Mirror/stable-diffusion-webui
1
0
Fork 0
mirror of https://github.com/AUTOMATIC1111/stable-diffusion-webui.git synced 2024-06-07 21:20:49 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #14300 from AUTOMATIC1111/oft_fixes

Browse Source 
Fix wrong implementation in network_oft
This commit is contained in:
AUTOMATIC1111 2023-12-14 09:39:57 +03:00 committed by GitHub
commit b943eebb1d
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 11 additions and 26 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

37
extensions-builtin/Lora/network_oft.py
View File

@ -21,6 +21,8 @@ class NetworkModuleOFT(network.NetworkModule):
self.lin_module = None self.lin_module = None
self.org_module: list[torch.Module] = [self.sd_module] self.org_module: list[torch.Module] = [self.sd_module]
self.scale = 1.0
# kohya-ss # kohya-ss
if "oft_blocks" in weights.w.keys(): if "oft_blocks" in weights.w.keys():
self.is_kohya = True self.is_kohya = True
@ -53,12 +55,18 @@ class NetworkModuleOFT(network.NetworkModule):
self.constraint = None self.constraint = None
self.block_size, self.num_blocks = factorization(self.out_dim, self.dim) self.block_size, self.num_blocks = factorization(self.out_dim, self.dim)
def calc_updown_kb(self, orig_weight, multiplier): def calc_updown(self, orig_weight):
oft_blocks = self.oft_blocks.to(orig_weight.device, dtype=orig_weight.dtype) oft_blocks = self.oft_blocks.to(orig_weight.device, dtype=orig_weight.dtype)
oft_blocks = oft_blocks - oft_blocks.transpose(1, 2) # ensure skew-symmetric orthogonal matrix eye = torch.eye(self.block_size, device=self.oft_blocks.device)
if self.is_kohya:
block_Q = oft_blocks - oft_blocks.transpose(1, 2) # ensure skew-symmetric orthogonal matrix
norm_Q = torch.norm(block_Q.flatten())
new_norm_Q = torch.clamp(norm_Q, max=self.constraint)
block_Q = block_Q * ((new_norm_Q + 1e-8) / (norm_Q + 1e-8))
oft_blocks = torch.matmul(eye + block_Q, (eye - block_Q).float().inverse())
R = oft_blocks.to(orig_weight.device, dtype=orig_weight.dtype) R = oft_blocks.to(orig_weight.device, dtype=orig_weight.dtype)
R = R * multiplier + torch.eye(self.block_size, device=orig_weight.device)
# This errors out for MultiheadAttention, might need to be handled up-stream # This errors out for MultiheadAttention, might need to be handled up-stream
merged_weight = rearrange(orig_weight, '(k n) ... -> k n ...', k=self.num_blocks, n=self.block_size) merged_weight = rearrange(orig_weight, '(k n) ... -> k n ...', k=self.num_blocks, n=self.block_size)
@ -72,26 +80,3 @@ class NetworkModuleOFT(network.NetworkModule):
updown = merged_weight.to(orig_weight.device, dtype=orig_weight.dtype) - orig_weight updown = merged_weight.to(orig_weight.device, dtype=orig_weight.dtype) - orig_weight
output_shape = orig_weight.shape output_shape = orig_weight.shape
return self.finalize_updown(updown, orig_weight, output_shape) return self.finalize_updown(updown, orig_weight, output_shape)
def calc_updown(self, orig_weight):
# if alpha is a very small number as in coft, calc_scale() will return a almost zero number so we ignore it
multiplier = self.multiplier()
return self.calc_updown_kb(orig_weight, multiplier)
# override to remove the multiplier/scale factor; it's already multiplied in get_weight
def finalize_updown(self, updown, orig_weight, output_shape, ex_bias=None):
if self.bias is not None:
updown = updown.reshape(self.bias.shape)
updown += self.bias.to(orig_weight.device, dtype=orig_weight.dtype)
updown = updown.reshape(output_shape)
if len(output_shape) == 4:
updown = updown.reshape(output_shape)
if orig_weight.size().numel() == updown.size().numel():
updown = updown.reshape(orig_weight.shape)
if ex_bias is not None:
ex_bias = ex_bias * self.multiplier()
return updown, ex_bias