""" /* Copyright (c) 2023 Amazon Written by Jan Buethe */ /* Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met: - Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer. - Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution. THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. */ """ """STFT-based Loss modules.""" import torch import torch.nn.functional as F from torch import nn import numpy as np import torchaudio def get_window(win_name, win_length, *args, **kwargs): window_dict = { 'bartlett_window' : torch.bartlett_window, 'blackman_window' : torch.blackman_window, 'hamming_window' : torch.hamming_window, 'hann_window' : torch.hann_window, 'kaiser_window' : torch.kaiser_window } if not win_name in window_dict: raise ValueError() return window_dict[win_name](win_length, *args, **kwargs) def stft(x, fft_size, hop_size, win_length, window): """Perform STFT and convert to magnitude spectrogram. Args: x (Tensor): Input signal tensor (B, T). fft_size (int): FFT size. hop_size (int): Hop size. win_length (int): Window length. window (str): Window function type. Returns: Tensor: Magnitude spectrogram (B, #frames, fft_size // 2 + 1). """ win = get_window(window, win_length).to(x.device) x_stft = torch.stft(x, fft_size, hop_size, win_length, win, return_complex=True) return torch.clamp(torch.abs(x_stft), min=1e-7) def spectral_convergence_loss(Y_true, Y_pred): dims=list(range(1, len(Y_pred.shape))) return torch.mean(torch.norm(torch.abs(Y_true) - torch.abs(Y_pred), p="fro", dim=dims) / (torch.norm(Y_pred, p="fro", dim=dims) + 1e-6)) def log_magnitude_loss(Y_true, Y_pred): Y_true_log_abs = torch.log(torch.abs(Y_true) + 1e-15) Y_pred_log_abs = torch.log(torch.abs(Y_pred) + 1e-15) return torch.mean(torch.abs(Y_true_log_abs - Y_pred_log_abs)) def spectral_xcorr_loss(Y_true, Y_pred): Y_true = Y_true.abs() Y_pred = Y_pred.abs() dims=list(range(1, len(Y_pred.shape))) xcorr = torch.sum(Y_true * Y_pred, dim=dims) / torch.sqrt(torch.sum(Y_true ** 2, dim=dims) * torch.sum(Y_pred ** 2, dim=dims) + 1e-9) return 1 - xcorr.mean() class MRLogMelLoss(nn.Module): def __init__(self, fft_sizes=[512, 256, 128, 64], overlap=0.5, fs=16000, n_mels=18 ): self.fft_sizes = fft_sizes self.overlap = overlap self.fs = fs self.n_mels = n_mels super().__init__() self.mel_specs = [] for fft_size in fft_sizes: hop_size = int(round(fft_size * (1 - self.overlap))) n_mels = self.n_mels if fft_size < 128: n_mels //= 2 self.mel_specs.append(torchaudio.transforms.MelSpectrogram(fs, fft_size, hop_length=hop_size, n_mels=n_mels)) for i, mel_spec in enumerate(self.mel_specs): self.add_module(f'mel_spec_{i+1}', mel_spec) def forward(self, y_true, y_pred): loss = torch.zeros(1, device=y_true.device) for mel_spec in self.mel_specs: Y_true = mel_spec(y_true) Y_pred = mel_spec(y_pred) loss = loss + log_magnitude_loss(Y_true, Y_pred) loss = loss / len(self.mel_specs) return loss def create_weight_matrix(num_bins, bins_per_band=10): m = torch.zeros((num_bins, num_bins), dtype=torch.float32) r0 = bins_per_band // 2 r1 = bins_per_band - r0 for i in range(num_bins): i0 = max(i - r0, 0) j0 = min(i + r1, num_bins) m[i, i0: j0] += 1 if i < r0: m[i, :r0 - i] += 1 if i > num_bins - r1: m[i, num_bins - r1 - i:] += 1 return m / bins_per_band def weighted_spectral_convergence(Y_true, Y_pred, w): # calculate sfm based weights logY = torch.log(torch.abs(Y_true) + 1e-9) Y = torch.abs(Y_true) avg_logY = torch.matmul(logY.transpose(1, 2), w) avg_Y = torch.matmul(Y.transpose(1, 2), w) sfm = torch.exp(avg_logY) / (avg_Y + 1e-9) weight = (torch.relu(1 - sfm) ** .5).transpose(1, 2) loss = torch.mean( torch.mean(weight * torch.abs(torch.abs(Y_true) - torch.abs(Y_pred)), dim=[1, 2]) / (torch.mean( weight * torch.abs(Y_true), dim=[1, 2]) + 1e-9) ) return loss def gen_filterbank(N, Fs=16000): in_freq = (np.arange(N+1, dtype='float32')/N*Fs/2)[None,:] out_freq = (np.arange(N, dtype='float32')/N*Fs/2)[:,None] #ERB from B.C.J Moore, An Introduction to the Psychology of Hearing, 5th Ed., page 73. ERB_N = 24.7 + .108*in_freq delta = np.abs(in_freq-out_freq)/ERB_N center = (delta<.5).astype('float32') R = -12*center*delta**2 + (1-center)*(3-12*delta) RE = 10.**(R/10.) norm = np.sum(RE, axis=1) RE = RE/norm[:, np.newaxis] return torch.from_numpy(RE) def smooth_log_mag(Y_true, Y_pred, filterbank): Y_true_smooth = torch.matmul(filterbank, torch.abs(Y_true)) Y_pred_smooth = torch.matmul(filterbank, torch.abs(Y_pred)) loss = torch.abs( torch.log(Y_true_smooth + 1e-9) - torch.log(Y_pred_smooth + 1e-9) ) loss = loss.mean() return loss class MRSTFTLoss(nn.Module): def __init__(self, fft_sizes=[2048, 1024, 512, 256, 128, 64], overlap=0.5, window='hann_window', fs=16000, log_mag_weight=0, sc_weight=0, wsc_weight=0, smooth_log_mag_weight=2, sxcorr_weight=1): super().__init__() self.fft_sizes = fft_sizes self.overlap = overlap self.window = window self.log_mag_weight = log_mag_weight self.sc_weight = sc_weight self.wsc_weight = wsc_weight self.smooth_log_mag_weight = smooth_log_mag_weight self.sxcorr_weight = sxcorr_weight self.fs = fs # weights for SFM weighted spectral convergence loss self.wsc_weights = torch.nn.ParameterDict() for fft_size in fft_sizes: width = min(11, int(1000 * fft_size / self.fs + .5)) width += width % 2 self.wsc_weights[str(fft_size)] = torch.nn.Parameter( create_weight_matrix(fft_size // 2 + 1, width), requires_grad=False ) # filterbanks for smooth log magnitude loss self.filterbanks = torch.nn.ParameterDict() for fft_size in fft_sizes: self.filterbanks[str(fft_size)] = torch.nn.Parameter( gen_filterbank(fft_size//2), requires_grad=False ) def __call__(self, y_true, y_pred): lm_loss = torch.zeros(1, device=y_true.device) sc_loss = torch.zeros(1, device=y_true.device) wsc_loss = torch.zeros(1, device=y_true.device) slm_loss = torch.zeros(1, device=y_true.device) sxcorr_loss = torch.zeros(1, device=y_true.device) for fft_size in self.fft_sizes: hop_size = int(round(fft_size * (1 - self.overlap))) win_size = fft_size Y_true = stft(y_true, fft_size, hop_size, win_size, self.window) Y_pred = stft(y_pred, fft_size, hop_size, win_size, self.window) if self.log_mag_weight > 0: lm_loss = lm_loss + log_magnitude_loss(Y_true, Y_pred) if self.sc_weight > 0: sc_loss = sc_loss + spectral_convergence_loss(Y_true, Y_pred) if self.wsc_weight > 0: wsc_loss = wsc_loss + weighted_spectral_convergence(Y_true, Y_pred, self.wsc_weights[str(fft_size)]) if self.smooth_log_mag_weight > 0: slm_loss = slm_loss + smooth_log_mag(Y_true, Y_pred, self.filterbanks[str(fft_size)]) if self.sxcorr_weight > 0: sxcorr_loss = sxcorr_loss + spectral_xcorr_loss(Y_true, Y_pred) total_loss = (self.log_mag_weight * lm_loss + self.sc_weight * sc_loss + self.wsc_weight * wsc_loss + self.smooth_log_mag_weight * slm_loss + self.sxcorr_weight * sxcorr_loss) / len(self.fft_sizes) return total_loss def td_l2_norm(y_true, y_pred): dims = list(range(1, len(y_true.shape))) loss = torch.mean((y_true - y_pred) ** 2, dim=dims) / (torch.mean(y_pred ** 2, dim=dims) ** .5 + 1e-6) return loss.mean() class LaceLoss(nn.Module): def __init__(self): super().__init__() self.stftloss = MRSTFTLoss(log_mag_weight=0, sc_weight=0, wsc_weight=0, smooth_log_mag_weight=2, sxcorr_weight=1) def forward(self, x, y): specloss = self.stftloss(x, y) phaseloss = td_l2_norm(x, y) total_loss = (specloss + 10 * phaseloss) / 13 return total_loss def compare(self, x_ref, x_deg): # trim items to same size n = min(len(x_ref), len(x_deg)) x_ref = x_ref[:n].copy() x_deg = x_deg[:n].copy() # pre-emphasis x_ref[1:] -= 0.85 * x_ref[:-1] x_deg[1:] -= 0.85 * x_deg[:-1] device = next(iter(self.parameters())).device x = torch.from_numpy(x_ref).to(device) y = torch.from_numpy(x_deg).to(device) with torch.no_grad(): dist = 10 * self.forward(x, y) return dist.cpu().numpy().item() lace_loss = LaceLoss() device = 'cuda' if torch.cuda.is_available() else 'cpu' lace_loss.to(device) def compare(x, y): return lace_loss.compare(x, y)