In-place operations can break gradient calculation

This commit is contained in:
brkirch 2023-01-07 07:04:59 -05:00
parent 151233399c
commit df3b31eb55

View File

@ -247,9 +247,9 @@ class FrozenCLIPEmbedderWithCustomWordsBase(torch.nn.Module):
# restoring original mean is likely not correct, but it seems to work well to prevent artifacts that happen otherwise
batch_multipliers = torch.asarray(batch_multipliers).to(devices.device)
original_mean = z.mean()
z *= batch_multipliers.reshape(batch_multipliers.shape + (1,)).expand(z.shape)
z = z * batch_multipliers.reshape(batch_multipliers.shape + (1,)).expand(z.shape)
new_mean = z.mean()
z *= original_mean / new_mean
z = z * (original_mean / new_mean)
return z