modeling_hifigan.py

# --------------------------------------------------------
# SenseTime
# Copyright (c) 2025 SenseTime
# Licensed under The MIT License [see LICENSE for details]
# --------------------------------------------------------
from typing import Dict, Optional, List
import numpy as np
from scipy.signal import get_window
import torch
import torch.nn.functional as F
from torch.nn import ConvTranspose1d, Conv1d, Parameter
from torch.nn.utils import remove_weight_norm
from torch.nn.utils.parametrizations import weight_norm
from torch.distributions.uniform import Uniform
from torch import nn, sin, pow
from transformers.modeling_utils import PreTrainedModel

from .configuration_hifigan import HiFiGanConfig

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)
    return

# Implementation adapted from https://github.com/EdwardDixon/snake under the MIT license.
#   LICENSE is in incl_licenses directory.
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 = Parameter(torch.zeros(in_features) * alpha)
        else:  # linear scale alphas initialized to ones
            self.alpha = 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)) * pow(sin(x * alpha), 2)

        return x
    
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 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 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_wavs = sine_wavs.to(self.l_linear.weight.data.dtype)  # noqa, TODO
        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(PreTrainedModel):
    """
    HiFTNet Generator: Neural Source Filter + ISTFTNet
    https://arxiv.org/abs/2309.09493
    """
    def __init__(
            self,
            config: HiFiGanConfig
    ):
        super(HiFTGenerator, self).__init__(config)

        self.out_channels = 1
        self.nb_harmonics = config.nb_harmonics
        self.sampling_rate = config.sampling_rate
        self.istft_params = config.istft_params
        self.lrelu_slope = config.lrelu_slope
        self.audio_limit = config.audio_limit

        self.num_kernels = len(config.resblock_kernel_sizes)
        self.num_upsamples = len(config.upsample_rates)
        self.m_source = SourceModuleHnNSF(
            sampling_rate=config.sampling_rate,
            upsample_scale=np.prod(config.upsample_rates) * config.istft_params["hop_len"],
            harmonic_num=config.nb_harmonics,
            sine_amp=config.nsf_alpha,
            add_noise_std=config.nsf_sigma,
            voiced_threshod=config.nsf_voiced_threshold)
        self.f0_upsamp = torch.nn.Upsample(scale_factor=np.prod(config.upsample_rates) * config.istft_params["hop_len"])

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

        # Up
        self.ups = nn.ModuleList()
        for i, (u, k) in enumerate(zip(config.upsample_rates, config.upsample_kernel_sizes)):
            self.ups.append(
                weight_norm(
                    ConvTranspose1d(
                        config.base_channels // (2**i),
                        config.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] + config.upsample_rates[::-1][:-1]
        downsample_cum_rates = np.cumprod(downsample_rates)
        for i, (u, k, d) in enumerate(zip(downsample_cum_rates[::-1], config.source_resblock_kernel_sizes, config.source_resblock_dilation_sizes)):
            if u == 1:
                self.source_downs.append(
                    Conv1d(config.istft_params["n_fft"] + 2, config.base_channels // (2 ** (i + 1)), 1, 1)
                )
            else:
                self.source_downs.append(
                    Conv1d(config.istft_params["n_fft"] + 2, config.base_channels // (2 ** (i + 1)), u * 2, u, padding=(u // 2))
                )

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

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

        self.conv_post = weight_norm(Conv1d(ch, config.istft_params["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", config.istft_params["n_fft"], fftbins=True).astype(np.float32))
        self.f0_predictor = ConvRNNF0Predictor(**config.f0_predictor_config)

    def remove_weight_norm(self):
        print('Removing weight norm...')
        for l in self.ups:
            remove_weight_norm(l)
        for l in self.resblocks:
            l.remove_weight_norm()
        remove_weight_norm(self.conv_pre)
        remove_weight_norm(self.conv_post)
        self.m_source.remove_weight_norm()
        for l in self.source_downs:
            remove_weight_norm(l)
        for l in self.source_resblocks:
            l.remove_weight_norm()

    def _stft(self, x):
        spec = torch.stft(
            x,
            self.istft_params["n_fft"], self.istft_params["hop_len"], self.istft_params["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_params["n_fft"], self.istft_params["hop_len"],
                                        self.istft_params["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)
        s_stft = s_stft.to(x)  # noqa TODO
        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_params["n_fft"] // 2 + 1, :])
        phase = torch.sin(x[:, self.istft_params["n_fft"] // 2 + 1:, :])  # actually, sin is redundancy

        magnitude = magnitude.to(torch.float)  # noqa TODO
        phase = phase.to(torch.float)  # noqa TODO
        
        x = self._istft(magnitude, phase)
        x = torch.clamp(x, -self.audio_limit, self.audio_limit)
        return x

    def forward(
            self,
            batch: dict,
            device: torch.device,
    ) -> Dict[str, Optional[torch.Tensor]]:
        speech_feat = batch['speech_feat'].transpose(1, 2).to(device)
        # 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, cache_source: torch.Tensor = torch.zeros(1, 1, 0)) -> torch.Tensor:
        # process data
        speech_feat = speech_feat.to(self.f0_predictor.classifier.weight.data.dtype)  # noqa, TODO
        cache_source = cache_source.to(self.f0_predictor.classifier.weight.data.dtype)  # noqa, TODO
        # 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)
        # use cache_source to avoid glitch
        if cache_source.shape[2] != 0:
            s[:, :, :cache_source.shape[2]] = cache_source
        generated_speech = self.decode(x=speech_feat, s=s)
        return generated_speech, s