Skip to main content

Example PyCaret deployment

In this example we will use PyCaret to predict whether someone has diabetes, using the PyCaret sample dataset diabetes. PyCaret has some unique behavior when it comes to creating deployment-ready models, which we'll focus on in this example.

First, in your notebook, import and log in to Modelbit:

import modelbit
mb = modelbit.login()

Now we'll ask PyCaret to make a model using the diabetes sample dataset:

from pycaret.regression import *
from pycaret.datasets import get_data

diabetes = get_data("diabetes")
s = setup(diabetes, target = 'Class variable')
model = create_model('et')

With our model created, we need to save and re-load it so it can be pickled for deployment. PyCaret models are not pickle-able right after they've been trained, and the save/load process forces the model into a pickle-able state.

save_model(model, 'pycaret_example_model')
model = load_model("pycaret_example_model", verbose=False)

With our model re-loaded, we'll make a deployment function with a unit test. Notice that getting the output of the PyCaret model requires calling predict_model, and then fetching the value from the Label column in the returned dataframe:

import pandas as pd

def predict_with_pycaret(input_dict):
Example pycaret predictor

>>> predict_with_pycaret({ "Number of times pregnant": 6, "Plasma glucose concentration a 2 hours in an oral glucose tolerance test": 148, "Diastolic blood pressure (mm Hg)": 72, "Triceps skin fold thickness (mm)": 35, "2-Hour serum insulin (mu U/ml)": 0, "Body mass index (weight in kg/(height in m)^2)": 33.6, "Diabetes pedigree function": 0.627, "Age (years)": 50 })
{ "has_diabetes": 1}
df = pd.DataFrame.from_records([input_dict])
return { "has_diabetes": predict_model(model, data=df)[['Label']].iloc[0][0] }

We can now deploy our PyCaret deployment. Include numpy's version in the python_packages as well, as it's an implicit dependency of PyCaret.

mb.deploy(predict_with_pycaret, python_packages=["pycaret==2.3.10", "numpy==1.20.0"])

Modelbit will package the PyCaret model and predict_with_pycaret function into a deployment that can then be called from REST, Snowflake, and Redshift.