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

Fix gapping bug in voice2voice clip

Browse Source
This commit is contained in:
James Betker 2021-12-29 14:44:46 -07:00
parent 51ce1b5007
commit 5ae7e0d9b0
1 changed files with 6 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
codes/models/gpt_voice/voice_voice_clip.py
View File

@ -60,8 +60,8 @@ class VoiceCLIP(nn.Module):
# Introduce a random gap between the two clips.
potential_gap = half_length // 4
if potential_gap > 0:
gap = random.randint(0, potential_gap)
gap = random.randint(0, potential_gap)
if gap > 0:
first_half = first_half[:, :, :-gap]
second_half = second_half[:, :, gap:]
@ -111,6 +111,7 @@ def register_voice_to_voice_clip(opt_net, opt):
if __name__ == '__main__':
clip = VoiceCLIP()
clip(torch.randn((2,80,200)),
torch.randint(0,200*1024,(2,)),
return_loss=True)
for k in range(1000):
clip(torch.randn((2,80,156)),
torch.randint(130*1024,156*1024,(2,)),
return_loss=True)