Why wrong code can work?

[poor1017]

Hi @qiuqiangkong

I notice you modified the code of resunet_subbandtime.py a couple of days ago.

Now the code is:

separated_subband_audio = torch.stack(
    [
        self.feature_maps_to_wav(
            input_tensor=x[:, j :: self.subbands_num, :, :],
            sp=mag[:, j :: self.subbands_num, :, :],
            sin_in=sin_in[:, j :: self.subbands_num, :, :],
            cos_in=cos_in[:, j :: self.subbands_num, :, :],
            audio_length=audio_length,
        )
        for j in range(self.subbands_num)
    ],
    dim=2,
)

I think this change is for pqmf.synthesis compatible.

The old code is:

separated_subband_audio = torch.cat(
    [
        self.feature_maps_to_wav(
            input_tensor=net_output[:, j * C1 : (j + 1) * C1, :, :],
            sp=mag[:, j * C2 : (j + 1) * C2, :, :],
            sin_in=sin_in[:, j * C2 : (j + 1) * C2, :, :],
            cos_in=cos_in[:, j * C2 : (j + 1) * C2, :, :],
            audio_length=audio_length,
        )
        for j in range(self.subbands_num)
    ],
    dim=1,
)

The old code is not compatible with pqmf.synthesis. pqmf.synthesis assumes subband first, while the resulting tensor generated by the old code is channel first.

It is strange that the old code still works, and can get pretty good model. Do you know the reason?

Thanks