This study presents a unified end-to-end framework for network equilibrium analysis framework. The end-to-end framework directly learns model supply- and demand-side components and equilibrium states from multi-day traffic state observations. It parametrizes unknown model components with computational graphs and embeds them in a variational inequality to enforce user equilibrium conditions. Each component can be model-based, model-free (i.e., neural network), or hybrid. By minimizing the differences between the estimated and observed traffic states, the framework simultaneously estimates the unknown parameters for supply- and demand-sides. Our study addresses key challenges in modeling and calibrating the unified end-to-end framework. We identify a novel neural network architecture that guarantees the existence of equilibrium traffic states and accommodates the potential changes in the road network topology for future what-if planning analysis. To train the model effectively, we leverage the computational power of computational graphs and design auto-differentiation-based gradient descent algorithms to handle both link- or path-based user equilibrium constraints. In forward propagation, we adopt recent developments in operator-splitting methods and differential optimization to solve a batch of VI problems. In backpropagation, iterated differentiation and implicit differentiation techniques are used to efficiently differentiate through the equilibrium states. The proposed framework and findings are validated using three synthesized datasets.