Announcing Determined 0.13.8

We are pleased to announce the 0.13.8 release of the Determined deep learning training platform!

Notable highlights include:

  • Determined now supports Keras models that subclass tf.keras.Model. Subclassing tf.keras.Model is a relatively new way to define Keras models that was introduced in TensorFlow 2.2, and provides a high degree of customization over what happens during model training. Model code that uses this technique is now supported in Determined.
  • Multiple fixes for issues with stateful Keras callbacks.
  • Support for how trial logs are managed in the system is being overhauled. The first set of changes have landed in this release, including the ability to filter trial logs by selected criteria.
  • When using det-deploy aws, users can specify which subnet to launch agents in, which effectively provides control over which availability zone is used.
  • The master can now use TLS certificates that are valid for a different hostname than the hostname that the agents use to connect to it. This ability is useful in situations where the master is accessed using multiple different addresses (e.g., private and public IP addresses of a cloud instance).

Complete release notes can be found here.

As always, we encourage you to install the new version. Please refer to our documentation for instructions on how to install the platform and associated CLI.

We welcome feedback in our community Slack channel, on GitHub, or on our mailing list.