Unverified Commit 4b893421 authored by djqualia's avatar djqualia Committed by GitHub
Browse files

Pad right side of audio vs. left 

This is more natural when listening to generated samples while training sound stream (prevents long silences when audio source < data_max_length).

Please confirm this doesn't have negative consequences on expectations elsewhere...
parent 60e4cd84

audiolm_pytorch/data.py

+1 −1
Original line number Diff line number Diff line
@@ -61,7 +61,7 @@ class SoundDataset(Dataset): 
            data = data[:, start:start + self.max_length]



        else:

            data = torch.nn.functional.pad(data, (self.max_length - data.size(1), 0), 'constant')

            data = torch.nn.functional.pad(data, (0, self.max_length - data.size(1)), 'constant')

        

        data = rearrange(data, '1 ... -> ...')