Hi, I am currently doing a research on semantic segmentation using adversarial loss or GAN.
I have a pre-trained segmentation (generator) and discriminator and I would like to use the pre-trained
discriminator to fine tune my generator. However, as it iterates, the loss or accuracy only changes very little, so I tried changing hyperparamter to extreme, but they are still the same.
For example, I once fine tuned my generator with learning rate 0.005, and the accuracy reached 86.3 at epoch 40 and I tried changing learning rate to extreme like 0.8 and the accuracy reached 86.3 at epoch 40 as well which doesn’t make any sense to me.
I’ve checked that requires_grad is True for all my generator and discriminator.
Can somebody help me with this problem?
Thank you!