stable-diffusion-webui/extensions-builtin/Lora/network_oft.py

82 lines
2.9 KiB
Python
Raw Normal View History

2023-10-18 06:35:50 +00:00
import torch
import network
class ModuleTypeOFT(network.ModuleType):
def create_module(self, net: network.Network, weights: network.NetworkWeights):
if all(x in weights.w for x in ["oft_blocks"]):
return NetworkModuleOFT(net, weights)
return None
# adapted from https://github.com/kohya-ss/sd-scripts/blob/main/networks/oft.py
class NetworkModuleOFT(network.NetworkModule):
def __init__(self, net: network.Network, weights: network.NetworkWeights):
2023-10-18 11:16:01 +00:00
2023-10-18 06:35:50 +00:00
super().__init__(net, weights)
self.oft_blocks = weights.w["oft_blocks"]
self.alpha = weights.w["alpha"]
self.dim = self.oft_blocks.shape[0]
self.num_blocks = self.dim
if "Linear" in self.sd_module.__class__.__name__:
self.out_dim = self.sd_module.out_features
elif "Conv" in self.sd_module.__class__.__name__:
self.out_dim = self.sd_module.out_channels
2023-10-18 11:16:01 +00:00
self.constraint = self.alpha
#self.constraint = self.alpha * self.out_dim
2023-10-18 06:35:50 +00:00
self.block_size = self.out_dim // self.num_blocks
2023-10-18 11:16:01 +00:00
self.org_module: list[torch.Module] = [self.sd_module]
self.R = self.get_weight()
2023-10-18 06:35:50 +00:00
2023-10-18 11:16:01 +00:00
self.apply_to()
# replace forward method of original linear rather than replacing the module
def apply_to(self):
self.org_forward = self.org_module[0].forward
self.org_module[0].forward = self.forward
2023-10-18 06:35:50 +00:00
2023-10-18 11:16:01 +00:00
def get_weight(self, multiplier=None):
if not multiplier:
multiplier = self.multiplier()
2023-10-18 06:35:50 +00:00
block_Q = self.oft_blocks - self.oft_blocks.transpose(1, 2)
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))
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())
2023-10-18 11:16:01 +00:00
block_R_weighted = multiplier * block_R + (1 - multiplier) * I
2023-10-18 06:35:50 +00:00
R = torch.block_diag(*block_R_weighted)
return R
def calc_updown(self, orig_weight):
2023-10-18 11:16:01 +00:00
R = self.R
if orig_weight.dim() == 4:
weight = torch.einsum("oihw, op -> pihw", orig_weight, R)
else:
weight = torch.einsum("oi, op -> pi", orig_weight, R)
updown = orig_weight @ R
output_shape = [orig_weight.size(0), R.size(1)]
#output_shape = [R.size(0), orig_weight.size(1)]
2023-10-18 06:35:50 +00:00
return self.finalize_updown(updown, orig_weight, output_shape)
2023-10-18 11:16:01 +00:00
def forward(self, x, y=None):
x = self.org_forward(x)
if self.multiplier() == 0.0:
return x
R = self.get_weight().to(x.device, dtype=x.dtype)
if x.dim() == 4:
x = x.permute(0, 2, 3, 1)
x = torch.matmul(x, R)
x = x.permute(0, 3, 1, 2)
else:
x = torch.matmul(x, R)
return x