Self-implemented training code could not reproduce the results in the paper

I tried to create the training code myself, but the results don't match the paper. Perhaps I missed something, but I'm not sure. I followed the exact hyperparameters and procedures mentioned in the paper. I would greatly appreciate it if you could provide the training code. Thank you.