| """ |
| WavCoch configuration for Hugging Face Transformers. |
| """ |
|
|
| from transformers import PretrainedConfig |
|
|
|
|
| class WavCochConfig(PretrainedConfig): |
| """Configuration class for WavCoch checkpoints with optional vocoder.""" |
|
|
| model_type = "wavcoch" |
|
|
| def __init__( |
| self, |
| window_size: int = 1001, |
| window_padding: int = 1000, |
| centered_left_padding: int = None, |
| centered_right_padding: int = None, |
| hop_length: int = 80, |
| out_channels: int = 211, |
| causal_convs: bool = True, |
| causal_pad_mode: str = "repeat", |
| encoder_layers: int = 8, |
| encoder_dim: int = 512, |
| encoder_kernel_size: int = 3, |
| decoder_layers: int = 8, |
| decoder_dim: int = 512, |
| decoder_kernel_size: int = 9, |
| quantizer: str = "FSQ", |
| channels=None, |
| vocab_size: int = None, |
| sample_rate: int = 16000, |
| has_vocoder: bool = False, |
| vocoder_upsample_rates=None, |
| vocoder_upsample_kernel_sizes=None, |
| vocoder_upsample_initial_channel: int = 512, |
| vocoder_resblock: str = "1", |
| vocoder_resblock_kernel_sizes=None, |
| vocoder_resblock_dilation_sizes=None, |
| **kwargs, |
| ): |
| channels = list(channels or [8, 8, 8, 4, 4]) |
| if vocab_size is None: |
| vocab_size = 1 |
| for level in channels: |
| vocab_size *= int(level) |
|
|
| self.window_size = int(window_size) |
| self.window_padding = int(window_padding) |
| half_window = (self.window_size - 1) // 2 |
| self.centered_right_padding = int( |
| half_window if centered_right_padding is None else centered_right_padding |
| ) |
| expected_left_padding = self.window_padding - self.centered_right_padding |
| self.centered_left_padding = int( |
| expected_left_padding if centered_left_padding is None else centered_left_padding |
| ) |
| if self.centered_left_padding + self.centered_right_padding != self.window_padding: |
| raise ValueError( |
| "Centered left/right padding must sum to window_padding so encoding length stays unchanged" |
| ) |
| self.hop_length = int(hop_length) |
| self.out_channels = int(out_channels) |
| self.causal_convs = bool(causal_convs) |
| self.causal_pad_mode = str(causal_pad_mode) |
| self.encoder_layers = int(encoder_layers) |
| self.encoder_dim = int(encoder_dim) |
| self.encoder_kernel_size = int(encoder_kernel_size) |
| self.decoder_layers = int(decoder_layers) |
| self.decoder_dim = int(decoder_dim) |
| self.decoder_kernel_size = int(decoder_kernel_size) |
| self.quantizer = str(quantizer) |
| self.channels = channels |
| self.vocab_size = int(vocab_size) |
| self.sample_rate = int(sample_rate) |
|
|
| self.has_vocoder = bool(has_vocoder) |
| self.vocoder_upsample_rates = list(vocoder_upsample_rates or [5, 4, 2, 2]) |
| self.vocoder_upsample_kernel_sizes = list(vocoder_upsample_kernel_sizes or [10, 8, 4, 4]) |
| self.vocoder_upsample_initial_channel = int(vocoder_upsample_initial_channel) |
| self.vocoder_resblock = str(vocoder_resblock) |
| self.vocoder_resblock_kernel_sizes = list(vocoder_resblock_kernel_sizes or [11, 7, 3]) |
| self.vocoder_resblock_dilation_sizes = [ |
| list(d) for d in (vocoder_resblock_dilation_sizes or [[1, 3, 5], [1, 3, 5], [1, 3, 5]]) |
| ] |
|
|
| super().__init__(**kwargs) |
|
|
