error: weight's shape is different, _IncompatibleKeys about a1111-sd-webui-locon

[KohakuBlueleaf]

@azumukupoe You should use more small dim (
if you resolution is different from the resolution that base model trained on
conv will learn more about "how to denoise"
and sometime it will make your training process broken

[KohakuBlueleaf]

You should update the extension

[azumukupoe]

it should be up to date according to ORIG_HEAD

9abae9e

[KohakuBlueleaf]

Are you using some other extension for lora?

[azumukupoe]

I'm using https://github.com/kohya-ss/sd-webui-additional-networks and that's it

[KohakuBlueleaf]

LoHa and LoCon with cp decomposition doesn't support addnet now

[azumukupoe]

oh I see but the outputs don't look right
it wasn't like this before the cp decomposition update

[KohakuBlueleaf]

your training is broken
maybe due to the change of weight init value
try to tune the training parameter

[azumukupoe]

I didn't see any change in args so I used the same setting to train
--network_args "conv_dim=64" "conv_alpha=1" "algo=lora"
or is this something I should tweak in the kohya's script?
sorry I'm still figuring out training