Commit e6eaabdb authored by Phil Wang's avatar Phil Wang
Browse files

revert changes to elu activation placement, as researcher @yigityu is hitting convergence issues

parent d06fcf7e

audiolm_pytorch/soundstream.py

+1 −3
Original line number Diff line number Diff line
@@ -246,10 +246,10 @@ 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, cycle_dilations = (1, 3, 9)):
@@ -258,7 +258,6 @@ def EncoderBlock(chan_in, chan_out, stride, cycle_dilations = (1, 3, 9)): 
        ResidualUnit(chan_in, chan_in, next(it)),

        ResidualUnit(chan_in, chan_in, next(it)),

        ResidualUnit(chan_in, chan_in, next(it)),

        nn.ELU(),

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

    )
@@ -269,7 +268,6 @@ def DecoderBlock(chan_in, chan_out, stride, cycle_dilations = (1, 3, 9)): 


    it = cycle(cycle_dilations)

    return nn.Sequential(

        nn.ELU(),

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

        ResidualUnit(chan_out, chan_out, next(it)),

        ResidualUnit(chan_out, chan_out, next(it)),

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 = '0.7.1',

  version = '0.7.2',

  license='MIT',

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

  author = 'Phil Wang',