Parameters and meta data

[nudles]

During training, the model may generate some meta data;
The training dataset may have some meta data, e.g., the mapping from the index to the label name;
These meta data is used during inference.
How to pass them to the predict function?

[NLGithubWP]

Each model have two methods called dump_parameters and load_parameters, which defined by user and store whatever user want. So after training, the dump_parameters will be called and the parameter will be stored under file "params". When we build an inference job 2 services will start, one is predictor, another one is called inference service. Predictor is a http server, getting images from user, and produce those images to kafka, inference service serve as consumer, which will load the model class from DB(postgresql) and then call the load_parameters method to load parameters we stored in "params".

[nudles]

1. Are the parameters dumped into a database or a file/folder? can the the meta data be dumped together?
2. We may consider to use a pipeline like Apache Airlow to replace the current implementation of the predictor, which is a bit ad-hoc. Pipeline is more general. We can put the models into a pipeline in different way.

[NLGithubWP]

1, The parameters dumped into file name as aaa.model for example, we can dump all meta-data together, no matter what we dump, we must load them in load_parameters methods with the same data structure.
2. ok