src/kanade_tokenizer/module/hift.py

# Adapted from: https://github.com/yl4579/HiFTNet/blob/main/models.py
# https://github.com/FunAudioLLM/CosyVoice/blob/main/cosyvoice/hifigan/generator.py

from typing import Dict, List, Optional

import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from scipy.signal import get_window
from torch.distributions.uniform import Uniform
from torch.nn import Conv1d, ConvTranspose1d
from torch.nn.utils import remove_weight_norm
from torch.nn.utils.parametrizations import weight_norm


def get_padding(kernel_size, dilation=1):
    return int((kernel_size * dilation - dilation) / 2)


def init_weights(m, mean=0.0, std=0.01):
    classname = m.__class__.__name__
    if classname.find("Conv") != -1:
        m.weight.data.normal_(mean, std)


def mel_spec_transform(
    audio: torch.Tensor,
    n_fft: int,
    n_mels: int,
    sample_rate: int,
    hop_size: int,
    win_size: int,
    fmin: int = 0,
    fmax: Optional[int] = None,
):
    from librosa.filters import mel as librosa_mel_fn

    # (n_mels, n_fft // 2 + 1)
    mel_basis = librosa_mel_fn(
        sr=sample_rate, n_fft=n_fft, n_mels=n_mels, norm="slaney", htk=False, fmin=fmin, fmax=fmax
    )
    mel_basis = torch.from_numpy(mel_basis).float()
    hann_window = torch.hann_window(win_size)

    # Pad so that the output length T = L // hop_length
    padding = (n_fft - hop_size) // 2
    audio = torch.nn.functional.pad(audio, (padding, padding), mode="reflect")
    audio = audio.reshape(-1, audio.shape[-1])

    # (B, n_fft // 2 + 1, T=1 + (L' - n_fft) // hop_length)
    # L' = L + n_fft - hop_length
    # T = L // hop_length
    spec = torch.stft(
        audio,
        n_fft=n_fft,
        hop_length=hop_size,
        win_length=win_size,
        window=hann_window,
        center=False,
        pad_mode="reflect",
        normalized=False,
        onesided=True,
        return_complex=True,
    )
    spec = spec.reshape(audio.shape[:-1] + spec.shape[-2:])

    spec = torch.sqrt(torch.view_as_real(spec).pow(2).sum(-1) + 1e-9)
    mel_spec = torch.matmul(mel_basis, spec)
    mel_spec = torch.log(torch.clamp(mel_spec, min=1e-5))

    return mel_spec


class Snake(nn.Module):
    """
    Implementation of a sine-based periodic activation function
    Shape:
        - Input: (B, C, T)
        - Output: (B, C, T), same shape as the input
    Parameters:
        - alpha - trainable parameter
    References:
        - This activation function is from this paper by Liu Ziyin, Tilman Hartwig, Masahito Ueda:
        https://arxiv.org/abs/2006.08195
    Examples:
        >>> a1 = snake(256)
        >>> x = torch.randn(256)
        >>> x = a1(x)
    """

    def __init__(self, in_features, alpha=1.0, alpha_trainable=True, alpha_logscale=False):
        """
        Initialization.
        INPUT:
            - in_features: shape of the input
            - alpha: trainable parameter
            alpha is initialized to 1 by default, higher values = higher-frequency.
            alpha will be trained along with the rest of your model.
        """
        super(Snake, self).__init__()
        self.in_features = in_features

        # initialize alpha
        self.alpha_logscale = alpha_logscale
        if self.alpha_logscale:  # log scale alphas initialized to zeros
            self.alpha = nn.Parameter(torch.zeros(in_features) * alpha)
        else:  # linear scale alphas initialized to ones
            self.alpha = nn.Parameter(torch.ones(in_features) * alpha)

        self.alpha.requires_grad = alpha_trainable

        self.no_div_by_zero = 0.000000001

    def forward(self, x):
        """
        Forward pass of the function.
        Applies the function to the input elementwise.
        Snake ∶= x + 1/a * sin^2 (xa)
        """
        alpha = self.alpha.unsqueeze(0).unsqueeze(-1)  # line up with x to [B, C, T]
        if self.alpha_logscale:
            alpha = torch.exp(alpha)
        x = x + (1.0 / (alpha + self.no_div_by_zero)) * torch.pow(torch.sin(x * alpha), 2)

        return x


class ResBlock(torch.nn.Module):
    """Residual block module in HiFiGAN/BigVGAN."""

    def __init__(
        self,
        channels: int = 512,
        kernel_size: int = 3,
        dilations: List[int] = [1, 3, 5],
    ):
        super(ResBlock, self).__init__()
        self.convs1 = nn.ModuleList()
        self.convs2 = nn.ModuleList()

        for dilation in dilations:
            self.convs1.append(
                weight_norm(
                    Conv1d(
                        channels,
                        channels,
                        kernel_size,
                        1,
                        dilation=dilation,
                        padding=get_padding(kernel_size, dilation),
                    )
                )
            )
            self.convs2.append(
                weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1, padding=get_padding(kernel_size, 1)))
            )
        self.convs1.apply(init_weights)
        self.convs2.apply(init_weights)
        self.activations1 = nn.ModuleList([Snake(channels, alpha_logscale=False) for _ in range(len(self.convs1))])
        self.activations2 = nn.ModuleList([Snake(channels, alpha_logscale=False) for _ in range(len(self.convs2))])

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        for idx in range(len(self.convs1)):
            xt = self.activations1[idx](x)
            xt = self.convs1[idx](xt)
            xt = self.activations2[idx](xt)
            xt = self.convs2[idx](xt)
            x = xt + x
        return x

    def remove_weight_norm(self):
        for idx in range(len(self.convs1)):
            remove_weight_norm(self.convs1[idx])
            remove_weight_norm(self.convs2[idx])


class ConvRNNF0Predictor(nn.Module):
    def __init__(self, num_class: int = 1, in_channels: int = 80, cond_channels: int = 512):
        super().__init__()

        self.num_class = num_class
        self.condnet = nn.Sequential(
            weight_norm(nn.Conv1d(in_channels, cond_channels, kernel_size=3, padding=1)),
            nn.ELU(),
            weight_norm(nn.Conv1d(cond_channels, cond_channels, kernel_size=3, padding=1)),
            nn.ELU(),
            weight_norm(nn.Conv1d(cond_channels, cond_channels, kernel_size=3, padding=1)),
            nn.ELU(),
            weight_norm(nn.Conv1d(cond_channels, cond_channels, kernel_size=3, padding=1)),
            nn.ELU(),
            weight_norm(nn.Conv1d(cond_channels, cond_channels, kernel_size=3, padding=1)),
            nn.ELU(),
        )
        self.classifier = nn.Linear(in_features=cond_channels, out_features=self.num_class)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x = self.condnet(x)
        x = x.transpose(1, 2)
        return torch.abs(self.classifier(x).squeeze(-1))


class SineGen(torch.nn.Module):
    """Definition of sine generator
    SineGen(samp_rate, harmonic_num = 0,
            sine_amp = 0.1, noise_std = 0.003,
            voiced_threshold = 0,
            flag_for_pulse=False)
    samp_rate: sampling rate in Hz
    harmonic_num: number of harmonic overtones (default 0)
    sine_amp: amplitude of sine-wavefrom (default 0.1)
    noise_std: std of Gaussian noise (default 0.003)
    voiced_thoreshold: F0 threshold for U/V classification (default 0)
    flag_for_pulse: this SinGen is used inside PulseGen (default False)
    Note: when flag_for_pulse is True, the first time step of a voiced
        segment is always sin(np.pi) or cos(0)
    """

    def __init__(self, samp_rate, harmonic_num=0, sine_amp=0.1, noise_std=0.003, voiced_threshold=0):
        super(SineGen, self).__init__()
        self.sine_amp = sine_amp
        self.noise_std = noise_std
        self.harmonic_num = harmonic_num
        self.sampling_rate = samp_rate
        self.voiced_threshold = voiced_threshold

    def _f02uv(self, f0):
        # generate uv signal
        uv = (f0 > self.voiced_threshold).type(torch.float32)
        return uv

    @torch.no_grad()
    def forward(self, f0):
        """
        :param f0: [B, 1, sample_len], Hz
        :return: [B, 1, sample_len]
        """

        F_mat = torch.zeros((f0.size(0), self.harmonic_num + 1, f0.size(-1))).to(f0.device)
        for i in range(self.harmonic_num + 1):
            F_mat[:, i : i + 1, :] = f0 * (i + 1) / self.sampling_rate

        theta_mat = 2 * np.pi * (torch.cumsum(F_mat, dim=-1) % 1)
        u_dist = Uniform(low=-np.pi, high=np.pi)
        phase_vec = u_dist.sample(sample_shape=(f0.size(0), self.harmonic_num + 1, 1)).to(F_mat.device)
        phase_vec[:, 0, :] = 0

        # generate sine waveforms
        sine_waves = self.sine_amp * torch.sin(theta_mat + phase_vec)

        # generate uv signal
        uv = self._f02uv(f0)

        # noise: for unvoiced should be similar to sine_amp
        #        std = self.sine_amp/3 -> max value ~ self.sine_amp
        # .       for voiced regions is self.noise_std
        noise_amp = uv * self.noise_std + (1 - uv) * self.sine_amp / 3
        noise = noise_amp * torch.randn_like(sine_waves)

        # first: set the unvoiced part to 0 by uv
        # then: additive noise
        sine_waves = sine_waves * uv + noise
        return sine_waves, uv, noise


class SourceModuleHnNSF(torch.nn.Module):
    """SourceModule for hn-nsf
    SourceModule(sampling_rate, harmonic_num=0, sine_amp=0.1,
                 add_noise_std=0.003, voiced_threshod=0)
    sampling_rate: sampling_rate in Hz
    harmonic_num: number of harmonic above F0 (default: 0)
    sine_amp: amplitude of sine source signal (default: 0.1)
    add_noise_std: std of additive Gaussian noise (default: 0.003)
        note that amplitude of noise in unvoiced is decided
        by sine_amp
    voiced_threshold: threhold to set U/V given F0 (default: 0)
    Sine_source, noise_source = SourceModuleHnNSF(F0_sampled)
    F0_sampled (batchsize, length, 1)
    Sine_source (batchsize, length, 1)
    noise_source (batchsize, length 1)
    uv (batchsize, length, 1)
    """

    def __init__(
        self, sampling_rate, upsample_scale, harmonic_num=0, sine_amp=0.1, add_noise_std=0.003, voiced_threshod=0
    ):
        super(SourceModuleHnNSF, self).__init__()

        self.sine_amp = sine_amp
        self.noise_std = add_noise_std

        # to produce sine waveforms
        self.l_sin_gen = SineGen(sampling_rate, harmonic_num, sine_amp, add_noise_std, voiced_threshod)

        # to merge source harmonics into a single excitation
        self.l_linear = torch.nn.Linear(harmonic_num + 1, 1)
        self.l_tanh = torch.nn.Tanh()

    def forward(self, x):
        """
        Sine_source, noise_source = SourceModuleHnNSF(F0_sampled)
        F0_sampled (batchsize, length, 1)
        Sine_source (batchsize, length, 1)
        noise_source (batchsize, length 1)
        """
        # source for harmonic branch
        with torch.no_grad():
            sine_wavs, uv, _ = self.l_sin_gen(x.transpose(1, 2))
            sine_wavs = sine_wavs.transpose(1, 2)
            uv = uv.transpose(1, 2)
        sine_merge = self.l_tanh(self.l_linear(sine_wavs))

        # source for noise branch, in the same shape as uv
        noise = torch.randn_like(uv) * self.sine_amp / 3
        return sine_merge, noise, uv


class SineGen2(torch.nn.Module):
    """Definition of sine generator
    SineGen(samp_rate, harmonic_num = 0,
            sine_amp = 0.1, noise_std = 0.003,
            voiced_threshold = 0,
            flag_for_pulse=False)
    samp_rate: sampling rate in Hz
    harmonic_num: number of harmonic overtones (default 0)
    sine_amp: amplitude of sine-wavefrom (default 0.1)
    noise_std: std of Gaussian noise (default 0.003)
    voiced_thoreshold: F0 threshold for U/V classification (default 0)
    flag_for_pulse: this SinGen is used inside PulseGen (default False)
    Note: when flag_for_pulse is True, the first time step of a voiced
        segment is always sin(np.pi) or cos(0)
    """

    def __init__(
        self,
        samp_rate,
        upsample_scale,
        harmonic_num=0,
        sine_amp=0.1,
        noise_std=0.003,
        voiced_threshold=0,
        flag_for_pulse=False,
    ):
        super(SineGen2, self).__init__()
        self.sine_amp = sine_amp
        self.noise_std = noise_std
        self.harmonic_num = harmonic_num
        self.dim = self.harmonic_num + 1
        self.sampling_rate = samp_rate
        self.voiced_threshold = voiced_threshold
        self.flag_for_pulse = flag_for_pulse
        self.upsample_scale = upsample_scale

    def _f02uv(self, f0):
        # generate uv signal
        uv = (f0 > self.voiced_threshold).type(torch.float32)
        return uv

    def _f02sine(self, f0_values):
        """f0_values: (batchsize, length, dim)
        where dim indicates fundamental tone and overtones
        """
        # convert to F0 in rad. The interger part n can be ignored
        # because 2 * np.pi * n doesn't affect phase
        rad_values = (f0_values / self.sampling_rate) % 1

        # initial phase noise (no noise for fundamental component)
        rand_ini = torch.rand(f0_values.shape[0], f0_values.shape[2], device=f0_values.device)
        rand_ini[:, 0] = 0
        rad_values[:, 0, :] = rad_values[:, 0, :] + rand_ini

        # instantanouse phase sine[t] = sin(2*pi \sum_i=1 ^{t} rad)
        if not self.flag_for_pulse:
            rad_values = torch.nn.functional.interpolate(
                rad_values.transpose(1, 2), scale_factor=1 / self.upsample_scale, mode="linear"
            ).transpose(1, 2)

            phase = torch.cumsum(rad_values, dim=1) * 2 * np.pi
            phase = torch.nn.functional.interpolate(
                phase.transpose(1, 2) * self.upsample_scale, scale_factor=self.upsample_scale, mode="linear"
            ).transpose(1, 2)
            sines = torch.sin(phase)
        else:
            # If necessary, make sure that the first time step of every
            # voiced segments is sin(pi) or cos(0)
            # This is used for pulse-train generation

            # identify the last time step in unvoiced segments
            uv = self._f02uv(f0_values)
            uv_1 = torch.roll(uv, shifts=-1, dims=1)
            uv_1[:, -1, :] = 1
            u_loc = (uv < 1) * (uv_1 > 0)

            # get the instantanouse phase
            tmp_cumsum = torch.cumsum(rad_values, dim=1)
            # different batch needs to be processed differently
            for idx in range(f0_values.shape[0]):
                temp_sum = tmp_cumsum[idx, u_loc[idx, :, 0], :]
                temp_sum[1:, :] = temp_sum[1:, :] - temp_sum[0:-1, :]
                # stores the accumulation of i.phase within
                # each voiced segments
                tmp_cumsum[idx, :, :] = 0
                tmp_cumsum[idx, u_loc[idx, :, 0], :] = temp_sum

            # rad_values - tmp_cumsum: remove the accumulation of i.phase
            # within the previous voiced segment.
            i_phase = torch.cumsum(rad_values - tmp_cumsum, dim=1)

            # get the sines
            sines = torch.cos(i_phase * 2 * np.pi)
        return sines

    def forward(self, f0):
        """sine_tensor, uv = forward(f0)
        input F0: tensor(batchsize=1, length, dim=1)
                  f0 for unvoiced steps should be 0
        output sine_tensor: tensor(batchsize=1, length, dim)
        output uv: tensor(batchsize=1, length, 1)
        """
        # fundamental component
        fn = torch.multiply(f0, torch.FloatTensor([[range(1, self.harmonic_num + 2)]]).to(f0.device))

        # generate sine waveforms
        sine_waves = self._f02sine(fn) * self.sine_amp

        # generate uv signal
        uv = self._f02uv(f0)

        # noise: for unvoiced should be similar to sine_amp
        #        std = self.sine_amp/3 -> max value ~ self.sine_amp
        # .       for voiced regions is self.noise_std
        noise_amp = uv * self.noise_std + (1 - uv) * self.sine_amp / 3
        noise = noise_amp * torch.randn_like(sine_waves)

        # first: set the unvoiced part to 0 by uv
        # then: additive noise
        sine_waves = sine_waves * uv + noise
        return sine_waves, uv, noise


class SourceModuleHnNSF2(torch.nn.Module):
    """SourceModule for hn-nsf
    SourceModule(sampling_rate, harmonic_num=0, sine_amp=0.1,
                 add_noise_std=0.003, voiced_threshod=0)
    sampling_rate: sampling_rate in Hz
    harmonic_num: number of harmonic above F0 (default: 0)
    sine_amp: amplitude of sine source signal (default: 0.1)
    add_noise_std: std of additive Gaussian noise (default: 0.003)
        note that amplitude of noise in unvoiced is decided
        by sine_amp
    voiced_threshold: threhold to set U/V given F0 (default: 0)
    Sine_source, noise_source = SourceModuleHnNSF(F0_sampled)
    F0_sampled (batchsize, length, 1)
    Sine_source (batchsize, length, 1)
    noise_source (batchsize, length 1)
    uv (batchsize, length, 1)
    """

    def __init__(
        self, sampling_rate, upsample_scale, harmonic_num=0, sine_amp=0.1, add_noise_std=0.003, voiced_threshod=0
    ):
        super(SourceModuleHnNSF2, self).__init__()

        self.sine_amp = sine_amp
        self.noise_std = add_noise_std

        # to produce sine waveforms
        self.l_sin_gen = SineGen2(sampling_rate, upsample_scale, harmonic_num, sine_amp, add_noise_std, voiced_threshod)

        # to merge source harmonics into a single excitation
        self.l_linear = torch.nn.Linear(harmonic_num + 1, 1)
        self.l_tanh = torch.nn.Tanh()

    def forward(self, x):
        """
        Sine_source, noise_source = SourceModuleHnNSF(F0_sampled)
        F0_sampled (batchsize, length, 1)
        Sine_source (batchsize, length, 1)
        noise_source (batchsize, length 1)
        """
        # source for harmonic branch
        with torch.no_grad():
            sine_wavs, uv, _ = self.l_sin_gen(x)
        sine_merge = self.l_tanh(self.l_linear(sine_wavs))

        # source for noise branch, in the same shape as uv
        noise = torch.randn_like(uv) * self.sine_amp / 3
        return sine_merge, noise, uv


class HiFTGenerator(nn.Module):
    """
    HiFTNet Generator: Neural Source Filter + ISTFTNet
    https://arxiv.org/abs/2309.09493
    """

    def __init__(
        self,
        in_channels: int = 80,
        base_channels: int = 512,
        nb_harmonics: int = 8,
        sampling_rate: int = 24000,
        nsf_alpha: float = 0.1,
        nsf_sigma: float = 0.003,
        nsf_voiced_threshold: float = 10,
        upsample_rates: list[int] = [8, 5, 3],
        upsample_kernel_sizes: list[int] = [16, 11, 7],
        istft_n_fft: int = 16,
        istft_hop_len: int = 4,
        resblock_kernel_sizes: list[int] = [3, 7, 11],
        resblock_dilation_sizes: list[list[int]] = [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
        source_resblock_kernel_sizes: list[int] = [7, 7, 11],
        source_resblock_dilation_sizes: list[list[int]] = [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
        lrelu_slope: float = 0.1,
        audio_limit: float = 0.99,
        f0_predictor_channels: int = 512,
    ):
        super(HiFTGenerator, self).__init__()

        self.out_channels = 1
        self.nb_harmonics = nb_harmonics
        self.sampling_rate = sampling_rate
        self.istft_n_fft = istft_n_fft
        self.istft_hop_len = istft_hop_len
        self.lrelu_slope = lrelu_slope
        self.audio_limit = audio_limit

        self.num_kernels = len(resblock_kernel_sizes)
        self.num_upsamples = len(upsample_rates)
        self.m_source = SourceModuleHnNSF2(
            sampling_rate=sampling_rate,
            upsample_scale=np.prod(upsample_rates) * istft_hop_len,
            harmonic_num=nb_harmonics,
            sine_amp=nsf_alpha,
            add_noise_std=nsf_sigma,
            voiced_threshod=nsf_voiced_threshold,
        )
        self.f0_upsamp = torch.nn.Upsample(scale_factor=np.prod(upsample_rates) * istft_hop_len)

        self.conv_pre = weight_norm(Conv1d(in_channels, base_channels, 7, 1, padding=3))

        # Up
        self.ups = nn.ModuleList()
        for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
            self.ups.append(
                weight_norm(
                    ConvTranspose1d(
                        base_channels // (2**i), base_channels // (2 ** (i + 1)), k, u, padding=(k - u) // 2
                    )
                )
            )

        # Down
        self.source_downs = nn.ModuleList()
        self.source_resblocks = nn.ModuleList()
        downsample_rates = [1] + upsample_rates[::-1][:-1]
        downsample_cum_rates = np.cumprod(downsample_rates)
        for i, (u, k, d) in enumerate(
            zip(downsample_cum_rates[::-1], source_resblock_kernel_sizes, source_resblock_dilation_sizes)
        ):
            if u == 1:
                self.source_downs.append(Conv1d(istft_n_fft + 2, base_channels // (2 ** (i + 1)), 1, 1))
            else:
                self.source_downs.append(
                    Conv1d(istft_n_fft + 2, base_channels // (2 ** (i + 1)), u * 2, u, padding=(u // 2))
                )

            self.source_resblocks.append(ResBlock(base_channels // (2 ** (i + 1)), k, d))

        self.resblocks = nn.ModuleList()
        for i in range(len(self.ups)):
            ch = base_channels // (2 ** (i + 1))
            for _, (k, d) in enumerate(zip(resblock_kernel_sizes, resblock_dilation_sizes)):
                self.resblocks.append(ResBlock(ch, k, d))

        self.conv_post = weight_norm(Conv1d(ch, istft_n_fft + 2, 7, 1, padding=3))

        self.ups.apply(init_weights)
        self.conv_post.apply(init_weights)
        self.reflection_pad = nn.ReflectionPad1d((1, 0))
        self.stft_window = torch.from_numpy(get_window("hann", istft_n_fft, fftbins=True).astype(np.float32))

        self.f0_predictor = ConvRNNF0Predictor(
            num_class=1, in_channels=in_channels, cond_channels=f0_predictor_channels
        )

    def remove_weight_norm(self):
        for layer in self.ups:
            remove_weight_norm(layer)
        for layer in self.resblocks:
            layer.remove_weight_norm()
        remove_weight_norm(self.conv_pre)
        remove_weight_norm(self.conv_post)
        self.m_source.remove_weight_norm()
        for layer in self.source_downs:
            remove_weight_norm(layer)
        for layer in self.source_resblocks:
            layer.remove_weight_norm()

    def _stft(self, x):
        spec = torch.stft(
            x,
            self.istft_n_fft,
            self.istft_hop_len,
            self.istft_n_fft,
            window=self.stft_window.to(x.device),
            return_complex=True,
        )
        spec = torch.view_as_real(spec)  # [B, F, TT, 2]
        return spec[..., 0], spec[..., 1]

    def _istft(self, magnitude, phase):
        magnitude = torch.clip(magnitude, max=1e2)
        real = magnitude * torch.cos(phase)
        img = magnitude * torch.sin(phase)
        inverse_transform = torch.istft(
            torch.complex(real, img),
            self.istft_n_fft,
            self.istft_hop_len,
            self.istft_n_fft,
            window=self.stft_window.to(magnitude.device),
        )
        return inverse_transform

    def decode(self, x: torch.Tensor, s: torch.Tensor = torch.zeros(1, 1, 0)) -> torch.Tensor:
        s_stft_real, s_stft_imag = self._stft(s.squeeze(1))
        s_stft = torch.cat([s_stft_real, s_stft_imag], dim=1)

        x = self.conv_pre(x)
        for i in range(self.num_upsamples):
            x = F.leaky_relu(x, self.lrelu_slope)
            x = self.ups[i](x)

            if i == self.num_upsamples - 1:
                x = self.reflection_pad(x)

            # fusion
            si = self.source_downs[i](s_stft)
            si = self.source_resblocks[i](si)
            x = x + si

            xs = None
            for j in range(self.num_kernels):
                if xs is None:
                    xs = self.resblocks[i * self.num_kernels + j](x)
                else:
                    xs += self.resblocks[i * self.num_kernels + j](x)
            x = xs / self.num_kernels

        x = F.leaky_relu(x)
        x = self.conv_post(x)
        magnitude = torch.exp(x[:, : self.istft_n_fft // 2 + 1, :])
        phase = torch.sin(x[:, self.istft_n_fft // 2 + 1 :, :])  # actually, sin is redundancy

        x = self._istft(magnitude, phase)
        x = torch.clamp(x, -self.audio_limit, self.audio_limit)
        return x

    def forward(self, speech_feat: torch.Tensor) -> Dict[str, Optional[torch.Tensor]]:
        speech_feat = speech_feat.transpose(1, 2)
        # mel->f0
        f0 = self.f0_predictor(speech_feat)
        # f0->source
        s = self.f0_upsamp(f0[:, None]).transpose(1, 2)  # bs,n,t
        s, _, _ = self.m_source(s)
        s = s.transpose(1, 2)
        # mel+source->speech
        generated_speech = self.decode(x=speech_feat, s=s)
        return generated_speech, f0

    @torch.inference_mode()
    def inference(self, speech_feat: torch.Tensor) -> torch.Tensor:
        # mel->f0
        f0 = self.f0_predictor(speech_feat)
        # f0->source
        s = self.f0_upsamp(f0[:, None]).transpose(1, 2)  # bs,n,t
        s, _, _ = self.m_source(s)
        s = s.transpose(1, 2)
        generated_speech = self.decode(x=speech_feat, s=s)
        return generated_speech

    def load_weights(self, weights_path: str):
        checkpoint = torch.load(weights_path, map_location="cpu")
        state_dict = {k.replace("generator.", ""): v for k, v in checkpoint.items()}
        self.load_state_dict(state_dict, strict=True)