Hello, Jeffrey, thanks for this insight. Good to know about this work-around.

I’m having an issue with a similar model, but it’s earlier in the development process. I’m looking for tutorials/tips on initially building a model based on USE-lite and SentencePiece. What does the Tensorflow (Keras) code look like for defining the first two layers: the input and the pretrained USE language model? I would add a dense layer after that for classification, which would be trainable. Thanks.