted masking in loss
calculation.
segment_size (int): Segment size for random windowed discriminator
generator_out_channels (int): Number of output channels.
generator_channels (int): Number of hidden representation channels.
generator_global_channels (int): Number of global conditioning channels.
generator_kernel_size (int): Kernel size of initial and final conv layer.
generator_upsample_scales (List[int]): List of upsampling scales.
generator_upsample_kernel_sizes (List[int]): List of kernel sizes for
upsample layers.
generator_resblock_kernel_sizes (List[int]): List of kernel sizes for
residual blocks.
generator_resblock_dilations (List[List[int]]): List of list of dilations
for residual blocks.
generator_use_additional_convs (bool): Whether to use additional conv layers
in residual blocks.
generator_bias (bool): Whether to add bias parameter in convolution layers.
generator_nonlinear_activation (str): Activation function module name.
generator_nonlinear_activation_params (Dict[str, Any]): Hyperparameters for
activation function.
generator_use_weight_norm (bool): Whether to use weight norm.
If set to true, it will be applied to all of the conv layers.
"""
super().__init__()
self.segment_size = segment_size
self.upsample_factor = int(np.prod(generator_upsample_scales))
self.idim = idim
self.odim = odim
self.reduction_factor = reduction_factor
self.encoder_type = encoder_type
self.decoder_type = decoder_type
self.stop_gradient_from_pitch_predictor = stop_gradient_from_pitch_predictor
self.stop_gradient_from_energy_predictor = stop_gradient_from_energy_predictor
self.use_scaled_pos_enc = use_scaled_pos_enc
self.use_gst = use_gst
# use idx 0 as padding idx
self.padding_idx = 0
# get positional encoding class
pos_enc_class = (ScaledPositionalEncoding
if self.use_scaled_pos_enc else PositionalEncoding)
# check relative positional encoding compatibility
if "conformer" in [encoder_type, decoder_type]:
if conformer_rel_pos_type == "legacy":
if conformer_pos_enc_layer_type == "rel_pos":
conformer_pos_enc_layer_type = "legacy_rel_pos"
logging.warning(
"Fallback to conformer_pos_enc_layer_type = 'legacy_rel_pos' "
"due to the compatibility. If you want to use the new one, "
"please use conformer_pos_enc_layer_type = 'latest'.")
if conformer_self_attn_layer_type == "rel_selfattn":
conformer_self_attn_layer_type = "legacy_rel_selfattn"
logging.warning(
"Fallback to "
"conformer_self_attn_layer_type = 'legacy_rel_selfattn' "
"due to the compatibility. If you want to use the new one, "
"please use conformer_pos_enc_layer_type = 'latest'.")
elif conformer_rel_pos_type == "latest":
assert conformer_pos_enc_layer_type != "legacy_rel_pos"
assert conformer_self_attn_layer_type != "legacy_rel_selfattn"
else:
raise ValueError(
f"Unknown rel_pos_type: {conformer_rel_pos_type}")
# define encoder
encoder_input_layer = torch.nn.Embedding(num_embeddings=idim,
embedding_dim=adim,
padding_idx=self.padding_idx)
if encoder_type == "transformer":
self.encoder = TransformerEncoder(
idim=idim,
attention_dim=adim,
attention_heads=aheads,
linear_units=eunits,
num_blocks=elayers,
input_layer=encoder_input_layer,
dropout_rate=transformer_enc_dropout_rate,
positional_dropout_rate=transformer_enc_positional_dropout_rate,
attention_dropout_rate=transformer_enc_attn_dropout_rate,
pos_enc_class=pos_enc_class,
normalize_before=encoder_normalize_before,
concat_after=encoder_concat_after,
positionwise_layer_type=positionwise_layer_type,
positionwise_conv_kernel_size=positionwise_conv_kernel_size,
)
elif encoder_type == "conformer":
self.encoder = ConformerEncoder(
idim=idim,
attention_dim=adim,
attention_heads=aheads,
linear_units=eunits,
num_blocks=elayers,
input_layer=encoder_input_layer,
dropout_rate=transformer_enc_dropout_rate,
positional_dropout_rate=transformer_enc_positional_dropout_rate,
attention_dropout_rate=transformer_enc_attn_dropout_rate,
normalize_before=encoder_normalize_before,
concat_after=encoder_concat_after,
positionwise_layer_type=positionwise_layer_type,
positionwise_conv_kernel_size=positionwise_conv_kernel_size,
macaron_style=use_macaron_style_in_conformer,
pos_enc_layer_type=conformer_pos_enc_layer_type,
selfattention_layer_type=conformer_self_attn_layer_type,
activation_type=conformer_activation_type,
use_cnn_module=use_cnn_in_conformer,
cnn_module_kernel=conformer_enc_kernel_size,
zero_triu=zero_triu,
)
else:
raise ValueError(f"{encoder_type} is not supported.")
# define GST
if self.use_gst:
self.gst = StyleEncoder(
idim=odim, # the input is mel-spectrogram
gst_tokens=gst_tokens,
gst_token_dim=adim,
gst_heads=gst_heads,
conv_layers=gst_conv_layers,
conv_chans_list=gst_conv_chans_list,
conv_kernel_size=gst_conv_kernel_size,
conv_stride=gst_conv_stride,
gru_layers=gst_gru_layers,
gru_units=gst_gru_units,
)
# define spk and lang embedding
self.spks = None
if spks is not None and spks > 1:
self.spks = spks
self.sid_emb = torch.nn.Embedding(spks, adim)
self.langs = None
if langs is not None and langs > 1:
self.langs = langs
self.lid_emb = torch.nn.Embedding(langs, adim)
# define additional projection for speaker embedding
self.spk_embed_dim = None
if spk_embed_dim is not None and spk_embed_dim > 0:
self.spk_embed_dim = spk_embed_dim
self.spk_embed_integration_type = spk_embed_integration_type
if self.spk_embed_dim is not None:
if self.spk_embed_integration_type == "add":
self.projection = torch.nn.Linear(self.spk_embed_dim, adim)
else:
self.projection = torch.nn.Linear(adim + self.spk_embed_dim,
adim)
# define duration predictor
self.duration_predictor = DurationPredictor(
idim=adim,
n_layers=duration_predictor_layers,
n_chans=duration_predictor_chans,
kernel_size=duration_predictor_kernel_size,
dropout_rate=duration_predictor_dropout_rate,
)
# define pitch predictor
self.pitch_predictor = VariancePredictor(
idim=adim,
n_layers=pitch_predictor_layers,
n_chans=pitch_predictor_chans,
kernel_size=pitch_predictor_kernel_size,
dropout_rate=pitch_predictor_dropout,
)
# NOTE(kan-bayashi): We use continuous pitch + FastPitch style avg
self.pitch_embed = torch.nn.Sequential(
torch.nn.Conv1d(
in_channels=1,
out_channels=adim,
kernel_size=pitch_embed_kernel_size,
padding=(pitch_embed_kernel_size - 1) // 2,
),
torch.nn.Dropout(pitch_embed_dropout),
)
# define energy predictor
self.energy_predictor = VariancePredictor(
idim=adim,
n_layers=energy_predictor_layers,
n_chans=energy_predictor_chans,
kernel_size=energy_predictor_kernel_size,
dropout_rate=energy_predictor_dropout,
)
# NOTE(kan-bayashi): We use continuous enegy + FastPitch style avg
self.energy_embed = torch.nn.Sequential(
torch.nn.Conv1d(
in_channels=1,
out_channels=adim,
kernel_size=energy_embed_kernel_size,
padding=(energy_embed_kernel_size - 1) // 2,
),
torch.nn.Dropout(energy_embed_dropout),
)
# define AlignmentModule
self.alignment_module = AlignmentModule(adim, odim)
# define length regulator
self.length_regulator = GaussianUpsampling()
# define decoder
# NOTE: we use encoder as decoder
# because fastspeech's decoder is the same as encoder
if decoder_type == "transformer":
self.decoder = TransformerEncoder(
idim=0,
attention_dim=adim,
attention_heads=aheads,
linear_units=dunits,
num_blocks=dlayers,
input_layer=None,
dropout_rate=transformer_dec_dropout_rate,
positional_dropout_rate=transformer_dec_positional_dropout_rate,
attention_dropout_rate=transformer_dec_attn_dropout_rate,
pos_enc_class=pos_enc_class,
normalize_before=decoder_normalize_before,
concat_after=decoder_concat_after,
positionwise_layer_type=positionwise_layer_type,
positionwise_conv_kernel_size=positionwise_conv_kernel_size,
)
elif decoder_type == "conformer":
self.decoder = ConformerEncoder(
idim=0,
attention_dim=adim,
attention_heads=aheads,
linear_units=dunits,
num_blocks=dlayers,
input_layer=None,
dropout_rate=transformer_dec_dropout_rate,
positional_dropout_rate=transformer_dec_positional_dropout_rate,
attention_dropout_rate=transformer_dec_attn_dropout_rate,
normalize_before=decoder_normalize_before,
concat_after=decoder_concat_after,
positionwise_layer_type=positionwise_layer_type,
positionwise_conv_kernel_size=positionwise_conv_kernel_size,
macaron_style=use_macaron_style_in_conformer,
pos_enc_layer_type=conformer_pos_enc_layer_type,
selfattention_layer_type=conformer_self_attn_layer_type,
activation_type=conformer_activation_type,
use_cnn_module=use_cnn_in_conformer,
cnn_module_kernel=conformer_dec_kernel_size,
)
else:
raise ValueError(f"{decoder_type} is not supported.")
# define hifigan generator
self.generator = HiFiGANGenerator(
in_channels=adim,
out_channels=generator_out_channels,
channels=generator_channels,
global_channels=generator_global_channels,
kernel_size=generator_kernel_size,
upsample_scales=generator_upsample_scales,
upsample_kernel_sizes=generator_upsample_kernel_sizes,
resblock_kernel_sizes=generator_resblock_kernel_sizes,
resblock_dilations=generator_resblock_dilations,
use_additional_convs=generator_use_additional_convs,
bias=generator_bias,
nonlinear_activation=generator_nonlinear_activation,
nonlinear_activation_params=generator_nonlinear_activation_params,
use_weight_norm=generator_use_weight_norm,
)
# initialize parameters
self._reset_parameters(
init_type=init_type,
init_enc_alpha=init_enc_alpha,
init_dec_alpha=init_dec_alpha,
)
def __init__(
self,
# network structure related
idim: int,
odim: int,
adim: int = 384,
aheads: int = 4,
elayers: int = 6,
eunits: int = 1536,
dlayers: int = 6,
dunits: int = 1536,
postnet_layers: int = 5,
postnet_chans: int = 512,
postnet_filts: int = 5,
postnet_dropout_rate: float = 0.5,
positionwise_layer_type: str = "conv1d",
positionwise_conv_kernel_size: int = 1,
use_scaled_pos_enc: bool = True,
use_batch_norm: bool = True,
encoder_normalize_before: bool = True,
decoder_normalize_before: bool = True,
encoder_concat_after: bool = False,
decoder_concat_after: bool = False,
duration_predictor_layers: int = 2,
duration_predictor_chans: int = 384,
duration_predictor_kernel_size: int = 3,
duration_predictor_dropout_rate: float = 0.1,
reduction_factor: int = 1,
encoder_type: str = "transformer",
decoder_type: str = "transformer",
transformer_enc_dropout_rate: float = 0.1,
transformer_enc_positional_dropout_rate: float = 0.1,
transformer_enc_attn_dropout_rate: float = 0.1,
transformer_dec_dropout_rate: float = 0.1,
transformer_dec_positional_dropout_rate: float = 0.1,
transformer_dec_attn_dropout_rate: float = 0.1,
# only for conformer
conformer_rel_pos_type: str = "legacy",
conformer_pos_enc_layer_type: str = "rel_pos",
conformer_self_attn_layer_type: str = "rel_selfattn",
conformer_activation_type: str = "swish",
use_macaron_style_in_conformer: bool = True,
use_cnn_in_conformer: bool = True,
conformer_enc_kernel_size: int = 7,
conformer_dec_kernel_size: int = 31,
zero_triu: bool = False,
# extra embedding related
spks: Optional[int] = None,
langs: Optional[int] = None,
spk_embed_dim: Optional[int] = None,
spk_embed_integration_type: str = "add",
use_gst: bool = False,
gst_tokens: int = 10,
gst_heads: int = 4,
gst_conv_layers: int = 6,
gst_conv_chans_list: Sequence[int] = (32, 32, 64, 64, 128, 128),
gst_conv_kernel_size: int = 3,
gst_conv_stride: int = 2,
gst_gru_layers: int = 1,
gst_gru_units: int = 128,
# training related
init_type: str = "xavier_uniform",
init_enc_alpha: float = 1.0,
init_dec_alpha: float = 1.0,
use_masking: bool = False,
use_weighted_masking: bool = False,
):
"""Initialize FastSpeech module.
Args:
idim (int): Dimension of the inputs.
odim (int): Dimension of the outputs.
elayers (int): Number of encoder layers.
eunits (int): Number of encoder hidden units.
dlayers (int): Number of decoder layers.
dunits (int): Number of decoder hidden units.
postnet_layers (int): Number of postnet layers.
postnet_chans (int): Number of postnet channels.
postnet_filts (int): Kernel size of postnet.
postnet_dropout_rate (float): Dropout rate in postnet.
use_scaled_pos_enc (bool): Whether to use trainable scaled pos encoding.
use_batch_norm (bool): Whether to use batch normalization in encoder prenet.
encoder_normalize_before (bool): Whether to apply layernorm layer before
encoder block.
decoder_normalize_before (bool): Whether to apply layernorm layer before
decoder block.
encoder_concat_after (bool): Whether to concatenate attention layer's input
and output in encoder.
decoder_concat_after (bool): Whether to concatenate attention layer's input
and output in decoder.
duration_predictor_layers (int): Number of duration predictor layers.
duration_predictor_chans (int): Number of duration predictor channels.
duration_predictor_kernel_size (int): Kernel size of duration predictor.
duration_predictor_dropout_rate (float): Dropout rate in duration predictor.
reduction_factor (int): Reduction factor.
encoder_type (str): Encoder type ("transformer" or "conformer").
decoder_type (str): Decoder type ("transformer" or "conformer").
transformer_enc_dropout_rate (float): Dropout rate in encoder except
attention and positional encoding.
transformer_enc_positional_dropout_rate (float): Dropout rate after encoder
positional encoding.
transformer_enc_attn_dropout_rate (float): Dropout rate in encoder
self-attention module.
transformer_dec_dropout_rate (float): Dropout rate in decoder except
attention & positional encoding.
transformer_dec_positional_dropout_rate (float): Dropout rate after decoder
positional encoding.
transformer_dec_attn_dropout_rate (float): Dropout rate in decoder
self-attention module.
conformer_rel_pos_type (str): Relative pos encoding type in conformer.
conformer_pos_enc_layer_type (str): Pos encoding layer type in conformer.
conformer_self_attn_layer_type (str): Self-attention layer type in conformer
conformer_activation_type (str): Activation function type in conformer.
use_macaron_style_in_conformer: Whether to use macaron style FFN.
use_cnn_in_conformer: Whether to use CNN in conformer.
conformer_enc_kernel_size: Kernel size of encoder conformer.
conformer_dec_kernel_size: Kernel size of decoder conformer.
zero_triu: Whether to use zero triu in relative self-attention module.
spks (Optional[int]): Number of speakers. If set to > 1, assume that the
sids will be provided as the input and use sid embedding layer.
langs (Optional[int]): Number of languages. If set to > 1, assume that the
lids will be provided as the input and use sid embedding layer.
spk_embed_dim (Optional[int]): Speaker embedding dimension. If set to > 0,
assume that spembs will be provided as the input.
spk_embed_integration_type: How to integrate speaker embedding.
use_gst (str): Whether to use global style token.
gst_tokens (int): The number of GST embeddings.
gst_heads (int): The number of heads in GST multihead attention.
gst_conv_layers (int): The number of conv layers in GST.
gst_conv_chans_list: (Sequence[int]):
List of the number of channels of conv layers in GST.
gst_conv_kernel_size (int): Kernel size of conv layers in GST.
gst_conv_stride (int): Stride size of conv layers in GST.
gst_gru_layers (int): The number of GRU layers in GST.
gst_gru_units (int): The number of GRU units in GST.
init_type (str): How to initialize transformer parameters.
init_enc_alpha (float): Initial value of alpha in scaled pos encoding of the
encoder.
init_dec_alpha (float): Initial value of alpha in scaled pos encoding of the
decoder.
use_masking (bool): Whether to apply masking for padded part in loss
calculation.
use_weighted_masking (bool): Whether to apply weighted masking in loss
calculation.
"""
assert check_argument_types()
super().__init__()
# store hyperparameters
self.idim = idim
self.odim = odim
self.eos = idim - 1
self.reduction_factor = reduction_factor
self.encoder_type = encoder_type
self.decoder_type = decoder_type
self.use_scaled_pos_enc = use_scaled_pos_enc
self.use_gst = use_gst
# use idx 0 as padding idx
self.padding_idx = 0
# get positional encoding class
pos_enc_class = (ScaledPositionalEncoding
if self.use_scaled_pos_enc else PositionalEncoding)
# check relative positional encoding compatibility
if "conformer" in [encoder_type, decoder_type]:
if conformer_rel_pos_type == "legacy":
if conformer_pos_enc_layer_type == "rel_pos":
conformer_pos_enc_layer_type = "legacy_rel_pos"
logging.warning(
"Fallback to conformer_pos_enc_layer_type = 'legacy_rel_pos' "
"due to the compatibility. If you want to use the new one, "
"please use conformer_pos_enc_layer_type = 'latest'.")
if conformer_self_attn_layer_type == "rel_selfattn":
conformer_self_attn_layer_type = "legacy_rel_selfattn"
logging.warning(
"Fallback to "
"conformer_self_attn_layer_type = 'legacy_rel_selfattn' "
"due to the compatibility. If you want to use the new one, "
"please use conformer_pos_enc_layer_type = 'latest'.")
elif conformer_rel_pos_type == "latest":
assert conformer_pos_enc_layer_type != "legacy_rel_pos"
assert conformer_self_attn_layer_type != "legacy_rel_selfattn"
else:
raise ValueError(
f"Unknown rel_pos_type: {conformer_rel_pos_type}")
# define encoder
encoder_input_layer = torch.nn.Embedding(num_embeddings=idim,
embedding_dim=adim,
padding_idx=self.padding_idx)
if encoder_type == "transformer":
self.encoder = TransformerEncoder(
idim=idim,
attention_dim=adim,
attention_heads=aheads,
linear_units=eunits,
num_blocks=elayers,
input_layer=encoder_input_layer,
dropout_rate=transformer_enc_dropout_rate,
positional_dropout_rate=transformer_enc_positional_dropout_rate,
attention_dropout_rate=transformer_enc_attn_dropout_rate,
pos_enc_class=pos_enc_class,
normalize_before=encoder_normalize_before,
concat_after=encoder_concat_after,
positionwise_layer_type=positionwise_layer_type,
positionwise_conv_kernel_size=positionwise_conv_kernel_size,
)
elif encoder_type == "conformer":
self.encoder = ConformerEncoder(
idim=idim,
attention_dim=adim,
attention_heads=aheads,
linear_units=eunits,
num_blocks=elayers,
input_layer=encoder_input_layer,
dropout_rate=transformer_enc_dropout_rate,
positional_dropout_rate=transformer_enc_positional_dropout_rate,
attention_dropout_rate=transformer_enc_attn_dropout_rate,
normalize_before=encoder_normalize_before,
concat_after=encoder_concat_after,
positionwise_layer_type=positionwise_layer_type,
positionwise_conv_kernel_size=positionwise_conv_kernel_size,
macaron_style=use_macaron_style_in_conformer,
pos_enc_layer_type=conformer_pos_enc_layer_type,
selfattention_layer_type=conformer_self_attn_layer_type,
activation_type=conformer_activation_type,
use_cnn_module=use_cnn_in_conformer,
cnn_module_kernel=conformer_enc_kernel_size,
)
else:
raise ValueError(f"{encoder_type} is not supported.")
# define GST
if self.use_gst:
self.gst = StyleEncoder(
idim=odim, # the input is mel-spectrogram
gst_tokens=gst_tokens,
gst_token_dim=adim,
gst_heads=gst_heads,
conv_layers=gst_conv_layers,
conv_chans_list=gst_conv_chans_list,
conv_kernel_size=gst_conv_kernel_size,
conv_stride=gst_conv_stride,
gru_layers=gst_gru_layers,
gru_units=gst_gru_units,
)
# define spk and lang embedding
self.spks = None
if spks is not None and spks > 1:
self.spks = spks
self.sid_emb = torch.nn.Embedding(spks, adim)
self.langs = None
if langs is not None and langs > 1:
self.langs = langs
self.lid_emb = torch.nn.Embedding(langs, adim)
# define additional projection for speaker embedding
self.spk_embed_dim = None
if spk_embed_dim is not None and spk_embed_dim > 0:
self.spk_embed_dim = spk_embed_dim
self.spk_embed_integration_type = spk_embed_integration_type
if self.spk_embed_dim is not None:
if self.spk_embed_integration_type == "add":
self.projection = torch.nn.Linear(self.spk_embed_dim, adim)
else:
self.projection = torch.nn.Linear(adim + self.spk_embed_dim,
adim)
# define duration predictor
self.duration_predictor = DurationPredictor(
idim=adim,
n_layers=duration_predictor_layers,
n_chans=duration_predictor_chans,
kernel_size=duration_predictor_kernel_size,
dropout_rate=duration_predictor_dropout_rate,
)
# define length regulator
self.length_regulator = LengthRegulator()
# define decoder
# NOTE: we use encoder as decoder
# because fastspeech's decoder is the same as encoder
if decoder_type == "transformer":
self.decoder = TransformerEncoder(
idim=0,
attention_dim=adim,
attention_heads=aheads,
linear_units=dunits,
num_blocks=dlayers,
input_layer=None,
dropout_rate=transformer_dec_dropout_rate,
positional_dropout_rate=transformer_dec_positional_dropout_rate,
attention_dropout_rate=transformer_dec_attn_dropout_rate,
pos_enc_class=pos_enc_class,
normalize_before=decoder_normalize_before,
concat_after=decoder_concat_after,
positionwise_layer_type=positionwise_layer_type,
positionwise_conv_kernel_size=positionwise_conv_kernel_size,
)
elif decoder_type == "conformer":
self.decoder = ConformerEncoder(
idim=0,
attention_dim=adim,
attention_heads=aheads,
linear_units=dunits,
num_blocks=dlayers,
input_layer=None,
dropout_rate=transformer_dec_dropout_rate,
positional_dropout_rate=transformer_dec_positional_dropout_rate,
attention_dropout_rate=transformer_dec_attn_dropout_rate,
normalize_before=decoder_normalize_before,
concat_after=decoder_concat_after,
positionwise_layer_type=positionwise_layer_type,
positionwise_conv_kernel_size=positionwise_conv_kernel_size,
macaron_style=use_macaron_style_in_conformer,
pos_enc_layer_type=conformer_pos_enc_layer_type,
selfattention_layer_type=conformer_self_attn_layer_type,
activation_type=conformer_activation_type,
use_cnn_module=use_cnn_in_conformer,
cnn_module_kernel=conformer_dec_kernel_size,
)
else:
raise ValueError(f"{decoder_type} is not supported.")
# define final projection
self.feat_out = torch.nn.Linear(adim, odim * reduction_factor)
# define postnet
self.postnet = (None if postnet_layers == 0 else Postnet(
idim=idim,
odim=odim,
n_layers=postnet_layers,
n_chans=postnet_chans,
n_filts=postnet_filts,
use_batch_norm=use_batch_norm,
dropout_rate=postnet_dropout_rate,
))
# initialize parameters
self._reset_parameters(
init_type=init_type,
init_enc_alpha=init_enc_alpha,
init_dec_alpha=init_dec_alpha,
)
# define criterions
self.criterion = FastSpeechLoss(
use_masking=use_masking, use_weighted_masking=use_weighted_masking)
def __init__(
self,
# network structure related
idim: int,
odim: int,
adim: int = 384,
aheads: int = 4,
elayers: int = 6,
eunits: int = 1536,
dlayers: int = 6,
dunits: int = 1536,
postnet_layers: int = 5,
postnet_chans: int = 512,
postnet_filts: int = 5,
positionwise_layer_type: str = "conv1d",
positionwise_conv_kernel_size: int = 1,
use_scaled_pos_enc: bool = True,
use_batch_norm: bool = True,
encoder_normalize_before: bool = False,
decoder_normalize_before: bool = False,
encoder_concat_after: bool = False,
decoder_concat_after: bool = False,
reduction_factor: int = 1,
encoder_type: str = "transformer",
decoder_type: str = "transformer",
# only for conformer
conformer_pos_enc_layer_type: str = "rel_pos",
conformer_self_attn_layer_type: str = "rel_selfattn",
conformer_activation_type: str = "swish",
use_macaron_style_in_conformer: bool = True,
use_cnn_in_conformer: bool = True,
conformer_enc_kernel_size: int = 7,
conformer_dec_kernel_size: int = 31,
# duration predictor
duration_predictor_layers: int = 2,
duration_predictor_chans: int = 384,
duration_predictor_kernel_size: int = 3,
# energy predictor
energy_predictor_layers: int = 2,
energy_predictor_chans: int = 384,
energy_predictor_kernel_size: int = 3,
energy_predictor_dropout: float = 0.5,
energy_embed_kernel_size: int = 9,
energy_embed_dropout: float = 0.5,
stop_gradient_from_energy_predictor: bool = False,
# pitch predictor
pitch_predictor_layers: int = 2,
pitch_predictor_chans: int = 384,
pitch_predictor_kernel_size: int = 3,
pitch_predictor_dropout: float = 0.5,
pitch_embed_kernel_size: int = 9,
pitch_embed_dropout: float = 0.5,
stop_gradient_from_pitch_predictor: bool = False,
# pretrained spk emb
spk_embed_dim: int = None,
spk_embed_integration_type: str = "add",
# GST
use_gst: bool = False,
gst_tokens: int = 10,
gst_heads: int = 4,
gst_conv_layers: int = 6,
gst_conv_chans_list: Sequence[int] = (32, 32, 64, 64, 128, 128),
gst_conv_kernel_size: int = 3,
gst_conv_stride: int = 2,
gst_gru_layers: int = 1,
gst_gru_units: int = 128,
# training related
transformer_enc_dropout_rate: float = 0.1,
transformer_enc_positional_dropout_rate: float = 0.1,
transformer_enc_attn_dropout_rate: float = 0.1,
transformer_dec_dropout_rate: float = 0.1,
transformer_dec_positional_dropout_rate: float = 0.1,
transformer_dec_attn_dropout_rate: float = 0.1,
duration_predictor_dropout_rate: float = 0.1,
postnet_dropout_rate: float = 0.5,
init_type: str = "xavier_uniform",
init_enc_alpha: float = 1.0,
init_dec_alpha: float = 1.0,
use_masking: bool = False,
use_weighted_masking: bool = False,
):
"""Initialize FastSpeech2 module."""
assert check_argument_types()
super().__init__()
# store hyperparameters
self.idim = idim
self.odim = odim
self.eos = idim - 1
self.reduction_factor = reduction_factor
self.encoder_type = encoder_type
self.decoder_type = decoder_type
self.stop_gradient_from_pitch_predictor = stop_gradient_from_pitch_predictor
self.stop_gradient_from_energy_predictor = stop_gradient_from_energy_predictor
self.use_scaled_pos_enc = use_scaled_pos_enc
self.use_gst = use_gst
self.spk_embed_dim = spk_embed_dim
if self.spk_embed_dim is not None:
self.spk_embed_integration_type = spk_embed_integration_type
# use idx 0 as padding idx
self.padding_idx = 0
# get positional encoding class
pos_enc_class = (ScaledPositionalEncoding
if self.use_scaled_pos_enc else PositionalEncoding)
# define encoder
encoder_input_layer = torch.nn.Embedding(num_embeddings=idim,
embedding_dim=adim,
padding_idx=self.padding_idx)
if encoder_type == "transformer":
self.encoder = TransformerEncoder(
idim=idim,
attention_dim=adim,
attention_heads=aheads,
linear_units=eunits,
num_blocks=elayers,
input_layer=encoder_input_layer,
dropout_rate=transformer_enc_dropout_rate,
positional_dropout_rate=transformer_enc_positional_dropout_rate,
attention_dropout_rate=transformer_enc_attn_dropout_rate,
pos_enc_class=pos_enc_class,
normalize_before=encoder_normalize_before,
concat_after=encoder_concat_after,
positionwise_layer_type=positionwise_layer_type,
positionwise_conv_kernel_size=positionwise_conv_kernel_size,
)
elif encoder_type == "conformer":
self.encoder = ConformerEncoder(
idim=idim,
attention_dim=adim,
attention_heads=aheads,
linear_units=eunits,
num_blocks=elayers,
input_layer=encoder_input_layer,
dropout_rate=transformer_enc_dropout_rate,
positional_dropout_rate=transformer_enc_positional_dropout_rate,
attention_dropout_rate=transformer_enc_attn_dropout_rate,
normalize_before=encoder_normalize_before,
concat_after=encoder_concat_after,
positionwise_layer_type=positionwise_layer_type,
positionwise_conv_kernel_size=positionwise_conv_kernel_size,
macaron_style=use_macaron_style_in_conformer,
pos_enc_layer_type=conformer_pos_enc_layer_type,
selfattention_layer_type=conformer_self_attn_layer_type,
activation_type=conformer_activation_type,
use_cnn_module=use_cnn_in_conformer,
cnn_module_kernel=conformer_enc_kernel_size,
)
else:
raise ValueError(f"{encoder_type} is not supported.")
# define GST
if self.use_gst:
self.gst = StyleEncoder(
idim=odim, # the input is mel-spectrogram
gst_tokens=gst_tokens,
gst_token_dim=adim,
gst_heads=gst_heads,
conv_layers=gst_conv_layers,
conv_chans_list=gst_conv_chans_list,
conv_kernel_size=gst_conv_kernel_size,
conv_stride=gst_conv_stride,
gru_layers=gst_gru_layers,
gru_units=gst_gru_units,
)
# define additional projection for speaker embedding
if self.spk_embed_dim is not None:
if self.spk_embed_integration_type == "add":
self.projection = torch.nn.Linear(self.spk_embed_dim, adim)
else:
self.projection = torch.nn.Linear(adim + self.spk_embed_dim,
adim)
# define duration predictor
self.duration_predictor = DurationPredictor(
idim=adim,
n_layers=duration_predictor_layers,
n_chans=duration_predictor_chans,
kernel_size=duration_predictor_kernel_size,
dropout_rate=duration_predictor_dropout_rate,
)
# define pitch predictor
self.pitch_predictor = VariancePredictor(
idim=adim,
n_layers=pitch_predictor_layers,
n_chans=pitch_predictor_chans,
kernel_size=pitch_predictor_kernel_size,
dropout_rate=pitch_predictor_dropout,
)
# NOTE(kan-bayashi): We use continuous pitch + FastPitch style avg
self.pitch_embed = torch.nn.Sequential(
torch.nn.Conv1d(
in_channels=1,
out_channels=adim,
kernel_size=pitch_embed_kernel_size,
padding=(pitch_embed_kernel_size - 1) // 2,
),
torch.nn.Dropout(pitch_embed_dropout),
)
# define energy predictor
self.energy_predictor = VariancePredictor(
idim=adim,
n_layers=energy_predictor_layers,
n_chans=energy_predictor_chans,
kernel_size=energy_predictor_kernel_size,
dropout_rate=energy_predictor_dropout,
)
# NOTE(kan-bayashi): We use continuous enegy + FastPitch style avg
self.energy_embed = torch.nn.Sequential(
torch.nn.Conv1d(
in_channels=1,
out_channels=adim,
kernel_size=energy_embed_kernel_size,
padding=(energy_embed_kernel_size - 1) // 2,
),
torch.nn.Dropout(energy_embed_dropout),
)
# define length regulator
self.length_regulator = LengthRegulator()
# define decoder
# NOTE: we use encoder as decoder
# because fastspeech's decoder is the same as encoder
if decoder_type == "transformer":
self.decoder = TransformerEncoder(
idim=0,
attention_dim=adim,
attention_heads=aheads,
linear_units=dunits,
num_blocks=dlayers,
input_layer=None,
dropout_rate=transformer_dec_dropout_rate,
positional_dropout_rate=transformer_dec_positional_dropout_rate,
attention_dropout_rate=transformer_dec_attn_dropout_rate,
pos_enc_class=pos_enc_class,
normalize_before=decoder_normalize_before,
concat_after=decoder_concat_after,
positionwise_layer_type=positionwise_layer_type,
positionwise_conv_kernel_size=positionwise_conv_kernel_size,
)
elif decoder_type == "conformer":
self.decoder = ConformerEncoder(
idim=0,
attention_dim=adim,
attention_heads=aheads,
linear_units=dunits,
num_blocks=dlayers,
input_layer=None,
dropout_rate=transformer_dec_dropout_rate,
positional_dropout_rate=transformer_dec_positional_dropout_rate,
attention_dropout_rate=transformer_dec_attn_dropout_rate,
normalize_before=decoder_normalize_before,
concat_after=decoder_concat_after,
positionwise_layer_type=positionwise_layer_type,
positionwise_conv_kernel_size=positionwise_conv_kernel_size,
macaron_style=use_macaron_style_in_conformer,
pos_enc_layer_type=conformer_pos_enc_layer_type,
selfattention_layer_type=conformer_self_attn_layer_type,
activation_type=conformer_activation_type,
use_cnn_module=use_cnn_in_conformer,
cnn_module_kernel=conformer_dec_kernel_size,
)
else:
raise ValueError(f"{decoder_type} is not supported.")
# define final projection
self.feat_out = torch.nn.Linear(adim, odim * reduction_factor)
# define postnet
self.postnet = (None if postnet_layers == 0 else Postnet(
idim=idim,
odim=odim,
n_layers=postnet_layers,
n_chans=postnet_chans,
n_filts=postnet_filts,
use_batch_norm=use_batch_norm,
dropout_rate=postnet_dropout_rate,
))
# initialize parameters
self._reset_parameters(
init_type=init_type,
init_enc_alpha=init_enc_alpha,
init_dec_alpha=init_dec_alpha,
)
# define criterions
self.criterion = FastSpeech2Loss(
use_masking=use_masking, use_weighted_masking=use_weighted_masking)
def __init__(
self,
input_dim: int,
num_spk: int = 2,
adim: int = 384,
aheads: int = 4,
layers: int = 6,
linear_units: int = 1536,
positionwise_layer_type: str = "linear",
positionwise_conv_kernel_size: int = 1,
normalize_before: bool = False,
concat_after: bool = False,
dropout_rate: float = 0.1,
positional_dropout_rate: float = 0.1,
attention_dropout_rate: float = 0.1,
use_scaled_pos_enc: bool = True,
nonlinear: str = "relu",
):
"""Transformer separator.
Args:
input_dim: input feature dimension
num_spk: number of speakers
adim (int): Dimension of attention.
aheads (int): The number of heads of multi head attention.
linear_units (int): The number of units of position-wise feed forward.
layers (int): The number of transformer blocks.
dropout_rate (float): Dropout rate.
attention_dropout_rate (float): Dropout rate in attention.
positional_dropout_rate (float): Dropout rate after adding
positional encoding.
normalize_before (bool): Whether to use layer_norm before the first block.
concat_after (bool): Whether to concat attention layer's input and output.
if True, additional linear will be applied.
i.e. x -> x + linear(concat(x, att(x)))
if False, no additional linear will be applied. i.e. x -> x + att(x)
positionwise_layer_type (str): "linear", "conv1d", or "conv1d-linear".
positionwise_conv_kernel_size (int): Kernel size of
positionwise conv1d layer.
use_scaled_pos_enc (bool) : use scaled positional encoding or not
nonlinear: the nonlinear function for mask estimation,
select from 'relu', 'tanh', 'sigmoid'
"""
super().__init__()
self._num_spk = num_spk
pos_enc_class = (ScaledPositionalEncoding
if use_scaled_pos_enc else PositionalEncoding)
self.transformer = TransformerEncoder(
idim=input_dim,
attention_dim=adim,
attention_heads=aheads,
linear_units=linear_units,
num_blocks=layers,
input_layer="linear",
dropout_rate=dropout_rate,
positional_dropout_rate=positional_dropout_rate,
attention_dropout_rate=attention_dropout_rate,
pos_enc_class=pos_enc_class,
normalize_before=normalize_before,
concat_after=concat_after,
positionwise_layer_type=positionwise_layer_type,
positionwise_conv_kernel_size=positionwise_conv_kernel_size,
)
self.linear = torch.nn.ModuleList(
[torch.nn.Linear(adim, input_dim) for _ in range(self.num_spk)])
if nonlinear not in ("sigmoid", "relu", "tanh"):
raise ValueError("Not supporting nonlinear={}".format(nonlinear))
self.nonlinear = {
"sigmoid": torch.nn.Sigmoid(),
"relu": torch.nn.ReLU(),
"tanh": torch.nn.Tanh(),
}[nonlinear]
def __init__(
self,
idim,
odim,
adim: int = 192,
aheads: int = 2,
elayers: int = 4,
eunits: int = 768,
dlayers: int = 2,
dunits: int = 768,
use_scaled_pos_enc: bool = True,
positionwise_layer_type: str = "conv1d",
positionwise_conv_kernel_size: int = 1,
encoder_normalize_before: bool = True,
decoder_normalize_before: bool = True,
encoder_concat_after: bool = False,
decoder_concat_after: bool = False,
duration_predictor_layers: int = 2,
duration_predictor_chans: int = 192,
duration_predictor_kernel_size: int = 3,
duration_predictor_dropout_rate: float = 0.1,
transformer_enc_dropout_rate: float = 0.1,
transformer_enc_positional_dropout_rate: float = 0.1,
transformer_enc_attn_dropout_rate: float = 0.1,
transformer_dec_dropout_rate: float = 0.1,
transformer_dec_positional_dropout_rate: float = 0.1,
transformer_dec_attn_dropout_rate: float = 0.1,
init_type: str = "xavier_uniform",
init_enc_alpha: float = 1.0,
init_dec_alpha: float = 1.0,
use_masking: bool = False,
use_weighted_masking: bool = False,
ddim: int = 64,
beta_min: float = 0.05,
beta_max: float = 20.0,
pe_scale: int = 1000,
# energy predictor
energy_predictor_layers: int = 2,
energy_predictor_chans: int = 384,
energy_predictor_kernel_size: int = 3,
energy_predictor_dropout: float = 0.5,
energy_embed_kernel_size: int = 9,
energy_embed_dropout: float = 0.5,
stop_gradient_from_energy_predictor: bool = False,
# pitch predictor
pitch_predictor_layers: int = 2,
pitch_predictor_chans: int = 384,
pitch_predictor_kernel_size: int = 3,
pitch_predictor_dropout: float = 0.5,
pitch_embed_kernel_size: int = 9,
pitch_embed_dropout: float = 0.5,
stop_gradient_from_pitch_predictor: bool = False,
):
super(GradTTS, self).__init__()
self.idim = idim
self.odim = odim
self.eos = idim - 1
self.stop_gradient_from_pitch_predictor = stop_gradient_from_pitch_predictor
self.stop_gradient_from_energy_predictor = stop_gradient_from_energy_predictor
self.padding_idx = 0
self.use_scaled_pos_enc = use_scaled_pos_enc
encoder_input_layer = torch.nn.Embedding(num_embeddings=idim,
embedding_dim=adim,
padding_idx=self.padding_idx)
pos_enc_class = (ScaledPositionalEncoding
if self.use_scaled_pos_enc else PositionalEncoding)
self.encoder = TransformerEncoder(
idim=adim,
attention_dim=adim,
attention_heads=aheads,
linear_units=eunits,
num_blocks=elayers,
input_layer=encoder_input_layer,
dropout_rate=transformer_enc_dropout_rate,
positional_dropout_rate=transformer_enc_positional_dropout_rate,
attention_dropout_rate=transformer_enc_attn_dropout_rate,
pos_enc_class=pos_enc_class,
normalize_before=encoder_normalize_before,
concat_after=encoder_concat_after,
positionwise_layer_type=positionwise_layer_type,
positionwise_conv_kernel_size=positionwise_conv_kernel_size,
)
self.duration_predictor = DurationPredictor(
idim=adim,
n_layers=duration_predictor_layers,
n_chans=duration_predictor_chans,
kernel_size=duration_predictor_kernel_size,
dropout_rate=duration_predictor_dropout_rate,
)
self.pitch_predictor = VariancePredictor(
idim=adim,
n_layers=pitch_predictor_layers,
n_chans=pitch_predictor_chans,
kernel_size=pitch_predictor_kernel_size,
dropout_rate=pitch_predictor_dropout,
)
# NOTE(kan-bayashi): We use continuous pitch + FastPitch style avg
self.pitch_embed = torch.nn.Sequential(
torch.nn.Conv1d(
in_channels=1,
out_channels=adim,
kernel_size=pitch_embed_kernel_size,
padding=(pitch_embed_kernel_size - 1) // 2,
),
torch.nn.Dropout(pitch_embed_dropout),
)
# define energy predictor
self.energy_predictor = VariancePredictor(
idim=adim,
n_layers=energy_predictor_layers,
n_chans=energy_predictor_chans,
kernel_size=energy_predictor_kernel_size,
dropout_rate=energy_predictor_dropout,
)
# NOTE(kan-bayashi): We use continuous enegy + FastPitch style avg
self.energy_embed = torch.nn.Sequential(
torch.nn.Conv1d(
in_channels=1,
out_channels=adim,
kernel_size=energy_embed_kernel_size,
padding=(energy_embed_kernel_size - 1) // 2,
),
torch.nn.Dropout(energy_embed_dropout),
)
self.length_regulator = LengthRegulator()
self.pre_decoder = TransformerEncoder(
idim=0,
attention_dim=adim,
attention_heads=aheads,
linear_units=dunits,
num_blocks=dlayers,
input_layer=None,
dropout_rate=transformer_dec_dropout_rate,
positional_dropout_rate=transformer_dec_positional_dropout_rate,
attention_dropout_rate=transformer_dec_attn_dropout_rate,
pos_enc_class=pos_enc_class,
normalize_before=decoder_normalize_before,
concat_after=decoder_concat_after,
positionwise_layer_type=positionwise_layer_type,
positionwise_conv_kernel_size=positionwise_conv_kernel_size,
)
self.feat_out = torch.nn.Linear(adim, odim)
self.decoder = Diffusion(ddim, beta_min, beta_max, pe_scale)
self._reset_parameters(
init_type=init_type,
init_enc_alpha=init_enc_alpha,
init_dec_alpha=init_dec_alpha,
)
self.criterion = GradTTSLoss(
odim,
use_masking=use_masking,
use_weighted_masking=use_weighted_masking,
)
