Finetuning stopping criteria

[aliab3d]

Hi guys, amazing repository. Ive been playing around to finetune the models on my own datasets, and Ive got a few questions I was hoping you could help me with:

1. when finetuning, what is your recommendation for stopping the training process? I've noticed that fine-tuning for longer periods does not necessarily result in better SR outputs, but also deteriorates them by introducing artefacts or losing color information. should we manually inspect the super-resolved images at different iterations to decide on when to stop?

2. following the previous question, are there any visualizations integrated in the codebase during training? I can see a experiments/visualization directory being created but no outputs are being stored. can you comment on how to utilize this option?

3. in the finetuning recipe and based on options/finetune_realesrgan_x4plus.yml, the generator network is initialized with RealESRNet but the discriminator is initialized with RealESRGAN weights. Is there a reason why we wouldnt want to initialize the Generator with RealESRGAN generator when finetuning?

would love to hear your thoughts,

cheers

[dersugiritlioglu]

Hi @aliab3d , any findings about these?