From 1c6efdbba774d603c592debaccd6f5ad827bd1b2 Mon Sep 17 00:00:00 2001
From: v0xie <28695009+v0xie@users.noreply.github.com>
Date: Wed, 18 Oct 2023 04:16:01 -0700
Subject: inference working but SLOW

---
 extensions-builtin/Lora/network_oft.py | 73 +++++++++++++++++-----------------
 1 file changed, 36 insertions(+), 37 deletions(-)

(limited to 'extensions-builtin/Lora/network_oft.py')

diff --git a/extensions-builtin/Lora/network_oft.py b/extensions-builtin/Lora/network_oft.py
index 9ddb175c..f085eca5 100644
--- a/extensions-builtin/Lora/network_oft.py
+++ b/extensions-builtin/Lora/network_oft.py
@@ -12,6 +12,7 @@ class ModuleTypeOFT(network.ModuleType):
 # 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):
+
         super().__init__(net, weights)
 
         self.oft_blocks = weights.w["oft_blocks"]
@@ -20,24 +21,29 @@ class NetworkModuleOFT(network.NetworkModule):
         self.dim = self.oft_blocks.shape[0]
         self.num_blocks = self.dim
 
-        #if type(self.alpha) == torch.Tensor:
-        #    self.alpha = self.alpha.detach().numpy()
-
         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
 
-        self.constraint = self.alpha * self.out_dim
+        self.constraint = self.alpha
+        #self.constraint = self.alpha * self.out_dim
         self.block_size = self.out_dim // self.num_blocks
 
-        self.oft_multiplier = self.multiplier()
+        self.org_module: list[torch.Module] = [self.sd_module]
+
+        self.R = self.get_weight()
 
-        # replace forward method of original linear rather than replacing the module
-        # self.org_forward = self.sd_module.forward
-        # self.sd_module.forward = self.forward
+        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
     
-    def get_weight(self):
+    def get_weight(self, multiplier=None):
+        if not multiplier:
+            multiplier = self.multiplier()
         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)
@@ -45,38 +51,31 @@ class NetworkModuleOFT(network.NetworkModule):
         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_weighted = self.oft_multiplier * block_R + (1 - self.oft_multiplier) * I
+        block_R_weighted = multiplier * block_R + (1 - multiplier) * I
         R = torch.block_diag(*block_R_weighted)
 
         return R
 
     def calc_updown(self, orig_weight):
-        oft_blocks = self.oft_blocks.to(orig_weight.device, dtype=orig_weight.dtype)
-        block_Q = oft_blocks - 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=oft_blocks.device).unsqueeze(0).repeat(self.num_blocks, 1, 1)
-        block_R = torch.matmul(I + block_Q, (I - block_Q).inverse())
-
-        block_R_weighted = self.oft_multiplier * block_R + (1 - self.oft_multiplier) * I
-        R = torch.block_diag(*block_R_weighted)
-        #R = self.get_weight().to(orig_weight.device, dtype=orig_weight.dtype)
-        # W = R*W_0
-        updown = orig_weight + R
-        output_shape = [R.size(0), orig_weight.size(1)]
+        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)]
         return self.finalize_updown(updown, orig_weight, output_shape)
     
-    # def forward(self, x, y=None):
-    #     x = self.org_forward(x)
-    #     if self.oft_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
+    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
-- 
cgit v1.2.3