Chi Tran

perm is unnecessary in training step

add `self.state.save_to_json(os.path.join(output_dir, TRAINER_STATE_NAME))` in llava/train/llava_trainer.py in _save_checkpoint method

  Here is the result of JointBERT(86%) and JointIDSF(81%)script, I followed the original script in repo without modification https://colab.research.google.com/drive/1Ua7otJZR-cDpjvvFa0mVLVkbMJoPnCmv#scrollTo=L89Kex6VFwQ-