Is it possible to use Vertex AI experiments without Tensorboard?

[maxromanovsky]

Is it possible to use Vertex AI experiments without Tensorboard?

When I programmatically init the SDK and start experiment run as described in docs, Tensorboard is automatically provisioned.

I would like to use only summary metrics and parameters tracking (i.e. not time-series metrics), and AFAIK it doesn't require Tensorboard.

Here's the code snippet:

aiplatform.init(experiment="foo", location="us-central1")
aiplatform.start_run(run="bar")
aiplatform.log_metrics({"score": 0.42})

[kvandres]

Good day @maxromanovsky ,

Welcome to Google Cloud Community!

If you have assigned a backing tensorboard, you can delete the backing Vertex AI TensorBoard using this parameter delete_backing_tensorboard_run, I assume that it is not related to your provided snippet since you did not specify the tensorboard or experiment_tensorboard in your run or init. You can check this link to learn more: https://cloud.google.com/python/docs/reference/aiplatform/1.15.0/google.cloud.aiplatform.ExperimentR...

Hope this will help!

[sascha_heyer]

As kvandre said 
Yes you can use Vertex AI Experiments without TensorBoard instance (as long as you don't need Time Series Metrics).

The TensorBoard Instance is not automatically created. 
To create it, you have to call explicitly:

experiment_tensorboard = vertex_ai.Tensorboard.create()

I wrote a deep dive article on Vertex AI Experiments a few weeks ago:
https://medium.com/google-cloud/machine-learning-experiment-tracking-with-vertex-ai-8406f8d44376

Best regards

[maxromanovsky]

@sascha_heyer I definitely didn't call vertex_ai.Tensorboard.create() explicitly.

All I did was:

• Created an experiment using console UI. By that time there was no link to Tensorboard in the experiment, and no sign of tensorboard being created under the hood (i.e. no spinning wheel), even after page refresh
• Ran locally the script with the snipped above. Once script started running, tensorboard started provisioning.

SDK version was google-cloud-aiplatform 1.25.0.

This was counterintuitive, unexpected and costly 🙂

[mmozer]

This appears to be a "feature" of 1.25.0 of google-cloud-aiplatform:
https://github.com/googleapis/python-aiplatform/commit/fa7d3a0e3cd5040eb4ab1c3b0df4e494dc84bac3

[jubebo]

Its true that the default behaviour was changed with version 1.25.0 of google-cloud-aiplatform. In my article on medium I provide some background on why this was changed:

https://medium.com/@j.berzborn/what-you-can-expect-from-vertex-tensorboard-53f5995bbac3