Hello World - Python BentoML

A simple machine learning model with API serving that is written in python and using BentoML. BentoML is an open source framework for high performance ML model serving, which supports all major machine learning frameworks including Keras, Tensorflow, PyTorch, Fast.ai, XGBoost and etc.

This sample will walk you through the steps of creating and deploying a machine learning model using python. It will use BentoML to package a classifier model trained on the Iris dataset. Afterward, it will create a container image and deploy the image to Knative.

Knative deployment guide with BentoML is also available in the BentoML documentation

Before you begin

  • A Kubernetes cluster with Knative installed. Follow the installation instructions if you need to create one.
  • Docker installed and running on your local machine, and a Docker Hub account configured. Docker Hub will be used for a container registry).
  • Python 3.6 or above installed and running on your local machine.

    • Install scikit-learn and bentoml packages:

      pip install scikit-learn
pip install bentoml

Recreating sample code

Run the following code on your local machine, to train a machine learning model and deploy it as API endpoint with KNative Serving.

  1. BentoML creates a model API server, via prediction service abstraction. In iris_classifier.py, it defines a prediction service that requires a scikit-learn model, asks BentoML to figure out the required pip dependencies, also defines an API, which is the entry point for accessing this machine learning service.

from bentoml import env, artifacts, api, BentoService from bentoml.handlers import DataframeHandler from bentoml.artifact import SklearnModelArtifact

@env(auto_pip_dependencies=True) @artifacts([SklearnModelArtifact(‘model’)]) class IrisClassifier(BentoService):

@api(DataframeHandler)
def predict(self, df):
    return self.artifacts.model.predict(df)
  1. In main.py, it uses the classic iris flower data set to train a classification model which can predict the species of an iris flower with given data and then save the model with BentoML to local disk.

from sklearn import svm from sklearn import datasets

from iris_classifier import IrisClassifier

if name == “main": # Load training data iris = datasets.load_iris() X, y = iris.data, iris.target

# Model Training
clf = svm.SVC(gamma='scale')
clf.fit(X, y)

# Create a iris classifier service instance
iris_classifier_service = IrisClassifier()

# Pack the newly trained model artifact
iris_classifier_service.pack('model', clf)

# Save the prediction service to disk for model serving
saved_path = iris_classifier_service.save()








Run the `main.py` file to train and save the model:

```shell
python main.py
```

  1. Use BentoML CLI to check saved model’s information.

    bentoml get IrisClassifier:latest

    Example:

    > bentoml get IrisClassifier:latest
{
  "name": "IrisClassifier",
  "version": "20200305171229_0A1411",
  "uri": {
    "type": "LOCAL",
    "uri": "/Users/bozhaoyu/bentoml/repository/IrisClassifier/20200305171229_0A1411"
  },
  "bentoServiceMetadata": {
    "name": "IrisClassifier",
    "version": "20200305171229_0A1411",
    "createdAt": "2020-03-06T01:12:49.431011Z",
    "env": {
      "condaEnv": "name: bentoml-IrisClassifier\nchannels:\n- defaults\ndependencies:\n- python=3.7.3\n- pip\n",
      "pipDependencies": "bentoml==0.6.2\nscikit-learn",
      "pythonVersion": "3.7.3"
    },
    "artifacts": [
      {
        "name": "model",
        "artifactType": "SklearnModelArtifact"
      }
    ],
    "apis": [
      {
        "name": "predict",
        "handlerType": "DataframeHandler",
        "docs": "BentoService API",
        "handlerConfig": {
          "orient": "records",
          "typ": "frame",
          "input_dtypes": null,
          "output_orient": "records"
        }
      }
    ]
  }
}

  2. Test run API server. BentoML can start an API server from the saved model. Use BentoML CLI command to start an API server locally and test it with the curl command.

    bentoml serve IrisClassifier:latest

    In another terminal window, make curl request with sample data to the API server and get prediction result:

    curl -v -i \
--header "Content-Type: application/json" \
--request POST \
--data '[[5.1, 3.5, 1.4, 0.2]]' \
127.0.0.1:5000/predict

Building and deploying the sample

BentoML supports creating an API server docker image from its saved model directory, where a Dockerfile is automatically generated when saving the model.

  1. To build an API model server docker image, replace {username} with your Docker Hub username and run the following commands.

    # jq might not be installed on your local system, please follow jq install
# instruction at https://stedolan.github.io/jq/download/
saved_path=$(bentoml get IrisClassifier:latest -q | jq -r ".uri.uri")

# Build the container on your local machine
docker build - t {username}/iris-classifier $saved_path

# Push the container to docker registry
docker push {username}/iris-classifier

  2. In service.yaml, replace {username} with your Docker hub username, and then deploy the service to Knative Serving with kubectl:

apiVersion: serving.knative.dev/v1 kind: Service metadata: name: iris-classifier namespace: default spec: template: spec: containers: - image: docker.io/{username}/iris-classifier ports: - containerPort: 5000 # Port to route to livenessProbe: httpGet: path: /healthz initialDelaySeconds: 3 periodSeconds: 5 readinessProbe: httpGet: path: /healthz initialDelaySeconds: 3 periodSeconds: 5 failureThreshold: 3 timeoutSeconds: 60

```shell
kubectl apply --filename service.yaml
```

  1. Now that your service is created, Knative performs the following steps:

    • Create a new immutable revision for this version of the app.
    • Network programming to create a route, ingress, service, and load balance for your application.
    • Automatically scale your pods up and down (including to zero active pods).

  2. Run the following command to find the domain URL for your service:

    kubectl get ksvc iris-classifier --output=custom-columns=NAME:.metadata.name,URL:.status.url

NAME              URL
iris-classifier   http://iris-classifer.default.example.com

  3. Replace the request URL with the URL return in the previous command, and execute the command to get prediction result from the deployed model API endpoint.

    curl -v -i \
  --header "Content-Type: application/json" \
  --request POST \
  --data '[[5.1, 3.5, 1.4, 0.2]]' \
  http://iris-classifier.default.example.com/predict

[0]

Removing the sample app deployment

To remove the application from your cluster, delete the service record:

kubectl delete --filename service.yaml