tweak linear attention

Author: lucidrains

denoising_diffusion_pytorch/classifier_free_guidance.py

@@ -214,7 +214,6 @@ def forward(self, x):
         k = k.softmax(dim = -1)
 
         q = q * self.scale
-        v = v / (h * w)
 
         context = torch.einsum('b h d n, b h e n -> b h d e', k, v)
 

denoising_diffusion_pytorch/denoising_diffusion_pytorch.py

@@ -218,7 +218,6 @@ def forward(self, x):
         k = k.softmax(dim = -1)
 
         q = q * self.scale
-        v = v / (h * w)
 
         context = torch.einsum('b h d n, b h e n -> b h d e', k, v)
 

denoising_diffusion_pytorch/guided_diffusion.py

@@ -212,7 +212,6 @@ def forward(self, x):
         k = k.softmax(dim = -1)
 
         q = q * self.scale
-        v = v / (h * w)
 
         context = torch.einsum('b h d n, b h e n -> b h d e', k, v)
 

denoising_diffusion_pytorch/simple_diffusion.py

@@ -187,7 +187,6 @@ def forward(self, x):
         k = k.softmax(dim = -1)
 
         q = q * self.scale
-        v = v / (h * w)
 
         context = torch.einsum('b h d n, b h e n -> b h d e', k, v)
 

denoising_diffusion_pytorch/version.py

@@ -1 +1 @@
-__version__ = '1.7.3'
+__version__ = '1.7.4'