Commit 36c39540 authored by Phil Wang's avatar Phil Wang
Browse files

use the same activation placements as encodec, suggested by @inspirit

parent daeedb27

audiolm_pytorch/soundstream.py

+3 −1
Original line number Diff line number Diff line
@@ -224,10 +224,10 @@ class CausalConvTranspose1d(nn.Module): 


def ResidualUnit(chan_in, chan_out, dilation, kernel_size = 7):

    return Residual(nn.Sequential(

        nn.ELU(),

        CausalConv1d(chan_in, chan_out, kernel_size, dilation = dilation),

        nn.ELU(),

        CausalConv1d(chan_out, chan_out, 1),

        nn.ELU()

    ))



def EncoderBlock(chan_in, chan_out, stride):
@@ -235,6 +235,7 @@ def EncoderBlock(chan_in, chan_out, stride): 
        ResidualUnit(chan_in, chan_in, 1),

        ResidualUnit(chan_in, chan_in, 3),

        ResidualUnit(chan_in, chan_in, 9),

        nn.ELU(),

        CausalConv1d(chan_in, chan_out, 2 * stride, stride = stride)

    )
@@ -244,6 +245,7 @@ def DecoderBlock(chan_in, chan_out, stride): 
    output_padding = 0 if even_stride else 1



    return nn.Sequential(

        nn.ELU(),

        CausalConvTranspose1d(chan_in, chan_out, 2 * stride, stride = stride),

        ResidualUnit(chan_out, chan_out, 1),

        ResidualUnit(chan_out, chan_out, 3),

setup.py

+1 −1
Original line number Diff line number Diff line
@@ -3,7 +3,7 @@ from setuptools import setup, find_packages 
setup(

  name = 'audiolm-pytorch',

  packages = find_packages(exclude=[]),

  version = '0.5.1',

  version = '0.6.0',

  license='MIT',

  description = 'AudioLM - Language Modeling Approach to Audio Generation from Google Research - Pytorch',

  author = 'Phil Wang',