models/mel.py

import torch
import torch.nn as nn
import torchaudio


class AugmentMelSTFT(nn.Module):
    def __init__(self, n_mels=128, sr=32000, win_length=800, hopsize=320, n_fft=1024, freqm=48, timem=192,
                 fmin=0.0, fmax=None, fmin_aug_range=10, fmax_aug_range=2000, norm_mel=False):
        """
        :param n_mels: number of mel bins
        :param sr: sampling rate used (same as passed as argument to dataset)
        :param win_length: fft window length in samples
        :param hopsize: fft hop size in samples
        :param n_fft: length of fft
        :param freqm: maximum possible length of mask along frequency dimension
        :param timem: maximum possible length of mask along time dimension
        :param fmin: minimum frequency used
        :param fmax: maximum frequency used
        :param fmin_aug_range: randomly changes min frequency
        :param fmax_aug_range: randomly changes max frequency
        """
        torch.nn.Module.__init__(self)
        # adapted from: https://github.com/CPJKU/kagglebirds2020/commit/70f8308b39011b09d41eb0f4ace5aa7d2b0e806e

        self.win_length = win_length
        self.n_mels = n_mels
        self.n_fft = n_fft
        self.sr = sr
        self.fmin = fmin
        if fmax is None:
            # fmax is by default set to sampling_rate/2 -> Nyquist!
            fmax = sr // 2 - fmax_aug_range // 2
            print(f"Warning: FMAX is None setting to {fmax} ")
        self.fmax = fmax
        self.hopsize = hopsize
        # buffers are not trained by the optimizer, persistent=False also avoids adding it to the model's state dict
        self.register_buffer('window',
                             torch.hann_window(win_length, periodic=False),
                             persistent=False)
        assert fmin_aug_range >= 1, f"fmin_aug_range={fmin_aug_range} should be >=1; 1 means no augmentation"
        assert fmax_aug_range >= 1, f"fmax_aug_range={fmax_aug_range} should be >=1; 1 means no augmentation"
        self.fmin_aug_range = fmin_aug_range
        self.fmax_aug_range = fmax_aug_range

        self.register_buffer("preemphasis_coefficient", torch.as_tensor([[[-.97, 1]]]), persistent=False)
        if freqm == 0:
            self.freqm = torch.nn.Identity()
        else:
            self.freqm = torchaudio.transforms.FrequencyMasking(freqm, iid_masks=True)
        if timem == 0:
            self.timem = torch.nn.Identity()
        else:
            self.timem = torchaudio.transforms.TimeMasking(timem, iid_masks=True)
        self.norm_mel = norm_mel

    def forward(self, x):
        # shape: batch size x samples
        # majority of energy located in lower end of the spectrum, pre-emphasis compensates for the average spectral
        # shape
        x = nn.functional.conv1d(x.unsqueeze(1), self.preemphasis_coefficient).squeeze(1)

        # Short-Time Fourier Transform using Hanning window
        x = torch.stft(x, self.n_fft, hop_length=self.hopsize, win_length=self.win_length,
                       center=True, normalized=False, window=self.window, return_complex=False)
        # shape: batch size x freqs (n_fft/2 + 1) x timeframes (samples/hop_length) x 2 (real and imaginary components)

        # calculate power spectrum
        x = (x ** 2).sum(dim=-1)
        fmin = self.fmin + torch.randint(self.fmin_aug_range, (1,)).item()
        fmax = self.fmax + self.fmax_aug_range // 2 - torch.randint(self.fmax_aug_range, (1,)).item()

        if not self.training:
            # don't augment eval data
            fmin = self.fmin
            fmax = self.fmax

        # create mel filterbank
        mel_basis, _ = torchaudio.compliance.kaldi.get_mel_banks(self.n_mels, self.n_fft, self.sr,
                                                                 fmin, fmax, vtln_low=100.0, vtln_high=-500.,
                                                                 vtln_warp_factor=1.0)
        mel_basis = torch.as_tensor(torch.nn.functional.pad(mel_basis, (0, 1), mode='constant', value=0),
                                    device=x.device)
        if self.norm_mel:
            mel_basis = mel_basis / mel_basis.sum(1)[:, None]
        # apply mel filterbank to power spectrogram
        with torch.cuda.amp.autocast(enabled=False):
            melspec = torch.matmul(mel_basis, x)
        # calculate log mel spectrogram
        melspec = (melspec + 0.00001).log()

        if self.training:
            # don't augment eval data
            melspec = self.freqm(melspec)
            melspec = self.timem(melspec)

        melspec = (melspec + 4.5) / 5.  # fast normalization
        return melspec