Deploying a model
Once a model has completed training, you can deploy it in production. Navigate to the Production section and click on the Create a new deployment button to open the wizard for deploying a model.
You must start out by entering a suitable name for your model. Choose a name that makes it easy to remember the purpose of the deployment in the future.
You also have to choose whether the deployment should be an online deployment or an offline deployment. If you choose an online deployment, the model will be deployed in Google Cloud ML Engine and you will be able to interact with it using a RESTful interface. If you choose an offline deployment, the model will not be deployed in Google Cloud ML Engine. Rather, you will be able to download the model to run in an offline setting (e.g., using TensorFlow Serving on an edge computing device).
In this example, we will choose an online deployment.
Second, you must choose the model you wish to deploy. Only models that have completed training and are eligible for deployment will be visible in the dropdown menu.
Finally, you will see an overview of the information you have provided for the deployment. Click on Create deployment if you are satisfied with the summary and wish to create the deployment.
Creating the deployment can take a couple of minutes. This is due to the fact that copy of the model will be created and made immutable so that it cannot be changed in the future. A deployment can have multiple versions and each version will be an immutable copy of the model at the specific point in time when the version was created.
On the deployment details page you can see an overview of the deployment's versions. You will also be able to download each of them and run them in an offline setting (e.g., using TensorFlow Serving).
An online deployment takes a little longer to create than an offline deployment, as creating the model in Google Cloud ML Engine takes an additional couple of minutes. Check back after a little while to see whether the creation of the model has completed.
Once the Ready field of the version changes to Yes, the model has been deployed successfully.
This concludes the article on deployment of models.