G function training for other blur kernel family

[Laviness]

In Readme, the author said “To do image deblurring, data augmentation, and blur generation, you first need to train the blur encoding network (The F function in the paper). This is the only network that you need to train.“ However, after I retrained the F function using out-of-focus blur- sharp image pairs and used it in the image deblurring script "generic_deblur.py" to deblur out-of-focus images, the results are not yet satisfying.

I trained with a batch_size of 13 and iteration of 600,000. The final loss is around 1e03, and the training datasets volume is around 3000 image pairs.

I wonder if I need to retrain the G function as well, if I am using the network to deblur out-of-focus images. If needed, is it possible to provide a script that trains both G and F functions?

Thank you!

[P0lyFish]

Thank you for your interest!
The functions F and G are trained simultaneously. So if you use the provided code to train F, G is trained too. You should test the blur transferring code first to see if the functions F and G are trained correctly.
Also, to do image deblurring, it is required to carefully tune hyper-parameters, and we haven't tried the method with out-of-focus blur yet.

[Jimzhou82sub]

Me too, the train loss is around 1e03, I use the GoPro dataset, but the model can't converge