indextts/utils/maskgct/models/codec/speechtokenizer/modules/seanet.py

# Copyright (c) 2023 Amphion.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
# This source file is copied from https://github.com/facebookresearch/encodec

# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.

"""Encodec SEANet-based encoder and decoder implementation."""

import typing as tp

import numpy as np
import torch.nn as nn
import torch

from . import SConv1d, SConvTranspose1d, SLSTM


@torch.jit.script
def snake(x, alpha):
    shape = x.shape
    x = x.reshape(shape[0], shape[1], -1)
    x = x + (alpha + 1e-9).reciprocal() * torch.sin(alpha * x).pow(2)
    x = x.reshape(shape)
    return x


class Snake1d(nn.Module):
    def __init__(self, channels):
        super().__init__()
        self.alpha = nn.Parameter(torch.ones(1, channels, 1))

    def forward(self, x):
        return snake(x, self.alpha)


class SEANetResnetBlock(nn.Module):
    """Residual block from SEANet model.
    Args:
        dim (int): Dimension of the input/output
        kernel_sizes (list): List of kernel sizes for the convolutions.
        dilations (list): List of dilations for the convolutions.
        activation (str): Activation function.
        activation_params (dict): Parameters to provide to the activation function
        norm (str): Normalization method.
        norm_params (dict): Parameters to provide to the underlying normalization used along with the convolution.
        causal (bool): Whether to use fully causal convolution.
        pad_mode (str): Padding mode for the convolutions.
        compress (int): Reduced dimensionality in residual branches (from Demucs v3)
        true_skip (bool): Whether to use true skip connection or a simple convolution as the skip connection.
    """

    def __init__(
        self,
        dim: int,
        kernel_sizes: tp.List[int] = [3, 1],
        dilations: tp.List[int] = [1, 1],
        activation: str = "ELU",
        activation_params: dict = {"alpha": 1.0},
        norm: str = "weight_norm",
        norm_params: tp.Dict[str, tp.Any] = {},
        causal: bool = False,
        pad_mode: str = "reflect",
        compress: int = 2,
        true_skip: bool = True,
    ):
        super().__init__()
        assert len(kernel_sizes) == len(
            dilations
        ), "Number of kernel sizes should match number of dilations"
        act = getattr(nn, activation) if activation != "Snake" else Snake1d
        hidden = dim // compress
        block = []
        for i, (kernel_size, dilation) in enumerate(zip(kernel_sizes, dilations)):
            in_chs = dim if i == 0 else hidden
            out_chs = dim if i == len(kernel_sizes) - 1 else hidden
            block += [
                act(**activation_params) if activation != "Snake" else act(in_chs),
                SConv1d(
                    in_chs,
                    out_chs,
                    kernel_size=kernel_size,
                    dilation=dilation,
                    norm=norm,
                    norm_kwargs=norm_params,
                    causal=causal,
                    pad_mode=pad_mode,
                ),
            ]
        self.block = nn.Sequential(*block)
        self.shortcut: nn.Module
        if true_skip:
            self.shortcut = nn.Identity()
        else:
            self.shortcut = SConv1d(
                dim,
                dim,
                kernel_size=1,
                norm=norm,
                norm_kwargs=norm_params,
                causal=causal,
                pad_mode=pad_mode,
            )

    def forward(self, x):
        return self.shortcut(x) + self.block(x)


class SEANetEncoder(nn.Module):
    """SEANet encoder.
    Args:
        channels (int): Audio channels.
        dimension (int): Intermediate representation dimension.
        n_filters (int): Base width for the model.
        n_residual_layers (int): nb of residual layers.
        ratios (Sequence[int]): kernel size and stride ratios. The encoder uses downsampling ratios instead of
            upsampling ratios, hence it will use the ratios in the reverse order to the ones specified here
            that must match the decoder order
        activation (str): Activation function.
        activation_params (dict): Parameters to provide to the activation function
        norm (str): Normalization method.
        norm_params (dict): Parameters to provide to the underlying normalization used along with the convolution.
        kernel_size (int): Kernel size for the initial convolution.
        last_kernel_size (int): Kernel size for the initial convolution.
        residual_kernel_size (int): Kernel size for the residual layers.
        dilation_base (int): How much to increase the dilation with each layer.
        causal (bool): Whether to use fully causal convolution.
        pad_mode (str): Padding mode for the convolutions.
        true_skip (bool): Whether to use true skip connection or a simple
            (streamable) convolution as the skip connection in the residual network blocks.
        compress (int): Reduced dimensionality in residual branches (from Demucs v3).
        lstm (int): Number of LSTM layers at the end of the encoder.
    """

    def __init__(
        self,
        channels: int = 1,
        dimension: int = 128,
        n_filters: int = 32,
        n_residual_layers: int = 1,
        ratios: tp.List[int] = [8, 5, 4, 2],
        activation: str = "ELU",
        activation_params: dict = {"alpha": 1.0},
        norm: str = "weight_norm",
        norm_params: tp.Dict[str, tp.Any] = {},
        kernel_size: int = 7,
        last_kernel_size: int = 7,
        residual_kernel_size: int = 3,
        dilation_base: int = 2,
        causal: bool = False,
        pad_mode: str = "reflect",
        true_skip: bool = False,
        compress: int = 2,
        lstm: int = 2,
        bidirectional: bool = False,
    ):
        super().__init__()
        self.channels = channels
        self.dimension = dimension
        self.n_filters = n_filters
        self.ratios = list(reversed(ratios))
        del ratios
        self.n_residual_layers = n_residual_layers
        self.hop_length = np.prod(self.ratios)  # 计算乘积

        act = getattr(nn, activation) if activation != "Snake" else Snake1d
        mult = 1
        model: tp.List[nn.Module] = [
            SConv1d(
                channels,
                mult * n_filters,
                kernel_size,
                norm=norm,
                norm_kwargs=norm_params,
                causal=causal,
                pad_mode=pad_mode,
            )
        ]
        # Downsample to raw audio scale
        for i, ratio in enumerate(self.ratios):
            # Add residual layers
            for j in range(n_residual_layers):
                model += [
                    SEANetResnetBlock(
                        mult * n_filters,
                        kernel_sizes=[residual_kernel_size, 1],
                        dilations=[dilation_base**j, 1],
                        norm=norm,
                        norm_params=norm_params,
                        activation=activation,
                        activation_params=activation_params,
                        causal=causal,
                        pad_mode=pad_mode,
                        compress=compress,
                        true_skip=true_skip,
                    )
                ]

            # Add downsampling layers
            model += [
                (
                    act(**activation_params)
                    if activation != "Snake"
                    else act(mult * n_filters)
                ),
                SConv1d(
                    mult * n_filters,
                    mult * n_filters * 2,
                    kernel_size=ratio * 2,
                    stride=ratio,
                    norm=norm,
                    norm_kwargs=norm_params,
                    causal=causal,
                    pad_mode=pad_mode,
                ),
            ]
            mult *= 2

        if lstm:
            model += [
                SLSTM(mult * n_filters, num_layers=lstm, bidirectional=bidirectional)
            ]

        mult = mult * 2 if bidirectional else mult
        model += [
            (
                act(**activation_params)
                if activation != "Snake"
                else act(mult * n_filters)
            ),
            SConv1d(
                mult * n_filters,
                dimension,
                last_kernel_size,
                norm=norm,
                norm_kwargs=norm_params,
                causal=causal,
                pad_mode=pad_mode,
            ),
        ]

        self.model = nn.Sequential(*model)

    def forward(self, x):
        return self.model(x)


class SEANetDecoder(nn.Module):
    """SEANet decoder.
    Args:
        channels (int): Audio channels.
        dimension (int): Intermediate representation dimension.
        n_filters (int): Base width for the model.
        n_residual_layers (int): nb of residual layers.
        ratios (Sequence[int]): kernel size and stride ratios
        activation (str): Activation function.
        activation_params (dict): Parameters to provide to the activation function
        final_activation (str): Final activation function after all convolutions.
        final_activation_params (dict): Parameters to provide to the activation function
        norm (str): Normalization method.
        norm_params (dict): Parameters to provide to the underlying normalization used along with the convolution.
        kernel_size (int): Kernel size for the initial convolution.
        last_kernel_size (int): Kernel size for the initial convolution.
        residual_kernel_size (int): Kernel size for the residual layers.
        dilation_base (int): How much to increase the dilation with each layer.
        causal (bool): Whether to use fully causal convolution.
        pad_mode (str): Padding mode for the convolutions.
        true_skip (bool): Whether to use true skip connection or a simple
            (streamable) convolution as the skip connection in the residual network blocks.
        compress (int): Reduced dimensionality in residual branches (from Demucs v3).
        lstm (int): Number of LSTM layers at the end of the encoder.
        trim_right_ratio (float): Ratio for trimming at the right of the transposed convolution under the causal setup.
            If equal to 1.0, it means that all the trimming is done at the right.
    """

    def __init__(
        self,
        channels: int = 1,
        dimension: int = 128,
        n_filters: int = 32,
        n_residual_layers: int = 1,
        ratios: tp.List[int] = [8, 5, 4, 2],
        activation: str = "ELU",
        activation_params: dict = {"alpha": 1.0},
        final_activation: tp.Optional[str] = None,
        final_activation_params: tp.Optional[dict] = None,
        norm: str = "weight_norm",
        norm_params: tp.Dict[str, tp.Any] = {},
        kernel_size: int = 7,
        last_kernel_size: int = 7,
        residual_kernel_size: int = 3,
        dilation_base: int = 2,
        causal: bool = False,
        pad_mode: str = "reflect",
        true_skip: bool = False,
        compress: int = 2,
        lstm: int = 2,
        trim_right_ratio: float = 1.0,
        bidirectional: bool = False,
    ):
        super().__init__()
        self.dimension = dimension
        self.channels = channels
        self.n_filters = n_filters
        self.ratios = ratios
        del ratios
        self.n_residual_layers = n_residual_layers
        self.hop_length = np.prod(self.ratios)

        act = getattr(nn, activation) if activation != "Snake" else Snake1d
        mult = int(2 ** len(self.ratios))
        model: tp.List[nn.Module] = [
            SConv1d(
                dimension,
                mult * n_filters,
                kernel_size,
                norm=norm,
                norm_kwargs=norm_params,
                causal=causal,
                pad_mode=pad_mode,
            )
        ]

        if lstm:
            model += [
                SLSTM(mult * n_filters, num_layers=lstm, bidirectional=bidirectional)
            ]

        # Upsample to raw audio scale
        for i, ratio in enumerate(self.ratios):
            # Add upsampling layers
            model += [
                (
                    act(**activation_params)
                    if activation != "Snake"
                    else act(mult * n_filters)
                ),
                SConvTranspose1d(
                    mult * n_filters,
                    mult * n_filters // 2,
                    kernel_size=ratio * 2,
                    stride=ratio,
                    norm=norm,
                    norm_kwargs=norm_params,
                    causal=causal,
                    trim_right_ratio=trim_right_ratio,
                ),
            ]
            # Add residual layers
            for j in range(n_residual_layers):
                model += [
                    SEANetResnetBlock(
                        mult * n_filters // 2,
                        kernel_sizes=[residual_kernel_size, 1],
                        dilations=[dilation_base**j, 1],
                        activation=activation,
                        activation_params=activation_params,
                        norm=norm,
                        norm_params=norm_params,
                        causal=causal,
                        pad_mode=pad_mode,
                        compress=compress,
                        true_skip=true_skip,
                    )
                ]

            mult //= 2

        # Add final layers
        model += [
            act(**activation_params) if activation != "Snake" else act(n_filters),
            SConv1d(
                n_filters,
                channels,
                last_kernel_size,
                norm=norm,
                norm_kwargs=norm_params,
                causal=causal,
                pad_mode=pad_mode,
            ),
        ]
        # Add optional final activation to decoder (eg. tanh)
        if final_activation is not None:
            final_act = getattr(nn, final_activation)
            final_activation_params = final_activation_params or {}
            model += [final_act(**final_activation_params)]
        self.model = nn.Sequential(*model)

    def forward(self, z):
        y = self.model(z)
        return y


def test():
    import torch

    encoder = SEANetEncoder()
    decoder = SEANetDecoder()
    x = torch.randn(1, 1, 24000)
    z = encoder(x)
    print("z ", z.shape)
    assert 1 == 2
    assert list(z.shape) == [1, 128, 75], z.shape
    y = decoder(z)
    assert y.shape == x.shape, (x.shape, y.shape)


if __name__ == "__main__":
    test()