About training sr

[mrgreen3325]

Hello.
Thanks for your work.
I have some questions about training sr.

1. Is that I need to train both dresnet and han models and provide both trained models to train the sr?
2. I notice that the config_cont.json and config_cont2.json have different train_s_path1, train_s_path2, train_target_path, train_label_path, should I use the same training data for that?
3. There are two sr model g1 and g2 during the training, how do I choose to use during the inference?
   Thanks.

[sangyun884]

Hi,

1. Yes.
2. The code only uses the paths in config_cont.json
3. There is no specific way. You may choose the best performing one or use both.

[mrgreen3325]

Hi,

1. Yes.
2. The code only uses the paths in config_cont.json
3. There is no specific way. You may choose the best performing one or use both.

Thanks for your reply.
May I know can I train the sr model from scratch, i.e. not use the realesrgan pretrain?
Also, is there any psnr comparison between realesrgan and the mssr?
Thanks.

[sangyun884]

May I know can I train the sr model from scratch, i.e. not use the realesrgan pretrain?

Yes.

Also, is there any psnr comparison between realesrgan and the mssr?

No.

[mrgreen3325]

May I know can I train the sr model from scratch, i.e. not use the realesrgan pretrain?

Yes.

Also, is there any psnr comparison between realesrgan and the mssr?

No.

Thanks.
Will the result be degraded if I don't use the realesrgan pretrain model?
For training the degradation models, if I want to use my own camera data, is that the train-source refers to my noisy phone camera images, train-target refers to perfect noiseless images, e.g. ntire2020 target 800 images?

[mrgreen3325]

May I know can I train the sr model from scratch, i.e. not use the realesrgan pretrain?

Yes.

Also, is there any psnr comparison between realesrgan and the mssr?

No.

Hello.
I follow the guide to train.
When I train the rrdb model (use the Realesrgan pretrain) with NTIRE2020 dataset, I found that at about 80000 iteration the val output image is a bit noisy, as shown:
Is that I miss some setting, or it is normal?