How to convert all classical neural networks into QNN

The picture shows a simple 4-layer classical neural network. How to transform it into a corresponding QNN? And design the corresponding quantum circuit for training, adding the input data set element size is within 100？

First, I use quantum encoding technology to convert angular encoding into a phase representation, then how should I choose ANSATZ and design-related parametric U gates? In the last measurement, what I read was actually the parameterized angle value correct? Then iterate on the parameterized angle value to get the best angle. But f(x,theta), does theta optimal represent the optimal output result x?

The title is slightly misleading, as you can't convert classical NN to run on a quantum computer, at least not on NISQ. It's very hard to run arithmetic on a quantum computer.
What one can do is to design QNN that are similar to classical NN in function. A QNN is typically a variational circuit with angles that determine the rotations of various operations, see medium.com/xanaduai/training-quantum-neural-networks-with-pennylane-pytorch-and-tensorflow-c669108118cc. You can't read the angle values from measurements, those are inputs to the QNN and are being optimized, just like you have trainable parameters for classical NN, like parameters in linear layer, and they are not your NN outputs.

What is the measurement output of VQE? Can the minimum Hamiltonian be found by optimizing the angle? Which eigenvector's smallest eigenvalue is VQE looking for?