How to reproduce the metrics in your paper?

[tiankuan93]

Questions:
I tried different config settings to train the image codec model, and at the lowest quality, I got similar results to the paper. But when using the No.4 model setting, I could not get similar results to the paper. When validating MS-SSIM in the CLIC validation set, the trained model differed from the metrics reported in the paper by about 5-6 dB, which is a very serious difference. The result I reproduced is approximately the red symbol in the figure.

I also tried training with an earlier version of cac2921, with the model configuration shown below, and this result still differs from the paper's results by about 5 dB.

Request:
I would like to reproduce the metrics in the paper, is it convenient to provide more training details of the model in the paper?