Masking immutable matricies/tensors to perform same functionality as numpy where assignment?

Question

I am working on creating a custom tensorflow loss function. This loss function is used to train a neural network on my GPU. The loss function is provided with two rank 4 tensors