Parakeet/parakeet/models/parallel_wavegan.py

# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import math
from typing import Any
from typing import Dict
from typing import List
from typing import Optional

import numpy as np
import paddle
from paddle import nn
from paddle.nn import functional as F


class Stretch2D(nn.Layer):
    def __init__(self, w_scale: int, h_scale: int, mode: str="nearest"):
        """Strech an image (or image-like object) with some interpolation.

        Parameters
        ----------
        w_scale : int
            Scalar of width.
        h_scale : int
            Scalar of the height.
        mode : str, optional
            Interpolation mode, modes suppored are "nearest", "bilinear", 
            "trilinear", "bicubic", "linear" and "area",by default "nearest"

            For more details about interpolation, see 
            `paddle.nn.functional.interpolate <https://www.paddlepaddle.org.cn/documentation/docs/en/api/paddle/nn/functional/interpolate_en.html>`_.
        """
        super().__init__()
        self.w_scale = w_scale
        self.h_scale = h_scale
        self.mode = mode

    def forward(self, x):
        """
        Parameters
        ----------
        x : Tensor
            Shape (N, C, H, W)

        Returns
        -------
        Tensor
            Shape (N, C, H', W'), where ``H'=h_scale * H``, ``W'=w_scale * W``.
            The stretched image.
        """
        out = F.interpolate(
            x, scale_factor=(self.h_scale, self.w_scale), mode=self.mode)
        return out


class UpsampleNet(nn.Layer):
    """A Layer to upsample spectrogram by applying consecutive stretch and
    convolutions.

    Parameters
    ----------
    upsample_scales : List[int]
        Upsampling factors for each strech.
    nonlinear_activation : Optional[str], optional
        Activation after each convolution, by default None
    nonlinear_activation_params : Dict[str, Any], optional
        Parameters passed to construct the activation, by default {}
    interpolate_mode : str, optional
        Interpolation mode of the strech, by default "nearest"
    freq_axis_kernel_size : int, optional
        Convolution kernel size along the frequency axis, by default 1
    use_causal_conv : bool, optional
        Whether to use causal padding before convolution, by default False

        If True, Causal padding is used along the time axis, i.e. padding
        amount is ``receptive field - 1`` and 0 for before and after,
        respectively.

        If False, "same" padding is used along the time axis.
    """

    def __init__(self,
                 upsample_scales: List[int],
                 nonlinear_activation: Optional[str]=None,
                 nonlinear_activation_params: Dict[str, Any]={},
                 interpolate_mode: str="nearest",
                 freq_axis_kernel_size: int=1,
                 use_causal_conv: bool=False):
        super().__init__()
        self.use_causal_conv = use_causal_conv
        self.up_layers = nn.LayerList()
        for scale in upsample_scales:
            stretch = Stretch2D(scale, 1, interpolate_mode)
            assert freq_axis_kernel_size % 2 == 1
            freq_axis_padding = (freq_axis_kernel_size - 1) // 2
            kernel_size = (freq_axis_kernel_size, scale * 2 + 1)
            if use_causal_conv:
                padding = (freq_axis_padding, scale * 2)
            else:
                padding = (freq_axis_padding, scale)
            conv = nn.Conv2D(
                1, 1, kernel_size, padding=padding, bias_attr=False)
            self.up_layers.extend([stretch, conv])
            if nonlinear_activation is not None:
                nonlinear = getattr(
                    nn, nonlinear_activation)(**nonlinear_activation_params)
                self.up_layers.append(nonlinear)

    def forward(self, c):
        """
        Parameters
        ----------
        c : Tensor
            Shape (N, F, T), spectrogram

        Returns
        -------
        Tensor
            Shape (N, F, T'), where ``T' = upsample_factor * T``, upsampled 
            spectrogram
        """
        c = c.unsqueeze(1)
        for f in self.up_layers:
            if self.use_causal_conv and isinstance(f, nn.Conv2D):
                c = f(c)[:, :, :, c.shape[-1]]
            else:
                c = f(c)
        return c.squeeze(1)


class ConvInUpsampleNet(nn.Layer):
    """A Layer to upsample spectrogram composed of a convolution and an 
    UpsampleNet.

    Parameters
    ----------
    upsample_scales : List[int]
        Upsampling factors for each strech.
    nonlinear_activation : Optional[str], optional
        Activation after each convolution, by default None
    nonlinear_activation_params : Dict[str, Any], optional
        Parameters passed to construct the activation, by default {}
    interpolate_mode : str, optional
        Interpolation mode of the strech, by default "nearest"
    freq_axis_kernel_size : int, optional
        Convolution kernel size along the frequency axis, by default 1
    aux_channels : int, optional
        Feature size of the input, by default 80
    aux_context_window : int, optional
        Context window of the first 1D convolution applied to the input. It 
        related to the kernel size of the convolution, by default 0

        If use causal convolution, the kernel size is ``window + 1``, else
        the kernel size is ``2 * window + 1``.
    use_causal_conv : bool, optional
        Whether to use causal padding before convolution, by default False

        If True, Causal padding is used along the time axis, i.e. padding 
        amount is ``receptive field - 1`` and 0 for before and after, 
        respectively.

        If False, "same" padding is used along the time axis.
    """

    def __init__(self,
                 upsample_scales: List[int],
                 nonlinear_activation: Optional[str]=None,
                 nonlinear_activation_params: Dict[str, Any]={},
                 interpolate_mode: str="nearest",
                 freq_axis_kernel_size: int=1,
                 aux_channels: int=80,
                 aux_context_window: int=0,
                 use_causal_conv: bool=False):
        super().__init__()
        self.aux_context_window = aux_context_window
        self.use_causal_conv = use_causal_conv and aux_context_window > 0
        kernel_size = aux_context_window + 1 if use_causal_conv else 2 * aux_context_window + 1
        self.conv_in = nn.Conv1D(
            aux_channels,
            aux_channels,
            kernel_size=kernel_size,
            bias_attr=False)
        self.upsample = UpsampleNet(
            upsample_scales=upsample_scales,
            nonlinear_activation=nonlinear_activation,
            nonlinear_activation_params=nonlinear_activation_params,
            interpolate_mode=interpolate_mode,
            freq_axis_kernel_size=freq_axis_kernel_size,
            use_causal_conv=use_causal_conv)

    def forward(self, c):
        """
        Parameters
        ----------
        c : Tensor
            Shape (N, F, T), spectrogram

        Returns
        -------
        Tensors
            Shape (N, F, T'), where ``T' = upsample_factor * T``, upsampled 
            spectrogram
        """
        c_ = self.conv_in(c)
        c = c_[:, :, :-self.aux_context_window] if self.use_causal_conv else c_
        return self.upsample(c)


class ResidualBlock(nn.Layer):
    """A gated activation unit composed of an 1D convolution, a gated tanh
    unit and parametric redidual and skip connections. For more details, 
    refer to `WaveNet: A Generative Model for Raw Audio <https://arxiv.org/abs/1609.03499>`_.

    Parameters
    ----------
    kernel_size : int, optional
        Kernel size of the 1D convolution, by default 3
    residual_channels : int, optional
        Feature size of the resiaudl output(and also the input), by default 64
    gate_channels : int, optional
        Output feature size of the 1D convolution, by default 128
    skip_channels : int, optional
        Feature size of the skip output, by default 64
    aux_channels : int, optional
        Feature size of the auxiliary input (e.g. spectrogram), by default 80
    dropout : float, optional
        Probability of the dropout before the 1D convolution, by default 0.
    dilation : int, optional
        Dilation of the 1D convolution, by default 1
    bias : bool, optional
        Whether to use bias in the 1D convolution, by default True
    use_causal_conv : bool, optional
        Whether to use causal padding for the 1D convolution, by default False
    """

    def __init__(self,
                 kernel_size: int=3,
                 residual_channels: int=64,
                 gate_channels: int=128,
                 skip_channels: int=64,
                 aux_channels: int=80,
                 dropout: float=0.,
                 dilation: int=1,
                 bias: bool=True,
                 use_causal_conv: bool=False):
        super().__init__()
        self.dropout = dropout
        if use_causal_conv:
            padding = (kernel_size - 1) * dilation
        else:
            assert kernel_size % 2 == 1
            padding = (kernel_size - 1) // 2 * dilation
        self.use_causal_conv = use_causal_conv

        self.conv = nn.Conv1D(
            residual_channels,
            gate_channels,
            kernel_size,
            padding=padding,
            dilation=dilation,
            bias_attr=bias)
        if aux_channels is not None:
            self.conv1x1_aux = nn.Conv1D(
                aux_channels, gate_channels, kernel_size=1, bias_attr=False)
        else:
            self.conv1x1_aux = None

        gate_out_channels = gate_channels // 2
        self.conv1x1_out = nn.Conv1D(
            gate_out_channels, residual_channels, kernel_size=1, bias_attr=bias)
        self.conv1x1_skip = nn.Conv1D(
            gate_out_channels, skip_channels, kernel_size=1, bias_attr=bias)

    def forward(self, x, c):
        """
        Parameters
        ----------
        x : Tensor
            Shape (N, C_res, T), the input features.
        c : Tensor
            Shape (N, C_aux, T), the auxiliary input.

        Returns
        -------
        res : Tensor
            Shape (N, C_res, T), the residual output, which is used as the 
            input of the next ResidualBlock in a stack of ResidualBlocks.
        skip : Tensor
            Shape (N, C_skip, T), the skip output, which is collected among
            each layer in a stack of ResidualBlocks.
        """
        x_input = x
        x = F.dropout(x, self.dropout, training=self.training)
        x = self.conv(x)
        x = x[:, :, x_input.shape[-1]] if self.use_causal_conv else x
        if c is not None:
            c = self.conv1x1_aux(c)
            x += c

        a, b = paddle.chunk(x, 2, axis=1)
        x = paddle.tanh(a) * F.sigmoid(b)

        skip = self.conv1x1_skip(x)
        res = (self.conv1x1_out(x) + x_input) * math.sqrt(0.5)
        return res, skip


class PWGGenerator(nn.Layer):
    """Wave Generator for Parallel WaveGAN

    Parameters
    ----------
    in_channels : int, optional
        Number of channels of the input waveform, by default 1
    out_channels : int, optional
        Number of channels of the output waveform, by default 1
    kernel_size : int, optional
        Kernel size of the residual blocks inside, by default 3
    layers : int, optional
        Number of residual blocks inside, by default 30
    stacks : int, optional
        The number of groups to split the residual blocks into, by default 3

        Within each group, the dilation of the residual block grows 
        exponentially.
    residual_channels : int, optional
        Residual channel of the residual blocks, by default 64
    gate_channels : int, optional
        Gate channel of the residual blocks, by default 128
    skip_channels : int, optional
        Skip channel of the residual blocks, by default 64
    aux_channels : int, optional
        Auxiliary channel of the residual blocks, by default 80
    aux_context_window : int, optional
        The context window size of the first convolution applied to the 
        auxiliary input, by default 2
    dropout : float, optional
        Dropout of the residual blocks, by default 0.
    bias : bool, optional
        Whether to use bias in residual blocks, by default True
    use_weight_norm : bool, optional
        Whether to use weight norm in all convolutions, by default True
    use_causal_conv : bool, optional
        Whether to use causal padding in the upsample network and residual 
        blocks, by default False
    upsample_scales : List[int], optional
        Upsample scales of the upsample network, by default [4, 4, 4, 4]
    nonlinear_activation : Optional[str], optional
        Non linear activation in upsample network, by default None
    nonlinear_activation_params : Dict[str, Any], optional
        Parameters passed to the linear activation in the upsample network, 
        by default {}
    interpolate_mode : str, optional
        Interpolation mode of the upsample network, by default "nearest"
    freq_axis_kernel_size : int, optional
        Kernel size along the frequency axis of the upsample network, by default 1
    """

    def __init__(self,
                 in_channels: int=1,
                 out_channels: int=1,
                 kernel_size: int=3,
                 layers: int=30,
                 stacks: int=3,
                 residual_channels: int=64,
                 gate_channels: int=128,
                 skip_channels: int=64,
                 aux_channels: int=80,
                 aux_context_window: int=2,
                 dropout: float=0.,
                 bias: bool=True,
                 use_weight_norm: bool=True,
                 use_causal_conv: bool=False,
                 upsample_scales: List[int]=[4, 4, 4, 4],
                 nonlinear_activation: Optional[str]=None,
                 nonlinear_activation_params: Dict[str, Any]={},
                 interpolate_mode: str="nearest",
                 freq_axis_kernel_size: int=1):
        super().__init__()
        self.in_channels = in_channels
        self.out_channels = out_channels
        self.aux_channels = aux_channels
        self.aux_context_window = aux_context_window
        self.layers = layers
        self.stacks = stacks
        self.kernel_size = kernel_size

        assert layers % stacks == 0
        layers_per_stack = layers // stacks

        self.first_conv = nn.Conv1D(
            in_channels, residual_channels, 1, bias_attr=True)
        self.upsample_net = ConvInUpsampleNet(
            upsample_scales=upsample_scales,
            nonlinear_activation=nonlinear_activation,
            nonlinear_activation_params=nonlinear_activation_params,
            interpolate_mode=interpolate_mode,
            freq_axis_kernel_size=freq_axis_kernel_size,
            aux_channels=aux_channels,
            aux_context_window=aux_context_window,
            use_causal_conv=use_causal_conv)
        self.upsample_factor = np.prod(upsample_scales)

        self.conv_layers = nn.LayerList()
        for layer in range(layers):
            dilation = 2**(layer % layers_per_stack)
            conv = ResidualBlock(
                kernel_size=kernel_size,
                residual_channels=residual_channels,
                gate_channels=gate_channels,
                skip_channels=skip_channels,
                aux_channels=aux_channels,
                dilation=dilation,
                dropout=dropout,
                bias=bias,
                use_causal_conv=use_causal_conv)
            self.conv_layers.append(conv)

        self.last_conv_layers = nn.Sequential(nn.ReLU(),
                                              nn.Conv1D(
                                                  skip_channels,
                                                  skip_channels,
                                                  1,
                                                  bias_attr=True),
                                              nn.ReLU(),
                                              nn.Conv1D(
                                                  skip_channels,
                                                  out_channels,
                                                  1,
                                                  bias_attr=True))

        if use_weight_norm:
            self.apply_weight_norm()

    def forward(self, x, c):
        """Generate waveform.

        Parameters
        ----------
        x : Tensor
            Shape (N, C_in, T), The input waveform.
        c : Tensor
            Shape (N, C_aux, T'). The auxiliary input (e.g. spectrogram). It 
            is upsampled to match the time resolution of the input.

        Returns
        -------
        Tensor
            Shape (N, C_out, T), the generated waveform.
        """
        c = self.upsample_net(c)
        assert c.shape[-1] == x.shape[-1]

        x = self.first_conv(x)
        skips = 0
        for f in self.conv_layers:
            x, s = f(x, c)
            skips += s
        skips *= math.sqrt(1.0 / len(self.conv_layers))

        x = self.last_conv_layers(skips)
        return x

    def apply_weight_norm(self):
        """Recursively apply weight normalization to all the Convolution layers
        in the sublayers.
        """

        def _apply_weight_norm(layer):
            if isinstance(layer, (nn.Conv1D, nn.Conv2D)):
                nn.utils.weight_norm(layer)

        self.apply(_apply_weight_norm)

    def remove_weight_norm(self):
        """Recursively remove weight normalization from all the Convolution 
        layers in the sublayers.
        """

        def _remove_weight_norm(layer):
            try:
                nn.utils.remove_weight_norm(layer)
            except ValueError:
                pass

        self.apply(_remove_weight_norm)

    def inference(self, c=None):
        """Waveform generation. This function is used for single instance 
        inference.

        Parameters
        ----------
        c : Tensor, optional
            Shape (T', C_aux), the auxiliary input, by default None
        x : Tensor, optional
            Shape (T, C_in), the noise waveform, by default None
            If not provided, a sample is drawn from a gaussian distribution.

        Returns
        -------
        Tensor
            Shape (T, C_out), the generated waveform
        """
        x = paddle.randn(
            [1, self.in_channels, paddle.shape(c)[0] * self.upsample_factor])
        c = paddle.transpose(c, [1, 0]).unsqueeze(0)  # pseudo batch
        c = nn.Pad1D(self.aux_context_window, mode='replicate')(c)
        out = self(x, c).squeeze(0).transpose([1, 0])
        return out


class PWGDiscriminator(nn.Layer):
    """A convolutional discriminator for audio.

    Parameters
    ----------
    in_channels : int, optional
        Number of channels of the input audio, by default 1
    out_channels : int, optional
        Output feature size, by default 1
    kernel_size : int, optional
        Kernel size of convolutional sublayers, by default 3
    layers : int, optional
        Number of layers, by default 10
    conv_channels : int, optional
        Feature size of the convolutional sublayers, by default 64
    dilation_factor : int, optional
        The factor with which dilation of each convolutional sublayers grows 
        exponentially if it is greater than 1, else the dilation of each 
        convolutional sublayers grows linearly, by default 1
    nonlinear_activation : str, optional
        The activation after each convolutional sublayer, by default "LeakyReLU"
    nonlinear_activation_params : Dict[str, Any], optional
        The parameters passed to the activation's initializer, by default 
        {"negative_slope": 0.2}
    bias : bool, optional
        Whether to use bias in convolutional sublayers, by default True
    use_weight_norm : bool, optional
        Whether to use weight normalization at all convolutional sublayers, 
        by default True
    """

    def __init__(
            self,
            in_channels: int=1,
            out_channels: int=1,
            kernel_size: int=3,
            layers: int=10,
            conv_channels: int=64,
            dilation_factor: int=1,
            nonlinear_activation: str="LeakyReLU",
            nonlinear_activation_params: Dict[str, Any]={"negative_slope": 0.2},
            bias: bool=True,
            use_weight_norm: bool=True):
        super().__init__()
        assert kernel_size % 2 == 1
        assert dilation_factor > 0
        conv_layers = []
        conv_in_channels = in_channels
        for i in range(layers - 1):
            if i == 0:
                dilation = 1
            else:
                dilation = i if dilation_factor == 1 else dilation_factor**i
                conv_in_channels = conv_channels
            padding = (kernel_size - 1) // 2 * dilation
            conv_layer = nn.Conv1D(
                conv_in_channels,
                conv_channels,
                kernel_size,
                padding=padding,
                dilation=dilation,
                bias_attr=bias)
            nonlinear = getattr(
                nn, nonlinear_activation)(**nonlinear_activation_params)
            conv_layers.append(conv_layer)
            conv_layers.append(nonlinear)
        padding = (kernel_size - 1) // 2
        last_conv = nn.Conv1D(
            conv_in_channels,
            out_channels,
            kernel_size,
            padding=padding,
            bias_attr=bias)
        conv_layers.append(last_conv)
        self.conv_layers = nn.Sequential(*conv_layers)

        if use_weight_norm:
            self.apply_weight_norm()

    def forward(self, x):
        """
        Parameters
        ----------
        x : Tensor
            Shape (N, in_channels, num_samples), the input audio.

        Returns
        -------
        Tensor
            Shape (N, out_channels, num_samples), the predicted logits.
        """
        return self.conv_layers(x)

    def apply_weight_norm(self):
        def _apply_weight_norm(layer):
            if isinstance(layer, (nn.Conv1D, nn.Conv2D)):
                nn.utils.weight_norm(layer)

        self.apply(_apply_weight_norm)

    def remove_weight_norm(self):
        def _remove_weight_norm(layer):
            try:
                nn.utils.remove_weight_norm(layer)
            except ValueError:
                pass

        self.apply(_remove_weight_norm)


class ResidualPWGDiscriminator(nn.Layer):
    """A wavenet-style discriminator for audio.

    Parameters
    ----------
    in_channels : int, optional
        Number of channels of the input audio, by default 1
    out_channels : int, optional
        Output feature size, by default 1
    kernel_size : int, optional
        Kernel size of residual blocks, by default 3
    layers : int, optional
        Number of residual blocks, by default 30
    stacks : int, optional
        Number of groups of residual blocks, within which the dilation 
        of each residual blocks grows exponentially, by default 3
    residual_channels : int, optional
        Residual channels of residual blocks, by default 64
    gate_channels : int, optional
        Gate channels of residual blocks, by default 128
    skip_channels : int, optional
        Skip channels of residual blocks, by default 64
    dropout : float, optional
        Dropout probability of residual blocks, by default 0.
    bias : bool, optional
        Whether to use bias in residual blocks, by default True
    use_weight_norm : bool, optional
        Whether to use weight normalization in all convolutional layers, 
        by default True
    use_causal_conv : bool, optional
        Whether to use causal convolution in residual blocks, by default False
    nonlinear_activation : str, optional
        Activation after convolutions other than those in residual blocks, 
        by default "LeakyReLU"
    nonlinear_activation_params : Dict[str, Any], optional
        Parameters to pass to the activation, by default {"negative_slope": 0.2}
    """

    def __init__(self,
                 in_channels: int=1,
                 out_channels: int=1,
                 kernel_size: int=3,
                 layers: int=30,
                 stacks: int=3,
                 residual_channels: int=64,
                 gate_channels: int=128,
                 skip_channels: int=64,
                 dropout: float=0.,
                 bias: bool=True,
                 use_weight_norm: bool=True,
                 use_causal_conv: bool=False,
                 nonlinear_activation: str="LeakyReLU",
                 nonlinear_activation_params: Dict[
                     str, Any]={"negative_slope": 0.2}):
        super().__init__()
        assert kernel_size % 2 == 1
        self.in_channels = in_channels
        self.out_channels = out_channels
        self.layers = layers
        self.stacks = stacks
        self.kernel_size = kernel_size

        assert layers % stacks == 0
        layers_per_stack = layers // stacks

        self.first_conv = nn.Sequential(
            nn.Conv1D(in_channels, residual_channels, 1, bias_attr=True),
            getattr(nn, nonlinear_activation)(**nonlinear_activation_params))

        self.conv_layers = nn.LayerList()
        for layer in range(layers):
            dilation = 2**(layer % layers_per_stack)
            conv = ResidualBlock(
                kernel_size=kernel_size,
                residual_channels=residual_channels,
                gate_channels=gate_channels,
                skip_channels=skip_channels,
                aux_channels=None,  # no auxiliary input
                dropout=dropout,
                dilation=dilation,
                bias=bias,
                use_causal_conv=use_causal_conv)
            self.conv_layers.append(conv)

        self.last_conv_layers = nn.Sequential(
            getattr(nn, nonlinear_activation)(**nonlinear_activation_params),
            nn.Conv1D(skip_channels, skip_channels, 1, bias_attr=True),
            getattr(nn, nonlinear_activation)(**nonlinear_activation_params),
            nn.Conv1D(skip_channels, out_channels, 1, bias_attr=True))

        if use_weight_norm:
            self.apply_weight_norm()

    def forward(self, x):
        """
        Parameters
        ----------
        x : Tensor
            Shape (N, in_channels, num_samples), the input audio.

        Returns
        -------
        Tensor
            Shape (N, out_channels, num_samples), the predicted logits.
        """
        x = self.first_conv(x)
        skip = 0
        for f in self.conv_layers:
            x, h = f(x, None)
            skip += h
        skip *= math.sqrt(1 / len(self.conv_layers))

        x = skip
        x = self.last_conv_layers(x)
        return x

    def apply_weight_norm(self):
        def _apply_weight_norm(layer):
            if isinstance(layer, (nn.Conv1D, nn.Conv2D)):
                nn.utils.weight_norm(layer)

        self.apply(_apply_weight_norm)

    def remove_weight_norm(self):
        def _remove_weight_norm(layer):
            try:
                nn.utils.remove_weight_norm(layer)
            except ValueError:
                pass

        self.apply(_remove_weight_norm)


class PWGInference(nn.Layer):
    def __init__(self, normalizer, pwg_generator):
        super().__init__()
        self.normalizer = normalizer
        self.pwg_generator = pwg_generator

    def forward(self, logmel):
        normalized_mel = self.normalizer(logmel)
        wav = self.pwg_generator.inference(normalized_mel)
        return wav