stable-diffusion-webui
stable-diffusion-webui copied to clipboard
AUTOMATIC1111
Add missing support for linear activation in hypernetwork
In hypernetwork, the linear activation_func that the old implementation used is missing from ui.
Add the linear activation in hypernetwork back to ui.
Just a heads-up: unless I am mistaken, from my testing of this branch, it seems the changes that remove activation_func=None (38, 52) will cause prior trained hypernets to be unable to load.
RuntimeError: hypernetwork uses an unsupported activation function: None
Just a heads-up: unless I am mistaken, from my testing of this branch, it seems the changes that remove
activation_func=None(38, 52) will cause prior trained hypernets to be unable to load.
RuntimeError: hypernetwork uses an unsupported activation function: None
Sorry for that... I only test the new hypernetwork created with 'linear' activation before last update. Now the old hypernetworks should also work.
Old hypernetworks still do not work
raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format(
RuntimeError: Error(s) in loading state_dict for HypernetworkModule:
Missing key(s) in state_dict: "linear.2.weight", "linear.2.bias".
Unexpected key(s) in state_dict: "linear.1.weight", "linear.1.bias".
And hypernetwork create by this commit can not work on latest master commit:737eb28faca8be2bb996ee0930ec77d1f7ebd939
raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format(
RuntimeError: Error(s) in loading state_dict for HypernetworkModule:
Missing key(s) in state_dict: "linear.1.weight", "linear.1.bias", "linear.3.weight", "linear.3.bias".
Unexpected key(s) in state_dict: "linear.4.weight", "linear.4.bias", "linear.6.weight", "linear.6.bias".
size mismatch for linear.2.weight: copying a param with shape torch.Size([3072, 1536]) from checkpoint, the shape in current model is torch.Size([1536, 3072]).
size mismatch for linear.2.bias: copying a param with shape torch.Size([3072]) from checkpoint, the shape in current model is torch.Size([1536])
the keys of state_dict should be like: linear.0.weight linear.0.bias linear.2.weight linear.2.bias linear.4.weight linear.4.bias linear.5.weight linear.5.bias
but in this commit keys of state_dict are linear.0.weight linear.0.bias linear.2.weight linear.2.bias linear.4.weight linear.4.bias linear.6.weight linear.6.bias
Old hypernetworks still do not work
raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format( RuntimeError: Error(s) in loading state_dict for HypernetworkModule: Missing key(s) in state_dict: "linear.2.weight", "linear.2.bias". Unexpected key(s) in state_dict: "linear.1.weight", "linear.1.bias".And hypernetwork create by this commit can not work on latest master commit:737eb28faca8be2bb996ee0930ec77d1f7ebd939
raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format( RuntimeError: Error(s) in loading state_dict for HypernetworkModule: Missing key(s) in state_dict: "linear.1.weight", "linear.1.bias", "linear.3.weight", "linear.3.bias". Unexpected key(s) in state_dict: "linear.4.weight", "linear.4.bias", "linear.6.weight", "linear.6.bias". size mismatch for linear.2.weight: copying a param with shape torch.Size([3072, 1536]) from checkpoint, the shape in current model is torch.Size([1536, 3072]). size mismatch for linear.2.bias: copying a param with shape torch.Size([3072]) from checkpoint, the shape in current model is torch.Size([1536])
Now the default behaviour should be the same as commit:737eb28faca8be2bb996ee0930ec77d1f7ebd939. Let's check if this work on your old hypernetworks.
Now it is working on my old hypernetworks , but can not work with hypernetwork created by3302fbd. But It's all right, I will retrain my hypernetwork
Okay I checked the code, apparently "linear": torch.nn.Identity, would add additional Linear layer instead of leaving it empty, which will break old HNs. If Linear was created with that function, that might be problem... so I'll somehow find a way to make both of them work.
In https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/3771 https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/3771/commits/f361e804ebaa5af4a10711ece2522869fb64a4c6 fixes it.
@nekoyama32767 @benkyoujouzu
if keyError is happening : it's because HN were created without skipping linear, somehow without this line
if activation_func == "linear" or activation_func is None: pass
Its not normal in main branch, so forcing Identity won't be supported.