Add files via upload
This commit is contained in:
Sam Khoze
Binary file not shown.
@@ -0,0 +1,278 @@
|
||||
from dataclasses import dataclass, field
|
||||
from typing import List, Optional
|
||||
|
||||
from coqpit import Coqpit
|
||||
|
||||
from TTS.vc.configs.shared_configs import BaseVCConfig
|
||||
|
||||
|
||||
@dataclass
|
||||
class FreeVCAudioConfig(Coqpit):
|
||||
"""Audio configuration
|
||||
|
||||
Args:
|
||||
max_wav_value (float):
|
||||
The maximum value of the waveform.
|
||||
|
||||
input_sample_rate (int):
|
||||
The sampling rate of the input waveform.
|
||||
|
||||
output_sample_rate (int):
|
||||
The sampling rate of the output waveform.
|
||||
|
||||
filter_length (int):
|
||||
The length of the filter.
|
||||
|
||||
hop_length (int):
|
||||
The hop length.
|
||||
|
||||
win_length (int):
|
||||
The window length.
|
||||
|
||||
n_mel_channels (int):
|
||||
The number of mel channels.
|
||||
|
||||
mel_fmin (float):
|
||||
The minimum frequency of the mel filterbank.
|
||||
|
||||
mel_fmax (Optional[float]):
|
||||
The maximum frequency of the mel filterbank.
|
||||
"""
|
||||
|
||||
max_wav_value: float = field(default=32768.0)
|
||||
input_sample_rate: int = field(default=16000)
|
||||
output_sample_rate: int = field(default=24000)
|
||||
filter_length: int = field(default=1280)
|
||||
hop_length: int = field(default=320)
|
||||
win_length: int = field(default=1280)
|
||||
n_mel_channels: int = field(default=80)
|
||||
mel_fmin: float = field(default=0.0)
|
||||
mel_fmax: Optional[float] = field(default=None)
|
||||
|
||||
|
||||
@dataclass
|
||||
class FreeVCArgs(Coqpit):
|
||||
"""FreeVC model arguments
|
||||
|
||||
Args:
|
||||
spec_channels (int):
|
||||
The number of channels in the spectrogram.
|
||||
|
||||
inter_channels (int):
|
||||
The number of channels in the intermediate layers.
|
||||
|
||||
hidden_channels (int):
|
||||
The number of channels in the hidden layers.
|
||||
|
||||
filter_channels (int):
|
||||
The number of channels in the filter layers.
|
||||
|
||||
n_heads (int):
|
||||
The number of attention heads.
|
||||
|
||||
n_layers (int):
|
||||
The number of layers.
|
||||
|
||||
kernel_size (int):
|
||||
The size of the kernel.
|
||||
|
||||
p_dropout (float):
|
||||
The dropout probability.
|
||||
|
||||
resblock (str):
|
||||
The type of residual block.
|
||||
|
||||
resblock_kernel_sizes (List[int]):
|
||||
The kernel sizes for the residual blocks.
|
||||
|
||||
resblock_dilation_sizes (List[List[int]]):
|
||||
The dilation sizes for the residual blocks.
|
||||
|
||||
upsample_rates (List[int]):
|
||||
The upsample rates.
|
||||
|
||||
upsample_initial_channel (int):
|
||||
The number of channels in the initial upsample layer.
|
||||
|
||||
upsample_kernel_sizes (List[int]):
|
||||
The kernel sizes for the upsample layers.
|
||||
|
||||
n_layers_q (int):
|
||||
The number of layers in the quantization network.
|
||||
|
||||
use_spectral_norm (bool):
|
||||
Whether to use spectral normalization.
|
||||
|
||||
gin_channels (int):
|
||||
The number of channels in the global conditioning vector.
|
||||
|
||||
ssl_dim (int):
|
||||
The dimension of the self-supervised learning embedding.
|
||||
|
||||
use_spk (bool):
|
||||
Whether to use external speaker encoder.
|
||||
"""
|
||||
|
||||
spec_channels: int = field(default=641)
|
||||
inter_channels: int = field(default=192)
|
||||
hidden_channels: int = field(default=192)
|
||||
filter_channels: int = field(default=768)
|
||||
n_heads: int = field(default=2)
|
||||
n_layers: int = field(default=6)
|
||||
kernel_size: int = field(default=3)
|
||||
p_dropout: float = field(default=0.1)
|
||||
resblock: str = field(default="1")
|
||||
resblock_kernel_sizes: List[int] = field(default_factory=lambda: [3, 7, 11])
|
||||
resblock_dilation_sizes: List[List[int]] = field(default_factory=lambda: [[1, 3, 5], [1, 3, 5], [1, 3, 5]])
|
||||
upsample_rates: List[int] = field(default_factory=lambda: [10, 8, 2, 2])
|
||||
upsample_initial_channel: int = field(default=512)
|
||||
upsample_kernel_sizes: List[int] = field(default_factory=lambda: [16, 16, 4, 4])
|
||||
n_layers_q: int = field(default=3)
|
||||
use_spectral_norm: bool = field(default=False)
|
||||
gin_channels: int = field(default=256)
|
||||
ssl_dim: int = field(default=1024)
|
||||
use_spk: bool = field(default=False)
|
||||
num_spks: int = field(default=0)
|
||||
segment_size: int = field(default=8960)
|
||||
|
||||
|
||||
@dataclass
|
||||
class FreeVCConfig(BaseVCConfig):
|
||||
"""Defines parameters for FreeVC End2End TTS model.
|
||||
|
||||
Args:
|
||||
model (str):
|
||||
Model name. Do not change unless you know what you are doing.
|
||||
|
||||
model_args (FreeVCArgs):
|
||||
Model architecture arguments. Defaults to `FreeVCArgs()`.
|
||||
|
||||
audio (FreeVCAudioConfig):
|
||||
Audio processing configuration. Defaults to `FreeVCAudioConfig()`.
|
||||
|
||||
grad_clip (List):
|
||||
Gradient clipping thresholds for each optimizer. Defaults to `[1000.0, 1000.0]`.
|
||||
|
||||
lr_gen (float):
|
||||
Initial learning rate for the generator. Defaults to 0.0002.
|
||||
|
||||
lr_disc (float):
|
||||
Initial learning rate for the discriminator. Defaults to 0.0002.
|
||||
|
||||
lr_scheduler_gen (str):
|
||||
Name of the learning rate scheduler for the generator. One of the `torch.optim.lr_scheduler.*`. Defaults to
|
||||
`ExponentialLR`.
|
||||
|
||||
lr_scheduler_gen_params (dict):
|
||||
Parameters for the learning rate scheduler of the generator. Defaults to `{'gamma': 0.999875, "last_epoch":-1}`.
|
||||
|
||||
lr_scheduler_disc (str):
|
||||
Name of the learning rate scheduler for the discriminator. One of the `torch.optim.lr_scheduler.*`. Defaults to
|
||||
`ExponentialLR`.
|
||||
|
||||
lr_scheduler_disc_params (dict):
|
||||
Parameters for the learning rate scheduler of the discriminator. Defaults to `{'gamma': 0.999875, "last_epoch":-1}`.
|
||||
|
||||
scheduler_after_epoch (bool):
|
||||
If true, step the schedulers after each epoch else after each step. Defaults to `False`.
|
||||
|
||||
optimizer (str):
|
||||
Name of the optimizer to use with both the generator and the discriminator networks. One of the
|
||||
`torch.optim.*`. Defaults to `AdamW`.
|
||||
|
||||
kl_loss_alpha (float):
|
||||
Loss weight for KL loss. Defaults to 1.0.
|
||||
|
||||
disc_loss_alpha (float):
|
||||
Loss weight for the discriminator loss. Defaults to 1.0.
|
||||
|
||||
gen_loss_alpha (float):
|
||||
Loss weight for the generator loss. Defaults to 1.0.
|
||||
|
||||
feat_loss_alpha (float):
|
||||
Loss weight for the feature matching loss. Defaults to 1.0.
|
||||
|
||||
mel_loss_alpha (float):
|
||||
Loss weight for the mel loss. Defaults to 45.0.
|
||||
|
||||
return_wav (bool):
|
||||
If true, data loader returns the waveform as well as the other outputs. Do not change. Defaults to `True`.
|
||||
|
||||
compute_linear_spec (bool):
|
||||
If true, the linear spectrogram is computed and returned alongside the mel output. Do not change. Defaults to `True`.
|
||||
|
||||
use_weighted_sampler (bool):
|
||||
If true, use weighted sampler with bucketing for balancing samples between datasets used in training. Defaults to `False`.
|
||||
|
||||
weighted_sampler_attrs (dict):
|
||||
Key retuned by the formatter to be used for weighted sampler. For example `{"root_path": 2.0, "speaker_name": 1.0}` sets sample probabilities
|
||||
by overweighting `root_path` by 2.0. Defaults to `{}`.
|
||||
|
||||
weighted_sampler_multipliers (dict):
|
||||
Weight each unique value of a key returned by the formatter for weighted sampling.
|
||||
For example `{"root_path":{"/raid/datasets/libritts-clean-16khz-bwe-coqui_44khz/LibriTTS/train-clean-100/":1.0, "/raid/datasets/libritts-clean-16khz-bwe-coqui_44khz/LibriTTS/train-clean-360/": 0.5}`.
|
||||
It will sample instances from `train-clean-100` 2 times more than `train-clean-360`. Defaults to `{}`.
|
||||
|
||||
r (int):
|
||||
Number of spectrogram frames to be generated at a time. Do not change. Defaults to `1`.
|
||||
|
||||
add_blank (bool):
|
||||
If true, a blank token is added in between every character. Defaults to `True`.
|
||||
|
||||
test_sentences (List[List]):
|
||||
List of sentences with speaker and language information to be used for testing.
|
||||
|
||||
language_ids_file (str):
|
||||
Path to the language ids file.
|
||||
|
||||
use_language_embedding (bool):
|
||||
If true, language embedding is used. Defaults to `False`.
|
||||
|
||||
Note:
|
||||
Check :class:`TTS.tts.configs.shared_configs.BaseTTSConfig` for the inherited parameters.
|
||||
|
||||
Example:
|
||||
|
||||
>>> from TTS.vc.configs.freevc_config import FreeVCConfig
|
||||
>>> config = FreeVCConfig()
|
||||
"""
|
||||
|
||||
model: str = "freevc"
|
||||
# model specific params
|
||||
model_args: FreeVCArgs = field(default_factory=FreeVCArgs)
|
||||
audio: FreeVCAudioConfig = field(default_factory=FreeVCAudioConfig)
|
||||
|
||||
# optimizer
|
||||
# TODO with training support
|
||||
|
||||
# loss params
|
||||
# TODO with training support
|
||||
|
||||
# data loader params
|
||||
return_wav: bool = True
|
||||
compute_linear_spec: bool = True
|
||||
|
||||
# sampler params
|
||||
use_weighted_sampler: bool = False # TODO: move it to the base config
|
||||
weighted_sampler_attrs: dict = field(default_factory=lambda: {})
|
||||
weighted_sampler_multipliers: dict = field(default_factory=lambda: {})
|
||||
|
||||
# overrides
|
||||
r: int = 1 # DO NOT CHANGE
|
||||
add_blank: bool = True
|
||||
|
||||
# multi-speaker settings
|
||||
# use speaker embedding layer
|
||||
num_speakers: int = 0
|
||||
speakers_file: str = None
|
||||
speaker_embedding_channels: int = 256
|
||||
|
||||
# use d-vectors
|
||||
use_d_vector_file: bool = False
|
||||
d_vector_file: List[str] = None
|
||||
d_vector_dim: int = None
|
||||
|
||||
def __post_init__(self):
|
||||
for key, val in self.model_args.items():
|
||||
if hasattr(self, key):
|
||||
self[key] = val
|
||||
@@ -0,0 +1,155 @@
|
||||
from dataclasses import asdict, dataclass, field
|
||||
from typing import Dict, List
|
||||
|
||||
from coqpit import Coqpit, check_argument
|
||||
|
||||
from TTS.config import BaseAudioConfig, BaseDatasetConfig, BaseTrainingConfig
|
||||
|
||||
|
||||
@dataclass
|
||||
class BaseVCConfig(BaseTrainingConfig):
|
||||
"""Shared parameters among all the tts models.
|
||||
|
||||
Args:
|
||||
|
||||
audio (BaseAudioConfig):
|
||||
Audio processor config object instance.
|
||||
|
||||
batch_group_size (int):
|
||||
Size of the batch groups used for bucketing. By default, the dataloader orders samples by the sequence
|
||||
length for a more efficient and stable training. If `batch_group_size > 1` then it performs bucketing to
|
||||
prevent using the same batches for each epoch.
|
||||
|
||||
loss_masking (bool):
|
||||
enable / disable masking loss values against padded segments of samples in a batch.
|
||||
|
||||
min_text_len (int):
|
||||
Minimum length of input text to be used. All shorter samples will be ignored. Defaults to 0.
|
||||
|
||||
max_text_len (int):
|
||||
Maximum length of input text to be used. All longer samples will be ignored. Defaults to float("inf").
|
||||
|
||||
min_audio_len (int):
|
||||
Minimum length of input audio to be used. All shorter samples will be ignored. Defaults to 0.
|
||||
|
||||
max_audio_len (int):
|
||||
Maximum length of input audio to be used. All longer samples will be ignored. The maximum length in the
|
||||
dataset defines the VRAM used in the training. Hence, pay attention to this value if you encounter an
|
||||
OOM error in training. Defaults to float("inf").
|
||||
|
||||
compute_f0 (int):
|
||||
(Not in use yet).
|
||||
|
||||
compute_energy (int):
|
||||
(Not in use yet).
|
||||
|
||||
compute_linear_spec (bool):
|
||||
If True data loader computes and returns linear spectrograms alongside the other data.
|
||||
|
||||
precompute_num_workers (int):
|
||||
Number of workers to precompute features. Defaults to 0.
|
||||
|
||||
use_noise_augment (bool):
|
||||
Augment the input audio with random noise.
|
||||
|
||||
start_by_longest (bool):
|
||||
If True, the data loader will start loading the longest batch first. It is useful for checking OOM issues.
|
||||
Defaults to False.
|
||||
|
||||
shuffle (bool):
|
||||
If True, the data loader will shuffle the dataset when there is not sampler defined. Defaults to True.
|
||||
|
||||
drop_last (bool):
|
||||
If True, the data loader will drop the last batch if it is not complete. It helps to prevent
|
||||
issues that emerge from the partial batch statistics. Defaults to True.
|
||||
|
||||
add_blank (bool):
|
||||
Add blank characters between each other two characters. It improves performance for some models at expense
|
||||
of slower run-time due to the longer input sequence.
|
||||
|
||||
datasets (List[BaseDatasetConfig]):
|
||||
List of datasets used for training. If multiple datasets are provided, they are merged and used together
|
||||
for training.
|
||||
|
||||
optimizer (str):
|
||||
Optimizer used for the training. Set one from `torch.optim.Optimizer` or `TTS.utils.training`.
|
||||
Defaults to ``.
|
||||
|
||||
optimizer_params (dict):
|
||||
Optimizer kwargs. Defaults to `{"betas": [0.8, 0.99], "weight_decay": 0.0}`
|
||||
|
||||
lr_scheduler (str):
|
||||
Learning rate scheduler for the training. Use one from `torch.optim.Scheduler` schedulers or
|
||||
`TTS.utils.training`. Defaults to ``.
|
||||
|
||||
lr_scheduler_params (dict):
|
||||
Parameters for the generator learning rate scheduler. Defaults to `{"warmup": 4000}`.
|
||||
|
||||
test_sentences (List[str]):
|
||||
List of sentences to be used at testing. Defaults to '[]'
|
||||
|
||||
eval_split_max_size (int):
|
||||
Number maximum of samples to be used for evaluation in proportion split. Defaults to None (Disabled).
|
||||
|
||||
eval_split_size (float):
|
||||
If between 0.0 and 1.0 represents the proportion of the dataset to include in the evaluation set.
|
||||
If > 1, represents the absolute number of evaluation samples. Defaults to 0.01 (1%).
|
||||
|
||||
use_speaker_weighted_sampler (bool):
|
||||
Enable / Disable the batch balancer by speaker. Defaults to ```False```.
|
||||
|
||||
speaker_weighted_sampler_alpha (float):
|
||||
Number that control the influence of the speaker sampler weights. Defaults to ```1.0```.
|
||||
|
||||
use_language_weighted_sampler (bool):
|
||||
Enable / Disable the batch balancer by language. Defaults to ```False```.
|
||||
|
||||
language_weighted_sampler_alpha (float):
|
||||
Number that control the influence of the language sampler weights. Defaults to ```1.0```.
|
||||
|
||||
use_length_weighted_sampler (bool):
|
||||
Enable / Disable the batch balancer by audio length. If enabled the dataset will be divided
|
||||
into 10 buckets considering the min and max audio of the dataset. The sampler weights will be
|
||||
computed forcing to have the same quantity of data for each bucket in each training batch. Defaults to ```False```.
|
||||
|
||||
length_weighted_sampler_alpha (float):
|
||||
Number that control the influence of the length sampler weights. Defaults to ```1.0```.
|
||||
"""
|
||||
|
||||
audio: BaseAudioConfig = field(default_factory=BaseAudioConfig)
|
||||
# training params
|
||||
batch_group_size: int = 0
|
||||
loss_masking: bool = None
|
||||
# dataloading
|
||||
min_audio_len: int = 1
|
||||
max_audio_len: int = float("inf")
|
||||
min_text_len: int = 1
|
||||
max_text_len: int = float("inf")
|
||||
compute_f0: bool = False
|
||||
compute_energy: bool = False
|
||||
compute_linear_spec: bool = False
|
||||
precompute_num_workers: int = 0
|
||||
use_noise_augment: bool = False
|
||||
start_by_longest: bool = False
|
||||
shuffle: bool = False
|
||||
drop_last: bool = False
|
||||
# dataset
|
||||
datasets: List[BaseDatasetConfig] = field(default_factory=lambda: [BaseDatasetConfig()])
|
||||
# optimizer
|
||||
optimizer: str = "radam"
|
||||
optimizer_params: dict = None
|
||||
# scheduler
|
||||
lr_scheduler: str = None
|
||||
lr_scheduler_params: dict = field(default_factory=lambda: {})
|
||||
# testing
|
||||
test_sentences: List[str] = field(default_factory=lambda: [])
|
||||
# evaluation
|
||||
eval_split_max_size: int = None
|
||||
eval_split_size: float = 0.01
|
||||
# weighted samplers
|
||||
use_speaker_weighted_sampler: bool = False
|
||||
speaker_weighted_sampler_alpha: float = 1.0
|
||||
use_language_weighted_sampler: bool = False
|
||||
language_weighted_sampler_alpha: float = 1.0
|
||||
use_length_weighted_sampler: bool = False
|
||||
length_weighted_sampler_alpha: float = 1.0
|
||||
Reference in New Issue
Block a user