Release notes

This page documents production updates to all AI Platform products. You can periodically check this page for announcements about new or updated features, bug fixes, known issues, and deprecated functionality.

For more detailed information, read the documentation for each product.

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/ai-platform-release-notes.xml

February 11, 2020

AI Platform Training

You can now set a maximum running time when you create a training job. If your training job is still running after this duration, AI Platform Training cancels the job. Set the maximum running time by specifying the scheduling.maxRunningTime field.

February 10, 2020

AI Platform Prediction

The known issue with using custom prediction routines together with runtime version 1.15 and Python 3.7 has been fixed. The issue was described in a January 23, 2020 release note.

You can now use custom prediction routines with runtime version 1.15 and Python 3.7.

February 05, 2020

AI Platform Prediction

The GPU compatibility issue that was described in the January 7, 2020 release note has been resolved. You can now use GPUs to accelerate prediction on runtime version 1.15.

AI Platform Training

The GPU compatibility issue that was described in the January 7, 2020 release note has been resolved. You can now use GPUs to accelerate training on runtime version 1.15.

February 04, 2020

AI Platform Notebooks

VPC Service Controls now supports AI Platform Notebooks. Learn how to use a notebook instance within a service perimeter. This functionality is in beta.

February 03, 2020

AI Platform Notebooks

AI Platform Notebooks now supports Access Transparency. Access Transparency provides you with logs of actions that Google staff have taken when accessing your data. To learn more about Access Transparency, see the Overview of Access Transparency.

January 29, 2020

AI Platform Prediction

AI Platform Prediction documentation has been reorganized. The new information architecture only includes documents that are relevant to AI Platform Prediction.

Previously, documentation for AI Platform Prediction and AI Platform Training were grouped together. You can now view AI Platform Training documentation separately. Some overviews and tutorials that are relevant to both products are available in the overall AI Platform documentation.

AI Platform Training

AI Platform Training documentation has been reorganized. The new information architecture only includes documents that are relevant to AI Platform Training.

Previously, documentation for AI Platform Training and AI Platform Prediction were grouped together. You can now view AI Platform Prediction documentation separately. Some overviews and tutorials that are relevant to both products are available in the overall AI Platform documentation.

January 28, 2020

AI Platform Training

AI Platform Training runtime version 1.15 now supports training with TPUs using TensorFlow 1.15.

January 23, 2020

AI Platform Prediction

Creating an AI Platform Prediction custom prediction routine that uses runtime version 1.15 and Python 3.7 might fail due to a problem with a dependency.

As a workaround, use runtime version 1.15 with Python 2.7 or use a different runtime version when you create your model version.

January 22, 2020

AI Platform Prediction

AI Explanations no longer supports AI Platform Prediction runtime version 1.13. AI Explanations now supports runtime versions 1.14 and 1.15. Learn more about AI Platform Prediction runtime versions supported by AI Explanations.

January 21, 2020

AI Platform Deep Learning VM Image

M41 release

TensorFlow Enterprise 2.1 images are now available.

MXNet upgraded to 1.5.1.

PyTorch upgraded to 1.4.0.

XGBoost upgraded to 0.90.

January 15, 2020

AI Platform Prediction

The price of using NVIDIA Tesla T4 GPUs for online prediction has changed from $0.9500 per hour to $0.3500 per hour.

GPUs for online prediction are currently only available when you deploy your model in the us-central1 region and use a Compute Engine (N1) machine type.

January 14, 2020

AI Platform Training

The price of using NVIDIA Tesla T4 GPUs for training has changed. The following table describes the pricing change for various geographic areas:

Geographic area   Old price per hour   New price per hour  
Americas $0.9500 $0.3500
Europe $1.0300 $0.3800
Asia Pacific $1.0300 $0.3900

Read more about using GPUs for training.

January 07, 2020

AI Platform Prediction

Model versions that use both runtime version 1.15 and GPUs fail due to a compatibility issue with the CuDNN library, which TensorFlow depends on.

As a workaround, do one of the following:

AI Platform Training

Training jobs that use both runtime version 1.15 and GPUs fail due to a compatibility issue with the CuDNN library, which TensorFlow depends on.

As a workaround, do one of the following:

December 20, 2019

AI Platform Training

VPC Service Controls now supports AI Platform Training. Learn how to use a service perimeter to protect your training jobs. This functionality is in beta.

December 19, 2019

AI Platform Prediction

AI Platform runtime version 1.15 is now available for prediction. This version supports TensorFlow 1.15.0 and includes other packages as listed in the runtime version list.

Runtime version 1.15 is the first runtime version to support serving predictions using Python 3.7, instead of Python 3.5. Runtime version 1.15 also still supports Python 2.7. Learn about specifying the Python version for prediction.

AI Platform Training

AI Platform Training now offers two built-in algorithms to train a machine learning model on image data without writing your own training code:

Both image algorithms are available in beta.

AI Platform runtime version 1.15 is now available for training. This version supports TensorFlow 1.15.0 and includes other packages as listed in the runtime version list.

Runtime version 1.15 is the first runtime version to support training using Python 3.7, instead of Python 3.5. Runtime version 1.15 also still supports Python 2.7. Learn about specifying the Python version for training.

Training with TPUs is not supported in runtime version 1.15 at this time.

December 10, 2019

AI Platform Prediction

Starting January 1, 2020, the Python Software Foundation will no longer support Python 2.7. Accordingly, any runtime versions released after January 1, 2020 will not support Python 2.7.

Starting on January 13, 2020, AI Platform Training and AI Platform Prediction will support each runtime version for one year after its release date. You can find the release date of each runtime version in the runtime version list.

Support for each runtime version changes according to the following schedule:

  1. Starting on the release date: You can create training jobs, batch prediction jobs, and model versions that use the runtime version.

  2. Starting 12 months after the release date: You can no longer create training jobs, batch prediction jobs, or model versions that use the runtime version.

    Existing model versions that have been deployed to AI Platform Prediction continue to function.

  3. 24 months after the release date: AI Platform Prediction automatically deletes all model versions that use the runtime version.

This policy will be applied retroactively on January 13, 2020. For example, since runtime version 1.0 was released over 24 months ago, AI Platform Training and AI Platform Prediction no longer support it. There will be a three-month grace period (until April 13, 2020) before AI Platform Prediction automatically deletes model versions that use the oldest runtime versions.

The following table describes the first two important dates that mark the end of support for runtime versions:

Date  Runtime versions affected   Change in functionality  
January 13, 2020   1.0, 1.1, 1.2, 1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 1.10, 1.11, 1.12  You can no longer create training jobs, batch prediction jobs, or model versions using these runtime versions.
April 13, 2020   1.0, 1.1, 1.2, 1.4, 1.5, 1.6  AI Platform Prediction automatically deletes any model versions using these runtime versions.

To learn about when availability ends for every runtime version, see the runtime version list.

Starting on January 13, 2020, runtimeVersion and pythonVersion will become required fields when you create Job or Version resources. Previously, runtimeVersion defaulted to 1.0 and pythonVersion defaulted to 2.7.

AI Platform Training

Starting January 1, 2020, the Python Software Foundation will no longer support Python 2.7. Accordingly, any runtime versions released after January 1, 2020 will not support Python 2.7.

Starting on January 13, 2020, AI Platform Training and AI Platform Prediction will support each runtime version for one year after its release date. You can find the release date of each runtime version in the runtime version list.

Support for each runtime version changes according to the following schedule:

  1. Starting on the release date: You can create training jobs, batch prediction jobs, and model versions that use the runtime version.

  2. Starting 12 months after the release date: You can no longer create training jobs, batch prediction jobs, or model versions that use the runtime version.

    Existing model versions that have been deployed to AI Platform Prediction continue to function.

  3. 24 months after the release date: AI Platform Prediction automatically deletes all model versions that use the runtime version.

This policy will be applied retroactively on January 13, 2020. For example, since runtime version 1.0 was released over 24 months ago, AI Platform Training and AI Platform Prediction no longer support it. There will be a three-month grace period (until April 13, 2020) before AI Platform Prediction automatically deletes model versions that use the oldest runtime versions.

The following table describes the first two important dates that mark the end of support for runtime versions:

Date  Runtime versions affected  Change in functionality
January 13, 2020  1.0, 1.1, 1.2, 1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 1.10, 1.11, 1.12 You can no longer create training jobs, batch prediction jobs, or model versions using these runtime versions.
April 13, 2020  1.0, 1.1, 1.2, 1.4, 1.5, 1.6 AI Platform Prediction automatically deletes any model versions using these runtime versions.

To learn about when availability ends for every runtime version, see the runtime version list.

Starting on January 13, 2020, AI Platform Training will automatically delete the history of each training job 120 days after it is completed. A training job is considered completed when the job enters the SUCCEEDED, FAILED, or CANCELLED state.

This policy will be applied retroactively on January 13, 2020: all jobs that were completed September 15, 2019 or earlier will be deleted.

Starting on January 13, 2020, runtimeVersion and pythonVersion will become required fields when you create Job or Version resources. Previously, runtimeVersion defaulted to 1.0 and pythonVersion defaulted to 2.7.

December 03, 2019

AI Platform Prediction

You cannot enable request-response logging for AI Platform Prediction when you create a model version. Instead, you must first create a model version without request-response logging enabled, then enable request-response logging by sending a projects.models.versions.patch request to the REST API.

November 27, 2019

AI Platform Training

AI Platform Training no longer supports TPUs in runtime version 1.12. You can still train using TPUs in runtime versions 1.13 and 1.14.

November 20, 2019

AI Platform Prediction

AI Explanations now offers feature attributions through AI Platform Prediction. This feature is available in Beta. To gain more insight on your model's predictions, you can use feature attributions based on the sampled Shapley and integrated gradients methods. Try the example notebooks to get started, and refer to the AI Explainability Whitepaper to learn more.

AI Platform Training

AI Platform Training now offers a built-in distributed XGBoost algorithm to train a machine learning model without writing your own training code. This algorithm is available in beta.

The built-in distributed XGBoost algorithm provides functionality similar to the existing single-replica version of the built-in XGBoost algorithm, but it lets you speed up training on large datasets by using multiple virtual machines in parallel. The algorithm also lets you use GPUs for training.

The built-in distributed XGBoost algorithm does not support automatic preprocessing of data.

November 01, 2019

AI Platform Deep Learning VM Image

You can now create a TensorFlow Enterprise Deep Learning VM Image. TensorFlow Enterprise image families provide users with a Google Cloud Platform optimized distribution of TensorFlow with long-term version support. To learn more about TensorFlow Enterprise, read the TensorFlow Enterprise overview.

October 28, 2019

AI Platform Training

We now recommend that you use Compute Engine machine types when you create new AI Platform Training jobs. These machine types offer the greatest flexibility for customizing the virtual CPU (vCPU), RAM, GPU, and TPU resources that your jobs use.

The older machine types available for training, which were previously referred to as "AI Platform Training machine types," are now called "legacy machine types" in the AI Platform Training documentation.

October 24, 2019

AI Platform Prediction

Many Compute Engine (N1) machine types are now available for online prediction in beta, in addition to the existing legacy (MLS1) machine types. When you create a model version with a Compute Engine machine type, you can allocate virtual machines with more vCPU and memory resources for your online prediction nodes, improving throughput of predictions or reducing latency. Additionally, you can use GPUs with these new machine types and deploy TensorFlow models up to 2 GB in size. The machine types are currently only available in the us-central1 region.

Learn more about the features, limitations, and usage of Compute Engine (N1) machine types. Model versions that use Compute Engine (N1) machine types, including with GPUs, are available at no charge until November 14, 2019. Read about the pricing for these machine types that goes into effect on November 14, 2019.

Model versions that use one of the new Compute Engine (N1) machine types and scale to use more than 40 prediction nodes may exhibit high latency when handling online prediction requests. In this case, AI Platform Prediction may also drop requests.

For the best performance until this issue is resolved, do not scale your model version to use more than 40 nodes.

The default max model size for model versions that use a legacy (MLS1) machine type has increased from 250 MB to 500 MB.

October 11, 2019

AI Platform Deep Learning VM Image

M36 release

The TensorFlow 2.0 image is out of experimental.

What-If Tool (witwidget) upgraded to 1.4.2 for TensorFlow 1.x images.

October 04, 2019

AI Platform Prediction

The us-west2 (Los Angeles), us-east4 (N. Virginia), and europe-north1 (Finland) regions are now available for batch prediction. Note that us-east4 was already available for online prediction.

Additionally, the us-west1 (Oregon) and europe-west4 (Netherlands) regions, which were already available for training, are now available for batch prediction.

Read about pricing for batch prediction in these regions.

AI Platform Training

The us-west2 (Los Angeles), us-east4 (N. Virginia), and europe-north1 (Finland) regions are now available for training. You can use NVIDIA Tesla P4 GPUs for training in us-west2 and us-east4.

Read about pricing for training in these regions, including pricing for accelerators.

September 16, 2019

AI Platform Prediction

The What-If Tool can be used to inspect models deployed on AI Platform Prediction, and to compare two models. Learn how to use the What-If Tool with AI Platform Prediction.

September 09, 2019

AI Platform Notebooks

AI Platform Notebooks now provides more ways for you to customize your network settings, encrypt your notebook content, and grant access to your notebook instance. These options are available when you create a notebook.

Now you can implement AI Platform Notebooks using custom containers. Use a Deep Learning Containers image or create a derivative container of your own, then create a new notebook instance using your custom container.

AI Platform Training

Runtime version 1.14 now supports training with TPUs using TensorFlow 1.14.

September 06, 2019

AI Platform Prediction

When you deploy a model version to AI Platform Prediction, you can now configure AI Platform Prediction to log a sample of online prediction requests received by the model together with the responses it sends to these requests. AI Platform Prediction saves these request-response pairs to BigQuery. This feature is in beta.

Learn how to how to enable request-response logging and read about the configuration options for this type of logging.

August 28, 2019

AI Platform Prediction

The documentation for AI Platform Notebooks has moved to a new location.

AI Platform Training

Training with custom containers is now generally available.

NVIDIA Tesla P4 and NVIDIA Tesla T4 GPUs are now generally available for training. Read about using GPUs for training and learn about GPU pricing.

The documentation for AI Platform Notebooks has moved to a new location.

August 26, 2019

AI Platform Deep Learning VM Image

M34 release

JupyterLab upgraded to 1.0 on all images.

PyTorch upgraded to 1.2.

AI Platform Training

AI Platform Training now supports using Cloud TPU devices with TPU v3 configurations to accelerate your training jobs. TPU v3 accelerators for AI Platform Training are available in beta.

Learn more about how to configure your training job to use TPU v3 accelerators and read about TPU v3 pricing.

August 22, 2019

AI Platform Prediction

Continuous evaluation for AI Platform Prediction is now available in beta. When you create a continuous evaluation job, AI Platform Data Labeling Service assigns human reviewers to provide ground truth labels for a portion of your model version's online predictions; alternatively, you can provide your own ground truth labels. Then Data Labeling Service compares these labels to your model version's predictions to calculate daily evaluation metrics.

Learn more about continuous evaluation.

August 16, 2019

AI Platform Prediction

AI Platform runtime versions 1.13 and 1.14 now include numpy 1.16.4 instead of numpy 1.16.0. View the runtime version list for the full list of packages included in runtime versions 1.13 and 1.14.

AI Platform Training

AI Platform runtime versions 1.13 and 1.14 now include numpy 1.16.4 instead of numpy 1.16.0. View the runtime version list for the full list of packages included in runtime versions 1.13 and 1.14.

August 01, 2019

AI Platform Prediction

The AI Platform Prediction Training and Prediction documentation has been reorganized. Previously, documentation for using AI Platform Prediction with specific machine learning frameworks was separated into sections. You can now navigate to all Training and Prediction documentation from the AI Platform documentation home page.

AI Platform Training

The AI Platform Training Training and Prediction documentation has been reorganized. Previously, documentation for using AI Platform Training with specific machine learning frameworks was separated into sections. You can now navigate to all Training and Prediction documentation from the AI Platform documentation home page.

July 19, 2019

AI Platform Prediction

AI Platform runtime version 1.14 is now available for prediction. This version supports TensorFlow 1.14.0 and includes other packages as listed in the runtime version list.

AI Platform runtime version 1.12 now supports TensorFlow 1.12.3. View the runtime version list for the full list of packages included in runtime version 1.12.

AI Platform Training

AI Platform runtime version 1.14 is now available for training. This version supports TensorFlow 1.14.0 and includes other packages as listed in the runtime version list.

Training with TPUs is not supported in runtime version 1.14 at this time.

AI Platform runtime version 1.12 now supports TensorFlow 1.12.3. View the runtime version list for the full list of packages included in runtime version 1.12.

July 17, 2019

AI Platform Prediction

The prediction input format for the following built-in algorithms has changed:

Instead of a raw string, make sure to format each instance as a JSON with a "csv_row" key and "key" key. This "key" is useful for doing batch predictions using AI Platform Prediction. For online predictions, you can pass in a dummy value to the "key" key in your input JSON request. For example:

{"csv_row": "1, 2, 3, 4, 0, abc", "key" : "dummy-key"}

See the Census Income tutorial for an example.

AI Platform Training

The prediction input format for the following built-in algorithms has changed:

Instead of a raw string, make sure to format each instance as a JSON with a "csv_row" key and "key" key. This "key" is useful for doing batch predictions using AI Platform Training. For online predictions, you can pass in a dummy value to the "key" key in your input JSON request. For example:

{"csv_row": "1, 2, 3, 4, 0, abc", "key" : "dummy-key"}

See the Census Income tutorial for an example.

July 12, 2019

AI Platform Deep Learning VM Image

M30 release

R upgraded to version 3.6.

TensorFlow: added support for using Python 3.7.

R Notebooks are no longer dependent on a Conda environment.

Fix for the bug when Nvidia driver is not installed if the user does not have the Google Cloud Storage API enabled.

What-If Tool (witwidget) fixes for TensorFlow 1.14.

Miscellaneous bug fixes.

AI Platform Notebooks

R upgraded to version 3.6.

R Notebooks are no longer dependent on a Conda environment.

July 01, 2019

AI Platform Deep Learning VM Image

M28 release

What-If Tool (witwidget) added to DLVM.

Fixed TensorFlow 1.14 issues.

Miscellaneous bug fixes.

June 24, 2019

AI Platform Deep Learning Containers

AI Platform Deep Learning Containers is now available in beta. AI Platform Deep Learning Containers lets you quickly prototype with a portable and consistent environment for developing, testing, and deploying your AI applications.

Visit the AI Platform Deep Learning Containers overview and the guide to getting started with a local deep learning container.

June 20, 2019

AI Platform Deep Learning VM Image

M27.1 release updates

TensorFlow upgraded to: 1.14.0.

TensorFlow 2.0 upgraded to: Beta 1.

Miscellaneous bug fixes.

June 19, 2019

AI Platform Prediction

The asia-southeast1 (Singapore) region is now available for batch prediction.

AI Platform Training

The asia-southeast1 (Singapore) region is now available for training. You can use P4 or T4 GPUs for training in the region. Read about pricing for training in asia-southeast1, including pricing for accelerators.

June 18, 2019

AI Platform Training

Runtime version 1.13 now supports training with TPUs using TensorFlow 1.13.

Support for training with TPUs in runtime version 1.11 ended on June 6, 2019.

June 17, 2019

AI Platform Deep Learning VM Image

M27 release

New ML framework added: CNTK 2.7 from Microsoft.

New ML framework added: Caffe 1.0 BVLC from UC Berkeley.

Updated TensorFlow 2.0 Beta0.

Miscellaneous bug fixes.

June 12, 2019

AI Platform Training

You can now view monitoring data for training jobs directly within the AI Platform Training Job Details page in the Cloud Console. The following charts are available:

  • CPU, GPU, and memory utilization, broken down by master, worker, and parameter servers.
  • Network usage: the rate per second of bytes sent and received.

Learn more about how to monitor resource utilization for your training jobs.

There are new options for filtering jobs within the AI Platform Training Jobs page in the Cloud Console. You can filter jobs by Type and by whether or not the job used HyperTune.

Learn more about how to filter your training jobs.

You can now view and sort hyperparameter tuning trials within the AI Platform Training Job Details page in the Cloud Console. If your training job uses hyperparameter tuning, your Job Details page includes a HyperTune trials table, where you can view metrics such as RMSE, learning rate, and training steps. You can also access logs for each trial. This table makes it easier to compare individual trials.

Learn more about how to view your hyperparameter tuning trials.

June 05, 2019

AI Platform Prediction

You can now specify a service account for your model version to use when you deploy a custom prediction routine to AI Platform Prediction. This feature is in beta.

June 03, 2019

AI Platform Notebooks

You can now create AI Platform Notebooks instances with R and core R packages installed. Learn how to install R dependencies, and read guides for using R with BigQuery in AI Platform Notebooks and using R and Python in the same notebook.

AI Platform Training

You can now create AI Platform Notebooks instances with R and core R packages installed. Learn how to install R dependencies, and read guides for using R with BigQuery in AI Platform Notebooks and using R and Python in the same notebook.

May 29, 2019

AI Platform Deep Learning VM Image

M26 release

RAPIDS updated to 0.7.

Faster driver installation time for common TensorFlow and PyTorch images.

You can now use Deep Learning VMs without a public IP address if you have enabled Google Private Access.

Miscellaneous bug fixes.

May 03, 2019

AI Platform Deep Learning VM Image

M25 release

New image added: CUDA 10.1.

PyTorch upgraded to 1.1.0.

fastai upgraded to 1.0.52.

MXNet upgraded to 1.4.0 (and now based on CUDA 10.0 images).

Chainer upgraded to 5.4.0.

AI Platform Prediction

AI Platform runtime version 1.12 now supports TensorFlow 1.12.2. View the runtime version list for the full list of packages included in runtime version 1.12.

AI Platform Training

T4 GPUs are now in beta for AI Platform Training. For more information, see the guides to using GPUs, their regional availability, and their pricing.

AI Platform runtime version 1.12 now supports TensorFlow 1.12.2. View the runtime version list for the full list of packages included in runtime version 1.12.

April 26, 2019

AI Platform Deep Learning VM Image

M24 release

We now support two authorization modes in the new release: single user mode and service account mode3.

rpy2 is now pre-installed in the R image.

A plugin for editing metadata of cells is now pre-installed.

jupyterlab-celltags JupyterLab extension is now pre-installed.

Fixed bug with sudo (now you can use sudo from the JupyterLab terminal).

Downloading files from JupyterLab file browser is now working.

April 25, 2019

AI Platform Prediction

AI Platform Prediction now supports custom prediction routines in beta. Custom prediction routines let you provide AI Platform Prediction with custom code to use when it serves online predictions from your deployed model. This can be useful for preprocessing prediction input, postprocessing your model's predictions, and more.

Work through a tutorial on deploying a custom prediction routine with Keras or one on deploying a custom prediction routine with scikit-learn.

AI Platform Prediction now supports custom transformers for scikit-learn pipelines in beta. This lets you provide AI Platform Prediction with custom code to use during online prediction. Your deployed scikit-learn pipeline uses this code when it serves predictions.

Work through a tutorial on training and deploying a custom scikit-learn pipeline.

AI Platform Prediction now supports logging of your prediction nodes' stderr and stdout streams to Stackdriver logging during online prediction. Stream logging is in beta. You can enable this type of logging in addition to—or in place of—the access logging that was already available. It can be useful for understanding how your deployment handles prediction requests.

April 10, 2019

AI Platform Data Labeling Service

AI Platform Data Labeling Service Beta has been released.

March 15, 2019

AI Platform Deep Learning VM Image

M22 release

Tensorflow upgraded to version 1.13.

Fairing now preinstalled.

cookiecutter and seaborn now preinstalled.

More descriptive serial logs to help customers debug common issues.

Misc bug fixes.

Due to incompatibilities between Tensorflow 1.13 (which requires Numpy 1.16.2 or greater) and the latest Intel optimized version of Numpy (which is 1.15) we are not using the intel optimized versions of Numpy and Scipy for this release.

March 01, 2019

AI Platform Notebooks

AI Platform Notebooks is now available in beta. AI Platform Notebooks enables you to create and manage virtual machine (VM) instances that are pre-packaged with JupyterLab and a suite of deep learning software.

Visit the AI Platform Notebooks overview and the guide to creating a new notebook instance to learn more.

February 21, 2019

AI Platform Deep Learning VM Image

M20 release

TensorFlow and Pytorch GPU images switch between CPU-only/GPU-enabled binaries at startup depending on whether GPUs are attached.

SSH is not disabled during NVIDIA driver installation on GPU images.

Due to incompatibilities between the latest kernel update (Debian 9.8) and Docker, we have put a hold on the kernel updates for this release (that is, apt-mark hold linux-image-4.9.0-8-amd64). If you require the latest kernel, you can run sudo apt-mark unhold linux-image-4.9.0-8-amd64 && sudo apt upgrade, but we cannot guarantee that Docker or our direct JupyterLab link from Marketplace will function correctly if you force the upgrade.

January 29, 2019

AI Platform Deep Learning VM Image

M19 release

New TensorFlow 2.0 (experimental) flavor is added.

New experimental ability to use Deep Learning VMs with special Web proxy, instead of SSHing to the VM.

January 14, 2019

AI Platform Deep Learning VM Image

M16 release

New MXNet 1.3 (experimental) flavor is added.

December 19, 2018

AI Platform Deep Learning VM Image

General Availability

Launched the new 1.0 version of AI Platform Deep Learning VM Image.

M15 release

BigQuery magic plugin now preloaded all the time.

Jupyter SQL integration now pre-installed and SQL plugin now preloaded.

TensorFlow images now include bazel pre-installed.

Python Dataproc client now pre-installed on all our images.

fastai updated to the latest version 1.0.38.

December 10, 2018

AI Platform Deep Learning VM Image

M14 release

Fixed bug that was resulting in a broken Git UI in some cases.

Fast.Ai updated to 1.0.36.

December 05, 2018

AI Platform Deep Learning VM Image

M13 release

Integrates fix for speed regression in linear models when using TensorFlow with Intel® MKL DNN.

Adds Git-Jupyter integration.

November 20, 2018

AI Platform Deep Learning VM Image

M12 release

Chainer is now upgraded to 5.0.0 (and CuPy to 5.0.0).

CuDNN updated to 7.4.

TensorRT5 updated to GA.

XGBoost updated to 0.81.

Images now have papermill pre-installed.

Ability to change Jupyter UI that is running on the port 8080, currently supported: Lab and Notebook.

November 13, 2018

AI Platform Deep Learning VM Image

M11.1 release

Fixed an issue where users were locked out of apt after startup due to a package needing configuration. If you are using an M11 image and are experiencing issues with apt, please either recreate your VM or run sudo dpkg --configure -a to clear the lock.

November 08, 2018

AI Platform Deep Learning VM Image

M11 release

All GPU images install NVIDIA driver 410.72.

TensorFlow updated to v1.12.0.

PyTorch 0.4 image now uses conda for package management.

October 23, 2018

AI Platform Deep Learning VM Image

M10 release

PyTorch 1.0 updated to the latest build as of October 23.

fastai updated to 1.0.12.

fastai course materials are now available at $HOME/tutorials/fastai/.

Chainer UI updated to 0.6.0.

Chainer MN updated to 1.3.1.

Fixed a bug that was causing Intel packages to be overwritten.

October 10, 2018

AI Platform Deep Learning VM Image

M9 release

Intel Optimized Python packages are installed in all distributions:

  • NumPy
  • SciPy
  • scikit-learn
  • TensorFlow (when applicable)

PyTorch 1.0 (Experimental) images include support for [conda](https://conda.io/) and [fastai](http://fast.ai/).

Chainer updated from v4.4.0 to v4.5.0.

September 27, 2018

AI Platform Deep Learning VM Image

M8 release

New XGBoost images:

  • xgboost-<var>VERSION</var>-cu92-experimental
  • xgboost-<var>VERSION</var>-cpu-experimental

New CUDA 10.0 image (common-cu100) with the following NVIDIA stack in it:

  • CuDNN 7.3
  • NCCL 2.3.4
  • Driver 410.48
  • TensorRT 5

TensorFlow updated from v1.10.1 to v1.11.0.

TensorFlow now compiled with CUDA 10.0 and CuDNN 7.3.

Common CUDA 9.2 image now has latest NCCL 2.3.4

Common CUDA 9.0 image now has:

  • latest NCCL 2.3.4
  • latest CuDNN 7.3
  • TensorRT 5.0.0

Following packages are now pre-installed on the images:

  • htop
  • protobuf-compiler
  • tree

After SSHing to the instance you now will see the exact revision of the image in the header.

September 18, 2018

AI Platform Deep Learning VM Image

M7.1 release

Introducing new experimental images with PyTorch 1.0RC. New image families are:

  • pytorch-1-0-cu92-experimental
  • pytorch-1-0-cpu-experimental

September 12, 2018

AI Platform Deep Learning VM Image

M7 release

Chainer updated from v4.3.0 to v4.4.0.

Better integration with BigQuery.

Pillow has been replaced with the faster Pillow-SIMD package.

minikube is now pre-installed.

New simplified image families introduced:

  • tf-latest-gpu
  • pytorch-latest-gpu
  • chainer-latest-gpu-experimental

Jupyter now running on behalf of its own user (not root).

August 30, 2018

AI Platform Deep Learning VM Image

M6 release

Introducing experimental images: these images bring new frameworks for you to try out, but they come with no guarantees of future support. Current experimental images:

  • Chainer (4.3)

All images now have kubectl installed.

TensorFlow updated from v1.10.0 to v1.10.1.

August 14, 2018

AI Platform Deep Learning VM Image

M5 release

All images now have Docker and/or NVIDIA Docker pre-installed.

TensorFlow and PyTorch images now include pre-baked tutorials.

GPU flavors of TensorFlow and PyTorch images now swap binaries to the CPU optimized binaries during the first boot if the instance does not have a GPU.

July 31, 2018

AI Platform Deep Learning VM Image

M4 release

Includes Tensorfow Serving: model server binary at /usr/local/bin/tensorflow_model_server and tensorflow-serving-api preinstalled.

Integration with Colab: default JupyterLab instance can be connected as a Colab backend.

Upgraded to support CUDA 9.2 (note this changes the pytorch family name).

Fixed an issue with CUDA linking in the build process, binaries up to 10% faster now.

July 17, 2018

AI Platform Deep Learning VM Image

M3 release

New common image with CUDA 9.0 has been introduced.

The following changes are included in this release:

  • All images now include OpenMPI.
  • TensorFlow GPU images now include Horovod.
  • CUDA 9.2 stack now includes latest NCCL 2.2.13.

Bug that was preventing Jupyter Notebook from working correctly has been resolved.

July 11, 2018

AI Platform Deep Learning VM Image

M2 release

TensorFlow updated to version 1.9.0.

New public Google Group for users: google-dl-platform

July 02, 2018

AI Platform Deep Learning VM Image

Beta launch

AI Platform Deep Learning VM Image is available as a beta release.