Release Notes

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

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.

Training with custom containers is now generally available.

Using Compute Engine machine types for your training configuration 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.

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.

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.

AI Platform runtime version 1.14 is now available for training and prediction. 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.

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

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

You can now view monitoring data for training jobs directly within the AI Platform Job Details page in the GCP 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 Jobs page in the GCP 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 Job Details page in the GCP 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.

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.

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.

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 Prediction now supports custom prediction routines in beta. Custom prediction routines let you provide AI Platform 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 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.

AI Platform now offers reduced pricing for training, online prediction and batch prediction.

Learn more about AI Platform pricing.

AI Platform now offers training with built-in algorithms. You can submit your data for automatic preprocessing, and train a model on the TensorFlow linear learner, TensorFlow wide and deep, and XGBoost algorithms without writing any code.

Learn more about training with built-in algorithms.

Support for training with TPUs in AI Platform runtime version 1.9 ended on March 8, 2019. See the currently supported versions in the runtime version list.

AI Platform runtime version 1.13 is now available for training and prediction. This version supports TensorFlow 1.13 and includes other packages as listed in the runtime version list.

Training with TPUs is not supported in runtime version 1.13 at this time.

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.

Cloud TPU is now generally available for training TensorFlow models. Tensor Processing Units (TPUs) are Google's custom-developed accelerators for machine-learning workloads.

See how to use TPUs to train your models on AI Platform, and read more about their pricing.

Training with custom containers is now available in Beta. This feature allows you to run your training application on AI Platform using a custom Docker image. You can build your custom container with the ML frameworks of your choice. Get started with training a PyTorch model by using custom containers.

You can now configure training jobs with certain Compute Engine machine types. This provides additional flexibility for allocating computing resources to your training jobs. This feature is available in Beta.

When you configure your job with Compute Engine machine types, you may attach a custom set of GPUs.

Read more about Compute Engine machine types, GPU attachments, and their pricing.

P4 GPUs are now in Beta for training. For more information, see the guides to using GPUs, their regional availability, and their pricing.

Online prediction is now available in the us-east4 region. See the guide to region availability.

V100 GPUs are now generally available for training. For more information, see the guides to using GPUs and pricing.

The AI Platform runtime versions 1.11 and 1.12 are now available for training and prediction. These versions support TensorFlow 1.11 and 1.12 respectively, and other packages as listed in the runtime version list.

TPU training support has been added for AI Platform runtime versions 1.11 and 1.12. Version 1.10 is not supported. See the currently supported versions in the runtime version list.

The AI Platform runtime version 1.11 is now available for training and prediction. This version supports TensorFlow 1.11 and other packages as listed in the runtime version list.

The AI Platform runtime version 1.10 is now available for training and prediction. This version supports TensorFlow 1.10 and other packages as listed in the runtime version list.

V100 GPUs are now in Beta for training. Using V100 GPUs now incurs charges. For more information, see the guides to using GPUs and pricing.

P100 GPUs are now generally available for training. For more information, see the guides to using GPUs and pricing.

Two new regions: us-west1, europe-west4 are now available for training. See regions page for info.

The AI Platform runtime version 1.9 is now available for training and prediction. This version supports TensorFlow 1.9 and other packages as listed in the runtime version list.

AI Platform now supports scikit-learn and XGBoost for training. This feature is generally available. See the guide to training with scikit-learn and XGBoost on AI Platform.

Online prediction support for scikit-learn and XGBoost is now generally available.

• Set framework on projects.models.versions.create to specify your machine learning framework when creating a model version. Valid values are TENSORFLOW, SCIKIT_LEARN, and XGBOOST. The default is TENSORFLOW. If you specify SCIKIT_LEARN or XGBOOST, you must also set runtimeVersion to 1.4 or greater on the model version.
• See the guide to local training and online predictions with scikit-learn and XGBoost.

You can add labels to your AI Platform resources—jobs, models, and model versions—then use those labels to organize the resources into categories. Labels are also available on operations—in this case the labels are derived from the resource to which the operation applies. Read more about adding and using labels.

The AI Platform runtime version 1.8 is now available for training and prediction. This version supports TensorFlow 1.8 and other packages as listed in the runtime version list.

You can now update the minimum number of nodes for autoscaling on an existing model version, as well as specifying the attribute when creating a new version.

AI Platform now offers Cloud TPU (Beta) for training TensorFlow models. Tensor Processing Units (TPUs) are Google’s custom-developed ASICs, used to accelerate machine-learning workloads. See how to use TPUs to train your models on AI Platform.

The AI Platform runtime version 1.7 is now available for training and prediction. This version supports TensorFlow 1.7 and other packages as listed in the runtime version list.

Hyperparameter algorithms: When tuning the hyperparameters in your training job, you can now specify a search algorithm in the HyperparameterSpec. Available values are:

• GRID_SEARCH: A simple grid search within the feasible space. This option is particularly useful if you want to specify a number of trials that is more than the number of points in the feasible space. In such cases, if you do not specify a grid search, the AI Platform default algorithm may generate duplicate suggestions. To use grid search, all parameters must be of type INTEGER, CATEGORICAL, or DISCRETE.
• RANDOM_SEARCH: A simple random search within the feasible space.

If you do not specify an algorithm, your job uses the default AI Platform algorithm, which drives the parameter search to arrive at the optimal solution with a more effective search over the parameter space. For more about hyperparameter tuning, see the hyperparameter tuning overview.

AI Platform now supports scikit-learn and XGBoost for online prediction. This feature is in Beta.

• Set framework on projects.models.versions.create to specify your machine learning framework when creating a model version. Valid values are TENSORFLOW, SCIKIT_LEARN, and XGBOOST. The default is TENSORFLOW. If you specify SCIKIT_LEARN or XGBOOST, you must also set runtimeVersion to 1.4 or greater on the model version.
• See the guide to scikit-learn and XGBoost on AI Platform.

Python 3.5 is available for online prediction.

• Set pythonVersion on projects.models.versions.create to specify your version of Python when creating a model version. The default is Python 2.7.
• For details of all available packages in AI Platform, see the runtime version list.

The AI Platform runtime version 1.6 is now available for training and prediction. This version supports TensorFlow 1.6 and other packages as listed in the runtime version list.

The AI Platform runtime version for TensorFlow 1.5 is now available for training and prediction. For more information, see the Runtime Version List.

Added new features for hyperparameter tuning: automated early stopping of trials, resuming a previous hyperparameter tuning job, and additional efficiency optimizations when you run similar jobs. For more information, See the hyperparameter tuning overview.

The AI Platform runtime version for TensorFlow 1.4 is now available for training and prediction. For more information, see the Runtime Version List.

Python 3 is now available for training as part of the AI Platform runtime version for TensorFlow 1.4. For more information, see the Runtime Version List.

Online prediction is now generally available for single core serving. See the guide to online prediction and the blog post.

Pricing has been reduced and simplified for both training and prediction. See the pricing details, the blog post, and the comparison of old and current prices in the pricing FAQ.

P100 GPUs are now in Beta. Using P100 GPUs now incurs charges. For more information, see Using GPUs and Pricing.

Audit logging for AI Platform is now in Beta. For more information, see Viewing Audit Logs.

Predefined IAM roles for AI Platform are available for general use. For more information, see Access Control.

The AI Platform runtime version for TensorFlow 1.2 is now available for training and prediction. For more information, see the Runtime Version List.

Announced general availability of GPU-enabled machines. For more information, see Using GPUs for Training Models in the Cloud.

GPUs are now available in the us-central1 region. For the full list of regions that support GPUs, see Using GPUs for Training Models in the Cloud.

Announced general availability of AI Platform. Version 1 of AI Platform is available for general use for training models, deploying models, and generating batch predictions. The hyperparameter tuning feature is also available for general use, but online prediction and GPU-enabled machines remain in beta.

Online prediction is now in the Beta launch stage. Its use is now subject to the AI Platform pricing policy, and follows the same pricing formula as batch prediction. While it remains in Beta, online prediction is not intended for use in critical applications.

Online prediction is an Alpha feature. Though AI Platform overall is in its Beta phase, online prediction is still undergoing significant changes to improve performance. You will not be charged for online prediction while it remains in Alpha.

Preprocessing and the rest of the AI Platform SDK are Alpha features. The SDK is undergoing active development to better integrate AI Platform with Apache Beam.