Polyaxon integrates with Tensorboard to visualize and debug deep learning models.


Polyaxon provides several ways for using Tensorboard, you can check the component on Polyaxon for more details.

Note: The default version latest targets a single run, and it has the same component’s content as tensorboard:single-run.

Create a new tensorboard using the CLI

polyaxon run --hub tensorboard -P uuid=UUID

The UUID is the uuid of the experiment to start a tensorboard for.

To provide a specific tag:

 polyaxon run --hub tensorboard:multi-run -P uuids=UUID1,UUID2

To start the session on a different project

polyaxon run --hub tensorboard -p project-name

Viewing the service

Go to the UI under the service tab:

polyaxon ops dashboard [-uid] [-p]

Or to get to the service directly:

polyaxon ops service [-uid] [-p]

Or to get the service in full-screen mode:

polyaxon ops service --external [-uid] [-p]
  • Single-run tensorboard



  • Multi-run tensorboard



Queues, presets, node scheduling, custom resources, …

You can provide more information before scheduling the service, like the queue, presets, …

polyaxon run --hub tensorboard -q agent-name/queue-name --presets preset-name1,preset-name2

Local preset

polyaxon run --hub tensorboard -f path/to/preset.yaml

You can also provide a full operation manifest to customize the environment section, node selector, connections, initializers, resources requirements, …

version: 1.1
kind: operation
hubRef: tensorboard
  connections: [...]
        memory: 300Mi

Forking and customizing the component

If you need to expose the tensorboard component with your predefined configuration without requiring end-users to create operations, we suggest that you clone the component and customize it.

On Polyaxon CE, you will need to create a new .yaml file where you will host the content of the component, and users can either start new sessions using:

  • polyaxon run -f my-custom-tensorboard.yaml
  • polyaxon run --url https://path/to/my-custom-tensorboard.yaml

On Polyaxon Cloud or Polyaxon EE, you just need to add a new component hub. The end users will need to run with org-name/tensorboard instead of tensorboard:

polyaxon run --hub acme/tensorboard

Note: In order to use acme/tensorboard without :tag you need to name the version latest.


All tensorboard versions can be found on the component hub

polyaxon hub ls -c tensorboard