How to get same result on multiple runs ?

[Dongwoo-Im]

Hi, thanks for your sharing and contribution!

I tried to reproduce same result about training loss on my custom dataset, but it didn't work.

So, I wonder if HAT can return the exactly same result about training loss.

Any help would be much appreciated, thanks.

My environments

• windows 10
• python : 3.7.13
• pytorch : 1.12.1+cu113
• torchvision : 0.13.1+cu113
• cuda : 11.3
• cudnn : 8.4.1
• basicsr : both 1.3.4.9 and 1.4.2 (latest version)

Methods I tried

• use_hflip = False
• use_rot = False
• use_shuffle = False
• num_worker_per_gpu = 0

[Dongwoo-Im]

When I set print_freq to 1, we can check that there's some negligible difference in loss

[chxy95]

I do not understand your question. But you can refer to the original training log of HAT for SRx4 on DF2K.
train_416_train_Final_Model_SRx4_scratch_DF2K_CR3W16P64_500k_B4G8_20220301_002049.log

[Dongwoo-Im]

@chxy95 Thanks for quick response.
When I run same yaml file twice on my custom dataset, loss is not same. Maybe this is caused by some non-deterministic algorhitm in HAT. Anyway I want to reproduce same result on same yaml file, but I do not know how to do it.

[chxy95]

@Dongwoo-Im I'm not sure the reason for this either. The ONLY random factor in HAT appears to be drop_path_rate. It should have no effect on the performance. You can set it to 0 and observe the training process. If it does not work, there may be some random factors that the current version of BasicSR does not take into account.

[Dongwoo-Im]

@chxy95 Thanks your advice. I haven't solved yet, but I'll re-open when I get some insights or find solution.