mrq/DL-Art-School
1
3
Fork 5
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Zero out spectrogram code inputs initially.

Browse Source
This commit is contained in:
James Betker 2021-10-15 12:10:11 -06:00
parent 1d0b44ebc2
commit 62c8c5d93e
1 changed files with 4 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
codes/models/gpt_voice/unet_diffusion_vocoder_with_ref.py
View File

@ -19,7 +19,10 @@ class DiscreteSpectrogramConditioningBlock(nn.Module):
self.intg = nn.Sequential(nn.Conv1d(channels*2, channels*2, kernel_size=1),
normalization(channels*2),
nn.SiLU(),
nn.Conv1d(channels*2, channels, kernel_size=3, padding=1))
nn.Conv1d(channels*2, channels, kernel_size=3, padding=1),
normalization(channels),
nn.SiLU(),
zero_module(nn.Conv1d(channels, channels, kernel_size=1)))
"""
Embeds the given codes and concatenates them onto x. Return shape is the same as x.shape.