Chapter 08

Scenario 3: Multiple data scientists working on multiple ML models

NotebookPython 3.9.12 ('exp-tracking-env')12 cells

Scenario 3: Multiple data scientists working on multiple ML models

MLflow setup:

Tracking server: yes, remote server (EC2).
Backend store: postgresql database.
Artifacts store: s3 bucket.

The experiments can be explored by accessing the remote server.

The example uses AWS to host a remote server. In order to run the example you'll need an AWS account. Follow the steps described in the file mlflow_on_aws.md to create a new AWS account and launch the tracking server.

In [ ]python · cell 2

python

import mlflow
import os

os.environ["AWS_PROFILE"] = "" # fill in with your AWS profile. More info: https://docs.aws.amazon.com/sdk-for-java/latest/developer-guide/setup.html#setup-credentials

TRACKING_SERVER_HOST = "" # fill in with the public DNS of the EC2 instance
mlflow.set_tracking_uri(f"http://{TRACKING_SERVER_HOST}:5000")

In [ ]python · cell 3

python

print(f"tracking URI: '{mlflow.get_tracking_uri()}'")

In [ ]python · cell 4

python

mlflow.search_experiments() # list_experiments API has been removed, you can use search_experiments instead.()

In [ ]python · cell 5

python

from sklearn.linear_model import LogisticRegression
from sklearn.datasets import load_iris
from sklearn.metrics import accuracy_score

mlflow.set_experiment("my-experiment-1")

with mlflow.start_run():

    X, y = load_iris(return_X_y=True)

    params = {"C": 0.1, "random_state": 42}
    mlflow.log_params(params)

    lr = LogisticRegression(**params).fit(X, y)
    y_pred = lr.predict(X)
    mlflow.log_metric("accuracy", accuracy_score(y, y_pred))

    mlflow.sklearn.log_model(lr, artifact_path="models")
    print(f"default artifacts URI: '{mlflow.get_artifact_uri()}'")

In [ ]python · cell 6

python

mlflow.search_experiments()

In [ ]python · cell 7

python

Interacting with the model registry

In [ ]python · cell 9

python

from mlflow.tracking import MlflowClient


client = MlflowClient(f"http://{TRACKING_SERVER_HOST}:5000")

In [ ]python · cell 10

python

client.search_registered_models()

In [ ]python · cell 11

python

run_id = client.search_runs(experiment_ids=['1'])[0].info.run_id
mlflow.register_model(
    model_uri=f"runs:/{run_id}/models",
    name='iris-classifier'
)

In [ ]python · cell 12

python