Training from scratch (200k iters + 100k iters) doesn't achieve the reported performance (PSNR 35.46) with GoPro

[ohjinjin]

Hello, thank you for your good research first of all.

I was trying to reproduce the performance reported in your paper with SCER-GoPro dataset that you shared as a link.
(Before I started training from scratch, I had checked that your pre-trained weights gave me PSNR 35.44.
I thought this difference was not that big.)

Since I trained by myself with SCER-GoPro dataset and the code implementation here, total number of iterations was set as 200k which was not equal to the explanation in your paper. (The paper said total num of iter was 300k). Hence I thought this inconsistency was from that difference.
Thus I proceeded to train an additional 100k iters again, however, the performance became lower than that of 200k iter.

Is it because of issues from loading the resume training or should I have to modify some part of the experimental setting in code implementation to obtain the same performance?
Result performance is like below:

psnr 35.46 ssim 0.972 at 200K iter
psnr: 34.4930 ssim: 0.9662 at 200K + 100K iter

It would be really appreciated if you answer my question!

[ohjinjin]

I figured out that i made a mistake about setting batchsize and the number of gpus. Thanks!

[shkids]

hi, @ohjinjin
How did you set up yml? If you're trying to train, you'll need to run the test after 258 iterations, depending on your setup. However, unlike the settings, there is a problem where the test starts as soon as iter rotates 8 times.