Add files via upload
This commit is contained in:
Sam Khoze
@@ -0,0 +1,17 @@
|
||||
import importlib
|
||||
import os
|
||||
from inspect import isclass
|
||||
|
||||
# import all files under configs/
|
||||
configs_dir = os.path.dirname(__file__)
|
||||
for file in os.listdir(configs_dir):
|
||||
path = os.path.join(configs_dir, file)
|
||||
if not file.startswith("_") and not file.startswith(".") and (file.endswith(".py") or os.path.isdir(path)):
|
||||
config_name = file[: file.find(".py")] if file.endswith(".py") else file
|
||||
module = importlib.import_module("TTS.vocoder.configs." + config_name)
|
||||
for attribute_name in dir(module):
|
||||
attribute = getattr(module, attribute_name)
|
||||
|
||||
if isclass(attribute):
|
||||
# Add the class to this package's variables
|
||||
globals()[attribute_name] = attribute
|
||||
Binary file not shown.
Binary file not shown.
Binary file not shown.
Binary file not shown.
Binary file not shown.
Binary file not shown.
Binary file not shown.
Binary file not shown.
Binary file not shown.
Binary file not shown.
@@ -0,0 +1,106 @@
|
||||
from dataclasses import dataclass, field
|
||||
|
||||
from .shared_configs import BaseGANVocoderConfig
|
||||
|
||||
|
||||
@dataclass
|
||||
class FullbandMelganConfig(BaseGANVocoderConfig):
|
||||
"""Defines parameters for FullBand MelGAN vocoder.
|
||||
|
||||
Example:
|
||||
|
||||
>>> from TTS.vocoder.configs import FullbandMelganConfig
|
||||
>>> config = FullbandMelganConfig()
|
||||
|
||||
Args:
|
||||
model (str):
|
||||
Model name used for selecting the right model at initialization. Defaults to `fullband_melgan`.
|
||||
discriminator_model (str): One of the discriminators from `TTS.vocoder.models.*_discriminator`. Defaults to
|
||||
'melgan_multiscale_discriminator`.
|
||||
discriminator_model_params (dict): The discriminator model parameters. Defaults to
|
||||
'{"base_channels": 16, "max_channels": 1024, "downsample_factors": [4, 4, 4, 4]}`
|
||||
generator_model (str): One of the generators from TTS.vocoder.models.*`. Every other non-GAN vocoder model is
|
||||
considered as a generator too. Defaults to `melgan_generator`.
|
||||
batch_size (int):
|
||||
Batch size used at training. Larger values use more memory. Defaults to 16.
|
||||
seq_len (int):
|
||||
Audio segment length used at training. Larger values use more memory. Defaults to 8192.
|
||||
pad_short (int):
|
||||
Additional padding applied to the audio samples shorter than `seq_len`. Defaults to 0.
|
||||
use_noise_augment (bool):
|
||||
enable / disable random noise added to the input waveform. The noise is added after computing the
|
||||
features. Defaults to True.
|
||||
use_cache (bool):
|
||||
enable / disable in memory caching of the computed features. It can cause OOM error if the system RAM is
|
||||
not large enough. Defaults to True.
|
||||
use_stft_loss (bool):
|
||||
enable / disable use of STFT loss originally used by ParallelWaveGAN model. Defaults to True.
|
||||
use_subband_stft (bool):
|
||||
enable / disable use of subband loss computation originally used by MultiBandMelgan model. Defaults to True.
|
||||
use_mse_gan_loss (bool):
|
||||
enable / disable using Mean Squeare Error GAN loss. Defaults to True.
|
||||
use_hinge_gan_loss (bool):
|
||||
enable / disable using Hinge GAN loss. You should choose either Hinge or MSE loss for training GAN models.
|
||||
Defaults to False.
|
||||
use_feat_match_loss (bool):
|
||||
enable / disable using Feature Matching loss originally used by MelGAN model. Defaults to True.
|
||||
use_l1_spec_loss (bool):
|
||||
enable / disable using L1 spectrogram loss originally used by HifiGAN model. Defaults to False.
|
||||
stft_loss_params (dict): STFT loss parameters. Default to
|
||||
`{"n_ffts": [1024, 2048, 512], "hop_lengths": [120, 240, 50], "win_lengths": [600, 1200, 240]}`
|
||||
stft_loss_weight (float): STFT loss weight that multiplies the computed loss before summing up the total
|
||||
model loss. Defaults to 0.5.
|
||||
subband_stft_loss_weight (float):
|
||||
Subband STFT loss weight that multiplies the computed loss before summing up the total loss. Defaults to 0.
|
||||
mse_G_loss_weight (float):
|
||||
MSE generator loss weight that multiplies the computed loss before summing up the total loss. faults to 2.5.
|
||||
hinge_G_loss_weight (float):
|
||||
Hinge generator loss weight that multiplies the computed loss before summing up the total loss. Defaults to 0.
|
||||
feat_match_loss_weight (float):
|
||||
Feature matching loss weight that multiplies the computed loss before summing up the total loss. faults to 108.
|
||||
l1_spec_loss_weight (float):
|
||||
L1 spectrogram loss weight that multiplies the computed loss before summing up the total loss. Defaults to 0.
|
||||
"""
|
||||
|
||||
model: str = "fullband_melgan"
|
||||
|
||||
# Model specific params
|
||||
discriminator_model: str = "melgan_multiscale_discriminator"
|
||||
discriminator_model_params: dict = field(
|
||||
default_factory=lambda: {"base_channels": 16, "max_channels": 512, "downsample_factors": [4, 4, 4]}
|
||||
)
|
||||
generator_model: str = "melgan_generator"
|
||||
generator_model_params: dict = field(
|
||||
default_factory=lambda: {"upsample_factors": [8, 8, 2, 2], "num_res_blocks": 4}
|
||||
)
|
||||
|
||||
# Training - overrides
|
||||
batch_size: int = 16
|
||||
seq_len: int = 8192
|
||||
pad_short: int = 2000
|
||||
use_noise_augment: bool = True
|
||||
use_cache: bool = True
|
||||
|
||||
# LOSS PARAMETERS - overrides
|
||||
use_stft_loss: bool = True
|
||||
use_subband_stft_loss: bool = False
|
||||
use_mse_gan_loss: bool = True
|
||||
use_hinge_gan_loss: bool = False
|
||||
use_feat_match_loss: bool = True # requires MelGAN Discriminators (MelGAN and HifiGAN)
|
||||
use_l1_spec_loss: bool = False
|
||||
|
||||
stft_loss_params: dict = field(
|
||||
default_factory=lambda: {
|
||||
"n_ffts": [1024, 2048, 512],
|
||||
"hop_lengths": [120, 240, 50],
|
||||
"win_lengths": [600, 1200, 240],
|
||||
}
|
||||
)
|
||||
|
||||
# loss weights - overrides
|
||||
stft_loss_weight: float = 0.5
|
||||
subband_stft_loss_weight: float = 0
|
||||
mse_G_loss_weight: float = 2.5
|
||||
hinge_G_loss_weight: float = 0
|
||||
feat_match_loss_weight: float = 108
|
||||
l1_spec_loss_weight: float = 0.0
|
||||
@@ -0,0 +1,136 @@
|
||||
from dataclasses import dataclass, field
|
||||
|
||||
from TTS.vocoder.configs.shared_configs import BaseGANVocoderConfig
|
||||
|
||||
|
||||
@dataclass
|
||||
class HifiganConfig(BaseGANVocoderConfig):
|
||||
"""Defines parameters for FullBand MelGAN vocoder.
|
||||
|
||||
Example:
|
||||
|
||||
>>> from TTS.vocoder.configs import HifiganConfig
|
||||
>>> config = HifiganConfig()
|
||||
|
||||
Args:
|
||||
model (str):
|
||||
Model name used for selecting the right model at initialization. Defaults to `hifigan`.
|
||||
discriminator_model (str): One of the discriminators from `TTS.vocoder.models.*_discriminator`. Defaults to
|
||||
'hifigan_discriminator`.
|
||||
generator_model (str): One of the generators from TTS.vocoder.models.*`. Every other non-GAN vocoder model is
|
||||
considered as a generator too. Defaults to `hifigan_generator`.
|
||||
generator_model_params (dict): Parameters of the generator model. Defaults to
|
||||
`
|
||||
{
|
||||
"upsample_factors": [8, 8, 2, 2],
|
||||
"upsample_kernel_sizes": [16, 16, 4, 4],
|
||||
"upsample_initial_channel": 512,
|
||||
"resblock_kernel_sizes": [3, 7, 11],
|
||||
"resblock_dilation_sizes": [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
|
||||
"resblock_type": "1",
|
||||
}
|
||||
`
|
||||
batch_size (int):
|
||||
Batch size used at training. Larger values use more memory. Defaults to 16.
|
||||
seq_len (int):
|
||||
Audio segment length used at training. Larger values use more memory. Defaults to 8192.
|
||||
pad_short (int):
|
||||
Additional padding applied to the audio samples shorter than `seq_len`. Defaults to 0.
|
||||
use_noise_augment (bool):
|
||||
enable / disable random noise added to the input waveform. The noise is added after computing the
|
||||
features. Defaults to True.
|
||||
use_cache (bool):
|
||||
enable / disable in memory caching of the computed features. It can cause OOM error if the system RAM is
|
||||
not large enough. Defaults to True.
|
||||
use_stft_loss (bool):
|
||||
enable / disable use of STFT loss originally used by ParallelWaveGAN model. Defaults to True.
|
||||
use_subband_stft (bool):
|
||||
enable / disable use of subband loss computation originally used by MultiBandMelgan model. Defaults to True.
|
||||
use_mse_gan_loss (bool):
|
||||
enable / disable using Mean Squeare Error GAN loss. Defaults to True.
|
||||
use_hinge_gan_loss (bool):
|
||||
enable / disable using Hinge GAN loss. You should choose either Hinge or MSE loss for training GAN models.
|
||||
Defaults to False.
|
||||
use_feat_match_loss (bool):
|
||||
enable / disable using Feature Matching loss originally used by MelGAN model. Defaults to True.
|
||||
use_l1_spec_loss (bool):
|
||||
enable / disable using L1 spectrogram loss originally used by HifiGAN model. Defaults to False.
|
||||
stft_loss_params (dict):
|
||||
STFT loss parameters. Default to
|
||||
`{
|
||||
"n_ffts": [1024, 2048, 512],
|
||||
"hop_lengths": [120, 240, 50],
|
||||
"win_lengths": [600, 1200, 240]
|
||||
}`
|
||||
l1_spec_loss_params (dict):
|
||||
L1 spectrogram loss parameters. Default to
|
||||
`{
|
||||
"use_mel": True,
|
||||
"sample_rate": 22050,
|
||||
"n_fft": 1024,
|
||||
"hop_length": 256,
|
||||
"win_length": 1024,
|
||||
"n_mels": 80,
|
||||
"mel_fmin": 0.0,
|
||||
"mel_fmax": None,
|
||||
}`
|
||||
stft_loss_weight (float): STFT loss weight that multiplies the computed loss before summing up the total
|
||||
model loss. Defaults to 0.5.
|
||||
subband_stft_loss_weight (float):
|
||||
Subband STFT loss weight that multiplies the computed loss before summing up the total loss. Defaults to 0.
|
||||
mse_G_loss_weight (float):
|
||||
MSE generator loss weight that multiplies the computed loss before summing up the total loss. faults to 2.5.
|
||||
hinge_G_loss_weight (float):
|
||||
Hinge generator loss weight that multiplies the computed loss before summing up the total loss. Defaults to 0.
|
||||
feat_match_loss_weight (float):
|
||||
Feature matching loss weight that multiplies the computed loss before summing up the total loss. faults to 108.
|
||||
l1_spec_loss_weight (float):
|
||||
L1 spectrogram loss weight that multiplies the computed loss before summing up the total loss. Defaults to 0.
|
||||
"""
|
||||
|
||||
model: str = "hifigan"
|
||||
# model specific params
|
||||
discriminator_model: str = "hifigan_discriminator"
|
||||
generator_model: str = "hifigan_generator"
|
||||
generator_model_params: dict = field(
|
||||
default_factory=lambda: {
|
||||
"upsample_factors": [8, 8, 2, 2],
|
||||
"upsample_kernel_sizes": [16, 16, 4, 4],
|
||||
"upsample_initial_channel": 512,
|
||||
"resblock_kernel_sizes": [3, 7, 11],
|
||||
"resblock_dilation_sizes": [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
|
||||
"resblock_type": "1",
|
||||
}
|
||||
)
|
||||
|
||||
# LOSS PARAMETERS - overrides
|
||||
use_stft_loss: bool = False
|
||||
use_subband_stft_loss: bool = False
|
||||
use_mse_gan_loss: bool = True
|
||||
use_hinge_gan_loss: bool = False
|
||||
use_feat_match_loss: bool = True # requires MelGAN Discriminators (MelGAN and HifiGAN)
|
||||
use_l1_spec_loss: bool = True
|
||||
|
||||
# loss weights - overrides
|
||||
stft_loss_weight: float = 0
|
||||
subband_stft_loss_weight: float = 0
|
||||
mse_G_loss_weight: float = 1
|
||||
hinge_G_loss_weight: float = 0
|
||||
feat_match_loss_weight: float = 108
|
||||
l1_spec_loss_weight: float = 45
|
||||
l1_spec_loss_params: dict = field(
|
||||
default_factory=lambda: {
|
||||
"use_mel": True,
|
||||
"sample_rate": 22050,
|
||||
"n_fft": 1024,
|
||||
"hop_length": 256,
|
||||
"win_length": 1024,
|
||||
"n_mels": 80,
|
||||
"mel_fmin": 0.0,
|
||||
"mel_fmax": None,
|
||||
}
|
||||
)
|
||||
|
||||
# optimizer parameters
|
||||
lr: float = 1e-4
|
||||
wd: float = 1e-6
|
||||
@@ -0,0 +1,106 @@
|
||||
from dataclasses import dataclass, field
|
||||
|
||||
from TTS.vocoder.configs.shared_configs import BaseGANVocoderConfig
|
||||
|
||||
|
||||
@dataclass
|
||||
class MelganConfig(BaseGANVocoderConfig):
|
||||
"""Defines parameters for MelGAN vocoder.
|
||||
|
||||
Example:
|
||||
|
||||
>>> from TTS.vocoder.configs import MelganConfig
|
||||
>>> config = MelganConfig()
|
||||
|
||||
Args:
|
||||
model (str):
|
||||
Model name used for selecting the right model at initialization. Defaults to `melgan`.
|
||||
discriminator_model (str): One of the discriminators from `TTS.vocoder.models.*_discriminator`. Defaults to
|
||||
'melgan_multiscale_discriminator`.
|
||||
discriminator_model_params (dict): The discriminator model parameters. Defaults to
|
||||
'{"base_channels": 16, "max_channels": 1024, "downsample_factors": [4, 4, 4, 4]}`
|
||||
generator_model (str): One of the generators from TTS.vocoder.models.*`. Every other non-GAN vocoder model is
|
||||
considered as a generator too. Defaults to `melgan_generator`.
|
||||
batch_size (int):
|
||||
Batch size used at training. Larger values use more memory. Defaults to 16.
|
||||
seq_len (int):
|
||||
Audio segment length used at training. Larger values use more memory. Defaults to 8192.
|
||||
pad_short (int):
|
||||
Additional padding applied to the audio samples shorter than `seq_len`. Defaults to 0.
|
||||
use_noise_augment (bool):
|
||||
enable / disable random noise added to the input waveform. The noise is added after computing the
|
||||
features. Defaults to True.
|
||||
use_cache (bool):
|
||||
enable / disable in memory caching of the computed features. It can cause OOM error if the system RAM is
|
||||
not large enough. Defaults to True.
|
||||
use_stft_loss (bool):
|
||||
enable / disable use of STFT loss originally used by ParallelWaveGAN model. Defaults to True.
|
||||
use_subband_stft (bool):
|
||||
enable / disable use of subband loss computation originally used by MultiBandMelgan model. Defaults to True.
|
||||
use_mse_gan_loss (bool):
|
||||
enable / disable using Mean Squeare Error GAN loss. Defaults to True.
|
||||
use_hinge_gan_loss (bool):
|
||||
enable / disable using Hinge GAN loss. You should choose either Hinge or MSE loss for training GAN models.
|
||||
Defaults to False.
|
||||
use_feat_match_loss (bool):
|
||||
enable / disable using Feature Matching loss originally used by MelGAN model. Defaults to True.
|
||||
use_l1_spec_loss (bool):
|
||||
enable / disable using L1 spectrogram loss originally used by HifiGAN model. Defaults to False.
|
||||
stft_loss_params (dict): STFT loss parameters. Default to
|
||||
`{"n_ffts": [1024, 2048, 512], "hop_lengths": [120, 240, 50], "win_lengths": [600, 1200, 240]}`
|
||||
stft_loss_weight (float): STFT loss weight that multiplies the computed loss before summing up the total
|
||||
model loss. Defaults to 0.5.
|
||||
subband_stft_loss_weight (float):
|
||||
Subband STFT loss weight that multiplies the computed loss before summing up the total loss. Defaults to 0.
|
||||
mse_G_loss_weight (float):
|
||||
MSE generator loss weight that multiplies the computed loss before summing up the total loss. faults to 2.5.
|
||||
hinge_G_loss_weight (float):
|
||||
Hinge generator loss weight that multiplies the computed loss before summing up the total loss. Defaults to 0.
|
||||
feat_match_loss_weight (float):
|
||||
Feature matching loss weight that multiplies the computed loss before summing up the total loss. faults to 108.
|
||||
l1_spec_loss_weight (float):
|
||||
L1 spectrogram loss weight that multiplies the computed loss before summing up the total loss. Defaults to 0.
|
||||
"""
|
||||
|
||||
model: str = "melgan"
|
||||
|
||||
# Model specific params
|
||||
discriminator_model: str = "melgan_multiscale_discriminator"
|
||||
discriminator_model_params: dict = field(
|
||||
default_factory=lambda: {"base_channels": 16, "max_channels": 1024, "downsample_factors": [4, 4, 4, 4]}
|
||||
)
|
||||
generator_model: str = "melgan_generator"
|
||||
generator_model_params: dict = field(
|
||||
default_factory=lambda: {"upsample_factors": [8, 8, 2, 2], "num_res_blocks": 3}
|
||||
)
|
||||
|
||||
# Training - overrides
|
||||
batch_size: int = 16
|
||||
seq_len: int = 8192
|
||||
pad_short: int = 2000
|
||||
use_noise_augment: bool = True
|
||||
use_cache: bool = True
|
||||
|
||||
# LOSS PARAMETERS - overrides
|
||||
use_stft_loss: bool = True
|
||||
use_subband_stft_loss: bool = False
|
||||
use_mse_gan_loss: bool = True
|
||||
use_hinge_gan_loss: bool = False
|
||||
use_feat_match_loss: bool = True # requires MelGAN Discriminators (MelGAN and HifiGAN)
|
||||
use_l1_spec_loss: bool = False
|
||||
|
||||
stft_loss_params: dict = field(
|
||||
default_factory=lambda: {
|
||||
"n_ffts": [1024, 2048, 512],
|
||||
"hop_lengths": [120, 240, 50],
|
||||
"win_lengths": [600, 1200, 240],
|
||||
}
|
||||
)
|
||||
|
||||
# loss weights - overrides
|
||||
stft_loss_weight: float = 0.5
|
||||
subband_stft_loss_weight: float = 0
|
||||
mse_G_loss_weight: float = 2.5
|
||||
hinge_G_loss_weight: float = 0
|
||||
feat_match_loss_weight: float = 108
|
||||
l1_spec_loss_weight: float = 0
|
||||
@@ -0,0 +1,144 @@
|
||||
from dataclasses import dataclass, field
|
||||
|
||||
from TTS.vocoder.configs.shared_configs import BaseGANVocoderConfig
|
||||
|
||||
|
||||
@dataclass
|
||||
class MultibandMelganConfig(BaseGANVocoderConfig):
|
||||
"""Defines parameters for MultiBandMelGAN vocoder.
|
||||
|
||||
Example:
|
||||
|
||||
>>> from TTS.vocoder.configs import MultibandMelganConfig
|
||||
>>> config = MultibandMelganConfig()
|
||||
|
||||
Args:
|
||||
model (str):
|
||||
Model name used for selecting the right model at initialization. Defaults to `multiband_melgan`.
|
||||
discriminator_model (str): One of the discriminators from `TTS.vocoder.models.*_discriminator`. Defaults to
|
||||
'melgan_multiscale_discriminator`.
|
||||
discriminator_model_params (dict): The discriminator model parameters. Defaults to
|
||||
'{
|
||||
"base_channels": 16,
|
||||
"max_channels": 512,
|
||||
"downsample_factors": [4, 4, 4]
|
||||
}`
|
||||
generator_model (str): One of the generators from TTS.vocoder.models.*`. Every other non-GAN vocoder model is
|
||||
considered as a generator too. Defaults to `melgan_generator`.
|
||||
generator_model_param (dict):
|
||||
The generator model parameters. Defaults to `{"upsample_factors": [8, 4, 2], "num_res_blocks": 4}`.
|
||||
use_pqmf (bool):
|
||||
enable / disable PQMF modulation for multi-band training. Defaults to True.
|
||||
lr_gen (float):
|
||||
Initial learning rate for the generator model. Defaults to 0.0001.
|
||||
lr_disc (float):
|
||||
Initial learning rate for the discriminator model. Defaults to 0.0001.
|
||||
optimizer (torch.optim.Optimizer):
|
||||
Optimizer used for the training. Defaults to `AdamW`.
|
||||
optimizer_params (dict):
|
||||
Optimizer kwargs. Defaults to `{"betas": [0.8, 0.99], "weight_decay": 0.0}`
|
||||
lr_scheduler_gen (torch.optim.Scheduler):
|
||||
Learning rate scheduler for the generator. Defaults to `MultiStepLR`.
|
||||
lr_scheduler_gen_params (dict):
|
||||
Parameters for the generator learning rate scheduler. Defaults to
|
||||
`{"gamma": 0.5, "milestones": [100000, 200000, 300000, 400000, 500000, 600000]}`.
|
||||
lr_scheduler_disc (torch.optim.Scheduler):
|
||||
Learning rate scheduler for the discriminator. Defaults to `MultiStepLR`.
|
||||
lr_scheduler_dict_params (dict):
|
||||
Parameters for the discriminator learning rate scheduler. Defaults to
|
||||
`{"gamma": 0.5, "milestones": [100000, 200000, 300000, 400000, 500000, 600000]}`.
|
||||
batch_size (int):
|
||||
Batch size used at training. Larger values use more memory. Defaults to 16.
|
||||
seq_len (int):
|
||||
Audio segment length used at training. Larger values use more memory. Defaults to 8192.
|
||||
pad_short (int):
|
||||
Additional padding applied to the audio samples shorter than `seq_len`. Defaults to 0.
|
||||
use_noise_augment (bool):
|
||||
enable / disable random noise added to the input waveform. The noise is added after computing the
|
||||
features. Defaults to True.
|
||||
use_cache (bool):
|
||||
enable / disable in memory caching of the computed features. It can cause OOM error if the system RAM is
|
||||
not large enough. Defaults to True.
|
||||
steps_to_start_discriminator (int):
|
||||
Number of steps required to start training the discriminator. Defaults to 0.
|
||||
use_stft_loss (bool):`
|
||||
enable / disable use of STFT loss originally used by ParallelWaveGAN model. Defaults to True.
|
||||
use_subband_stft (bool):
|
||||
enable / disable use of subband loss computation originally used by MultiBandMelgan model. Defaults to True.
|
||||
use_mse_gan_loss (bool):
|
||||
enable / disable using Mean Squeare Error GAN loss. Defaults to True.
|
||||
use_hinge_gan_loss (bool):
|
||||
enable / disable using Hinge GAN loss. You should choose either Hinge or MSE loss for training GAN models.
|
||||
Defaults to False.
|
||||
use_feat_match_loss (bool):
|
||||
enable / disable using Feature Matching loss originally used by MelGAN model. Defaults to True.
|
||||
use_l1_spec_loss (bool):
|
||||
enable / disable using L1 spectrogram loss originally used by HifiGAN model. Defaults to False.
|
||||
stft_loss_params (dict): STFT loss parameters. Default to
|
||||
`{"n_ffts": [1024, 2048, 512], "hop_lengths": [120, 240, 50], "win_lengths": [600, 1200, 240]}`
|
||||
stft_loss_weight (float): STFT loss weight that multiplies the computed loss before summing up the total
|
||||
model loss. Defaults to 0.5.
|
||||
subband_stft_loss_weight (float):
|
||||
Subband STFT loss weight that multiplies the computed loss before summing up the total loss. Defaults to 0.
|
||||
mse_G_loss_weight (float):
|
||||
MSE generator loss weight that multiplies the computed loss before summing up the total loss. faults to 2.5.
|
||||
hinge_G_loss_weight (float):
|
||||
Hinge generator loss weight that multiplies the computed loss before summing up the total loss. Defaults to 0.
|
||||
feat_match_loss_weight (float):
|
||||
Feature matching loss weight that multiplies the computed loss before summing up the total loss. faults to 108.
|
||||
l1_spec_loss_weight (float):
|
||||
L1 spectrogram loss weight that multiplies the computed loss before summing up the total loss. Defaults to 0.
|
||||
"""
|
||||
|
||||
model: str = "multiband_melgan"
|
||||
|
||||
# Model specific params
|
||||
discriminator_model: str = "melgan_multiscale_discriminator"
|
||||
discriminator_model_params: dict = field(
|
||||
default_factory=lambda: {"base_channels": 16, "max_channels": 512, "downsample_factors": [4, 4, 4]}
|
||||
)
|
||||
generator_model: str = "multiband_melgan_generator"
|
||||
generator_model_params: dict = field(default_factory=lambda: {"upsample_factors": [8, 4, 2], "num_res_blocks": 4})
|
||||
use_pqmf: bool = True
|
||||
|
||||
# optimizer - overrides
|
||||
lr_gen: float = 0.0001 # Initial learning rate.
|
||||
lr_disc: float = 0.0001 # Initial learning rate.
|
||||
optimizer: str = "AdamW"
|
||||
optimizer_params: dict = field(default_factory=lambda: {"betas": [0.8, 0.99], "weight_decay": 0.0})
|
||||
lr_scheduler_gen: str = "MultiStepLR" # one of the schedulers from https:#pytorch.org/docs/stable/optim.html
|
||||
lr_scheduler_gen_params: dict = field(
|
||||
default_factory=lambda: {"gamma": 0.5, "milestones": [100000, 200000, 300000, 400000, 500000, 600000]}
|
||||
)
|
||||
lr_scheduler_disc: str = "MultiStepLR" # one of the schedulers from https:#pytorch.org/docs/stable/optim.html
|
||||
lr_scheduler_disc_params: dict = field(
|
||||
default_factory=lambda: {"gamma": 0.5, "milestones": [100000, 200000, 300000, 400000, 500000, 600000]}
|
||||
)
|
||||
|
||||
# Training - overrides
|
||||
batch_size: int = 64
|
||||
seq_len: int = 16384
|
||||
pad_short: int = 2000
|
||||
use_noise_augment: bool = False
|
||||
use_cache: bool = True
|
||||
steps_to_start_discriminator: bool = 200000
|
||||
|
||||
# LOSS PARAMETERS - overrides
|
||||
use_stft_loss: bool = True
|
||||
use_subband_stft_loss: bool = True
|
||||
use_mse_gan_loss: bool = True
|
||||
use_hinge_gan_loss: bool = False
|
||||
use_feat_match_loss: bool = False # requires MelGAN Discriminators (MelGAN and HifiGAN)
|
||||
use_l1_spec_loss: bool = False
|
||||
|
||||
subband_stft_loss_params: dict = field(
|
||||
default_factory=lambda: {"n_ffts": [384, 683, 171], "hop_lengths": [30, 60, 10], "win_lengths": [150, 300, 60]}
|
||||
)
|
||||
|
||||
# loss weights - overrides
|
||||
stft_loss_weight: float = 0.5
|
||||
subband_stft_loss_weight: float = 0
|
||||
mse_G_loss_weight: float = 2.5
|
||||
hinge_G_loss_weight: float = 0
|
||||
feat_match_loss_weight: float = 108
|
||||
l1_spec_loss_weight: float = 0
|
||||
@@ -0,0 +1,134 @@
|
||||
from dataclasses import dataclass, field
|
||||
|
||||
from .shared_configs import BaseGANVocoderConfig
|
||||
|
||||
|
||||
@dataclass
|
||||
class ParallelWaveganConfig(BaseGANVocoderConfig):
|
||||
"""Defines parameters for ParallelWavegan vocoder.
|
||||
|
||||
Args:
|
||||
model (str):
|
||||
Model name used for selecting the right configuration at initialization. Defaults to `gan`.
|
||||
discriminator_model (str): One of the discriminators from `TTS.vocoder.models.*_discriminator`. Defaults to
|
||||
'parallel_wavegan_discriminator`.
|
||||
discriminator_model_params (dict): The discriminator model kwargs. Defaults to
|
||||
'{"num_layers": 10}`
|
||||
generator_model (str): One of the generators from TTS.vocoder.models.*`. Every other non-GAN vocoder model is
|
||||
considered as a generator too. Defaults to `parallel_wavegan_generator`.
|
||||
generator_model_param (dict):
|
||||
The generator model kwargs. Defaults to `{"upsample_factors": [4, 4, 4, 4], "stacks": 3, "num_res_blocks": 30}`.
|
||||
batch_size (int):
|
||||
Batch size used at training. Larger values use more memory. Defaults to 16.
|
||||
seq_len (int):
|
||||
Audio segment length used at training. Larger values use more memory. Defaults to 8192.
|
||||
pad_short (int):
|
||||
Additional padding applied to the audio samples shorter than `seq_len`. Defaults to 0.
|
||||
use_noise_augment (bool):
|
||||
enable / disable random noise added to the input waveform. The noise is added after computing the
|
||||
features. Defaults to True.
|
||||
use_cache (bool):
|
||||
enable / disable in memory caching of the computed features. It can cause OOM error if the system RAM is
|
||||
not large enough. Defaults to True.
|
||||
steps_to_start_discriminator (int):
|
||||
Number of steps required to start training the discriminator. Defaults to 0.
|
||||
use_stft_loss (bool):`
|
||||
enable / disable use of STFT loss originally used by ParallelWaveGAN model. Defaults to True.
|
||||
use_subband_stft (bool):
|
||||
enable / disable use of subband loss computation originally used by MultiBandMelgan model. Defaults to True.
|
||||
use_mse_gan_loss (bool):
|
||||
enable / disable using Mean Squeare Error GAN loss. Defaults to True.
|
||||
use_hinge_gan_loss (bool):
|
||||
enable / disable using Hinge GAN loss. You should choose either Hinge or MSE loss for training GAN models.
|
||||
Defaults to False.
|
||||
use_feat_match_loss (bool):
|
||||
enable / disable using Feature Matching loss originally used by MelGAN model. Defaults to True.
|
||||
use_l1_spec_loss (bool):
|
||||
enable / disable using L1 spectrogram loss originally used by HifiGAN model. Defaults to False.
|
||||
stft_loss_params (dict): STFT loss parameters. Default to
|
||||
`{"n_ffts": [1024, 2048, 512], "hop_lengths": [120, 240, 50], "win_lengths": [600, 1200, 240]}`
|
||||
stft_loss_weight (float): STFT loss weight that multiplies the computed loss before summing up the total
|
||||
model loss. Defaults to 0.5.
|
||||
subband_stft_loss_weight (float):
|
||||
Subband STFT loss weight that multiplies the computed loss before summing up the total loss. Defaults to 0.
|
||||
mse_G_loss_weight (float):
|
||||
MSE generator loss weight that multiplies the computed loss before summing up the total loss. faults to 2.5.
|
||||
hinge_G_loss_weight (float):
|
||||
Hinge generator loss weight that multiplies the computed loss before summing up the total loss. Defaults to 0.
|
||||
feat_match_loss_weight (float):
|
||||
Feature matching loss weight that multiplies the computed loss before summing up the total loss. faults to 0.
|
||||
l1_spec_loss_weight (float):
|
||||
L1 spectrogram loss weight that multiplies the computed loss before summing up the total loss. Defaults to 0.
|
||||
lr_gen (float):
|
||||
Generator model initial learning rate. Defaults to 0.0002.
|
||||
lr_disc (float):
|
||||
Discriminator model initial learning rate. Defaults to 0.0002.
|
||||
optimizer (torch.optim.Optimizer):
|
||||
Optimizer used for the training. Defaults to `AdamW`.
|
||||
optimizer_params (dict):
|
||||
Optimizer kwargs. Defaults to `{"betas": [0.8, 0.99], "weight_decay": 0.0}`
|
||||
lr_scheduler_gen (torch.optim.Scheduler):
|
||||
Learning rate scheduler for the generator. Defaults to `ExponentialLR`.
|
||||
lr_scheduler_gen_params (dict):
|
||||
Parameters for the generator learning rate scheduler. Defaults to `{"gamma": 0.5, "step_size": 200000, "last_epoch": -1}`.
|
||||
lr_scheduler_disc (torch.optim.Scheduler):
|
||||
Learning rate scheduler for the discriminator. Defaults to `ExponentialLR`.
|
||||
lr_scheduler_dict_params (dict):
|
||||
Parameters for the discriminator learning rate scheduler. Defaults to `{"gamma": 0.5, "step_size": 200000, "last_epoch": -1}`.
|
||||
"""
|
||||
|
||||
model: str = "parallel_wavegan"
|
||||
|
||||
# Model specific params
|
||||
discriminator_model: str = "parallel_wavegan_discriminator"
|
||||
discriminator_model_params: dict = field(default_factory=lambda: {"num_layers": 10})
|
||||
generator_model: str = "parallel_wavegan_generator"
|
||||
generator_model_params: dict = field(
|
||||
default_factory=lambda: {"upsample_factors": [4, 4, 4, 4], "stacks": 3, "num_res_blocks": 30}
|
||||
)
|
||||
|
||||
# Training - overrides
|
||||
batch_size: int = 6
|
||||
seq_len: int = 25600
|
||||
pad_short: int = 2000
|
||||
use_noise_augment: bool = False
|
||||
use_cache: bool = True
|
||||
steps_to_start_discriminator: int = 200000
|
||||
target_loss: str = "loss_1"
|
||||
|
||||
# LOSS PARAMETERS - overrides
|
||||
use_stft_loss: bool = True
|
||||
use_subband_stft_loss: bool = False
|
||||
use_mse_gan_loss: bool = True
|
||||
use_hinge_gan_loss: bool = False
|
||||
use_feat_match_loss: bool = False # requires MelGAN Discriminators (MelGAN and HifiGAN)
|
||||
use_l1_spec_loss: bool = False
|
||||
|
||||
stft_loss_params: dict = field(
|
||||
default_factory=lambda: {
|
||||
"n_ffts": [1024, 2048, 512],
|
||||
"hop_lengths": [120, 240, 50],
|
||||
"win_lengths": [600, 1200, 240],
|
||||
}
|
||||
)
|
||||
|
||||
# loss weights - overrides
|
||||
stft_loss_weight: float = 0.5
|
||||
subband_stft_loss_weight: float = 0
|
||||
mse_G_loss_weight: float = 2.5
|
||||
hinge_G_loss_weight: float = 0
|
||||
feat_match_loss_weight: float = 0
|
||||
l1_spec_loss_weight: float = 0
|
||||
|
||||
# optimizer overrides
|
||||
lr_gen: float = 0.0002 # Initial learning rate.
|
||||
lr_disc: float = 0.0002 # Initial learning rate.
|
||||
optimizer: str = "AdamW"
|
||||
optimizer_params: dict = field(default_factory=lambda: {"betas": [0.8, 0.99], "weight_decay": 0.0})
|
||||
lr_scheduler_gen: str = "StepLR" # one of the schedulers from https:#pytorch.org/docs/stable/optim.html
|
||||
lr_scheduler_gen_params: dict = field(default_factory=lambda: {"gamma": 0.5, "step_size": 200000, "last_epoch": -1})
|
||||
lr_scheduler_disc: str = "StepLR" # one of the schedulers from https:#pytorch.org/docs/stable/optim.html
|
||||
lr_scheduler_disc_params: dict = field(
|
||||
default_factory=lambda: {"gamma": 0.5, "step_size": 200000, "last_epoch": -1}
|
||||
)
|
||||
scheduler_after_epoch: bool = False
|
||||
@@ -0,0 +1,182 @@
|
||||
from dataclasses import dataclass, field
|
||||
|
||||
from TTS.config import BaseAudioConfig, BaseTrainingConfig
|
||||
|
||||
|
||||
@dataclass
|
||||
class BaseVocoderConfig(BaseTrainingConfig):
|
||||
"""Shared parameters among all the vocoder models.
|
||||
Args:
|
||||
audio (BaseAudioConfig):
|
||||
Audio processor config instance. Defaultsto `BaseAudioConfig()`.
|
||||
use_noise_augment (bool):
|
||||
Augment the input audio with random noise. Defaults to False/
|
||||
eval_split_size (int):
|
||||
Number of instances used for evaluation. Defaults to 10.
|
||||
data_path (str):
|
||||
Root path of the training data. All the audio files found recursively from this root path are used for
|
||||
training. Defaults to `""`.
|
||||
feature_path (str):
|
||||
Root path to the precomputed feature files. Defaults to None.
|
||||
seq_len (int):
|
||||
Length of the waveform segments used for training. Defaults to 1000.
|
||||
pad_short (int):
|
||||
Extra padding for the waveforms shorter than `seq_len`. Defaults to 0.
|
||||
conv_path (int):
|
||||
Extra padding for the feature frames against convolution of the edge frames. Defaults to MISSING.
|
||||
Defaults to 0.
|
||||
use_cache (bool):
|
||||
enable / disable in memory caching of the computed features. If the RAM is not enough, if may cause OOM.
|
||||
Defaults to False.
|
||||
epochs (int):
|
||||
Number of training epochs to. Defaults to 10000.
|
||||
wd (float):
|
||||
Weight decay.
|
||||
optimizer (torch.optim.Optimizer):
|
||||
Optimizer used for the training. Defaults to `AdamW`.
|
||||
optimizer_params (dict):
|
||||
Optimizer kwargs. Defaults to `{"betas": [0.8, 0.99], "weight_decay": 0.0}`
|
||||
"""
|
||||
|
||||
audio: BaseAudioConfig = field(default_factory=BaseAudioConfig)
|
||||
# dataloading
|
||||
use_noise_augment: bool = False # enable/disable random noise augmentation in spectrograms.
|
||||
eval_split_size: int = 10 # number of samples used for evaluation.
|
||||
# dataset
|
||||
data_path: str = "" # root data path. It finds all wav files recursively from there.
|
||||
feature_path: str = None # if you use precomputed features
|
||||
seq_len: int = 1000 # signal length used in training.
|
||||
pad_short: int = 0 # additional padding for short wavs
|
||||
conv_pad: int = 0 # additional padding against convolutions applied to spectrograms
|
||||
use_cache: bool = False # use in memory cache to keep the computed features. This might cause OOM.
|
||||
# OPTIMIZER
|
||||
epochs: int = 10000 # total number of epochs to train.
|
||||
wd: float = 0.0 # Weight decay weight.
|
||||
optimizer: str = "AdamW"
|
||||
optimizer_params: dict = field(default_factory=lambda: {"betas": [0.8, 0.99], "weight_decay": 0.0})
|
||||
|
||||
|
||||
@dataclass
|
||||
class BaseGANVocoderConfig(BaseVocoderConfig):
|
||||
"""Base config class used among all the GAN based vocoders.
|
||||
Args:
|
||||
use_stft_loss (bool):
|
||||
enable / disable the use of STFT loss. Defaults to True.
|
||||
use_subband_stft_loss (bool):
|
||||
enable / disable the use of Subband STFT loss. Defaults to True.
|
||||
use_mse_gan_loss (bool):
|
||||
enable / disable the use of Mean Squared Error based GAN loss. Defaults to True.
|
||||
use_hinge_gan_loss (bool):
|
||||
enable / disable the use of Hinge GAN loss. Defaults to True.
|
||||
use_feat_match_loss (bool):
|
||||
enable / disable feature matching loss. Defaults to True.
|
||||
use_l1_spec_loss (bool):
|
||||
enable / disable L1 spectrogram loss. Defaults to True.
|
||||
stft_loss_weight (float):
|
||||
Loss weight that multiplies the computed loss value. Defaults to 0.
|
||||
subband_stft_loss_weight (float):
|
||||
Loss weight that multiplies the computed loss value. Defaults to 0.
|
||||
mse_G_loss_weight (float):
|
||||
Loss weight that multiplies the computed loss value. Defaults to 1.
|
||||
hinge_G_loss_weight (float):
|
||||
Loss weight that multiplies the computed loss value. Defaults to 0.
|
||||
feat_match_loss_weight (float):
|
||||
Loss weight that multiplies the computed loss value. Defaults to 100.
|
||||
l1_spec_loss_weight (float):
|
||||
Loss weight that multiplies the computed loss value. Defaults to 45.
|
||||
stft_loss_params (dict):
|
||||
Parameters for the STFT loss. Defaults to `{"n_ffts": [1024, 2048, 512], "hop_lengths": [120, 240, 50], "win_lengths": [600, 1200, 240]}`.
|
||||
l1_spec_loss_params (dict):
|
||||
Parameters for the L1 spectrogram loss. Defaults to
|
||||
`{
|
||||
"use_mel": True,
|
||||
"sample_rate": 22050,
|
||||
"n_fft": 1024,
|
||||
"hop_length": 256,
|
||||
"win_length": 1024,
|
||||
"n_mels": 80,
|
||||
"mel_fmin": 0.0,
|
||||
"mel_fmax": None,
|
||||
}`
|
||||
target_loss (str):
|
||||
Target loss name that defines the quality of the model. Defaults to `G_avg_loss`.
|
||||
grad_clip (list):
|
||||
A list of gradient clipping theresholds for each optimizer. Any value less than 0 disables clipping.
|
||||
Defaults to [5, 5].
|
||||
lr_gen (float):
|
||||
Generator model initial learning rate. Defaults to 0.0002.
|
||||
lr_disc (float):
|
||||
Discriminator model initial learning rate. Defaults to 0.0002.
|
||||
lr_scheduler_gen (torch.optim.Scheduler):
|
||||
Learning rate scheduler for the generator. Defaults to `ExponentialLR`.
|
||||
lr_scheduler_gen_params (dict):
|
||||
Parameters for the generator learning rate scheduler. Defaults to `{"gamma": 0.999, "last_epoch": -1}`.
|
||||
lr_scheduler_disc (torch.optim.Scheduler):
|
||||
Learning rate scheduler for the discriminator. Defaults to `ExponentialLR`.
|
||||
lr_scheduler_disc_params (dict):
|
||||
Parameters for the discriminator learning rate scheduler. Defaults to `{"gamma": 0.999, "last_epoch": -1}`.
|
||||
scheduler_after_epoch (bool):
|
||||
Whether to update the learning rate schedulers after each epoch. Defaults to True.
|
||||
use_pqmf (bool):
|
||||
enable / disable PQMF for subband approximation at training. Defaults to False.
|
||||
steps_to_start_discriminator (int):
|
||||
Number of steps required to start training the discriminator. Defaults to 0.
|
||||
diff_samples_for_G_and_D (bool):
|
||||
enable / disable use of different training samples for the generator and the discriminator iterations.
|
||||
Enabling it results in slower iterations but faster convergance in some cases. Defaults to False.
|
||||
"""
|
||||
|
||||
model: str = "gan"
|
||||
|
||||
# LOSS PARAMETERS
|
||||
use_stft_loss: bool = True
|
||||
use_subband_stft_loss: bool = True
|
||||
use_mse_gan_loss: bool = True
|
||||
use_hinge_gan_loss: bool = True
|
||||
use_feat_match_loss: bool = True # requires MelGAN Discriminators (MelGAN and HifiGAN)
|
||||
use_l1_spec_loss: bool = True
|
||||
|
||||
# loss weights
|
||||
stft_loss_weight: float = 0
|
||||
subband_stft_loss_weight: float = 0
|
||||
mse_G_loss_weight: float = 1
|
||||
hinge_G_loss_weight: float = 0
|
||||
feat_match_loss_weight: float = 100
|
||||
l1_spec_loss_weight: float = 45
|
||||
|
||||
stft_loss_params: dict = field(
|
||||
default_factory=lambda: {
|
||||
"n_ffts": [1024, 2048, 512],
|
||||
"hop_lengths": [120, 240, 50],
|
||||
"win_lengths": [600, 1200, 240],
|
||||
}
|
||||
)
|
||||
|
||||
l1_spec_loss_params: dict = field(
|
||||
default_factory=lambda: {
|
||||
"use_mel": True,
|
||||
"sample_rate": 22050,
|
||||
"n_fft": 1024,
|
||||
"hop_length": 256,
|
||||
"win_length": 1024,
|
||||
"n_mels": 80,
|
||||
"mel_fmin": 0.0,
|
||||
"mel_fmax": None,
|
||||
}
|
||||
)
|
||||
|
||||
target_loss: str = "loss_0" # loss value to pick the best model to save after each epoch
|
||||
|
||||
# optimizer
|
||||
grad_clip: float = field(default_factory=lambda: [5, 5])
|
||||
lr_gen: float = 0.0002 # Initial learning rate.
|
||||
lr_disc: float = 0.0002 # Initial learning rate.
|
||||
lr_scheduler_gen: str = "ExponentialLR" # one of the schedulers from https:#pytorch.org/docs/stable/optim.html
|
||||
lr_scheduler_gen_params: dict = field(default_factory=lambda: {"gamma": 0.999, "last_epoch": -1})
|
||||
lr_scheduler_disc: str = "ExponentialLR" # one of the schedulers from https:#pytorch.org/docs/stable/optim.html
|
||||
lr_scheduler_disc_params: dict = field(default_factory=lambda: {"gamma": 0.999, "last_epoch": -1})
|
||||
scheduler_after_epoch: bool = True
|
||||
|
||||
use_pqmf: bool = False # enable/disable using pqmf for multi-band training. (Multi-band MelGAN)
|
||||
steps_to_start_discriminator = 0 # start training the discriminator after this number of steps.
|
||||
diff_samples_for_G_and_D: bool = False # use different samples for G and D training steps.
|
||||
@@ -0,0 +1,161 @@
|
||||
from dataclasses import dataclass, field
|
||||
from typing import Dict
|
||||
|
||||
from TTS.vocoder.configs.shared_configs import BaseGANVocoderConfig
|
||||
|
||||
|
||||
@dataclass
|
||||
class UnivnetConfig(BaseGANVocoderConfig):
|
||||
"""Defines parameters for UnivNet vocoder.
|
||||
|
||||
Example:
|
||||
|
||||
>>> from TTS.vocoder.configs import UnivNetConfig
|
||||
>>> config = UnivNetConfig()
|
||||
|
||||
Args:
|
||||
model (str):
|
||||
Model name used for selecting the right model at initialization. Defaults to `UnivNet`.
|
||||
discriminator_model (str): One of the discriminators from `TTS.vocoder.models.*_discriminator`. Defaults to
|
||||
'UnivNet_discriminator`.
|
||||
generator_model (str): One of the generators from TTS.vocoder.models.*`. Every other non-GAN vocoder model is
|
||||
considered as a generator too. Defaults to `UnivNet_generator`.
|
||||
generator_model_params (dict): Parameters of the generator model. Defaults to
|
||||
`
|
||||
{
|
||||
"use_mel": True,
|
||||
"sample_rate": 22050,
|
||||
"n_fft": 1024,
|
||||
"hop_length": 256,
|
||||
"win_length": 1024,
|
||||
"n_mels": 80,
|
||||
"mel_fmin": 0.0,
|
||||
"mel_fmax": None,
|
||||
}
|
||||
`
|
||||
batch_size (int):
|
||||
Batch size used at training. Larger values use more memory. Defaults to 32.
|
||||
seq_len (int):
|
||||
Audio segment length used at training. Larger values use more memory. Defaults to 8192.
|
||||
pad_short (int):
|
||||
Additional padding applied to the audio samples shorter than `seq_len`. Defaults to 0.
|
||||
use_noise_augment (bool):
|
||||
enable / disable random noise added to the input waveform. The noise is added after computing the
|
||||
features. Defaults to True.
|
||||
use_cache (bool):
|
||||
enable / disable in memory caching of the computed features. It can cause OOM error if the system RAM is
|
||||
not large enough. Defaults to True.
|
||||
use_stft_loss (bool):
|
||||
enable / disable use of STFT loss originally used by ParallelWaveGAN model. Defaults to True.
|
||||
use_subband_stft (bool):
|
||||
enable / disable use of subband loss computation originally used by MultiBandMelgan model. Defaults to True.
|
||||
use_mse_gan_loss (bool):
|
||||
enable / disable using Mean Squeare Error GAN loss. Defaults to True.
|
||||
use_hinge_gan_loss (bool):
|
||||
enable / disable using Hinge GAN loss. You should choose either Hinge or MSE loss for training GAN models.
|
||||
Defaults to False.
|
||||
use_feat_match_loss (bool):
|
||||
enable / disable using Feature Matching loss originally used by MelGAN model. Defaults to True.
|
||||
use_l1_spec_loss (bool):
|
||||
enable / disable using L1 spectrogram loss originally used by univnet model. Defaults to False.
|
||||
stft_loss_params (dict):
|
||||
STFT loss parameters. Default to
|
||||
`{
|
||||
"n_ffts": [1024, 2048, 512],
|
||||
"hop_lengths": [120, 240, 50],
|
||||
"win_lengths": [600, 1200, 240]
|
||||
}`
|
||||
l1_spec_loss_params (dict):
|
||||
L1 spectrogram loss parameters. Default to
|
||||
`{
|
||||
"use_mel": True,
|
||||
"sample_rate": 22050,
|
||||
"n_fft": 1024,
|
||||
"hop_length": 256,
|
||||
"win_length": 1024,
|
||||
"n_mels": 80,
|
||||
"mel_fmin": 0.0,
|
||||
"mel_fmax": None,
|
||||
}`
|
||||
stft_loss_weight (float): STFT loss weight that multiplies the computed loss before summing up the total
|
||||
model loss. Defaults to 0.5.
|
||||
subband_stft_loss_weight (float):
|
||||
Subband STFT loss weight that multiplies the computed loss before summing up the total loss. Defaults to 0.
|
||||
mse_G_loss_weight (float):
|
||||
MSE generator loss weight that multiplies the computed loss before summing up the total loss. faults to 2.5.
|
||||
hinge_G_loss_weight (float):
|
||||
Hinge generator loss weight that multiplies the computed loss before summing up the total loss. Defaults to 0.
|
||||
feat_match_loss_weight (float):
|
||||
Feature matching loss weight that multiplies the computed loss before summing up the total loss. faults to 108.
|
||||
l1_spec_loss_weight (float):
|
||||
L1 spectrogram loss weight that multiplies the computed loss before summing up the total loss. Defaults to 0.
|
||||
"""
|
||||
|
||||
model: str = "univnet"
|
||||
batch_size: int = 32
|
||||
# model specific params
|
||||
discriminator_model: str = "univnet_discriminator"
|
||||
generator_model: str = "univnet_generator"
|
||||
generator_model_params: Dict = field(
|
||||
default_factory=lambda: {
|
||||
"in_channels": 64,
|
||||
"out_channels": 1,
|
||||
"hidden_channels": 32,
|
||||
"cond_channels": 80,
|
||||
"upsample_factors": [8, 8, 4],
|
||||
"lvc_layers_each_block": 4,
|
||||
"lvc_kernel_size": 3,
|
||||
"kpnet_hidden_channels": 64,
|
||||
"kpnet_conv_size": 3,
|
||||
"dropout": 0.0,
|
||||
}
|
||||
)
|
||||
|
||||
# LOSS PARAMETERS - overrides
|
||||
use_stft_loss: bool = True
|
||||
use_subband_stft_loss: bool = False
|
||||
use_mse_gan_loss: bool = True
|
||||
use_hinge_gan_loss: bool = False
|
||||
use_feat_match_loss: bool = False # requires MelGAN Discriminators (MelGAN and univnet)
|
||||
use_l1_spec_loss: bool = False
|
||||
|
||||
# loss weights - overrides
|
||||
stft_loss_weight: float = 2.5
|
||||
stft_loss_params: Dict = field(
|
||||
default_factory=lambda: {
|
||||
"n_ffts": [1024, 2048, 512],
|
||||
"hop_lengths": [120, 240, 50],
|
||||
"win_lengths": [600, 1200, 240],
|
||||
}
|
||||
)
|
||||
subband_stft_loss_weight: float = 0
|
||||
mse_G_loss_weight: float = 1
|
||||
hinge_G_loss_weight: float = 0
|
||||
feat_match_loss_weight: float = 0
|
||||
l1_spec_loss_weight: float = 0
|
||||
l1_spec_loss_params: Dict = field(
|
||||
default_factory=lambda: {
|
||||
"use_mel": True,
|
||||
"sample_rate": 22050,
|
||||
"n_fft": 1024,
|
||||
"hop_length": 256,
|
||||
"win_length": 1024,
|
||||
"n_mels": 80,
|
||||
"mel_fmin": 0.0,
|
||||
"mel_fmax": None,
|
||||
}
|
||||
)
|
||||
|
||||
# optimizer parameters
|
||||
lr_gen: float = 1e-4 # Initial learning rate.
|
||||
lr_disc: float = 1e-4 # Initial learning rate.
|
||||
lr_scheduler_gen: str = None # one of the schedulers from https:#pytorch.org/docs/stable/optim.html
|
||||
# lr_scheduler_gen_params: dict = field(default_factory=lambda: {"gamma": 0.999, "last_epoch": -1})
|
||||
lr_scheduler_disc: str = None # one of the schedulers from https:#pytorch.org/docs/stable/optim.html
|
||||
# lr_scheduler_disc_params: dict = field(default_factory=lambda: {"gamma": 0.999, "last_epoch": -1})
|
||||
optimizer_params: Dict = field(default_factory=lambda: {"betas": [0.5, 0.9], "weight_decay": 0.0})
|
||||
steps_to_start_discriminator: int = 200000
|
||||
|
||||
def __post_init__(self):
|
||||
super().__post_init__()
|
||||
self.generator_model_params["cond_channels"] = self.audio.num_mels
|
||||
@@ -0,0 +1,90 @@
|
||||
from dataclasses import dataclass, field
|
||||
|
||||
from TTS.vocoder.configs.shared_configs import BaseVocoderConfig
|
||||
from TTS.vocoder.models.wavegrad import WavegradArgs
|
||||
|
||||
|
||||
@dataclass
|
||||
class WavegradConfig(BaseVocoderConfig):
|
||||
"""Defines parameters for WaveGrad vocoder.
|
||||
Example:
|
||||
|
||||
>>> from TTS.vocoder.configs import WavegradConfig
|
||||
>>> config = WavegradConfig()
|
||||
|
||||
Args:
|
||||
model (str):
|
||||
Model name used for selecting the right model at initialization. Defaults to `wavegrad`.
|
||||
generator_model (str): One of the generators from TTS.vocoder.models.*`. Every other non-GAN vocoder model is
|
||||
considered as a generator too. Defaults to `wavegrad`.
|
||||
model_params (WavegradArgs): Model parameters. Check `WavegradArgs` for default values.
|
||||
target_loss (str):
|
||||
Target loss name that defines the quality of the model. Defaults to `avg_wavegrad_loss`.
|
||||
epochs (int):
|
||||
Number of epochs to traing the model. Defaults to 10000.
|
||||
batch_size (int):
|
||||
Batch size used at training. Larger values use more memory. Defaults to 96.
|
||||
seq_len (int):
|
||||
Audio segment length used at training. Larger values use more memory. Defaults to 6144.
|
||||
use_cache (bool):
|
||||
enable / disable in memory caching of the computed features. It can cause OOM error if the system RAM is
|
||||
not large enough. Defaults to True.
|
||||
mixed_precision (bool):
|
||||
enable / disable mixed precision training. Default is True.
|
||||
eval_split_size (int):
|
||||
Number of samples used for evalutaion. Defaults to 50.
|
||||
train_noise_schedule (dict):
|
||||
Training noise schedule. Defaults to
|
||||
`{"min_val": 1e-6, "max_val": 1e-2, "num_steps": 1000}`
|
||||
test_noise_schedule (dict):
|
||||
Inference noise schedule. For a better performance, you may need to use `bin/tune_wavegrad.py` to find a
|
||||
better schedule. Defaults to
|
||||
`
|
||||
{
|
||||
"min_val": 1e-6,
|
||||
"max_val": 1e-2,
|
||||
"num_steps": 50,
|
||||
}
|
||||
`
|
||||
grad_clip (float):
|
||||
Gradient clipping threshold. If <= 0.0, no clipping is applied. Defaults to 1.0
|
||||
lr (float):
|
||||
Initila leraning rate. Defaults to 1e-4.
|
||||
lr_scheduler (str):
|
||||
One of the learning rate schedulers from `torch.optim.scheduler.*`. Defaults to `MultiStepLR`.
|
||||
lr_scheduler_params (dict):
|
||||
kwargs for the scheduler. Defaults to `{"gamma": 0.5, "milestones": [100000, 200000, 300000, 400000, 500000, 600000]}`
|
||||
"""
|
||||
|
||||
model: str = "wavegrad"
|
||||
# Model specific params
|
||||
generator_model: str = "wavegrad"
|
||||
model_params: WavegradArgs = field(default_factory=WavegradArgs)
|
||||
target_loss: str = "loss" # loss value to pick the best model to save after each epoch
|
||||
|
||||
# Training - overrides
|
||||
epochs: int = 10000
|
||||
batch_size: int = 96
|
||||
seq_len: int = 6144
|
||||
use_cache: bool = True
|
||||
mixed_precision: bool = True
|
||||
eval_split_size: int = 50
|
||||
|
||||
# NOISE SCHEDULE PARAMS
|
||||
train_noise_schedule: dict = field(default_factory=lambda: {"min_val": 1e-6, "max_val": 1e-2, "num_steps": 1000})
|
||||
|
||||
test_noise_schedule: dict = field(
|
||||
default_factory=lambda: { # inference noise schedule. Try TTS/bin/tune_wavegrad.py to find the optimal values.
|
||||
"min_val": 1e-6,
|
||||
"max_val": 1e-2,
|
||||
"num_steps": 50,
|
||||
}
|
||||
)
|
||||
|
||||
# optimizer overrides
|
||||
grad_clip: float = 1.0
|
||||
lr: float = 1e-4 # Initial learning rate.
|
||||
lr_scheduler: str = "MultiStepLR" # one of the schedulers from https:#pytorch.org/docs/stable/optim.html
|
||||
lr_scheduler_params: dict = field(
|
||||
default_factory=lambda: {"gamma": 0.5, "milestones": [100000, 200000, 300000, 400000, 500000, 600000]}
|
||||
)
|
||||
@@ -0,0 +1,102 @@
|
||||
from dataclasses import dataclass, field
|
||||
|
||||
from TTS.vocoder.configs.shared_configs import BaseVocoderConfig
|
||||
from TTS.vocoder.models.wavernn import WavernnArgs
|
||||
|
||||
|
||||
@dataclass
|
||||
class WavernnConfig(BaseVocoderConfig):
|
||||
"""Defines parameters for Wavernn vocoder.
|
||||
Example:
|
||||
|
||||
>>> from TTS.vocoder.configs import WavernnConfig
|
||||
>>> config = WavernnConfig()
|
||||
|
||||
Args:
|
||||
model (str):
|
||||
Model name used for selecting the right model at initialization. Defaults to `wavernn`.
|
||||
mode (str):
|
||||
Output mode of the WaveRNN vocoder. `mold` for Mixture of Logistic Distribution, `gauss` for a single
|
||||
Gaussian Distribution and `bits` for quantized bits as the model's output.
|
||||
mulaw (bool):
|
||||
enable / disable the use of Mulaw quantization for training. Only applicable if `mode == 'bits'`. Defaults
|
||||
to `True`.
|
||||
generator_model (str):
|
||||
One of the generators from TTS.vocoder.models.*`. Every other non-GAN vocoder model is
|
||||
considered as a generator too. Defaults to `WaveRNN`.
|
||||
wavernn_model_params (dict):
|
||||
kwargs for the WaveRNN model. Defaults to
|
||||
`{
|
||||
"rnn_dims": 512,
|
||||
"fc_dims": 512,
|
||||
"compute_dims": 128,
|
||||
"res_out_dims": 128,
|
||||
"num_res_blocks": 10,
|
||||
"use_aux_net": True,
|
||||
"use_upsample_net": True,
|
||||
"upsample_factors": [4, 8, 8]
|
||||
}`
|
||||
batched (bool):
|
||||
enable / disable the batched inference. It speeds up the inference by splitting the input into segments and
|
||||
processing the segments in a batch. Then it merges the outputs with a certain overlap and smoothing. If
|
||||
you set it False, without CUDA, it is too slow to be practical. Defaults to True.
|
||||
target_samples (int):
|
||||
Size of the segments in batched mode. Defaults to 11000.
|
||||
overlap_sampels (int):
|
||||
Size of the overlap between consecutive segments. Defaults to 550.
|
||||
batch_size (int):
|
||||
Batch size used at training. Larger values use more memory. Defaults to 256.
|
||||
seq_len (int):
|
||||
Audio segment length used at training. Larger values use more memory. Defaults to 1280.
|
||||
|
||||
use_noise_augment (bool):
|
||||
enable / disable random noise added to the input waveform. The noise is added after computing the
|
||||
features. Defaults to True.
|
||||
use_cache (bool):
|
||||
enable / disable in memory caching of the computed features. It can cause OOM error if the system RAM is
|
||||
not large enough. Defaults to True.
|
||||
mixed_precision (bool):
|
||||
enable / disable mixed precision training. Default is True.
|
||||
eval_split_size (int):
|
||||
Number of samples used for evalutaion. Defaults to 50.
|
||||
num_epochs_before_test (int):
|
||||
Number of epochs waited to run the next evalution. Since inference takes some time, it is better to
|
||||
wait some number of epochs not ot waste training time. Defaults to 10.
|
||||
grad_clip (float):
|
||||
Gradient clipping threshold. If <= 0.0, no clipping is applied. Defaults to 4.0
|
||||
lr (float):
|
||||
Initila leraning rate. Defaults to 1e-4.
|
||||
lr_scheduler (str):
|
||||
One of the learning rate schedulers from `torch.optim.scheduler.*`. Defaults to `MultiStepLR`.
|
||||
lr_scheduler_params (dict):
|
||||
kwargs for the scheduler. Defaults to `{"gamma": 0.5, "milestones": [200000, 400000, 600000]}`
|
||||
"""
|
||||
|
||||
model: str = "wavernn"
|
||||
|
||||
# Model specific params
|
||||
model_args: WavernnArgs = field(default_factory=WavernnArgs)
|
||||
target_loss: str = "loss"
|
||||
|
||||
# Inference
|
||||
batched: bool = True
|
||||
target_samples: int = 11000
|
||||
overlap_samples: int = 550
|
||||
|
||||
# Training - overrides
|
||||
epochs: int = 10000
|
||||
batch_size: int = 256
|
||||
seq_len: int = 1280
|
||||
use_noise_augment: bool = False
|
||||
use_cache: bool = True
|
||||
mixed_precision: bool = True
|
||||
eval_split_size: int = 50
|
||||
num_epochs_before_test: int = (
|
||||
10 # number of epochs to wait until the next test run (synthesizing a full audio clip).
|
||||
)
|
||||
|
||||
# optimizer overrides
|
||||
grad_clip: float = 4.0
|
||||
lr: float = 1e-4 # Initial learning rate.
|
||||
lr_scheduler: str = "MultiStepLR" # one of the schedulers from https:#pytorch.org/docs/stable/optim.html
|
||||
lr_scheduler_params: dict = field(default_factory=lambda: {"gamma": 0.5, "milestones": [200000, 400000, 600000]})
|
||||
Reference in New Issue
Block a user