How to avoid dead ReLU and NaN output values?

Question

I recently got an assignment that requires me to code a dense neural network from scratch in Python. We are supposed to solve some regression problems while using Sigmoid, T