Determined’s model-hub library makes it easy to train models from popular third-party libraries with a Determined cluster. With model-hub, use trusted implementations of model architectures with Determined’s ability to easily scale to distributed training, track experiments, share resources, and perform hyperparameter searches.
Each supported third-party library in model-hub is accompanied by:
Official examples checked for correctness and thoroughly tested for use with Determined.
A base Determined Trial class with common functionality implemented for the user.
A prebuilt docker environment with all dependencies installed and versioned for reproducibility.
A suite of helper functions (if applicable) to allow users to easily write their own Trial classes for use with the third-party library.
For a given task, deep learning practitioners often adapt existing model implementations from a trusted third-party library, such as HuggingFace Transformers. When beginning your deep learning project in this way, we suggest using model-hub with the following steps:
Check for a Model Hub library that supports model implementations for your task.
If the Model Hub Library includes an official example fit for your task, copy, customize, and deploy it.
If the Model Hub Library does not include example fit for your task, copy the base Determined Trial class and customize it.
For detailed instructions, check out the documentation for your Model Hub library of choice.
Future libraries on our roadmap
Our initial release of model-hub includes support for the Huggingface transformers library. We are actively working on releasing new third-party libraries. Please check back for updates. If you have additional libraries you want to see supported in model-hub please let us know by filing an issue on GitHub or reaching out on our community Slack.
For next steps, learn more about Model Hub Transformers!