Design of Zephyr hyperparameters

[sarahmish]

Currently Supported Design

current approach to change hyperparameters is through a dictionary we pass to the zephyr api. This dictionary specifies the primitive and the hyperparameter of that primitive that the user wishes to modify.

An example code for an xgb pipeline

"hyperparameters = {
    "xgboost.XGBClassifier#1": {
        "n_estimators": 50
    }
}

zephyr = Zephyr('xgb', hyperparameters)

Ideal Design

the proposed design is that hyperparameters can be exposed at a pipeline level without the need of a dictionary.

An example code that should achieve the same changes in the previous example

zephyr = Zephyr('xgb', n_estimators=50)

under the hood, these hyperparameters should be mapped to the right primitive and altered.

Challenges

there are several cases that need to be resolved with this strategy:

• if the hyperparameter name belongs to more than one primitive, which one is the user altering?
• these hyperparameters are dynamic in a sense that they change from pipeline to pipeline, how do we support them? how would the user know what hyperparameters they can change?
• can these hyperparameters be modified in only instantiation phase or should we allow other phases like fit? e.g. epochs.