Questions about training

[wangqiuwen1006]

Hello! I am glad that you are willing to share your work. When reproducing the final model of the paper, I followed the instructions in the README training section and used the Re10k-subset. Without changing the config folder, I found that the final trained model achieved a PSNR of only 21.653383681648656 during testing. The difference between your setup and mine is that I only have one Nvidia RTX 4090. Therefore, I adjusted the data_loader.train.batch_size to 4 to prevent CUDA OUT OF MEMORY errors. Additionally, I noticed that the results obtained at 300k steps are similar to those at 120k steps. Is it reasonable to achieve this result? If not, what could be the possible reasons for this discrepancy, and what should I modify? I'm looking forward to hearing back from you.