Custom sized images for training

Hi

I would like to train on image size 1920x1080. I have followed How can I use DetectNet for custom size data ? · Issue #980 · NVIDIA/DIGITS · GitHub
There are people succeeding in training with this image size but they had to change the size from 1920x1080 to 1920x1088 because 1080 divided by 16 is not an integer value.

In my case the training wont even start because I got the error of shape error: cannot force shape(84,5) into shape(50,5).

Has somebody succeeded in training with high def images?