Thanks for your work of the code! I used the [CheckerboardAutogressivev2.py] based on the Minnen2018 in the version2 folder to train a model, but the evaluation results on Kodak dataset is worse than the results given in CompressAI. The PSNR is about 4dB lower than CompressAI‘s while at the same bit rate. Have you evaluate your model on Kodak? Thank you!