Text model weight names cause error in loading.

#1
by ostris - opened

First, thank you for doing this! Super cool and I am experimenting with it. I wanted to bring to your attention that the text model weight keys are

text_model.text_model.embeddings.position_embedding.weight

vs

text_model.embeddings.position_embedding.weight

There is an extra text_model on all the weights that prevents it from loading with transformers.

Thank you for pointing that out! I goofed on the checkpoint conversion. Should be fixed now. Though no guarantee the model works correctly; I'll have to get back to this project in the future and double check everything.

Another issue I ran into is transformers complaining about max_length not being specified, even though it's in the tokenizer's config. So I had to run with: inputs = processor(text=texts, images=image, padding="max_length", return_tensors="pt", truncation=True, max_length=256)

fancyfeast changed discussion status to closed

Sign up or log in to comment