Shortcuts

Native API Tutorial

The Native API allows developers to seamlessly move between training in a local development environment and training at cluster-scale on a Determined cluster. It also provides an interface to train tf.keras and tf.estimator models using idiomatic framework patterns, reducing (or eliminating) the effort to port model code for use with Determined.

This tutorial describes how a minimal tf.keras example can be quickly ported to the Native API, and augmented to launch hyperparameter searches and/or distributed training jobs.

For a more detailed discussion of how the Native API is implemented, refer to the Native API Topic Guide.

Prerequisites

  • You will need access to a Determined cluster to train your model. If you have not yet installed Determined, refer to the installation instructions.

  • The determined Python package should be installed in your local development environment, including TensorFlow: pip install 'determined[tensorflow]'. For more detailed instructions on setting up an IDE like PyCharm for local development with Determined, refer to the IDE Setup How-To.

Gallery generated by Sphinx-Gallery