I'm trying to understand OutputModel config_text and config_dict. I'd like to include the python code that can be copypasted to load the model definition that can accept the weights stored in the registered model. Is this a common practice? Is there a better way to link the model weights and the model architecture? I seem to be able to get it to work, but when I rerun my training loop and resave my model weights, the initialized config_dict information gets lost and replaced with the string of the filename where the weights are stored. Am I misusing this somehow?