Seldon Core Release 0.4.1

A summary of the main contributions to the Seldon Core release 0.4.1.

Black Box Model Explanations

By utlizing Seldon’sopen source Model Explanation library Alibi we provide the ability to launch a model and an associated explainer for that model. At present we support the Anchors explanation technique for tabular text and image examples.

An example SeldonDeployment for an image model with associated explainer is shown below:

kind: SeldonDeployment
  name: image
  annotations: "10000000" "10000000" "1000000000"
  name: image
  - graph:
      children: []
      implementation: TENSORFLOW_SERVER
      modelUri: gs://seldon-models/tfserving/imagenet/model
      name: classifier
        type: GRPC
        - name: model_name
          type: STRING
          value: classifier
        - name: model_input
          type: STRING
          value: input_image
        - name: model_output
          type: STRING
          value: predictions/Softmax:0
        memory: 1Gi
      type: anchor_images
      modelUri: gs://seldon-models/tfserving/imagenet/explainer
    name: default
    replicas: 1

The Tensorflow model has a anchor_images explainer associated with it. An example input showing a persian cat along with an example explanation for that image showing the segment of the image the model focused on for providing the classifcation result can be seen below.

cat cat-explanation

We provide an example notebook with tabular, text and image model examples.

Misc. Updates

Join our slack community to discuss.