pretrained.encodec

Defines a simple API for interacting with Meta’s pretrained Encodec model.

This API only supports the causal version of the model, which is better suited for streaming applications.

from pretrained.encodec import pretrained_encodec

model = pretrained_encodec("24khz")
encoder, decoder = model.get_encoder(), model.get_decoder()

# Get the tokens for a waveform.
tokens = encoder.encode(torch.randn(1, 24_000))

# Reconstructs the waveform from the tokens.
reconstructed = decoder.decode(tokens)

pretrained.encodec.cast_pretrained_encodec_type(s: str) → Literal['24khz']

pretrained.encodec.get_extra_padding_for_conv1d(x: Tensor, kernel_size: int, stride: int, padding_total: int = 0) → int

pretrained.encodec.pad1d(x: Tensor, paddings: tuple[int, int], mode: Literal['reflect', 'replicate', 'circular', 'constant'] = 'constant', value: float = 0.0) → Tensor

pretrained.encodec.unpad1d(x: Tensor, paddings: tuple[int, int]) → Tensor

class pretrained.encodec.NormConv1d(conv: Conv1d, norm: Literal['no_norm', 'batch', 'batch_affine', 'instance', 'instance_affine', 'group', 'group_affine', 'layer', 'layer_affine'] | Literal['no_norm', 'weight', 'spectral'], *, groups: int = 1)

Bases: Module

Initializes internal Module state, shared by both nn.Module and ScriptModule.

forward(x: Tensor) → Tensor

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class pretrained.encodec.NormConvTranspose1d(conv: ConvTranspose1d, norm: Literal['no_norm', 'batch', 'batch_affine', 'instance', 'instance_affine', 'group', 'group_affine', 'layer', 'layer_affine'] | Literal['no_norm', 'weight', 'spectral'], *, groups: int = 1)

Bases: Module

Initializes internal Module state, shared by both nn.Module and ScriptModule.

forward(x: Tensor) → Tensor

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class pretrained.encodec.SConv1d(in_channels: int, out_channels: int, kernel_size: int, stride: int = 1, dilation: int = 1, groups: int = 1, bias: bool = True, causal: bool = False, norm: Literal['no_norm', 'batch', 'batch_affine', 'instance', 'instance_affine', 'group', 'group_affine', 'layer', 'layer_affine'] | Literal['no_norm', 'weight', 'spectral'] = 'no_norm', norm_groups: int = 1, pad_mode: Literal['reflect', 'replicate', 'circular', 'constant'] = 'reflect')

Bases: Module

Initializes internal Module state, shared by both nn.Module and ScriptModule.

forward(x: Tensor) → Tensor

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class pretrained.encodec.SConvTranspose1d(in_channels: int, out_channels: int, kernel_size: int, stride: int = 1, causal: bool = False, norm: Literal['no_norm', 'batch', 'batch_affine', 'instance', 'instance_affine', 'group', 'group_affine', 'layer', 'layer_affine'] | Literal['no_norm', 'weight', 'spectral'] = 'no_norm', norm_groups: int = 1, trim_right_ratio: float = 1.0)

Bases: Module

Initializes internal Module state, shared by both nn.Module and ScriptModule.

forward(x: Tensor) → Tensor

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class pretrained.encodec.SLSTM(dimension: int, num_layers: int = 2, skip: bool = True)

Bases: Module

Initializes internal Module state, shared by both nn.Module and ScriptModule.

forward(x: Tensor) → Tensor

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class pretrained.encodec.SEANetResnetBlock(dim: int, kernel_sizes: Sequence[int] = [3, 1], dilations: Sequence[int] = [1, 1], activation: Literal['no_act', 'relu', 'relu6', 'relu2', 'clamp6', 'leaky_relu', 'elu', 'celu', 'selu', 'gelu', 'gelu_fast', 'sigmoid', 'log_sigmoid', 'hard_sigomid', 'tanh', 'softsign', 'softplus', 'silu', 'mish', 'swish', 'hard_swish', 'soft_shrink', 'hard_shrink', 'tanh_shrink', 'soft_sign', 'relu_squared', 'laplace'] = 'elu', norm: Literal['no_norm', 'batch', 'batch_affine', 'instance', 'instance_affine', 'group', 'group_affine', 'layer', 'layer_affine'] | Literal['no_norm', 'weight', 'spectral'] = 'weight', causal: bool = False, pad_mode: Literal['reflect', 'replicate', 'circular', 'constant'] = 'reflect', compress: int = 2, true_skip: bool = True)

Bases: Module

Initializes internal Module state, shared by both nn.Module and ScriptModule.

forward(x: Tensor) → Tensor

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class pretrained.encodec.SEANetEncoder(channels: int = 1, dimension: int = 128, n_filters: int = 32, n_residual_layers: int = 1, ratios: Sequence[int] = [8, 5, 4, 2], activation: Literal['no_act', 'relu', 'relu6', 'relu2', 'clamp6', 'leaky_relu', 'elu', 'celu', 'selu', 'gelu', 'gelu_fast', 'sigmoid', 'log_sigmoid', 'hard_sigomid', 'tanh', 'softsign', 'softplus', 'silu', 'mish', 'swish', 'hard_swish', 'soft_shrink', 'hard_shrink', 'tanh_shrink', 'soft_sign', 'relu_squared', 'laplace'] = 'elu', norm: Literal['no_norm', 'batch', 'batch_affine', 'instance', 'instance_affine', 'group', 'group_affine', 'layer', 'layer_affine'] | Literal['no_norm', 'weight', 'spectral'] = 'weight', kernel_size: int = 7, last_kernel_size: int = 7, residual_kernel_size: int = 3, dilation_base: int = 2, causal: bool = False, pad_mode: Literal['reflect', 'replicate', 'circular', 'constant'] = 'reflect', true_skip: bool = False, compress: int = 2, lstm: int = 2)

Bases: Module

Initializes internal Module state, shared by both nn.Module and ScriptModule.

forward(x: Tensor) → Tensor

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class pretrained.encodec.SEANetDecoder(channels: int = 1, dimension: int = 128, n_filters: int = 32, n_residual_layers: int = 1, ratios: Sequence[int] = [8, 5, 4, 2], activation: Literal['no_act', 'relu', 'relu6', 'relu2', 'clamp6', 'leaky_relu', 'elu', 'celu', 'selu', 'gelu', 'gelu_fast', 'sigmoid', 'log_sigmoid', 'hard_sigomid', 'tanh', 'softsign', 'softplus', 'silu', 'mish', 'swish', 'hard_swish', 'soft_shrink', 'hard_shrink', 'tanh_shrink', 'soft_sign', 'relu_squared', 'laplace'] = 'elu', final_activation: Literal['no_act', 'relu', 'relu6', 'relu2', 'clamp6', 'leaky_relu', 'elu', 'celu', 'selu', 'gelu', 'gelu_fast', 'sigmoid', 'log_sigmoid', 'hard_sigomid', 'tanh', 'softsign', 'softplus', 'silu', 'mish', 'swish', 'hard_swish', 'soft_shrink', 'hard_shrink', 'tanh_shrink', 'soft_sign', 'relu_squared', 'laplace'] = 'no_act', norm: Literal['no_norm', 'batch', 'batch_affine', 'instance', 'instance_affine', 'group', 'group_affine', 'layer', 'layer_affine'] | Literal['no_norm', 'weight', 'spectral'] = 'weight', kernel_size: int = 7, last_kernel_size: int = 7, residual_kernel_size: int = 3, dilation_base: int = 2, causal: bool = False, pad_mode: Literal['reflect', 'replicate', 'circular', 'constant'] = 'reflect', true_skip: bool = False, compress: int = 2, lstm: int = 2, trim_right_ratio: float = 1.0)

Bases: Module

Initializes internal Module state, shared by both nn.Module and ScriptModule.

forward(z: Tensor) → Tensor

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class pretrained.encodec.Encodec(encoder: SEANetEncoder, decoder: SEANetDecoder, quantizer: ResidualVectorQuantization, channels: int = 1, sample_rate: int | None = None)

Bases: Module

EnCodec model operating on the raw waveform.

Parameters:

• encoder – Encoder network.

• decoder – Decoder network.

• quantizer – The residual quantizer module.

• channels – Number of audio channels.

• sample_rate – The sample rate of the input audio.

Initializes internal Module state, shared by both nn.Module and ScriptModule.

get_encoder() → Encoder

get_decoder() → Decoder

encode(waveform: Tensor, n_q: int | None = None) → Tensor

decode(tokens: Tensor) → Tensor

forward(x: Tensor) → tuple[torch.Tensor, torch.Tensor]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class pretrained.encodec.Encoder(encodec: Encodec)

Bases: Module

Initializes internal Module state, shared by both nn.Module and ScriptModule.

encode(waveform: Tensor, n_q: int | None = None) → Tensor

forward(waveform: Tensor) → Tensor

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class pretrained.encodec.Decoder(encodec: Encodec)

Bases: Module

Initializes internal Module state, shared by both nn.Module and ScriptModule.

decode(tokens: Tensor) → Tensor

forward(waveform: Tensor) → Tensor

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class pretrained.encodec.EncodecConfig(num_quantizers: int, channels: int, causal: bool, norm: Literal['no_norm', 'batch', 'batch_affine', 'instance', 'instance_affine', 'group', 'group_affine', 'layer', 'layer_affine'] | Literal['no_norm', 'weight', 'spectral'], sample_rate: int | None = None)

Bases: object

num_quantizers: int

channels: int

causal: bool

norm: Literal['no_norm', 'batch', 'batch_affine', 'instance', 'instance_affine', 'group', 'group_affine', 'layer', 'layer_affine'] | Literal['no_norm', 'weight', 'spectral']

sample_rate: int | None = None

pretrained.encodec.pretrained_encodec(size: str | Literal['24khz'], load_weights: bool = True) → Encodec

pretrained.encodec.test_encodec_adhoc() → None