how can we specify version of tensorflow serving in kubeflow?

Question

I am trying to use tensorflow serving to serve a model. when I try to apply serve component using ksonnet, I see that workload created on kubernetes(gke) is using tensorflow 1.7(gcr.io/kubeflow-images-public/tensorflow-serving-1.7:v20180604-0da89b8a).

Is there a way to specify what version of tensorflow serving in Kubeflow ?

Answer 1

The version of TensorFlow is determined by the TensorFlow Serving Docker image you use. If you are using the Kubeflow ksonnet prototype for TFServing the parameter modelServerImage can be used to set the image.

Answer 2

Additionally, after creating the serving prototype, you can run ks show <env_name> -c <prototype_name> to see the actual YAML file that will be deployed. This might give you a more readable view of what you actually will deploy and then you can look deeper inside of the jsonnet to see how the specific variables (in this case the image that reference the tf serving version) are set with the params that you specify in the params.jsonnet.