Training result not reproducible

[Cheeun]

Thanks for sharing the code. But unfortunately, the training result from setting in run.sh doesn't meet the results of the paper. Training for 8bit EDSR x4 only reaches 31.45 for set 5, which is far behind the reported results.

Are there any specific settings to reproduce the pretrained 4bit_EDSR_x4 or 8bit_EDSR_x4?

[songyonger]

Have you solved this problem?

[Cheeun]

Have you solved this problem?

To get the model of similar performance with the paper, I trained the model with 300 epochs & decay at 150 epochs. (a bit strange since the paper reports only training for 30 epochs). But maybe this will help.

[songyonger]

thanks