Hi, I am trying to finetune a github’s (https://github.com/ericsujw/InstColorization/blob/master/README_TRAIN.md) image colorization pretrained model. The problem is that the training loss quickly (during the first epoch) plateaus around 0.17-0.19 depending on the dataset. There is not much colorization that happens, even inputting the images, during the training, into the model outputs poor results (not much changes in the LAB color space).
I was able to finetune a model over 50 epochs, but it was only able to colorize sky (blue) and grass (green). I am not too sure why my model does not learn and any input is much appreciated. I have introduced a LeakyRelu and tried increasing and decreasing the learning rate.