Training All Mask R-CNN layers - Val Loss not improving

I am trying to train a Mask R-CNN to detect power lines but the validation loss keeps fluctuating. I am using a dataset size of ~1300. Any help?