search

Nerogar / OneTrainer

OneTrainer is a one-stop solution for all your stable diffusion training needs.
GNU Affero General Public License v3.0
1.78k stars 148 forks source link

[Bug]: Trained embeddings have different token count than my specified value #475

Closed kurophali closed 1 month ago

kurophali commented 1 month ago

What happened?

Not sure if it's a bug or not. But I finetuned a model with some additional embeddings in shape (42, 768) for the first encoder. Then in the trainer's ui I specified 1 as the token count. The saved embeddings are in (42, 768) instead of (1, 768). Is it by designed?

What did you expect would happen?

Saved embeddings should be in (1, 768)

Relevant log output

No response

Output of pip freeze

No response

Nerogar commented 1 month ago

I finetuned a model with some additional embeddings in shape (42, 768) for the first encoder

I'm not quite sure what you mean by this. Did you load an existing embedding with token count 42 into an additional embedding with count 1? In that case... just don't do that.

kurophali commented 1 month ago

I finetuned a model with some additional embeddings in shape (42, 768) for the first encoder

I'm not quite sure what you mean by this. Did you load an existing embedding with token count 42 into an additional embedding with count 1? In that case... just don't do that.

Yes that's what I tried. I'm now reshaping my own files to match the train ui token count.