Onnx models lack linear layer

#6
by jmzzomg - opened

Models in onnx directory produce outputs of shape (x, 384) instead of (x, 96), like with the torch counterpart.
However, vespa_colbert.onnx produce outputs of the correct shape

Is it a bug, or is it supposed to be this way?

Thank you

Answer.AI org

Whoops no this isn't intended behaviour! My guess as to what happened is that @Xenova converted the weights as a normal BERTModel, without wrapping the linear downcasting layer at the end. Can you confirm @Xenova ? Is there any way I can assist you in generating proper ones?

Thanks for flagging @jmzzomg !

Sign up or log in to comment