amansrivastava17/embedding-as-service

XLNet encoding issue

harrypb opened this issue · 1 comments

harrypb commented

Please consider the following code.

from embedding_as_service.text.encode import Encoder
en = Encoder(embedding='xlnet', model='xlnet_base_cased', download=True)
vecs = en.encode(texts=["hello, how are you?"])

The above code gives the following error. I'm using python 3.6 (I've run using spyder)

ValueError: Trying to share variable model/transformer/r_w_bias, but specified shape (12, 12, 64) and found shape (24, 16, 64)

By the way, I'm able to generate the embeddings through bert.

Thanks,
Hari

ashutoshsingh0223 commented

The code seems fine.
I wasn't able to reproduce this error though. The setting you are trying to use is a default, where it should share the bias variable. I tried it on CPU, GPU, and TPU runtimes.

Can you share your tensorflow and keras versions?
Also whether you were using CPU or not?