It seems that encoder model is not updated and saved on checkpoint when training.
It seems that encoder model is not updated and saved on checkpoint when training.