Closed michaelschleiss closed 4 years ago
You can modify self.project
to self.project.float()
. It will work. Thanks~
More specifically, x = F.linear(x, self.project).reshape(-1, 4)
-> x = F.linear(x, self.project.float()).reshape(-1, 4)
Thanks, it worked.
When I try to train with fp16 precision I get the following errors:
Is there a way to fix this?