Model Explainers¶

cat explanation

Seldon provides model explanations using its Alibi library.

The v1 explainer server supports explainers saved with Python 3.7. However, for the Open Inference Protocol (or V2 protocol) using MLServer, this requirement is no longer necessary.

Package

Version

alibi

0.6.4

Available Methods¶

Seldon Core supports a subset of the methods currently available in Alibi. Presently this the following:

Method

Explainer Key

Anchor Tabular

AnchorTabular

Anchor Text

AnchorText

Anchor Images

AnchorImages

kernel Shap

KernelShap

Integrated Gradients

IntegratedGradients

Tree Shap

TreeShap

Creating your explainer¶

For Alibi explainers that need to be trained you should

  1. Use python 3.7 as the Seldon Alibi Explain Server also runs in python 3.7.10 when it loads your explainer.

  2. Follow the Alibi docs for your particular desired explainer. The Seldon Wrapper presently supports: Anchors (Tabular, Text and Image), KernelShap and Integrated Gradients.

  3. Save your explainer using explainer.save method and store in the object store or PVC in your cluster. We support various cloud storage solutions through our init container.

The runtime environment in our Alibi Explain Server is locked using Poetry. See our e2e example here on how to use that definition to train your explainers.

Open Inference Protocol for explainer using MLServer¶

The support for Open Inference Protocol is now handled with MLServer moving forward. This is experimental and only works for black-box explainers.

For an e2e example, please check AnchorTabular notebook here.

Explain API¶

For the Seldon Protocol an endpoint path will be exposed for:

http://<ingress-gateway>/seldon/<namespace>/<deployment name>/<predictor name>/api/v1.0/explain

So for example if you deployed:

apiVersion: machinelearning.seldon.io/v1
kind: SeldonDeployment
metadata:
  name: income
  namespace: seldon
spec:
  name: income
  annotations:
    seldon.io/rest-timeout: "100000"
  predictors:
  - graph:
      children: []
      implementation: SKLEARN_SERVER
      modelUri: gs://seldon-models/sklearn/income/model-0.23.2
      name: classifier
    explainer:
      type: AnchorTabular
      modelUri: gs://seldon-models/sklearn/income/explainer-py36-0.5.2
    name: default
    replicas: 1

If you were port forwarding to Ambassador or istio on localhost:8003 then the API call would be:

http://localhost:8003/seldon/seldon/income-explainer/default/api/v1.0/explain

The explain method is also supported for tensorflow and Open Inference protocols. The full list of endpoint URIs is:

Protocol

URI

seldon

http://<host>/<ingress-path>/api/v1.0/explain

tensorflow

http://<host>/<ingress-path>/v1/models/<model-name>:explain

v2

http://<host>/<ingress-path>/v2/models/<model-name>/infer

Note: for tensorflow protocol we support similar non-standard extension as for the prediction API, http://<host>/<ingress-path>/v1/models/:explain.