search

NJU-Jet / SR_Mobile_Quantization

Winner solution of mobile AI (CVPRW 2021).
148 stars 41 forks source link

a bug that during training the x2 model, 0811.pt is not generated? #4

Open xiaoxiongli opened 3 years ago

xiaoxiongli commented 3 years ago

Hi Dear @NJU-Jet : it seems have a bug, please see below figure: image

image

and I check the folder data/DIV2K/bin/DIV2K_train_HR/, below figure:

image

you can see that the it have 0001.pt ~ 0810.pt, and no 0811.pt ... , so the train.py crashed...

How to fix this? thank you~

NJU-Jet commented 3 years ago

If you want to only use ten images for fast validation, you can modify DIV2K_val.txt file. Retain the first ten lines and remove other lines.

xiaoxiongli commented 3 years ago

OK, I got it, thank you and below warning is normal or not? image

NJU-Jet commented 3 years ago

It's normal. Sometimes PSNR doesn't increase and you should retrain the model.

xiaoxiongli commented 3 years ago

And For 0811.pt is not generated, I find the reason is the README.md's data tree is not same as the base7.yaml, the base7.yaml have another "bin" while README.md not have. I change the base7.yaml, it works.

xiaoxiongli commented 3 years ago

It's normal. Sometimes PSNR doesn't increase and you should retrain the model.

you mean that just retrain it and do not change any super-parameters such as lr or batch-size? just re-run the same command "python train.py --opt options/train/base7.yaml --name base7_D4C28_bs16ps64_lr1e-2 --scale 2 --bs 16 --ps 64 --lr 1e-3 --gpu_ids 0"?

NJU-Jet commented 3 years ago

Yes. In my experiment, simple retrain will solve this problem.