search

aria1th / Hypernetwork-MonkeyPatch-Extension

Extension that patches Hypernetwork structures and training
116 stars 10 forks source link

Training embeddings results in unrecognizable characters #39

Closed siotss96 closed 1 year ago

siotss96 commented 1 year ago

I'm not sure if it was Automatic111 updates that caused the issue or something else but every time I try to create a new embedding with train_gamma, the preview images it reproduces during any Steps is always different from the subject.

Example: Previously I was able to make a Lucy (edgerunners) embedding pretty easily and even during step 100 the preview image had her likeness down but now even at step 500 I would just get a random anime girl without Lucy's characteristics.

anyone else have this issue?

aria1th commented 1 year ago

Unfortunately I cannot find direct change to training code - so I cannot directly compare between patches. But I suspect its fault of default adamW weight decay parameter, you can set it to 0 instead of 0.01. Now it shares parameter with hypernetwork, so decay parameter's default value has changed from 0.

siotss96 commented 1 year ago

doesn't seem to be that, do you know any other settings that may have been touched? I just tried again with a Lucy embedding still different. image

siotss96 commented 1 year ago

the torch and cuda versions was the issue. Needed to downgrade them since the newer versions couldn't train properly.

Apologies for the trouble!