Face Difficulties in Training Models

[wwxn]

Hi, thank you for your great work and we are very interesting in it!

However, we faced some difficulties in reproducing the results reported in the paper using flicker 2W as our training set (https://github.com/liujiaheng/CompressionData). Our training settings are listed below.

Our result tested on Kodak is 0.13bpp/27.18dB, which is about about 2dB lower than the results reported in the paper. Is there some mistake in the training settings?