Release Notes

This page documents production updates to AutoML Vision. We recommend that AutoML Vision developers periodically check this list for any new announcements.

To get the latest product updates delivered to you, add the URL of this page to your feed reader, or add the feed URL directly: https://cloud.google.com/feeds/automl-vision-image-classification-icn-release-notes.xml

November 20, 2019

Image Classification General Availability (GA) release

Some notes about the GA release:

  • Cloud AutoML Vision and AutoML Vision Edge are now in General Availability (GA). The Service Level Agreement has been published.
  • We’ve introduced a new version (v1) of the Cloud AutoML API. The current version (v1beta1) is deprecated and is scheduled for deletion within six months. As a result, you will need to migrate your applications from the v1beta1 version to the v1 version. If you have integrated with the Create Dataset API, you will also need to change your code to reflect changes with v1 version integration.
    • Creating dataset will now return an Operation, whose status must be requested to detect completion.
    • Both dataset and model can be updated using patch, with a field mask that can include labels and displayName for both, and description only for dataset.
    • In ModelEvaluation, baseAuPrc has been deprecated. This is the Area Under Precision-Recall Curve metric based on priors.
    • Specifying an Image is the only option for ExamplePayload of online prediction. The inputConfig option was not supported and has been deleted.
  • The maximum supported lifespan for custom models created in Cloud AutoML Vision and AutoML Vision Edge has changed from two years to 18 months.
  • The new pricing for Image Classification is effective November 21, 2019 at 12 AM Pacific Time. For further information, including pricing examples, please refer to our pricing page.
  • Image Classification models trained prior to the Beta Refresh release of October 9, 2019, will serve online predictions under the old pricing structure until you redeploy. Online predictions from these models will stop on February 21, 2020, unless you redeploy.

For more information, see the product documentation.

November 12, 2019

GA pricing published

The new pricing for Image Classification is effective the day after GA release later this year. For further information, including pricing examples, please refer to our pricing page.

Old prediction service live through February 20, 2020 with beta pricing

We’ve migrated your datasets to the GA (Integrated) UI. As a result, your Classification models trained prior to the migration are shown as deployed with '0 nodes' until you redeploy them.

To allow customers enough time to review their models, any model deployed with '0' nodes at GA will:

  • continue to serve online predictions from the old service until February 20, 2020.
  • be charged according to our current pricing, which is $3 per 1,000 images
  • be excluded from our Service Level Agreement (SLA) offered at GA

If no action is taken, these existing models will be undeployed on February 21, 2020. Online predictions will fail until you redeploy by following instructions published below on October 17, 2019.

October 17, 2019

Avoid Unnecessary Classification Model Deployment Charges at GA

  1. Review each project’s Models. You can review your models by:
    1. Visit the Google Cloud Console project selector dashboard to access each of your projects in turn.
    2. From each project page, click on the top left navigation menu (hamburger icon), scroll down and select Vision dashboard (Artificial Intelligence > Vision).
    3. From the Vision dashboard, select the Image Classification card under AutoML Vision. If dataset migration has completed for your project, you will be redirected to access the Integrated UI shared with Object Detection. Some customers will not be redirected in this step today, and must complete the following steps when redirection works. If you cannot complete them by October 31st, 2019, please join the cloud-vision-discuss Google group to request assistance.
  2. The Integrated UI will display your existing Classification models as deployed with ‘0 nodes’. This indicates that they are still deployed on the old online prediction service.
  3. Redeploy active Image Classification models or update model’s node number to adjust online prediction queries served per second, Each node that is selected for deployment will incur hourly charge of $1.25 (USD) at GA.
  4. When you are done with these steps, no Classification model should be left in initial state, deployed with '0 nodes'. Each model left in this state will serve online predictions from the old service, but will be excluded from our Service Level Agreement offered at GA, and will be undeployed on February 21, 2020.

October 14, 2019

Batch prediction Beta

The following Beta feature is available in API version v1beta1:

In addition to online predictions in our Cloud and predictions at the Edge, we now support batch predictions in our Cloud with your AutoML Vision models.

Making batch predictions is documented for both Image Classification and Object Detection.

TensorFlow.js Integration Beta

This integration allows AutoML Vision Edge models to be exported as TensorFlow.js package.

The edge model can be deployed in an increasing number of platforms with TensorFlow.js support, including all major browsers and server-side in Node.js. You may want to view the documented user journey for exporting to web, and TensorFlow.js tutorial for Edge Image Classification models.

October 09, 2019

Image Classification Beta Refresh

Image Classification Beta Refresh has been released. The unified infrastructure now supports both Classification and Object Detection as model objectives.

  • Dataset migration for existing Classification projects is ongoing. Unless your project has migrated, the issue published below on September 30, 2019 will continue to be relevant for you. Please follow instructions there.
  • New projects created after this release will not have this restriction. Selecting Classification as an option while creating a new dataset in the User Interface for Object Detection will succeed for them.

The following features that Classification customers had been used to will not be available in the unified infrastructure:

  • Integrated Human labeling; alternatively, you can use Cloud AI Platform Data Labeling Service.
  • Automated deployment of models will happen if you opt-in; you can manually deploy and set the number of nodes on which your model(s) are deployed. Both the UI and the API support this feature.

Model deployment & undeployment

Starting with this release, users must deploy their trained models before they are able to make predictions with them. They will also need to undeploy models to avoid incurring deployment charges when Classification models switch to deployment-based pricing for online predictions at General Availability later in 2019.

For more information, see the Deploying models and Undeploying models topics.

September 30, 2019

New Dataset Issue for Image Classification

On September 30, we begin migrating Classification datasets, prioritizing the projects created since July 2018 beta release of this feature. Following migration, creation of Classification datasets will succeed in the UI for Object Detection for these projects even before the Beta Refresh release. So the above caveats stop applying when dataset migrations is done, and the project is redirected to the unified UI.

The How-to Guides for Object Detection can be used temporarily to understand how tasks such as deploying and undeploying models are done in the unified UI. These changes will be reflected in the Classification documentation by the Beta Refresh release date.

Some caveats until AutoML Vision announces Beta Refresh release:

  • Selecting Classification as an option while creating a new dataset in the User Interface for Object Detection will succeed, but will block further progress.
  • Importing data into this dataset will fail. These release notes will be updated when this issue has been resolved.
  • To avoid this issue, please continue creating new Classification datasets in the UI for Image Classification.
  • Until our Beta Refresh release, separate selection cards for Image Classification and Object Detection will be available in the Vision Dashboard.
  • Following the release, these two cards will be replaced by single AutoML Vision card, and new Classification datasets will not be subject to this issue.

April 10, 2019

AutoML Vision Edge Beta Release

AutoML Vision Edge Beta has been released. Some notes about the Beta release:

  • You can access the updated AutoML Vision Beta UI with with AutoML Vision Edge support at https://cloud.google.com/automl/ui/vision.
  • You can now perform all of your AutoML Vision Edge tasks using the AutoML API. Create datasets, train models, make predictions, and export models. For details, see the how-to guides.
  • You can currently only supply base64-encoded image content to the predict method. For an example, see the Annotating prediction images page.

For more information, see the product documentation.

Microsoft Edge and Microsoft Internet Explorer do not support all features of AutoML Vision. If you are having problems, try Google Chrome, Safari, or Firefox.

November 01, 2018

The AutoML Vision service account requires additional permissions to access project resources when using the AutoML Vision UI. You must add the role serviceusage.serviceUsageAdmin. See step 8 in Before You Begin.

August 24, 2018

The AutoML Vision service account requires additional permissions to access project resources when using the AutoML Vision UI. You must add the role serviceusage.serviceUsageAdmin. See step 8 in Before You Begin.

July 24, 2018

Beta release

AutoML Vision Beta has been released.

Beta changes

Models created prior to July 19 must be recreated/retrained as soon as possible to avoid service interruptions. Models created prior to July 19 will only accept prediction requests until October 19.

The format of resource IDs (dataset IDs, model IDs, and operation IDs) is changing. If you are using hardcoded resource IDs, you need to obtain new IDs for those resources (using the get or list method) by August 31, 2018.

During the beta period, you can no longer use the Cloud ML Engine API to query your models.

May 03, 2018

Alpha release

AutoML Vision Alpha has been released. Some notes about the alpha release:

  • You can access the AutoML Vision Alpha UI at http://beta-dot-custom-vision.appspot.com.
  • You can now perform all of your AutoML Vision tasks using the AutoML API. Create datasets, train models, make predictions, and much more. For details, see Cloud-hosted model quickstart.
  • Making a prediction for an image is no longer supported by the images:annotate method of the Cloud Vision API. Instead, use the predict method of the AutoML API.
  • The AutoML Vision Alpha UI does not currently include the ability to list the models that have been trained for a dataset.
  • You can currently only supply base64-encoded image content to the predict method. For an example, see Make a prediction.

For more information, see the product documentation.

February 01, 2018

The vision.googleapis.com service must be enabled from the Google Cloud Platform Console. New projects automatically enable the API when you use the Do it for me button during setup.

Pricing is now in effect for advanced models.

The vision.googleapis.com service must be enabled from the Google Cloud Platform Console. New projects automatically enable the API when you use the Do it for me button during setup.

Was this page helpful? Let us know how we did:

Send feedback about...

Cloud AutoML Vision
Need help? Visit our support page.