Commit 241bae31 authored by Phil Wang's avatar Phil Wang
Browse files

address issue in soundstream with causal conv padding not taking into account... 

address issue in soundstream with causal conv padding not taking into account strides https://github.com/lucidrains/audiolm-pytorch/issues/166
parent e50305de

README.md

+2 −0
Original line number Diff line number Diff line
@@ -36,6 +36,8 @@ In the future, <a href="https://www.youtube.com/watch?v=olNvmUCmY8o">this movie 


- <a href="https://github.com/LWprogramming">LWprogramming</a> for adding Encodec compatibility!



- <a href="https://github.com/YoungloLee">@YoungloLee</a> for identifying a big bug in the 1d causal convolution for soundstream related to padding not accounting for strides!



## Install



```bash

audiolm_pytorch/soundstream.py

+3 −2
Original line number Diff line number Diff line
@@ -305,12 +305,13 @@ class CausalConv1d(nn.Module): 
        super().__init__()

        kernel_size = kernel_size

        dilation = kwargs.get('dilation', 1)

        self.causal_padding = dilation * (kernel_size - 1)

        stride = kwargs.get('stride', 1)

        self.causal_padding = dilation * (kernel_size - 1) + (1 - stride)



        self.conv = nn.Conv1d(chan_in, chan_out, kernel_size, **kwargs)



    def forward(self, x):

        x = F.pad(x, (self.causal_padding, 0))

        x = F.pad(x, (self.causal_padding, 0), mode = 'reflect')

        return self.conv(x)



class CausalConvTranspose1d(nn.Module):

audiolm_pytorch/trainer.py

+1 −1
Original line number Diff line number Diff line
@@ -122,7 +122,7 @@ class SoundStreamTrainer(nn.Module): 
        num_train_steps: int,

        batch_size: int,

        data_max_length: int = None,

        data_max_length_seconds: float = None,

        data_max_length_seconds: Union[int, float] = None,

        folder: str = None,

        train_dataloader: DataLoader = None,

        val_dataloader: DataLoader = None,

audiolm_pytorch/version.py

+1 −1
Original line number Diff line number Diff line 
__version__ = '0.26.2'
 
__version__ = '0.26.5'