Class Base

models/base.py:75–152

class Base:
    def __init__(self):
        pass

    def spectrogram(self, input, eps=0.):
        (real, imag) = self.stft(input)
        return torch.clamp(real ** 2 + imag ** 2, eps, np.inf) ** 0.5

    def spectrogram_phase(self, input, eps=0.):
        (real, imag) = self.stft(input)
        mag = torch.clamp(real ** 2 + imag ** 2, eps, np.inf) ** 0.5
        cos = real / mag
        sin = imag / mag
        return mag, cos, sin


    def wav_to_spectrogram_phase(self, input, eps=1e-10):
        """Waveform to spectrogram.

        Args:
          input: (batch_size, segment_samples, channels_num)

        Outputs:
          output: (batch_size, channels_num, time_steps, freq_bins)
        """
        sp_list = []
        cos_list = []
        sin_list = []
        channels_num = input.shape[1]
        for channel in range(channels_num):
            mag, cos, sin = self.spectrogram_phase(input[:, channel, :], eps=eps)
            sp_list.append(mag)
            cos_list.append(cos)
            sin_list.append(sin)

        sps = torch.cat(sp_list, dim=1)
        coss = torch.cat(cos_list, dim=1)
        sins = torch.cat(sin_list, dim=1)
        return sps, coss, sins

    def wav_to_spectrogram(self, input, eps=0.):
        """Waveform to spectrogram.

        Args:
          input: (batch_size, segment_samples, channels_num)

        Outputs:
          output: (batch_size, channels_num, time_steps, freq_bins)
        """
        sp_list = []
        channels_num = input.shape[1]
        for channel in range(channels_num):
            sp_list.append(self.spectrogram(input[:, channel, :], eps=eps))

        output = torch.cat(sp_list, dim=1)
        return output