(Advanced) Model Serving by Seldon

In this section, we want to post our MNIST classifier online, thus allowing everyone to try our model via a URL.

To do this, you must have:

  • Full control of your Kubernetes cluster

  • Full control of your PrimeHub account

  • A Google Cloud account

  1. Install Seldon (https://docs.seldon.io/projects/seldon-core/en/latest/).

     helm install seldon-core-operator --name seldon-core --repo https://storage.googleapis.com/seldon-charts --set usageMetrics.enabled=true --namespace seldon-system --version 0.5.1

  2. Create a new python file train_mnist_serving.py in JupyterLab.

     import tensorflow as tf
 import argparse
 import os
 import shutil
 
 parser = argparse.ArgumentParser(description='Process some integers.')
 parser.add_argument('--dropout', type=float, default=0.2)
 args = parser.parse_args()
 
 mnist = tf.keras.datasets.mnist
 
 (x_train, y_train),(x_test, y_test) = mnist.load_data()
 x_train, x_test = x_train / 255.0, x_test / 255.0
 
 model = tf.keras.models.Sequential([
   tf.keras.layers.Flatten(input_shape=(28, 28)),
   tf.keras.layers.Dense(512, activation=tf.nn.relu),
   tf.keras.layers.Dropout(args.dropout),
   tf.keras.layers.Dense(10, activation=tf.nn.softmax)
 ])
 
 model.compile(optimizer='adam',
               loss='sparse_categorical_crossentropy',
               metrics=['accuracy'])
 
 model.fit(x_train, y_train, epochs=5)
 model.evaluate(x_test, y_test)
 
 export_path = "1"
 if os.path.isdir(export_path):
     print('Cleaning up\n')
     shutil.rmtree(export_path)
 
 tf.saved_model.simple_save(
     tf.keras.backend.get_session(),
     export_path,
     inputs={'input_image': model.input},
     outputs={t.name:t for t in model.outputs})

  3. Run this file in job submission with command.

     cd /project/<group name>/
 python -u train_mnist_serving.py

  4. Upload the outputted 1 folder into your Google Bucket in <your_bucket_name>/mnist

  5. Create a service account and a JSON key.

  6. Download the key in Google Cloud.

  7. Create a secret in your cluster. <LOCALFILE> is the key file you just downloaded.

     kubectl create secret generic user-gcp-sa --from-file=gcloud-application-credentials.json=<LOCALFILE>

  8. Create a service account in your cluster.

     apiVersion: v1
 kind: ServiceAccount
 metadata:
   name: user-gcp-sa
 secrets:
   - name: user-gcp-sa

  9. Create a Seldon deployment in your cluster with your bucket. Replace gs://<your_bucket_name>/mnist.

     apiVersion: machinelearning.seldon.io/v1alpha2
 kind: SeldonDeployment
 metadata:
   name: tfserving
 spec:
   name: mnist
   predictors:
   - graph:
       children: []
       implementation: TENSORFLOW_SERVER
       modelUri: gs://<your_bucket_name>/mnist
       serviceAccountName: user-gcp-sa
       name: mnist-model
       parameters:
         - name: signature_name
           type: STRING
           value: serving_default
         - name: model_name
           type: STRING
           value: mnist-model
     name: default
     replicas: 1

  10. Create an ingress rule with your service. Find out your newly created service name starting with mnist-xxx by kubectl get svc. Replae <check_svc_name_which_created_by_seldondeployment> and *.<your_host>'.

    apiVersion: extensions/v1beta1
kind: Ingress
metadata:
  name: mnist-classifier
  annotations:
    certmanager.k8s.io/acme-challenge-type: dns01
    certmanager.k8s.io/acme-dns01-provider: clouddns
    certmanager.k8s.io/cluster-issuer: letsencrypt-prod-dns
    kubernetes.io/ingress.allow-http: "true"
    kubernetes.io/ingress.class: "nginx"
    kubernetes.io/tls-acme: "true"
  namespace: default
spec:
  rules:
  - host: mnist.<your_host>
    http:
      paths:
      - backend:
          serviceName: <check_svc_name_which_created_by_seldondeployment>
          servicePort: 9000
        path: /
  tls:
  - hosts:
    - '*.<your_host>'
    secretName: dns01-tls

  11. Make a prediction through curl. Replace https://mnist.<your_host>/predict.

    curl -X POST https://mnist.<your_host>/predict -d '{"data":{"tensor":{"shape":[1,784],"values":[0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.32941176470588235, 0.7254901960784313, 0.6235294117647059, 0.592156862745098, 0.23529411764705882, 0.1411764705882353, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.8705882352941177, 0.996078431372549, 0.996078431372549, 0.996078431372549, 0.996078431372549, 0.9450980392156862, 0.7764705882352941, 0.7764705882352941, 0.7764705882352941, 0.7764705882352941, 0.7764705882352941, 0.7764705882352941, 0.7764705882352941, 0.7764705882352941, 0.6666666666666666, 0.20392156862745098, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.2627450980392157, 0.4470588235294118, 0.2823529411764706, 0.4470588235294118, 0.6392156862745098, 0.8901960784313725, 0.996078431372549, 0.8823529411764706, 0.996078431372549, 0.996078431372549, 0.996078431372549, 0.9803921568627451, 0.8980392156862745, 0.996078431372549, 0.996078431372549, 0.5490196078431373, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.06666666666666667, 0.25882352941176473, 0.054901960784313725, 0.2627450980392157, 0.2627450980392157, 0.2627450980392157, 0.23137254901960785, 0.08235294117647059, 0.9254901960784314, 0.996078431372549, 0.41568627450980394, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.3254901960784314, 0.9921568627450981, 0.8196078431372549, 0.07058823529411765, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.08627450980392157, 0.9137254901960784, 1.0, 0.3254901960784314, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.5058823529411764, 0.996078431372549, 0.9333333333333333, 0.17254901960784313, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.23137254901960785, 0.9764705882352941, 0.996078431372549, 0.24313725490196078, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.5215686274509804, 0.996078431372549, 0.7333333333333333, 0.0196078431372549, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.03529411764705882, 0.803921568627451, 0.9725490196078431, 0.22745098039215686, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.49411764705882355, 0.996078431372549, 0.7137254901960784, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.29411764705882354, 0.984313725490196, 0.9411764705882353, 0.2235294117647059, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.07450980392156863, 0.8666666666666667, 0.996078431372549, 0.6509803921568628, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.011764705882352941, 0.796078431372549, 0.996078431372549, 0.8588235294117647, 0.13725490196078433, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.14901960784313725, 0.996078431372549, 0.996078431372549, 0.30196078431372547, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.12156862745098039, 0.8784313725490196, 0.996078431372549, 0.45098039215686275, 0.00392156862745098, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.5215686274509804, 0.996078431372549, 0.996078431372549, 0.20392156862745098, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.23921568627450981, 0.9490196078431372, 0.996078431372549, 0.996078431372549, 0.20392156862745098, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.4745098039215686, 0.996078431372549, 0.996078431372549, 0.8588235294117647, 0.1568627450980392, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.4745098039215686, 0.996078431372549, 0.8117647058823529, 0.07058823529411765, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0]}}}' -H "Content-Type: application/json"

    The data is normalized vector of this image:

  12. You will get output like the screenshot below. The values field has a total of ten values (separated by commas). The first value represents the probability that the image is a '0', the second value represents the probability that the image is a '1', and so on. Thus, the eighth value represents the highest probability that this picture is '7', which is in line with our expectations (since the image is surely a '7'!). Our model makes a correct prediction!

In the future, we will provide even better model serving functions. Stay tuned!

Previous(Advanced) Use Job Submission to Tune HyperparametersNextJob Artifacts Simple Usecase