Outlier Detection on CIFAR10

demo

In this example we will deploy an image classification model along with an outlier detector trained on the same dataset. For in depth details on creating an outlier detection model for your own dataset see the alibi-detect project and associated documentation. You can find details for this CIFAR10 example in their documentation as well.

Prequisites:

• Knative eventing installed

  • Ensure the istio-ingressgateway is exposed as a loadbalancer (no auth in this demo)

• Seldon Core installed

  • Ensure you install for istio, e.g. for the helm chart --set istio.enabled=true

  Tested on GKE and Kind with Knative 0.18 and Istio 1.7.3

!pip install -r requirements_notebook.txt

Ensure istio gateway installed

!kubectl apply -f ../../../notebooks/resources/seldon-gateway.yaml

!cat ../../../notebooks/resources/seldon-gateway.yaml

Setup Resources

!kubectl create namespace cifar10

%%writefile broker.yaml
apiVersion: eventing.knative.dev/v1
kind: broker
metadata:
 name: default
 namespace: cifar10

!kubectl create -f broker.yaml

%%writefile event-display.yaml
apiVersion: apps/v1
kind: Deployment
metadata:
  name: hello-display
  namespace: cifar10
spec:
  replicas: 1
  selector:
    matchLabels: &labels
      app: hello-display
  template:
    metadata:
      labels: *labels
    spec:
      containers:
        - name: event-display
          image: gcr.io/knative-releases/knative.dev/eventing-contrib/cmd/event_display

---

kind: Service
apiVersion: v1
metadata:
  name: hello-display
  namespace: cifar10
spec:
  selector:
    app: hello-display
  ports:
  - protocol: TCP
    port: 80
    targetPort: 8080

!kubectl apply -f event-display.yaml

Create the SeldonDeployment image classification model for Cifar10. We add in a logger for requests - the default destination is the namespace Knative Broker.

%%writefile cifar10.yaml
apiVersion: machinelearning.seldon.io/v1
kind: SeldonDeployment
metadata:
  name: tfserving-cifar10
  namespace: cifar10
spec:
  protocol: tensorflow
  transport: rest
  predictors:
  - componentSpecs:
    - spec:
        containers:
        - args: 
          - --port=8500
          - --rest_api_port=8501
          - --model_name=resnet32
          - --model_base_path=gs://seldon-models/tfserving/cifar10/resnet32
          image: tensorflow/serving
          name: resnet32
          ports:
          - containerPort: 8501
            name: http
            protocol: TCP
    graph:
      name: resnet32
      type: MODEL
      endpoint:
        service_port: 8501
      logger:
        mode: all
        url: http://broker-ingress.knative-eventing.svc.cluster.local/cifar10/default
    name: model
    replicas: 1

!kubectl apply -f cifar10.yaml

Create the pretrained VAE Cifar10 Outlier Detector. We forward replies to the message-dumper we started.

Here we configure seldonio/alibi-detect-server to use rclone for downloading the artifact. If RCLONE_ENABLED=true environmental variable is set or any of the environmental variables contain RCLONE_CONFIG in their name then rclone will be used to download the artifacts. If RCLONE_ENABLED=false or no RCLONE_CONFIG variables are present then kfserving storage.py logic will be used to download the artifacts.

%%writefile cifar10od.yaml

apiVersion: v1
kind: Secret
metadata:
  name: seldon-rclone-secret
  namespace: cifar10    
type: Opaque
stringData:
  RCLONE_CONFIG_GS_TYPE: google cloud storage
  RCLONE_CONFIG_GS_ANONYMOUS: "true"

---            

apiVersion: serving.knative.dev/v1
kind: Service
metadata:
  name: vae-outlier
  namespace: cifar10
spec:
  template:
    metadata:
      annotations:
        autoscaling.knative.dev/minScale: "1"
    spec:
      containers:
      - image: seldonio/alibi-detect-server:1.8.0-dev
        imagePullPolicy: IfNotPresent
        args:
        - --model_name
        - cifar10od
        - --http_port
        - '8080'
        - --protocol
        - tensorflow.http
        - --storage_uri
        - gs://seldon-models/alibi-detect/od/OutlierVAE/cifar10
        - --reply_url
        - http://hello-display.cifar10
        - --event_type
        - io.seldon.serving.inference.outlier
        - --event_source
        - io.seldon.serving.cifar10od
        - OutlierDetector
        envFrom:
        - secretRef:
            name: seldon-rclone-secret

!kubectl apply -f cifar10od.yaml

Create a Knative trigger to forward logging events to our Outlier Detector.

%%writefile trigger.yaml
apiVersion: eventing.knative.dev/v1
kind: Trigger
metadata:
  name: vaeoutlier-trigger
  namespace: cifar10
spec:
  broker: default
  filter:
    attributes:
      type: io.seldon.serving.inference.request
  subscriber:
    ref:
      apiVersion: serving.knative.dev/v1
      kind: Service
      name: vae-outlier
      namespace: cifar10

!kubectl apply -f trigger.yaml

Get the IP address of the Istio Ingress Gateway. This assumes you have installed istio with a LoadBalancer.

CLUSTER_IPS = !(kubectl -n istio-system get service istio-ingressgateway -o jsonpath='{.status.loadBalancer.ingress[0].ip}')
CLUSTER_IP = CLUSTER_IPS[0]
print(CLUSTER_IP)

Optionally add an authorization token here if you need one.Acquiring this token will be dependent on your auth setup.

TOKEN = "Bearer <my token>"

If you are using Kind or Minikube you will need to port-forward to the istio ingressgateway and uncomment the following

# CLUSTER_IP="localhost:8004"

SERVICE_HOSTNAMES = !(kubectl get ksvc -n cifar10 vae-outlier -o jsonpath='{.status.url}' | cut -d "/" -f 3)
SERVICE_HOSTNAME_VAEOD = SERVICE_HOSTNAMES[0]
print(SERVICE_HOSTNAME_VAEOD)

import json

import matplotlib.pyplot as plt
import numpy as np
import tensorflow as tf

tf.keras.backend.clear_session()

import requests
from alibi_detect.od.vae import OutlierVAE
from alibi_detect.utils.perturbation import apply_mask
from alibi_detect.utils.visualize import plot_feature_outlier_image

train, test = tf.keras.datasets.cifar10.load_data()
X_train, y_train = train
X_test, y_test = test

X_train = X_train.astype("float32") / 255
X_test = X_test.astype("float32") / 255
print(X_train.shape, y_train.shape, X_test.shape, y_test.shape)
classes = (
    "plane",
    "car",
    "bird",
    "cat",
    "deer",
    "dog",
    "frog",
    "horse",
    "ship",
    "truck",
)


def show(X):
    plt.imshow(X.reshape(32, 32, 3))
    plt.axis("off")
    plt.show()


def predict(X):
    formData = {"instances": X.tolist()}
    headers = {"Authorization": TOKEN}
    res = requests.post(
        "http://"
        + CLUSTER_IP
        + "/seldon/cifar10/tfserving-cifar10/v1/models/resnet32/:predict",
        json=formData,
        headers=headers,
    )
    if res.status_code == 200:
        return classes[np.array(res.json()["predictions"])[0].argmax()]
    else:
        print("Failed with ", res.status_code)
        return []


def outlier(X):
    formData = {"instances": X.tolist()}
    headers = {
        "Alibi-Detect-Return-Feature-Score": "true",
        "Alibi-Detect-Return-Instance-Score": "true",
        "ce-namespace": "default",
        "ce-modelid": "cifar10",
        "ce-type": "io.seldon.serving.inference.request",
        "ce-id": "1234",
        "ce-source": "localhost",
        "ce-specversion": "1.0",
    }
    headers["Host"] = SERVICE_HOSTNAME_VAEOD
    headers["Authorization"] = TOKEN
    res = requests.post("http://" + CLUSTER_IP + "/", json=formData, headers=headers)
    if res.status_code == 200:
        od = res.json()
        od["data"]["feature_score"] = np.array(od["data"]["feature_score"])
        od["data"]["instance_score"] = np.array(od["data"]["instance_score"])
        return od
    else:
        print("Failed with ", res.status_code)
        return []

Normal Prediction

idx = 1
X = X_train[idx : idx + 1]
show(X)
predict(X)

Lets check the message dumper for an outlier detection prediction. This should be false.

!kubectl logs -n cifar10 $(kubectl get pod -n cifar10 -l app=hello-display -o jsonpath='{.items[0].metadata.name}')

Outlier Prediction

np.random.seed(0)
X_mask, mask = apply_mask(
    X.reshape(1, 32, 32, 3),
    mask_size=(10, 10),
    n_masks=1,
    channels=[0, 1, 2],
    mask_type="normal",
    noise_distr=(0, 1),
    clip_rng=(0, 1),
)

show(X_mask)
predict(X_mask)

Now lets check the message dumper for a new message. This should show we have found an outlier.

!kubectl logs -n cifar10 $(kubectl get pod -n cifar10 -l app=hello-display -o jsonpath='{.items[0].metadata.name}')

We will now call our outlier detector directly and ask for the feature scores to gain more information about why it predicted this instance was an outlier.

od_preds = outlier(X_mask)

We now plot those feature scores returned by the outlier detector along with our original image.

plot_feature_outlier_image(od_preds, X_mask, X_recon=None)

Tear Down

!kubectl delete ns cifar10