This section describes how to use the training APIs and contains the API reference information.
Determined leverages specific APIs for each Deep Learning framework. In general, users convert their existing training code by subclassing a Trial class and implementing methods that advertise components of the user’s model - e.g., model architecture, data loader, optimizer, learning rate scheduler, callbacks, etc. This is called the Trial definition and by structuring their code in this way, Determined is able to run the training loop and provide advanced training and model management capabilities.
Once users’ model code are ported to Determined’s APIs, they can use an experiment configuration to configure how Determined should train the model - e.g., multi-GPU, hyperparameter search, etc.
If you have existing model code that you’d like to train with Determined, continue to one of the API docs below depending on your ML Framework.
If you’d like a review of implementing the Determined APIs on simple models, please take a look at our Tutorials. Or, if you’d like to build off of an existing model that already runs on Determined, take a look at our examples to see if the model you’d like to train is already available.
TensorFlow Core Models¶
Determined has support for TensorFlow models that use the Keras or Estimator APIs. For models that use the low-level TensorFlow Core APIs, we recommend wrapping your model in Keras, as recommended by the official TensorFlow documentation.
TensorFlow 1 vs 2¶
Determined supports both TensorFlow 1 and 2. The version of TensorFlow that is used for a particular experiment is controlled by the container image that has been configured for that experiment. Determined provides prebuilt Docker images that include TensorFlow 2.8, 1.15, 2.5, 2.6, and 2.7, respectively:
We also provide lightweight CPU-only counterparts:
To change the container image used for an experiment, specify environment.image in the experiment configuration file. Please see Container Images for more details about configuring training environments and a more complete list of prebuilt Docker images.
AMD ROCm Support¶
Determined has experimental support for ROCm. Determined provides a prebuilt Docker image that includes ROCm 4.2, PyTorch 1.9 and Tensorflow 2.5:
Only agent-based deployments are available; Kubernetes is not yet supported.
GPU profiling is not yet supported.