Skip to content

Overview

While building machine learning models, it is common to perform a lot of experimentation to improve model performance. Tensorboard is an extremely popular machine learning visualization toolkit that helps you visualize metrics such as loss and accuracy in training and validation data, weights and biases, model graphs, etc. TensorBoard is an open source tool built by Tensorflow that runs as a web application.

Model Visualization

You cannot improve what you don’t measure. This is true in machine learning as well!

Integrated Tensorboard

Rafay's MLOps platform for Kubeflow provides a turnkey integration with Tensorboard by making it available to users right in their dashboard. In the image below, as you can see, users just need to click on the Tensorboard menu on the left to access it.

Integrated Tensorboard

Note

Users can always use their preferred, commercial tools as well such as Neptune AI, Weights and Biases, Guild AI etc.

Benefits

Pre-Integrated

Tensorboard comes pre-integrated right in every user's dashboard. There is nothing to install, configure or maintain.

Integrated Experience

No Authentication Required

Access to Tensorboard does not require the user (i.e. data scientist or ML engineer) to be provisioned with special authentication and access credentials.

Multi Tenancy

TensorBoard inside Kubeflow has two distinct components

  • The Tensorboard CRD and Tensorboard controller.
  • Tensorboard CRUD WebApp

Clicking on “Start TensorBoard” create a TensorBoard CR in the namespace associated with the user's workspace. For example, in John's workspace, a separate TensorBoard CRD is created. By default, other users do not have access to this namespace. In summary, unless the user has access to the workspace, they will not have access to the TensorBoard associated with that workspace. 

kubectl get tensorboard -A
NAMESPACE   NAME               AGE
john-ns     tensorboard-demo   60d