How to calculate mask-image gradient consistency loss in tensorflow

Question

I want to create custom loss function as described in this article Real-time deep hair matting on mobile devices. I have implemented it in TF2.0 (Google Colab) it worked with no