When saving an entire deep learning network to disk using standard built-in functions, the system serializes only the model's parameters (such as its weights and biases), rather than the model's architecture itself. This separation occurs because neural network models often contain arbitrary control flow and complex code, making native serialization of the architecture difficult. Consequently, to fully reinstate a saved model, a practitioner must first regenerate the exact architecture in code and then load the stored parameters from the file into this newly instantiated model.