Hi everyone, what are the architecture changes on the pix2pix (Unet) one can do when the input image ( a cloud of 2 to 4 white spots) is very different from the target image ( the 2D phase profile created by Fourier transforms) ?
Hi Xavier! Welcome to the community. Hope you are doing well!
When the input image is very different from the target image, it can be very challenging to train a successful pix2pix model using the standard architecture. But what kind of architecture? Well, that cannot be generalized in my opinion. It depends on your experiment. You can start with the standard architecture (the one mentioned in the course) and then can trying making changes according to the performance like increasing the depth of encoders and decoders, increasing the complexity of the generator, discriminator etc.
So there is no fixed architecture for some problems, one has to try them out and see. I hope that it makes some sense. If not, feel free to share your opinion.
Have a great day!
Regards,
Nithin