[NotForMerging]Add mlflow loader for glassbox models

[eedeleon]

• Add loader functions for mlflow log_model and load_model support.

Allows for operationalizing EBM along with all glassbox models that support explain_global.

Example:

import json
import os

import pytest

from sklearn.datasets import ,load_boston

from interpret.glassbox import ExplainableBoostingRegressor
from interpret.glassbox.mlflow import log_model


boston = load_boston()
glassbox_model_ebm = ExplainableBoostingRegressor(feature_names=boston.feature_names)

X, y = boston.data, boston.target

glassbox_model_ebm.fit(X, y)

log_model("ebm", glassbox_model_ebm)

import mlflow
run_id = mlflow.active_run().info.run_id

# Starts up the server with {"columns":columns, "data":data} as inputs with returned predict(data)
print("mlflow models serve -m runs:/{}/{} &".format(run_id, "ebm"))
pred_data = [[x.item() for x in row] for row in X[:2]]

# Wait until the service is running before executing the curl command
print("curl http://127.0.0.1:5000/invocations -H 'Content-Type: application/json' -d '{}'".format({"data": str(pred_data).replace("'", '"').replace(" ", "")}))
# [26.34637146659037, 22.031728888383835]

# The two printed commands can be copied into commandline to test the service.

Code follows a similar paradigm to: https://www.mlflow.org/docs/latest/models.html#example-creating-a-custom-add-n-model without enforcing interpret take a dependency on mlflow.

mlflow is a soft dependency to allow for an enhanced mlflow experience without affecting existing user experiences

[paulbkoch]

Hi @eedeleon -- I see this PR has been labeled "NotForMerging" for a while now. What's the current status?