Once you've selected the features, transform them as needed, chosen your model architecture, applied any regularization as necessary to ensure good performance, trained your model, and iterated through this process a couple of times, it's time to serve the model for a prediction. Of course, making individual predictions is not really realistic, because we can't expect client code to have our model object in memory. For others to use our trained model, or to save or export our model to a file and then expect client code to instantiate the model from that exported file. We'll export the model to a TensorFlow SavedModel format. Once we have a model in this format, we have lots of ways to serve the model, web application, code like JavaScript from a mobile application, etc. Let's learn a little bit more about SavedModel. SavedModel, it's a universal serialization format for TensorFlow models. SavedModel provides a language-neutral format to save your Machine Learning models that is both recoverable and hermetic. It enables higher-level systems and tools to produce, consume, and transform your TensorFlow models. The resulting SavedModel is then subtle. Models saved in this format can be restored using the tf.keras.models.load_model and they're compatible with TensorFlow Serving. One of the ways that we can serve the model, is to utilize the Cloud AI Platform managed service. The AI platform service also performs scale to training, but for now, we going to be focusing on just serving a trained model. You can start by creating a model object in AI Platform. Let's give ours a creative name, I just decided to call it propertyprice. Again with the house price prediction example here. Next, and this is really important. We're going to create a version for our model, we'll call this model version dnn, because it's going to be a deep neural network. You can also utilize a timestamp or another differentiator in case you have multiple versions of the same model type. Now once the version and the model are created, you can just run this command, to push the model to the Cloud. Remember to point to the output directory in which the saved model was actually saved to. The command to push the model also takes other flags, such as Python and TensorFlow run-time versions, the framework in case you're using something like scikit-learn or XGBoost. That flag here defaults to TensorFlow and a bucket in which to save the staging training archives. A staging bucket is only required if a file upload is necessary. That is, other flags include local paths. Once the model is created and then pushed to AI Platform, you can just use this command, gcloud ai-platform predict, and that will actually do the performing your prediction. Make sure that the flags include the model name, its version, and the path to a file containing the examples that you want to get predictions on.