Skip to content

Latest commit

 

History

History
76 lines (51 loc) · 2.76 KB

README.md

File metadata and controls

76 lines (51 loc) · 2.76 KB

Web service deployment with model from model registry

Setup

  • Run an experiment with RandomForestRegressor
  • Save tracking data in local storage
  • Save artifacts in GCP Bucket
  • Load the model from model registry
  • Deploy the model as web service

Create a GCP Bucket

Create a new Google Cloud Storage Bucket.

gcloud storage buckets create gs://BUCKET_NAME --project=PROJECT_ID --default-storage-class=STANDARD --location=BUCKET_LOCATION --uniform-bucket-level-access

gcloud storage buckets create gs://mlops-zoomcamp-artifacts --project=mlops-demo-408506 --default-storage-class=STANDARD --location=asia-northeast3 --uniform-bucket-level-access

Start MLFlow server

mlflow server --host=0.0.0.0 \
--backend-store-uri=sqlite:///mfllow.db \
--default-artifact-root=gs://pytholic-mlops-zoomcamp-artifacts/

Train the model and save in registry

Follow 04-deployment/web-service-model-registry/random-forest.ipynb. We will make use of sklearn pipeline. Save the model in the registry bucket. We can also save the dict_vectorizer.bin and use it later for preprocessing, byt with sklearn-pipeline this process because embedded and cleaner for us. Tehrefore, we do not need to store it separately.

Run Prediction

Make changes as shown in 04-deployment/web-service-mlflow/predict.py to load model and dict vectorizer from mlflow server. I am loading model using the RUN ID but we can also use model stage for it.

We also need to install mlflow and boto3 in our deployment environment.

pip install mlflow boto3 google-cloud-storage

Test.

python test.py


{'duration': 45.50965007660853, 'model_version': '553def03f5224f649fe56bc1567daccc'}

Summarized steps

  • Run the tracking server i.e. mlflow server --host=0.0.0.0 --backend-store-uri=sqlite:///mfllow.db --default-artifact-root=gs://pytholic-mlops-zoomcamp-artifacts/
  • Run the prediction service i.e. python predict.py
  • Send prediction request i.e. python test.py

Removing dependence on the tracking server

What is the tracking serving is down? We cannot create a new instance of the web-service since we cannot connect to the tracking server. Consequently, we cannot send prediction requests to the service. So, we have become dependent on the tracking server, and this is a problem.

Instead of using MLFLOW_TRACKIING_URI, we can directly point it to load the model from GCS bucket.

# MLFLOW_TRACKING_URI = "http://0.0.0.0:5000"
# mlflow.set_tracking_uri(MLFLOW_TRACKING_URI)

# Load model
EXPERIMENT_ID = 1
RUN_ID = "553def03f5224f649fe56bc1567daccc"
logged_model = f"gs://pytholic-mlops-zoomcamp-artifacts/{EXPERIMENT_ID}/{RUN_ID}/artifacts/model"
# logged_model = f'runs:/{RUN_ID}/model'
model = mlflow.pyfunc.load_model(logged_model)