Frequency dimension ISSUE

[StefanoGiacomelli]

Hi, and first of all congratulations for the splendid job. I was "playing" with the example you provided in the main page of the repository and i found a bug in the DiffRes model initialization. I simply changed the frequency resolution of the data generator and of the in_f_dim attribute too (to 512, or generally bigger than 128) and the model output an error about target tensor dimensions, during inference.

I suppose it's simply a dimensionality initialization issue (easy to fix), but I take this opportunity to ask: wouldn't it be better to accept any dimension for the frequency axis? It should also be possible to process standard spectra of positive modules (I am referring to dimensions such as "dft_size // 2 + 1") or any kind of filtered (arbitrary channels) data.

Thank you