How about saving the model parameters for the backbone part of the training process as a .pkl file, like you provided in the example