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

refactor: fix constraint, re-use get_weight

Browse Source
This commit is contained in:
v0xie 2023-10-19 12:41:17 -07:00
parent eb01d7f0e0
commit 321680ccd0
1 changed files with 16 additions and 24 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

40
extensions-builtin/Lora/network_oft.py
View File

@ -9,7 +9,7 @@ class ModuleTypeOFT(network.ModuleType):
return None return None
# adapted from https://github.com/kohya-ss/sd-scripts/blob/main/networks/oft.py # adapted from kohya's implementation https://github.com/kohya-ss/sd-scripts/blob/main/networks/oft.py
class NetworkModuleOFT(network.NetworkModule): class NetworkModuleOFT(network.NetworkModule):
def __init__(self, net: network.Network, weights: network.NetworkWeights): def __init__(self, net: network.Network, weights: network.NetworkWeights):
@ -17,7 +17,6 @@ class NetworkModuleOFT(network.NetworkModule):
self.oft_blocks = weights.w["oft_blocks"] self.oft_blocks = weights.w["oft_blocks"]
self.alpha = weights.w["alpha"] self.alpha = weights.w["alpha"]
self.dim = self.oft_blocks.shape[0] self.dim = self.oft_blocks.shape[0]
self.num_blocks = self.dim self.num_blocks = self.dim
@ -26,64 +25,57 @@ class NetworkModuleOFT(network.NetworkModule):
elif "Conv" in self.sd_module.__class__.__name__: elif "Conv" in self.sd_module.__class__.__name__:
self.out_dim = self.sd_module.out_channels self.out_dim = self.sd_module.out_channels
self.constraint = self.alpha self.constraint = self.alpha * self.out_dim
#self.constraint = self.alpha * self.out_dim
self.block_size = self.out_dim // self.num_blocks self.block_size = self.out_dim // self.num_blocks
self.org_module: list[torch.Module] = [self.sd_module] self.org_module: list[torch.Module] = [self.sd_module]
self.R = self.get_weight(self.oft_blocks)
self.R = self.get_weight()
self.apply_to() self.apply_to()
# replace forward method of original linear rather than replacing the module # replace forward method of original linear rather than replacing the module
# how do we revert this to unload the weights?
def apply_to(self): def apply_to(self):
self.org_forward = self.org_module[0].forward self.org_forward = self.org_module[0].forward
self.org_module[0].forward = self.forward self.org_module[0].forward = self.forward
def get_weight(self, multiplier=None): def get_weight(self, oft_blocks, multiplier=None):
if not multiplier: block_Q = oft_blocks - oft_blocks.transpose(1, 2)
multiplier = self.multiplier()
block_Q = self.oft_blocks - self.oft_blocks.transpose(1, 2)
norm_Q = torch.norm(block_Q.flatten()) norm_Q = torch.norm(block_Q.flatten())
new_norm_Q = torch.clamp(norm_Q, max=self.constraint) new_norm_Q = torch.clamp(norm_Q, max=self.constraint)
block_Q = block_Q * ((new_norm_Q + 1e-8) / (norm_Q + 1e-8)) block_Q = block_Q * ((new_norm_Q + 1e-8) / (norm_Q + 1e-8))
I = torch.eye(self.block_size, device=self.oft_blocks.device).unsqueeze(0).repeat(self.num_blocks, 1, 1) I = torch.eye(self.block_size, device=self.oft_blocks.device).unsqueeze(0).repeat(self.num_blocks, 1, 1)
block_R = torch.matmul(I + block_Q, (I - block_Q).inverse()) block_R = torch.matmul(I + block_Q, (I - block_Q).inverse())
#block_R_weighted = multiplier * block_R + (1 - multiplier) * I
block_R_weighted = multiplier * block_R + (1 - multiplier) * I #R = torch.block_diag(*block_R_weighted)
R = torch.block_diag(*block_R_weighted) R = torch.block_diag(*block_R)
return R return R
def calc_updown(self, orig_weight): def calc_updown(self, orig_weight):
# this works oft_blocks = self.oft_blocks.to(orig_weight.device, dtype=orig_weight.dtype)
# R = self.R
self.R = self.get_weight(self.multiplier())
# sending R to device causes major deepfrying i.e. just doesn't work R = self.get_weight(oft_blocks)
# R = self.R.to(orig_weight.device, dtype=orig_weight.dtype) self.R = R
# if orig_weight.dim() == 4: # if orig_weight.dim() == 4:
# weight = torch.einsum("oihw, op -> pihw", orig_weight, R) # weight = torch.einsum("oihw, op -> pihw", orig_weight, R)
# else: # else:
# weight = torch.einsum("oi, op -> pi", orig_weight, R) # weight = torch.einsum("oi, op -> pi", orig_weight, R)
updown = orig_weight @ self.R updown = orig_weight @ R
output_shape = self.oft_blocks.shape output_shape = self.oft_blocks.shape
## this works
# updown = orig_weight @ R
# output_shape = [orig_weight.size(0), R.size(1)]
return self.finalize_updown(updown, orig_weight, output_shape) return self.finalize_updown(updown, orig_weight, output_shape)
def forward(self, x, y=None): def forward(self, x, y=None):
x = self.org_forward(x) x = self.org_forward(x)
if self.multiplier() == 0.0: if self.multiplier() == 0.0:
return x return x
# calculating R here is excruciatingly slow
#R = self.get_weight().to(x.device, dtype=x.dtype) #R = self.get_weight().to(x.device, dtype=x.dtype)
R = self.R.to(x.device, dtype=x.dtype) R = self.R.to(x.device, dtype=x.dtype)
if x.dim() == 4: if x.dim() == 4:
x = x.permute(0, 2, 3, 1) x = x.permute(0, 2, 3, 1)
x = torch.matmul(x, R) x = torch.matmul(x, R)