Significant performance gap between re-trained model and official checkpoints on BIWI dataset

[pepper725]

I followed the instructions in the README.md to re-train the Stage 2 model and evaluate it on the BIWI dataset. However, I observed a significant performance gap between my re-trained model and the provided official checkpoints.

1.The first image shows the result from my re-trained model.

2.The second image shows the result using the official checkpoints.