Using TensorFlow Models

To provide increased scalability, flexibility, and ease of use for machine learning, Vertica supports importing and exporting TensorFlow models, and using TensorFlow models to predict on data in Vertica.

TensorFlow is one of the most popular neural network and deep learning frameworks, and an increasing number of data scientists are using it. Many Vertica customers and prospects have either already started exploring it or plan to do so in the near future. Its capability to run on GPUs to train neural networks can be a great fit for many users, but these users still need to score their data in Vertica. Having the capability to import models from TensorFlow and run scoring on large data sets in Vertica provides users with much more analytic capability in cases where speed of prediction and visualization or dashboarding are extremely important.

Some of the advantages of using TensorFlow with Vertica are:

  • You can import TensorFlow models into Vertica.
  • You can run the model to do scoring on data inside Vertica without having to move your data.
  • You can manage all your imported TensorFlow models, just as you can manage Vertica models.

TensorFlow is a framework for creating neural networks. The Tensorflow library implements basic linear algebra and multi-variable calculus operations in a scalable fashion, and allows users to easily chain these operations into a "computation graph".

When you run a TensorFlow model to predict on data in the database, Vertica calls a TensorFlow process to run the model. This allows Vertica to support any model you can create and train using TensorFlow. Vertica just provides the inputs - your data in the Vertica database - and stores the outputs.

Getting Started with TensorFlow Models in Vertica

See Setting up TensorFlow Support in Vertica for the prerequisites and steps to set up your Vertica database for working with TensorFlow models, and an example that you can try yourself. The example shows how to import a TensorFlow model, use it to score the data stored in Vertica, and optionally export the model again.

In This Section