9cf2bcdb24f23a17ec11e69b8885851771dfd3d8,cube/models/vocoder.py,BeeCoder,learn,#BeeCoder#,121

Before Change

                    sys.stdout.flush()
            output_power = self._predict_one(mgc[mgc_index], runtime=False)
            // print(np.abs(signal_fft[mgc_index]))
            fft_pow = 20 * np.log10(np.maximum(1e-5, np.abs(signal_fft[mgc_index])))  // np.abs(signal_fft[mgc_index])
            min_level_db = -100.0
            fft_pow = np.clip((fft_pow - min_level_db) / -min_level_db, 0, 1)
            // print (fft_pow)
            // print (fft_angle)
            // print("")
            losses.append(dy.binary_log_loss(output_power, dy.inputVector(fft_pow)))

After Change

            pred_output = self._predict_one(mgc[mgc_index], runtime=False)

            if mgc_index != len(mgc) - 1:
                losses.append(dy.squared_distance(pred_output, dy.inputVector(
                    wave[mgc_index * self.UPSAMPLE_COUNT:mgc_index * self.UPSAMPLE_COUNT + self.UPSAMPLE_COUNT])))

            if len(losses) >= batch_size:
                loss = dy.esum(losses)
                total_loss += loss.value()
                loss.backward()

Instances