Polyaxon & Tensorboard

How to use Polyaxon and Tensorboard together

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

Overview

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
runPatch:
  connections: [...]
  environment:
    ...
  container:
    resources:
      requests:
        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.

Versions

All tensorboard versions can be found on the component hub

polyaxon hub ls -c tensorboard