I saved a BERT model previously fine tuned on Google Cloud TPU (using tensorflow 1.14) to run it locally on CPU. I used this method:
def serving_input_fn():
