def __init__(self, config: GlowTTSConfig):
super().__init__()
# pass all config fields to `self`
# for fewer code change
self.config = config
for key in config:
setattr(self, key, config[key])
_, self.config, self.num_chars = self.get_characters(config)
self.decoder_output_dim = config.out_channels
self.init_multispeaker(config)
# if is a multispeaker and c_in_channels is 0, set to 256
self.c_in_channels = 0
if self.num_speakers > 1:
if self.d_vector_dim:
self.c_in_channels = self.d_vector_dim
elif self.c_in_channels == 0 and not self.d_vector_dim:
# TODO: make this adjustable
self.c_in_channels = 256
self.run_data_dep_init = config.data_dep_init_steps > 0
self.encoder = Encoder(
self.num_chars,
out_channels=self.out_channels,
hidden_channels=self.hidden_channels_enc,
hidden_channels_dp=self.hidden_channels_dp,
encoder_type=self.encoder_type,
encoder_params=self.encoder_params,
mean_only=self.mean_only,
use_prenet=self.use_encoder_prenet,
dropout_p_dp=self.dropout_p_dp,
c_in_channels=self.c_in_channels,
)
self.decoder = Decoder(
self.out_channels,
self.hidden_channels_dec,
self.kernel_size_dec,
self.dilation_rate,
self.num_flow_blocks_dec,
self.num_block_layers,
dropout_p=self.dropout_p_dec,
num_splits=self.num_splits,
num_squeeze=self.num_squeeze,
sigmoid_scale=self.sigmoid_scale,
c_in_channels=self.c_in_channels,
)
def __init__(
self,
config: GlowTTSConfig,
ap: "AudioProcessor" = None,
tokenizer: "TTSTokenizer" = None,
speaker_manager: SpeakerManager = None,
):
super().__init__(config, ap, tokenizer, speaker_manager)
# pass all config fields to `self`
# for fewer code change
self.config = config
for key in config:
setattr(self, key, config[key])
self.decoder_output_dim = config.out_channels
# init multi-speaker layers if necessary
self.init_multispeaker(config)
self.run_data_dep_init = config.data_dep_init_steps > 0
self.encoder = Encoder(
self.num_chars,
out_channels=self.out_channels,
hidden_channels=self.hidden_channels_enc,
hidden_channels_dp=self.hidden_channels_dp,
encoder_type=self.encoder_type,
encoder_params=self.encoder_params,
mean_only=self.mean_only,
use_prenet=self.use_encoder_prenet,
dropout_p_dp=self.dropout_p_dp,
c_in_channels=self.c_in_channels,
)
self.decoder = Decoder(
self.out_channels,
self.hidden_channels_dec,
self.kernel_size_dec,
self.dilation_rate,
self.num_flow_blocks_dec,
self.num_block_layers,
dropout_p=self.dropout_p_dec,
num_splits=self.num_splits,
num_squeeze=self.num_squeeze,
sigmoid_scale=self.sigmoid_scale,
c_in_channels=self.c_in_channels,
)
def __init__(
self,
num_chars,
hidden_channels_enc,
hidden_channels_dec,
use_encoder_prenet,
hidden_channels_dp,
out_channels,
num_flow_blocks_dec=12,
inference_noise_scale=0.33,
kernel_size_dec=5,
dilation_rate=5,
num_block_layers=4,
dropout_p_dp=0.1,
dropout_p_dec=0.05,
num_speakers=0,
c_in_channels=0,
num_splits=4,
num_squeeze=1,
sigmoid_scale=False,
mean_only=False,
encoder_type="transformer",
encoder_params=None,
speaker_embedding_dim=None,
):
super().__init__()
self.num_chars = num_chars
self.hidden_channels_dp = hidden_channels_dp
self.hidden_channels_enc = hidden_channels_enc
self.hidden_channels_dec = hidden_channels_dec
self.out_channels = out_channels
self.num_flow_blocks_dec = num_flow_blocks_dec
self.kernel_size_dec = kernel_size_dec
self.dilation_rate = dilation_rate
self.num_block_layers = num_block_layers
self.dropout_p_dec = dropout_p_dec
self.num_speakers = num_speakers
self.c_in_channels = c_in_channels
self.num_splits = num_splits
self.num_squeeze = num_squeeze
self.sigmoid_scale = sigmoid_scale
self.mean_only = mean_only
self.use_encoder_prenet = use_encoder_prenet
self.inference_noise_scale = inference_noise_scale
# model constants.
self.noise_scale = 0.33 # defines the noise variance applied to the random z vector at inference.
self.length_scale = 1.0 # scaler for the duration predictor. The larger it is, the slower the speech.
self.speaker_embedding_dim = speaker_embedding_dim
# if is a multispeaker and c_in_channels is 0, set to 256
if num_speakers > 1:
if self.c_in_channels == 0 and not self.speaker_embedding_dim:
# TODO: make this adjustable
self.c_in_channels = 256
elif self.speaker_embedding_dim:
self.c_in_channels = self.speaker_embedding_dim
self.encoder = Encoder(
num_chars,
out_channels=out_channels,
hidden_channels=hidden_channels_enc,
hidden_channels_dp=hidden_channels_dp,
encoder_type=encoder_type,
encoder_params=encoder_params,
mean_only=mean_only,
use_prenet=use_encoder_prenet,
dropout_p_dp=dropout_p_dp,
c_in_channels=self.c_in_channels,
)
self.decoder = Decoder(
out_channels,
hidden_channels_dec,
kernel_size_dec,
dilation_rate,
num_flow_blocks_dec,
num_block_layers,
dropout_p=dropout_p_dec,
num_splits=num_splits,
num_squeeze=num_squeeze,
sigmoid_scale=sigmoid_scale,
c_in_channels=self.c_in_channels,
)
if num_speakers > 1 and not speaker_embedding_dim:
# speaker embedding layer
self.emb_g = nn.Embedding(num_speakers, self.c_in_channels)
nn.init.uniform_(self.emb_g.weight, -0.1, 0.1)
def __init__(self,
num_chars,
hidden_channels,
filter_channels,
filter_channels_dp,
out_channels,
kernel_size=3,
num_heads=2,
num_layers_enc=6,
dropout_p=0.1,
num_flow_blocks_dec=12,
kernel_size_dec=5,
dilation_rate=5,
num_block_layers=4,
dropout_p_dec=0.,
num_speakers=0,
c_in_channels=0,
num_splits=4,
num_sqz=1,
sigmoid_scale=False,
rel_attn_window_size=None,
input_length=None,
mean_only=False,
hidden_channels_enc=None,
hidden_channels_dec=None,
use_encoder_prenet=False,
encoder_type="transformer"):
super().__init__()
self.num_chars = num_chars
self.hidden_channels = hidden_channels
self.filter_channels = filter_channels
self.filter_channels_dp = filter_channels_dp
self.out_channels = out_channels
self.kernel_size = kernel_size
self.num_heads = num_heads
self.num_layers_enc = num_layers_enc
self.dropout_p = dropout_p
self.num_flow_blocks_dec = num_flow_blocks_dec
self.kernel_size_dec = kernel_size_dec
self.dilation_rate = dilation_rate
self.num_block_layers = num_block_layers
self.dropout_p_dec = dropout_p_dec
self.num_speakers = num_speakers
self.c_in_channels = c_in_channels
self.num_splits = num_splits
self.num_sqz = num_sqz
self.sigmoid_scale = sigmoid_scale
self.rel_attn_window_size = rel_attn_window_size
self.input_length = input_length
self.mean_only = mean_only
self.hidden_channels_enc = hidden_channels_enc
self.hidden_channels_dec = hidden_channels_dec
self.use_encoder_prenet = use_encoder_prenet
self.noise_scale = 0.66
self.length_scale = 1.
self.encoder = Encoder(num_chars,
out_channels=out_channels,
hidden_channels=hidden_channels,
filter_channels=filter_channels,
filter_channels_dp=filter_channels_dp,
encoder_type=encoder_type,
num_heads=num_heads,
num_layers=num_layers_enc,
kernel_size=kernel_size,
dropout_p=dropout_p,
mean_only=mean_only,
use_prenet=use_encoder_prenet,
c_in_channels=c_in_channels)
self.decoder = Decoder(out_channels,
hidden_channels_dec or hidden_channels,
kernel_size_dec,
dilation_rate,
num_flow_blocks_dec,
num_block_layers,
dropout_p=dropout_p_dec,
num_splits=num_splits,
num_sqz=num_sqz,
sigmoid_scale=sigmoid_scale,
c_in_channels=c_in_channels)
if num_speakers > 1:
self.emb_g = nn.Embedding(num_speakers, c_in_channels)
nn.init.uniform_(self.emb_g.weight, -0.1, 0.1)