Using google bert for RoBERTa first, using convert_pytorch_checkpoint_to_tf.py convert wieght second, edit config file third, modified modelling:using_token_type=False start=2 Fourth, change bert tokenizer to roberta tokenizer